From 234a566b810ebf289ae4c6f082fff1f0b30bc21d Mon Sep 17 00:00:00 2001 From: Changming Sun Date: Wed, 13 Nov 2024 22:16:40 +0000 Subject: [PATCH] cleanup code --- cmake/CMakeLists.txt | 54 ---- cmake/onnxruntime.cmake | 1 - cmake/onnxruntime_codegen_tvm.cmake | 25 -- cmake/onnxruntime_providers.cmake | 7 - cmake/onnxruntime_providers_tvm.cmake | 64 ---- cmake/onnxruntime_python.cmake | 33 -- cmake/onnxruntime_unittests.cmake | 15 - onnxruntime/core/codegen/common/common.cc | 284 ---------------- onnxruntime/core/codegen/common/common.h | 153 --------- onnxruntime/core/codegen/common/creator.h | 76 ----- onnxruntime/core/codegen/common/dispatcher.h | 76 ----- onnxruntime/core/codegen/common/dump_array.h | 62 ---- onnxruntime/core/codegen/common/handle.h | 22 -- onnxruntime/core/codegen/common/op_macro.h | 101 ------ onnxruntime/core/codegen/common/profile.h | 37 --- onnxruntime/core/codegen/common/registry.h | 72 ----- onnxruntime/core/codegen/common/settings.cc | 78 ----- onnxruntime/core/codegen/common/settings.h | 40 --- onnxruntime/core/codegen/common/target_info.h | 33 -- onnxruntime/core/codegen/common/utils.cc | 99 ------ onnxruntime/core/codegen/common/utils.h | 45 --- onnxruntime/core/codegen/mti/common.h | 16 - .../core/codegen/mti/debug/tvm_print.cc | 83 ----- .../core/codegen/mti/debug/tvm_print.h | 19 -- .../core/codegen/mti/math/binary_ops.cc | 70 ---- .../core/codegen/mti/math/binary_ops.h | 42 --- onnxruntime/core/codegen/mti/math/gemm.cc | 30 -- onnxruntime/core/codegen/mti/math/gemm.h | 16 - .../core/codegen/mti/math/logsoftmax.cc | 18 -- .../core/codegen/mti/math/logsoftmax.h | 11 - .../core/codegen/mti/math/matmul_ops.cc | 161 ---------- .../core/codegen/mti/math/matmul_ops.h | 23 -- .../core/codegen/mti/math/reduce_ops.cc | 90 ------ .../core/codegen/mti/math/reduce_ops.h | 72 ----- onnxruntime/core/codegen/mti/math/softmax.cc | 18 -- onnxruntime/core/codegen/mti/math/softmax.h | 11 - .../core/codegen/mti/math/unary_ops.cc | 155 --------- onnxruntime/core/codegen/mti/math/unary_ops.h | 36 --- onnxruntime/core/codegen/mti/mti_tvm_utils.cc | 203 ------------ onnxruntime/core/codegen/mti/mti_tvm_utils.h | 71 ---- onnxruntime/core/codegen/mti/nn/conv_ops.cc | 193 ----------- onnxruntime/core/codegen/mti/nn/conv_ops.h | 39 --- onnxruntime/core/codegen/mti/nn/lstm.cc | 140 -------- onnxruntime/core/codegen/mti/nn/lstm.h | 35 -- onnxruntime/core/codegen/mti/nn/pool_ops.cc | 63 ---- onnxruntime/core/codegen/mti/nn/pool_ops.h | 36 --- .../core/codegen/mti/tensor/cast_ops.cc | 37 --- .../core/codegen/mti/tensor/cast_ops.h | 15 - .../core/codegen/mti/tensor/concat_ops.cc | 83 ----- .../core/codegen/mti/tensor/concat_ops.h | 15 - onnxruntime/core/codegen/mti/tensor/crop.cc | 58 ---- onnxruntime/core/codegen/mti/tensor/crop.h | 17 - onnxruntime/core/codegen/mti/tensor/expand.cc | 30 -- onnxruntime/core/codegen/mti/tensor/expand.h | 14 - onnxruntime/core/codegen/mti/tensor/gather.cc | 55 ---- onnxruntime/core/codegen/mti/tensor/gather.h | 17 - .../codegen/mti/tensor/gather_elements.cc | 45 --- .../core/codegen/mti/tensor/gather_elements.h | 17 - .../core/codegen/mti/tensor/pad_ops.cc | 121 ------- onnxruntime/core/codegen/mti/tensor/pad_ops.h | 34 -- .../core/codegen/mti/tensor/reshape_ops.cc | 48 --- .../core/codegen/mti/tensor/reshape_ops.h | 16 - .../core/codegen/mti/tensor/shape_op.cc | 25 -- .../core/codegen/mti/tensor/shape_op.h | 14 - onnxruntime/core/codegen/mti/tensor/slice.cc | 91 ------ onnxruntime/core/codegen/mti/tensor/slice.h | 19 -- onnxruntime/core/codegen/mti/tensor/split.cc | 72 ----- onnxruntime/core/codegen/mti/tensor/split.h | 25 -- onnxruntime/core/codegen/mti/tensor/tile.cc | 40 --- onnxruntime/core/codegen/mti/tensor/tile.h | 16 - .../core/codegen/mti/tensor/transpose.cc | 16 - .../core/codegen/mti/tensor/transpose.h | 16 - onnxruntime/core/codegen/mti/tensor/where.cc | 36 --- onnxruntime/core/codegen/mti/tensor/where.h | 17 - .../codegen/passes/op_ir_creator/all_ops.h | 47 --- .../passes/op_ir_creator/math/binary_ops.cc | 46 --- .../codegen/passes/op_ir_creator/math/clip.cc | 48 --- .../codegen/passes/op_ir_creator/math/gemm.cc | 39 --- .../passes/op_ir_creator/math/logsoftmax.cc | 32 -- .../passes/op_ir_creator/math/matmul.cc | 23 -- .../math/quantize/matmul_integer.cc | 37 --- .../passes/op_ir_creator/math/reduce_ops.cc | 111 ------- .../passes/op_ir_creator/math/softmax.cc | 32 -- .../passes/op_ir_creator/math/unary_funcs.h | 51 --- .../passes/op_ir_creator/math/unary_ops.cc | 93 ------ .../passes/op_ir_creator/math/variadic_ops.cc | 36 --- .../codegen/passes/op_ir_creator/nn/conv.cc | 131 -------- .../codegen/passes/op_ir_creator/nn/lstm.cc | 64 ---- .../passes/op_ir_creator/nn/pool_ops.cc | 51 --- .../passes/op_ir_creator/tensor/cast.cc | 40 --- .../passes/op_ir_creator/tensor/concat.cc | 30 -- .../passes/op_ir_creator/tensor/crop.cc | 46 --- .../passes/op_ir_creator/tensor/expand.cc | 26 -- .../passes/op_ir_creator/tensor/gather.cc | 30 -- .../op_ir_creator/tensor/gather_elements.cc | 32 -- .../passes/op_ir_creator/tensor/pad.cc | 49 --- .../op_ir_creator/tensor/reshape_ops.cc | 99 ------ .../passes/op_ir_creator/tensor/shape_op.cc | 26 -- .../passes/op_ir_creator/tensor/slice.cc | 71 ---- .../passes/op_ir_creator/tensor/split.cc | 66 ---- .../passes/op_ir_creator/tensor/transpose.cc | 48 --- .../passes/op_ir_creator/tensor/where.cc | 28 -- .../passes/op_ir_creator/tvm_ir_builder.cc | 125 ------- .../passes/op_ir_creator/tvm_ir_builder.h | 64 ---- .../passes/op_ir_creator/tvm_op_creator.cc | 37 --- .../passes/op_ir_creator/tvm_op_creator.h | 84 ----- .../codegen/passes/scheduler/all_schedules.h | 20 -- .../passes/scheduler/ort_type_schedule.cc | 22 -- .../passes/scheduler/schedule_utils.cc | 178 ---------- .../codegen/passes/scheduler/schedule_utils.h | 65 ---- .../passes/scheduler/tvm_rule_schedule.cc | 41 --- .../passes/scheduler/tvm_schedule_builder.cc | 104 ------ .../passes/scheduler/tvm_schedule_builder.h | 46 --- .../codegen/passes/scheduler/tvm_scheduler.cc | 79 ----- .../codegen/passes/scheduler/tvm_scheduler.h | 128 -------- .../codegen/passes/utils/codegen_context.cc | 27 -- .../codegen/passes/utils/codegen_context.h | 44 --- .../codegen/passes/utils/ort_tvm_utils.cc | 194 ----------- .../core/codegen/passes/utils/ort_tvm_utils.h | 31 -- .../codegen/passes/weight_layout/tiling_2d.cc | 105 ------ .../codegen/passes/weight_layout/tiling_2d.h | 43 --- .../passes/weight_layout/transpose_2d.cc | 64 ---- .../passes/weight_layout/transpose_2d.h | 33 -- .../weight_layout/vertical_stripes_2d.cc | 77 ----- .../weight_layout/vertical_stripes_2d.h | 40 --- .../passes/weight_layout/weight_layout.cc | 91 ------ .../passes/weight_layout/weight_layout.h | 68 ---- onnxruntime/core/framework/utils.cc | 1 - .../core/platform/windows/stacktrace.cc | 1 - .../core/providers/get_execution_providers.cc | 8 - .../providers/provider_factory_creators.h | 4 - .../core/providers/tvm/custom_logging.cc | 52 --- .../core/providers/tvm/hash_alg/hasher.cc | 30 -- .../core/providers/tvm/hash_alg/hasher.h | 32 -- .../providers/tvm/hash_alg/hasher_impl.cc | 39 --- .../core/providers/tvm/hash_alg/hasher_impl.h | 42 --- onnxruntime/core/providers/tvm/symbols.txt | 1 - .../core/providers/tvm/tvm_allocator.cc | 29 -- .../core/providers/tvm/tvm_allocator.h | 45 --- onnxruntime/core/providers/tvm/tvm_api.cc | 303 ----------------- onnxruntime/core/providers/tvm/tvm_api.h | 38 --- onnxruntime/core/providers/tvm/tvm_common.h | 22 -- .../core/providers/tvm/tvm_compiler.cc | 48 --- onnxruntime/core/providers/tvm/tvm_compiler.h | 63 ---- onnxruntime/core/providers/tvm/tvm_defaults.h | 35 -- .../core/providers/tvm/tvm_ep_options.cc | 273 ---------------- .../core/providers/tvm/tvm_ep_options.h | 76 ----- .../providers/tvm/tvm_execution_provider.cc | 304 ------------------ .../providers/tvm/tvm_execution_provider.h | 71 ---- .../providers/tvm/tvm_provider_factory.cc | 53 --- .../tvm/tvm_provider_factory_creator.h | 19 -- onnxruntime/core/providers/tvm/tvm_runner.cc | 26 -- onnxruntime/core/providers/tvm/tvm_runner.h | 34 -- .../core/providers/tvm/tvm_runner_impl.cc | 175 ---------- .../core/providers/tvm/tvm_runner_impl.h | 126 -------- .../tvm/tvm_so_execution_provider.cc | 284 ---------------- .../providers/tvm/tvm_so_execution_provider.h | 72 ----- onnxruntime/core/providers/tvm/tvm_utils.cc | 31 -- onnxruntime/core/providers/tvm/tvm_utils.h | 70 ---- .../core/providers/tvm/xpu_data_transfer.cc | 84 ----- .../core/providers/tvm/xpu_data_transfer.h | 38 --- .../core/session/provider_registration.cc | 9 - .../python/onnxruntime_pybind_state.cc | 10 - .../python/onnxruntime_pybind_state_common.h | 15 +- onnxruntime/python/providers/tvm/__init__.py | 10 - .../providers/tvm/extend_python_file.py | 54 ---- onnxruntime/python/providers/tvm/ort.py | 140 -------- onnxruntime/test/framework/function_test.cc | 6 - .../test/platform/windows/stacktrace_test.cc | 1 - .../python/onnxruntime_test_python_tvm.py | 242 -------------- onnxruntime/test/util/default_providers.cc | 8 - .../test/util/include/default_providers.h | 1 - onnxruntime/test/util/include/providers.h | 3 - tools/ci_build/build.py | 62 +--- ...orttraining-linux-gpu-test-ci-pipeline.yml | 41 --- ...py-packaging-training-cuda-stage-steps.yml | 209 ------------ .../linux/docker/scripts/install_os_deps.sh | 1 - .../docker/scripts/install_python_deps.sh | 16 - .../stage1/requirements_rocm/requirements.txt | 2 - .../requirements.txt | 8 - .../requirements.txt | 7 - .../requirements_torch_cpu/requirements.txt | 3 - .../requirements.txt | 5 - .../stage1/torch_eager_cpu/requirements.txt | 11 - .../ortmodule/stage2/requirements.txt | 15 - .../ci_build/github/linux/run_dockerbuild.sh | 26 -- .../pai/pai_huggingface_bert_large_test.sh | 43 --- tools/ci_build/set-trigger-rules.py | 5 - .../nuget/generate_nuspec_for_native_nuget.py | 51 +-- tools/scripts/python_test.sh | 8 +- 190 files changed, 5 insertions(+), 10815 deletions(-) delete mode 100644 cmake/onnxruntime_codegen_tvm.cmake delete mode 100644 cmake/onnxruntime_providers_tvm.cmake delete mode 100644 onnxruntime/core/codegen/common/common.cc delete mode 100644 onnxruntime/core/codegen/common/common.h delete mode 100644 onnxruntime/core/codegen/common/creator.h delete mode 100644 onnxruntime/core/codegen/common/dispatcher.h delete mode 100644 onnxruntime/core/codegen/common/dump_array.h delete mode 100644 onnxruntime/core/codegen/common/handle.h delete mode 100644 onnxruntime/core/codegen/common/op_macro.h delete mode 100644 onnxruntime/core/codegen/common/profile.h delete mode 100644 onnxruntime/core/codegen/common/registry.h delete mode 100644 onnxruntime/core/codegen/common/settings.cc delete mode 100644 onnxruntime/core/codegen/common/settings.h delete mode 100644 onnxruntime/core/codegen/common/target_info.h delete mode 100644 onnxruntime/core/codegen/common/utils.cc delete mode 100644 onnxruntime/core/codegen/common/utils.h delete mode 100644 onnxruntime/core/codegen/mti/common.h delete mode 100644 onnxruntime/core/codegen/mti/debug/tvm_print.cc delete mode 100644 onnxruntime/core/codegen/mti/debug/tvm_print.h delete mode 100644 onnxruntime/core/codegen/mti/math/binary_ops.cc delete mode 100644 onnxruntime/core/codegen/mti/math/binary_ops.h delete mode 100644 onnxruntime/core/codegen/mti/math/gemm.cc delete mode 100644 onnxruntime/core/codegen/mti/math/gemm.h delete mode 100644 onnxruntime/core/codegen/mti/math/logsoftmax.cc delete mode 100644 onnxruntime/core/codegen/mti/math/logsoftmax.h delete mode 100644 onnxruntime/core/codegen/mti/math/matmul_ops.cc delete mode 100644 onnxruntime/core/codegen/mti/math/matmul_ops.h delete mode 100644 onnxruntime/core/codegen/mti/math/reduce_ops.cc delete mode 100644 onnxruntime/core/codegen/mti/math/reduce_ops.h delete mode 100644 onnxruntime/core/codegen/mti/math/softmax.cc delete mode 100644 onnxruntime/core/codegen/mti/math/softmax.h delete mode 100644 onnxruntime/core/codegen/mti/math/unary_ops.cc delete mode 100644 onnxruntime/core/codegen/mti/math/unary_ops.h delete mode 100644 onnxruntime/core/codegen/mti/mti_tvm_utils.cc delete mode 100644 onnxruntime/core/codegen/mti/mti_tvm_utils.h delete mode 100644 onnxruntime/core/codegen/mti/nn/conv_ops.cc delete mode 100644 onnxruntime/core/codegen/mti/nn/conv_ops.h delete mode 100644 onnxruntime/core/codegen/mti/nn/lstm.cc delete mode 100644 onnxruntime/core/codegen/mti/nn/lstm.h delete mode 100644 onnxruntime/core/codegen/mti/nn/pool_ops.cc delete mode 100644 onnxruntime/core/codegen/mti/nn/pool_ops.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/cast_ops.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/cast_ops.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/concat_ops.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/concat_ops.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/crop.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/crop.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/expand.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/expand.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/gather.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/gather.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/gather_elements.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/gather_elements.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/pad_ops.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/pad_ops.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/reshape_ops.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/reshape_ops.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/shape_op.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/shape_op.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/slice.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/slice.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/split.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/split.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/tile.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/tile.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/transpose.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/transpose.h delete mode 100644 onnxruntime/core/codegen/mti/tensor/where.cc delete mode 100644 onnxruntime/core/codegen/mti/tensor/where.h delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/all_ops.h delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/binary_ops.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/clip.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/gemm.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/logsoftmax.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/matmul.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/quantize/matmul_integer.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/reduce_ops.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/softmax.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/unary_funcs.h delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/unary_ops.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/math/variadic_ops.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/nn/conv.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/nn/lstm.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/nn/pool_ops.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/cast.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/concat.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/crop.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/expand.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/gather.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/gather_elements.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/pad.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/reshape_ops.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/shape_op.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/slice.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/split.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/transpose.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tensor/where.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tvm_ir_builder.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tvm_ir_builder.h delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tvm_op_creator.cc delete mode 100644 onnxruntime/core/codegen/passes/op_ir_creator/tvm_op_creator.h delete mode 100644 onnxruntime/core/codegen/passes/scheduler/all_schedules.h delete mode 100644 onnxruntime/core/codegen/passes/scheduler/ort_type_schedule.cc delete mode 100644 onnxruntime/core/codegen/passes/scheduler/schedule_utils.cc delete mode 100644 onnxruntime/core/codegen/passes/scheduler/schedule_utils.h delete mode 100644 onnxruntime/core/codegen/passes/scheduler/tvm_rule_schedule.cc delete mode 100644 onnxruntime/core/codegen/passes/scheduler/tvm_schedule_builder.cc delete mode 100644 onnxruntime/core/codegen/passes/scheduler/tvm_schedule_builder.h delete mode 100644 onnxruntime/core/codegen/passes/scheduler/tvm_scheduler.cc delete mode 100644 onnxruntime/core/codegen/passes/scheduler/tvm_scheduler.h delete mode 100644 onnxruntime/core/codegen/passes/utils/codegen_context.cc delete mode 100644 onnxruntime/core/codegen/passes/utils/codegen_context.h delete mode 100644 onnxruntime/core/codegen/passes/utils/ort_tvm_utils.cc delete mode 100644 onnxruntime/core/codegen/passes/utils/ort_tvm_utils.h delete mode 100644 onnxruntime/core/codegen/passes/weight_layout/tiling_2d.cc delete mode 100644 onnxruntime/core/codegen/passes/weight_layout/tiling_2d.h delete mode 100644 onnxruntime/core/codegen/passes/weight_layout/transpose_2d.cc delete mode 100644 onnxruntime/core/codegen/passes/weight_layout/transpose_2d.h delete mode 100644 onnxruntime/core/codegen/passes/weight_layout/vertical_stripes_2d.cc delete mode 100644 onnxruntime/core/codegen/passes/weight_layout/vertical_stripes_2d.h delete mode 100644 onnxruntime/core/codegen/passes/weight_layout/weight_layout.cc delete mode 100644 onnxruntime/core/codegen/passes/weight_layout/weight_layout.h delete mode 100644 onnxruntime/core/providers/tvm/custom_logging.cc delete mode 100644 onnxruntime/core/providers/tvm/hash_alg/hasher.cc delete mode 100644 onnxruntime/core/providers/tvm/hash_alg/hasher.h delete mode 100644 onnxruntime/core/providers/tvm/hash_alg/hasher_impl.cc delete mode 100644 onnxruntime/core/providers/tvm/hash_alg/hasher_impl.h delete mode 100644 onnxruntime/core/providers/tvm/symbols.txt delete mode 100644 onnxruntime/core/providers/tvm/tvm_allocator.cc delete mode 100644 onnxruntime/core/providers/tvm/tvm_allocator.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_api.cc delete mode 100644 onnxruntime/core/providers/tvm/tvm_api.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_common.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_compiler.cc delete mode 100644 onnxruntime/core/providers/tvm/tvm_compiler.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_defaults.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_ep_options.cc delete mode 100644 onnxruntime/core/providers/tvm/tvm_ep_options.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_execution_provider.cc delete mode 100644 onnxruntime/core/providers/tvm/tvm_execution_provider.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_provider_factory.cc delete mode 100644 onnxruntime/core/providers/tvm/tvm_provider_factory_creator.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_runner.cc delete mode 100644 onnxruntime/core/providers/tvm/tvm_runner.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_runner_impl.cc delete mode 100644 onnxruntime/core/providers/tvm/tvm_runner_impl.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_so_execution_provider.cc delete mode 100644 onnxruntime/core/providers/tvm/tvm_so_execution_provider.h delete mode 100644 onnxruntime/core/providers/tvm/tvm_utils.cc delete mode 100644 onnxruntime/core/providers/tvm/tvm_utils.h delete mode 100644 onnxruntime/core/providers/tvm/xpu_data_transfer.cc delete mode 100644 onnxruntime/core/providers/tvm/xpu_data_transfer.h delete mode 100644 onnxruntime/python/providers/tvm/__init__.py delete mode 100644 onnxruntime/python/providers/tvm/extend_python_file.py delete mode 100644 onnxruntime/python/providers/tvm/ort.py delete mode 100644 onnxruntime/test/python/onnxruntime_test_python_tvm.py delete mode 100644 tools/ci_build/github/azure-pipelines/templates/orttraining-linux-gpu-test-ci-pipeline.yml delete mode 100644 tools/ci_build/github/azure-pipelines/templates/py-packaging-training-cuda-stage-steps.yml delete mode 100644 tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_rocm/requirements.txt delete mode 100644 tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch2.0.0_cu11.8/requirements.txt delete mode 100644 tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch2.1.0_cu12.2/requirements.txt delete mode 100644 tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch_cpu/requirements.txt delete mode 100644 tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch_nightly/requirements.txt delete mode 100644 tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/torch_eager_cpu/requirements.txt delete mode 100644 tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage2/requirements.txt delete mode 100755 tools/ci_build/github/pai/pai_huggingface_bert_large_test.sh diff --git a/cmake/CMakeLists.txt b/cmake/CMakeLists.txt index 31ebf58b03152..af341aaead2d5 100644 --- a/cmake/CMakeLists.txt +++ b/cmake/CMakeLists.txt @@ -102,7 +102,6 @@ option(onnxruntime_BUILD_CSHARP "Build C# library" OFF) option(onnxruntime_BUILD_OBJC "Build Objective-C library" OFF) option(onnxruntime_USE_PREINSTALLED_EIGEN "Use pre-installed EIGEN. Need to provide eigen_SOURCE_PATH if turn this on." OFF) option(onnxruntime_BUILD_BENCHMARKS "Build ONNXRuntime micro-benchmarks" OFF) -option(onnxruntime_USE_LLVM "Build TVM with LLVM" OFF) option(onnxruntime_USE_VSINPU "Build with VSINPU support" OFF) cmake_dependent_option(onnxruntime_USE_FLASH_ATTENTION "Build flash attention kernel for scaled dot product attention" ON "onnxruntime_USE_CUDA" OFF) @@ -145,10 +144,6 @@ option(onnxruntime_USE_TELEMETRY "Build with Telemetry" OFF) cmake_dependent_option(onnxruntime_USE_MIMALLOC "Override new/delete and arena allocator with mimalloc" OFF "WIN32;NOT onnxruntime_USE_CUDA;NOT onnxruntime_USE_OPENVINO" OFF) option(onnxruntime_USE_CANN "Build with CANN support" OFF) option(onnxruntime_USE_ROCM "Build with AMD GPU support" OFF) -option(onnxruntime_USE_TVM "Build with TVM support" OFF) -option(onnxruntime_TVM_CUDA_RUNTIME "Build TVM with CUDA support" OFF) -option(onnxruntime_TVM_USE_LLVM "Build TVM with LLVM. Set customized path to llvm-config.exe here if need" OFF) -option(onnxruntime_TVM_USE_HASH "Build ipp-crypto library for support hash algorithm. It is defined for TVM only") option(onnxruntime_USE_XNNPACK "Build with XNNPACK support. Provides an alternative math library on ARM, WebAssembly and x86." OFF) option(onnxruntime_USE_WEBNN "Build with WebNN support. Enable hardware acceleration in web browsers." OFF) option(onnxruntime_USE_WEBGPU "Build with WebGPU support. Enable WebGPU via C/C++ interface." OFF) @@ -906,11 +901,6 @@ if (onnxruntime_USE_SNPE) list(APPEND ONNXRUNTIME_PROVIDER_NAMES snpe) list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_SNPE=1) endif() -if (onnxruntime_USE_TVM) - list(APPEND ORT_PROVIDER_FLAGS -DUSE_TVM=1) - list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_TVM=1) - list(APPEND ONNXRUNTIME_PROVIDER_NAMES tvm) -endif() if (onnxruntime_USE_WINML) list(APPEND ORT_PROVIDER_FLAGS -DUSE_WINML=1) list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_WINML=1) @@ -1313,50 +1303,6 @@ if (onnxruntime_USE_DNNL) add_compile_definitions(DNNL_OPENMP) endif() -# TVM EP -if (onnxruntime_USE_TVM) - if (NOT TARGET tvm) - message(STATUS "Include TVM(*).") - include(tvm) - endif() - - # ipp-crypto - if (onnxruntime_TVM_USE_HASH) - message(STATUS "Include ipp-crypto(*).") - include(ipp-crypto) - endif() - - # TVM - if (onnxruntime_TVM_USE_LLVM) - set(USE_LLVM "${onnxruntime_TVM_USE_LLVM}" CACHE STRING "Path to LLVM for correct TVM build") - elseif(onnxruntime_USE_LLVM) - set(USE_LLVM ON CACHE BOOL "Only defined for TVM") - endif() - - if (onnxruntime_TVM_CUDA_RUNTIME) - set(USE_CUDA ON CACHE BOOL "Only defined for TVM" FORCE) - endif() - - # TODO(vvchernov): customized tvm logger is hidden due to the issue on TVM side (https://github.com/apache/tvm/issues/10139) - # add_compile_definitions(TVM_LOG_CUSTOMIZE=1) - # add_library(tvm_custom_logger STATIC ${ONNXRUNTIME_ROOT}/core/providers/tvm/custom_logging.cc) - - set(USE_OPENMP gnu CACHE STRING "Only defined for TVM") - add_subdirectory(${tvm_SOURCE_DIR} ${tvm_BINARY_DIR} EXCLUDE_FROM_ALL) - - set_target_properties(tvm PROPERTIES FOLDER ${tvm_SOURCE_DIR}) - # target_link_libraries(tvm PUBLIC tvm_custom_logger) - - set(TVM_INCLUDES ${tvm_SOURCE_DIR}/include - ${tvm_SOURCE_DIR}/3rdparty/dmlc-core/include - ${tvm_SOURCE_DIR}/3rdparty/dlpack/include - $) - - set(onnxruntime_tvm_libs onnxruntime_providers_tvm) - list(APPEND onnxruntime_EXTERNAL_LIBRARIES tvm) - list(APPEND onnxruntime_EXTERNAL_DEPENDENCIES tvm) -endif() - # onnxruntime-extensions if (onnxruntime_USE_EXTENSIONS) include(extensions) diff --git a/cmake/onnxruntime.cmake b/cmake/onnxruntime.cmake index 003f7ad18286b..cb169c60eb3ca 100644 --- a/cmake/onnxruntime.cmake +++ b/cmake/onnxruntime.cmake @@ -221,7 +221,6 @@ set(onnxruntime_INTERNAL_LIBRARIES ${onnxruntime_winml} onnxruntime_optimizer onnxruntime_providers - ${onnxruntime_tvm_libs} onnxruntime_lora onnxruntime_framework onnxruntime_graph diff --git a/cmake/onnxruntime_codegen_tvm.cmake b/cmake/onnxruntime_codegen_tvm.cmake deleted file mode 100644 index 7b50d8f8603ae..0000000000000 --- a/cmake/onnxruntime_codegen_tvm.cmake +++ /dev/null @@ -1,25 +0,0 @@ -# Copyright (c) Microsoft Corporation. All rights reserved. -# Licensed under the MIT License. - -file(GLOB_RECURSE onnxruntime_codegen_common_srcs - "${ONNXRUNTIME_ROOT}/core/codegen/common/*.h" - "${ONNXRUNTIME_ROOT}/core/codegen/common/*.cc" -) - -file(GLOB_RECURSE onnxruntime_codegen_tvm_srcs CONFIGURE_DEPENDS - "${ONNXRUNTIME_ROOT}/core/codegen/mti/*.h" - "${ONNXRUNTIME_ROOT}/core/codegen/mti/*.cc" - "${ONNXRUNTIME_ROOT}/core/codegen/passes/*.h" - "${ONNXRUNTIME_ROOT}/core/codegen/passes/*.cc" -) - -source_group(TREE ${ONNXRUNTIME_ROOT}/core FILES ${onnxruntime_codegen_common_srcs} ${onnxruntime_codegen_tvm_srcs}) - -#onnxruntime_codegen_tvm depends on onnxruntime framework -onnxruntime_add_static_library(onnxruntime_codegen_tvm ${onnxruntime_codegen_common_srcs} ${onnxruntime_codegen_tvm_srcs}) -set_target_properties(onnxruntime_codegen_tvm PROPERTIES FOLDER "ONNXRuntime") -target_include_directories(onnxruntime_codegen_tvm PRIVATE ${ONNXRUNTIME_ROOT} ${TVM_INCLUDES} ${MKLML_INCLUDE_DIR} ${eigen_INCLUDE_DIRS}) -onnxruntime_add_include_to_target(onnxruntime_codegen_tvm onnxruntime_common onnxruntime_framework onnx onnx_proto ${PROTOBUF_LIB} flatbuffers::flatbuffers safeint_interface Boost::mp11) -target_compile_options(onnxruntime_codegen_tvm PRIVATE ${DISABLED_WARNINGS_FOR_TVM}) -# need onnx to build to create headers that this project includes -add_dependencies(onnxruntime_codegen_tvm ${onnxruntime_EXTERNAL_DEPENDENCIES}) diff --git a/cmake/onnxruntime_providers.cmake b/cmake/onnxruntime_providers.cmake index 9666877cdc206..582491de9503d 100644 --- a/cmake/onnxruntime_providers.cmake +++ b/cmake/onnxruntime_providers.cmake @@ -101,9 +101,6 @@ endif() if(onnxruntime_USE_ROCM) set(PROVIDERS_ROCM onnxruntime_providers_rocm) endif() -if (onnxruntime_USE_TVM) - set(PROVIDERS_TVM onnxruntime_providers_tvm) -endif() if (onnxruntime_USE_XNNPACK) set(PROVIDERS_XNNPACK onnxruntime_providers_xnnpack) endif() @@ -194,10 +191,6 @@ if (onnxruntime_USE_ROCM) include(onnxruntime_providers_rocm.cmake) endif() -if (onnxruntime_USE_TVM) - include(onnxruntime_providers_tvm.cmake) -endif() - if (onnxruntime_USE_VSINPU) include(onnxruntime_providers_vsinpu.cmake) endif() diff --git a/cmake/onnxruntime_providers_tvm.cmake b/cmake/onnxruntime_providers_tvm.cmake deleted file mode 100644 index 8fd50c70dd5d7..0000000000000 --- a/cmake/onnxruntime_providers_tvm.cmake +++ /dev/null @@ -1,64 +0,0 @@ -# Copyright (c) Microsoft Corporation. All rights reserved. -# Licensed under the MIT License. - - add_definitions(-DUSE_TVM=1) - if (onnxruntime_TVM_USE_HASH) - add_definitions(-DUSE_TVM_HASH=1) - endif() - - if (onnxruntime_TVM_USE_HASH) - file (GLOB_RECURSE onnxruntime_providers_tvm_cc_srcs CONFIGURE_DEPENDS - "${ONNXRUNTIME_ROOT}/core/providers/tvm/*.h" - "${ONNXRUNTIME_ROOT}/core/providers/tvm/*.cc" - ) - else() - file (GLOB onnxruntime_providers_tvm_cc_srcs CONFIGURE_DEPENDS - "${ONNXRUNTIME_ROOT}/core/providers/tvm/*.h" - "${ONNXRUNTIME_ROOT}/core/providers/tvm/*.cc" - ) - endif() - - source_group(TREE ${ONNXRUNTIME_ROOT}/core FILES ${onnxruntime_providers_tvm_cc_srcs}) - onnxruntime_add_static_library(onnxruntime_providers_tvm ${onnxruntime_providers_tvm_cc_srcs}) - - if ( CMAKE_COMPILER_IS_GNUCC ) - target_compile_options(onnxruntime_providers_tvm PRIVATE -Wno-unused-parameter -Wno-missing-field-initializers) - endif() - - target_include_directories(onnxruntime_providers_tvm PRIVATE - ${TVM_INCLUDES} - ${PYTHON_INCLUDE_DIRS}) - onnxruntime_add_include_to_target(onnxruntime_providers_tvm onnxruntime_common onnxruntime_framework onnx onnx_proto ${PROTOBUF_LIB} flatbuffers::flatbuffers Boost::mp11 safeint_interface) - - add_dependencies(onnxruntime_providers_tvm ${onnxruntime_EXTERNAL_DEPENDENCIES}) - - if (onnxruntime_TVM_USE_HASH) - add_dependencies(onnxruntime_providers_tvm ippcp_s) - target_include_directories(onnxruntime_providers_tvm PRIVATE ${IPP_CRYPTO_INCLUDE_DIR}) - target_link_libraries(onnxruntime_providers_tvm PRIVATE ippcp_s) - endif() - - set_target_properties(onnxruntime_providers_tvm PROPERTIES FOLDER "ONNXRuntime") - set_target_properties(onnxruntime_providers_tvm PROPERTIES LINKER_LANGUAGE CXX) - - if (WIN32 AND MSVC) - # wd4100: identifier' : unreferenced formal parameter - # wd4127: conditional expression is constant - # wd4244: conversion from 'int' to 'char', possible loss of data - # TODO: 4244 should not be disabled - target_compile_options(onnxruntime_providers_tvm PRIVATE "/wd4100" "/wd4127" "/wd4244") - else() - target_compile_options(onnxruntime_providers_tvm PRIVATE "-Wno-error=type-limits") - endif() - target_compile_definitions(onnxruntime_providers_tvm PUBLIC DMLC_USE_LOGGING_LIBRARY=) - - install(FILES ${PROJECT_SOURCE_DIR}/../include/onnxruntime/core/providers/tvm/tvm_provider_factory.h - DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/onnxruntime/) - - if (NOT onnxruntime_BUILD_SHARED_LIB) - install(TARGETS onnxruntime_providers_tvm - ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR} - LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR} - RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR} - FRAMEWORK DESTINATION ${CMAKE_INSTALL_BINDIR}) - endif() \ No newline at end of file diff --git a/cmake/onnxruntime_python.cmake b/cmake/onnxruntime_python.cmake index 7239b245a7245..d2c022e4e0269 100644 --- a/cmake/onnxruntime_python.cmake +++ b/cmake/onnxruntime_python.cmake @@ -169,7 +169,6 @@ endif() target_link_libraries(onnxruntime_pybind11_state PRIVATE onnxruntime_session ${onnxruntime_libs} - ${PROVIDERS_TVM} ${PROVIDERS_NNAPI} ${PROVIDERS_XNNPACK} ${PROVIDERS_COREML} @@ -184,7 +183,6 @@ target_link_libraries(onnxruntime_pybind11_state PRIVATE onnxruntime_optimizer onnxruntime_providers onnxruntime_util - ${onnxruntime_tvm_libs} onnxruntime_lora onnxruntime_framework onnxruntime_util @@ -965,37 +963,6 @@ if (onnxruntime_USE_ROCM) ) endif() -if (onnxruntime_USE_TVM) - file(GLOB onnxruntime_python_providers_tvm_srcs CONFIGURE_DEPENDS - "${ONNXRUNTIME_ROOT}/python/providers/tvm/*.py" - ) - add_custom_command( - TARGET onnxruntime_pybind11_state POST_BUILD - COMMAND ${CMAKE_COMMAND} -E make_directory $/onnxruntime/providers - COMMAND ${CMAKE_COMMAND} -E make_directory $/onnxruntime/providers/tvm - COMMAND ${CMAKE_COMMAND} -E copy - ${onnxruntime_python_providers_tvm_srcs} - $/onnxruntime/providers/tvm - COMMAND ${CMAKE_COMMAND} -E copy - $ - $/onnxruntime/capi/ - ) - - add_custom_command( - TARGET onnxruntime_pybind11_state POST_BUILD - WORKING_DIRECTORY ${tvm_SOURCE_DIR}/python - COMMAND ${Python_EXECUTABLE} setup.py bdist_wheel - ) - - add_custom_command( - TARGET onnxruntime_pybind11_state POST_BUILD - COMMAND ${Python_EXECUTABLE} - $/onnxruntime/providers/tvm/extend_python_file.py - --target_file $/onnxruntime/capi/_ld_preload.py - ) - -endif() - if (onnxruntime_USE_DML) if (NOT onnxruntime_USE_CUSTOM_DIRECTML) set(dml_shared_lib_path ${DML_PACKAGE_DIR}/bin/${onnxruntime_target_platform}-win/${DML_SHARED_LIB}) diff --git a/cmake/onnxruntime_unittests.cmake b/cmake/onnxruntime_unittests.cmake index 561f65a33b89c..f9075d2e92911 100644 --- a/cmake/onnxruntime_unittests.cmake +++ b/cmake/onnxruntime_unittests.cmake @@ -9,9 +9,6 @@ set(TEST_INC_DIR ${ONNXRUNTIME_ROOT}) if (onnxruntime_ENABLE_TRAINING) list(APPEND TEST_INC_DIR ${ORTTRAINING_ROOT}) endif() -if (onnxruntime_USE_TVM) - list(APPEND TEST_INC_DIR ${TVM_INCLUDES}) -endif() set(disabled_warnings) function(AddTest) @@ -111,7 +108,6 @@ function(AddTest) endif() target_compile_options(${_UT_TARGET} PRIVATE ${disabled_warnings}) else() - target_compile_options(${_UT_TARGET} PRIVATE ${DISABLED_WARNINGS_FOR_TVM}) target_compile_options(${_UT_TARGET} PRIVATE "$<$:SHELL:--compiler-options -Wno-error=sign-compare>" "$<$>:-Wno-error=sign-compare>") if (${HAS_NOERROR}) @@ -641,13 +637,11 @@ set(ONNXRUNTIME_TEST_LIBS ${PROVIDERS_ACL} ${PROVIDERS_ARMNN} ${PROVIDERS_COREML} - # ${PROVIDERS_TVM} ${PROVIDERS_XNNPACK} ${PROVIDERS_AZURE} onnxruntime_optimizer onnxruntime_providers onnxruntime_util - ${onnxruntime_tvm_libs} onnxruntime_lora onnxruntime_framework onnxruntime_util @@ -749,12 +743,6 @@ if(onnxruntime_USE_AZURE) list(APPEND onnxruntime_test_providers_libs onnxruntime_providers_azure) endif() -if(WIN32) - if (onnxruntime_USE_TVM) - list(APPEND disabled_warnings ${DISABLED_WARNINGS_FOR_TVM}) - endif() -endif() - file(GLOB onnxruntime_test_framework_src CONFIGURE_DEPENDS ${onnxruntime_test_framework_src_patterns} ) @@ -855,9 +843,6 @@ if (onnxruntime_ENABLE_TRAINING_APIS) list(APPEND all_tests ${onnxruntime_test_training_api_src}) endif() -if (onnxruntime_USE_TVM) - list(APPEND all_tests ${onnxruntime_test_tvm_src}) -endif() if (onnxruntime_USE_OPENVINO) list(APPEND all_tests ${onnxruntime_test_openvino_src}) diff --git a/onnxruntime/core/codegen/common/common.cc b/onnxruntime/core/codegen/common/common.cc deleted file mode 100644 index 818b919e99ef2..0000000000000 --- a/onnxruntime/core/codegen/common/common.cc +++ /dev/null @@ -1,284 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/common/common.h" - -#include "core/framework/tensorprotoutils.h" -#include "core/common/inlined_containers.h" -#include "core/graph/graph.h" -#include "core/graph/schema_registry.h" -#include -#include - -namespace onnxruntime { - -NodeKey GetKey(const onnxruntime::Node* node) { - ORT_ENFORCE(nullptr != node); - ORT_ENFORCE(node->OutputDefs().size() > 0); - return node->OutputDefs()[0]->Name(); -} - -NodeKey GetKey(const onnxruntime::Node& node) { - ORT_ENFORCE(node.OutputDefs().size() > 0); - return node.OutputDefs()[0]->Name(); -} - -NodeKey GetKey(const onnxruntime::NodeArg* def) { - // NodeArg's name is unique. - ORT_ENFORCE(nullptr != def); - return def->Name(); -} - -bool IsRecurrentNode(const onnxruntime::Node& node) { - auto op_type = node.OpType(); - return (op_type == "LSTM" || op_type == "RNN" || op_type == "GRU" || - op_type == "Scan" || op_type == "Loop"); -} - -bool IsAliasNode(const onnxruntime::Node& node) { - auto op_type = node.OpType(); - if (op_type == "Transpose") { - // Treat Transpose (1,N) -> (N,1) as Alias - const auto shape = node.OutputDefs()[0]->Shape(); - if (shape != nullptr && shape->dim_size() == 2) { - for (int i = 0; i < 2; ++i) { - if (shape->dim(i).has_dim_value() && shape->dim(i).dim_value() == 1) { - return true; - } - } - } - return false; - } - - return (op_type == "Flatten" || op_type == "Identity" || op_type == "Reshape" || - op_type == "Squeeze" || op_type == "Unsqueeze"); -} - -std::string NormalizeCppName(const std::string& name) { - std::string normalized_name = name; - for (char c : {'.', ' ', '+', '-', '*', '/', '\\', '='}) - std::replace(normalized_name.begin(), normalized_name.end(), c, '_'); - return normalized_name; -} - -std::string NormalizeNodeArgName(const NodeArg* def) { - return NormalizeCppName(def->Name()); -} - -bool IsFusedNode(const Node& node) { - if (node.NodeType() == Node::Type::Fused) { - return true; - } - return false; -} - -// A unified API to get Subgraph -const Graph* GetSubgraph(const Node& node) { - if (node.NodeType() == Node::Type::Fused) { - return &(node.GetFunctionBody()->Body()); - } else if (node.OpType() == "Scan") { - return node.GetGraphAttribute("body"); - } - // return nullptr implying no subgraph - return nullptr; -} - -bool HasLoop(const Node& node) { - auto op_type = node.OpType(); - if (op_type == "LSTM" || - op_type == "GRU" || - op_type == "RNN" || - op_type == "Scan") { - return true; - } - return false; -} - -// Return the corresponding input node for the NodeArg of the given node -const onnxruntime::Node* GetInputNode(const Node& node, const NodeArg* def) { - const auto& input_name = def->Name(); - const onnxruntime::Node* input_node = nullptr; - // search input node set to see if input_name is in their outputs (weights are not from node) - for (auto iter = node.InputNodesBegin(); iter != node.InputNodesEnd(); ++iter) { - const onnxruntime::Node& p = *iter; - bool found = false; - ORT_THROW_IF_ERROR(p.ForEachWithIndex( - p.OutputDefs(), - [&found, &input_name](const onnxruntime::NodeArg& out_def, size_t) { - if (input_name == out_def.Name()) { - found = true; - } - return Status::OK(); - })); - if (found) - input_node = &p; - } - return input_node; -} - -// create capacity from subgraph -std::unique_ptr ToCapacity(const onnxruntime::GraphViewer& graph, - int fused_count, - std::unique_ptr& subgraph) { - auto meta_def = std::make_unique<::onnxruntime::IndexedSubGraph::MetaDef>(); - meta_def->name = "Fuse" + std::to_string(fused_count); - meta_def->domain = "Fuse"; - - std::set node_indices(subgraph->nodes.begin(), subgraph->nodes.end()); - - const auto& start_node_index = subgraph->nodes.front(); - const auto& start_node = *graph.GetNode(start_node_index); - const auto& end_node_index = subgraph->nodes.back(); - const auto& end_node = *graph.GetNode(end_node_index); - meta_def->name += start_node.OpType() + std::to_string(start_node_index); - meta_def->name += "_With" + std::to_string(subgraph->nodes.size()) + "Nodes_"; - meta_def->name += end_node.OpType() + std::to_string(end_node_index); - - InlinedHashSet real_output_names; - real_output_names.reserve(graph.GetOutputs().size()); - for (const auto* def : graph.GetOutputs()) { - real_output_names.insert(def->Name()); - } - - for (const auto& node_index : subgraph->nodes) { - const auto& node = *graph.GetNode(node_index); - auto process_input_fn = - [&meta_def, &node, &node_indices](const onnxruntime::NodeArg& def, size_t) { - const onnxruntime::Node* input_node = GetInputNode(node, &def); - bool input_from_subgraph = (input_node && node_indices.count(input_node->Index())); - if (!input_from_subgraph) { - // input is from weights or outside of graph - meta_def->inputs.push_back(def.Name()); - } - return Status::OK(); - }; - // handle current graph's inputs - ORT_THROW_IF_ERROR(node.ForEachWithIndex(node.InputDefs(), process_input_fn)); - // nodes' implicit inputs also need to be collected. They need to - // be promoted to being explicit inputs for everything to work. - ORT_THROW_IF_ERROR(node.ForEachWithIndex(node.ImplicitInputDefs(), process_input_fn)); - - // Handle outouts - // two cases are considered as outputs - // 1. Output NodeArg is not used by any Node - // 2. Output NodeArg is used by at least one Node out of this subgraph. - // Note a NodeArg can be used by Nodes in and out of the subgraph at the same time. - // 3. Output NodeArg is one of real outputs of an Ort graph. - - auto InsertOutputToSubgraph = [&meta_def](const NodeArg* def) { - if (std::find(meta_def->outputs.begin(), meta_def->outputs.end(), def->Name()) == - meta_def->outputs.end()) { - meta_def->outputs.push_back(def->Name()); - } - }; - - InlinedHashSet input_names_from_the_output_node; - - for (auto o_iter = node.OutputEdgesBegin(); o_iter != node.OutputEdgesEnd(); ++o_iter) { - const auto& p = *o_iter; - const Node& out_node = p.GetNode(); - - // preprocess for the case 1 - ORT_THROW_IF_ERROR(out_node.ForEachWithIndex( - out_node.InputDefs(), - [&input_names_from_the_output_node](const onnxruntime::NodeArg& in_def, size_t) { - input_names_from_the_output_node.insert(in_def.Name()); - return Status::OK(); - })); - - // handle the case 2 - if (node_indices.count(out_node.Index()) == 0) { - const NodeArg* def = node.OutputDefs()[p.GetSrcArgIndex()]; - InsertOutputToSubgraph(def); - } - } - - // handle case 1 and 3 - ORT_THROW_IF_ERROR(node.ForEachWithIndex( - node.OutputDefs(), - [&](const onnxruntime::NodeArg& def, size_t) { - if (input_names_from_the_output_node.count(def.Name()) == 0 || - real_output_names.count(def.Name()) > 0) { - InsertOutputToSubgraph(&def); - } - return Status::OK(); - })); - } - - // Handle subgraph's initializers - const auto& all_initializers = graph.GetAllInitializedTensors(); - for (const auto& node_index : subgraph->nodes) { - const auto& node = *graph.GetNode(node_index); - // check whether it is an immediate nested subgraph - auto immediate_nested_subgraph = GetSubgraph(node); - // If so, copy the immediate nested subgraph's initializers to meta_def->inputs. - // Note we don't need recursion here, since Ort did recursion for us by handling subgraph early than the current graph. - // Therefore, the all inner nested subgraph's initializers should be already in the immediate nested subgraph's inputs. - if (nullptr != immediate_nested_subgraph) { - for (auto& n : immediate_nested_subgraph->Nodes()) { - auto add_input_fn = - [&meta_def, &all_initializers](const onnxruntime::NodeArg& def, size_t) { - auto iter = all_initializers.find(def.Name()); - if (iter != all_initializers.end()) { - meta_def->inputs.push_back(def.Name()); - } - return Status::OK(); - }; - ORT_THROW_IF_ERROR(n.ForEachWithIndex(n.InputDefs(), add_input_fn)); - ORT_THROW_IF_ERROR(n.ForEachWithIndex(n.ImplicitInputDefs(), add_input_fn)); - } - } - } - - meta_def->since_version = 1; - meta_def->status = ONNX_NAMESPACE::EXPERIMENTAL; - std::unique_ptr finished_subgraph(subgraph.release()); - finished_subgraph->SetMetaDef(std::move(meta_def)); - return std::make_unique(std::move(finished_subgraph)); -} - -int64_t ShapeRank(const NodeArg* def) { - ORT_ENFORCE_DEBUG(nullptr != def); - return gsl::narrow_cast(def->Shape()->dim_size()); -} - -bool ShapeHasValue(const NodeArg* def, int i) { - ORT_ENFORCE_DEBUG(nullptr != def); - ORT_ENFORCE_DEBUG(i >= 0); - ORT_ENFORCE_DEBUG(i < def->Shape()->dim_size()); - return utils::HasDimValue(def->Shape()->dim(i)); -} - -bool ShapeHasSymbol(const NodeArg* def, int i) { - ORT_ENFORCE_DEBUG(nullptr != def); - ORT_ENFORCE_DEBUG(i >= 0); - ORT_ENFORCE_DEBUG(i < def->Shape()->dim_size()); - return utils::HasDimParam(def->Shape()->dim(i)); -} - -int64_t ShapeValue(const NodeArg* def, int i) { - ORT_ENFORCE_DEBUG(ShapeHasValue(def, i)); - return def->Shape()->dim(i).dim_value(); -} - -const std::string& ShapeSymbol(const NodeArg* def, int i) { - ORT_ENFORCE_DEBUG(ShapeHasSymbol(def, i)); - return def->Shape()->dim(i).dim_param(); -} - -ONNX_NAMESPACE::TensorProto_DataType TensorProtoDataType(const NodeArg* def) { - ORT_ENFORCE_DEBUG(nullptr != def); - return static_cast(def->TypeAsProto()->tensor_type().elem_type()); -} - -// Convert GraphNodes to internal NodePtrs without check lifetime. -// Please use it only locally when GraphNodes still exist -InlinedVector ConvertGraphNodesToNodePtrs(const ConstGraphNodes& graph_nodes) { - InlinedVector nodes; - for (auto& node : graph_nodes) { - nodes.push_back(&node); - } - return nodes; -} - -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/common.h b/onnxruntime/core/codegen/common/common.h deleted file mode 100644 index 81b74daf6f711..0000000000000 --- a/onnxruntime/core/codegen/common/common.h +++ /dev/null @@ -1,153 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/common/inlined_containers.h" -#include "core/framework/compute_capability.h" -#include "core/framework/tensor.h" -#include "core/graph/graph_nodes.h" -#include "core/graph/graph_viewer.h" - -#ifndef NDEBUG -#define ORT_ENFORCE_DEBUG(...) ORT_ENFORCE(__VA_ARGS__) -#else -#define ORT_ENFORCE_DEBUG(...) -#endif // !NDEBUG - -// DYN_PROMOTE is a simplified llvm::dyn_cast, which does not need RTTI -// DYN_PROMOTE is faster than dynamic_cast and also has smaller binary size -// Please use DYN_PROMOTE in a critical path. -#define DYN_PROMOTE(BASE) \ - template \ - inline const ToType* Promote(const BASE* base) { \ - if (ToType::IsType(base)) \ - return static_cast(base); \ - return nullptr; \ - } \ - \ - template \ - inline ToType* Promote(BASE* base) { \ - if (ToType::IsType(base)) \ - return static_cast(base); \ - return nullptr; \ - } \ - \ - template \ - inline ToType* Promote(const std::unique_ptr& base) { \ - if (ToType::IsType(base.get())) \ - return static_cast(base); \ - return nullptr; \ - } \ - \ - template \ - inline ToType* Promote(const std::shared_ptr& base) { \ - if (ToType::IsType(base.get())) \ - return static_cast(base); \ - return nullptr; \ - } - -// DYN_PROMOTE_BASE is a macro inserted in the base class to support DYN_PROMOTE -// TYPE_ID is required for DYN_PROMOTE and TYPE_ID is a enum class -// TYPE_ID_VAR is a corresponding variable name for in the base class -#define DYN_PROMOTE_BASE(BASE, TYPE_ID, TYPE_ID_VAR) \ - inline const TYPE_ID TypeID() const { \ - return TYPE_ID_VAR; \ - } \ - \ - static inline bool IsType(const BASE*) { \ - return true; \ - } - -// DYN_PROMOTE_DERIVED is a macro inserted in a derived class to support DYN_PROMOTE -// TYPE_ID is required for DYN_PROMOTE and TYPE_ID is a enum class -// TYPE_ID_VALUE is corresponding TYPE_ID::value of a derived class. -#define DYN_PROMOTE_DERIVED(BASE, TYPE_ID, TYPE_ID_VALUE) \ - static inline bool IsType(const BASE* base) { \ - ORT_ENFORCE_DEBUG(nullptr != base); \ - return base->TypeID() == TYPE_ID::TYPE_ID_VALUE; \ - } - -// DYNAMIC_PROMOTE is a dynamic_cast needing RTTI -// DYNAMIC_PROMOTE is usually slower than than DYN_PROMOTE. -// Please use DYNAMIC_PROMOTE in a non-critical path. -#define DYNAMIC_PROMOTE(BASE) \ - template \ - inline const X* Promote(const BASE* base) { \ - auto derived = dynamic_cast(base); \ - ORT_ENFORCE(nullptr != derived); \ - return derived; \ - } \ - \ - template \ - inline X* Promote(BASE* base) { \ - auto derived = dynamic_cast(base); \ - ORT_ENFORCE(nullptr != derived); \ - return derived; \ - } \ - \ - template \ - inline X* Promote(const std::unique_ptr& base) { \ - auto derived = dynamic_cast(base.get()); \ - ORT_ENFORCE(nullptr != derived); \ - return derived; \ - } \ - \ - template \ - inline X* Promote(const std::shared_ptr& base) { \ - auto derived = dynamic_cast(base.get()); \ - ORT_ENFORCE(nullptr != derived); \ - return derived; \ - } - -namespace onnxruntime { - -// Nodekey is used as a key for maps -using NodeKey = std::string; - -NodeKey GetKey(const onnxruntime::Node* node); -NodeKey GetKey(const onnxruntime::Node& node); -NodeKey GetKey(const onnxruntime::NodeArg* def); - -bool IsRecurrentNode(const onnxruntime::Node& node); - -bool IsAliasNode(const onnxruntime::Node& node); - -// Helper function that creates ComputeCapability for subgraphs -std::unique_ptr ToCapacity(const onnxruntime::GraphViewer& graph, - int fused_count, - std::unique_ptr& subgraph); - -bool IsFusedNode(const Node& node); - -bool HasLoop(const Node& node); - -const Graph* GetSubgraph(const Node& node); - -std::string NormalizeCppName(const std::string& name); - -std::string NormalizeNodeArgName(const NodeArg* def); - -// Return the corresponding input node for the NodeArg of the given node -const onnxruntime::Node* GetInputNode(const Node& node, const NodeArg* def); - -int64_t ShapeRank(const NodeArg* def); - -bool ShapeHasValue(const NodeArg* def, int i); - -bool ShapeHasSymbol(const NodeArg* def, int i); - -int64_t ShapeValue(const NodeArg* def, int i); - -const std::string& ShapeSymbol(const NodeArg* def, int i); - -ONNX_NAMESPACE::TensorProto_DataType TensorProtoDataType(const NodeArg* def); - -// Convert ConstGraphNodes to internal NodePtrs without check lifetime. -// Please use it only locally when GraphNodes still exist -InlinedVector ConvertGraphNodesToNodePtrs(const ConstGraphNodes& graph_nodes); - -enum : int { - Dimension_Unknown = -1, -}; - -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/creator.h b/onnxruntime/core/codegen/common/creator.h deleted file mode 100644 index b31a12db4875b..0000000000000 --- a/onnxruntime/core/codegen/common/creator.h +++ /dev/null @@ -1,76 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/common/dispatcher.h" - -// TODO rename this file to creator_base -namespace onnxruntime { -namespace codegen { - -// It is a base class for TVM Op IR builder, weight layout builder, TVM scheduler -// CreatorBase is a template class of compiler pass -// for 1) TVM IR builder -// 2) Weight layout transformer -// 3) TVM Scheduler, etc. -// CreatorBase is similor to OpXXCreate in llvm IR builder - -template -class CreatorBase { - public: - CreatorBase(const std::string& name) - : name_(name) {} - - virtual ~CreatorBase() = default; - - virtual RETURN_TYPE Evaluate(INPUT_TYPE, - NODE_TYPE, - CONTEXT_TYPE, - OUTPUT_TYPE) = 0; - - const std::string& Name() const { - return name_; - } - - protected: - std::string name_; - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(CreatorBase); -}; - -// macro to stringize -#define STRINGIZE_NX(OP) #OP -#define STRINGIZE(OP) STRINGIZE_NX(OP) - -// macro returns class name -#define CREATOR_CLASS(OP, POSTFIX) \ - OP##POSTFIX - -// macro returns class name as string -#define CREATOR_STRING(OP, POSTFIX) \ - STRINGIZE(CREATOR_CLASS(OP, POSTFIX)) - -// macro returns class constructor name -#define CREATOR_CLASS_FUNC(OP, POSTFIX) \ - OP##POSTFIX() - -// macro declares a creator class inheriting the template class CreatorBase -// with corresponding template parameters -#define DECLARE_CREATOR_CLASS(OP, POSTFIX, INPUT, NODE, CONTEXT, OUTPUT, RETURN) \ - class CREATOR_CLASS(OP, POSTFIX) : public onnxruntime::codegen::CreatorBase { \ - public: \ - CREATOR_CLASS_FUNC(OP, POSTFIX) : CreatorBase(CREATOR_STRING(OP, POSTFIX)) {} \ - RETURN Evaluate(INPUT, \ - NODE, \ - CONTEXT, \ - OUTPUT) override; \ - \ - private: \ - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(CREATOR_CLASS(OP, POSTFIX)); \ - }; - -} // namespace codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/dispatcher.h b/onnxruntime/core/codegen/common/dispatcher.h deleted file mode 100644 index 80a854a06977c..0000000000000 --- a/onnxruntime/core/codegen/common/dispatcher.h +++ /dev/null @@ -1,76 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/common/common.h" -#include -#include -#include - -namespace onnxruntime { -namespace codegen { - -// DispatcherBase is a customized unordered_map -// that provides all codegen-related functionality -// including 1) dispatching a pass -// 2) dump corresponding name -// DispatcherBase may or may not keep ownership, -// depending on the template parameter, CONTENT_TYPE. -// Note DispatcherBase has a protected destructor - -template -class DispatcherBase { - public: - DispatcherBase(const std::string& name) - : name_(name) {} - - const std::string& Name() const { - return name_; - } - - bool Contains(const std::string& name) const { - return contents_.count(name) > 0; - } - - void ForEach(std::function - func) { - for (auto& p : contents_) { - func(p.first, p.second); - } - } - - bool Register(const std::string& name, - CONTENT_TYPE op) { - if (!Contains(name)) { - contents_.emplace(name, op); - return true; - } - return false; - } - - CONTENT_TYPE Get(const std::string& key) const { - auto iter = contents_.find(key); - if (iter != contents_.end()) { - return iter->second; - } - return nullptr; - } - - const std::unordered_map GetContents() const { - return contents_; - } - - std::unordered_map GetMutableContents() { - return contents_; - } - - protected: - std::string name_; - std::unordered_map contents_; - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(DispatcherBase); - ~DispatcherBase() = default; -}; - -} // namespace codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/dump_array.h b/onnxruntime/core/codegen/common/dump_array.h deleted file mode 100644 index 8e51cd36d0087..0000000000000 --- a/onnxruntime/core/codegen/common/dump_array.h +++ /dev/null @@ -1,62 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include -#include -#include -#include - -namespace onnxruntime { - -template -void DumpArrayRecursive(const T1* data, int64_t& data_offset, const std::vector& shape, int idx) { - int dim = static_cast(shape.size()); - if (dim == 0) { - std::cout << "[]\n"; - return; - } - - assert(idx < dim); - int sz = shape[idx]; - - std::cout << "["; - if (idx < dim - 1) { - for (auto i = 0; i < sz; ++i) { - DumpArrayRecursive(data, data_offset, shape, idx + 1); - if (i < sz - 1) { - std::cout << ","; - // print multiple newlines after ',' when necessary - for (int j = idx + 1; j < dim; j++) - std::cout << "\n"; - // print leading spaces before "[" when necessary - for (int j = 0; j < idx + 1; ++j) - std::cout << " "; - } - } - } else { - for (auto i = 0; i < sz; ++i) { - if (std::is_same::value || std::is_same::value) - std::cout << std::setw(3) << static_cast(*(data + data_offset)); - else - std::cout << std::setw(12) << std::setprecision(8) << *(data + data_offset); - data_offset++; - if (i < sz - 1) - std::cout << ","; - } - } - std::cout << "]"; -} - -// A helper function to dump multidimensional arrays in a way similar to numpy -template -void DumpArray(const std::string& tag, const T1* data, const std::vector& shape) { - std::cout << tag << "\n"; - int64_t data_offset = 0; - DumpArrayRecursive(data, data_offset, shape, 0); - assert(data_offset == TotalSize(shape)); - std::cout << std::endl; -} - -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/handle.h b/onnxruntime/core/codegen/common/handle.h deleted file mode 100644 index 7caad27dcbe01..0000000000000 --- a/onnxruntime/core/codegen/common/handle.h +++ /dev/null @@ -1,22 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/common/target_info.h" -#include -#include - -namespace onnxruntime { -namespace codegen { - -using DomainVersionLookupFunc = std::function; - -struct CodeGenHandle { - CodeGenTarget* codegen_target; - DomainVersionLookupFunc domain_version_lookup_func = - // by default, always uses the latest opset implemented - [](const std::string&) { return INT_MAX; }; -}; - -} // namespace codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/op_macro.h b/onnxruntime/core/codegen/common/op_macro.h deleted file mode 100644 index 04305c4aa47b0..0000000000000 --- a/onnxruntime/core/codegen/common/op_macro.h +++ /dev/null @@ -1,101 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -namespace onnxruntime { - -#define LIST_BINARY_OPS() \ - BINARY_OP(Add) \ - BINARY_OP(Div) \ - BINARY_OP(Mul) \ - BINARY_OP(PRelu) \ - BINARY_OP(Sub) - -#define LIST_BINARY_CMP_OPS() \ - BINARY_CMP_OP(Equal) \ - BINARY_CMP_OP(Greater) \ - BINARY_CMP_OP(Less) - -#define LIST_POOL_OPS() \ - POOL_OP(MaxPool) \ - POOL_OP(AveragePool) \ - POOL_OP(GlobalMaxPool) \ - POOL_OP(GlobalAveragePool) - -#define LIST_REDUCE_OPS() \ - REDUCE_INDEXED_OP(ArgMax) \ - REDUCE_INDEXED_OP(ArgMin) \ - REDUCE_OP(ReduceL1) \ - REDUCE_OP(ReduceL2) \ - REDUCE_OP(ReduceLogSum) \ - REDUCE_OP(ReduceLogSumExp) \ - REDUCE_OP(ReduceMax) \ - REDUCE_OP(ReduceMean) \ - REDUCE_OP(ReduceMin) \ - REDUCE_OP(ReduceProd) \ - REDUCE_OP(ReduceSum) \ - REDUCE_OP(ReduceSumSquare) - -#define LIST_UNARY_OPS() \ - UNARY_OP(Abs) \ - UNARY_OP(Affine) \ - UNARY_OP(Ceil) \ - UNARY_OP(Elu) \ - UNARY_OP(Exp) \ - UNARY_OP(Floor) \ - UNARY_OP(HardSigmoid) \ - UNARY_OP(LeakyRelu) \ - UNARY_OP(Log) \ - UNARY_OP(Neg) \ - UNARY_OP(ParametricSoftplus) \ - UNARY_OP(Reciprocal) \ - UNARY_OP(Relu) \ - UNARY_OP(ScaledTanh) \ - UNARY_OP(Selu) \ - UNARY_OP(Sigmoid) \ - UNARY_OP(Softplus) \ - UNARY_OP(Softsign) \ - UNARY_OP(Sqrt) \ - UNARY_OP(Tanh) \ - UNARY_OP(ThresholdedRelu) - -#define LIST_VARIADIC_OPS() \ - VARIADIC_OP(Max) \ - VARIADIC_OP(Min) \ - VARIADIC_OP(Sum) - -#define LIST_ALL_GENERIC_OPS() \ - LIST_BINARY_OPS() \ - LIST_BINARY_CMP_OPS() \ - LIST_REDUCE_OPS() \ - LIST_POOL_OPS() \ - LIST_UNARY_OPS() \ - LIST_VARIADIC_OPS() \ - ADD_OP_ITEM(Cast) \ - ADD_OP_ITEM(Clip) \ - ADD_OP_ITEM(Concat) \ - ADD_OP_ITEM(Conv) \ - ADD_OP_ITEM(Crop) \ - ADD_OP_ITEM(Dropout) \ - ADD_OP_ITEM(Expand) \ - ADD_OP_ITEM(Flatten) \ - ADD_OP_ITEM(Gather) \ - ADD_OP_ITEM(GatherElements) \ - ADD_OP_ITEM(Gemm) \ - ADD_OP_ITEM(Identity) \ - ADD_OP_ITEM(LogSoftmax) \ - ADD_OP_ITEM(LSTM) \ - ADD_OP_ITEM(MatMul) \ - ADD_OP_ITEM(MatMulInteger) \ - ADD_OP_ITEM(Pad) \ - ADD_OP_ITEM(Reshape) \ - ADD_OP_ITEM(Shape) \ - ADD_OP_ITEM(Slice) \ - ADD_OP_ITEM(Softmax) \ - ADD_OP_ITEM(Split) \ - ADD_OP_ITEM(Squeeze) \ - ADD_OP_ITEM(Transpose) \ - ADD_OP_ITEM(Unsqueeze) \ - ADD_OP_ITEM(Where) - -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/profile.h b/onnxruntime/core/codegen/common/profile.h deleted file mode 100644 index 31c9e764320d0..0000000000000 --- a/onnxruntime/core/codegen/common/profile.h +++ /dev/null @@ -1,37 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once - -// uncomment this line or use -DCODEGEN_ENABLE_PROFILER in compiler options to enable profiler events in codegen -// #define CODEGEN_ENABLE_PROFILER - -#ifdef CODEGEN_ENABLE_PROFILER -#include "core/common/profiler.h" - -namespace onnxruntime { - -class ProfilerEvent { - public: - ProfilerEvent(const std::string& name) : name_(name) { - ts_ = profiling::Profiler::Instance().StartTime(); - } - - ~ProfilerEvent() { - profiling::Profiler::Instance().EndTimeAndRecordEvent(profiling::EventCategory::NODE_EVENT, name_, ts_); - } - - private: - TimePoint ts_; - const std::string name_; -}; - -} // namespace onnxruntime - -#define CODEGEN_PROFILER_EVENT(name) onnxruntime::ProfilerEvent profiler_event(name) - -#else - -#define CODEGEN_PROFILER_EVENT(name) - -#endif diff --git a/onnxruntime/core/codegen/common/registry.h b/onnxruntime/core/codegen/common/registry.h deleted file mode 100644 index c1642e76e2120..0000000000000 --- a/onnxruntime/core/codegen/common/registry.h +++ /dev/null @@ -1,72 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/common/common.h" -#include -#include -#include - -namespace onnxruntime { -namespace codegen { - -// RegistryBase is a customized unordered_map -// that keep ownership of passes, -// including 1) IR builder passes -// 2) Weight layout transformer passes -// 3) Scheduler passses, etc. - -template -class RegistryBase { - public: - RegistryBase() = default; - - virtual ~RegistryBase() = default; - - bool Contains(const std::string& name) const { - return contents_.count(name) > 0; - } - - CONTENT_TYPE* Get(const std::string& name) const { - if (contents_.find(name) != contents_.end()) - return contents_.at(name).get(); - return nullptr; - } - - CONTENT_TYPE* RegisterOrGet( - const std::string& name, - std::unique_ptr&& ptr) { - if (!Contains(name)) - contents_.emplace(name, std::move(ptr)); - return Get(name); - } - - CONTENT_TYPE* RegisterOrGet( - std::unique_ptr&& ptr) { - return RegisterOrGet(ptr->Name(), std::move(ptr)); - } - - bool Register( - const std::string& name, - std::unique_ptr&& ptr) { - if (!Contains(name)) { - contents_.emplace(name, std::move(ptr)); - return true; - } - return false; - } - - bool Register( - std::unique_ptr&& ptr) { - return Register(ptr->Name(), std::move(ptr)); - } - - protected: - std::unordered_map> contents_; - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(RegistryBase); -}; - -// Put common Registry Management utilities if these is any - -} // namespace codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/settings.cc b/onnxruntime/core/codegen/common/settings.cc deleted file mode 100644 index 529cb654f922c..0000000000000 --- a/onnxruntime/core/codegen/common/settings.cc +++ /dev/null @@ -1,78 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/common/settings.h" - -#include "core/common/logging/logging.h" -#include -#include - -namespace onnxruntime { -namespace codegen { - -CodeGenSettings& CodeGenSettings::Instance() { - static CodeGenSettings settings; - return settings; -} - -CodeGenSettings::CodeGenSettings() {} - -void CodeGenSettings::InsertOptions(const std::map& options) { - for (const auto& option : options) { - const auto& key = option.first; - const auto& value = option.second; - - auto iter = options_.find(key); - // found existing ones - if (iter != options_.end()) { - if (iter->second != value) { - LOGS_DEFAULT(CODEGEN_SETTINGS_LOG_LEVEL) << "CodeGenSettings: option" - << key << " is overridded from: " - << iter->second << " to: " << value; - iter->second = value; - } - } else { - options_.insert(std::make_pair(key, value)); - } - } -} - -void CodeGenSettings::DumpOptions() const { - std::ostringstream stream; - stream << "CodeGenSettings: dump all options" << std::endl; - for (const auto& option : options_) { - stream << " " << option.first << " = " << option.second << std::endl; - } - LOGS_DEFAULT(CODEGEN_SETTINGS_LOG_LEVEL) << stream.str(); -} - -std::string CodeGenSettings::GetOptionValue(const std::string& key) const { - const auto& iter = options_.find(key); - if (iter == options_.end()) { - LOGS_DEFAULT(CODEGEN_SETTINGS_LOG_LEVEL) << "CodeGenSettings::GetOptionValue: unrecognized option" << key; - return ""; - } - return iter->second; -} - -bool CodeGenSettings::HasOption(const std::string& key) const { - return options_.count(key) > 0; -} - -bool CodeGenSettings::OptionMatches(const std::string& key, const std::string& value) const { - if (!HasOption(key)) - return false; - -#ifdef _WIN32 - return 0 == _stricmp(options_.at(key).c_str(), value.c_str()); -#else - return 0 == strcasecmp(options_.at(key).c_str(), value.c_str()); -#endif -} - -void CodeGenSettings::Clear() { - options_.clear(); -} - -} // namespace codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/settings.h b/onnxruntime/core/codegen/common/settings.h deleted file mode 100644 index e327b0e207cc2..0000000000000 --- a/onnxruntime/core/codegen/common/settings.h +++ /dev/null @@ -1,40 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace codegen { - -// use log level warning as default to make sure logs are outputted -#define CODEGEN_SETTINGS_LOG_LEVEL WARNING - -// This stores codegen settings to control dumps, execution preference, etc. -// CodeGenSettings could come from command line options or environment variables -// Or could come from a static variables in source code -class CodeGenSettings { - public: - // generic built-in options - constexpr static const char* kDumpAllOptions = "dump_all_options"; - constexpr static const char* kCodeGenDumpModule = "codegen_dump_module"; // dump tvm module - constexpr static const char* kCodeGenDumpLower = "codegen_dump_lower"; // dump lowered func - constexpr static const char* kCodeGenDumpSchedule = "codegen_dump_schedule"; // dump scheduler - - void InsertOptions(const std::map& options); - void DumpOptions() const; - std::string GetOptionValue(const std::string& key) const; - bool HasOption(const std::string& key) const; - bool OptionMatches(const std::string& key, const std::string& value) const; - void Clear(); - static CodeGenSettings& Instance(); - - private: - CodeGenSettings(); - - std::map options_; -}; - -} // namespace codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/target_info.h b/onnxruntime/core/codegen/common/target_info.h deleted file mode 100644 index da063545f0a1e..0000000000000 --- a/onnxruntime/core/codegen/common/target_info.h +++ /dev/null @@ -1,33 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { - -// CodeGenTarget holds meta info for backend code generation -// and will be lowered to a target of corresponding backend -// code generation, e.g. TVM's Target. -class CodeGenTarget { - public: - CodeGenTarget() {} - CodeGenTarget(const std::string& target_name) - : target_name_(target_name) {} - - virtual int NaturalVectorWidth(int /*bits*/) const { - return 1; - } - - const std::string& GetTargetName() const { - return target_name_; - } - - virtual ~CodeGenTarget() = default; - - private: - std::string target_name_{"unknown"}; // default name is unknown -}; - -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/utils.cc b/onnxruntime/core/codegen/common/utils.cc deleted file mode 100644 index f4140a411bddf..0000000000000 --- a/onnxruntime/core/codegen/common/utils.cc +++ /dev/null @@ -1,99 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/common/utils.h" -#include "core/common/cpuid_info.h" -#include "core/common/safeint.h" - -#include -#include - -namespace onnxruntime { - -std::unique_ptr GetEnv(const char* var) { - char* val = nullptr; -#if _MSC_VER - size_t len; - - if (_dupenv_s(&val, &len, var)) { - // Something went wrong, just return nullptr. - return nullptr; - } -#else - val = getenv(var); -#endif // _MSC_VER - - if (val == nullptr) { - return nullptr; - } - - // On windows, we will have to explicitly free val. Instead of returning val - // to its caller and make distinguish between windows and linux, we return - // a unique_ptr, and it will be destroyed automatically after the caller - // completes. - size_t len_val = strnlen(val, onnxruntime::kMaxStrLen) + 1; - auto p = std::make_unique(len_val); - // use explicit loop to get ride of VC's warning on unsafe copy - for (size_t i = 0; i < len_val; ++i) { - p[i] = val[i]; - } - return p; -} - -bool IsEnvVarDefined(const char* var) { - auto val = GetEnv(var); - return val != nullptr; -} - -int64_t TotalSize(const std::vector& shape) { - SafeInt total = 1; - for (auto s : shape) { - total *= s; - } - return total; -} - -// Return the strides for the input shape, i.e. the number of -// elements contained by a single element of current dimension. -// For example, for shape[3][4][5][6], strides will be -// [4*5*6, 5*6, 6, 1], i.e. [120, 30, 6, 1] -void GetStrides(const int64_t* shape, int ndim, std::vector& strides) { - strides.resize(ndim); - strides[ndim - 1] = 1; - for (int64_t i = ndim - 2; i >= 0; i--) { - strides[i] = strides[i + 1] * shape[i + 1]; - } -} - -// Common utils to get target option -TargetFeature GetTargetInfo(const codegen::CodeGenSettings& settings) { - TargetFeature feature; - - std::string target_str = ""; - - bool isAVX = false; - bool isAVX2 = false; - bool isAVX512 = false; - if (target_str == "avx") { - isAVX = true; - } else if (target_str == "avx2") { - isAVX = true; - isAVX2 = true; - } else if (target_str == "avx512") { - isAVX = true; - isAVX2 = true; - isAVX512 = true; - } else { - isAVX = CPUIDInfo::GetCPUIDInfo().HasAVX(); - isAVX2 = CPUIDInfo::GetCPUIDInfo().HasAVX2(); - isAVX512 = CPUIDInfo::GetCPUIDInfo().HasAVX512Skylake(); - } - - feature.hasAVX = isAVX; - feature.hasAVX2 = isAVX2; - feature.hasAVX512 = isAVX512; - - return feature; -} - -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/common/utils.h b/onnxruntime/core/codegen/common/utils.h deleted file mode 100644 index ef06b5b72dc2c..0000000000000 --- a/onnxruntime/core/codegen/common/utils.h +++ /dev/null @@ -1,45 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/common/common.h" -#include -#include -#include - -namespace onnxruntime { - -// Holding utility functions that are not tied to TVM and ORT - -std::unique_ptr GetEnv(const char* var); - -// Check if an environment variable is set -bool IsEnvVarDefined(const char* var); - -int64_t TotalSize(const std::vector& shape); - -void GetStrides(const int64_t* shape, int ndim, std::vector& strides); - -struct TargetFeature { - bool hasAVX; - bool hasAVX2; - bool hasAVX512; -}; - -TargetFeature GetTargetInfo(const codegen::CodeGenSettings& setttings); - -// GCD (Greatest Common Divisor) -template -T GCD(T a, T b) { - ORT_ENFORCE(a >= 0); - ORT_ENFORCE(b >= 0); - if (a < b) std::swap(a, b); - if (b == 0) return a; - while (a % b != 0) { - a = a % b; - std::swap(a, b); - } - return b; -} - -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/common.h b/onnxruntime/core/codegen/mti/common.h deleted file mode 100644 index d71e740b9284a..0000000000000 --- a/onnxruntime/core/codegen/mti/common.h +++ /dev/null @@ -1,16 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once - -#include -#include - -#define MTI_ASSERT(condition) \ - if (!(condition)) { \ - std::string error_msg = "Not satisfied: " #condition \ - ": line " + \ - std::to_string(__LINE__) + \ - " in file " + std::string(__FILE__) + "\n"; \ - throw std::runtime_error(error_msg); \ - } diff --git a/onnxruntime/core/codegen/mti/debug/tvm_print.cc b/onnxruntime/core/codegen/mti/debug/tvm_print.cc deleted file mode 100644 index 0491636032b47..0000000000000 --- a/onnxruntime/core/codegen/mti/debug/tvm_print.cc +++ /dev/null @@ -1,83 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/debug/tvm_print.h" - -#include "core/codegen/common/utils.h" -#include "core/codegen/common/dump_array.h" -#include "core/codegen/mti/common.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -TVM_REGISTER_GLOBAL("tvm.contrib.onnxruntime.print") - .set_body([](tvm::TVMArgs args, tvm::TVMRetValue* /*ret*/) { - DLTensor* X = args[0]; - DLTensor* Y = args[1]; - - DLDataType dtype = X->dtype; - std::vector shape; - int64_t total_size = 1; - for (int i = 0; i < X->ndim; ++i) { - shape.push_back(X->shape[i]); - total_size *= X->shape[i]; - } - - // pass X to Y - memcpy(static_cast(Y->data) + Y->byte_offset, - static_cast(X->data) + X->byte_offset, - total_size * dtype.bits / 8); - - if (tvm::runtime::TypeMatch(dtype, kDLFloat, 32)) { - float* data = reinterpret_cast(static_cast(X->data) + X->byte_offset); - DumpArray("float tensor:", data, shape); - } else if (tvm::runtime::TypeMatch(dtype, kDLInt, 8)) { - int8_t* data = reinterpret_cast(static_cast(X->data) + X->byte_offset); - DumpArray("int8 tensor:", data, shape); - } else if (tvm::runtime::TypeMatch(dtype, kDLInt, 16)) { - int16_t* data = reinterpret_cast(static_cast(X->data) + X->byte_offset); - DumpArray("int16 tensor:", data, shape); - } else if (tvm::runtime::TypeMatch(dtype, kDLInt, 32)) { - int32_t* data = reinterpret_cast(static_cast(X->data) + X->byte_offset); - DumpArray("int32 tensor:", data, shape); - } else if (tvm::runtime::TypeMatch(dtype, kDLUInt, 8)) { - uint8_t* data = reinterpret_cast(static_cast(X->data) + X->byte_offset); - DumpArray("uint8 tensor:", data, shape); - } else if (tvm::runtime::TypeMatch(dtype, kDLUInt, 16)) { - uint16_t* data = reinterpret_cast(static_cast(X->data) + X->byte_offset); - DumpArray("uint16 tensor:", data, shape); - } else if (tvm::runtime::TypeMatch(dtype, kDLUInt, 32)) { - uint32_t* data = reinterpret_cast(static_cast(X->data) + X->byte_offset); - DumpArray("uint32 tensor:", data, shape); - } else { - MTI_ASSERT(0 && "not implemented!"); - } - }); - -tvm::Array -PrintTVMTensorExtern(const tvm::Tensor& X, - const std::string& name) { - return topi::detail::make_extern( - {X->shape}, - {X->dtype}, - {X}, - [&](tvm::Array ins, tvm::Array outs) { - return topi::detail::call_packed({tvm::Expr("tvm.contrib.onnxruntime.print"), - topi::detail::pack_buffer(ins[0]), - topi::detail::pack_buffer(outs[0])}); - }, - name + "_print", "", {}); -} - -tvm::Tensor PrintImmutable(const tvm::Tensor& X) { - auto outputs = PrintTVMTensorExtern(X, X->op->name + "_print"); - return outputs[0]; -} - -void Print(tvm::Tensor& X) { - X = PrintImmutable(X); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/debug/tvm_print.h b/onnxruntime/core/codegen/mti/debug/tvm_print.h deleted file mode 100644 index 91a334785a2a4..0000000000000 --- a/onnxruntime/core/codegen/mti/debug/tvm_print.h +++ /dev/null @@ -1,19 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Array PrintTVMTensorExtern( - const tvm::Tensor& X, - const std::string& name = "PrintTVM2DTensorExtern"); - -tvm::Tensor PrintImmutable(const tvm::Tensor& X); - -void Print(tvm::Tensor& X); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/binary_ops.cc b/onnxruntime/core/codegen/mti/math/binary_ops.cc deleted file mode 100644 index f3048799458f4..0000000000000 --- a/onnxruntime/core/codegen/mti/math/binary_ops.cc +++ /dev/null @@ -1,70 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/math/binary_ops.h" - -#include "core/codegen/mti/math/unary_ops.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/cast_ops.h" -#include - -// Using namespace topi for override operator +-*/ -using namespace topi; - -namespace onnxruntime { -namespace tvm_codegen { - -#define TVM_BINARY_OP1(op, expr) \ - tvm::Tensor op(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name) { \ - return Rename(expr, name); \ - } \ - tvm::Tensor op(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name) { \ - return Rename(expr, name); \ - } - -#define TVM_BINARY_OP(op, expr) \ - TVM_BINARY_OP1(op, expr) \ - tvm::Tensor op(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name) { \ - return Rename(expr, name); \ - } - -TVM_BINARY_OP(Add, lhs + rhs); -TVM_BINARY_OP(Div, lhs / rhs); -TVM_BINARY_OP(Max, maximum(lhs, rhs)); -TVM_BINARY_OP(Min, minimum(lhs, rhs)); -TVM_BINARY_OP(Mul, lhs* rhs); -TVM_BINARY_OP1(PRelu, Relu(lhs) - rhs * Relu(0 - lhs)); -TVM_BINARY_OP(Sub, lhs - rhs); - -tvm::Tensor Equal(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name) { - return topi::equal(lhs, rhs, name); -} -tvm::Tensor Equal(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name) { - return topi::equal(lhs, rhs, name); -} -tvm::Tensor Equal(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name) { - return topi::equal(lhs, rhs, name); -} - -tvm::Tensor Greater(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name) { - return topi::greater(lhs, rhs, name); -} -tvm::Tensor Greater(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name) { - return topi::greater(lhs, rhs, name); -} -tvm::Tensor Greater(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name) { - return topi::greater(lhs, rhs, name); -} - -tvm::Tensor Less(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name) { - return topi::less(lhs, rhs, name); -} -tvm::Tensor Less(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name) { - return topi::less(lhs, rhs, name); -} -tvm::Tensor Less(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name) { - return topi::less(lhs, rhs, name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/binary_ops.h b/onnxruntime/core/codegen/mti/math/binary_ops.h deleted file mode 100644 index dd51ce5e7917d..0000000000000 --- a/onnxruntime/core/codegen/mti/math/binary_ops.h +++ /dev/null @@ -1,42 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Add(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name = "add"); -tvm::Tensor Add(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name = "add"); -tvm::Tensor Add(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name = "add"); -tvm::Tensor Div(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name = "div"); -tvm::Tensor Div(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name = "div"); -tvm::Tensor Div(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name = "div"); -tvm::Tensor Equal(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name = "equal"); -tvm::Tensor Equal(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name = "equal"); -tvm::Tensor Equal(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name = "equal"); -tvm::Tensor Greater(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name = "greater"); -tvm::Tensor Greater(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name = "greater"); -tvm::Tensor Greater(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name = "greater"); -tvm::Tensor Less(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name = "less"); -tvm::Tensor Less(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name = "less"); -tvm::Tensor Less(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name = "less"); -tvm::Tensor Max(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name = "max"); -tvm::Tensor Max(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name = "max"); -tvm::Tensor Max(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name = "max"); -tvm::Tensor Min(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name = "min"); -tvm::Tensor Min(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name = "min"); -tvm::Tensor Min(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name = "min"); -tvm::Tensor Mul(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name = "mul"); -tvm::Tensor Mul(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name = "mul"); -tvm::Tensor Mul(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name = "mul"); -tvm::Tensor PRelu(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name = "prelu"); -tvm::Tensor PRelu(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name = "prelu"); -tvm::Tensor Sub(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name = "sub"); -tvm::Tensor Sub(const tvm::Tensor& lhs, const tvm::Expr& rhs, const std::string& name = "sub"); -tvm::Tensor Sub(const tvm::Expr& lhs, const tvm::Tensor& rhs, const std::string& name = "sub"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/gemm.cc b/onnxruntime/core/codegen/mti/math/gemm.cc deleted file mode 100644 index 7a79513ccaa97..0000000000000 --- a/onnxruntime/core/codegen/mti/math/gemm.cc +++ /dev/null @@ -1,30 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/math/gemm.h" - -#include "core/codegen/mti/math/matmul_ops.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include - -// Using namespace topi for override operator +-*/ -using namespace topi; - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Gemm(const tvm::Tensor& A, const tvm::Tensor& B, const tvm::Tensor& C, - bool trans_A, bool trans_B, float alpha, float beta, - const std::string& name) { - auto A_dot_B = MatMul2D(A, B, trans_A, trans_B, name + "_matmul2d"); - tvm::Expr alphaExpr = tvm::make_const(A->dtype, alpha); - if (beta != 0) { - tvm::Expr betaExpr = tvm::make_const(A->dtype, beta); - return Rename(alphaExpr * A_dot_B + (betaExpr * C), name); - } else { - return Rename(alphaExpr * A_dot_B, name); - } -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/gemm.h b/onnxruntime/core/codegen/mti/math/gemm.h deleted file mode 100644 index 3bb205c13fdc9..0000000000000 --- a/onnxruntime/core/codegen/mti/math/gemm.h +++ /dev/null @@ -1,16 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Gemm(const tvm::Tensor& p_A, const tvm::Tensor& p_B, const tvm::Tensor& p_C, - bool trans_A, bool trans_B, float alpha, float beta, - const std::string& name = "gemm"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/logsoftmax.cc b/onnxruntime/core/codegen/mti/math/logsoftmax.cc deleted file mode 100644 index cd8c2edae6959..0000000000000 --- a/onnxruntime/core/codegen/mti/math/logsoftmax.cc +++ /dev/null @@ -1,18 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/math/logsoftmax.h" - -#include "core/codegen/mti/tensor/reshape_ops.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor LogSoftmax(const tvm::Tensor& input, int64_t axis, const std::string& name) { - tvm::Tensor flatten_t = Flatten(input, axis, "logsoftmax_flatten"); - return Reshape(topi::nn::log_softmax(flatten_t, name), input->shape, "logsoftmax_reshape"); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/logsoftmax.h b/onnxruntime/core/codegen/mti/math/logsoftmax.h deleted file mode 100644 index 606a32806434b..0000000000000 --- a/onnxruntime/core/codegen/mti/math/logsoftmax.h +++ /dev/null @@ -1,11 +0,0 @@ -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor LogSoftmax(const tvm::Tensor& input, int64_t axis, const std::string& name = "logsoftmax"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/matmul_ops.cc b/onnxruntime/core/codegen/mti/math/matmul_ops.cc deleted file mode 100644 index 6ecf2f69a9c25..0000000000000 --- a/onnxruntime/core/codegen/mti/math/matmul_ops.cc +++ /dev/null @@ -1,161 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/math/matmul_ops.h" - -#include "core/codegen/mti/common.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor MatMul2D(const tvm::Tensor& A, const tvm::Tensor& B, bool trans_a, bool trans_b, const std::string& name) { - return topi::matmul(A, B, trans_a, trans_b, name); -} - -/* - * Generic Matrix Multiplication - * - * If both arguments are 2-D, they are multiplied like conventional matrices. - * - * If either argument is N-D and N > 2, it is treated as a stack of matrices residing in the last two indexes and broadcast accordingly. - * - * If the first argument is 1-D, it is promoted to a matrix by prepending a 1 to its dimensions. - * After matrix multiplication the prepended 1 is removed. - * - * If the second argument is 1-D, it is promoted to a matrix by appending a 1 to its dimensions. - * After matrix multiplication the appended 1 is removed. - */ -tvm::Tensor MatMul(const tvm::Tensor& A, const tvm::Tensor& B, const std::string& name) { - int64_t a_rank = static_cast(A->shape.size()); - int64_t b_rank = static_cast(B->shape.size()); - const auto& A_shape = A->shape; - const auto& B_shape = B->shape; - if (a_rank == 2 && b_rank == 2) { - // 2-D X 2-D - return MatMul2D(A, B); - } else if (a_rank == 1 && b_rank == 1) { - // 1-D X 1-D - auto k = tvm::reduce_axis(tvm::Range(0, A_shape[0]), "k"); - - return tvm::compute( - {}, - [&](const tvm::Array& /*indices*/) { - return tvm::sum(A[k] * B[k], {k}); - }, - name); - } else if (a_rank == 1) { - // 1-D X n-D - auto k = tvm::reduce_axis(tvm::Range(0, A_shape[0]), "k"); - - auto l = [&](const tvm::Array& indices) { - auto ndims = indices.size(); - MTI_ASSERT(ndims >= 1); - tvm::Array b_indices; - for (size_t bi = 0; bi < ndims - 1; ++bi) { - b_indices.push_back(indices[bi]); - } - b_indices.push_back(k); - b_indices.push_back(indices[ndims - 1]); - return tvm::sum(A({k}) * B(b_indices), {k}); - }; - return tvm::compute(ConcatShapes(SliceShapeToDimension(B_shape, -2), SliceShapeFromDimension(B_shape, -1)), l, name); - } else if (b_rank == 1) { - // n-D X 1-D - auto k = tvm::reduce_axis(tvm::Range(0, B_shape[0]), "k"); - - auto l = [&](const tvm::Array& indices) { - tvm::Array a_indices(indices.begin(), indices.end()); - a_indices.push_back(k); - return tvm::sum(A(a_indices) * B({k}), {k}); - }; - return tvm::compute(SliceShapeToDimension(A->shape, -1), l, name); - } else { - // n-D X m-D - MTI_ASSERT(a_rank >= 2 && b_rank >= 2); - auto k = tvm::reduce_axis(tvm::Range(0, A_shape[a_rank - 1]), "k"); - - auto l = [&](const tvm::Array& indices) { - auto ndims = static_cast(indices.size()); - MTI_ASSERT(ndims > 2); - tvm::Array a_indices, b_indices; - - // handle broadcasting - int i = 0, a_idx = 0, b_idx = 0; - bool a_greater = a_rank > b_rank; - for (; i < std::abs(a_rank - b_rank); ++i) { - if (a_greater) { - a_indices.push_back(indices[i]); - a_idx++; - } else { - b_indices.push_back(indices[i]); - b_idx++; - } - } - for (; i < ndims - 2; ++i, ++a_idx, ++b_idx) { - auto tp = indices[i].type(); - if (IsOne(A_shape, a_idx)) { - a_indices.push_back(tvm::make_zero(tp)); - b_indices.push_back(indices[i]); - } else if (IsOne(B_shape, b_idx)) { - b_indices.push_back(tvm::make_zero(tp)); - a_indices.push_back(indices[i]); - } else { - a_indices.push_back(indices[i]); - b_indices.push_back(indices[i]); - } - } - - MTI_ASSERT(a_idx == a_rank - 2 && b_idx == b_rank - 2); - a_indices.push_back(indices[ndims - 2]); - a_indices.push_back(k); - - b_indices.push_back(k); - b_indices.push_back(indices[ndims - 1]); - - return tvm::sum(A(a_indices) * B(b_indices), {k}); - }; - - return tvm::compute(ComputeMatMulShape(A_shape, B_shape), l, name); - } -} - -tvm::Array -ComputeMatMulShape( - const tvm::Array& A_shape, - const tvm::Array& B_shape, - bool trans_a, - bool trans_b) { - auto a_rank = A_shape.size(); - auto b_rank = B_shape.size(); - tvm::Array output_shape; - int64_t output_rank = std::max(a_rank, b_rank); - MTI_ASSERT(a_rank > 0 && b_rank > 0); - if (a_rank == 1 && b_rank == 1) { - MTI_ASSERT(!trans_a && !trans_b); - // reduction, output shape is empty - } else if (a_rank == 1) { - MTI_ASSERT(!trans_a && !trans_b); - output_shape = SliceShapeToDimension(B_shape, b_rank - 2); - output_shape.push_back(B_shape[b_rank - 1]); - } else if (b_rank == 1) { - MTI_ASSERT(!trans_a && !trans_b); - output_shape = SliceShapeToDimension(A_shape, a_rank - 1); - } else { - for (int64_t i = 0; i < output_rank - 2; i++) { - tvm::Expr broadcasted_dim = tvm::make_const(HalideIR::Int(32), 1); - bool broadcasted = - BroadcastDim(A_shape, i, output_rank, broadcasted_dim) && - BroadcastDim(B_shape, i, output_rank, broadcasted_dim); - MTI_ASSERT(broadcasted); - output_shape.push_back(broadcasted_dim); - } - output_shape.push_back(A_shape[a_rank - (trans_a ? 1 : 2)]); - output_shape.push_back(B_shape[b_rank - (trans_b ? 2 : 1)]); - } - return output_shape; -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/matmul_ops.h b/onnxruntime/core/codegen/mti/math/matmul_ops.h deleted file mode 100644 index ab9986132d34a..0000000000000 --- a/onnxruntime/core/codegen/mti/math/matmul_ops.h +++ /dev/null @@ -1,23 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Array -ComputeMatMulShape( - const tvm::Array& A_shape, - const tvm::Array& B_shape, - bool trans_a = false, - bool trans_b = false); - -tvm::Tensor MatMul2D(const tvm::Tensor& A, const tvm::Tensor& B, bool trans_a = false, bool trans_b = false, const std::string& name = "matmul2d"); - -tvm::Tensor MatMul(const tvm::Tensor& A, const tvm::Tensor& B, const std::string& name = "matmul"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/reduce_ops.cc b/onnxruntime/core/codegen/mti/math/reduce_ops.cc deleted file mode 100644 index 7d179e2b04316..0000000000000 --- a/onnxruntime/core/codegen/mti/math/reduce_ops.cc +++ /dev/null @@ -1,90 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/math/reduce_ops.h" - -#include "core/codegen/mti/math/binary_ops.h" -#include "core/codegen/mti/math/unary_ops.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor ArgMax(const tvm::Tensor& X, int64_t axis, bool keep_dims, const std::string& name) { - return Rename(topi::argmax(X, ToTvmArrayInt({axis}), keep_dims), name); -} - -tvm::Tensor ArgMin(const tvm::Tensor& X, int64_t axis, bool keep_dims, const std::string& name) { - return Rename(topi::argmin(X, ToTvmArrayInt({axis}), keep_dims), name); -} - -tvm::Tensor ReduceL1(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name) { - return ReduceSum(Abs(X), axes, keep_dims, name); -} - -tvm::Tensor ReduceL2(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name) { - return Sqrt(ReduceSumSquare(X, axes, keep_dims), name); -} - -tvm::Tensor ReduceLogSum(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name) { - return Log(ReduceSum(X, axes, keep_dims), name); -} - -tvm::Tensor ReduceLogSumExp(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name) { - tvm::Tensor reduce_max = ReduceMax(X, axes, true); - tvm::Tensor exp_delta = Exp(Sub(X, reduce_max)); - tvm::Tensor reduce_max_keep_dims = ReduceMax(X, axes, keep_dims); - return Add(ReduceLogSum(exp_delta, axes, keep_dims), reduce_max_keep_dims, name); -} - -tvm::Tensor ReduceMax(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name) { - return Rename(topi::max(X, ToTvmArrayInt(axes), keep_dims), name); -} - -tvm::Tensor ReduceMean(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name) { - tvm::Tensor reduce_sum = ReduceSum(X, axes, keep_dims); - tvm::Expr count = tvm::make_const(reduce_sum->dtype, 1.0f); - if (axes.empty()) { - for (const auto& dim : X->shape) - count = count * dim; - } else { - for (int64_t axis : axes) { - int64_t i = HandleNegativeAxis(axis, X->shape.size()); - count = count * X->shape[i]; - } - } - return tvm::compute( - reduce_sum->shape, - [&](const tvm::Array& i) { - return reduce_sum(i) / count; - }, - name); -} - -tvm::Tensor ReduceMin(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name) { - return Rename(topi::min(X, ToTvmArrayInt(axes), keep_dims), name); -} - -tvm::Tensor ReduceProd(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name) { - auto prod = [](tvm::Expr source, tvm::Array rdom) { - tvm::Var x("x", source.type()), y("y", source.type()); - tvm::Expr Rename_element = tvm::make_const(source.type(), 1.0f); - tvm::ir::CommReducer combiner = - tvm::ir::CommReducerNode::make({x}, {y}, {x * y}, {Rename_element}); - return tvm::ir::Reduce::make(combiner, {source}, rdom, tvm::make_const(tvm::Bool(1), true), 0); - }; - - return Rename(topi::CommReduce(X, ToTvmArrayInt(axes), prod, keep_dims, true), name); -} - -tvm::Tensor ReduceSum(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name) { - return Rename(topi::sum(X, ToTvmArrayInt(axes), keep_dims), name); -} - -tvm::Tensor ReduceSumSquare(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name) { - return Rename(topi::sum(Mul(X, X), ToTvmArrayInt(axes), keep_dims), name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/reduce_ops.h b/onnxruntime/core/codegen/mti/math/reduce_ops.h deleted file mode 100644 index f782df5e6515f..0000000000000 --- a/onnxruntime/core/codegen/mti/math/reduce_ops.h +++ /dev/null @@ -1,72 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor ArgMax(const tvm::Tensor& X, - int64_t axis, - bool keep_dims, - const std::string& name = "argmax"); - -tvm::Tensor ArgMin(const tvm::Tensor& X, - int64_t axis, - bool keep_dims, - const std::string& name = "argmin"); - -tvm::Tensor ReduceL1(const tvm::Tensor& X, - const std::vector& axes, - bool keep_dims, - const std::string& name = "reduce_l1"); - -tvm::Tensor ReduceL2(const tvm::Tensor& X, - const std::vector& axes, - bool keep_dims, - const std::string& name = "reduce_l2"); - -tvm::Tensor ReduceLogSum(const tvm::Tensor& X, - const std::vector& axes, - bool keep_dims, - const std::string& name = "reduce_log_sum"); - -tvm::Tensor ReduceLogSumExp(const tvm::Tensor& X, - const std::vector& axes, - bool keep_dims, - const std::string& name = "argmareduce_log_sum_exp"); - -tvm::Tensor ReduceMax(const tvm::Tensor& X, - const std::vector& axes, - bool keep_dims, - const std::string& name = "reduce_max"); - -tvm::Tensor ReduceMean(const tvm::Tensor& X, - const std::vector& axes, - bool keep_dims, - const std::string& name = "reduce_mean"); - -tvm::Tensor ReduceMin(const tvm::Tensor& X, - const std::vector& axes, - bool keep_dims, - const std::string& name = "reduce_min"); - -tvm::Tensor ReduceProd(const tvm::Tensor& X, - const std::vector& axes, - bool keep_dims, - const std::string& name = "reduce_prod"); - -tvm::Tensor ReduceSum(const tvm::Tensor& X, - const std::vector& axes, - bool keep_dims, - const std::string& name = "reduce_sum"); - -tvm::Tensor ReduceSumSquare(const tvm::Tensor& X, - const std::vector& axes, - bool keep_dims, - const std::string& name = "reduce_sum_square"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/softmax.cc b/onnxruntime/core/codegen/mti/math/softmax.cc deleted file mode 100644 index d7404137bb873..0000000000000 --- a/onnxruntime/core/codegen/mti/math/softmax.cc +++ /dev/null @@ -1,18 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/math/softmax.h" - -#include "core/codegen/mti/tensor/reshape_ops.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Softmax(const tvm::Tensor& input, int64_t axis, const std::string& name) { - tvm::Tensor flatten_t = Flatten(input, axis, "softmax_flatten"); - return Reshape(topi::nn::softmax(flatten_t, 1, name), input->shape, "softmax_reshape"); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/softmax.h b/onnxruntime/core/codegen/mti/math/softmax.h deleted file mode 100644 index fb16fbaeb56a2..0000000000000 --- a/onnxruntime/core/codegen/mti/math/softmax.h +++ /dev/null @@ -1,11 +0,0 @@ -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Softmax(const tvm::Tensor& input, int64_t axis, const std::string& name = "softmax"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/unary_ops.cc b/onnxruntime/core/codegen/mti/math/unary_ops.cc deleted file mode 100644 index ae732ea33e670..0000000000000 --- a/onnxruntime/core/codegen/mti/math/unary_ops.cc +++ /dev/null @@ -1,155 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/math/unary_ops.h" - -#include "core/codegen/common/settings.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include -#include -#include -#include - -// Using namespace topi for override operator +-*/ -using namespace topi; - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Abs(const tvm::Tensor& X, const std::string& name) { - return abs(X, name); -} - -tvm::Tensor Affine(const tvm::Tensor& X, float alpha, float beta, const std::string& name) { - tvm::Expr alphaExpr = tvm::make_const(X->dtype, alpha); - tvm::Expr betaExpr = tvm::make_const(X->dtype, beta); - return Rename(alphaExpr * X + betaExpr, name); -} - -tvm::Tensor Ceil(const tvm::Tensor& X, const std::string& name) { - return topi::ceil(X, name); -} - -tvm::Tensor Clip(const tvm::Tensor& X, tvm::Expr min_value, tvm::Expr max_value, const std::string& name) { - auto Y = tvm::compute( - X->shape, - [&](const tvm::Array& indices) { - return tvm::min(tvm::max(X(indices), min_value), max_value); - }, - name); - return Y; -} - -tvm::Tensor Elu(const tvm::Tensor& X, float alpha, const std::string& name) { - tvm::Expr alphaExpr = tvm::make_const(X->dtype, alpha); - return Rename(Relu(X) - alphaExpr * Relu(1 - Exp(X)), name); -} - -tvm::Tensor Exp(const tvm::Tensor& X, const std::string& name) { - return tvm::compute( - X->shape, - [&](const tvm::Array& indices) { - return tvm::exp(X(indices)); - }, - name); -} - -tvm::Tensor Floor(const tvm::Tensor& X, const std::string& name) { - return topi::floor(X, name); -} - -tvm::Tensor HardSigmoid(const tvm::Tensor& X, float alpha, float beta, const std::string& name) { - tvm::Expr alphaExpr = tvm::make_const(X->dtype, alpha); - tvm::Expr betaExpr = tvm::make_const(X->dtype, beta); - return maximum(0, minimum(1, alphaExpr * X + betaExpr), name); -} - -tvm::Tensor LeakyRelu(const tvm::Tensor& X, float alpha, const std::string& name) { - tvm::Expr alphaExpr = tvm::make_const(X->dtype, alpha); - return Rename(Relu(X) - alphaExpr * Relu(0 - X), name); -} - -tvm::Tensor Log(const tvm::Tensor& X, const std::string& name) { - return tvm::compute( - X->shape, - [&](const tvm::Array& indices) { - return tvm::log(X(indices)); - }, - name); -} - -tvm::Tensor Neg(const tvm::Tensor& X, const std::string& name) { - return negative(X, name); -} - -tvm::Tensor ParametricSoftplus(const tvm::Tensor& X, float alpha, float beta, const std::string& name) { - tvm::Expr alphaExpr = tvm::make_const(X->dtype, alpha); - tvm::Expr betaExpr = tvm::make_const(X->dtype, beta); - return Rename(alphaExpr * Softplus(betaExpr * X), name); -} - -tvm::Tensor Reciprocal(const tvm::Tensor& X, const std::string& name) { - return Rename(1 / X, name); -} - -tvm::Tensor Relu(const tvm::Tensor& X, const std::string& name) { - return maximum(X, 0, name); -} - -tvm::Tensor ScaledTanh(const tvm::Tensor& X, float alpha, float beta, const std::string& name) { - tvm::Expr alphaExpr = tvm::make_const(X->dtype, alpha); - tvm::Expr betaExpr = tvm::make_const(X->dtype, beta); - return Rename(alphaExpr * Tanh(betaExpr * X), name); -} - -tvm::Tensor Selu(const tvm::Tensor& X, float alpha, float gamma, const std::string& name) { - tvm::Expr alphaExpr = tvm::make_const(X->dtype, alpha); - tvm::Expr gammaExpr = tvm::make_const(X->dtype, gamma); - return Rename(gammaExpr * (-alphaExpr * Relu(1 - Exp(X)) + Relu(X)), name); -} - -tvm::Tensor Sigmoid(const tvm::Tensor& X, const std::string& name) { - return tvm::compute( - X->shape, - [&](const tvm::Array& indices) { - return tvm::ir::Select::make(X(indices) > 0, - 1 / (1 + tvm::exp(-X(indices))), - tvm::exp(X(indices)) / (tvm::exp(X(indices)) + 1)); - }, - name); -} - -tvm::Tensor SignNoZero(const tvm::Tensor& X, const std::string& name) { - return Rename(greater_equal(X, 0) * 2 - 1, name); -} - -tvm::Tensor Softplus(const tvm::Tensor& X, const std::string& name) { - return Rename(Log(1 + Exp(Neg(Abs(X)))) + Relu(X), name); -} - -tvm::Tensor Softsign(const tvm::Tensor& X, const std::string& name) { - return Rename(X / (1 + Abs(X)), name); -} - -tvm::Tensor Sqrt(const tvm::Tensor& X, const std::string& name) { - return sqrt(X, name); -} - -tvm::Tensor Tanh(const tvm::Tensor& X, const std::string& name) { - return tvm::compute( - X->shape, - [&](const tvm::Array& indices) { - return tvm::ir::Select::make(X(indices) < 0, - (tvm::exp(2 * X(indices)) - 1) / (tvm::exp(2 * X(indices)) + 1), - (1 - tvm::exp(-2 * X(indices))) / (1 + tvm::exp(-2 * X(indices)))); - }, - name); -} - -tvm::Tensor ThresholdedRelu(const tvm::Tensor& X, float alpha, const std::string& name) { - tvm::Expr alphaExpr = tvm::make_const(X->dtype, alpha); - return topi::where(greater(X, alphaExpr), X, topi::full_like(X, tvm::make_zero(X->dtype)), name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/math/unary_ops.h b/onnxruntime/core/codegen/mti/math/unary_ops.h deleted file mode 100644 index aeb336262e547..0000000000000 --- a/onnxruntime/core/codegen/mti/math/unary_ops.h +++ /dev/null @@ -1,36 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Abs(const tvm::Tensor& X, const std::string& name = "abs"); -tvm::Tensor Affine(const tvm::Tensor& X, float alpha, float beta, const std::string& name = "affine"); -tvm::Tensor Ceil(const tvm::Tensor& X, const std::string& name = "ceil"); -tvm::Tensor Clip(const tvm::Tensor& X, tvm::Expr min_value, tvm::Expr max_value, const std::string& name = "clip"); -tvm::Tensor Elu(const tvm::Tensor& X, float alpha, const std::string& name = "elu"); -tvm::Tensor Exp(const tvm::Tensor& X, const std::string& name = "exp"); -tvm::Tensor Floor(const tvm::Tensor& X, const std::string& name = "floor"); -tvm::Tensor HardSigmoid(const tvm::Tensor& X, float alpha, float beta, const std::string& name = "hard_sigmoid"); -tvm::Tensor LeakyRelu(const tvm::Tensor& X, float alpha, const std::string& name = "leaky_relu"); -tvm::Tensor Log(const tvm::Tensor& X, const std::string& name = "log"); -tvm::Tensor Neg(const tvm::Tensor& X, const std::string& name = "neg"); -tvm::Tensor ParametricSoftplus(const tvm::Tensor& X, float alpha, float beta, const std::string& name = "parametric_softplus"); -tvm::Tensor Reciprocal(const tvm::Tensor& X, const std::string& name = "reciprocal"); -tvm::Tensor Relu(const tvm::Tensor& X, const std::string& name = "relu"); -tvm::Tensor ScaledTanh(const tvm::Tensor& X, float alpha, float beta, const std::string& name = "scaled_tanh"); -tvm::Tensor Selu(const tvm::Tensor& X, float alpha, float gamma, const std::string& name = "selu"); -tvm::Tensor Sigmoid(const tvm::Tensor& X, const std::string& name = "sigmoid"); -tvm::Tensor SignNoZero(const tvm::Tensor& X, const std::string& name = "sign_no_zero"); -tvm::Tensor Softplus(const tvm::Tensor& X, const std::string& name = "softplus"); -tvm::Tensor Softsign(const tvm::Tensor& X, const std::string& name = "softsign"); -tvm::Tensor Sqrt(const tvm::Tensor& X, const std::string& name = "sqrt"); -tvm::Tensor Tanh(const tvm::Tensor& X, const std::string& name = "tanh"); -tvm::Tensor ThresholdedRelu(const tvm::Tensor& X, float alpha, const std::string& name = "thresholded_relu"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/mti_tvm_utils.cc b/onnxruntime/core/codegen/mti/mti_tvm_utils.cc deleted file mode 100644 index 8e73629c05614..0000000000000 --- a/onnxruntime/core/codegen/mti/mti_tvm_utils.cc +++ /dev/null @@ -1,203 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/mti_tvm_utils.h" - -#include "core/codegen/common/settings.h" -#include "core/codegen/mti/tensor/reshape_ops.h" -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Array ToTvmArray(gsl::span shape) { - tvm::Array arr; - for (size_t i = 0; i < shape.size(); ++i) { - arr.push_back(tvm::Expr(static_cast(shape[i]))); - } - return arr; -} - -tvm::Array ToTvmArrayInt(gsl::span shape) { - tvm::Array arr; - for (size_t i = 0; i < shape.size(); ++i) { - arr.push_back(shape[i]); - } - return arr; -} - -tvm::Expr SizeToDimension(const tvm::Array& shape, int64_t axis) { - tvm::Expr size(1); - auto rank = shape.size(); - if (static_cast(axis) != rank) { - axis = HandleNegativeAxis(axis, rank); - } - for (size_t d = 0; d < std::min(rank, static_cast(axis)); ++d) - size = tvm::ir::Simplify(size * shape[d]); - return size; -} - -tvm::Expr SizeFromDimension(const tvm::Array& shape, int64_t axis) { - tvm::Expr size(1); - auto rank = shape.size(); - if (static_cast(axis) != rank) { - axis = HandleNegativeAxis(axis, rank); - } - for (size_t d = static_cast(axis); d < rank; ++d) - size = tvm::ir::Simplify(size * shape[d]); - return size; -} - -tvm::Expr RoundUp(tvm::Expr value, tvm::Expr alignment) { - return tvm::ir::Simplify((value + alignment - 1) / alignment * alignment); -} - -tvm::Array ConcatShapes( - const tvm::Array& shape1, - const tvm::Array& shape2) { - tvm::Array result; - for (size_t i = 0; i < shape1.size(); i++) - result.push_back(shape1[i]); - for (size_t i = 0; i < shape2.size(); i++) - result.push_back(shape2[i]); - return result; -} - -tvm::Tensor Rename(tvm::Tensor X, const std::string& name) { - const_cast(X->op->name) = name; - return X; -} - -tvm::Array SliceShape(const tvm::Array& shape, const std::vector& axes) { - tvm::Array new_shape; - for (auto axis : axes) { - CHECK(axis < static_cast(shape.size())); - new_shape.push_back(shape[axis]); - } - return new_shape; -} - -tvm::Array SliceShapeFromDimension(const tvm::Array& shape, int64_t axis) { - int64_t rank = static_cast(shape.size()); - axis = HandleNegativeAxis(axis, rank); - std::vector axes; - for (auto i = axis; i < rank; ++i) - axes.push_back(i); - return SliceShape(shape, axes); -} - -tvm::Array SliceShapeToDimension(const tvm::Array& shape, int64_t axis) { - int64_t rank = static_cast(shape.size()); - axis = HandleNegativeAxis(axis, rank); - std::vector axes; - for (auto i = 0; i < axis; ++i) - axes.push_back(i); - return SliceShape(shape, axes); -} - -bool IsOne(const tvm::Array& shape, int64_t axis) { - int64_t rank = static_cast(shape.size()); - axis = HandleNegativeAxis(axis, rank); - const auto& dim = shape[axis]; - auto* p = tvm::as_const_int(dim); - return p != nullptr && *p == 1; -} - -tvm::Tensor Promote(const tvm::Expr& expr, const tvm::Array& shape, const std::string& name) { - return tvm::compute( - shape, - [&](const tvm::Array&) { - return expr; - }, - name); -} - -void DumpTVMModuleToFile(const std::string& filename, tvm::runtime::Module& module) { - const codegen::CodeGenSettings& settings = codegen::CodeGenSettings::Instance(); - if (!settings.HasOption(codegen::CodeGenSettings::kCodeGenDumpModule)) - return; - - // ISSUE: note that all option values are converted to lower case. It doesn't cause - // any issue currently, because all supported formats (i.e. file exts) are of lower case. - // Just keep in mind that we might have issue if somehow we started to support dump - // formats with upper case, although it's quite unlikely. - std::string format = settings.GetOptionValue(codegen::CodeGenSettings::kCodeGenDumpModule); - std::string module_filename = filename + "." + format; - module->SaveToFile(module_filename, format); -} - -tvm::Tensor MakeZeroTensor(const tvm::Array& shape, - HalideIR::Type type, - const std::string& name) { - auto l = [&](const tvm::Array& /*indices*/) { - return tvm::make_zero(type); - }; - return tvm::compute(shape, l, name); -} - -bool BroadcastDim(const tvm::Array& shape, size_t i, size_t output_rank, tvm::Expr& dim) { - if (i >= output_rank - shape.size()) { - auto new_dim = shape[shape.size() - output_rank + i]; - if (tvm::ir::Equal(new_dim, dim)) - return true; - - const int64_t* p_new = tvm::as_const_int(new_dim); - if (p_new != nullptr && *p_new == 1) { - return true; - } else { - const int64_t* p_old = tvm::as_const_int(dim); - if (p_old != nullptr && *p_old == 1) { - dim = new_dim; - return true; - } - } - return false; - } - // auto broadcast to outer dims - return true; -} - -tvm::Array MakeInputsForExtern(const tvm::Array& inputs, const std::string& name) { - // note that currently TVM StorageFlatten creates strides like max(symbolic_dim, 1) - // which is not zero when checking symbolic_dim - max(symbolic_dim, 1) - // then triggers error like: Trying to bind compact buffer to strided one - // here's a workaround to reshape inputs to avoid that - tvm::Array fixed_inputs; - for (size_t idx_input = 0; idx_input < inputs.size(); ++idx_input) { - const auto& input = inputs[idx_input]; - tvm::Array fixed_shape; - if (input->shape.size() > 0) { - // stride compute does not use dim 0, so directly push to fixed_shape - fixed_shape.push_back(input->shape[0]); - bool need_fix = false; - for (size_t idx_dim = 1; idx_dim < input->shape.size(); ++idx_dim) { - const auto& dim = input->shape[idx_dim]; - if (tvm::as_const_int(dim) == nullptr) { - fixed_shape.push_back(tvm::max(dim, tvm::make_const(HalideIR::Int(32), 1))); - need_fix = true; - } else { - fixed_shape.push_back(dim); - } - } - if (need_fix) { - fixed_inputs.push_back(tvm_codegen::Reshape(input, fixed_shape, name + "_" + std::to_string(idx_input))); - continue; - } - } - // no fix needed - fixed_inputs.push_back(input); - } - return fixed_inputs; -} - -// Make sure idx is clamped in the range of [-bound, bound - 1] -tvm::Expr ClampIndex(const tvm::Expr& idx, const tvm::Expr& bound) { - // when idx >= 0, we take tvm::max(..., 0), because (idx < 0) is 0 - // when idx < 0, we take bound + tvm::max(...), because tvm::max(idx, 0) is 0 - return tvm::max(tvm::min(idx, bound - 1), 0) + - (idx < 0) * (bound + tvm::max(idx, -bound)); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/mti_tvm_utils.h b/onnxruntime/core/codegen/mti/mti_tvm_utils.h deleted file mode 100644 index c2a14106c1686..0000000000000 --- a/onnxruntime/core/codegen/mti/mti_tvm_utils.h +++ /dev/null @@ -1,71 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once - -#include -#include -#include -#include -#include "core/codegen/mti/common.h" - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Array ToTvmArray(gsl::span shape); - -tvm::Array ToTvmArrayInt(gsl::span shape); - -// Helper function to compute sub shape size to axis (not included) -tvm::Expr SizeToDimension(const tvm::Array& shape, int64_t axis); - -// Helper function to compute sub shape size from axis (included) -tvm::Expr SizeFromDimension(const tvm::Array& shape, int64_t axis); - -// Helper function to align -tvm::Expr RoundUp(tvm::Expr value, tvm::Expr alignment); - -tvm::Array ConcatShapes( - const tvm::Array& shape1, - const tvm::Array& shape2); - -// Helper function to rename tvm::Tensor -tvm::Tensor Rename(tvm::Tensor X, const std::string& name); - -// Helper function to slice TVM shape -tvm::Array SliceShape(const tvm::Array& shape, const std::vector& axes); - -// Helper function to slice TVM shape from axis (inclusive). -// Basically, this function returns the shape of [axis, shape.size()-1] -tvm::Array SliceShapeFromDimension(const tvm::Array& shape, int64_t axis); - -// this function returns the shape of [0, axis-1] -tvm::Array SliceShapeToDimension(const tvm::Array& shape, int64_t axis); - -// check if dimension is 1 -bool IsOne(const tvm::Array& shape, int64_t axis); - -// Helper function to convert tvm::Expr to tvm::Tensor -tvm::Tensor Promote(const tvm::Expr& expr, - const tvm::Array& shape, - const std::string& name = "PromoteExpr"); - -tvm::Tensor MakeZeroTensor(const tvm::Array& shape, HalideIR::Type type, const std::string& name); - -void DumpTVMModuleToFile(const std::string& filename, tvm::runtime::Module& module); - -bool BroadcastDim(const tvm::Array& shape, size_t i, size_t output_rank, tvm::Expr& dim); - -inline int64_t HandleNegativeAxis(int64_t axis, int64_t rank) { - MTI_ASSERT(axis >= -rank && axis <= rank - 1); - return axis = axis < 0 ? (axis + rank) : axis; -} - -// Make sure idx is clamped in the range of [-bound, bound - 1] -tvm::Expr ClampIndex(const tvm::Expr& idx, const tvm::Expr& bound); - -// Helper function to workaround tvm ExternOp issue when input has symbolic dimensions -tvm::Array MakeInputsForExtern(const tvm::Array& inputs, const std::string& name = "make_inputs_for_extern"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/nn/conv_ops.cc b/onnxruntime/core/codegen/mti/nn/conv_ops.cc deleted file mode 100644 index e2d4acc8843ad..0000000000000 --- a/onnxruntime/core/codegen/mti/nn/conv_ops.cc +++ /dev/null @@ -1,193 +0,0 @@ -#include "core/codegen/mti/nn/conv_ops.h" - -#include "core/codegen/mti/math/matmul_ops.h" -#include "core/codegen/mti/tensor/pad_ops.h" -#include "core/codegen/mti/tensor/reshape_ops.h" -#include "core/codegen/mti/tensor/transpose.h" - -namespace onnxruntime { -namespace tvm_codegen { - -static tvm::Tensor PadTensor1D(const tvm::Tensor& input, - const tvm::Array& padding, - size_t width_axis, - const std::string& name) { - auto pad_left = padding[0]; - auto pad_right = padding[1]; - - tvm::Array pad_before(std::vector(input->shape.size(), 0)); - pad_before.Set(width_axis, pad_left); - tvm::Array pad_after(std::vector(input->shape.size(), 0)); - pad_after.Set(width_axis, pad_right); - - const int64_t* padding_w0 = tvm::as_const_int(pad_left); - const int64_t* padding_w1 = tvm::as_const_int(pad_right); - - const bool do_pad = ((padding_w0 != nullptr && *padding_w0) || - (padding_w1 != nullptr && *padding_w1)); - - return do_pad ? Pad(input, pad_before, pad_after, - 0, "constant", name + "_input_padded") - : input; -} - -tvm::Tensor Conv1D(const tvm::Tensor& input, - const tvm::Tensor& filter, - const tvm::Array& out_shape, - const tvm::Array& stride, - const tvm::Array& padding, - const std::string& name) { - size_t channel_axis = 1; - size_t width_axis = 2; - - auto stride_width = stride[width_axis - 2]; - - auto input_padded = PadTensor1D(input, padding, width_axis, name); - auto rc = tvm::reduce_axis((tvm::Range(0, filter->shape[1])), "rc"); - auto rx = tvm::reduce_axis((tvm::Range(0, filter->shape[2])), "rx"); - - return tvm::compute( - out_shape, - [&](const tvm::Array& output) { - tvm::Array indices; - for (const tvm::Var& var : output) { - indices.push_back(var); - } - indices.Set(channel_axis, rc); - indices.Set(width_axis, output[width_axis] * stride_width + rx); - - return tvm::sum(input_padded(indices) * filter({output[1], rc, rx}), - {rc, rx}); - }, - name); -} - -tvm::Tensor Conv2D(const tvm::Tensor& input, - const tvm::Tensor& filter, - const tvm::Array& output_shape, - const tvm::Array& stride, - const tvm::Array& padding, - const std::string& name) { - return Conv2D_native(input, filter, output_shape, stride, padding); -} - -static tvm::Tensor PadTensor2D(const tvm::Tensor& input, - const tvm::Array& padding, - size_t height_axis, - size_t width_axis, - const std::string& name) { - auto pad_top = padding[0]; - auto pad_left = padding[1]; - auto pad_bottom = padding[2]; - auto pad_right = padding[3]; - - tvm::Array pad_before(std::vector(input->shape.size(), 0)); - pad_before.Set(height_axis, pad_top); - pad_before.Set(width_axis, pad_left); - - tvm::Array pad_after(std::vector(input->shape.size(), 0)); - pad_after.Set(height_axis, pad_bottom); - pad_after.Set(width_axis, pad_right); - - const int64_t* padding_h0 = tvm::as_const_int(pad_top); - const int64_t* padding_w0 = tvm::as_const_int(pad_left); - const int64_t* padding_h1 = tvm::as_const_int(pad_bottom); - const int64_t* padding_w1 = tvm::as_const_int(pad_right); - - const bool do_pad = ((padding_h0 != nullptr && *padding_h0) || - (padding_w0 != nullptr && *padding_w0)) || - ((padding_h1 != nullptr && *padding_h1) || - (padding_w1 != nullptr && *padding_w1)); - - return do_pad ? Pad(input, pad_before, pad_after, - 0, "constant", name + "_input_padded") - : input; -} - -tvm::Tensor Conv2D_native(const tvm::Tensor& input, - const tvm::Tensor& filter, - const tvm::Array& out_shape, - const tvm::Array& stride, - const tvm::Array& padding, - const std::string& name) { - size_t channel_axis = 1; - size_t height_axis = 2; - size_t width_axis = 3; - - auto stride_height = stride[height_axis - 2]; - auto stride_width = stride[width_axis - 2]; - - auto input_padded = PadTensor2D(input, padding, height_axis, width_axis, name); - - auto rc = tvm::reduce_axis((tvm::Range(0, filter->shape[1])), "rc"); - auto ry = tvm::reduce_axis((tvm::Range(0, filter->shape[2])), "ry"); - auto rx = tvm::reduce_axis((tvm::Range(0, filter->shape[3])), "rx"); - - return tvm::compute( - out_shape, - [&](const tvm::Array& output) { - tvm::Array indices; - for (const tvm::Var& var : output) { - indices.push_back(var); - } - indices.Set(channel_axis, rc); - indices.Set(height_axis, output[height_axis] * stride_height + ry); - indices.Set(width_axis, output[width_axis] * stride_width + rx); - - return tvm::sum(input_padded(indices) * filter({output[1], rc, ry, rx}), - {rc, ry, rx}); - }, - name); -} - -tvm::Tensor Conv2D_gemm(const tvm::Tensor& input, - const tvm::Tensor& filter, - const tvm::Array& out_shape, - const tvm::Array& stride, - const tvm::Array& padding, - const std::string& name) { - size_t height_axis = 2; - size_t width_axis = 3; - - auto stride_height = stride[height_axis - 2]; - auto stride_width = stride[width_axis - 2]; - - auto input_padded = PadTensor2D(input, padding, height_axis, width_axis, name); - - tvm::Array img_col_tmp(std::vector(6, 0)); - img_col_tmp.Set(0, out_shape[0]); - img_col_tmp.Set(1, out_shape[2]); - img_col_tmp.Set(2, out_shape[3]); - img_col_tmp.Set(3, filter->shape[1]); - img_col_tmp.Set(4, filter->shape[2]); - img_col_tmp.Set(5, filter->shape[3]); - - auto img_col = tvm::compute( - img_col_tmp, - [&](const tvm::Array& output) { - tvm::Array indices; - indices.push_back(output[0]); - indices.push_back(output[3]); - indices.push_back(output[1] * stride_height + output[4]); - indices.push_back(output[2] * stride_width + output[5]); - return input_padded(indices); - }, - name); - - tvm::Array input_col_shape(std::vector(2, 0)); - input_col_shape.Set(0, img_col_tmp[1] * img_col_tmp[2]); - input_col_shape.Set(1, img_col_tmp[3] * img_col_tmp[4] * img_col_tmp[5]); - auto input_col = Reshape(img_col, input_col_shape); - - tvm::Array filter_row_shape(std::vector(2, 0)); - filter_row_shape.Set(0, filter->shape[0]); - filter_row_shape.Set(1, filter->shape[1] * filter->shape[2] * filter->shape[3]); - auto filter_row = Reshape(filter, filter_row_shape, name); - - auto Y = MatMul2D(input_col, filter_row, false, true, name); - auto Y_T = Transpose(Y, /*axes=*/{}, name); - return Reshape(Y_T, out_shape, name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/nn/conv_ops.h b/onnxruntime/core/codegen/mti/nn/conv_ops.h deleted file mode 100644 index 1396c216865a7..0000000000000 --- a/onnxruntime/core/codegen/mti/nn/conv_ops.h +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Conv1D(const tvm::Tensor& input, - const tvm::Tensor& filter, - const tvm::Array& output_shape, - const tvm::Array& stride, - const tvm::Array& padding, - const std::string& name = "conv1d"); - -tvm::Tensor Conv2D(const tvm::Tensor& input, - const tvm::Tensor& filter, - const tvm::Array& output_shape, - const tvm::Array& stride, - const tvm::Array& padding, - const std::string& name = "conv2d"); - -tvm::Tensor Conv2D_native(const tvm::Tensor& input, - const tvm::Tensor& filter, - const tvm::Array& output_shape, - const tvm::Array& stride, - const tvm::Array& padding, - const std::string& name = "conv2d_native"); - -tvm::Tensor Conv2D_gemm(const tvm::Tensor& input, - const tvm::Tensor& filter, - const tvm::Array& output_shape, - const tvm::Array& stride, - const tvm::Array& padding, - const std::string& name = "conv2d_gemm"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/nn/lstm.cc b/onnxruntime/core/codegen/mti/nn/lstm.cc deleted file mode 100644 index 1148b0924e869..0000000000000 --- a/onnxruntime/core/codegen/mti/nn/lstm.cc +++ /dev/null @@ -1,140 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/nn/lstm.h" - -#include "core/codegen/mti/math/binary_ops.h" -#include "core/codegen/mti/math/unary_ops.h" -#include "core/codegen/mti/math/matmul_ops.h" -#include "core/codegen/mti/math/reduce_ops.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/reshape_ops.h" -#include "core/codegen/mti/tensor/split.h" - -namespace onnxruntime { -namespace tvm_codegen { - -/* -`X` - input tensor -`i` - input gate -`o` - output gate -`f` - forget gate -`c` - cell gate -`t` - time step (t-1 means previous time step) - -`W[iofc]` - W parameter weight matrix for input, output, forget, and cell gates -`R[iofc]` - R recurrence weight matrix for input, output, forget, and cell gates -`Wb[iofc]` - W bias vectors for input, output, forget, and cell gates -`Rb[iofc]` - R bias vectors for input, output, forget, and cell gates -`P[iof]` - P peephole weight vector for input, output, and forget gates -`WB[iofc]` - W parameter weight matrix for backward input, output, forget, and cell gates -`RB[iofc]` - R recurrence weight matrix for backward input, output, forget, and cell gates -`WBb[iofc]` - W bias vectors for backward input, output, forget, and cell gates -`RBb[iofc]` - R bias vectors for backward input, output, forget, and cell gates -`PB[iof]` - P peephole weight vector for backward input, output, and forget gates - -`H` - Hidden state -`num_directions` - 2 if direction == bidirectional else 1 - -Equations (Default: f=Sigmoid, g=Tanh, h=Tanh): - it = f(Xt*(Wi^T) + Ht-1*(Ri^T) + Pi (.) Ct-1 + Wbi + Rbi) - ft = f(Xt*(Wf^T) + Ht-1*(Rf^T) + Pf (.) Ct-1 + Wbf + Rbf) - ct = g(Xt*(Wc^T) + Ht-1*(Rc^T) + Wbc + Rbc) - Ct = ft (.) Ct-1 + it (.) ct - ot = f(Xt*(Wo^T) + Ht-1*(Ro^T) + Po (.) Ct + Wbo + Rbo) - Ht = ot (.) h(Ct) -*/ - -void LSTM_cell( - const LSTMAttributes& lstm_attrs, - const tvm::Tensor& X, - const tvm::Tensor& W, - const tvm::Tensor& R, - const tvm::Tensor& B, - bool has_B, - const tvm::Tensor& prev_H, - const tvm::Tensor& prev_C, - const tvm::Tensor& P, - bool has_P, - tvm::Tensor& Y_h, - tvm::Tensor& Y_c) { - // Input projection: Xt*(W[iofc]^T) for forward direction or Xt*(WB[iofc]^T) for reverse direction - // (batch_size, input_size) * trans(4 * hidden_size, input_size) => (batch_size, 4 * hidden_size) - tvm::Tensor input_proj = MatMul2D(X, W, /*trans_a*/ false, /*trans_b*/ true); - - // Hidden projection: Ht-1*(R[iofc]^T) for forward direction or Ht-1*(RB[iofc]^T) for reverse direction - // (batch_size, hidden_size) * trans(4 * hidden_size, hidden_size) => (batch_size, 4 * hidden_size) - tvm::Tensor hidden_proj = MatMul2D(prev_H, R, /*trans_a*/ false, /*trans_b*/ true); - - // (batch_size, 4 * hidden_size) - tvm::Tensor sum_proj = Add(input_proj, hidden_proj); - - // Concatenation of [Wb[iofc], Rb[iofc]] or [WBb[iofc], RBb[iofc]] - if (has_B) { - // (8 * hidden_size) -> (2, 4 * hidden_size) -> (1, 4 * hidden_size), should be done in const folding - tvm::Tensor reduce_B = - ReduceSum(Reshape(B, {2, 4 * static_cast(lstm_attrs.hidden_size)}), {0}, /*keep_dims*/ true); - // (batch_size, 4 * hidden_size) via broadcasting reduce_B - sum_proj = Add(sum_proj, reduce_B); - } - - std::vector iofc_sum_split_sizes(4, lstm_attrs.hidden_size); - // Split sum_proj into iofc, where each gate proj is of (batch_size, hidden_size) - tvm::Array iofc_sum_projs = Split(sum_proj, ToTvmArray(iofc_sum_split_sizes), /*axis*/ 1); - MTI_ASSERT(iofc_sum_projs.size() == 4); - tvm::Tensor i_proj = iofc_sum_projs[0], - o_proj = iofc_sum_projs[1], - f_proj = iofc_sum_projs[2], - c_proj = iofc_sum_projs[3]; - - tvm::Tensor P_i, P_o, P_f; - if (has_P) { - std::vector iof_p_split_sizes(3, lstm_attrs.hidden_size); - // Split P into P_i, P_o, P_f, in const pre-processing (P_i, P_f might be merged?) - // where each P_[iof] has the shape of (hidden_size) - tvm::Array iof_P_projs = Split(P, ToTvmArray(iof_p_split_sizes), /*axis*/ 0); - MTI_ASSERT(iof_P_projs.size() == 3); - P_i = iof_P_projs[0], - P_o = iof_P_projs[1], - P_f = iof_P_projs[2]; - - // (batch_size, hidden_size) via broadcasting P_[if] - i_proj = Add(i_proj, Mul(P_i, prev_C)); - f_proj = Add(f_proj, Mul(P_f, prev_C)); - } - - // TODO: handle more general cases for activations f, h, g and activation_alpha and - // activation_beta. We may consider to move some code such as ActivationInfo from deep_cpu_lstm - // into a common header file, because the code can be used here. - - // Note that by default f = Sigmoid, g = Tanh, h = Tanh - - // it = f(Xt*(Wi^T) + Ht-1*(Ri^T) + Pi (.) Ct-1 + Wbi + Rbi) - // shape: (batch_size, hidden_size) - tvm::Tensor i_t = Sigmoid(i_proj); - // ft = f(Xt*(Wf^T) + Ht-1*(Rf^T) + Pf (.) Ct-1 + Wbf + Rbf) - // shape: (batch_size, hidden_size) - tvm::Tensor f_t = Sigmoid(f_proj); - // ct = g(Xt*(Wc^T) + Ht-1*(Rc^T) + Wbc + Rbc) - // shape: (batch_size, hidden_size) - tvm::Tensor c_t = Tanh(c_proj); - - // Ct = ft (.) Ct-1 + it (.) ct - // shape: (batch_size, hidden_size) - Y_c = Add(Mul(f_t, prev_C), Mul(i_t, c_t), Y_c->op->name); - - // ot = f(Xt*(Wo^T) + Ht-1*(Ro^T) + Po (.) Ct + Wbo + Rbo) - // shape: (batch_size, hidden_size) - if (has_P) { - o_proj = Add(o_proj, Mul(P_o, Y_c)); - } - // ot = f(Xt*(Wo^T) + Ht-1*(Ro^T) + Po (.) Ct + Wbo + Rbo) - // shape: (batch_size, hidden_size) - o_proj = Sigmoid(o_proj); - // Ht = ot (.) h(Ct) - // shape: (batch_size, hidden_size) - Y_h = Mul(o_proj, Tanh(Y_c), Y_h->op->name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/nn/lstm.h b/onnxruntime/core/codegen/mti/nn/lstm.h deleted file mode 100644 index 851fa880c4427..0000000000000 --- a/onnxruntime/core/codegen/mti/nn/lstm.h +++ /dev/null @@ -1,35 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -// A bubble now. But don't remove it -// TODO: refactor the LSTMcell building to a tvm function -// and move it here - -namespace onnxruntime { -namespace tvm_codegen { - -struct LSTMAttributes { - LSTMAttributes(int64_t hidden_size_p) : hidden_size(hidden_size_p) {} - int64_t hidden_size; -}; - -void LSTM_cell( - const LSTMAttributes& lstm_attrs, - const tvm::Tensor& X, - const tvm::Tensor& W, - const tvm::Tensor& R, - const tvm::Tensor& B, - bool has_B, - const tvm::Tensor& prev_H, - const tvm::Tensor& prev_C, - const tvm::Tensor& P, - bool has_P, - tvm::Tensor& Y_h, - tvm::Tensor& Y_c); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/nn/pool_ops.cc b/onnxruntime/core/codegen/mti/nn/pool_ops.cc deleted file mode 100644 index 868a14748cabc..0000000000000 --- a/onnxruntime/core/codegen/mti/nn/pool_ops.cc +++ /dev/null @@ -1,63 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/nn/pool_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/mlas/inc/mlas.h" -#include "core/providers/cpu/nn/pool_attributes.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// TODO: topi only support 2d-pool, MaxPool1d and MaxPool3d will need to be added if necessary. -// only support version < 8 for topi doesn't come with implementation to output index tensor -tvm::Tensor MaxPool(const tvm::Tensor& input, - const PoolAttributes& pool_attrs, - const tvm::Array& /*output_shape*/, - const std::string& /*name*/) { - return topi::nn::pool(input, - ToTvmArray(pool_attrs.kernel_shape), - ToTvmArray(pool_attrs.strides), - ToTvmArray(pool_attrs.pads), - /*pool_type*/ topi::nn::kMaxPool, - /*ceil_mode*/ false, - /*layout*/ pool_attrs.storage_order == 0 ? "NCWH" : "NCHW", - pool_attrs.count_include_pad); -} - -tvm::Tensor AveragePool(const tvm::Tensor& input, - const PoolAttributes& pool_attrs, - const tvm::Array& /*output_shape*/, - const std::string& /*name*/) { - return topi::nn::pool(input, - ToTvmArray(pool_attrs.kernel_shape), - ToTvmArray(pool_attrs.strides), - ToTvmArray(pool_attrs.pads), - /*pool_type*/ topi::nn::kAvgPool, - /*ceil_mode*/ false, - /*layout*/ "NCHW", - pool_attrs.count_include_pad); -} - -tvm::Tensor GlobalMaxPool(const tvm::Tensor& input, - const PoolAttributes& /*pool_attrs*/, - const tvm::Array& /*output_shape*/, - const std::string& /*name*/) { - return topi::nn::global_pool(input, - /*pool_type*/ topi::nn::kMaxPool, - /*layout*/ "NCHW"); -} - -tvm::Tensor GlobalAveragePool(const tvm::Tensor& input, - const PoolAttributes& /*pool_attrs*/, - const tvm::Array& /*output_shape*/, - const std::string& /*name*/) { - return topi::nn::global_pool(input, - /*pool_type*/ topi::nn::kAvgPool, - /*layout*/ "NCHW"); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/nn/pool_ops.h b/onnxruntime/core/codegen/mti/nn/pool_ops.h deleted file mode 100644 index d381f9ddff859..0000000000000 --- a/onnxruntime/core/codegen/mti/nn/pool_ops.h +++ /dev/null @@ -1,36 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { - -// Forward declaration -struct PoolAttributes; - -namespace tvm_codegen { - -tvm::Tensor MaxPool(const tvm::Tensor& input, - const PoolAttributes& pool_attrs, - const tvm::Array& output_shape, - const std::string& name = "max_pool"); - -tvm::Tensor AveragePool(const tvm::Tensor& input, - const PoolAttributes& pool_attrs, - const tvm::Array& output_shape, - const std::string& name = "average_pool"); - -tvm::Tensor GlobalMaxPool(const tvm::Tensor& input, - const PoolAttributes& pool_attrs, - const tvm::Array& output_shape, - const std::string& name = "global_max_pool"); - -tvm::Tensor GlobalAveragePool(const tvm::Tensor& input, - const PoolAttributes& pool_attrs, - const tvm::Array& output_shape, - const std::string& name = "global_average_pool"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/cast_ops.cc b/onnxruntime/core/codegen/mti/tensor/cast_ops.cc deleted file mode 100644 index a8fc86488d82b..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/cast_ops.cc +++ /dev/null @@ -1,37 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/cast_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Cast(const tvm::Tensor& X, tvm::Type type, const std::string& name) { - return topi::cast(X, type, name); -} - -// handle cases where bool is reprented as uint8 (e.g. in ONNX). -tvm::Tensor CastToUInt8Bool(const tvm::Tensor& X, const std::string& name) { - return tvm::compute( - X->shape, - [&](const tvm::Array& indices) { - auto val = X(indices); - // A special cast from float16 to bool, first cast up to float32, - // to workaround a float16 bug in many TVM backends. - // Intel Skylake is one of them. https://github.com/dmlc/tvm/issues/2959 - // TODO: remove it, after TVM is fixed - if (X->dtype == HalideIR::Float(16)) - val = tvm::cast(HalideIR::Float(32), val); - return tvm::ir::Select::make(topi::equal(val, tvm::make_zero(val.type())), - tvm::make_zero(HalideIR::UInt(8)), - tvm::make_const(HalideIR::UInt(8), 1)); - }, - name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/cast_ops.h b/onnxruntime/core/codegen/mti/tensor/cast_ops.h deleted file mode 100644 index 02f6f9cb1fde7..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/cast_ops.h +++ /dev/null @@ -1,15 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Cast(const tvm::Tensor& X, tvm::Type type, const std::string& name = "cast"); -tvm::Tensor CastToUInt8Bool(const tvm::Tensor& X, const std::string& name = "cast_uint8_bool"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/concat_ops.cc b/onnxruntime/core/codegen/mti/tensor/concat_ops.cc deleted file mode 100644 index 3394d5b7e00a2..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/concat_ops.cc +++ /dev/null @@ -1,83 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/concat_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Concat(const tvm::Array& inputs, - int64_t axis, - const std::string& name) { - return ConcatSafe(inputs, axis, name); -} - -// Note topi's implementation requires control flow within iterations to avoid out-of-bound access. -// Therefore, MTI implements a ConcatSafe that does not have out-of-bound access, -// and does not requires control or predicate. -tvm::Tensor ConcatSafe(const tvm::Array& inputs, - int64_t axis, - const std::string& name) { - axis = HandleNegativeAxis(axis, gsl::narrow(inputs[0]->shape.size())); - MTI_ASSERT(axis < gsl::narrow(inputs[0]->shape.size()) && "axis out of bounds"); - - tvm::Array axis_sizes; - for (auto t : inputs) { - axis_sizes.push_back(t->shape[axis]); - } - - tvm::Expr join_size = axis_sizes[0]; - for (size_t i = 1; i < axis_sizes.size(); ++i) { - join_size += axis_sizes[i]; - } - join_size = tvm::ir::Simplify(join_size); - tvm::Array out_shape; - for (size_t i = 0; i < inputs[0]->shape.size(); ++i) { - out_shape.push_back(i == gsl::narrow(axis) ? join_size : inputs[0]->shape[i]); - } - - return tvm::compute( - out_shape, [&](const tvm::Array& ovars) { - tvm::Array indices; - - // preset - tvm::Expr min = 0; - tvm::Expr extent = axis_sizes[0]; - tvm::Expr offset = 0; - tvm::Expr ret; - - // input i = 0 - for (size_t j = 0; j < ovars.size(); ++j) { - if (j == gsl::narrow(axis)) { - tvm::Expr ivar = ovars[j]; - indices.push_back(tvm::max(tvm::min(ivar, min + extent - 1), min)); - } else { - indices.push_back(ovars[j]); - } - } - ret = inputs[0](indices); - - for (size_t i = 1; i < inputs.size(); ++i) { - offset += extent; - tvm::Expr min = 0; - extent = axis_sizes[i]; - auto j = gsl::narrow(axis); - tvm::Expr ivar = ovars[j] - offset; - indices.Set(j, tvm::max(tvm::min(ivar, min + extent - 1), min)); - - ret = tvm::ir::Select::make(ivar >= 0, - inputs[i](indices), - ret); - } - - return ret; - }, - name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/concat_ops.h b/onnxruntime/core/codegen/mti/tensor/concat_ops.h deleted file mode 100644 index 153afebb44615..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/concat_ops.h +++ /dev/null @@ -1,15 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Concat(const tvm::Array& inputs, int64_t axis, const std::string& name = "concat"); -tvm::Tensor ConcatSafe(const tvm::Array& inputs, int64_t axis, const std::string& name = "concat_safe"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/crop.cc b/onnxruntime/core/codegen/mti/tensor/crop.cc deleted file mode 100644 index 3fe569100df12..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/crop.cc +++ /dev/null @@ -1,58 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/crop.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Crop(const tvm::Tensor& t, - const tvm::Array& border, - const tvm::Array& scale, - const std::string& name) { - MTI_ASSERT(t->shape.size() == 4); - tvm::Expr N = t->shape[0]; - tvm::Expr C = t->shape[1]; - tvm::Expr H = t->shape[2]; - tvm::Expr W = t->shape[3]; - - MTI_ASSERT(border.size() == 4); - tvm::Expr leftBorder = border[0]; - tvm::Expr topBorder = border[1]; - tvm::Expr rightBorder = border[2]; - tvm::Expr bottomBorder = border[3]; - - tvm::Expr bottomLimit = H - bottomBorder; - tvm::Expr rightLimit = W - rightBorder; - - if (!scale.empty()) { - CHECK_EQ(scale.size(), 2); - bottomLimit = topBorder + scale[0]; - rightLimit = leftBorder + scale[1]; - } - - tvm::Array output_shape; - output_shape.push_back(tvm::ir::Simplify(N)); - output_shape.push_back(tvm::ir::Simplify(C)); - output_shape.push_back(tvm::ir::Simplify(bottomLimit - topBorder)); - output_shape.push_back(tvm::ir::Simplify(rightLimit - leftBorder)); - - auto l = [&](const tvm::Array& ovars) { - tvm::Array indices; - - indices.push_back(tvm::min(ovars[0], output_shape[0] - 1)); - indices.push_back(tvm::min(ovars[1], output_shape[1] - 1)); - indices.push_back(tvm::min(topBorder + ovars[2], topBorder + output_shape[2] - 1)); - indices.push_back(tvm::min(leftBorder + ovars[3], leftBorder + output_shape[3] - 1)); - - return t(indices); - }; - - return tvm::compute(output_shape, l, name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/crop.h b/onnxruntime/core/codegen/mti/tensor/crop.h deleted file mode 100644 index ffb6a05c70504..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/crop.h +++ /dev/null @@ -1,17 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Crop(const tvm::Tensor& t, - const tvm::Array& border, - const tvm::Array& scale = {}, - const std::string& name = "crop"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/expand.cc b/onnxruntime/core/codegen/mti/tensor/expand.cc deleted file mode 100644 index cdac4f56e1f9f..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/expand.cc +++ /dev/null @@ -1,30 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/expand.h" -#include "core/codegen/mti/common.h" - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Expand(const tvm::Tensor& X, const tvm::Array& new_shape, const std::string& name) { - MTI_ASSERT(new_shape.size() >= X->shape.size()); - return tvm::compute( - new_shape, - [&](const tvm::Array& out_indices) { - tvm::Array indices; - size_t broadcasted_rank = new_shape.size() - X->shape.size(); - for (size_t d = broadcasted_rank; d < new_shape.size(); ++d) { - if (tvm::is_const_int(X->shape[d - broadcasted_rank], 1)) { - indices.push_back(tvm::make_zero(HalideIR::Int(32))); - } else { - indices.push_back(out_indices[d]); - } - } - return X(indices); - }, - name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/expand.h b/onnxruntime/core/codegen/mti/tensor/expand.h deleted file mode 100644 index d66d41aeb0194..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/expand.h +++ /dev/null @@ -1,14 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Expand(const tvm::Tensor& X, const tvm::Array& new_shape, const std::string& name = "expand"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/gather.cc b/onnxruntime/core/codegen/mti/tensor/gather.cc deleted file mode 100644 index 152b3981f1623..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/gather.cc +++ /dev/null @@ -1,55 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/gather.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Gather(const tvm::Tensor& t, - int64_t axis, - const tvm::Tensor& indices, - const std::string& name) { - // handle negative axis - axis = HandleNegativeAxis(axis, gsl::narrow(t->shape.size())); - size_t axis_t = gsl::narrow(axis); - - tvm::Array output_shape; - for (size_t i = 0; i < axis_t; ++i) - output_shape.push_back(t->shape[i]); - - for (size_t i = 0; i < indices->shape.size(); ++i) - output_shape.push_back(indices->shape[i]); - - for (size_t i = axis_t + 1; i < t->shape.size(); ++i) - output_shape.push_back(t->shape[i]); - - tvm::Expr idx_upper_bound = t->shape[axis_t]; - auto l = [&](const tvm::Array& ovars) { - tvm::Array ivars; - for (size_t i = 0; i < t->shape.size(); ++i) { - if (i < axis_t) { - ivars.push_back(ovars[i]); - } else if (i == axis_t) { - tvm::Array idx_vars; - for (size_t d = 0; d < indices->shape.size(); ++d) - idx_vars.push_back(ovars[axis_t + d]); - // make sure idx is clamped in the range of [-idx_upper_bound, idx_upper_bound - 1] - tvm::Expr real_idx = tvm_codegen::ClampIndex(indices(idx_vars), idx_upper_bound); - ivars.push_back(tvm::cast(tvm::Int(32), real_idx)); // tvm indices must be Int32 - } else { - ivars.push_back(ovars[i - 1 + indices->shape.size()]); - } - } - return t(ivars); - }; - - return tvm::compute(output_shape, l, name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/gather.h b/onnxruntime/core/codegen/mti/tensor/gather.h deleted file mode 100644 index a44bf3e4127d5..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/gather.h +++ /dev/null @@ -1,17 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Gather(const tvm::Tensor& t, - int64_t axis, - const tvm::Tensor& indices, - const std::string& name = "gather"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/gather_elements.cc b/onnxruntime/core/codegen/mti/tensor/gather_elements.cc deleted file mode 100644 index 12d2983335890..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/gather_elements.cc +++ /dev/null @@ -1,45 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/gather_elements.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor GatherElements(const tvm::Tensor& t, - int64_t axis, - const tvm::Tensor& indices, - const std::string& name) { - tvm::Array output_shape; - int64_t indices_rank = static_cast(indices->shape.size()); - // output shape is the same as indices - for (int64_t i = 0; i < indices_rank; ++i) - output_shape.push_back(indices->shape[i]); - - tvm::Expr idx_upper_bound = t->shape[axis]; - auto l = [&](const tvm::Array& ovars) { - tvm::Array ivars; - for (int64_t i = 0; i < indices_rank; i++) { - if (i == axis) { - tvm::Array idx_vars; - for (int64_t j = 0; j < indices_rank; j++) - idx_vars.push_back(ovars[j]); - // make sure idx is clamped in the range of [-idx_upper_bound, idx_upper_bound - 1] - tvm::Expr real_idx = tvm_codegen::ClampIndex(indices(idx_vars), idx_upper_bound); - // tvm idx must be of Int(32) - ivars.push_back(tvm::cast(tvm::Int(32), real_idx)); - } else { - ivars.push_back(ovars[i]); - } - } - return t(ivars); - }; - - return tvm::compute(output_shape, l, name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/gather_elements.h b/onnxruntime/core/codegen/mti/tensor/gather_elements.h deleted file mode 100644 index 650086f0f2e87..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/gather_elements.h +++ /dev/null @@ -1,17 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor GatherElements(const tvm::Tensor& t, - int64_t axis, - const tvm::Tensor& indices, - const std::string& name = "gather_elements"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/pad_ops.cc b/onnxruntime/core/codegen/mti/tensor/pad_ops.cc deleted file mode 100644 index 2f688290d109e..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/pad_ops.cc +++ /dev/null @@ -1,121 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/pad_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// Note topi::pad does not support modes {edge, reflect} -// Therefore, MTI implements a generic Pad -tvm::Tensor Pad(const tvm::Tensor& t, - const tvm::Array& pad_before, - const tvm::Array& pad_after, - float pad_value, - const std::string& mode, - const std::string& name) { - MTI_ASSERT(pad_before.size() >= 1); - MTI_ASSERT(pad_before.size() == pad_after.size()); - MTI_ASSERT(pad_before.size() == t->shape.size()); - - tvm::Array output_shape; - for (size_t i = 0; i < t->shape.size(); ++i) { - output_shape.push_back( - tvm::ir::Simplify(t->shape[i] + pad_before[i] + pad_after[i])); - } - - auto l = [&](const tvm::Array& ovars) { - tvm::Array conds; - tvm::Array indices; - tvm::Array coords; - - for (size_t i = 0; i < t->shape.size(); ++i) { - tvm::Expr ivar = ovars[i] - pad_before[i]; - tvm::Expr min = 0; - tvm::Expr extent = t->shape[i]; - - conds.push_back(ivar < min); - conds.push_back(ivar >= min + extent); - indices.push_back(tvm::max(tvm::min(ivar, min + extent - 1), min)); - - if (mode == "reflect") { - // calculate indices for reflect mode - tvm::Expr limit = extent - 1; - tvm::Expr coord = ivar - min; - // Avoid mod zero when tensor shape has 1, - // e.g. input shape is [1, 3, 3] instead of [3, 3] - auto* p_limit = tvm::as_const_int(limit); - if (p_limit != nullptr && *p_limit != 0) - coord = (coord + 2 * limit) % (2 * limit); // avoid negative value - coord = coord - limit; - coord = tvm::abs(coord); - coord = limit - coord; - coord = coord + min; - coords.push_back(coord); - } - } - - if (mode == "reflect") { - return tvm::ir::Select::make(topi::detail::Map(conds, tvm::ir::Or::make), - t(coords), t(indices)); - } else if (mode == "constant") { - return tvm::ir::Select::make(topi::detail::Map(conds, tvm::ir::Or::make), - tvm::make_const(t->dtype, pad_value), t(indices)); - } - - // default mode is edge - return t(indices); - }; - - return tvm::compute(output_shape, l, name); -} - -tvm::Tensor Pad(const tvm::Tensor& t, - const tvm::Array& output_shape, - const tvm::Expr& pad_value, - const std::string& name) { - MTI_ASSERT(t->dtype == pad_value.type()); - - auto l = [&](const tvm::Array& ovars) { - tvm::Array conds; - tvm::Array indices; - - for (size_t i = 0; i < t->shape.size(); ++i) { - tvm::Expr ivar = ovars[i]; - tvm::Expr min = 0; - tvm::Expr extent = t->shape[i]; - - conds.push_back(ivar < min); - conds.push_back(ivar >= min + extent); - indices.push_back(tvm::max(tvm::min(ivar, min + extent - 1), min)); - } - - return tvm::ir::Select::make(topi::detail::Map(conds, tvm::ir::Or::make), - pad_value, t(indices)); - }; - - return tvm::compute(output_shape, l, name); -} - -tvm::Tensor PadLastDim(const tvm::Tensor& t, - const int32_t align_size, - const tvm::Expr& pad_value, - const std::string& name) { - auto input_shape = t->shape; - tvm::Array out_shape; - size_t input_shape_rank = input_shape.size(); - for (size_t i = 0; i < input_shape_rank - 1; ++i) { - out_shape.push_back(input_shape[i]); - } - out_shape.push_back( - (input_shape[input_shape_rank - 1] + align_size - 1) / - align_size * align_size); - - return Pad(t, out_shape, pad_value, name + "_pad"); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/pad_ops.h b/onnxruntime/core/codegen/mti/tensor/pad_ops.h deleted file mode 100644 index 6e8e350d71e97..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/pad_ops.h +++ /dev/null @@ -1,34 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// ONNX Pad semantics -tvm::Tensor Pad(const tvm::Tensor& t, - const tvm::Array& pad_before, - const tvm::Array& pad_after, - float pad_value = 0.0f, - const std::string& mode = "constant", - const std::string& name = "pad"); - -// Other common Pad interfaces -// Pad for a given shape -tvm::Tensor Pad(const tvm::Tensor& t, - const tvm::Array& output_shape, - const tvm::Expr& pad_value, - const std::string& name = "pad"); - -// Pad for the last dim only. -// This is widely used for weight layout to guard alignment -tvm::Tensor PadLastDim(const tvm::Tensor& t, - const int32_t align_size, - const tvm::Expr& pad_value, - const std::string& name = "pad_last_dim"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/reshape_ops.cc b/onnxruntime/core/codegen/mti/tensor/reshape_ops.cc deleted file mode 100644 index 817fb32c2837a..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/reshape_ops.cc +++ /dev/null @@ -1,48 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/reshape_ops.h" - -#include "core/codegen/mti/common.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Flatten(const tvm::Tensor& X, int64_t axis, const std::string& name) { - const auto& input_shape = X->shape; - return Reshape(X, {SizeToDimension(input_shape, axis), SizeFromDimension(input_shape, axis)}, name); -} - -tvm::Tensor Identity(const tvm::Tensor& X, const std::string& name) { - return Reshape(X, X->shape, name); -} - -tvm::Tensor Reshape(const tvm::Tensor& X, const tvm::Array& new_shape, const std::string& name) { - if (new_shape.size() > 0) { - auto X_dim = SizeToDimension(X->shape, X->shape.size()); - auto new_dim = SizeToDimension(new_shape, new_shape.size()); - auto* pX_dim = tvm::as_const_int(X_dim); - auto* pNew_dim = tvm::as_const_int(new_dim); - - if (pX_dim != nullptr && pNew_dim != nullptr) { - MTI_ASSERT(*pX_dim == *pNew_dim); - } - return topi::reshape(X, new_shape, name); - } else { - // generate empty dim tensor with origial input data value - tvm::Array tmp_shape; - tmp_shape.push_back(1); - auto tmp_tensor = topi::reshape(X, tmp_shape); - return tvm::compute( - new_shape, - [&](const tvm::Array&) { - return tmp_tensor[0]; - }, - name); - } -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/reshape_ops.h b/onnxruntime/core/codegen/mti/tensor/reshape_ops.h deleted file mode 100644 index e23d62e4c57b0..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/reshape_ops.h +++ /dev/null @@ -1,16 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Flatten(const tvm::Tensor& X, int64_t axis, const std::string& name = "flatten"); -tvm::Tensor Identity(const tvm::Tensor& X, const std::string& name = "identity"); -tvm::Tensor Reshape(const tvm::Tensor& X, const tvm::Array& new_shape, const std::string& name = "reshape"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/shape_op.cc b/onnxruntime/core/codegen/mti/tensor/shape_op.cc deleted file mode 100644 index b51bd67a8b2dc..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/shape_op.cc +++ /dev/null @@ -1,25 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/shape_op.h" - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Shape(const tvm::Tensor& X, const std::string& name) { - int ndim = static_cast(X->shape.size()); - tvm::Array out_shape{ndim}; - return tvm::compute( - out_shape, [&](const tvm::Array& indices) { - auto idx = indices[0]; - tvm::Expr ret = 0; - for (int i = 0; i < ndim; ++i) { - ret = tvm::ir::Select::make(idx == i, X->shape[i], ret); - } - return tvm::cast(HalideIR::Int(64), ret); - }, - name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/shape_op.h b/onnxruntime/core/codegen/mti/tensor/shape_op.h deleted file mode 100644 index 67ee2de50eca9..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/shape_op.h +++ /dev/null @@ -1,14 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Shape(const tvm::Tensor& X, const std::string& name = "shape"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/slice.cc b/onnxruntime/core/codegen/mti/tensor/slice.cc deleted file mode 100644 index 6cbab43584d4b..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/slice.cc +++ /dev/null @@ -1,91 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/slice.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include -#include -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// local constexpr for INT_MAX -constexpr int64_t max_range = INT_MAX; - -tvm::Expr position(const tvm::Expr& dim, const tvm::Integer& offset, bool allow_out_of_bound = false) { - if (offset->value >= max_range) { - return allow_out_of_bound ? dim : dim - 1; - } else if (offset->value <= -max_range) { - return tvm::make_const(HalideIR::Int(32), allow_out_of_bound ? -1 : 0); - } else { - if (offset->value >= 0) { - return tvm::ir::Simplify(tvm::ir::Min::make(offset, dim + (allow_out_of_bound ? 0 : -1))); - } else { - return tvm::ir::Simplify(dim + tvm::ir::Max::make(offset, -dim + (allow_out_of_bound ? -1 : 0))); - } - } -} - -tvm::Tensor Slice(const tvm::Tensor& X, - const std::vector& starts, - const std::vector& ends, - const std::vector& axes1, - const std::vector& steps, - const std::string& name) { - MTI_ASSERT(starts.size() == ends.size()); - MTI_ASSERT(starts.size() == axes1.size()); - MTI_ASSERT(starts.size() == steps.size()); - - std::vector axes; - for (const auto& i : axes1) { - axes.push_back(HandleNegativeAxis(i, X->shape.size())); - } - - tvm::Array output_shape; - bool empty = false; - for (int64_t i = 0; i < gsl::narrow(X->shape.size()); ++i) { - auto axes_iter = std::find(axes.begin(), axes.end(), i); - if (axes_iter != axes.end()) { - auto axis = axes_iter - axes.begin(); - tvm::Expr start = position(X->shape[i], starts[axis]); - tvm::Expr end = position(X->shape[i], ends[axis], /*allow_out_of_bound*/ true); - auto dim = tvm::ir::Simplify((end - start + tvm::Integer(steps[axis] + (steps[axis] < 0 ? 1 : -1))) / tvm::Integer(steps[axis])); - auto int_dim = tvm::as_const_int(dim); - if (int_dim && *int_dim <= 0) { - output_shape.push_back(0); - empty = true; - } else { - output_shape.push_back(dim); - } - } else { - output_shape.push_back(X->shape[i]); - } - } - - if (empty) { - return MakeZeroTensor(output_shape, X->dtype, name); - } - - return tvm::compute( - output_shape, - [&](const tvm::Array& ovars) { - tvm::Array ivars; - for (size_t i = 0; i < X->shape.size(); ++i) { - auto axes_iter = std::find(axes.begin(), axes.end(), i); - if (axes_iter != axes.end()) { - auto axis = axes_iter - axes.begin(); - ivars.push_back(tvm::ir::Simplify(ovars[i] * tvm::Integer(steps[axis]) + position(X->shape[i], starts[axis]))); - } else { - ivars.push_back(ovars[i]); - } - } - return X(ivars); - }, - name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/slice.h b/onnxruntime/core/codegen/mti/tensor/slice.h deleted file mode 100644 index ac5c9437791f6..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/slice.h +++ /dev/null @@ -1,19 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Slice(const tvm::Tensor& X, - const std::vector& starts, - const std::vector& ends, - const std::vector& axes, - const std::vector& steps, - const std::string& name = "slice"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/split.cc b/onnxruntime/core/codegen/mti/tensor/split.cc deleted file mode 100644 index 6ee366314858f..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/split.cc +++ /dev/null @@ -1,72 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/split.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// Similar to numpy, topi::split takes split indices rather than the -// sizes of the splits. Thus we implement our own. -tvm::Array Split(const tvm::Tensor& X, - const tvm::Array& split_sizes, - int64_t axis, - const std::string& name) { - MTI_ASSERT(axis < gsl::narrow(X->shape.size())); - size_t axis_t = gsl::narrow(axis); - - tvm::Array> output_shapes; - int num_splits = gsl::narrow(split_sizes.size()); - for (auto& s : split_sizes) { - tvm::Array shape; - for (size_t i = 0; i < axis_t; i++) { - shape.push_back(X->shape[i]); - } - shape.push_back(s); - for (size_t i = axis_t + 1; i < X->shape.size(); i++) { - shape.push_back(X->shape[i]); - } - output_shapes.push_back(shape); - } - - tvm::Array res; - int idx = 0; - for (int i_split = 0; i_split < num_splits; ++i_split) { - tvm::Expr s = split_sizes[i_split]; - auto l = [&](const tvm::Array& indices) { - tvm::Array new_indices; - for (size_t i = 0; i < axis_t; i++) { - new_indices.push_back(indices[i]); - } - new_indices.push_back(indices[axis_t] + idx); - for (size_t i = axis_t + 1; i < X->shape.size(); i++) { - new_indices.push_back(indices[i]); - } - MTI_ASSERT(topi::detail::IsConstInt(s)); - MTI_ASSERT(new_indices.size() == X->shape.size()); - int size = topi::detail::GetConstInt(s); - idx += size; - return X(new_indices); - }; - res.push_back(tvm::compute(output_shapes[i_split], l, name)); - } - - MTI_ASSERT(topi::detail::IsConstInt(X->shape[axis_t])); - int size_of_splitted_axis = static_cast(topi::detail::GetConstInt(X->shape[axis_t])); - MTI_ASSERT(idx == size_of_splitted_axis); - return res; -} - -tvm::Array SplitWithIndices(const tvm::Tensor& X, - const tvm::Array& split_sizes, - int64_t axis, - const std::string& name) { - return topi::split(X, split_sizes, gsl::narrow(axis), name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/split.h b/onnxruntime/core/codegen/mti/tensor/split.h deleted file mode 100644 index bcb9c47d936dd..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/split.h +++ /dev/null @@ -1,25 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// ONNX Split semantics -tvm::Array Split(const tvm::Tensor& X, - const tvm::Array& split_sizes, - int64_t axis, - const std::string& name = "split"); - -// Another common Split interface -// Split with chunck indices -tvm::Array SplitWithIndices(const tvm::Tensor& X, - const tvm::Array& split_sizes, - int64_t axis, - const std::string& name = "split_with_indices"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/tile.cc b/onnxruntime/core/codegen/mti/tensor/tile.cc deleted file mode 100644 index 2fef86adcbaea..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/tile.cc +++ /dev/null @@ -1,40 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/tile.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Tile(const tvm::Tensor& t, - const std::vector& repeats, - const std::string& name) { - MTI_ASSERT(repeats.size() == t->shape.size()); - tvm::Array output_shape; - - bool repeats_zero = false; - for (size_t i = 0; i < t->shape.size(); ++i) { - if (repeats[i] == 0) - repeats_zero = true; - output_shape.push_back(t->shape[i] * gsl::narrow(repeats[i])); - } - - auto l = [&](const tvm::Array& ovars) { - if (repeats_zero) - return tvm::make_zero(t->dtype); - - tvm::Array ivars; - for (size_t i = 0; i < t->shape.size(); ++i) { - tvm::Expr ovar = ovars[i]; - ivars.push_back(ovar % t->shape[i]); - } - return t(ivars); - }; - - return tvm::compute(output_shape, l, name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/tile.h b/onnxruntime/core/codegen/mti/tensor/tile.h deleted file mode 100644 index 7ce331fb5ea95..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/tile.h +++ /dev/null @@ -1,16 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Tile(const tvm::Tensor& t, - const std::vector& repeats, - const std::string& name = "tile"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/transpose.cc b/onnxruntime/core/codegen/mti/tensor/transpose.cc deleted file mode 100644 index 873ff8d7f1708..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/transpose.cc +++ /dev/null @@ -1,16 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/tensor/transpose.h" - -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Transpose(const tvm::Tensor& X, const tvm::Array& axes, const std::string& name) { - return topi::transpose(X, axes, name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/transpose.h b/onnxruntime/core/codegen/mti/tensor/transpose.h deleted file mode 100644 index a2a98fedf1e79..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/transpose.h +++ /dev/null @@ -1,16 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Transpose(const tvm::Tensor& X, - const tvm::Array& axes, - const std::string& name = "transpose"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/where.cc b/onnxruntime/core/codegen/mti/tensor/where.cc deleted file mode 100644 index 2bdac3cae7ef5..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/where.cc +++ /dev/null @@ -1,36 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/where.h" - -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Where(const tvm::Tensor& B, - const tvm::Tensor& X, - const tvm::Tensor& Y, - const std::string& name) { - size_t rank = std::max(std::max(B->shape.size(), X->shape.size()), Y->shape.size()); - tvm::Array output_shape; - for (size_t i = 0; i < rank; ++i) { - tvm::Expr dim = tvm::make_const(HalideIR::Int(32), 1); - bool broadcasted = - BroadcastDim(B->shape, i, rank, dim) && - BroadcastDim(X->shape, i, rank, dim) && - BroadcastDim(Y->shape, i, rank, dim); - MTI_ASSERT(broadcasted); - output_shape.push_back(dim); - } - - return topi::where(topi::broadcast_to(B, output_shape), - topi::broadcast_to(X, output_shape), - topi::broadcast_to(Y, output_shape), - name); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/mti/tensor/where.h b/onnxruntime/core/codegen/mti/tensor/where.h deleted file mode 100644 index 68c5288eb3580..0000000000000 --- a/onnxruntime/core/codegen/mti/tensor/where.h +++ /dev/null @@ -1,17 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Where(const tvm::Tensor& B, - const tvm::Tensor& X, - const tvm::Tensor& Y, - const std::string& name = "where"); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/all_ops.h b/onnxruntime/core/codegen/passes/op_ir_creator/all_ops.h deleted file mode 100644 index 1463e50bd72fb..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/all_ops.h +++ /dev/null @@ -1,47 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/passes/utils/codegen_context.h" -#include "core/codegen/common/op_macro.h" -#include "core/codegen/passes/op_ir_creator/tvm_op_creator.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// This macro declares a TVM IR builder -// based on ORT OP type with postfix DefaultTVM -#define DECLARE_GENERIC_OP_IR_CREATOR_CLASS(OP) \ - DECLARE_OP_IR_CREATOR_CLASS(OP, DefaultTVM) - -// This macro returns a TVM IR builder class name -// based ORT OP type with postfix DefaultTVM -#define GENERIC_OP_IR_CREATOR_CLASS(OP) \ - CREATOR_CLASS(OP, DefaultTVM##IRCreator) - -#define GENERIC_OP_IR_CREATOR_STRING(OP) \ - STRINGIZE(GENERIC_OP_IR_CREATOR_CLASS(OP)) - -// define all ops for DefaultTVM -#define ADD_OP_ITEM(OP) DECLARE_GENERIC_OP_IR_CREATOR_CLASS(OP) -#define BINARY_OP(OP) ADD_OP_ITEM(OP) -#define BINARY_CMP_OP(OP) ADD_OP_ITEM(OP) -#define POOL_OP(OP) ADD_OP_ITEM(OP) -#define UNARY_OP(OP) ADD_OP_ITEM(OP) -#define VARIADIC_OP(OP) ADD_OP_ITEM(OP) -#define REDUCE_INDEXED_OP(OP) ADD_OP_ITEM(OP) -#define REDUCE_OP(OP) ADD_OP_ITEM(OP) - -LIST_ALL_GENERIC_OPS() - -#undef ADD_OP_ITEM -#undef BINARY_OP -#undef BINARY_CMP_OP -#undef POOL_OP -#undef REDUCE_OP -#undef REDUCE_INDEXED_OP -#undef UNARY_OP -#undef VARIADIC_OP - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/binary_ops.cc b/onnxruntime/core/codegen/passes/op_ir_creator/math/binary_ops.cc deleted file mode 100644 index 9452146621ac7..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/binary_ops.cc +++ /dev/null @@ -1,46 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/common/op_macro.h" -#include "core/codegen/mti/math/binary_ops.h" -#include "core/codegen/mti/tensor/cast_ops.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// helper local macro defines Evaluate of BINARY_OP OpIRCreators -#define BINARY_OP(name) \ - Status GENERIC_OP_IR_CREATOR_CLASS(name)::Evaluate( \ - const tvm::Array& inputs, \ - const Node& node, \ - CodeGenContext&, \ - tvm::Array& outputs) { \ - tvm::Tensor Y = name(inputs[0], inputs[1], node.Name()); \ - outputs.push_back(Y); \ - return Status::OK(); \ - } - -LIST_BINARY_OPS() - -#undef BINARY_OP - -// helper local macro defines Evaluate of BINARY_CMP_OP OpIRCreators -#define BINARY_CMP_OP(name) \ - Status GENERIC_OP_IR_CREATOR_CLASS(name)::Evaluate( \ - const tvm::Array& inputs, \ - const Node& node, \ - CodeGenContext&, \ - tvm::Array& outputs) { \ - tvm::Tensor Y = Cast(name(inputs[0], inputs[1], node.Name()), HalideIR::UInt(8), "cast_bool_" #name); \ - outputs.push_back(Y); \ - return Status::OK(); \ - } - -LIST_BINARY_CMP_OPS() - -#undef BINARY_CMP_OP - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/clip.cc b/onnxruntime/core/codegen/passes/op_ir_creator/math/clip.cc deleted file mode 100644 index bb33e6e70accf..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/clip.cc +++ /dev/null @@ -1,48 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/math/unary_ops.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Clip OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Clip)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper info(&ctx); - - int version = ctx_codegen.GetCodeGenHandle()->domain_version_lookup_func(node.Domain()); - tvm::Expr min_value, max_value; - if (version < 11) { - float max_v, min_v; - info.GetAttrOrDefault("min", &min_v, std::numeric_limits::lowest()); - info.GetAttrOrDefault("max", &max_v, std::numeric_limits::max()); - min_value = tvm::make_const(tvm::Float(32), min_v); - max_value = tvm::make_const(tvm::Float(32), max_v); - } else { - // for op_version >= 11, max and min are optional inputs - min_value = tvm::make_const(tvm::Float(32), std::numeric_limits::lowest()); - max_value = tvm::make_const(tvm::Float(32), std::numeric_limits::max()); - auto num_inputs = inputs.size(); - if (num_inputs >= 2 && inputs[1].defined()) { - min_value = inputs[1](); - } - if (num_inputs == 3 && inputs[2].defined()) { - max_value = inputs[2](); - } - } - - tvm::Tensor Y = Clip(inputs[0], min_value, max_value, node.Name() + "_Clip"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/gemm.cc b/onnxruntime/core/codegen/passes/op_ir_creator/math/gemm.cc deleted file mode 100644 index 64f995076e1bb..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/gemm.cc +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/math/gemm.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Gemm OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Gemm)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& /*ctx_codegen*/, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - - tvm::Tensor A = inputs[0]; - tvm::Tensor B = inputs[1]; - tvm::Tensor C = inputs[2]; - - int64_t trans_A, trans_B; - ORT_RETURN_IF_ERROR(attrs.GetAttr("transA", &trans_A)); - ORT_RETURN_IF_ERROR(attrs.GetAttr("transB", &trans_B)); - - float alpha, beta; - ORT_ENFORCE(attrs.GetAttr("alpha", &alpha).IsOK()); - ORT_ENFORCE(attrs.GetAttr("beta", &beta).IsOK()); - - tvm::Tensor Y = Gemm(A, B, C, trans_A != 0, trans_B != 0, alpha, beta, node.Name() + "_Gemm"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/logsoftmax.cc b/onnxruntime/core/codegen/passes/op_ir_creator/math/logsoftmax.cc deleted file mode 100644 index cb09518bf63d1..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/logsoftmax.cc +++ /dev/null @@ -1,32 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/math/logsoftmax.h" -#include "core/framework/op_kernel_info.h" -#include "core/providers/common.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of LogSoftmax OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(LogSoftmax)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper info(&ctx); - - int64_t axis_i64; - ORT_RETURN_IF_ERROR(info.GetAttr("axis", &axis_i64)); - axis_i64 = HandleNegativeAxis(axis_i64, gsl::narrow_cast(inputs[0]->shape.size())); - - tvm::Tensor Y = LogSoftmax(inputs[0], axis_i64, node.Name() + "_LogSoftmax"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/matmul.cc b/onnxruntime/core/codegen/passes/op_ir_creator/math/matmul.cc deleted file mode 100644 index ab1ac237bfa5d..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/matmul.cc +++ /dev/null @@ -1,23 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/math/matmul_ops.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of MatMul OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(MatMul)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - tvm::Tensor Y = MatMul(inputs[0], inputs[1], node.Name() + "_MatMul"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/quantize/matmul_integer.cc b/onnxruntime/core/codegen/passes/op_ir_creator/math/quantize/matmul_integer.cc deleted file mode 100644 index 6f66b1f1a2afb..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/quantize/matmul_integer.cc +++ /dev/null @@ -1,37 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/math/binary_ops.h" -#include "core/codegen/mti/math/matmul_ops.h" -#include "core/codegen/mti/tensor/cast_ops.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of MatMulInteger OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(MatMulInteger)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - const auto& A = inputs[0]; - const auto& B = inputs[1]; - auto& name = node.Name(); - - // A generic path, cast to int32 - // Support skipped trailing inputs - auto A_Int32 = (node.InputDefs().size() >= 3 && node.InputDefs()[2]->Exists()) - ? Sub(Cast(A, HalideIR::Int(32)), Cast(inputs[2], HalideIR::Int(32))) - : Cast(A, HalideIR::Int(32)); - auto B_Int32 = (node.InputDefs().size() >= 4 && node.InputDefs()[3]->Exists()) - ? Sub(Cast(B, HalideIR::Int(32)), Cast(inputs[3], HalideIR::Int(32))) - : Cast(B, HalideIR::Int(32)); - tvm::Tensor Y = MatMul(A_Int32, B_Int32, name + "_MatMulInteger"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/reduce_ops.cc b/onnxruntime/core/codegen/passes/op_ir_creator/math/reduce_ops.cc deleted file mode 100644 index f29a3f3e7cdf7..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/reduce_ops.cc +++ /dev/null @@ -1,111 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/common/op_macro.h" -#include "core/codegen/mti/math/reduce_ops.h" -#include "core/codegen/mti/tensor/cast_ops.h" -#include "core/codegen/mti/tensor/reshape_ops.h" -#include "core/framework/op_kernel_info.h" -#include "core/providers/common.h" - -namespace onnxruntime { -namespace tvm_codegen { - -using ReduceIndexedFunc = tvm::Tensor (*)(const tvm::Tensor& X, int64_t axis, bool keep_dims, const std::string& name); -using ReduceFunc = tvm::Tensor (*)(const tvm::Tensor& X, const std::vector& axes, bool keep_dims, const std::string& name); - -// helper class for for REDUCE_INDEXED_OP -class FuncReduceIndexed { - public: - FuncReduceIndexed(const Node& node, ReduceIndexedFunc func, const std::string& name) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper info(&ctx); - axis_ = info.GetAttrOrDefault("axis", 0); - int64_t keepdims_i = 1; - ORT_ENFORCE(info.GetAttr("keepdims", &keepdims_i).IsOK()); - keep_dims_ = (keepdims_i == 1); - func_ = func; - name_ = name; - } - - tvm::Tensor operator()(const tvm::Tensor& X) const { - auto axis = HandleNegativeAxis(axis_, gsl::narrow_cast(X->shape.size())); - tvm::Tensor index32 = func_(X, axis, keep_dims_, name_); - return Cast(index32, tvm::Int(64)); - } - - private: - int64_t axis_; - bool keep_dims_; - ReduceIndexedFunc func_; - std::string name_; -}; - -// helper class for REDUCE_OP -class FuncReduce { - public: - FuncReduce(const Node& node, ReduceFunc func, const std::string& name) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper info(&ctx); - axes_ = info.GetAttrsOrDefault("axes"); - int64_t keepdims_i = 1; - ORT_ENFORCE(info.GetAttr("keepdims", &keepdims_i).IsOK()); - keep_dims_ = (keepdims_i == 1); - func_ = func; - name_ = name; - } - - tvm::Tensor operator()(const tvm::Tensor& X) const { - std::vector axes; - for (auto i : axes_) - axes.push_back(HandleNegativeAxis(i, gsl::narrow_cast(X->shape.size()))); - - return func_(X, axes, keep_dims_, name_); - } - - private: - std::vector axes_; - bool keep_dims_; - ReduceFunc func_; - std::string name_; -}; - -// helper macro defines Evaluate of REDUCE_OP OpIRCreators -#define REDUCE_OP(name) \ - Status GENERIC_OP_IR_CREATOR_CLASS(name)::Evaluate( \ - const tvm::Array& inputs, \ - const Node& node, \ - CodeGenContext&, \ - tvm::Array& outputs) { \ - tvm::Tensor Y; \ - if (ShapeRank(node.OutputDefs()[0]) == 0) { \ - tvm::Tensor temp = FuncReduce(node, &name, #name)(inputs[0]); \ - Y = Reshape(temp, {}); \ - } else { \ - Y = FuncReduce(node, &name, #name)(inputs[0]); \ - } \ - outputs.push_back(Y); \ - return Status::OK(); \ - } - -// helper macro defines Evaluate of REDUCE_INDEXED_OP OpIRCreators -#define REDUCE_INDEXED_OP(name) \ - Status GENERIC_OP_IR_CREATOR_CLASS(name)::Evaluate( \ - const tvm::Array& inputs, \ - const Node& node, \ - CodeGenContext&, \ - tvm::Array& outputs) { \ - tvm::Tensor Y = FuncReduceIndexed(node, &name, #name)(inputs[0]); \ - outputs.push_back(Y); \ - return Status::OK(); \ - } - -LIST_REDUCE_OPS() - -#undef REDUCE_OP -#undef REDUCE_INDEXED_OP - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/softmax.cc b/onnxruntime/core/codegen/passes/op_ir_creator/math/softmax.cc deleted file mode 100644 index 7b13de5a94e48..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/softmax.cc +++ /dev/null @@ -1,32 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/math/softmax.h" -#include "core/framework/op_kernel_info.h" -#include "core/providers/common.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Softmax OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Softmax)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper info(&ctx); - - int64_t axis_i64; - ORT_RETURN_IF_ERROR(info.GetAttr("axis", &axis_i64)); - - axis_i64 = HandleNegativeAxis(axis_i64, gsl::narrow_cast(inputs[0]->shape.size())); - tvm::Tensor Y = Softmax(inputs[0], axis_i64, node.Name() + "_Softmax"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/unary_funcs.h b/onnxruntime/core/codegen/passes/op_ir_creator/math/unary_funcs.h deleted file mode 100644 index 29e6519af0ef1..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/unary_funcs.h +++ /dev/null @@ -1,51 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { -// helper class for unary_ops with alpha -class FuncWithAlpha { - public: - FuncWithAlpha(const Node& node) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - ORT_ENFORCE(attrs.GetAttr("alpha", &alpha_).IsOK()); - } - - protected: - float alpha_; -}; - -// helper class for unary_ops with alpha and beta -class FuncWithAlphaBeta { - public: - FuncWithAlphaBeta(const Node& node) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - ORT_ENFORCE(attrs.GetAttr("alpha", &alpha_).IsOK()); - ORT_ENFORCE(attrs.GetAttr("beta", &beta_).IsOK()); - } - - protected: - float alpha_; - float beta_; -}; - -// helper class for unary_ops with alpha and gamma -class FuncWithAlphaGamma { - public: - FuncWithAlphaGamma(const Node& node) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - ORT_ENFORCE(attrs.GetAttr("alpha", &alpha_).IsOK()); - ORT_ENFORCE(attrs.GetAttr("gamma", &gamma_).IsOK()); - } - - protected: - float alpha_; - float gamma_; -}; -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/unary_ops.cc b/onnxruntime/core/codegen/passes/op_ir_creator/math/unary_ops.cc deleted file mode 100644 index 0407c0a06abf6..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/unary_ops.cc +++ /dev/null @@ -1,93 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/common/op_macro.h" -#include "core/codegen/mti/math/unary_ops.h" -#include "core/codegen/passes/op_ir_creator/math/unary_funcs.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// helper macro declares unary_ops helper class without attribute -#define FuncClass(name) \ - class Func##name { \ - public: \ - Func##name(const Node&) {} \ - tvm::Tensor operator()(const tvm::Tensor& X) const { \ - return name(X); \ - } \ - } - -// helper macro declares unary_ops helper class with alpha -#define FuncClassAlpha(name) \ - class Func##name : public FuncWithAlpha { \ - public: \ - Func##name(const Node& node) : FuncWithAlpha(node) {} \ - tvm::Tensor operator()(const tvm::Tensor& X) const { \ - return name(X, alpha_); \ - } \ - } - -// helper macro declares unary_ops helper class with alpha and beta -#define FuncClassAlphaBeta(name) \ - class Func##name : public FuncWithAlphaBeta { \ - public: \ - Func##name(const Node& node) : FuncWithAlphaBeta(node) {} \ - tvm::Tensor operator()(const tvm::Tensor& X) const { \ - return name(X, alpha_, beta_); \ - } \ - } - -// helper macro declares unary_ops helper class with alpha and gamma -#define FuncClassAlphaGamma(name) \ - class Func##name : public FuncWithAlphaGamma { \ - public: \ - Func##name(const Node& node) : FuncWithAlphaGamma(node) {} \ - tvm::Tensor operator()(const tvm::Tensor& X) const { \ - return name(X, alpha_, gamma_); \ - } \ - } - -FuncClass(Abs); -FuncClassAlphaBeta(Affine); -FuncClass(Ceil); -FuncClassAlpha(Elu); -FuncClass(Exp); -FuncClass(Floor); -FuncClassAlphaBeta(HardSigmoid); -FuncClassAlpha(LeakyRelu); -FuncClass(Log); -FuncClass(Neg); -FuncClassAlphaBeta(ParametricSoftplus); -FuncClass(Reciprocal); -FuncClass(Relu); -FuncClassAlphaBeta(ScaledTanh); -FuncClassAlphaGamma(Selu); -FuncClass(Sigmoid); -FuncClass(Softplus); -FuncClass(Softsign); -FuncClass(Sqrt); -FuncClass(Tanh); -FuncClassAlpha(ThresholdedRelu); - -// helper macro defines Evaluate of UNARY_OP OpIRCreators -#define UNARY_OP(name) \ - Status GENERIC_OP_IR_CREATOR_CLASS(name)::Evaluate( \ - const tvm::Array& inputs, \ - const Node& node, \ - CodeGenContext&, \ - tvm::Array& outputs) { \ - tvm::Tensor Y = Func##name(node)(inputs[0]); \ - outputs.push_back(Y); \ - return Status::OK(); \ - } - -// helper local macros to replace some calls in LIST_UNARY_OPS -LIST_UNARY_OPS() - -#undef UNARY_OP - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/math/variadic_ops.cc b/onnxruntime/core/codegen/passes/op_ir_creator/math/variadic_ops.cc deleted file mode 100644 index 9559a713c2876..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/math/variadic_ops.cc +++ /dev/null @@ -1,36 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/math/binary_ops.h" -#include "core/codegen/mti/tensor/reshape_ops.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -tvm::Tensor Sum(const tvm::Tensor& lhs, const tvm::Tensor& rhs, const std::string& name) { - return Add(lhs, rhs, name); -} - -// helper local macro defines Evaluate of BINARY_OP OpIRCreators -#define VARIADIC_OP(name) \ - Status GENERIC_OP_IR_CREATOR_CLASS(name)::Evaluate( \ - const tvm::Array& inputs, \ - const Node& node, \ - CodeGenContext&, \ - tvm::Array& outputs) { \ - tvm::Tensor Y = Identity(inputs[0], node.Name() + "0"); \ - for (size_t i = 1; i < inputs.size(); ++i) \ - Y = name(Y, inputs[i], node.Name() + std::to_string(i)); \ - outputs.push_back(Y); \ - return Status::OK(); \ - } - -LIST_VARIADIC_OPS() - -#undef VARIADIC_OP - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/nn/conv.cc b/onnxruntime/core/codegen/passes/op_ir_creator/nn/conv.cc deleted file mode 100644 index 19545d1554405..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/nn/conv.cc +++ /dev/null @@ -1,131 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/nn/conv_ops.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/concat_ops.h" -#include "core/codegen/mti/tensor/split.h" -#include "core/codegen/passes/utils/ort_tvm_utils.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -Status GENERIC_OP_IR_CREATOR_CLASS(Conv)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper info(&ctx); - - // Attributes - int64_t group; - std::string auto_pad; - std::vector kernel_shape, strides, dilations, pads; - - info.GetAttrOrDefault("group", &group, 1); - info.GetAttrOrDefault("auto_pad", &auto_pad, "NOTSET"); - - ORT_THROW_IF_ERROR(info.GetAttrs("kernel_shape", kernel_shape)); - ORT_ENFORCE(kernel_shape.size() <= 2, "Only support 1D/2D convolution currently!"); - ORT_THROW_IF_ERROR(info.GetAttrs("strides", strides)); - - dilations = info.GetAttrs("dilations", dilations).IsOK() ? dilations : std::vector(kernel_shape.size(), 1); - ORT_ENFORCE(dilations == std::vector(kernel_shape.size(), 1), "Only support dilation is 1 currently"); - - pads = info.GetAttrs("pads", pads).IsOK() ? pads : std::vector(kernel_shape.size() * 2, 0); - - // auto_pad - if (auto_pad != "NOTSET") { - auto rank = inputs[0]->shape.size() - 2; - ORT_ENFORCE(rank > 0); - for (uint64_t i = 0; i < rank; i++) { - if (auto_pad == "VALID") { - pads[i] = 0; - pads[i + rank] = 0; - } else if (auto_pad == "SAME_UPPER" || auto_pad == "SAME_LOWER") { - // TODO: handle symbolic dim - ORT_ENFORCE(ShapeHasValue(node.InputDefs()[0], 2 + i)); - - int64_t input_dim_value = ShapeValue(node.InputDefs()[0], 2 + i); - int64_t output_dim_value = (input_dim_value + strides[i] - 1) / strides[i]; - int64_t pad_needed = (output_dim_value - 1) * strides[i] + kernel_shape[i] - input_dim_value; - - pads[i] = auto_pad == "SAME_LOWER" ? (pad_needed + 1) / 2 : pad_needed / 2; - pads[i + rank] = pad_needed - pads[i]; - } else { - ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Unknown auto_pad value"); - } - } - } - - // Inputs - tvm::Tensor X = inputs[0]; - tvm::Tensor W = inputs[1]; - // Outputs - tvm::Tensor Y; - tvm::Array Y_shape = ShapeToTvmArray(node.OutputDefs()[0], ctx_codegen); - - // 1-D convolution - if (kernel_shape.size() == 1) { - Y = Conv1D(X, W, Y_shape, ToTvmArray(strides), ToTvmArray(pads), node.Name() + "_Conv1D"); - } - // 2-D convolution - else if (kernel_shape.size() == 2) { - if (group == 1) { - Y = Conv2D(X, W, Y_shape, ToTvmArray(strides), ToTvmArray(pads), node.Name() + "_Conv2D"); - } else { - int64_t channel_out = ShapeValue(node.InputDefs()[1], 0); - int64_t channel_in = ShapeValue(node.InputDefs()[1], 1); - ORT_ENFORCE(channel_out % group == 0); - - int64_t cout_group = channel_out / group; - Y_shape.Set(1, Y_shape[1] / gsl::narrow_cast(group)); - - tvm::Array split_index0; - tvm::Array split_index1; - - for (int i = 1; i < group; i++) { - split_index0.push_back(i * channel_in); - split_index1.push_back(i * cout_group); - } - - auto input_groups = SplitWithIndices(X, split_index0, 1); - auto weight_groups = SplitWithIndices(W, split_index1, 0); - - // FIXME: This will trigger a llvm buffer overflow when group is too large - // TODO: fix this change it to batched gemm/conv - tvm::Array output_tensors; - for (int i = 0; i < group; i++) { - auto output_tensor = Conv2D(input_groups[i], - weight_groups[i], - Y_shape, - ToTvmArray(strides), - ToTvmArray(pads), - node.Name() + "_Conv2D"); - output_tensors.push_back(output_tensor); - } - Y = Concat(output_tensors, 1); - } - } - - // Add bias if provided - // Support skipped trailing inputs - if (node.InputDefs().size() > 2 && node.InputDefs()[2]->Exists()) { - tvm::Tensor B = inputs[2]; - Y = tvm::compute( - Y_shape, - [&](const tvm::Array& indices) { - return Y(indices) + B(indices[1]); - }); - } - - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/nn/lstm.cc b/onnxruntime/core/codegen/passes/op_ir_creator/nn/lstm.cc deleted file mode 100644 index 88170bb56dd2d..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/nn/lstm.cc +++ /dev/null @@ -1,64 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/nn/lstm.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// In the cell computation, we don't have the "direction" dimension and sequence dimension, -// which have been processed outside of the cell. -// Here we implement an LTSM cell. -// For those args (inputs/outputs) of hidden states we put AFTER regular args (inputs/outputs) -// with a pre-defined order -// In a LSTM, the order is H and then C. -// Ouputs of LSTM is Y_h and then Y_c -Status GENERIC_OP_IR_CREATOR_CLASS(LSTM)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - - std::string direction_attr; - ORT_RETURN_IF_ERROR(attrs.GetAttr("direction", &direction_attr)); - int64_t hidden_size; - ORT_RETURN_IF_ERROR(attrs.GetAttr("hidden_size", &hidden_size)); - - // input tensor with shape [seq_length, batch_size, input_size] - const tvm::Tensor& X = inputs[0]; // input tensor with shape [seq_length, batch_size, input_size] - const tvm::Tensor& W = inputs[1]; // weights tensor with shape [4*hidden_size, input_size] - const tvm::Tensor& R = inputs[2]; // recurrence tensor with shape [4*hidden_size, hidden_size] - const tvm::Tensor& B = inputs[3]; // optional bias tensor with shape [8*hidden_size] - bool has_B = node.InputDefs()[3]->Exists(); - - // Unsupported the 4th inputs - // optional tensor specifying sequence lengths in a batch, shape: [batch_size] - // const tvm::Tensor* seq_len = inputs[4] ? &inputs[4]->tensor : nullptr; - - const tvm::Tensor& prev_H = inputs[5]; // optional initial H, shape: [batch_size, hidden_size] - const tvm::Tensor& prev_C = inputs[6]; // optional initial C, shape: [batch_size, hidden_size] - - const tvm::Tensor& P = inputs[7]; // optional peepholes tensor with shape [3*hidde_size] - bool has_P = node.InputDefs()[7]->Exists(); - - tvm::Tensor Y_h; // shape: [batch_size, hidden_size] - tvm::Tensor Y_c; // shape: [batch_size, hidden_size] - LSTMAttributes lstm_attrs(hidden_size); - LSTM_cell(lstm_attrs, X, W, R, B, has_B, prev_H, prev_C, P, has_P, Y_h, Y_c); - - // Since we only generate lstm cell, lstm's states need to be always outputs, - // regardless whethere they are skipped or not. - // The skipped trailing outputs need to be handled by Execution - outputs.push_back(Y_h); - outputs.push_back(Y_c); - - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/nn/pool_ops.cc b/onnxruntime/core/codegen/passes/op_ir_creator/nn/pool_ops.cc deleted file mode 100644 index 84d3b7c1e0f79..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/nn/pool_ops.cc +++ /dev/null @@ -1,51 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/nn/pool_ops.h" -#include "core/framework/op_kernel_info.h" -#include "core/providers/cpu/nn/pool_attributes.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// A local macro to create Pool Ops - -// helper macro defines Evaluate of of POOL_OP OpIRCreators -#define POOL_OP(name) \ - Status GENERIC_OP_IR_CREATOR_CLASS(name)::Evaluate( \ - const tvm::Array& inputs, \ - const Node& node, \ - CodeGenContext& ctx_codegen, \ - tvm::Array& outputs) { \ - ORT_RETURN_IF_NOT(outputs.size() == 1, "multiple outputs are not supported yet!"); \ - ProtoHelperNodeContext ctx(node); \ - OpNodeProtoHelper info(&ctx); \ - int version = ctx_codegen.GetCodeGenHandle()->domain_version_lookup_func(node.Domain()); \ - PoolAttributes pool_attrs(info, #name, version); \ - for (auto n : pool_attrs.dilations) { \ - ORT_RETURN_IF_NOT(n <= 1, "dilations are not supported yet!"); \ - } \ - if (pool_attrs.global_pooling) { \ - if (inputs[0]->shape.size() != 4) { \ - ORT_NOT_IMPLEMENTED(gsl::narrow_cast(inputs[0]->shape.size()) - 2, "d global pooling is not implementated"); \ - } \ - } else { \ - if (pool_attrs.kernel_shape.size() != 2) { \ - ORT_NOT_IMPLEMENTED(pool_attrs.kernel_shape.size(), "d pooling is not implementated"); \ - } \ - } \ - tvm::Array dummy_output_shape; \ - tvm::Tensor Y = name(inputs[0], pool_attrs, dummy_output_shape); \ - outputs.push_back(Y); \ - return Status::OK(); \ - } - -LIST_POOL_OPS() - -#undef POOL_OP - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/cast.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/cast.cc deleted file mode 100644 index bd324fd359edf..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/cast.cc +++ /dev/null @@ -1,40 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/tensor/cast_ops.h" -#include "core/codegen/passes/utils/ort_tvm_utils.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Cast OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Cast)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - - int64_t to; - ORT_RETURN_IF_ERROR(attrs.GetAttr("to", &to)); - auto to_type_proto = gsl::narrow_cast(to); - - tvm::Tensor X = inputs[0]; - tvm::Tensor Y; - if (to_type_proto == ONNX_NAMESPACE::TensorProto_DataType_BOOL) { - // special case for bool as ONNX bool is uint8, while in tvm it's uint1 - Y = CastToUInt8Bool(X, node.Name() + "_Cast"); - } else { - Y = Cast(X, ToTvmType(to_type_proto), node.Name() + "_Cast"); - } - - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/concat.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/concat.cc deleted file mode 100644 index 418296889419e..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/concat.cc +++ /dev/null @@ -1,30 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/tensor/concat_ops.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Concat OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Concat)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper info(&ctx); - - int64_t axis; - ORT_RETURN_IF_ERROR(info.GetAttr("axis", &axis)); - - tvm::Tensor Y = Concat(inputs, axis, node.Name() + "_Concat"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/crop.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/crop.cc deleted file mode 100644 index 3b6a9a76f0723..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/crop.cc +++ /dev/null @@ -1,46 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/crop.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Crop OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Crop)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - - if (inputs[0]->shape.size() != 4) { - return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, - "Input is expected to have four dimensions corresponding to [N,C,H,W]"); - } - - std::vector border; - std::vector scale; - - ORT_ENFORCE(attrs.GetAttrs("border", border).IsOK()); - // scale is optional and status is false when omit - bool is_ok = attrs.GetAttrs("scale", scale).IsOK(); - ORT_UNUSED_PARAMETER(is_ok); - - if (border.size() != 4) { - return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, - "Attribute border needs to be specified with four border elements"); - } - - tvm::Tensor Y = Crop(inputs[0], ToTvmArray(border), ToTvmArray(scale), node.Name() + "_Crop"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/expand.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/expand.cc deleted file mode 100644 index 0f0e0cf0987b3..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/expand.cc +++ /dev/null @@ -1,26 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/expand.h" -#include "core/codegen/passes/utils/ort_tvm_utils.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Expand OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Expand)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - tvm::Tensor Y = Expand(inputs[0], ShapeToTvmArray(node.OutputDefs()[0], ctx_codegen), node.Name() + "_Expand"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/gather.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/gather.cc deleted file mode 100644 index 3a5d801b6839f..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/gather.cc +++ /dev/null @@ -1,30 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/tensor/gather.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Gather OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Gather)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - - int64_t axis; - ORT_ENFORCE(attrs.GetAttr("axis", &axis).IsOK()); - - tvm::Tensor Y = Gather(inputs[0], axis, inputs[1], node.Name() + "_Gather"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/gather_elements.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/gather_elements.cc deleted file mode 100644 index 0b71506cceed3..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/gather_elements.cc +++ /dev/null @@ -1,32 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/tensor/gather_elements.h" -#include "core/framework/op_kernel_info.h" -#include "core/providers/common.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of GatherElements OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(GatherElements)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - - int64_t axis; - ORT_ENFORCE(attrs.GetAttr("axis", &axis).IsOK()); - axis = HandleNegativeAxis(axis, gsl::narrow_cast(inputs[0]->shape.size())); - - tvm::Tensor Y = GatherElements(inputs[0], axis, inputs[1], node.Name() + "_GatherElements"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/pad.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/pad.cc deleted file mode 100644 index e9e20e8a43998..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/pad.cc +++ /dev/null @@ -1,49 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/pad_ops.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Pad OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Pad)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - - std::string mode; - std::vector pads; - float value; - - ORT_THROW_IF_ERROR(attrs.GetAttr("mode", &mode)); - ORT_THROW_IF_ERROR(attrs.GetAttrs("pads", pads)); - ORT_THROW_IF_ERROR(attrs.GetAttr("value", &value)); - - if (mode != "constant" && mode != "edge" && mode != "reflect") - return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Pad: Unsupported padding mode!"); - - if (pads.size() != 2 * inputs[0]->shape.size()) - return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Pad: pads rank does not match inputs rank!"); - - std::vector pad_before, pad_after; - size_t offset = pads.size() / 2; - for (size_t i = 0; i < offset; i++) { - pad_before.push_back(pads[i]); - pad_after.push_back(pads[i + offset]); - } - - tvm::Tensor Y = Pad(inputs[0], ToTvmArray(pad_before), ToTvmArray(pad_after), value, mode, node.Name() + "_Pad"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/reshape_ops.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/reshape_ops.cc deleted file mode 100644 index a83f598bc8ad1..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/reshape_ops.cc +++ /dev/null @@ -1,99 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/reshape_ops.h" -#include "core/codegen/passes/utils/ort_tvm_utils.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Dropout OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Dropout)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - tvm::Tensor Y = Identity(inputs[0]); - outputs.push_back(Y); - - // optional mask - // Support skipped trailing outputs - if (node.OutputDefs().size() > 1 && node.OutputDefs()[1]->Exists()) { - // A fake mask with all ones - auto l = [&](const tvm::Array& /*indices*/) { - return tvm::make_const(tvm::UInt(8), 1); - }; - tvm::Tensor mask = tvm::compute(inputs[0]->shape, l, "mask"); - outputs.push_back(mask); - } - - return Status::OK(); -} - -// Evaluate of Flatten OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Flatten)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - - int64_t axis; - ORT_RETURN_IF_ERROR(attrs.GetAttr("axis", &axis)); - - tvm::Tensor Y = Flatten(inputs[0], axis, node.Name() + "_Flatten"); - outputs.push_back(Y); - return Status::OK(); -} - -// Evaluate of Identity OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Identity)::Evaluate( - const tvm::Array& inputs, - const Node&, - CodeGenContext&, - tvm::Array& outputs) { - tvm::Tensor Y = Identity(inputs[0]); - outputs.push_back(Y); - return Status::OK(); -} - -// Evaluate of Reshape OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Reshape)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - tvm::Tensor Y = Reshape(inputs[0], ShapeToTvmArray(node.OutputDefs()[0], ctx_codegen), node.Name() + "_Reshape"); - outputs.push_back(Y); - return Status::OK(); -} - -// Evaluate of Squeeze OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Squeeze)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - tvm::Tensor Y = Reshape(inputs[0], ShapeToTvmArray(node.OutputDefs()[0], ctx_codegen), node.Name() + "_Squeeze"); - outputs.push_back(Y); - return Status::OK(); -} - -// Evaluate of Unsqueeze OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Unsqueeze)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - tvm::Tensor Y = Reshape(inputs[0], ShapeToTvmArray(node.OutputDefs()[0], ctx_codegen), node.Name() + "_Unsqueeze"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/shape_op.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/shape_op.cc deleted file mode 100644 index 84761ecac1397..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/shape_op.cc +++ /dev/null @@ -1,26 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/shape_op.h" -#include "core/codegen/passes/utils/ort_tvm_utils.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Expand OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Shape)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - tvm::Tensor Y = Shape(inputs[0], node.Name() + "_Expand"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/slice.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/slice.cc deleted file mode 100644 index 6a016580c41e4..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/slice.cc +++ /dev/null @@ -1,71 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" -#include "core/codegen/passes/utils/ort_tvm_utils.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/slice.h" -#include "core/framework/op_kernel_info.h" -#include "core/framework/tensorprotoutils.h" - -#include - -namespace onnxruntime { -namespace tvm_codegen { - -Status SliceCommon(const tvm::Array& inputs, - const Node& node, - tvm::Array& outputs, - const std::vector& starts, - const std::vector& ends, - const std::vector& axes1, - const std::vector& steps1) { - ORT_RETURN_IF_NOT(nullptr != node.InputDefs()[0], "nullptr == node.InputDefs()[0]"); - - std::vector axes; - if (axes1.size() == 0) { - for (size_t i = 0; i < starts.size(); ++i) { - axes.push_back(gsl::narrow_cast(i)); - } - } else { - axes = axes1; - } - - std::vector steps; - if (steps1.size() == 0) { - steps.resize(starts.size(), 1); - } else { - steps = steps1; - } - - tvm::Tensor Y = Slice(inputs[0], starts, ends, axes, steps, node.Name() + "_Slice"); - outputs.push_back(Y); - return Status::OK(); -} - -// Evaluate of Slice OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Slice)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper info(&ctx); - - // NOTE that in opset 10, Slice has changed starts/ends/axes from attribute to input - // which may lead to dynamic output shape. - int version = ctx_codegen.GetCodeGenHandle()->domain_version_lookup_func(node.Domain()); - ORT_RETURN_IF_NOT(version <= 9, "Dynamic Slice is not supported yet"); - - std::vector starts, ends, steps; - ORT_RETURN_IF_ERROR(info.GetAttrs("starts", starts)); - ORT_RETURN_IF_ERROR(info.GetAttrs("ends", ends)); - ORT_RETURN_IF_NOT(starts.size() == ends.size(), "starts.size() != ends.size()"); - - auto axes = info.GetAttrsOrDefault("axes"); - - return SliceCommon(inputs, node, outputs, starts, ends, axes, steps); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/split.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/split.cc deleted file mode 100644 index ec52d98b5bf96..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/split.cc +++ /dev/null @@ -1,66 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/split.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Split OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Split)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper info(&ctx); - - int64_t axis; - ORT_RETURN_IF_ERROR(info.GetAttr("axis", &axis)); - axis = HandleNegativeAxis(axis, gsl::narrow_cast(inputs[0]->shape.size())); - std::vector split_sizes; - - int64_t split_size_sum = 0; - if (info.GetAttrs("split", split_sizes).IsOK()) { - // optional - split_size_sum = std::accumulate(split_sizes.cbegin(), split_sizes.cend(), 0LL); - ORT_RETURN_IF_NOT(std::all_of(split_sizes.cbegin(), split_sizes.cend(), [](int64_t value) { return value > 0; }), - "Invalid value in 'split' attribute. All values must be > 0"); - - // check split sizes - for (size_t i = 0; i < node.OutputDefs().size(); ++i) { - ORT_RETURN_IF_NOT(split_sizes[i] == ShapeValue(node.OutputDefs()[i], gsl::narrow(axis)), - "split_sizes[i] != ShapeValue(node.OutputDefs()[i], axis)"); - } - - } else { - for (size_t i = 0; i < node.OutputDefs().size(); ++i) { - split_sizes.push_back(ShapeValue(node.OutputDefs()[i], gsl::narrow(axis))); - split_size_sum += split_sizes[i]; - } - } - - // check total size - if (ShapeHasValue(node.InputDefs()[0], axis)) { - int64_t input_axis_dim = ShapeValue(node.InputDefs()[0], axis); - if (split_size_sum != input_axis_dim) { - return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, - "Cannot split using values in 'split' attribute. Axis=", axis, - " Dim being splitted=", input_axis_dim, - " Sum of sizes in 'split' (must equal size of selected axis) was ", split_size_sum); - } - } - - tvm::Array output_tensors = Split(inputs[0], ToTvmArray(split_sizes), axis, node.Name() + "_Split"); - for (size_t i = 0; i < node.OutputDefs().size(); ++i) { - outputs.push_back(output_tensors[i]); - } - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/transpose.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/transpose.cc deleted file mode 100644 index 43999ebd1f465..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/transpose.cc +++ /dev/null @@ -1,48 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/transpose.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Transpose OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Transpose)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - - size_t input_0_shape_rank = inputs[0]->shape.size(); - std::vector permute; - bool is_ok = attrs.GetAttrs("perm", permute).IsOK(); - if (permute.size() != 0 && permute.size() != input_0_shape_rank) - return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Transpose: Incorrect permute size"); - - std::vector default_permute; - const std::vector* perm; - // either we don't have perm attribute or the perm attribute is empty - bool use_default_perm = !is_ok || permute.size() == 0; - if (use_default_perm) { - default_permute.resize(input_0_shape_rank); - for (size_t i = 0; i < input_0_shape_rank; ++i) { - default_permute[i] = gsl::narrow(input_0_shape_rank - 1 - i); - } - perm = &default_permute; - } else { - perm = &permute; - } - - tvm::Tensor Y = Transpose(inputs[0], ToTvmArrayInt(*perm), node.Name() + "_Transpose"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/where.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tensor/where.cc deleted file mode 100644 index 9d6df7c1c430d..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tensor/where.cc +++ /dev/null @@ -1,28 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/all_ops.h" - -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/mti/tensor/where.h" -#include "core/framework/op_kernel_info.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Evaluate of Transpose OpIRCreator -Status GENERIC_OP_IR_CREATOR_CLASS(Where)::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext&, - tvm::Array& outputs) { - ProtoHelperNodeContext ctx(node); - OpNodeProtoHelper attrs(&ctx); - - tvm::Tensor Y = Where(inputs[0], inputs[1], inputs[2], node.Name() + "_Where"); - outputs.push_back(Y); - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tvm_ir_builder.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tvm_ir_builder.cc deleted file mode 100644 index 7889e2add755e..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tvm_ir_builder.cc +++ /dev/null @@ -1,125 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/tvm_ir_builder.h" - -#include "core/codegen/common/op_macro.h" -#include "core/codegen/passes/op_ir_creator/all_ops.h" -#include "core/common/common.h" - -namespace onnxruntime { -namespace tvm_codegen { - -TVMIRBuilder::TVMIRBuilder(const std::string& name) - : name_(name) {} - -const std::string& TVMIRBuilder::Name() const { - return name_; -} - -void TVMIRBuilder::InsertDispatcher(std::unique_ptr&& ptr) { - dispatchers_.push_back(std::move(ptr)); -} - -void TVMIRBuilder::ClearAllDispatchers() { - dispatchers_.clear(); -} - -void TVMIRBuilder::DumpAllOpIRCreators() const { - int count = 0; - for (auto& d : dispatchers_) { - std::cout << "************ TVM OpIRDispatcher " - << count << " : " - << d->Name() - << " ************" << std::endl; - - d->ForEach([](const std::string& key, OpIRCreator* builder) { - std::cout << "Key " << key - << ", Creator " << builder->Name() << std::endl; - }); - - ++count; - } -} - -// Evaluate finds ONE proper OpIRCreator and build the corresponding OpIR -// If a TVMIRBuilder has more than one OpIRCreator for an ORT Op, -// the first one will be used. -// Please adjust registration order and dispatcher in TVMIRBuilder -// to make sure the proper OpIRCreator is called. -Status TVMIRBuilder::Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx_codegen, - tvm::Array& outputs) { - OpIRCreator* candidate = nullptr; - for (auto& d : dispatchers_) { - candidate = d->Find(node); - if (nullptr != candidate) - break; - } - - if (nullptr == candidate) { - return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Not implemented: ", node.OpType()); - } - - ORT_RETURN_IF_ERROR(candidate->Evaluate(inputs, node, ctx_codegen, outputs)); - - return Status::OK(); -} - -// BEGIN: Generic IR creator classes -#define ADD_OP_ITEM(name) \ - op_ir_registry->Register(std::make_unique()); - -#define BINARY_OP(name) ADD_OP_ITEM(name) -#define BINARY_CMP_OP(name) ADD_OP_ITEM(name) -#define POOL_OP(name) ADD_OP_ITEM(name) -#define REDUCE_OP(name) ADD_OP_ITEM(name) -#define REDUCE_INDEXED_OP(name) ADD_OP_ITEM(name) -#define UNARY_OP(name) ADD_OP_ITEM(name) -#define VARIADIC_OP(name) ADD_OP_ITEM(name) - -void RegisterAllGenericOpIRCreators(OpIRRegistry* op_ir_registry) { - LIST_ALL_GENERIC_OPS(); -} - -#undef ADD_OP_ITEM -#undef BINARY_OP -#undef BINARY_CMP_OP -#undef POOL_OP -#undef REDUCE_OP -#undef REDUCE_INDEXED_OP -#undef UNARY_OP -#undef VARIADIC_OP - -// BEGIN: Plugin Generic IR creator classes -#define ADD_OP_ITEM(name) \ - dispatcher->Register(#name, registry->Get(GENERIC_OP_IR_CREATOR_STRING(name))); - -#define BINARY_OP(name) ADD_OP_ITEM(name) -#define BINARY_CMP_OP(name) ADD_OP_ITEM(name) -#define POOL_OP(name) ADD_OP_ITEM(name) -#define REDUCE_OP(name) ADD_OP_ITEM(name) -#define REDUCE_INDEXED_OP(name) ADD_OP_ITEM(name) -#define UNARY_OP(name) ADD_OP_ITEM(name) -#define VARIADIC_OP(name) ADD_OP_ITEM(name) - -void RegisterGenericOrtOpTypeDispatcher(const std::shared_ptr& builder, - const OpIRRegistry* registry) { - auto dispatcher = std::make_unique("GenericOrtOpTypeOpIRCreators"); - LIST_ALL_GENERIC_OPS() - builder->InsertDispatcher(std::move(dispatcher)); -} - -#undef ADD_OP_ITEM -#undef BINARY_OP -#undef BINARY_CMP_OP -#undef POOL_OP -#undef REDUCE_OP -#undef REDUCE_INDEXED_OP -#undef UNARY_OP -// END: Generic IR creators classes - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tvm_ir_builder.h b/onnxruntime/core/codegen/passes/op_ir_creator/tvm_ir_builder.h deleted file mode 100644 index c80056e619d6d..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tvm_ir_builder.h +++ /dev/null @@ -1,64 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/passes/utils/codegen_context.h" -#include "core/codegen/passes/op_ir_creator/tvm_op_creator.h" -#include "core/common/common.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// TVMIRBuilder contains all applicable TVM OpIRCreators -// OpIRCreators are stored in multiple dispatchers -// that check different conditions of an ORT Node. - -// If an ORT Node satisfies more than one OpIRCreators, -// the first dispatched pass will be applied. - -class TVMIRBuilder { - public: - TVMIRBuilder(const std::string& name); - ~TVMIRBuilder() = default; - - // A debug dumps all existing in this TVMIRBuilders - void DumpAllOpIRCreators() const; - - // Evaluates an OpIRCreator that first satisfies condtions of all dispatchers - Status Evaluate( - const tvm::Array& inputs, - const Node& node, - CodeGenContext& ctx, - tvm::Array& outputs); - - // Inserts a dispatcher and move its ownership to this TVMIRBuilder - void InsertDispatcher(std::unique_ptr&& ptr); - - // Clears all dispatchers in this TVMIRBuilder - void ClearAllDispatchers(); - - // Dumps the name of this TVMIRBuilder - const std::string& Name() const; - - private: - std::vector> dispatchers_; - std::string name_; - - private: - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(TVMIRBuilder); -}; - -// Utility function to register all builtin generic OpIRCreators into an OpIRRegistry. -// It creates instances of all generic OpIRCreators -// and registers them to op_ir_registry -void RegisterAllGenericOpIRCreators(OpIRRegistry* op_ir_registry); - -// Utility function to bind all builtin generic OpIRCreators to a TVMIRBuilder. -// It creates an instance of a Dispatcher that contains all generic OpIRCreators created above -// and uses OrtOpType to dispatch OpIRCreators. -// Then, it registers the created Dispatcher to a TVMIRBuilder, builder. -void RegisterGenericOrtOpTypeDispatcher(const std::shared_ptr& builder, - const OpIRRegistry* registry); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tvm_op_creator.cc b/onnxruntime/core/codegen/passes/op_ir_creator/tvm_op_creator.cc deleted file mode 100644 index 992272753f5a4..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tvm_op_creator.cc +++ /dev/null @@ -1,37 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/op_ir_creator/tvm_op_creator.h" - -#include "core/codegen/common/common.h" -#include "core/codegen/common/dispatcher.h" -#include "core/codegen/passes/utils/codegen_context.h" - -namespace onnxruntime { -namespace codegen { -// Explicit instantiation for OpIRCreator -template class CreatorBase&, - const Node&, - tvm_codegen::CodeGenContext&, - tvm::Array&, - Status>; - -// Explicit instantiation for OpIRCreators' dispatcher -template class DispatcherBase; - -} // namespace codegen - -namespace tvm_codegen { - -// One dispatcher is based on ORT OpType -OpIRCreator* OP_IR_DISPATCHER_CLASS(OpType)::Find(const Node& node) { - return DispatcherBase::Get(node.OpType()); -} - -// Another dispatcher is based ORT NodeArg name (GetKey) -OpIRCreator* OP_IR_DISPATCHER_CLASS(NodeName)::Find(const Node& node) { - return DispatcherBase::Get(GetKey(&node)); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/op_ir_creator/tvm_op_creator.h b/onnxruntime/core/codegen/passes/op_ir_creator/tvm_op_creator.h deleted file mode 100644 index e29c4a9f20767..0000000000000 --- a/onnxruntime/core/codegen/passes/op_ir_creator/tvm_op_creator.h +++ /dev/null @@ -1,84 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/common/creator.h" -#include "core/codegen/common/dispatcher.h" -#include "core/codegen/common/registry.h" -#include "core/graph/graph.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -class CodeGenContext; - -// OpIRCreator lowers an Ort Node to its corresponding TVM IRs -using OpIRCreator = codegen::CreatorBase< - const tvm::Array&, - const Node&, - CodeGenContext&, - tvm::Array&, - Status>; - -// OpIRDispatcher is the base dispatcher for TVM IR Builder -// It checks whether an Ort Node satisfying a criteria (in Find) -// and dispatches a corresponding OpIRCreator. -class OpIRDispatcher : public codegen::DispatcherBase { - public: - OpIRDispatcher(const std::string& name) - : DispatcherBase(name) {} - - virtual ~OpIRDispatcher() = default; - - virtual OpIRCreator* Find(const Node&) = 0; - - private: - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(OpIRDispatcher); -}; - -// Macro returns an OpIRCreators' dispatcher's name -#define OP_IR_DISPATCHER_CLASS(OP) \ - TVM##OP##IRCreator - -// Macro declares an OpIRCreators' dispatcher -#define DECLARE_OP_IR_DISPATCHER_CLASS(OP) \ - class OP_IR_DISPATCHER_CLASS(OP) : public OpIRDispatcher { \ - public: \ - TVM##OP##IRCreator(const std::string& name) \ - : OpIRDispatcher(name) {} \ - ~TVM##OP##IRCreator() = default; \ - OpIRCreator* Find(const Node&) override; \ - \ - private: \ - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(OP_IR_DISPATCHER_CLASS(OP)); \ - }; - -// Declare two common dispatchers for TVM Op IR builders -// One dispatcher is based on Ort OpType -DECLARE_OP_IR_DISPATCHER_CLASS(OpType) -// Another dispatcher is based Ort NodeArg name -DECLARE_OP_IR_DISPATCHER_CLASS(NodeName) - -// OpIRCreator Registry is a registry holds all OpIRCreators -using OpIRRegistry = codegen::RegistryBase; - -// Macro declares an OpIRCreator -#define DECLARE_OP_IR_CREATOR_CLASS(OP, PREFIX) \ - DECLARE_CREATOR_CLASS(OP, PREFIX##IRCreator, \ - const tvm::Array&, \ - const Node&, \ - tvm_codegen::CodeGenContext&, \ - tvm::Array&, \ - Status) - -// Macro returns an OpIRCreator's name with prefix -#define OP_IR_CREATOR_CLASS_EX(OP, PREFIX, ARCH) \ - CREATOR_CLASS(OP, PREFIX##ARCH##IRCreator) - -// Macro declares an OpIRCreator with prefix and arch -#define DECLARE_OP_IR_CREATOR_CLASS_EX(OP, PREFIX, ARCH) \ - DECLARE_OP_IR_CREATOR_CLASS(OP, PREFIX##ARCH) - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/scheduler/all_schedules.h b/onnxruntime/core/codegen/passes/scheduler/all_schedules.h deleted file mode 100644 index fe4be90f9fc84..0000000000000 --- a/onnxruntime/core/codegen/passes/scheduler/all_schedules.h +++ /dev/null @@ -1,20 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/passes/scheduler/tvm_scheduler.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// AlwaysRoot is for debug purpose -DECLARE_TVM_SCHEDULER_CLASS(AlwaysRoot, GenericTVMRule) -// Create schedule for TVM Rule -DECLARE_TVM_SCHEDULER_CLASS(Extern, GenericTVMRule) -DECLARE_TVM_SCHEDULER_CLASS(Reduce, GenericTVMRule) - -// Crete scheduler for ORT OpType, Softmax -DECLARE_TVM_SCHEDULER_CLASS(Softmax, GenericOrtOpType) - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/scheduler/ort_type_schedule.cc b/onnxruntime/core/codegen/passes/scheduler/ort_type_schedule.cc deleted file mode 100644 index 59f492d164b14..0000000000000 --- a/onnxruntime/core/codegen/passes/scheduler/ort_type_schedule.cc +++ /dev/null @@ -1,22 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/scheduler/all_schedules.h" - -#include "core/codegen/passes/scheduler/schedule_utils.h" - -namespace onnxruntime { -namespace tvm_codegen { - -bool TVM_SCHEDULER_CLASS(Softmax, GenericOrtOpType)::Evaluate( - const tvm::Tensor& tensor, - const Node*, - CodeGenContext&, - ScheduleContext& ctx_sched) { - // compute root the exp since it is reused more than once - auto& tensor_exp = tensor->op->InputTensors()[0]; - return InsertRootSchedule(tensor_exp, ctx_sched); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/scheduler/schedule_utils.cc b/onnxruntime/core/codegen/passes/scheduler/schedule_utils.cc deleted file mode 100644 index 76c2ad509c401..0000000000000 --- a/onnxruntime/core/codegen/passes/scheduler/schedule_utils.cc +++ /dev/null @@ -1,178 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/common/utils.h" -#include "core/codegen/passes/scheduler/schedule_utils.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// Check the schedule of tensor -// If it has no compute_root, Insert compute_root to tensor, and record it to ctx.scheduled_tensors -bool InsertRootSchedule( - const tvm::Tensor& tensor, - ScheduleContext& ctx) { - auto it = ctx.scheduled_tensors.find(tensor->op.get()); - if (it != ctx.scheduled_tensors.end()) { - if (it->second == ScheduleType::ScheduleClosure || - it->second == ScheduleType::ScheduleRoot) { - return false; - } - it->second = ScheduleType::ScheduleRoot; - } else { - ctx.scheduled_tensors.insert(std::make_pair(tensor->op.get(), ScheduleType::ScheduleRoot)); - } - ctx.schedule[tensor->op].compute_root(); - return true; -} - -// Check the schedule of tensor -// If it is not labeled as closure, lable it. -bool InsertClosure(const tvm::Tensor& tensor, - ScheduleContext& ctx) { - auto it = ctx.scheduled_tensors.find(tensor->op.get()); - if (it != ctx.scheduled_tensors.end()) { - if (it->second == ScheduleType::ScheduleClosure) - return false; - it->second = ScheduleType::ScheduleClosure; - } else { - ctx.scheduled_tensors.insert(std::make_pair(tensor->op.get(), ScheduleType::ScheduleClosure)); - } - return true; -} - -// Combination of InsertRootSchedule and InsertClosure -bool InsertRootScheduleAndClosure( - const tvm::Tensor& tensor, - ScheduleContext& ctx) { - auto it = ctx.scheduled_tensors.find(tensor->op.get()); - if (it != ctx.scheduled_tensors.end()) { - if (it->second == ScheduleType::ScheduleClosure) { - return false; - } - it->second = ScheduleType::ScheduleClosure; - } else { - ctx.scheduled_tensors.insert(std::make_pair(tensor->op.get(), ScheduleType::ScheduleClosure)); - } - ctx.schedule[tensor->op].compute_root(); - return true; -} - -// Check precondition for vectorize schedule -bool ShouldTryVectorization( - const tvm::Tensor& tensor, - ScheduleContext& ctx) { - auto it = ctx.scheduled_tensors.find(tensor->op.get()); - if (it != ctx.scheduled_tensors.end()) { - if (it->second > ScheduleType::ScheduleInline) { - return false; - } - } - return true; -} - -// Check the schedule of tensor -// If it is not scheduled, try to vectorize it. -// Note TryVectorization has to use with compute_root. -// Therefore, there is a safety check of tensor's schedule -bool TryVectorization( - const tvm::Tensor& tensor, - int64_t natural_vector_size, - ScheduleContext& ctx) { - if (!ShouldTryVectorization(tensor, ctx)) - return false; - - auto shape = tensor->shape; - auto rank = shape.size(); - if (rank < 1) { - return false; - } - const int64_t* tail_dim = as_const_int(shape[rank - 1]); - - if (nullptr != tail_dim) { - auto extern_op = tensor->op.as(); - if (nullptr != extern_op) { - return false; - } - - auto compute_op = tensor->op.as(); - - if (nullptr != compute_op) { - auto axis = compute_op->axis; - tvm::IterVar x = axis[rank - 1]; - if ((*tail_dim) > natural_vector_size) { - if ((*tail_dim) % natural_vector_size != 0) { - natural_vector_size = GCD(natural_vector_size, (*tail_dim)); - } - - if (natural_vector_size > 1) { - tvm::IterVar xi, xo; - ctx.schedule[tensor->op].split(x, static_cast(natural_vector_size), &xo, &xi); - ctx.schedule[tensor->op].vectorize(xi); - return true; - } - } else if (*tail_dim > 0) { - // don't vectorize if dim is 0 - ctx.schedule[tensor->op].vectorize(x); - return true; - } - } - } - return false; -} - -// Check the schedule of tensor -// If it is not scheduled, try to add compute_inline on it. -// Note TryInlineSchedule cannot be used with compute_root. -// Therefore, there is a safety check of tensor's schedule. -bool TryInlineSchedule( - const tvm::Tensor& tensor, - ScheduleContext& ctx) { - auto it = ctx.scheduled_tensors.find(tensor->op.get()); - if (it != ctx.scheduled_tensors.end()) { - if ((int)it->second < (int)ScheduleType::ScheduleInline) { - ctx.schedule[tensor->op].compute_inline(); - it->second = ScheduleType::ScheduleInline; - return true; - } else { - return false; - } - } - ctx.schedule[tensor->op].compute_inline(); - ctx.scheduled_tensors.insert(std::make_pair(tensor->op.get(), ScheduleType::ScheduleInline)); - return true; -} - -// Check the schedule of tensor's inputs, and call InsertRootSchedule for each of them -bool InputRootSchedule( - const tvm::Tensor& tensor, - ScheduleContext& ctx) { - bool status = false; - for (auto& t : tensor->op->InputTensors()) { - if (t->op->InputTensors().size() > 0) { - bool status_root = InsertRootSchedule(t, ctx); - status = status || status_root; - } - } - return status; -} - -// Check the schedule of tensor's inputs, -// and call InsertRootSchedule and TryVectorization for each of them -bool InputRootScheduleWithVectorization( - const tvm::Tensor& tensor, - int64_t natural_vector_size, - ScheduleContext& ctx) { - bool status = false; - for (auto& t : tensor->op->InputTensors()) { - if (t->op->InputTensors().size() > 0) { - bool status_vec = TryVectorization(t, natural_vector_size, ctx); - bool status_root = InsertRootSchedule(t, ctx); - status = status || status_root || status_vec; - } - } - return status; -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/scheduler/schedule_utils.h b/onnxruntime/core/codegen/passes/scheduler/schedule_utils.h deleted file mode 100644 index 4a0781f94d385..0000000000000 --- a/onnxruntime/core/codegen/passes/scheduler/schedule_utils.h +++ /dev/null @@ -1,65 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// Check the schedule of tensor -// If it has no compute_root, Insert compute_root to tensor, -// and record it to ctx.scheduled_tensors -bool InsertRootSchedule( - const tvm::Tensor& tensor, - ScheduleContext& ctx); - -// Check the schedule of tensor -// If it is not labeled as closure, lable it. -bool InsertClosure( - const tvm::Tensor& tensor, - ScheduleContext& ctx); - -// Combination of InsertRootSchedule and InsertClosure -bool InsertRootScheduleAndClosure( - const tvm::Tensor& tensor, - ScheduleContext& ctx); - -// Check precondition for vectorize schedule -bool ShouldTryVectorization( - const tvm::Tensor& tensor, - ScheduleContext& ctx); - -// Check the schedule of tensor -// If it is not scheduled, try to vectorize it. -// Note TryVectorization has to use with compute_root. -// Therefore, there is a safety check of tensor's schedule -bool TryVectorization( - const tvm::Tensor& tensor, - int64_t natural_vector_size, - ScheduleContext& ctx); - -// Check the schedule of tensor -// If it is not scheduled, try to add compute_inline on it. -// Note TryInlineSchedule cannot be used with compute_root. -// Therefore, there is a safety check of tensor's schedule. -bool TryInlineSchedule( - const tvm::Tensor& tensor, - ScheduleContext& ctx); - -// Check the schedule of tensor's inputs, -// and call InsertRootSchedule for each of them -bool InputRootSchedule( - const tvm::Tensor& tensor, - ScheduleContext& ctx); - -// Check the schedule of tensor's inputs, -// and call InsertRootSchedule and TryVectorization for each of them -bool InputRootScheduleWithVectorization( - const tvm::Tensor& tensor, - int64_t natural_vector_size, - ScheduleContext& ctx); - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/scheduler/tvm_rule_schedule.cc b/onnxruntime/core/codegen/passes/scheduler/tvm_rule_schedule.cc deleted file mode 100644 index 33162deddc983..0000000000000 --- a/onnxruntime/core/codegen/passes/scheduler/tvm_rule_schedule.cc +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/scheduler/all_schedules.h" - -#include "core/codegen/passes/scheduler/schedule_utils.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// This is for debug -bool TVM_SCHEDULER_CLASS(AlwaysRoot, GenericTVMRule)::Evaluate( - const tvm::Tensor& tensor, - const Node*, - CodeGenContext&, - ScheduleContext& ctx_sched) { - return InsertRootSchedule(tensor, ctx_sched); -} - -// For External tvm::Tensor -bool TVM_SCHEDULER_CLASS(Extern, GenericTVMRule)::Evaluate( - const tvm::Tensor& tensor, - const Node*, - CodeGenContext&, - ScheduleContext& ctx_sched) { - bool status = InsertRootScheduleAndClosure(tensor, ctx_sched); - bool status_input = InputRootSchedule(tensor, ctx_sched); - return status || status_input; -} - -// For Reduce Compute tvm::Tensor -bool TVM_SCHEDULER_CLASS(Reduce, GenericTVMRule)::Evaluate( - const tvm::Tensor& tensor, - const Node*, - CodeGenContext&, - ScheduleContext& ctx_sched) { - return InsertRootScheduleAndClosure(tensor, ctx_sched); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/scheduler/tvm_schedule_builder.cc b/onnxruntime/core/codegen/passes/scheduler/tvm_schedule_builder.cc deleted file mode 100644 index 2c8250198fa5f..0000000000000 --- a/onnxruntime/core/codegen/passes/scheduler/tvm_schedule_builder.cc +++ /dev/null @@ -1,104 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/scheduler/tvm_schedule_builder.h" - -#include "core/codegen/common/op_macro.h" -#include "core/codegen/common/settings.h" -#include "core/common/common.h" -#include "core/common/logging/logging.h" - -namespace onnxruntime { -namespace tvm_codegen { - -TVMScheduleBuilder::TVMScheduleBuilder(const std::string& name) - : name_(name) { -} - -const std::string& TVMScheduleBuilder::Name() const { - return name_; -} - -void TVMScheduleBuilder::InsertDispatcher(std::unique_ptr&& ptr) { - dispatchers_.push_back(std::move(ptr)); -} - -void TVMScheduleBuilder::ClearDispatcher() { - dispatchers_.clear(); -} - -void TVMScheduleBuilder::DumpAllSchedulers() const { - std::ostringstream stream; - int count = 0; - stream << "[CODEGEN_DUMP_SCHEDULE]" << std::endl; - for (auto& d : dispatchers_) { - stream << "************ TVM Scheduler Dispatcher " - << count << " : " - << d->Name() - << " ************" << std::endl; - - d->ForEach([&stream](const std::string& key, Scheduler* op) { - stream << "Key " << key - << ", Creator " << op->Name() << std::endl; - }); - - ++count; - } - - LOGS_DEFAULT(CODEGEN_SETTINGS_LOG_LEVEL) << stream.str(); -} - -Status TVMScheduleBuilder::Evaluate( - const tvm::Tensor& tensor, - const Node* node, - CodeGenContext& ctx_codegen, - ScheduleContext& sched) { - Scheduler* candidate = nullptr; - - for (auto& d : dispatchers_) { - candidate = d->Find(tensor, node, ctx_codegen); - if (nullptr != candidate) - break; - } - - bool enable_dump_schedule = codegen::CodeGenSettings::Instance().HasOption(codegen::CodeGenSettings::kCodeGenDumpSchedule); - - if (nullptr == candidate) { - if (nullptr != node) - return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Not implemented: ", node->OpType()); - return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Not implemented an internal tvm::Tensor: ", tensor->op->name); - } - - bool status = candidate->Evaluate(tensor, node, ctx_codegen, sched); - - if (enable_dump_schedule) { - std::ostringstream stream; - if (nullptr != node) { - stream << std::endl; - stream << "[CODEGEN_DUMP_SCHEDULE] " - << "Schedule Node: " << node->Name() << std::endl; - } else { - stream << std::endl; - } - - if (status) { - stream << "[CODEGEN_DUMP_SCHEDULE] " - << "Schedule tvm::Tesnor " - << tensor->op->name - << " with " - << candidate->Name() << std::endl; - } else { - stream << "[CODEGEN_DUMP_SCHEDULE] " - << "Schedule tvm::Tesnor " - << tensor->op->name - << " is suppressed " << std::endl; - } - - LOGS_DEFAULT(CODEGEN_SETTINGS_LOG_LEVEL) << stream.str(); - } - - return Status::OK(); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/scheduler/tvm_schedule_builder.h b/onnxruntime/core/codegen/passes/scheduler/tvm_schedule_builder.h deleted file mode 100644 index 9f0a1b3ef45c2..0000000000000 --- a/onnxruntime/core/codegen/passes/scheduler/tvm_schedule_builder.h +++ /dev/null @@ -1,46 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/passes/scheduler/tvm_scheduler.h" -#include "core/common/common.h" - -namespace onnxruntime { -namespace tvm_codegen { - -// TVMScheduleBuilder contains all applicable TVM scheduler passes. -// Scheduler passes are stored in multiple dispatchers -// that check different conditions of a tvm::Tensor. - -// If a tvm::Tensor satisfies more than one TVM scheduler passes, -// the first dispatched pass will be applied. - -class TVMScheduleBuilder { - public: - // TODO: add more parameter in consructor to support different target - TVMScheduleBuilder(const std::string& name); - ~TVMScheduleBuilder() = default; - - void DumpAllSchedulers() const; - - Status Evaluate( - const tvm::Tensor& tensor, - const Node* node, - CodeGenContext& ctx, - ScheduleContext& sched); - - void InsertDispatcher(std::unique_ptr&& ptr); - void ClearDispatcher(); - - const std::string& Name() const; - - private: - std::vector> dispatchers_; - std::string name_; - - private: - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(TVMScheduleBuilder); -}; - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/scheduler/tvm_scheduler.cc b/onnxruntime/core/codegen/passes/scheduler/tvm_scheduler.cc deleted file mode 100644 index 071200a234e33..0000000000000 --- a/onnxruntime/core/codegen/passes/scheduler/tvm_scheduler.cc +++ /dev/null @@ -1,79 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/scheduler/tvm_scheduler.h" - -#include "core/codegen/common/common.h" -#include "core/codegen/common/dispatcher.h" -#include "core/codegen/passes/utils/codegen_context.h" - -namespace onnxruntime { -namespace codegen { -// explicit instantiation -template class CreatorBase; - -template class DispatcherBase; - -} // namespace codegen - -namespace tvm_codegen { - -static const std::string TMVOpRuleKey_Extern("TVMOpRule_Extern"); -static const std::string TMVOpRuleKey_ComputeReduce("TVMOpRule_ComputeReduce"); -static const std::string TMVOpRuleKey_ComputeRegular("TVMOpRule_ComputeRegular"); -static const std::string TMVOpRuleKey_AlwaysRoot("TMVOpRuleKey_AlwaysRoot"); -static const std::string TMVOpRuleKey_NoRule("TVMOpRule_NoRule"); - -const std::string& GetTVMOpRule(TVMOpRuleType rule) { - if (rule == TVMOpRuleType::Extern) { - return TMVOpRuleKey_Extern; - } else if (rule == TVMOpRuleType::ComputeReduce) { - return TMVOpRuleKey_ComputeReduce; - } else if (rule == TVMOpRuleType::AlwaysRoot) { - return TMVOpRuleKey_AlwaysRoot; - } - return TMVOpRuleKey_NoRule; -} - -const std::string& GetTVMOpRule(const tvm::Tensor& tensor) { - auto extern_op = tensor->op.as(); - - if (nullptr != extern_op) { - return TMVOpRuleKey_Extern; - } - - auto compute_op = tensor->op.as(); - if (nullptr != compute_op) { - if (compute_op->reduce_axis.size() > 0) { - return TMVOpRuleKey_ComputeReduce; - } - } - - return TMVOpRuleKey_NoRule; -} - -Scheduler* SCHEDULE_DISPATCHER_CLASS(OrtOpType):: - Find(const tvm::Tensor&, const Node* node, tvm_codegen::CodeGenContext&) { - if (nullptr == node) - return nullptr; - return DispatcherBase::Get(node->OpType()); -} - -Scheduler* SCHEDULE_DISPATCHER_CLASS(TVMOpRule):: - Find(const tvm::Tensor& tensor, const Node*, tvm_codegen::CodeGenContext&) { - return DispatcherBase::Get(GetTVMOpRule(tensor)); -} - -Scheduler* SCHEDULE_DISPATCHER_CLASS(OrtOpName):: - Find(const tvm::Tensor&, const Node* node, tvm_codegen::CodeGenContext&) { - if (nullptr == node) - return nullptr; - return DispatcherBase::Get(GetKey(node)); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/scheduler/tvm_scheduler.h b/onnxruntime/core/codegen/passes/scheduler/tvm_scheduler.h deleted file mode 100644 index d022497c77f7e..0000000000000 --- a/onnxruntime/core/codegen/passes/scheduler/tvm_scheduler.h +++ /dev/null @@ -1,128 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/common/common.h" -#include "core/codegen/common/creator.h" -#include "core/codegen/common/registry.h" -#include "core/codegen/passes/utils/codegen_context.h" -#include "core/graph/graph.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// These are current generic TVMOpRule we used. -enum class TVMOpRuleType : int { - Extern = 0, - ComputeReduce = 1, - ComputeRegular = 2, - AlwaysRoot = 3, // for debug - NoRule, -}; - -const std::string& GetTVMOpRule(const tvm::Tensor& tensor); -const std::string& GetTVMOpRule(TVMOpRuleType rule); - -// These are current generic ScheduleType in tvm_codegen -enum class ScheduleType : int { - ScheduleNone = 0, - ScheduleInline = 1, - ScheduleAt = 2, - ScheduleRoot = 3, - ScheduleClosure = 4, -}; - -// Data struct to bundle tvm::Schedule and scheduled tensor -struct ScheduleContext { - ScheduleContext(const tvm::Array& ops) { - schedule = tvm::create_schedule(ops); - } - tvm::Schedule schedule; - std::map scheduled_tensors; -}; - -// Scheduler inserts a tvm::Schedule content to a tvm::Tensor -using Scheduler = codegen::CreatorBase< - const tvm::Tensor&, - const Node*, - tvm_codegen::CodeGenContext&, - ScheduleContext&, - bool>; - -// TVMScheduleDispatcher is the base dispatcher for TVM Schedule Builder -// It checks whether a pair of {tvm::Tensor, Ort Node} satisfying a criteria (in Find) -// and dispatches a corresponding Scheduler. -class TVMScheduleDispatcher : public codegen::DispatcherBase { - public: - TVMScheduleDispatcher(const std::string& name) - : DispatcherBase(name) {} - - virtual ~TVMScheduleDispatcher() = default; - - virtual Scheduler* Find(const tvm::Tensor&, - const Node*, - tvm_codegen::CodeGenContext&) = 0; - - private: - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(TVMScheduleDispatcher); -}; - -// Macro returns an Schedulers' dispatcher's name -#define SCHEDULE_DISPATCHER_CLASS(TYPE) \ - TVM##TYPE##Schedulers - -// Macro declares an Schedulers' dispatcher -#define DECLARE_SCHEDULE_DISPATCHER_CLASS(TYPE) \ - class SCHEDULE_DISPATCHER_CLASS(TYPE) : public tvm_codegen::TVMScheduleDispatcher { \ - public: \ - TVM##TYPE##Schedulers(const std::string& name) \ - : TVMScheduleDispatcher(name) {} \ - ~TVM##TYPE##Schedulers() = default; \ - tvm_codegen::Scheduler* Find(const tvm::Tensor&, \ - const Node*, \ - tvm_codegen::CodeGenContext&) override; \ - \ - private: \ - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(TVM##TYPE##Schedulers); \ - }; - -// Common dispatchers are listed here -// For a special pattern, it can be created later. -// One dispatcher is based on Ort OpType -DECLARE_SCHEDULE_DISPATCHER_CLASS(OrtOpType) -// One dispatcher is based on TVMOpRule -DECLARE_SCHEDULE_DISPATCHER_CLASS(TVMOpRule) -// One dispatcher is based Ort NodeArg name -DECLARE_SCHEDULE_DISPATCHER_CLASS(OrtOpName) - -// Scheduler Registry is a registry holds all Schedulers -using TVMScheduleRegistry = codegen::RegistryBase; - -// Macro declares TVM scheduler class -#define DECLARE_TVM_SCHEDULER_CLASS(OP, PRETFIX) \ - DECLARE_CREATOR_CLASS(OP, PRETFIX##Scheduler, \ - const tvm::Tensor&, \ - const Node*, \ - tvm_codegen::CodeGenContext&, \ - tvm_codegen::ScheduleContext&, \ - bool) - -// Macro returns TVM scheduler's name with prefix -#define TVM_SCHEDULER_CLASS(OP, PREFIX) \ - CREATOR_CLASS(OP, PREFIX##Scheduler) - -// Macro returns TVM scheduler's name as string -#define TVM_SCHEDULER_STRING(OP, PREFIX) \ - STRINGIZE(TVM_SCHEDULER_CLASS(OP, PREFIX)) - -// Macro returns TVM scheduler's name with prefix and arch -#define TVM_SCHEDULER_CLASS_EX(OP, PREFIX, ARCH) \ - CREATOR_CLASS(OP, PREFIX##ARCH##Scheduler) - -// Macro declares TVM scheduler class with prefix and arch -#define DECLARE_TVM_SCHEDULER_CLASS_EX(OP, PREFIX, ARCH) \ - DECLARE_TVM_SCHEDULER_CLASS(OP, PREFIX##ARCH) - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/utils/codegen_context.cc b/onnxruntime/core/codegen/passes/utils/codegen_context.cc deleted file mode 100644 index 2f1a59b4a92eb..0000000000000 --- a/onnxruntime/core/codegen/passes/utils/codegen_context.cc +++ /dev/null @@ -1,27 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/utils/codegen_context.h" - -#include "core/codegen/common/common.h" - -namespace onnxruntime { -namespace tvm_codegen { - -CodeGenContext::CodeGenContext( - const codegen::CodeGenHandle* handle) - : handle_(handle), unname_symbol_counter_(0) {} - -tvm::Var CodeGenContext::GetOrCreateDynamicDim(const std::string& name) { - if (dynamic_dims_.count(name) == 0) - dynamic_dims_.emplace(name, tvm::Var(name)); - - return dynamic_dims_.at(name); -} - -std::string CodeGenContext::CreateUnnamedSymbol() { - return "unnamed_" + std::to_string(unname_symbol_counter_++); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/utils/codegen_context.h b/onnxruntime/core/codegen/passes/utils/codegen_context.h deleted file mode 100644 index 641552bd3b2e8..0000000000000 --- a/onnxruntime/core/codegen/passes/utils/codegen_context.h +++ /dev/null @@ -1,44 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/common/handle.h" -#include "core/codegen/common/common.h" -#include "core/common/common.h" -#include "core/framework/data_types.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// CodeGenContext is a data structure involving across passes -// Compiler developers can use it to store meta data -// to support fine-grained control of code generation -class CodeGenContext { - public: - CodeGenContext(const codegen::CodeGenHandle* handle); - - virtual ~CodeGenContext() = default; - - // returns tvm::Var for the dynamic dim - tvm::Var GetOrCreateDynamicDim(const std::string& name); - - const codegen::CodeGenHandle* GetCodeGenHandle() const { - return handle_; - } - - std::string CreateUnnamedSymbol(); - - protected: - std::unordered_map dynamic_dims_; - - const codegen::CodeGenHandle* handle_; - - int unname_symbol_counter_; -}; - -// Add Promote for CodeGenContext -DYNAMIC_PROMOTE(CodeGenContext) - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/utils/ort_tvm_utils.cc b/onnxruntime/core/codegen/passes/utils/ort_tvm_utils.cc deleted file mode 100644 index 55892974aa33f..0000000000000 --- a/onnxruntime/core/codegen/passes/utils/ort_tvm_utils.cc +++ /dev/null @@ -1,194 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/utils/ort_tvm_utils.h" - -#include "core/codegen/common/profile.h" -#include "core/codegen/passes/utils/codegen_context.h" -#include "core/framework/tensorprotoutils.h" -#include "core/providers/common.h" -#include - -#include - -namespace onnxruntime { -namespace tvm_codegen { - -#define RETURN_DLDATATYPE_IF_MATCH(type_enum, type, type_code) \ - case type_enum: \ - return {type_code, sizeof(type) * 8, 1}; \ - break; - -// DLDataType: {DLDataTypeCode, bits, lanes} -DLDataType ToTvmDLDataType(MLDataType ml_type) { - if (ml_type->IsTensorType()) { - ml_type = ml_type->AsTensorType()->GetElementType(); - } - auto prim_type = ml_type->AsPrimitiveDataType(); - if (prim_type == nullptr) { - ORT_NOT_IMPLEMENTED("converting MLDataType ", ml_type, " to tvm DLDataType is not implemented"); - } - - switch (prim_type->GetDataType()) { - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_INT8, int8_t, kDLInt); - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_UINT8, uint8_t, kDLUInt); - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_INT16, int16_t, kDLInt); - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_UINT16, uint16_t, kDLUInt); - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_INT32, int32_t, kDLInt); - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_UINT32, uint32_t, kDLUInt); - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_INT64, int64_t, kDLInt); - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_UINT64, uint64_t, kDLUInt); - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_BOOL, bool, kDLUInt); - - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_FLOAT, float, kDLFloat); - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_DOUBLE, double, kDLFloat); - RETURN_DLDATATYPE_IF_MATCH(ONNX_NAMESPACE::TensorProto_DataType_FLOAT16, MLFloat16, kDLFloat); - default: - ORT_NOT_IMPLEMENTED("converting MLDataType ", ml_type, " to tvm DLDataType is not implemented"); - } -} - -tvm::Type ToTvmType(ONNX_NAMESPACE::TensorProto_DataType proto_type) { - switch (proto_type) { - // Note that bool is uint1 in tvm, but uint8 in ONNX, so it always require special handling - // case ONNX_NAMESPACE::TensorProto_DataType_BOOL: - // return tvm::UInt(1); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_INT16: - return tvm::Int(16); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_INT32: - return tvm::Int(32); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_INT64: - return tvm::Int(64); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_UINT8: - return tvm::UInt(8); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_UINT16: - return tvm::UInt(16); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_UINT32: - return tvm::UInt(32); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_UINT64: - return tvm::UInt(64); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_FLOAT: - return tvm::Float(32); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_DOUBLE: - return tvm::Float(64); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_INT8: - return tvm::Int(8); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_FLOAT16: - return tvm::Float(16); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_STRING: - ORT_THROW("Casting to and from strings is not supported yet."); /*break;*/ - case ONNX_NAMESPACE::TensorProto_DataType_UNDEFINED: - ORT_THROW("Cast op must have 'to' argument of type DataType"); /*break;*/ - default: - ORT_THROW("Unexpected 'to' argument value: ", proto_type); - } -} - -tvm::Array ShapeToTvmArray(const NodeArg* def, CodeGenContext& ctx) { - ORT_ENFORCE(nullptr != def); - const ONNX_NAMESPACE::TensorShapeProto* shape_proto = def->Shape(); - ORT_ENFORCE(nullptr != shape_proto); - - tvm::Array arr; - for (int i = 0; i < shape_proto->dim_size(); ++i) { - arr.push_back(ShapeDimToTvmDim(shape_proto->dim(i), ctx)); - } - return arr; -} - -tvm::Expr ShapeDimToTvmDim(const ONNX_NAMESPACE::TensorShapeProto_Dimension& dim, CodeGenContext& ctx) { - if (utils::HasDimParam(dim)) { - return ctx.GetOrCreateDynamicDim(dim.dim_param()); - } else if (utils::HasDimValue(dim)) { - return tvm::Expr(gsl::narrow_cast(dim.dim_value())); - } - return ctx.GetOrCreateDynamicDim(ctx.CreateUnnamedSymbol()); -} - -#ifdef CODEGEN_ENABLE_PROFILER -struct event_in_bracket_and_id { - bool in_bracket; - size_t id; -}; -std::unordered_map g_codegen_profiler_event_ids; -std::vector> g_codegen_profiler_events(1024); - -TVM_REGISTER_GLOBAL("tvm.contrib.onnxruntime.profile_event") - .set_body([](tvm::TVMArgs args, tvm::TVMRetValue* ret) { - DLTensor* X = args[0]; - DLTensor* Y = args[1]; - size_t event_id = args[2]; - bool is_begin = args[3]; - if (!is_begin) { - DCHECK(event_id < g_codegen_profiler_event_ids.size()); - profiling::Profiler::Instance().EndTimeAndRecordEvent( - profiling::EventCategory::NODE_EVENT, - g_codegen_profiler_events[event_id].first, - g_codegen_profiler_events[event_id].second); - } - - { - CODEGEN_PROFILER_EVENT("profile_stub"); - int64_t elem_count = 1; - for (int i = 0; i < X->ndim; ++i) { - elem_count *= X->shape[i]; - } - // there's overhead in this copy, so put begin after copy and end before copy - memcpy(static_cast(Y->data) + Y->byte_offset, - static_cast(X->data) + X->byte_offset, - elem_count * X->dtype.bits / 8); - } - - if (is_begin) { - DCHECK(g_codegen_profiler_events.size() > event_id); - DCHECK(!g_codegen_profiler_events[event_id].first.empty()); - DCHECK(g_codegen_profiler_event_ids[g_codegen_profiler_events[event_id].first].id == event_id); - g_codegen_profiler_events[event_id].second = - profiling::Profiler::Instance().StartTime(); - } - }); - -tvm::Tensor ProfileBegin(tvm::Tensor X, const std::string& event_name) { - size_t event_id; - if (g_codegen_profiler_event_ids.count(event_name) == 0) { - event_id = g_codegen_profiler_event_ids.size(); - ORT_ENFORCE(event_id < g_codegen_profiler_events.size()); - } else { - ORT_ENFORCE(!g_codegen_profiler_event_ids[event_name].in_bracket); - event_id = g_codegen_profiler_event_ids[event_name].id; - } - g_codegen_profiler_event_ids[event_name] = {true, event_id}; - g_codegen_profiler_events[event_id].first = event_name; - return topi::detail::make_extern( - {X->shape}, {X->dtype}, {X}, - [&](tvm::Array ins, tvm::Array outs) { - return topi::detail::call_packed({tvm::Expr("tvm.contrib.onnxruntime.profile_event"), - topi::detail::pack_buffer(ins[0]), - topi::detail::pack_buffer(outs[0]), - gsl::narrow(event_id), - true}); - }, - event_name + "_begin", "", {})[0]; -} - -tvm::Tensor ProfileEnd(tvm::Tensor X, const std::string& event_name) { - ORT_ENFORCE(g_codegen_profiler_event_ids.at(event_name).in_bracket); - g_codegen_profiler_event_ids.at(event_name).in_bracket = false; - size_t event_id = g_codegen_profiler_event_ids.at(event_name).id; - ORT_ENFORCE(event_id < g_codegen_profiler_events.size()); - ORT_ENFORCE(g_codegen_profiler_events[event_id].first == event_name); - return topi::detail::make_extern( - {X->shape}, {X->dtype}, {X}, - [&](tvm::Array ins, tvm::Array outs) { - return topi::detail::call_packed({tvm::Expr("tvm.contrib.onnxruntime.profile_event"), - topi::detail::pack_buffer(ins[0]), - topi::detail::pack_buffer(outs[0]), - gsl::narrow(event_id), - false}); - }, - event_name + "_end", "", {})[0]; -} -#endif - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/utils/ort_tvm_utils.h b/onnxruntime/core/codegen/passes/utils/ort_tvm_utils.h deleted file mode 100644 index f13e91a2d5cea..0000000000000 --- a/onnxruntime/core/codegen/passes/utils/ort_tvm_utils.h +++ /dev/null @@ -1,31 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/common/common.h" -#include "core/framework/data_types.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -class CodeGenContext; - -// Helper function that converts a onnxruntime MLDataType to TVM DLDataType -DLDataType ToTvmDLDataType(MLDataType ml_type); - -tvm::Type ToTvmType(ONNX_NAMESPACE::TensorProto_DataType proto_type); - -tvm::Array ShapeToTvmArray(const NodeArg* def, CodeGenContext& ctx); - -tvm::Expr ShapeDimToTvmDim(const ONNX_NAMESPACE::TensorShapeProto_Dimension& dim, CodeGenContext& ctx); - -#ifdef CODEGEN_ENABLE_PROFILER -// Helper functions to inspect into lowered function -tvm::Tensor ProfileBegin(tvm::Tensor X, const std::string& event_name); - -tvm::Tensor ProfileEnd(tvm::Tensor X, const std::string& event_name); -#endif - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/weight_layout/tiling_2d.cc b/onnxruntime/core/codegen/passes/weight_layout/tiling_2d.cc deleted file mode 100644 index c65132f6d4bca..0000000000000 --- a/onnxruntime/core/codegen/passes/weight_layout/tiling_2d.cc +++ /dev/null @@ -1,105 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/weight_layout/tiling_2d.h" - -#include "core/codegen/passes/utils/codegen_context.h" - -namespace onnxruntime { -namespace tvm_codegen { - -constexpr auto local_name_prefix = "tiling_2d_"; -constexpr int num_bits = 8; - -const std::string WeightLayoutTiling2D::GetKey( - ONNX_NAMESPACE::TensorProto_DataType proto_type, - int vector_width) { - return WeightLayout::GetKey( - local_name_prefix + std::to_string(vector_width), - proto_type, 2, 0.0f); -} - -WeightLayoutTiling2D::WeightLayoutTiling2D( - ONNX_NAMESPACE::TensorProto_DataType proto_type, - int vector_width) - : WeightLayout( - local_name_prefix + std::to_string(vector_width), - proto_type, 2, 0.0f), - vector_width_(vector_width) {} - -CoordTransFunc WeightLayoutTiling2D::ToActual(const tvm::Tensor& /*X*/) const { - return [&](const tvm::Array& nominal_coord) { - ORT_ENFORCE(nominal_coord.size() == 2); - const auto& y = nominal_coord[0]; - const auto& x = nominal_coord[1]; - return tvm::Array{ - x, - y}; - }; -} - -CoordTransFunc WeightLayoutTiling2D::ToNominal(const tvm::Tensor& X) const { - return [&](const tvm::Array& actual_coord) { - ORT_ENFORCE(actual_coord.size() == 2); - ORT_ENFORCE(X->dtype == HalideIR::type_of() || - X->dtype == HalideIR::type_of()); - - int tile_row = (sizeof(int32_t) * num_bits) / X->dtype.bits(); - int tile_col = ((vector_width_ * num_bits) / X->dtype.bits()) / tile_row; - - const auto& x = actual_coord[0]; - const auto& y = actual_coord[1]; - - const int block_dimy = tile_row; - const int block_dimx = tile_col; - - const auto& y0 = y % block_dimy; - const auto& y1 = (y / block_dimy) % block_dimx; - const auto& y2 = y / block_dimy / block_dimx; - - const auto& x0 = x % block_dimx; - const auto& x1 = x / block_dimx; - - return tvm::Array{ - y0 + y2 * block_dimx * block_dimy + x0 * block_dimy, - y1 + x1 * block_dimx}; - }; -} - -tvm::Array WeightLayoutTiling2D::ToActualShape(const tvm::Tensor& X) const { - ORT_ENFORCE(X->dtype == HalideIR::type_of() || - X->dtype == HalideIR::type_of()); - - auto pad_row = tvm::make_const(tvm::Int(32), (vector_width_ * num_bits) / X->dtype.bits()); - auto pad_col = tvm::make_const(tvm::Int(32), vector_width_ / sizeof(int32_t)); - - auto new_shape0 = ((X->shape[1] + pad_col - 1) / pad_col) * pad_col; - auto new_shape1 = ((X->shape[0] + pad_row - 1) / pad_row) * pad_row; - - tvm::Array - new_shape = { - new_shape0, - new_shape1}; - return new_shape; -} - -std::vector WeightLayoutTiling2D::ToActualShape(const Tensor* X) const { - ORT_ENFORCE(X != nullptr); - ORT_ENFORCE(X->Shape().GetDims().size() == 2); - - int pad_row = vector_width_ / X->DataType()->Size(); - int pad_col = vector_width_ / sizeof(int32_t); - - auto old_shape = X->Shape().GetDims(); - auto new_shape0 = (old_shape[1] + pad_col - 1) / pad_col * pad_col; - auto new_shape1 = ((old_shape[0] + pad_row - 1) / pad_row) * pad_row; - - std::vector new_shape = { - new_shape0, - new_shape1}; - - return new_shape; -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/weight_layout/tiling_2d.h b/onnxruntime/core/codegen/passes/weight_layout/tiling_2d.h deleted file mode 100644 index 64334a069f94f..0000000000000 --- a/onnxruntime/core/codegen/passes/weight_layout/tiling_2d.h +++ /dev/null @@ -1,43 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once - -#include "core/codegen/passes/weight_layout/weight_layout.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -/* - * \class! WeightLayoutTiling2D - * \breif! Transform 2D weight to 4D by tiling both dimension, - * this layout is used for tensorization. - * [M, N] => [M/Tx, N/Ty, Tx, Ty] - */ - -class WeightLayoutTiling2D : public WeightLayout { - public: - static const std::string GetKey(ONNX_NAMESPACE::TensorProto_DataType proto_type, - int vector_width); - - public: - WeightLayoutTiling2D(ONNX_NAMESPACE::TensorProto_DataType proto_type, - int vector_width); - - ~WeightLayoutTiling2D() = default; - - CoordTransFunc ToNominal(const tvm::Tensor& X) const override; - CoordTransFunc ToActual(const tvm::Tensor& X) const override; - tvm::Array ToActualShape(const tvm::Tensor& X) const override; - std::vector ToActualShape(const Tensor* X) const override; - - private: - int vector_width_; - - private: - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(WeightLayoutTiling2D); -}; - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/weight_layout/transpose_2d.cc b/onnxruntime/core/codegen/passes/weight_layout/transpose_2d.cc deleted file mode 100644 index ea8597f7dd89d..0000000000000 --- a/onnxruntime/core/codegen/passes/weight_layout/transpose_2d.cc +++ /dev/null @@ -1,64 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/weight_layout/transpose_2d.h" - -#include "core/codegen/passes/utils/codegen_context.h" - -namespace onnxruntime { -namespace tvm_codegen { - -constexpr auto local_layout_name = "transpose_2d"; - -const std::string WeightLayoutTranspose2D::GetKey( - ONNX_NAMESPACE::TensorProto_DataType proto_type) { - return WeightLayout::GetKey(local_layout_name, proto_type, 2, 0.0f); -} - -WeightLayoutTranspose2D::WeightLayoutTranspose2D( - ONNX_NAMESPACE::TensorProto_DataType proto_type) - : WeightLayout(local_layout_name, proto_type, 2, 0.0f) {} - -CoordTransFunc WeightLayoutTranspose2D::ToActual(const tvm::Tensor& /*X*/) const { - return [&](const tvm::Array& nominal_coord) { - ORT_ENFORCE(nominal_coord.size() == 2); - const auto& y = nominal_coord[0]; - const auto& x = nominal_coord[1]; - return tvm::Array{ - x, - y}; - }; -} - -CoordTransFunc WeightLayoutTranspose2D::ToNominal(const tvm::Tensor& /*X*/) const { - return [&](const tvm::Array& actual_coord) { - ORT_ENFORCE(actual_coord.size() == 2); - const auto& y = actual_coord[0]; - const auto& x = actual_coord[1]; - return tvm::Array{ - x, - y}; - }; -} - -tvm::Array WeightLayoutTranspose2D::ToActualShape(const tvm::Tensor& X) const { - tvm::Array new_shape = { - X->shape[1], - X->shape[0]}; - return new_shape; -} - -std::vector WeightLayoutTranspose2D::ToActualShape(const Tensor* X) const { - ORT_ENFORCE(X != nullptr); - ORT_ENFORCE(X->Shape().GetDims().size() == 2); - auto old_shape = X->Shape().GetDims(); - - std::vector new_shape = { - old_shape[1], - old_shape[0]}; - - return new_shape; -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/weight_layout/transpose_2d.h b/onnxruntime/core/codegen/passes/weight_layout/transpose_2d.h deleted file mode 100644 index 65babaaec8dac..0000000000000 --- a/onnxruntime/core/codegen/passes/weight_layout/transpose_2d.h +++ /dev/null @@ -1,33 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once - -#include "core/codegen/passes/weight_layout/weight_layout.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// WeightLayoutTranspose2D for transposing a 2D weight -// [W, H] => [H, W] -class WeightLayoutTranspose2D : public WeightLayout { - public: - static const std::string GetKey(ONNX_NAMESPACE::TensorProto_DataType proto_type); - - public: - WeightLayoutTranspose2D(ONNX_NAMESPACE::TensorProto_DataType proto_type); - - ~WeightLayoutTranspose2D() = default; - - CoordTransFunc ToNominal(const tvm::Tensor& X) const override; - CoordTransFunc ToActual(const tvm::Tensor& X) const override; - tvm::Array ToActualShape(const tvm::Tensor& X) const override; - std::vector ToActualShape(const Tensor* X) const override; - - private: - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(WeightLayoutTranspose2D); -}; - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/weight_layout/vertical_stripes_2d.cc b/onnxruntime/core/codegen/passes/weight_layout/vertical_stripes_2d.cc deleted file mode 100644 index b1ddb791a3b3d..0000000000000 --- a/onnxruntime/core/codegen/passes/weight_layout/vertical_stripes_2d.cc +++ /dev/null @@ -1,77 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/weight_layout/vertical_stripes_2d.h" - -#include "core/codegen/passes/utils/codegen_context.h" - -namespace onnxruntime { -namespace tvm_codegen { - -constexpr auto local_name_prefix = "vertical_stripe_2d_"; - -const std::string WeightLayoutVerticalStripe2D::GetKey( - ONNX_NAMESPACE::TensorProto_DataType proto_type, - int stripe_width) { - return WeightLayout::GetKey( - local_name_prefix + std::to_string(stripe_width), - proto_type, 2, 0.0f); -} - -WeightLayoutVerticalStripe2D::WeightLayoutVerticalStripe2D( - ONNX_NAMESPACE::TensorProto_DataType proto_type, - int stripe_width) - : WeightLayout( - local_name_prefix + std::to_string(stripe_width), - proto_type, 2, 0.0f), - stripe_width_(stripe_width) { -} - -CoordTransFunc WeightLayoutVerticalStripe2D::ToActual(const tvm::Tensor& /*X*/) const { - return [&](const tvm::Array& nominal_coord) { - ORT_ENFORCE(nominal_coord.size() == 2); - const auto& y = nominal_coord[0]; - const auto& x = nominal_coord[1]; - return tvm::Array{ - x / stripe_width_, - y, - x % stripe_width_}; - }; -} - -CoordTransFunc WeightLayoutVerticalStripe2D::ToNominal(const tvm::Tensor& /*X*/) const { - return [&](const tvm::Array& actual_coord) { - ORT_ENFORCE(actual_coord.size() == 3); - const auto& z = actual_coord[0]; - const auto& y = actual_coord[1]; - const auto& x = actual_coord[2]; - return tvm::Array{ - y, - x + stripe_width_ * z}; - }; -} - -tvm::Array WeightLayoutVerticalStripe2D::ToActualShape(const tvm::Tensor& X) const { - tvm::Array new_shape = { - (X->shape[1] + stripe_width_ - 1) / stripe_width_, - X->shape[0], - stripe_width_}; - return new_shape; -} - -std::vector WeightLayoutVerticalStripe2D::ToActualShape(const Tensor* X) const { - ORT_ENFORCE(X != nullptr); - auto old_shape = X->Shape().GetDims(); - - ORT_ENFORCE(old_shape.size() == 2); - - std::vector new_shape = { - (old_shape[1] + stripe_width_ - 1) / stripe_width_, - old_shape[0], - stripe_width_}; - - return new_shape; -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/weight_layout/vertical_stripes_2d.h b/onnxruntime/core/codegen/passes/weight_layout/vertical_stripes_2d.h deleted file mode 100644 index b9b65025dc014..0000000000000 --- a/onnxruntime/core/codegen/passes/weight_layout/vertical_stripes_2d.h +++ /dev/null @@ -1,40 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/common/common.h" -#include "core/codegen/passes/weight_layout/weight_layout.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -// WeightLayoutVerticalStripe2D for making a 2D weight to 3D, by tiling the lowest (verteical) dimension -// [W, H] => [H/stripe, W, stripe] -class WeightLayoutVerticalStripe2D : public WeightLayout { - public: - static const std::string GetKey( - ONNX_NAMESPACE::TensorProto_DataType proto_type, - int stripe_width); - - public: - WeightLayoutVerticalStripe2D( - ONNX_NAMESPACE::TensorProto_DataType proto_type, - int stripe_width); - - ~WeightLayoutVerticalStripe2D() = default; - - virtual CoordTransFunc ToNominal(const tvm::Tensor& X) const override; - virtual CoordTransFunc ToActual(const tvm::Tensor& X) const override; - tvm::Array ToActualShape(const tvm::Tensor& X) const override; - std::vector ToActualShape(const Tensor* X) const override; - - private: - int stripe_width_; - - private: - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(WeightLayoutVerticalStripe2D); -}; - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/weight_layout/weight_layout.cc b/onnxruntime/core/codegen/passes/weight_layout/weight_layout.cc deleted file mode 100644 index ab3e647fd284a..0000000000000 --- a/onnxruntime/core/codegen/passes/weight_layout/weight_layout.cc +++ /dev/null @@ -1,91 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/codegen/passes/weight_layout/weight_layout.h" - -#include "core/codegen/common/common.h" -#include "core/codegen/common/utils.h" -#include "core/codegen/mti/mti_tvm_utils.h" -#include "core/codegen/passes/utils/ort_tvm_utils.h" - -namespace onnxruntime { -namespace tvm_codegen { - -static tvm::Tensor CreateTVMPlaceholder( - const std::string& name, - HalideIR::Type type, - int dim) { - tvm::Array shape; - if (dim > 0) { - for (int i = 0; i < dim; ++i) { - shape.push_back(tvm::Var(name + "_v" + std::to_string(i))); - } - } else { - shape.push_back(1); - } - return tvm::placeholder(shape, type, name + "_placeholder"); -} - -const std::string WeightLayout::GetKey( - const std::string& name, - ONNX_NAMESPACE::TensorProto_DataType proto_type, - int input_dim, - float pad_zero) { - std::ostringstream key; - key << name << "_type_" << static_cast(proto_type); - key << "_dim_" << input_dim; - key << "_pad_zero_" << pad_zero; - return NormalizeCppName(key.str()); -} - -WeightLayout::WeightLayout( - const std::string& name, - ONNX_NAMESPACE::TensorProto_DataType proto_type, - int input_dim, - float pad_zero) - : name_(GetKey(name, proto_type, input_dim, pad_zero)), - proto_type_(proto_type), - input_dim_(input_dim), - pad_zero_(pad_zero) {} - -const std::string& WeightLayout::Name() const { - return name_; -} - -void WeightLayout::CreateLayoutMarshallingTVMOp(tvm::Array& inputs, - tvm::Array& outputs) const { - HalideIR::Type halide_type = ToTvmType(proto_type_); - - tvm::Tensor placeholder = CreateTVMPlaceholder(name_, halide_type, input_dim_); - inputs.push_back(placeholder); - - tvm::Array new_shape = ToActualShape(placeholder); - CoordTransFunc new_coord_to_old_coord_func = ToNominal(placeholder); - tvm::Expr pad_zero_expr = tvm::make_const(halide_type, pad_zero_); - - tvm::Tensor output = tvm::compute( - new_shape, - [&](const tvm::Array& output_coord) { - tvm::Array output_coord1; - for (const auto& coord : output_coord) - output_coord1.push_back(coord); - auto input_coord = new_coord_to_old_coord_func(output_coord1); - ORT_ENFORCE(input_coord.size() == placeholder->shape.size()); - - if (input_coord.size() > 0) { - auto in_range = (input_coord[0] >= 0) && (input_coord[0] < placeholder->shape[0]); - for (size_t dim = 1; dim < input_coord.size(); ++dim) - in_range = in_range && (input_coord[dim] >= 0) && (input_coord[dim] < placeholder->shape[dim]); - - return tvm::if_then_else(in_range, placeholder(input_coord), pad_zero_expr); - } else { - // scalar - return placeholder(input_coord); - } - }); - - outputs.push_back(output); -} - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/codegen/passes/weight_layout/weight_layout.h b/onnxruntime/core/codegen/passes/weight_layout/weight_layout.h deleted file mode 100644 index 1b45a38e7e24e..0000000000000 --- a/onnxruntime/core/codegen/passes/weight_layout/weight_layout.h +++ /dev/null @@ -1,68 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once -#include "core/codegen/common/common.h" -#include "core/codegen/common/registry.h" -#include "core/common/common.h" -#include "core/framework/tensor.h" -#include - -namespace onnxruntime { -namespace tvm_codegen { - -using CoordTransFunc = std::function(const tvm::Array&)>; - -// WeightLayout is data layout transformer for weight/initializer -class WeightLayout { - public: - // Static function to return unique string as a key - static const std::string GetKey( - const std::string& name, - ONNX_NAMESPACE::TensorProto_DataType proto_type, - int input_dim, - float pad_zero); - - public: - WeightLayout( - const std::string& name, - ONNX_NAMESPACE::TensorProto_DataType proto_type, - int input_dim, - float pad_zero); - - virtual ~WeightLayout() = default; - - // Return a CoordTransFunc from actual (transformed) coordinate to normial (original) coordinate - virtual CoordTransFunc ToNominal(const tvm::Tensor& X) const = 0; - - // Return a CoordTransFunc from normial (original) coordinate to actual (transformed) coordinate - virtual CoordTransFunc ToActual(const tvm::Tensor& X) const = 0; - - // Return actual (transformed) shape in tvm::Array (tvm_codegen) - virtual tvm::Array ToActualShape(const tvm::Tensor& X) const = 0; - - // Return actual (transformed) shape in vector (ort) - virtual std::vector ToActualShape(const Tensor* X) const = 0; - - // Create Layout Marshalling op in outputs - void CreateLayoutMarshallingTVMOp(tvm::Array& inputs, - tvm::Array& outputs) const; - - // Layout name - const std::string& Name() const; - - protected: - std::string name_; - ONNX_NAMESPACE::TensorProto_DataType proto_type_; - int input_dim_; - float pad_zero_; - - private: - ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(WeightLayout); -}; - -// Weight Layout Registry is a registry holds all WeightLayout -using WeightLayoutRegistry = codegen::RegistryBase; - -} // namespace tvm_codegen -} // namespace onnxruntime diff --git a/onnxruntime/core/framework/utils.cc b/onnxruntime/core/framework/utils.cc index 9eed0249711f9..ff664c2c76703 100644 --- a/onnxruntime/core/framework/utils.cc +++ b/onnxruntime/core/framework/utils.cc @@ -57,7 +57,6 @@ void DestroyStrings(void* p_data, int64_t elements) { bool ProviderIsCpuBased(const std::string& provider_type) { return provider_type == onnxruntime::kCpuExecutionProvider || provider_type == onnxruntime::kDnnlExecutionProvider || - provider_type == onnxruntime::kTvmExecutionProvider || provider_type == onnxruntime::kVitisAIExecutionProvider || provider_type == onnxruntime::kOpenVINOExecutionProvider || provider_type == onnxruntime::kNnapiExecutionProvider || diff --git a/onnxruntime/core/platform/windows/stacktrace.cc b/onnxruntime/core/platform/windows/stacktrace.cc index 3401507ae911f..cc23d70c0f11f 100644 --- a/onnxruntime/core/platform/windows/stacktrace.cc +++ b/onnxruntime/core/platform/windows/stacktrace.cc @@ -30,7 +30,6 @@ class CaptureStackTrace { // Get the stack trace. Currently only enabled for a DEBUG build as we require the DbgHelp library. std::vector GetStackTrace() { #ifndef NDEBUG -// TVM need to run with shared CRT, so won't work with debug helper now #if (defined __cpp_lib_stacktrace) && !(defined _OPSCHEMA_LIB_) && !(defined _GAMING_XBOX) && !(defined ONNXRUNTIME_ENABLE_MEMLEAK_CHECK) return detail::CaptureStackTrace().Trace(); #else diff --git a/onnxruntime/core/providers/get_execution_providers.cc b/onnxruntime/core/providers/get_execution_providers.cc index d2a72c3a38b03..7d8c5525667b9 100644 --- a/onnxruntime/core/providers/get_execution_providers.cc +++ b/onnxruntime/core/providers/get_execution_providers.cc @@ -66,14 +66,6 @@ constexpr ProviderInfo kProvidersInPriorityOrder[] = true, #else false, -#endif - }, - { - kTvmExecutionProvider, -#ifdef USE_TVM - true, -#else - false, #endif }, { diff --git a/onnxruntime/core/providers/provider_factory_creators.h b/onnxruntime/core/providers/provider_factory_creators.h index 41e418d9eb97f..1c62c1a7a8d0b 100644 --- a/onnxruntime/core/providers/provider_factory_creators.h +++ b/onnxruntime/core/providers/provider_factory_creators.h @@ -78,10 +78,6 @@ #include "core/providers/tensorrt/tensorrt_provider_factory_creator.h" #endif -#if defined(USE_TVM) -#include "core/providers/tvm/tvm_provider_factory_creator.h" -#endif - #if defined(USE_VITISAI) #include "core/providers/vitisai/vitisai_provider_factory_creator.h" #endif diff --git a/onnxruntime/core/providers/tvm/custom_logging.cc b/onnxruntime/core/providers/tvm/custom_logging.cc deleted file mode 100644 index 1cabe81f8e87e..0000000000000 --- a/onnxruntime/core/providers/tvm/custom_logging.cc +++ /dev/null @@ -1,52 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. -// -// Enable custom logging - this will cause TVM to use a custom implementation -// of tvm::runtime::detail::LogMessage. We use this to change the absolute -// file path to relative file path. - -#include -#include -#include -#include -#include -#include - -// TODO(agladyshev): Make conditional choice of sep for Windows and UNIX -std::string GetFileName(const std::string& file_path, char sep = '/') { - return {std::next(file_path.begin(), file_path.find_last_of(sep) + 1), - file_path.end()}; -} - -std::string GetTimedLogMessage(const std::string& file, int lineno, const std::string& message) { - std::stringstream sstream; - std::string file_name = GetFileName(file); - std::time_t t = std::time(nullptr); - sstream << "[" -#ifdef _WIN32 -// TODO(vvchernov): use #include instead of and localtime_s() approach for WIN32 -#pragma warning(disable : 4996) // _CRT_SECURE_NO_WARNINGS -#endif - << std::put_time(std::localtime(&t), "%H:%M:%S") -#ifdef _WIN32 -#pragma warning(default : 4996) -#endif - << "][TVM] " - << file_name << ":" << lineno << ": " + message; - return sstream.str(); -} - -namespace tvm { -namespace runtime { -namespace detail { -void LogFatalImpl(const std::string& file, int lineno, const std::string& message) { - throw std::runtime_error(GetTimedLogMessage(file, lineno, message)); -} - -void LogMessageImpl(const std::string& file, int lineno, const std::string& message) { - std::cerr << GetTimedLogMessage(file, lineno, message) << std::endl; -} - -} // namespace detail -} // namespace runtime -} // namespace tvm diff --git a/onnxruntime/core/providers/tvm/hash_alg/hasher.cc b/onnxruntime/core/providers/tvm/hash_alg/hasher.cc deleted file mode 100644 index bb62b41c7aa85..0000000000000 --- a/onnxruntime/core/providers/tvm/hash_alg/hasher.cc +++ /dev/null @@ -1,30 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/common/common.h" - -#include "hasher.h" // NOLINT(build/include_subdir) -#include "hasher_impl.h" // NOLINT(build/include_subdir) - -namespace onnxruntime { -namespace tvm { - -Hasher::Hasher(const std::string& hash_type) { - hasher_ = getHasherImpl(hash_type); -} - -std::string Hasher::hash(const char* src, size_t size) const { - return hasher_->hash(src, size); -} - -std::shared_ptr Hasher::getHasherImpl(const std::string& hash_type) { - if (hash_type == "sha256") { - return std::make_shared(); - } else { - ORT_NOT_IMPLEMENTED("Hasher was not implemented for hash type: ", hash_type); - } - return nullptr; -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/hash_alg/hasher.h b/onnxruntime/core/providers/tvm/hash_alg/hasher.h deleted file mode 100644 index 7b2f50def2e36..0000000000000 --- a/onnxruntime/core/providers/tvm/hash_alg/hasher.h +++ /dev/null @@ -1,32 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef ONNXRUNTIME_CORE_PROVIDERS_TVM_HASH_ALG_HASHER_H_ -#define ONNXRUNTIME_CORE_PROVIDERS_TVM_HASH_ALG_HASHER_H_ - -#include -#include - -namespace onnxruntime { -namespace tvm { -class HasherImpl; - -class Hasher { - public: - Hasher() = delete; - explicit Hasher(const std::string& hash_type); - virtual ~Hasher() = default; - - std::string hash(const char* src, size_t size) const; - - private: - std::shared_ptr getHasherImpl(const std::string& hash_type); - - private: - std::shared_ptr hasher_; -}; - -} // namespace tvm -} // namespace onnxruntime - -#endif // ONNXRUNTIME_CORE_PROVIDERS_TVM_HASH_ALG_HASHER_H_ diff --git a/onnxruntime/core/providers/tvm/hash_alg/hasher_impl.cc b/onnxruntime/core/providers/tvm/hash_alg/hasher_impl.cc deleted file mode 100644 index 20aef66f3046a..0000000000000 --- a/onnxruntime/core/providers/tvm/hash_alg/hasher_impl.cc +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "hasher_impl.h" // NOLINT(build/include_subdir) - -namespace onnxruntime { -namespace tvm { - -std::string HasherSHA256Impl::hash(const char* src, size_t size) const { - return hexdigest(src, size); -} - -void HasherSHA256Impl::digest(const Ipp8u* src, int size, Ipp8u* dst) { - IppStatus status = ippStsNoErr; - const IppsHashMethod* hashMethod = ippsHashMethod_SHA256(); - status = ippsHashMessage_rmf(src, size, dst, hashMethod); - if (ippStsNoErr != status) { - ORT_THROW("Can't get SHA-256..."); - } -} - -std::string HasherSHA256Impl::digest(const char* src, size_t size) { - const int digest_size_byte = IPP_SHA256_DIGEST_BITSIZE / 8; - auto dst = std::unique_ptr(new char[digest_size_byte]); - digest(reinterpret_cast(src), static_cast(size), reinterpret_cast(dst.get())); - return std::string(dst.get(), digest_size_byte); -} - -std::string HasherSHA256Impl::hexdigest(const char* src, size_t size) { - std::string byte_digest = digest(src, size); - std::stringstream ss; - for (char c : byte_digest) { - ss << std::hex << std::setw(2) << std::setfill('0') << (0xff & c); - } - return ss.str(); -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/hash_alg/hasher_impl.h b/onnxruntime/core/providers/tvm/hash_alg/hasher_impl.h deleted file mode 100644 index 6c285dd0c78f3..0000000000000 --- a/onnxruntime/core/providers/tvm/hash_alg/hasher_impl.h +++ /dev/null @@ -1,42 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef ONNXRUNTIME_CORE_PROVIDERS_TVM_HASH_ALG_HASHER_IMPL_H_ -#define ONNXRUNTIME_CORE_PROVIDERS_TVM_HASH_ALG_HASHER_IMPL_H_ - -#include -#include -#include -#include -#include - -#include "core/common/common.h" - -namespace onnxruntime { -namespace tvm { - -class HasherImpl { - public: - HasherImpl() = default; - virtual ~HasherImpl() = default; - - virtual std::string hash(const char* src, size_t size) const = 0; -}; - -class HasherSHA256Impl : public HasherImpl { - public: - HasherSHA256Impl() = default; - virtual ~HasherSHA256Impl() = default; - - std::string hash(const char* src, size_t size) const final; - - private: - static void digest(const Ipp8u* src, int size, Ipp8u* dst); - static std::string digest(const char* src, size_t size); - static std::string hexdigest(const char* src, size_t size); -}; - -} // namespace tvm -} // namespace onnxruntime - -#endif // ONNXRUNTIME_CORE_PROVIDERS_TVM_HASH_ALG_HASHER_IMPL_H_ diff --git a/onnxruntime/core/providers/tvm/symbols.txt b/onnxruntime/core/providers/tvm/symbols.txt deleted file mode 100644 index 8d903acd9ea76..0000000000000 --- a/onnxruntime/core/providers/tvm/symbols.txt +++ /dev/null @@ -1 +0,0 @@ -OrtSessionOptionsAppendExecutionProvider_Tvm diff --git a/onnxruntime/core/providers/tvm/tvm_allocator.cc b/onnxruntime/core/providers/tvm/tvm_allocator.cc deleted file mode 100644 index 4b68f6432e8cc..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_allocator.cc +++ /dev/null @@ -1,29 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include - -#include "tvm_allocator.h" -#include "core/framework/session_state.h" -#include "xpu_data_transfer.h" - -namespace onnxruntime { -namespace tvm { - -void* TVMAllocator::Alloc(size_t size) { - void* p = nullptr; - if (size > 0) { - DLDataType dl_type{kDLInt, 8, 1}; - int err = TVMDeviceAllocDataSpace(ctx, size, ::tvm::runtime::kAllocAlignment, dl_type, reinterpret_cast(&p)); - CHECK_EQ(err, 0); - return p; - } - return p; -} - -void TVMAllocator::Free(void* p) { - TVMDeviceFreeDataSpace(ctx, p); -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/tvm_allocator.h b/onnxruntime/core/providers/tvm/tvm_allocator.h deleted file mode 100644 index f3ba544b8ac46..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_allocator.h +++ /dev/null @@ -1,45 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef TVM_ALLOCATOR -#define TVM_ALLOCATOR - -#include "core/framework/allocator.h" -#include "tvm_common.h" - -namespace onnxruntime { -namespace tvm { - -#define TVM_ALLOC_ALIGN 128 - -class TVMAllocator : public IAllocator { - public: - TVMAllocator() : TVMAllocator(OrtMemoryInfo("TVM", - OrtAllocatorType::OrtDeviceAllocator, - OrtDevice(OrtDevice::CPU, OrtDevice::MemType::DEFAULT, 0), - 0, - OrtMemTypeDefault)) {} - explicit TVMAllocator(const OrtMemoryInfo& info) - : IAllocator(info) { - switch (info.device.Type()) { - case OrtDevice::CPU: - ctx = {kDLCPU, info.device.Id()}; - break; - case OrtDevice::GPU: - ctx = {kDLVulkan, info.device.Id()}; - break; - default: - ORT_NOT_IMPLEMENTED("Unsupported device"); - break; - } - } - - virtual void* Alloc(size_t size) override; - virtual void Free(void* p) override; - DLDevice ctx; -}; - -} // namespace tvm -} // namespace onnxruntime - -#endif // TVM_ALLOCATOR diff --git a/onnxruntime/core/providers/tvm/tvm_api.cc b/onnxruntime/core/providers/tvm/tvm_api.cc deleted file mode 100644 index e9a7d002e77c8..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_api.cc +++ /dev/null @@ -1,303 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifdef _WIN32 -#include -#else -#include // glob(), globfree() -#endif -#include // memset() -#include -#include -#include - -#include -#include -#include - -#include "core/common/common.h" -#include - -#include "tvm_api.h" - -namespace onnxruntime { -namespace tvm { - -using TvmIntArray = ::tvm::Array<::tvm::Integer>; -using TvmPackedFunc = ::tvm::PackedFunc; -namespace tvm_rt = ::tvm::runtime; -namespace tvm_rt_vm = tvm_rt::vm; - -TvmModule TVMCompile(const TvmEPOptions& options, - const std::string& onnx_txt, - const std::string& model_path, - int opset, - const TVMTensorShapes& input_shapes) { - ::tvm::Array shapes; - for (size_t i = 0; i < input_shapes.size(); ++i) { - TvmIntArray shape; - for (auto& dim : input_shapes[i]) { - shape.push_back(::tvm::Integer(dim)); - } - shapes.push_back(shape); - } - - const TvmPackedFunc* compile = tvm_rt::Registry::Get("tvm_onnx_import_and_compile"); - ORT_ENFORCE(compile != nullptr, "Unable to retrieve 'tvm_onnx_import_and_compile'."); - TvmModule mod = (*compile)(TVMByteArray{onnx_txt.data(), onnx_txt.size()}, - model_path, - options.executor, - options.target, - options.target_host, - options.opt_level, - opset, - options.freeze_weights, - shapes, - options.to_nhwc, - options.tuning_file_path, - options.tuning_type); - ORT_ENFORCE(mod.get() != nullptr, "Compiled TVM Module is nullptr!"); - return mod; -} - -std::vector glob(const std::string& dir, const std::string& extension) { - std::vector filenames; -#ifdef _WIN32 - std::string pattern = dir + "/*." + extension; - WIN32_FIND_DATA fd; - HANDLE hFind = ::FindFirstFile(pattern.c_str(), &fd); - if (hFind != INVALID_HANDLE_VALUE) { - do { - if (!(fd.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY)) { - filenames.push_back( - dir + - ToUTF8String(PathString{k_preferred_path_separator}) + - fd.cFileName); - } - } while (::FindNextFile(hFind, &fd)); - ::FindClose(hFind); - } -#else - glob_t glob_result; - memset(&glob_result, 0, sizeof(glob_result)); - - std::string pattern = dir + "/*." + extension; - int return_value = glob(pattern.c_str(), GLOB_TILDE, NULL, &glob_result); - ORT_ENFORCE(return_value == 0, "No results of glob for pattern: " + pattern); - - for (size_t i = 0; i < glob_result.gl_pathc; ++i) { - filenames.push_back(std::string(glob_result.gl_pathv[i])); - } - globfree(&glob_result); -#endif - return filenames; -} - -std::string filter_lib_paths(const std::vector& lib_paths, const std::string& lib_ext) { - std::string lib_path; - size_t counter = 0; - for (const auto& path : lib_paths) { - if (path.find("libtvm_runtime." + lib_ext) != std::string::npos || - path.find("liboctomized_model." + lib_ext) != std::string::npos) { - ++counter; - } else { - lib_path = path; - } - } - ORT_ENFORCE((lib_paths.size() - counter) == 1, "It should be only one shared library for model after filtering"); - - return lib_path; -} - -static std::unordered_map str2dev_type = { - {"llvm", 1}, - {"stackvm", 1}, - {"cpu", 1}, - {"c", 1}, - {"hybrid", 1}, - {"composite", 1}, - {"cuda", 2}, - {"nvptx", 2}, - {"cl", 4}, - {"opencl", 4}, - {"sdaccel", 4}, - {"aocl", 5}, - {"aocl_sw_emu", 5}, - {"vulkan", 7}, - {"metal", 8}, - {"vpi", 9}, - {"rocm", 10}, - {"ext_dev", 12}, - {"hexagon", 14}, - {"webgpu", 15}}; - -TvmModule TVMSoCompile(const TvmEPOptions& options) { - const std::string& dir = options.so_folder; -#ifdef _WIN32 - std::string lib_ext = "dll"; -#else - std::string lib_ext = "so"; -#endif - const std::string lib_path = filter_lib_paths(glob(dir, lib_ext), lib_ext); - const std::string consts_path = dir + - ToUTF8String(PathString{k_preferred_path_separator}) + - "consts"; - const auto& ro_paths = glob(dir, "ro"); - ORT_ENFORCE(ro_paths.size() == 1, "It should be only one ro file in folder: " + dir); - const std::string vm_exec_code_path = ro_paths[0]; - - TvmModule lib = TvmModule::LoadFromFile(lib_path); - - std::ifstream code(vm_exec_code_path, std::ios::binary); - std::stringstream ss; - ss << code.rdbuf(); - - auto exec_mod = tvm_rt_vm::Executable::Load(ss.str(), lib); - const tvm_rt_vm::Executable* tmp = exec_mod.as(); - auto exec = tvm_rt::GetObjectPtr(const_cast(tmp)); - exec->LoadLateBoundConstantsFromFile(consts_path); - - auto vm = tvm_rt::make_object(); - vm->LoadExecutable(exec); - - size_t pos = options.target.find(" "); - const std::string dev_type_str = options.target.substr(0, pos); - ORT_ENFORCE(!dev_type_str.empty(), "Device was not found in target string"); - uint64_t dev_type = str2dev_type[dev_type_str]; - const uint64_t cpu_type = str2dev_type["cpu"]; - // Initialize the VM for the specified device. If the device is not a CPU, - // We'll need to add a CPU context to drive it. - int arity; - if (dev_type == cpu_type) { - arity = 3; - } else { - arity = 6; - } - uint64_t alloc_type = uint64_t(tvm_rt_vm::AllocatorType::kPooled); - // TODO(vchernov): multiple devices using and using device with specified id are not supported - // Always use the first device of the specified type. - uint64_t device_id = 0; - std::vector init_vals(arity); - std::vector codes(arity); - tvm_rt::TVMArgsSetter setter(init_vals.data(), codes.data()); - setter(0, dev_type); - setter(1, device_id); - setter(2, alloc_type); - // Also initialize a CPU device context. - if (dev_type != cpu_type) { - setter(3, cpu_type); - setter(4, device_id); - setter(5, alloc_type); - } - tvm_rt::TVMRetValue rv; - // Call the packed func with the init arguments. - vm->GetFunction("init", nullptr).CallPacked(tvm_rt::TVMArgs(init_vals.data(), codes.data(), arity), &rv); - - return TvmModule(vm); -} - -void TVMSetInputs(TvmModule& mod, - std::vector& inds, - std::vector& inputs) { - TvmPackedFunc set_input = mod.GetFunction("set_input", false); - TvmPackedFunc set_input_zero_copy = mod.GetFunction("set_input_zero_copy", false); - for (size_t i = 0; i < inds.size(); ++i) { - if (reinterpret_cast(inputs[i].data) % tvm_rt::kAllocAlignment == 0) { - set_input_zero_copy(inds[i], &inputs[i]); - } else { - set_input(inds[i], &inputs[i]); - } - } -} - -void TVM_VM_SetInputs(TvmModule& mod, - std::vector& inds, - std::vector& inputs) { - size_t num_total_args = inputs.size() + 1; - std::vector tvm_values(num_total_args); - std::vector tvm_type_codes(num_total_args); - ::tvm::runtime::TVMArgsSetter setter(tvm_values.data(), tvm_type_codes.data()); - const std::string func_name = "main"; - setter(0, func_name.c_str()); - for (size_t k = 0; k < num_total_args - 1; ++k) { - setter(inds[k] + 1, &inputs[k]); - } - - TvmPackedFunc set_input = mod.GetFunction("set_input", false); - ::tvm::runtime::TVMRetValue rv; - set_input.CallPacked(::tvm::runtime::TVMArgs(tvm_values.data(), tvm_type_codes.data(), gsl::narrow_cast(num_total_args)), &rv); -} - -void TVMSetOutputsZeroCopy(TvmModule& mod, - std::vector& outputs) { - TvmPackedFunc set_output = mod.GetFunction("set_output_zero_copy", false); - for (size_t i = 0; i < outputs.size(); ++i) { - set_output(i, &outputs[i]); - } -} - -void TVM_VM_SetOutputsZeroCopy(TvmModule& mod, - std::vector& outputs) { - size_t num_total_args = outputs.size() + 1; - std::vector tvm_values(num_total_args); - std::vector tvm_type_codes(num_total_args); - tvm_rt::TVMArgsSetter setter(tvm_values.data(), tvm_type_codes.data()); - const std::string func_name = "main"; - setter(0, func_name.c_str()); - for (size_t k = 0; k < num_total_args - 1; ++k) { - setter(k + 1, &outputs[k]); - } - - TvmPackedFunc set_output = mod.GetFunction("set_outputs", false); - tvm_rt::TVMRetValue rv; - set_output.CallPacked(tvm_rt::TVMArgs(tvm_values.data(), tvm_type_codes.data(), gsl::narrow_cast(num_total_args)), &rv); -} - -void TVMGetOutputs(TvmModule& mod, - std::vector& outputs) { - TvmPackedFunc get_output = mod.GetFunction("get_output", false); - for (size_t i = 0; i < outputs.size(); ++i) { - get_output(i, &outputs[i]); - } -} - -void TVM_VM_GetOutputs(TvmModule& mod, - std::vector& outputs) { - TvmPackedFunc get_output = mod.GetFunction("get_output", false); - for (size_t i = 0; i < outputs.size(); ++i) { - // TODO(vvchernov): think about improvement of memory management - tvm_rt::NDArray output_array = get_output(i); - output_array.CopyTo(&outputs[i]); - } -} - -void TVMGetOutputShapes(TvmModule& mod, - TVMTensorShapes& output_shapes) { - size_t size = output_shapes.size(); - TvmPackedFunc get_output = mod.GetFunction("get_output", false); - for (size_t i = 0; i < size; ++i) { - tvm_rt::NDArray output_array = get_output(i); - tvm_rt::ShapeTuple shape_tuple = output_array.Shape(); - size_t dims_num = shape_tuple.size(); - TensorShapeVector dims; - for (size_t j = 0; j < dims_num; ++j) { - dims.push_back(int64_t(shape_tuple[j])); - } - output_shapes[i] = dims; - } -} - -void TVMRun(TvmModule& mod) { - TvmPackedFunc run = mod.GetFunction("run", false); - ORT_ENFORCE(run != nullptr, "Unable to retrieve graph executor run."); - run(); -} - -void TVM_VM_Run(TvmModule& mod) { - TvmPackedFunc run = mod.GetFunction("invoke", false); - ORT_ENFORCE(run != nullptr, "Unable to retrieve virtual machine invoke."); - run("main"); -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/tvm_api.h b/onnxruntime/core/providers/tvm/tvm_api.h deleted file mode 100644 index bbf05f4fc06d9..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_api.h +++ /dev/null @@ -1,38 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef TVM_API_H -#define TVM_API_H - -#include -#include - -#include "tvm_common.h" -#include "tvm_defaults.h" -#include "tvm_ep_options.h" - -namespace onnxruntime { -namespace tvm { - -TvmModule TVMCompile(const TvmEPOptions& options, - const std::string& onnx_txt, - const std::string& model_path, - int opset, - const TVMTensorShapes& input_shapes); -TvmModule TVMSoCompile(const TvmEPOptions& options); - -void TVMSetInputs(TvmModule& mod, std::vector& inds, std::vector& inputs); -void TVM_VM_SetInputs(TvmModule& mod, std::vector& inds, std::vector& inputs); -void TVMSetOutputsZeroCopy(TvmModule& mod, std::vector& outputs); -void TVM_VM_SetOutputsZeroCopy(TvmModule& mod, std::vector& outputs); -void TVMGetOutputs(TvmModule& mod, std::vector& outputs); -void TVM_VM_GetOutputs(TvmModule& mod, std::vector& outputs); -void TVMGetOutputShapes(TvmModule& mod, - TVMTensorShapes& output_shapes); -void TVMRun(TvmModule& mod); -void TVM_VM_Run(TvmModule& mod); - -} // namespace tvm -} // namespace onnxruntime - -#endif // TVM_API_H diff --git a/onnxruntime/core/providers/tvm/tvm_common.h b/onnxruntime/core/providers/tvm/tvm_common.h deleted file mode 100644 index 68e3b6496328a..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_common.h +++ /dev/null @@ -1,22 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef TVM_COMMON_H -#define TVM_COMMON_H - -#include -#include - -#include -#include -#include - -namespace onnxruntime { -namespace tvm { - -using TvmModule = ::tvm::runtime::Module; - -} // namespace tvm -} // namespace onnxruntime - -#endif // TVM_COMMON_H diff --git a/onnxruntime/core/providers/tvm/tvm_compiler.cc b/onnxruntime/core/providers/tvm/tvm_compiler.cc deleted file mode 100644 index 8f4e7e7de9a36..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_compiler.cc +++ /dev/null @@ -1,48 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include - -#include "tvm_compiler.h" -#include "tvm_api.h" - -namespace onnxruntime { -namespace tvm { - -auto TVMCompilerBase::operator()(const TvmEPOptions& options, - const TVMTensorShapes& input_shapes) -> ModulePtr { - if (mod_) { - return mod_; - } - - mod_ = std::make_shared(); - this->compileTVMModule(options, input_shapes); - - return mod_; -} - -TVMCompiler::TVMCompiler(std::string&& onnx_model_str, - const std::string& model_path, - int opset) : onnx_model_str_(std::move(onnx_model_str)), - model_path_(model_path), - opset_(opset) { -} - -void TVMCompiler::compileTVMModule(const TvmEPOptions& options, - const TVMTensorShapes& input_shapes) { - *mod_ = tvm::TVMCompile(options, - onnx_model_str_, - model_path_, - opset_, - input_shapes); - - onnx_model_str_.clear(); -} - -void TVMSoCompiler::compileTVMModule(const TvmEPOptions& options, - [[maybe_unused]] const TVMTensorShapes& input_shapes) { - *mod_ = tvm::TVMSoCompile(options); -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/tvm_compiler.h b/onnxruntime/core/providers/tvm/tvm_compiler.h deleted file mode 100644 index bfc73d67aa07f..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_compiler.h +++ /dev/null @@ -1,63 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef TVM_COMPILER_H -#define TVM_COMPILER_H - -#include -#include - -#include "tvm_common.h" -#include "tvm_ep_options.h" - -namespace onnxruntime { -namespace tvm { - -class TVMCompilerBase { - public: - using ModulePtr = std::shared_ptr; - - TVMCompilerBase() = default; - virtual ~TVMCompilerBase() = default; - - ModulePtr operator()(const TvmEPOptions& options, - const TVMTensorShapes& input_shapes); - - virtual void compileTVMModule(const TvmEPOptions& options, - const TVMTensorShapes& input_shapes) = 0; - - protected: - ModulePtr mod_; -}; - -class TVMCompiler : public TVMCompilerBase { - public: - TVMCompiler() = delete; - ~TVMCompiler() = default; - - TVMCompiler(std::string&& onnx_model_str, - const std::string& model_path, - int opset); - - void compileTVMModule(const TvmEPOptions& options, - const TVMTensorShapes& input_shapes) final; - - private: - std::string onnx_model_str_; - std::string model_path_; - int opset_; -}; - -class TVMSoCompiler : public TVMCompilerBase { - public: - TVMSoCompiler() = default; - ~TVMSoCompiler() = default; - - void compileTVMModule(const TvmEPOptions& options, - const TVMTensorShapes& input_shapes) final; -}; - -} // namespace tvm -} // namespace onnxruntime - -#endif // TVM_COMPILER_H diff --git a/onnxruntime/core/providers/tvm/tvm_defaults.h b/onnxruntime/core/providers/tvm/tvm_defaults.h deleted file mode 100644 index 197d1f363c50d..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_defaults.h +++ /dev/null @@ -1,35 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef ONNXRUNTIME_CORE_PROVIDERS_TVM_TVM_DEFAULTS_H_ -#define ONNXRUNTIME_CORE_PROVIDERS_TVM_TVM_DEFAULTS_H_ - -#include - -namespace onnxruntime { -namespace tvm { - -namespace env_vars { -static const std::string kDumpSubgraphs = "ORT_TVM_DUMP_SUBGRAPHS"; -} // namespace env_vars - -constexpr const char* default_executor_type = "vm"; -constexpr const char* vm_executor_type = "vm"; -constexpr const char* graph_executor_type = "graph"; - -constexpr const char* default_target_str = "llvm"; -constexpr const char* llvm_target_str = "llvm"; - -constexpr const char* cpu_target_str = "cpu"; -constexpr const char* gpu_target_str = "gpu"; - -constexpr const char* default_tuning_type = "AutoTVM"; -constexpr const char* autotvm_tuning_type = "AutoTVM"; -constexpr const char* ansor_tuning_type = "Ansor"; - -constexpr const unsigned int default_opt_level = 3; - -} // namespace tvm -} // namespace onnxruntime - -#endif // ONNXRUNTIME_CORE_PROVIDERS_TVM_TVM_DEFAULTS_H_ diff --git a/onnxruntime/core/providers/tvm/tvm_ep_options.cc b/onnxruntime/core/providers/tvm/tvm_ep_options.cc deleted file mode 100644 index 70e99833cd78b..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_ep_options.cc +++ /dev/null @@ -1,273 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include -#include - -#include "core/common/common.h" -#include "core/common/cpuid_info.h" -#include "core/framework/provider_options_utils.h" - -#include "tvm_ep_options.h" - -namespace onnxruntime { -namespace tvm { - -namespace provider_option_names { -constexpr const char* kExecutor = "executor"; -constexpr const char* kSoFolder = "so_folder"; -constexpr const char* kCheckHash = "check_hash"; -constexpr const char* kHashFilePath = "hash_file_path"; -constexpr const char* kTarget = "target"; -constexpr const char* kTargetHost = "target_host"; -constexpr const char* kOptLevel = "opt_level"; -constexpr const char* kFreezeWeights = "freeze_weights"; -constexpr const char* kSetOutputZeroCopy = "set_output_zero_copy"; -constexpr const char* kToNHWC = "to_nhwc"; -constexpr const char* kTuningFilePath = "tuning_file_path"; -constexpr const char* kTuningType = "tuning_type"; -constexpr const char* kInputNames = "input_names"; -constexpr const char* kInputShapes = "input_shapes"; - -static const std::unordered_set valid_keys{ - std::string{kExecutor}, - std::string{kSoFolder}, - std::string{kCheckHash}, - std::string{kHashFilePath}, - std::string{kTarget}, - std::string{kTargetHost}, - std::string{kOptLevel}, - std::string{kFreezeWeights}, - std::string{kSetOutputZeroCopy}, - std::string{kToNHWC}, - std::string{kTuningFilePath}, - std::string{kTuningType}, - std::string{kInputNames}, - std::string{kInputShapes}}; - -} // namespace provider_option_names - -size_t split(const std::string& src, std::vector& dst, char ch) { - dst.clear(); - - size_t pos = src.find(ch); - size_t initialPos = 0; - while (pos != std::string::npos) { - dst.push_back(src.substr(initialPos, pos - initialPos)); - initialPos = pos + 1; - - pos = src.find(ch, initialPos); - } - dst.push_back(src.substr(initialPos, std::min(pos, src.size()) - initialPos + 1)); - - return dst.size(); -} - -TvmEPOptions TvmEPOptionsHelper::FromOptionsString(const char* opt_str) { - std::string settings{opt_str}; - ProviderOptions options; - if (!settings.empty()) { - const std::string& str = settings; - - // tokenize settings - std::regex reg("\\s*,\\s*"); - std::sregex_token_iterator iter(str.begin(), str.end(), reg, -1); - std::sregex_token_iterator iter_end; - std::vector pairs(iter, iter_end); - - ORT_ENFORCE(pairs.size() > 0); - - for (const auto& pair : pairs) { - auto pos_colon = pair.find(':'); - ORT_ENFORCE(pos_colon != std::string::npos, "Invalid key value pair."); - std::string key = pair.substr(0, pos_colon); - std::string value = pair.substr(pos_colon + 1); - - // trim leading and trailing spaces from key/value - key = whitespace_trimming(key); - value = whitespace_trimming(value); - - // Check keys of obtained options - if (tvm::provider_option_names::valid_keys.count(key) == 0) { - ORT_NOT_IMPLEMENTED("TvmOptions: unknown option (", key, ")"); - } - - options[key] = value; - } - } - - return TvmEPOptionsHelper::FromProviderOptions(options); -} - -std::string TvmEPOptionsHelper::whitespace_trimming(const std::string& str) { - const std::string WHITESPACE = " \n\r\t\f\v"; - size_t start = str.find_first_not_of(WHITESPACE); - if (start == std::string::npos) { - return ""; - } else { - size_t end = str.find_last_not_of(WHITESPACE); - ORT_ENFORCE(end != std::string::npos); - return str.substr(start, end + 1); - } -} - -TvmEPOptions TvmEPOptionsHelper::FromProviderOptions(const ProviderOptions& pr_options) { - TvmEPOptions options{}; - - ORT_THROW_IF_ERROR( - ProviderOptionsParser{} - .AddAssignmentToReference(tvm::provider_option_names::kExecutor, options.executor) - .AddAssignmentToReference(tvm::provider_option_names::kSoFolder, options.so_folder) - .AddAssignmentToReference(tvm::provider_option_names::kCheckHash, options.check_hash) - .AddAssignmentToReference(tvm::provider_option_names::kHashFilePath, options.hash_file_path) - .AddAssignmentToReference(tvm::provider_option_names::kTarget, options.target) - .AddAssignmentToReference(tvm::provider_option_names::kTargetHost, options.target_host) - .AddAssignmentToReference(tvm::provider_option_names::kOptLevel, options.opt_level) - .AddAssignmentToReference(tvm::provider_option_names::kFreezeWeights, options.freeze_weights) - .AddAssignmentToReference(tvm::provider_option_names::kSetOutputZeroCopy, options.set_output_zero_copy) - .AddAssignmentToReference(tvm::provider_option_names::kToNHWC, options.to_nhwc) - .AddAssignmentToReference(tvm::provider_option_names::kTuningFilePath, options.tuning_file_path) - .AddAssignmentToReference(tvm::provider_option_names::kTuningType, options.tuning_type) - .AddAssignmentToReference(tvm::provider_option_names::kInputNames, options.input_names_str) - .AddAssignmentToReference(tvm::provider_option_names::kInputShapes, options.input_shapes_str) - .Parse(pr_options)); - - optionsPostprocess(options); - - return options; -} - -void TvmEPOptionsHelper::optionsPostprocess(TvmEPOptions& options) { - setInputShapes(options); - targetPostprocess(options.target); - targetHostPostprocess(options.target, options.target_host); - optLevelPostprocess(options.opt_level); -} - -bool TvmEPOptionsHelper::checkCPUTarget(const std::string& target) { - bool check = target.find("llvm") != std::string::npos; - return check; -} - -bool TvmEPOptionsHelper::checkGPUTarget(const std::string& target) { - bool check = (target.find("cuda") != std::string::npos || - target.find("opencl") != std::string::npos || - target.find("metal") != std::string::npos || - target.find("vulkan") != std::string::npos); - return check; -} - -void TvmEPOptionsHelper::setInputShapes(TvmEPOptions& options) { - if (options.input_names_str.empty() && options.input_shapes_str.empty()) - return; - ORT_ENFORCE(!options.input_names_str.empty() && !options.input_shapes_str.empty(), - "Both provider options \"input_names\" and \"input_shapes\" should be empty or full"); - - std::vector name_set; - std::string trimmed_names = whitespace_trimming(options.input_names_str); - size_t inp_tensors_num = split(trimmed_names, name_set, ' '); - ORT_ENFORCE(inp_tensors_num, "There is no any input tensor names!"); - - std::string trimmed_shapes = whitespace_trimming(options.input_shapes_str); - size_t end_pos = trimmed_shapes.find_last_of(']'); - ORT_ENFORCE(end_pos != std::string::npos, "Invalid string for input shapes. Symbol ] is not found"); - ORT_ENFORCE(end_pos == (trimmed_shapes.size() - 1), - "Invalid string for input shapes. Symbol ] should be last after whitespace trimming"); - - std::vector shape_set; - split(trimmed_shapes, shape_set, ']'); - shape_set.pop_back(); - ORT_ENFORCE(shape_set.size() == inp_tensors_num, - "Number of shapes is not the same as number of input tensor names"); - - for (size_t i = 0; i < inp_tensors_num; ++i) { - size_t pos = shape_set[i].find('['); - ORT_ENFORCE(pos != std::string::npos, "There is no symbol [ as pair for ]"); - std::string numbers = shape_set[i].substr(pos + 1); - std::vector number_set; - ORT_ENFORCE(split(numbers, number_set, ' '), "There is no any number between [ and ] symbols"); - - TensorShapeVector dims; - for (const auto& number : number_set) { - dims.push_back(std::stoi(number)); - } - - options.input_shapes[name_set[i]] = dims; - } -} - -void TvmEPOptionsHelper::targetPostprocess(std::string& target) { - if (target == tvm::cpu_target_str || - target == tvm::llvm_target_str) { - ProcessCPUTarget(target); - } else if (target == tvm::gpu_target_str) { - ProcessGPUTarget(); - } else if (target.empty()) { - ORT_NOT_IMPLEMENTED("target option is empty!"); - } else { - // TODO(vvchernov): extend mechanism of auto-definition of target - // target is gotten from option set up by client - } -} - -void TvmEPOptionsHelper::ProcessCPUTarget(std::string& target) { - const auto& cpu_id_info = CPUIDInfo::GetCPUIDInfo(); - // auto detect from CPU ID - if (cpu_id_info.HasAVX512Skylake()) { - target = tvm::cpu_targets::LLVM_TARGET_SKYLAKE_AVX512; - } else if (cpu_id_info.HasAVX512f()) { - target = tvm::cpu_targets::LLVM_TARGET_AVX512; - } else if (cpu_id_info.HasAVX2()) { - target = tvm::cpu_targets::LLVM_TARGET_AVX2; - } else if (cpu_id_info.HasAVX()) { - target = tvm::cpu_targets::LLVM_TARGET_AVX; - } else { - // TODO(vvchernov): extend mechanism of auto-definition of cpu target - target = tvm::llvm_target_str; - } -} - -void TvmEPOptionsHelper::ProcessGPUTarget() { - ORT_NOT_IMPLEMENTED("GPU target auto-defenition is not implemented now!"); -} - -void TvmEPOptionsHelper::targetHostPostprocess(const std::string& target, std::string& target_host) { - if ((target_host == tvm::cpu_target_str || - target_host == tvm::llvm_target_str) && - target_host != target) { - target_host = target; - } else if (target_host.empty()) { - target_host = target; - } else { - // TODO(vvchernov): extend mechanism of auto-definition of target host - // target host is gotten from option set up by client - } -} - -void TvmEPOptionsHelper::optLevelPostprocess(unsigned int& opt_level) { - if (opt_level < 1) { - opt_level = tvm::default_opt_level; - } -} - -std::ostream& operator<<(std::ostream& out, const TvmEPOptions& options) { - out << "TVM EP options:\n" - << "executor type: " << options.executor << "\n" - << "so_folder: " << options.so_folder << "\n" - << "check_hash: " << options.check_hash << "\n" - << "hash_file_path: " << options.hash_file_path << "\n" - << "target: " << options.target << "\n" - << "target_host: " << options.target_host << "\n" - << "opt level: " << options.opt_level << "\n" - << "freeze weights: " << options.freeze_weights << "\n" - << "set_output_zero_copy: " << options.set_output_zero_copy << "\n" - << "tuning file path: " << options.tuning_file_path << "\n" - << "tuning type: " << options.tuning_type << "\n" - << "convert layout to NHWC: " << options.to_nhwc << "\n" - << "input tensor names: " << options.input_names_str << "\n" - << "input tensor shapes: " << options.input_shapes_str; - return out; -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/tvm_ep_options.h b/onnxruntime/core/providers/tvm/tvm_ep_options.h deleted file mode 100644 index 0f2db30a3b304..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_ep_options.h +++ /dev/null @@ -1,76 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef TVM_EXECUTION_PROVIDER_OPTIONS_H -#define TVM_EXECUTION_PROVIDER_OPTIONS_H - -#include -#include -#include -#include - -#include "core/framework/provider_options.h" -#include "core/framework/tensor_shape.h" - -#include "tvm_defaults.h" - -namespace onnxruntime { - -namespace tvm { -namespace cpu_targets { -// TODO(vvchernov): avx and avx512 need more careful differentiation for target -const std::string LLVM_TARGET_AVX = "llvm -mcpu=corei7-avx"; -const std::string LLVM_TARGET_AVX2 = "llvm -mcpu=core-avx2"; -const std::string LLVM_TARGET_SKYLAKE_AVX512 = "llvm -mcpu=skylake-avx512"; -const std::string LLVM_TARGET_AVX512 = "llvm -mcpu=skylake-avx512"; -} // namespace cpu_targets - -using TVMTensorShapes = std::vector; -using TVMInputShapes = std::unordered_map; -using InputsInfoMap = std::unordered_map; - -// Information needed to construct an TVM execution provider. -struct TvmEPOptions { - std::string executor{tvm::default_executor_type}; - std::string so_folder{""}; - bool check_hash = false; - std::string hash_file_path{""}; - std::string target{tvm::default_target_str}; - std::string target_host{tvm::default_target_str}; - unsigned int opt_level{tvm::default_opt_level}; - bool freeze_weights = true; - bool to_nhwc = false; - bool set_output_zero_copy = true; - std::string tuning_file_path{""}; - std::string tuning_type{tvm::default_tuning_type}; - std::string input_names_str{""}; - std::string input_shapes_str{""}; - TVMInputShapes input_shapes{}; - TVMTensorShapes output_shapes{}; -}; - -std::ostream& operator<<(std::ostream& out, const TvmEPOptions& options); - -class TvmEPOptionsHelper { - public: - static TvmEPOptions FromOptionsString(const char* options); - static TvmEPOptions FromProviderOptions(const ProviderOptions& options); - static std::string whitespace_trimming(const std::string& str); - - static bool checkCPUTarget(const std::string& target); - static bool checkGPUTarget(const std::string& target); - - private: - static void optionsPostprocess(TvmEPOptions& options); - static void setInputShapes(TvmEPOptions& options); - static void targetPostprocess(std::string& target); - static void ProcessCPUTarget(std::string& target); - static void ProcessGPUTarget(); - static void targetHostPostprocess(const std::string& target, std::string& target_host); - static void optLevelPostprocess(unsigned int& opt_level); -}; - -} // namespace tvm -} // namespace onnxruntime - -#endif // TVM_EXECUTION_PROVIDER_OPTIONS_H diff --git a/onnxruntime/core/providers/tvm/tvm_execution_provider.cc b/onnxruntime/core/providers/tvm/tvm_execution_provider.cc deleted file mode 100644 index 61ee8f899dbf1..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_execution_provider.cc +++ /dev/null @@ -1,304 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include -#include -#include - -#include "core/common/common.h" -#include "core/framework/execution_provider.h" -#include "core/framework/tensorprotoutils.h" -#include "core/framework/kernel_registry.h" -#include "core/framework/compute_capability.h" -#include "core/graph/graph_proto_serializer.h" -#include "core/platform/env.h" -#include "core/graph/model.h" - -#include "tvm_execution_provider.h" -#include "xpu_data_transfer.h" -#include "tvm_allocator.h" -#include "tvm_utils.h" -#include "tvm_api.h" - -using namespace ONNX_NAMESPACE; - -namespace onnxruntime { -namespace tvm { - -// Information to construct kernel function state. -struct TVMFuncState { - AllocateFunc allocate_func = nullptr; - DestroyFunc release_func = nullptr; - AllocatorHandle allocator = nullptr; - std::shared_ptr compiler = nullptr; -}; - -TvmExecutionProvider::TvmExecutionProvider(const TvmEPOptions& options) - : IExecutionProvider{kTvmExecutionProvider}, - options_{options} { - AllocatorCreationInfo default_memory_info = {[](int) { - return std::make_unique(); - }, - 0, false}; - // Get environment variables - const Env& env_instance = Env::Default(); - - const std::string dump_subgraphs_env = env_instance.GetEnvironmentVar(env_vars::kDumpSubgraphs); - if (!dump_subgraphs_env.empty()) { - dump_subgraphs_ = std::stoi(dump_subgraphs_env) != 0; - } -} - -std::vector TvmExecutionProvider::CreatePreferredAllocators() { - AllocatorCreationInfo default_memory_info = {[](int) { - return std::make_unique(); - }, - 0, false}; - return std::vector{CreateAllocator(default_memory_info)}; // TODO(leca): REVIEW: will CPU EP also use this? -} - -TvmExecutionProvider::~TvmExecutionProvider() {} - -std::vector> -TvmExecutionProvider::GetCapability(const GraphViewer& graph_viewer, - const IKernelLookup& /*kernel_lookup*/) const { - std::vector> result; - if (graph_viewer.IsSubgraph()) { - return result; - } - - const auto& init_tensors = graph_viewer.GetAllInitializedTensors(); - - std::unordered_set required_initializers; - const std::vector& sorted_nodes = graph_viewer.GetNodesInTopologicalOrder(); - std::unique_ptr sub_graph = std::make_unique(); - for (auto& node_idx : sorted_nodes) { - graph_viewer.GetNode(node_idx)->ForEachDef([&required_initializers, &init_tensors](const NodeArg& node_arg, bool is_input) { - if(is_input && init_tensors.count(node_arg.Name())) { - required_initializers.insert(node_arg.Name()); - } }, true); - } - - auto meta_def = std::make_unique<::onnxruntime::IndexedSubGraph::MetaDef>(); - meta_def->name = "TVMStandalone"; - meta_def->domain = "StandaloneTest"; - std::vector inputs; - std::vector outputs; - - for (auto& nodeArgPtr : graph_viewer.GetInputs()) { - inputs.push_back(nodeArgPtr->Name()); - } - - for (auto& name : required_initializers) { - inputs.push_back(name); - } - - for (auto& nodeArgPtr : graph_viewer.GetOutputs()) { - outputs.push_back(nodeArgPtr->Name()); - } - meta_def->inputs = inputs; - meta_def->outputs = outputs; - meta_def->since_version = 1; - meta_def->status = ONNX_NAMESPACE::EXPERIMENTAL; - sub_graph->SetMetaDef(std::move(meta_def)); - sub_graph->nodes = sorted_nodes; - result.push_back( - std::make_unique(std::move(sub_graph))); - return result; -} - -common::Status TvmExecutionProvider::Compile(const std::vector& fused_nodes_and_graphs, - std::vector& node_compute_funcs) { - printOptions(); - for (auto& fused_node_graph : fused_nodes_and_graphs) { - const GraphViewer& graph_body_viewer = fused_node_graph.filtered_graph; - const Node& fused_node = fused_node_graph.fused_node; - const std::string func_name = fused_node.Name(); - Model model(graph_body_viewer.Name(), true, ModelMetaData(), PathString(), - IOnnxRuntimeOpSchemaRegistryList(), graph_body_viewer.DomainToVersionMap(), - std::vector(), *GetLogger()); - ONNX_NAMESPACE::ModelProto model_proto = model.ToProto(); - // TVM EP is using static lib approach, so invoke serializer directly. - GraphViewerToProto(graph_body_viewer, *model_proto.mutable_graph(), true, true); - auto opset = model_proto.add_opset_import(); - opset->set_domain(kOnnxDomain); - opset->set_version(graph_body_viewer.DomainToVersionMap().at(kOnnxDomain)); - - std::string onnx_model_str; - model_proto.SerializeToString(&onnx_model_str); - compilers_[func_name] = std::make_shared(std::move(onnx_model_str), - ToUTF8String(fused_node.ModelPath().ToPathString()), - int(opset->version())); - InputsInfoMap all_input_shapes; - auto mod = compileModel(func_name, graph_body_viewer, all_input_shapes); - - std::vector output_tensors; - prepareOutputTensors(mod, output_tensors, graph_body_viewer.GetOutputs().size()); - - runners_[func_name] = std::make_shared(options_, mod, all_input_shapes, output_tensors); - - if (dump_subgraphs_) { - std::fstream dump("/tmp/" + func_name + ".onnx", - std::ios::out | std::ios::trunc | std::ios::binary); - model_proto.SerializeToOstream(&dump); - } - - // TODO(vvchernov): implement ops checking and mechanism of gracefully passing the responsibility to other EPs - // if the checking fails due to unsupported op(s) - NodeComputeInfo compute_info = prepareComputeInfo(func_name); - - node_compute_funcs.push_back(compute_info); - } - return Status::OK(); -} - -std::unique_ptr TvmExecutionProvider::GetDataTransfer() const { - // TODO(vvchernov): target or target host? - if (TvmEPOptionsHelper::checkGPUTarget(options_.target)) { - return std::make_unique(); - } else if (TvmEPOptionsHelper::checkCPUTarget(options_.target)) { - return std::make_unique(); - } else { - ORT_NOT_IMPLEMENTED("TVM GetDataTransfer is not implemented for target ", options_.target); - } -} - -void TvmExecutionProvider::printOptions() { - LOGS(*GetLogger(), INFO) << options_; -} - -std::shared_ptr TvmExecutionProvider::compileModel(const std::string& func_name, - const GraphViewer& graph_viewer, - InputsInfoMap& all_input_shapes) { - all_input_shapes.clear(); - - TVMTensorShapes input_shapes; - if (options_.freeze_weights) { - setInputShapesForFreezedNN(graph_viewer, input_shapes, all_input_shapes); - } else { - setInputShapesForUnfreezedNN(graph_viewer, input_shapes, all_input_shapes); - } - - std::shared_ptr mod = compilers_[func_name]->operator()(options_, input_shapes); - - return mod; -} - -void TvmExecutionProvider::setInputShapesForFreezedNN(const GraphViewer& graph_viewer, - TVMTensorShapes& input_shapes, - InputsInfoMap& all_input_shapes) { - const std::vector& all_nodes = graph_viewer.GetInputsIncludingInitializers(); - - size_t indx = 0; - for (const auto* node : all_nodes) { - if (!graph_viewer.IsInitializedTensor(node->Name())) { - TensorShapeVector shape = getInputShape(node); - all_input_shapes[indx++] = shape; - input_shapes.emplace_back(shape); - } - } -} - -void TvmExecutionProvider::setInputShapesForUnfreezedNN(const GraphViewer& graph_viewer, - TVMTensorShapes& input_shapes, - InputsInfoMap& all_input_shapes) { - const std::vector& all_nodes = graph_viewer.GetInputsIncludingInitializers(); - - size_t indx = 0; - for (const auto* node : all_nodes) { - TensorShapeVector shape = getInputShape(node); - all_input_shapes[indx++] = shape; - if (!graph_viewer.IsInitializedTensor(node->Name())) { - input_shapes.emplace_back(shape); - } - } -} - -TensorShapeVector TvmExecutionProvider::getInputShape(const NodeArg* node) { - TensorShapeVector shape; - const auto& node_name = node->Name(); - if (!options_.input_shapes.empty() && - options_.input_shapes.count(node_name)) { - shape = options_.input_shapes[node_name]; - } else { - shape = convertTensorShape(*node->Shape()); - } - - return shape; -} - -TensorShapeVector TvmExecutionProvider::convertTensorShape(const TensorShapeProto& shape_proto) { - TensorShape ort_shape = utils::GetTensorShapeFromTensorShapeProto(shape_proto); - size_t dims = ort_shape.NumDimensions(); - - TensorShapeVector shape(dims); - for (size_t j = 0; j < dims; ++j) { - int64_t dim = int64_t(ort_shape[j]); - ORT_ENFORCE(dim > 0, "Input dimension is not positive value (dim = " + std::to_string(dim) + "). " + - "Please use provider options to setup input_names and input_shapes"); - shape[j] = dim; - } - - return shape; -} - -void TvmExecutionProvider::prepareOutputTensors(const std::shared_ptr& mod, - std::vector& output_tensors, - size_t num) { - ORT_ENFORCE(mod != nullptr, "TVM module is not compiled"); - output_tensors.clear(); - options_.output_shapes.clear(); - options_.output_shapes.resize(num); - - if (options_.executor != "vm") { - TVMGetOutputShapes(*mod, options_.output_shapes); - } - - for (auto& output_shape : options_.output_shapes) { - DLTensor t; - // Draft for tensor, correct data is defined during inference - t.strides = nullptr; - t.byte_offset = 0; - t.data = nullptr; - if (options_.executor == "vm") { - t.ndim = 0; - t.shape = nullptr; - } else { - t.ndim = output_shape.size(); - t.shape = output_shape.data(); - } - - output_tensors.push_back(t); - } -} - -NodeComputeInfo TvmExecutionProvider::prepareComputeInfo(const std::string& func_name) { - NodeComputeInfo compute_info; - compute_info.create_state_func = std::bind(&TvmExecutionProvider::createStateFunc, - this, - std::placeholders::_1, - std::placeholders::_2); - - compute_info.release_state_func = [](FunctionState state) { - if (state) - delete static_cast(state); - }; - - compute_info.compute_func = *runners_[func_name].get(); - - return compute_info; -} - -int TvmExecutionProvider::createStateFunc(ComputeContext* context, FunctionState* state) { - auto* state_ptr = new TVMFuncState(); - *state_ptr = {context->allocate_func, - context->release_func, - context->allocator_handle, - compilers_[context->node_name]}; - // TODO(vvchernov): Who and when release state? - *state = state_ptr; - return 0; -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/tvm_execution_provider.h b/onnxruntime/core/providers/tvm/tvm_execution_provider.h deleted file mode 100644 index baa46c593fa07..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_execution_provider.h +++ /dev/null @@ -1,71 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef TVM_EXECUTION_PROVIDER_H -#define TVM_EXECUTION_PROVIDER_H - -#include -#include -#include -#include - -#include "core/common/logging/logging.h" -#include "core/framework/execution_provider.h" -#include - -#include "tvm_compiler.h" -#include "tvm_runner.h" - -namespace onnxruntime { -class Graph; -class NodeArg; -namespace tvm { - -class TvmExecutionProvider : public IExecutionProvider { - using Compiler = TVMCompilerBase; - using Compilers = std::unordered_map>; - using Runner = TVMRunner; - using Runners = std::unordered_map>; - - public: - explicit TvmExecutionProvider(const TvmEPOptions& options); - virtual ~TvmExecutionProvider(); - - std::vector> - GetCapability(const onnxruntime::GraphViewer& graph, - const IKernelLookup& /*kernel_lookup*/) const override; - - common::Status Compile(const std::vector& fused_nodes_and_graphs, - std::vector& node_compute_funcs) override; - std::unique_ptr GetDataTransfer() const override; - std::vector CreatePreferredAllocators() override; - - private: - void printOptions(); - std::shared_ptr compileModel(const std::string& func_name, - const GraphViewer& graph_viewer, - InputsInfoMap& inputs_info); // NOLINT - void setInputShapesForFreezedNN(const GraphViewer& graph_viewer, - TVMTensorShapes& input_shapes, // NOLINT - InputsInfoMap& all_input_shapes); // NOLINT - void setInputShapesForUnfreezedNN(const GraphViewer& graph_viewer, - TVMTensorShapes& input_shapes, // NOLINT - InputsInfoMap& all_input_shapes); // NOLINT - TensorShapeVector getInputShape(const NodeArg* node); - TensorShapeVector convertTensorShape(const ONNX_NAMESPACE::TensorShapeProto& shape_proto); - void prepareOutputTensors(const std::shared_ptr& mod, - std::vector& output_tensors, size_t num); // NOLINT - NodeComputeInfo prepareComputeInfo(const std::string& func_name); - int createStateFunc(ComputeContext*, FunctionState*); - - private: - TvmEPOptions options_; - Compilers compilers_; - Runners runners_; - bool dump_subgraphs_ = false; -}; - -} // namespace tvm -} // namespace onnxruntime - -#endif // TVM_EXECUTION_PROVIDER_H diff --git a/onnxruntime/core/providers/tvm/tvm_provider_factory.cc b/onnxruntime/core/providers/tvm/tvm_provider_factory.cc deleted file mode 100644 index d83fd8ee4d1cb..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_provider_factory.cc +++ /dev/null @@ -1,53 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include -#include - -#include "core/providers/tvm/tvm_provider_factory.h" -#include "core/session/abi_session_options_impl.h" - -#include "tvm_execution_provider.h" -#include "tvm_provider_factory_creator.h" -#include "tvm_so_execution_provider.h" // NOLINT(build/include_subdir) - -namespace onnxruntime { - -struct TvmProviderFactory : IExecutionProviderFactory { - TvmProviderFactory(const tvm::TvmEPOptions& options) : options_{options} {} - ~TvmProviderFactory() = default; - - std::unique_ptr CreateProvider() override { - std::unique_ptr provider = nullptr; - if (options_.so_folder != "") { - ORT_ENFORCE(options_.executor == "vm", - "Only virtual machine module is compiled from shared lib and dependences!"); - provider = std::move(std::make_unique(options_)); - } else { - provider = std::move(std::make_unique(options_)); - } - - return provider; - } - - private: - tvm::TvmEPOptions options_; -}; - -std::shared_ptr TVMProviderFactoryCreator::Create(const char* opt_str) { - tvm::TvmEPOptions options = tvm::TvmEPOptionsHelper::FromOptionsString(opt_str); - return std::make_shared(options); -} - -std::shared_ptr TVMProviderFactoryCreator::Create(const tvm::TvmEPOptions& options) { - return std::make_shared(options); -} -} // namespace onnxruntime - -ORT_API_STATUS_IMPL(OrtSessionOptionsAppendExecutionProvider_Tvm, - _In_ OrtSessionOptions* options, - _In_ const char* opt_str) { - onnxruntime::tvm::TvmEPOptions tvm_options = onnxruntime::tvm::TvmEPOptionsHelper::FromOptionsString(opt_str); - options->provider_factories.push_back(onnxruntime::TVMProviderFactoryCreator::Create(tvm_options)); - return nullptr; -} diff --git a/onnxruntime/core/providers/tvm/tvm_provider_factory_creator.h b/onnxruntime/core/providers/tvm/tvm_provider_factory_creator.h deleted file mode 100644 index 2d7e06b5b7c59..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_provider_factory_creator.h +++ /dev/null @@ -1,19 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#pragma once - -#include - -#include "core/providers/providers.h" - -namespace onnxruntime { -namespace tvm { -struct TvmEPOptions; -} - -struct TVMProviderFactoryCreator { - static std::shared_ptr Create(const tvm::TvmEPOptions& options); - static std::shared_ptr Create(const char* params); -}; -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/tvm_runner.cc b/onnxruntime/core/providers/tvm/tvm_runner.cc deleted file mode 100644 index 5dda8f5bf9c3e..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_runner.cc +++ /dev/null @@ -1,26 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/graph/model.h" -#include "core/framework/tensorprotoutils.h" - -#include "tvm_runner.h" - -using namespace ONNX_NAMESPACE; -namespace onnxruntime { -namespace tvm { - -TVMRunner::TVMRunner(const TvmEPOptions& options, - const std::shared_ptr& mod, - const InputsInfoMap& inputs_info, - const std::vector& output_tensors) { - runner_ = getTVMRunnerImpl(mod, options, inputs_info, output_tensors); -} - -common::Status TVMRunner::operator()(FunctionState state, const OrtApi* /*api*/, OrtKernelContext* context) { - Ort::KernelContext ctx(context); - return runner_->run(ctx); -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/tvm_runner.h b/onnxruntime/core/providers/tvm/tvm_runner.h deleted file mode 100644 index 4b7349ee3405e..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_runner.h +++ /dev/null @@ -1,34 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef TVM_RUNNER_H -#define TVM_RUNNER_H - -#include -#include - -#include "tvm_runner_impl.h" - -namespace onnxruntime { -namespace tvm { - -class TVMRunner { - public: - TVMRunner() = delete; - virtual ~TVMRunner() = default; - - TVMRunner(const TvmEPOptions& options, - const std::shared_ptr& mod, - const InputsInfoMap& inputs_info, - const std::vector& output_tensor); - - common::Status operator()(FunctionState state, const OrtApi* api, OrtKernelContext* context); - - private: - std::shared_ptr runner_; -}; - -} // namespace tvm -} // namespace onnxruntime - -#endif // TVM_TVM_RUNNER_H diff --git a/onnxruntime/core/providers/tvm/tvm_runner_impl.cc b/onnxruntime/core/providers/tvm/tvm_runner_impl.cc deleted file mode 100644 index c88de2652f14b..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_runner_impl.cc +++ /dev/null @@ -1,175 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/framework/tensorprotoutils.h" - -#include "tvm_runner_impl.h" -#include "tvm_utils.h" -#include "tvm_api.h" - -namespace onnxruntime { -namespace tvm { - -/* ------------------------------------ RunnerImplFactory ----------------------------- */ - -std::shared_ptr getTVMRunnerImpl(const std::shared_ptr& mod, - const TvmEPOptions& options, - const InputsInfoMap& inputs_info, - const std::vector output_tensors) { - const std::string& name = options.executor; - if (name == "graph") { - return std::make_shared(mod, inputs_info, options.output_shapes, - output_tensors, options.set_output_zero_copy); - } else if (name == "vm") { - return std::make_shared(mod, inputs_info, options.output_shapes, - output_tensors, options.set_output_zero_copy); - } - return nullptr; -} - -/* ------------------------------------ RunnerImpl ------------------------------------ */ - -RunnerImpl::RunnerImpl(const std::shared_ptr& mod, - const InputsInfoMap& inputs_info, - const TVMTensorShapes output_shapes, - const std::vector output_tensors, - bool set_output_zero_copy) : mod_(mod), - inputs_info_(inputs_info), - output_shapes_(output_shapes), - output_tensors_(output_tensors), - set_output_zero_copy_(set_output_zero_copy) { -} - -void RunnerImpl::convert_input_tensors2dl_tensors(Ort::KernelContext& context, - std::vector& dst, - std::vector& dst_inds) { - size_t num = inputs_info_.size(); - dst.reserve(num); - dst_inds.reserve(num); - for (auto& info : inputs_info_) { - // TODO(vvchernov): decomposition declaration only available with -std=c++1z or -std=gnu++1z - auto& i = info.first; - auto& shape = info.second; - - auto input_tensor = context.GetInput(i); - ORT_ENFORCE(input_tensor.IsTensor()); - - auto ort_device_type = input_tensor.GetTensorMemoryInfo().GetDeviceType(); - const auto tensor_type = input_tensor.GetTensorTypeAndShapeInfo().GetElementType(); - - DLTensor t; - t.device = GetDLDevice(ort_device_type); - t.dtype = GetDataType(tensor_type); - t.strides = nullptr; - t.byte_offset = 0; - t.data = const_cast(input_tensor.GetTensorRawData()); - t.ndim = shape.size(); - t.shape = shape.data(); - dst.emplace_back(t); - dst_inds.push_back(i); - } -} - -void RunnerImpl::add_device_type_data2output_tensors(Ort::KernelContext& context) { - size_t num_outputs = output_tensors_.size(); - for (auto i = 0u; i < num_outputs; i++) { - // setup output tensor property - auto output_tensor = context.GetOutput(i, - output_shapes_[i].data(), - output_shapes_[i].size()); - ORT_ENFORCE(output_tensor.IsTensor()); - - output_tensors_[i].device = - GetDLDevice(output_tensor.GetTensorMemoryInfo().GetDeviceType()); - output_tensors_[i].dtype = - GetDataType(output_tensor.GetTensorTypeAndShapeInfo().GetElementType()); - output_tensors_[i].data = output_tensor.GetTensorMutableRawData(); - } -} - -/* ------------------------------------ GERunnerImpl ------------------------------------ */ - -GERunnerImpl::GERunnerImpl(const std::shared_ptr& mod, - const InputsInfoMap& inputs_info, - const TVMTensorShapes output_shapes, - const std::vector output_tensors, - bool set_output_zero_copy) : RunnerImpl(mod, inputs_info, output_shapes, output_tensors, set_output_zero_copy) { -} - -void GERunnerImpl::set_input(Ort::KernelContext& context) { - std::vector inds; - std::vector dl_tensors_inputs; - convert_input_tensors2dl_tensors(context, dl_tensors_inputs, inds); - - tvm::TVMSetInputs(*mod_, inds, dl_tensors_inputs); -} - -void GERunnerImpl::connect_output_tensors2ort(Ort::KernelContext& context) { - add_device_type_data2output_tensors(context); -} - -void GERunnerImpl::set_output_zero_copy() { - tvm::TVMSetOutputsZeroCopy(*mod_, output_tensors_); -} - -void GERunnerImpl::run() { - tvm::TVMRun(*mod_); -} - -void GERunnerImpl::get_outputs() { - tvm::TVMGetOutputs(*mod_, output_tensors_); -} - -/* ------------------------------------ VMRunnerImpl ------------------------------------ */ - -VMRunnerImpl::VMRunnerImpl(const std::shared_ptr& mod, - const InputsInfoMap& inputs_info, - const TVMTensorShapes output_shapes, - const std::vector output_tensors, - bool set_output_zero_copy) : RunnerImpl(mod, inputs_info, output_shapes, output_tensors, set_output_zero_copy) { -} - -void VMRunnerImpl::set_input(Ort::KernelContext& context) { - std::vector inds; - std::vector dl_tensors_inputs; - convert_input_tensors2dl_tensors(context, dl_tensors_inputs, inds); - - tvm::TVM_VM_SetInputs(*mod_, inds, dl_tensors_inputs); -} - -void VMRunnerImpl::connect_output_tensors2ort(Ort::KernelContext& context) { - // TODO(vvchernov): try to find more flexible solution - if (!probe_infer_) { - infer_once_to_get_output_shapes(); - } - - add_device_type_data2output_tensors(context); -} - -void VMRunnerImpl::set_output_zero_copy() { - tvm::TVM_VM_SetOutputsZeroCopy(*mod_, output_tensors_); -} - -void VMRunnerImpl::run() { - tvm::TVM_VM_Run(*mod_); -} - -void VMRunnerImpl::get_outputs() { - tvm::TVM_VM_GetOutputs(*mod_, output_tensors_); -} - -void VMRunnerImpl::infer_once_to_get_output_shapes() { - run(); - size_t num_outputs = output_tensors_.size(); - // TODO(vvchernov): check it - output_shapes_.resize(num_outputs); - tvm::TVMGetOutputShapes(*mod_, output_shapes_); - for (size_t i = 0; i < num_outputs; ++i) { - output_tensors_[i].ndim = output_shapes_[i].size(); - output_tensors_[i].shape = output_shapes_[i].data(); - } - probe_infer_ = true; -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/tvm_runner_impl.h b/onnxruntime/core/providers/tvm/tvm_runner_impl.h deleted file mode 100644 index 8c325303673b6..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_runner_impl.h +++ /dev/null @@ -1,126 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef TVM_RUNNER_IMPL_H -#define TVM_RUNNER_IMPL_H - -#include -#include -#include - -#include "core/framework/func_api.h" -#include "core/session/onnxruntime_cxx_api.h" - -#include "tvm_common.h" -#include "tvm_ep_options.h" - -namespace onnxruntime { -namespace tvm { - -class RunnerImpl { - public: - RunnerImpl() = delete; - RunnerImpl(const std::shared_ptr& mod, - const InputsInfoMap& inputs_info, - const TVMTensorShapes output_shapes, - const std::vector tensors_outputs, - bool set_output_zero_copy); - virtual ~RunnerImpl() = default; - - virtual common::Status run(Ort::KernelContext& context) { - common::Status res; - if (set_output_zero_copy_) { - res = run_without_output_copying(context); - } else { - res = run_with_output_copying(context); - } - return res; - } - - virtual common::Status run_without_output_copying(Ort::KernelContext& context) { - set_input(context); - connect_output_tensors2ort(context); - set_output_zero_copy(); - run(); - - return Status::OK(); - } - - virtual common::Status run_with_output_copying(Ort::KernelContext& context) { - set_input(context); - connect_output_tensors2ort(context); - run(); - get_outputs(); - - return Status::OK(); - } - - virtual void set_input(Ort::KernelContext& context) = 0; - virtual void connect_output_tensors2ort(Ort::KernelContext& context) = 0; - virtual void set_output_zero_copy() = 0; - virtual void run() = 0; - virtual void get_outputs() = 0; - - protected: - void convert_input_tensors2dl_tensors(Ort::KernelContext& context, - std::vector& dst, - std::vector& dst_inds); - void add_device_type_data2output_tensors(Ort::KernelContext& context); - - protected: - std::shared_ptr mod_; - InputsInfoMap inputs_info_; - TVMTensorShapes output_shapes_; - std::vector output_tensors_; - bool set_output_zero_copy_; -}; - -class GERunnerImpl : public RunnerImpl { - public: - GERunnerImpl() = delete; - GERunnerImpl(const std::shared_ptr& mod, - const InputsInfoMap& inputs_info, - const TVMTensorShapes output_shapes, - const std::vector tensors_outputs, - bool set_output_zero_copy); - virtual ~GERunnerImpl() = default; - - void set_input(Ort::KernelContext& context) final; - void connect_output_tensors2ort(Ort::KernelContext& context) final; - void set_output_zero_copy() final; - void run() final; - void get_outputs() final; -}; - -class VMRunnerImpl : public RunnerImpl { - public: - VMRunnerImpl() = delete; - VMRunnerImpl(const std::shared_ptr& mod, - const InputsInfoMap& inputs_info, - const TVMTensorShapes output_shapes, - const std::vector tensors_outputs, - bool set_output_zero_copy); - virtual ~VMRunnerImpl() = default; - - void set_input(Ort::KernelContext& context) final; - void connect_output_tensors2ort(Ort::KernelContext& context) final; - void set_output_zero_copy() final; - void run() final; - void get_outputs() final; - - private: - void infer_once_to_get_output_shapes(); - - private: - bool probe_infer_ = false; -}; - -std::shared_ptr getTVMRunnerImpl(const std::shared_ptr& mod, - const TvmEPOptions& options, - const InputsInfoMap& inputs_info, - const std::vector output_tensors); - -} // namespace tvm -} // namespace onnxruntime - -#endif // TVM_TVM_RUNNER_IMPL_H diff --git a/onnxruntime/core/providers/tvm/tvm_so_execution_provider.cc b/onnxruntime/core/providers/tvm/tvm_so_execution_provider.cc deleted file mode 100644 index 029f25d6f292a..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_so_execution_provider.cc +++ /dev/null @@ -1,284 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include -#include -#include - -#include "core/framework/execution_provider.h" -#include "core/framework/tensorprotoutils.h" -#include "core/framework/kernel_registry.h" -#include "core/framework/compute_capability.h" -#include "core/platform/env.h" -#include "core/graph/model.h" - -#include "tvm_so_execution_provider.h" // NOLINT(build/include_subdir) -#include "xpu_data_transfer.h" // NOLINT(build/include_subdir) -#include "tvm_allocator.h" // NOLINT(build/include_subdir) -#include "tvm_utils.h" // NOLINT(build/include_subdir) -#include "tvm_api.h" // NOLINT(build/include_subdir) -#ifdef USE_TVM_HASH -#include "hash_alg/hasher.h" // NOLINT(build/include_subdir) -#endif - -using ONNX_NAMESPACE::TensorShapeProto; - -namespace onnxruntime { -namespace tvm { - -// Information to construct kernel function state. -struct TVMFuncState { - AllocateFunc allocate_func = nullptr; - DestroyFunc release_func = nullptr; - AllocatorHandle allocator = nullptr; - std::shared_ptr compiler = nullptr; -}; - -TvmSoExecutionProvider::TvmSoExecutionProvider(const TvmEPOptions& options) - : IExecutionProvider{kTvmExecutionProvider}, - options_{options} { - // Get environment variables - const Env& env_instance = Env::Default(); - - const std::string dump_subgraphs_env = env_instance.GetEnvironmentVar(env_vars::kDumpSubgraphs); - ORT_ENFORCE(dump_subgraphs_env.empty(), "TVM EP processing shared lib does not support subgraphs"); -} - -std::vector TvmSoExecutionProvider::CreatePreferredAllocators() { - AllocatorCreationInfo default_memory_info = {[](int) { - return std::make_unique(); - }, - 0, false}; - return std::vector{CreateAllocator(default_memory_info)}; -} - -TvmSoExecutionProvider::~TvmSoExecutionProvider() {} - -std::vector> -TvmSoExecutionProvider::GetCapability(const GraphViewer& graph_viewer, - const IKernelLookup& /*kernel_lookup*/) const { - std::vector> result; - if (graph_viewer.IsSubgraph()) { - return result; - } - - const auto& init_tensors = graph_viewer.GetAllInitializedTensors(); - - std::unordered_set required_initializers; - const std::vector& sorted_nodes = graph_viewer.GetNodesInTopologicalOrder(); - std::unique_ptr sub_graph = std::make_unique(); - for (auto& node_idx : sorted_nodes) { - graph_viewer.GetNode(node_idx)->ForEachDef([&required_initializers, &init_tensors](const NodeArg& node_arg, bool is_input) { - if (is_input && init_tensors.count(node_arg.Name())) { - required_initializers.insert(node_arg.Name()); - } }, true); - } - - auto meta_def = std::make_unique<::onnxruntime::IndexedSubGraph::MetaDef>(); - meta_def->name = "TVMStandalone"; - meta_def->domain = "StandaloneTest"; - std::vector inputs; - std::vector outputs; - - for (auto& nodeArgPtr : graph_viewer.GetInputs()) { - inputs.push_back(nodeArgPtr->Name()); - } - - for (auto& name : required_initializers) { - inputs.push_back(name); - } - - for (auto& nodeArgPtr : graph_viewer.GetOutputs()) { - outputs.push_back(nodeArgPtr->Name()); - } - meta_def->inputs = inputs; - meta_def->outputs = outputs; - meta_def->since_version = 1; - meta_def->status = ONNX_NAMESPACE::EXPERIMENTAL; - sub_graph->SetMetaDef(std::move(meta_def)); - sub_graph->nodes = sorted_nodes; - result.push_back( - std::make_unique(std::move(sub_graph))); - return result; -} - -common::Status TvmSoExecutionProvider::Compile(const std::vector& fused_nodes_and_graphs, - std::vector& node_compute_funcs) { - printOptions(); - for (auto& fused_node_graph : fused_nodes_and_graphs) { - const GraphViewer& graph_body_viewer = fused_node_graph.filtered_graph; - const Node& fused_node = fused_node_graph.fused_node; -#ifdef USE_TVM_HASH - if (options_.check_hash) { - ORT_ENFORCE(checkHash(ToUTF8String(fused_node.ModelPath().ToPathString())), - "Hash check shows that used tuning files were not obtained for the given onnx-model"); - } -#endif - const std::string func_name = fused_node.Name(); - - compilers_[func_name] = std::make_shared(); - InputsInfoMap all_input_shapes; - auto mod = compileModel(func_name, graph_body_viewer, all_input_shapes); - - std::vector output_tensors(graph_body_viewer.GetOutputs().size()); - prepareOutputTensors(output_tensors); - - runners_[func_name] = std::make_shared(options_, mod, all_input_shapes, output_tensors); - - // TODO(vvchernov): implement ops checking and mechanism of gracefully passing the responsibility to other EPs - // if the checking fails due to unsupported op(s) - NodeComputeInfo compute_info = prepareComputeInfo(func_name); - - node_compute_funcs.push_back(compute_info); - } - return Status::OK(); -} - -std::unique_ptr TvmSoExecutionProvider::GetDataTransfer() const { - // TODO(vvchernov): target or target host? - if (TvmEPOptionsHelper::checkGPUTarget(options_.target)) { - return std::make_unique(); - } else if (TvmEPOptionsHelper::checkCPUTarget(options_.target)) { - return std::make_unique(); - } else { - ORT_NOT_IMPLEMENTED("TVM GetDataTransfer is not implemented for target ", options_.target); - } -} - -void TvmSoExecutionProvider::printOptions() { - LOGS(*GetLogger(), INFO) << options_; -} - -#ifdef USE_TVM_HASH -bool TvmSoExecutionProvider::checkHash(const std::string& onnx_path) const { - auto hasher = Hasher("sha256"); - std::string onnx_str = readFromFile(onnx_path); - std::string onnx_hash = hasher.hash(onnx_str.c_str(), onnx_str.size()); - onnx_str.clear(); - std::string hash; - if (options_.hash_file_path.empty()) { - // TODO(vvchernov): align hash file name with OctoML team - hash = readFromFile(options_.so_folder + "/hash.txt"); - } else { - hash = readFromFile(options_.hash_file_path); - } - return onnx_hash == hash; -} -#endif - -std::shared_ptr TvmSoExecutionProvider::compileModel(const std::string& func_name, - const GraphViewer& graph_viewer, - InputsInfoMap& all_input_shapes) { - all_input_shapes.clear(); - - TVMTensorShapes input_shapes; - if (options_.freeze_weights) { - setInputShapesForFreezedNN(graph_viewer, input_shapes, all_input_shapes); - } else { - setInputShapesForUnfreezedNN(graph_viewer, input_shapes, all_input_shapes); - } - - std::shared_ptr mod = compilers_[func_name]->operator()(options_, input_shapes); - - return mod; -} - -void TvmSoExecutionProvider::setInputShapesForFreezedNN(const GraphViewer& graph_viewer, - TVMTensorShapes& input_shapes, - InputsInfoMap& all_input_shapes) { - const std::vector& all_nodes = graph_viewer.GetInputsIncludingInitializers(); - - size_t indx = 0; - for (const auto* node : all_nodes) { - if (!graph_viewer.IsInitializedTensor(node->Name())) { - TensorShapeVector shape = getInputShape(node); - all_input_shapes[indx++] = shape; - input_shapes.emplace_back(shape); - } - } -} - -void TvmSoExecutionProvider::setInputShapesForUnfreezedNN(const GraphViewer& graph_viewer, - TVMTensorShapes& input_shapes, - InputsInfoMap& all_input_shapes) { - const std::vector& all_nodes = graph_viewer.GetInputsIncludingInitializers(); - - size_t indx = 0; - for (const auto* node : all_nodes) { - TensorShapeVector shape = getInputShape(node); - all_input_shapes[indx++] = shape; - if (!graph_viewer.IsInitializedTensor(node->Name())) { - input_shapes.emplace_back(shape); - } - } -} - -TensorShapeVector TvmSoExecutionProvider::getInputShape(const NodeArg* node) { - TensorShapeVector shape; - const auto& node_name = node->Name(); - if (!options_.input_shapes.empty() && - options_.input_shapes.count(node_name)) { - shape = options_.input_shapes[node_name]; - } else { - shape = convertTensorShape(*node->Shape()); - } - - return shape; -} - -TensorShapeVector TvmSoExecutionProvider::convertTensorShape(const TensorShapeProto& shape_proto) { - TensorShape ort_shape = utils::GetTensorShapeFromTensorShapeProto(shape_proto); - size_t dims = ort_shape.NumDimensions(); - - TensorShapeVector shape(dims); - for (size_t j = 0; j < dims; ++j) { - int64_t dim = int64_t(ort_shape[j]); - ORT_ENFORCE(dim > 0, "Input dimension is not positive value (dim = " + std::to_string(dim) + "). " + - "Please use provider options to setup input_names and input_shapes"); - shape[j] = dim; - } - - return shape; -} - -void TvmSoExecutionProvider::prepareOutputTensors(std::vector& output_tensors) { - for (DLTensor& t : output_tensors) { - // Draft for tensor, correct data is defined during inference - t.strides = nullptr; - t.byte_offset = 0; - t.data = nullptr; - t.ndim = 0; - t.shape = nullptr; - } -} - -NodeComputeInfo TvmSoExecutionProvider::prepareComputeInfo(const std::string& func_name) { - NodeComputeInfo compute_info; - compute_info.create_state_func = std::bind(&TvmSoExecutionProvider::createStateFunc, - this, - std::placeholders::_1, - std::placeholders::_2); - - compute_info.release_state_func = [](FunctionState state) { - if (state) - delete static_cast(state); - }; - - compute_info.compute_func = *runners_[func_name].get(); - - return compute_info; -} - -int TvmSoExecutionProvider::createStateFunc(ComputeContext* context, FunctionState* state) { - auto* state_ptr = new TVMFuncState(); - *state_ptr = {context->allocate_func, - context->release_func, - context->allocator_handle, - compilers_[context->node_name]}; - // TODO(vvchernov): Who and when release state? - *state = state_ptr; - return 0; -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/tvm_so_execution_provider.h b/onnxruntime/core/providers/tvm/tvm_so_execution_provider.h deleted file mode 100644 index d3840f46b5b55..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_so_execution_provider.h +++ /dev/null @@ -1,72 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef ONNXRUNTIME_CORE_PROVIDERS_TVM_TVM_SO_EXECUTION_PROVIDER_H_ -#define ONNXRUNTIME_CORE_PROVIDERS_TVM_TVM_SO_EXECUTION_PROVIDER_H_ - -#include -#include -#include -#include - -#include "core/common/logging/logging.h" -#include "core/framework/execution_provider.h" -#include - -#include "tvm_compiler.h" // NOLINT(build/include_subdir) -#include "tvm_runner.h" // NOLINT(build/include_subdir) - -namespace onnxruntime { -class Graph; -class NodeArg; -namespace tvm { - -class TvmSoExecutionProvider : public IExecutionProvider { - using Compiler = TVMCompilerBase; - using Compilers = std::unordered_map>; - using Runner = TVMRunner; - using Runners = std::unordered_map>; - - public: - explicit TvmSoExecutionProvider(const TvmEPOptions& options); - virtual ~TvmSoExecutionProvider(); - - std::vector> - GetCapability(const onnxruntime::GraphViewer& graph, - const IKernelLookup& /*kernel_lookup*/) const override; - - common::Status Compile(const std::vector& fused_nodes_and_graphs, - std::vector& node_compute_funcs) override; - std::unique_ptr GetDataTransfer() const override; - std::vector CreatePreferredAllocators() override; - - private: - void printOptions(); -#ifdef USE_TVM_HASH - bool checkHash(const std::string& onnx_path) const; -#endif - std::shared_ptr compileModel(const std::string& func_name, - const GraphViewer& graph_viewer, - InputsInfoMap& inputs_info); // NOLINT - void setInputShapesForFreezedNN(const GraphViewer& graph_viewer, - TVMTensorShapes& input_shapes, // NOLINT - InputsInfoMap& all_input_shapes); // NOLINT - void setInputShapesForUnfreezedNN(const GraphViewer& graph_viewer, - TVMTensorShapes& input_shapes, // NOLINT - InputsInfoMap& all_input_shapes); // NOLINT - TensorShapeVector getInputShape(const NodeArg* node); - TensorShapeVector convertTensorShape(const ONNX_NAMESPACE::TensorShapeProto& shape_proto); - void prepareOutputTensors(std::vector& output_tensors); // NOLINT - NodeComputeInfo prepareComputeInfo(const std::string& func_name); - int createStateFunc(ComputeContext*, FunctionState*); - - private: - TvmEPOptions options_; - Compilers compilers_; - Runners runners_; -}; - -} // namespace tvm -} // namespace onnxruntime - -#endif // ONNXRUNTIME_CORE_PROVIDERS_TVM_TVM_SO_EXECUTION_PROVIDER_H_ diff --git a/onnxruntime/core/providers/tvm/tvm_utils.cc b/onnxruntime/core/providers/tvm/tvm_utils.cc deleted file mode 100644 index e0a5b566835c8..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_utils.cc +++ /dev/null @@ -1,31 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef TVM_UTILS_H -#define TVM_UTILS_H - -#include -#include - -#include "tvm_utils.h" // NOLINT(build/include_subdir) - -namespace onnxruntime { -namespace tvm { - -std::string readFromFile(const std::string& file_path) { - std::string str; - - std::ifstream t(file_path); - t.seekg(0, std::ios::end); - str.reserve(t.tellg()); - t.seekg(0, std::ios::beg); - - str.assign((std::istreambuf_iterator(t)), - std::istreambuf_iterator()); - return str; -} - -} // namespace tvm -} // namespace onnxruntime - -#endif // TVM_UTILS_H diff --git a/onnxruntime/core/providers/tvm/tvm_utils.h b/onnxruntime/core/providers/tvm/tvm_utils.h deleted file mode 100644 index de77368c715b9..0000000000000 --- a/onnxruntime/core/providers/tvm/tvm_utils.h +++ /dev/null @@ -1,70 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef TVM_UTILS_H -#define TVM_UTILS_H - -#include - -#include "tvm_common.h" - -#include "core/session/onnxruntime_cxx_api.h" -#include "core/framework/ortdevice.h" -#include "core/common/common.h" - -namespace onnxruntime { -namespace tvm { - -inline DLDataType GetDataType(ONNXTensorElementDataType type) { - switch (type) { - case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8: - return {kDLUInt, 8, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8: - return {kDLInt, 8, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16: - return {kDLUInt, 16, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16: - return {kDLInt, 16, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32: - return {kDLUInt, 32, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32: - return {kDLInt, 32, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64: - return {kDLUInt, 64, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64: - return {kDLInt, 64, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16: - return {kDLFloat, 16, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT: - return {kDLFloat, 32, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE: - return {kDLFloat, 64, 1}; - case ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL: - return {kDLUInt, 1, 1}; - default: - ORT_NOT_IMPLEMENTED("Unsupported data type"); - } -} - -inline DLDevice GetDLDevice(OrtMemoryInfoDeviceType device_type) { - DLDevice context; - switch (device_type) { - case OrtDevice::CPU: - context = {kDLCPU, 0}; - break; - case OrtDevice::GPU: - context = {kDLVulkan, 0}; - break; - default: - ORT_NOT_IMPLEMENTED("Unsupported device"); - break; - } - return context; -} - -std::string readFromFile(const std::string& file_path); - -} // namespace tvm -} // namespace onnxruntime - -#endif // TVM_UTILS_H diff --git a/onnxruntime/core/providers/tvm/xpu_data_transfer.cc b/onnxruntime/core/providers/tvm/xpu_data_transfer.cc deleted file mode 100644 index 4011dee7b7b7f..0000000000000 --- a/onnxruntime/core/providers/tvm/xpu_data_transfer.cc +++ /dev/null @@ -1,84 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#include "core/framework/tensor.h" - -#include "xpu_data_transfer.h" -#include "tvm_utils.h" - -namespace onnxruntime { -namespace tvm { - -XPUDataTransfer::XPUDataTransfer() { -} - -XPUDataTransfer::~XPUDataTransfer() { -} - -bool XPUDataTransfer::CanCopy(const OrtDevice& src_device, const OrtDevice& dst_device) const { - return (src_device.Type() == OrtDevice::CPU && dst_device.Type() == OrtDevice::CPU) || - (src_device.Type() == OrtDevice::GPU || dst_device.Type() == OrtDevice::GPU); -} - -common::Status XPUDataTransfer::CopyTensor(const Tensor& src, Tensor& dst) const { - size_t bytes = src.SizeInBytes(); - const void* src_data = src.DataRaw(); - void* dst_data = dst.MutableDataRaw(); - const auto src_device_type = src.Location().device.Type(); - const auto dst_device_type = dst.Location().device.Type(); - - if ((src_device_type == OrtDevice::CPU) && (dst_device_type == OrtDevice::CPU)) { - if (src_data == dst_data) { - // no need copying as both pointers are referring to same piece of memory. - return Status::OK(); - } - memcpy(dst_data, src_data, bytes); - } else { - DLTensor tvm_src, tvm_dst; - DLDataType dl_type{kDLInt, 8, 1}; - std::vector shape{int64_t(bytes)}; - // Construct source DLTensor - tvm_src.device = GetDLDevice(static_cast(src_device_type)); - tvm_src.dtype = dl_type; - tvm_src.strides = nullptr; - tvm_src.byte_offset = 0; - tvm_src.data = const_cast(src_data); - tvm_src.ndim = 1; - tvm_src.shape = shape.data(); - // Construct destination DLTensor - tvm_dst.device = GetDLDevice(static_cast(dst_device_type)); - tvm_dst.dtype = dl_type; - tvm_dst.strides = nullptr; - tvm_dst.byte_offset = 0; - tvm_dst.data = dst_data; - tvm_dst.ndim = 1; - tvm_dst.shape = shape.data(); - // Copying from src to dst - TVMDeviceCopyDataFromTo(&tvm_src, &tvm_dst, nullptr); - } - return Status::OK(); -} - -DLDevice XPUDataTransfer::get_context(const OrtDevice& device) const { - return GetDLDevice(static_cast(device.Type())); -} - -bool TvmCPUDataTransfer::CanCopy(const OrtDevice& src_device, const OrtDevice& dst_device) const { - return src_device.Type() == OrtDevice::CPU && dst_device.Type() == OrtDevice::CPU; -} - -common::Status TvmCPUDataTransfer::CopyTensor(const Tensor& src, Tensor& dst) const { - const void* src_data = src.DataRaw(); - void* dst_data = dst.MutableDataRaw(); - if (src_data == dst_data) { - // no need copying as both pointers are referring to same piece of memory. - return Status::OK(); - } - // Copying only happens between two same size tensors. - ORT_ENFORCE(src.SizeInBytes() == dst.SizeInBytes()); - memcpy(dst_data, src_data, src.SizeInBytes()); - return Status::OK(); -} - -} // namespace tvm -} // namespace onnxruntime diff --git a/onnxruntime/core/providers/tvm/xpu_data_transfer.h b/onnxruntime/core/providers/tvm/xpu_data_transfer.h deleted file mode 100644 index a2cf55b241bb1..0000000000000 --- a/onnxruntime/core/providers/tvm/xpu_data_transfer.h +++ /dev/null @@ -1,38 +0,0 @@ -// Copyright (c) Microsoft Corporation. All rights reserved. -// Licensed under the MIT License. - -#ifndef XPU_DATA_TRANSFER -#define XPU_DATA_TRANSFER - -#include "core/framework/data_transfer.h" -#include "tvm_common.h" - -namespace onnxruntime { -namespace tvm { - -class XPUDataTransfer : public IDataTransfer { - public: - XPUDataTransfer(); - ~XPUDataTransfer(); - - bool CanCopy(const OrtDevice& src_device, const OrtDevice& dst_device) const override; - - // Dumpen MSVC warning about not fully overriding - using IDataTransfer::CopyTensor; - common::Status CopyTensor(const Tensor& src, Tensor& dst) const override; - DLDevice get_context(const OrtDevice& device) const; -}; - -class TvmCPUDataTransfer : public IDataTransfer { - public: - TvmCPUDataTransfer() = default; - // Dampen MSVC warning about not fully overriding CopyTensor - using IDataTransfer::CopyTensor; - bool CanCopy(const OrtDevice& src_device, const OrtDevice& dst_device) const override; - common::Status CopyTensor(const Tensor& src, Tensor& dst) const override; -}; - -} // namespace tvm -} // namespace onnxruntime - -#endif // XPU_DATA_TRANSFER diff --git a/onnxruntime/core/session/provider_registration.cc b/onnxruntime/core/session/provider_registration.cc index 8c512c561ea8c..8bea347c85280 100644 --- a/onnxruntime/core/session/provider_registration.cc +++ b/onnxruntime/core/session/provider_registration.cc @@ -205,15 +205,6 @@ ORT_API_STATUS_IMPL(OrtSessionOptionsAppendExecutionProvider_Nnapi, } #endif -#ifndef USE_TVM -ORT_API_STATUS_IMPL(OrtSessionOptionsAppendExecutionProvider_Tvm, - _In_ OrtSessionOptions* options, _In_ const char* settings) { - ORT_UNUSED_PARAMETER(options); - ORT_UNUSED_PARAMETER(settings); - return CreateNotEnabledStatus("Tvm"); -} -#endif - #ifdef __cplusplus } #endif diff --git a/onnxruntime/python/onnxruntime_pybind_state.cc b/onnxruntime/python/onnxruntime_pybind_state.cc index 05ca3c6c15793..09bbccb23e207 100644 --- a/onnxruntime/python/onnxruntime_pybind_state.cc +++ b/onnxruntime/python/onnxruntime_pybind_state.cc @@ -1128,16 +1128,6 @@ std::unique_ptr CreateExecutionProviderInstance( LOGS_DEFAULT(WARNING) << "Failed to create " << type << ". Please refer https://onnxruntime.ai/docs/execution-providers/OpenVINO-ExecutionProvider.html#requirements to ensure all dependencies are met."; } } -#endif - } else if (type == kTvmExecutionProvider) { -#if USE_TVM - onnxruntime::tvm::TvmEPOptions info{}; - const auto it = provider_options_map.find(type); - if (it != provider_options_map.end()) { - info = onnxruntime::tvm::TvmEPOptionsHelper::FromProviderOptions(it->second); - } - - return onnxruntime::TVMProviderFactoryCreator::Create(info)->CreateProvider(); #endif } else if (type == kVitisAIExecutionProvider) { #ifdef USE_VITISAI diff --git a/onnxruntime/python/onnxruntime_pybind_state_common.h b/onnxruntime/python/onnxruntime_pybind_state_common.h index fa4916f8922f2..b71081bf20efc 100644 --- a/onnxruntime/python/onnxruntime_pybind_state_common.h +++ b/onnxruntime/python/onnxruntime_pybind_state_common.h @@ -24,7 +24,7 @@ struct OrtStatus { char msg[1]; // a null-terminated string }; -#define BACKEND_DEVICE BACKEND_PROC BACKEND_DNNL BACKEND_OPENVINO BACKEND_TVM BACKEND_OPENBLAS BACKEND_MIGRAPHX BACKEND_ACL BACKEND_ARMNN BACKEND_DML BACKEND_CANN BACKEND_WEBGPU +#define BACKEND_DEVICE BACKEND_PROC BACKEND_DNNL BACKEND_OPENVINO BACKEND_OPENBLAS BACKEND_MIGRAPHX BACKEND_ACL BACKEND_ARMNN BACKEND_DML BACKEND_CANN BACKEND_WEBGPU #include "core/session/onnxruntime_cxx_api.h" #include "core/providers/providers.h" #include "core/providers/provider_factory_creators.h" @@ -75,12 +75,6 @@ struct OrtStatus { #define BACKEND_OPENVINO "" #endif -#ifdef USE_TVM -#define BACKEND_TVM "-TVM" -#else -#define BACKEND_TVM "" -#endif - #if USE_OPENBLAS #define BACKEND_OPENBLAS "-OPENBLAS" #else @@ -141,9 +135,6 @@ extern std::string openvino_device_type; } } // namespace onnxruntime #endif -#ifdef USE_TVM -#include "core/providers/tvm/tvm_ep_options.h" -#endif #ifdef USE_ACL #include "core/providers/acl/acl_provider_factory.h" #endif @@ -444,10 +435,6 @@ std::shared_ptr CreateExecutionProviderFactory_MIGrap std::shared_ptr CreateExecutionProviderFactory_MIGraphX(int device_id); std::shared_ptr CreateExecutionProviderFactory_Cuda(const OrtCUDAProviderOptions* params); std::shared_ptr CreateExecutionProviderFactory_Dnnl(const OrtDnnlProviderOptions* params); -#ifdef USE_TVM -std::shared_ptr CreateExecutionProviderFactory_Tvm(const tvm::TvmEPOptions& info); -std::shared_ptr CreateExecutionProviderFactory_Tvm(const char* params); -#endif std::shared_ptr CreateExecutionProviderFactory_ACL(bool enable_fast_math); std::shared_ptr CreateExecutionProviderFactory_ArmNN(int use_arena); std::shared_ptr CreateExecutionProviderFactory_DML(int device_id); diff --git a/onnxruntime/python/providers/tvm/__init__.py b/onnxruntime/python/providers/tvm/__init__.py deleted file mode 100644 index 4bcbc0bfef586..0000000000000 --- a/onnxruntime/python/providers/tvm/__init__.py +++ /dev/null @@ -1,10 +0,0 @@ -# ------------------------------------------------------------------------- -# Copyright (c) Microsoft Corporation. All rights reserved. -# Licensed under the MIT License. See License.txt in the project root for -# license information. -# -------------------------------------------------------------------------- -""" -JIT interface implementing packed functions that -import and compile frontend models -""" -from .ort import ANSOR_TYPE, AUTO_TVM_TYPE, onnx_compile # noqa: F401 diff --git a/onnxruntime/python/providers/tvm/extend_python_file.py b/onnxruntime/python/providers/tvm/extend_python_file.py deleted file mode 100644 index 65902619f8150..0000000000000 --- a/onnxruntime/python/providers/tvm/extend_python_file.py +++ /dev/null @@ -1,54 +0,0 @@ -# ------------------------------------------------------------------------- -# Copyright (c) Microsoft Corporation. All rights reserved. -# Licensed under the MIT License. See License.txt in the project root for -# license information. -# -------------------------------------------------------------------------- - -import argparse -import textwrap - - -def rewrite_target_file(target): - with open(target, "a") as f: - f.write( - textwrap.dedent( - """ - import warnings - - try: - # This import is necessary in order to delegate the loading of libtvm.so to TVM. - import tvm - except ImportError as e: - warnings.warn( - f"WARNING: Failed to import TVM, libtvm.so was not loaded. More details: {e}" - ) - try: - # Working between the C++ and Python parts in TVM EP is done using the PackedFunc and - # Registry classes. In order to use a Python function in C++ code, it must be registered in - # the global table of functions. Registration is carried out through the JIT interface, - # so it is necessary to call special functions for registration. - # To do this, we need to make the following import. - import onnxruntime.providers.tvm - except ImportError as e: - warnings.warn( - f"WARNING: Failed to register python functions to work with TVM EP. More details: {e}" - ) - """ - ) - ) - - -def main(): - parser = argparse.ArgumentParser() - parser.add_argument( - "--target_file", - type=str, - required=True, - help="Path to the file to be expanded.", - ) - args = parser.parse_args() - rewrite_target_file(args.target_file) - - -if __name__ == "__main__": - main() diff --git a/onnxruntime/python/providers/tvm/ort.py b/onnxruntime/python/providers/tvm/ort.py deleted file mode 100644 index be6d23f39c532..0000000000000 --- a/onnxruntime/python/providers/tvm/ort.py +++ /dev/null @@ -1,140 +0,0 @@ -# ------------------------------------------------------------------------- -# Copyright (c) Microsoft Corporation. All rights reserved. -# Licensed under the MIT License. See License.txt in the project root for -# license information. -# -------------------------------------------------------------------------- - -import collections -import copy -import logging -import os - -import onnx -import tvm -from tvm import auto_scheduler, autotvm, relay -from tvm.contrib import graph_executor -from tvm.relay import vm - -log = logging.getLogger("tvm_ep") - -ANSOR_TYPE = "Ansor" -AUTO_TVM_TYPE = "AutoTVM" - - -@tvm.register_func("tvm_onnx_import_and_compile") -def onnx_compile( - model_string, - model_path, - executor, - target, - target_host, - opt_level, - opset, - freeze_params, - input_shapes, - nhwc=False, - tuning_logfile="", - tuning_type=AUTO_TVM_TYPE, -): - def get_tvm_executor(irmod, executor, target, params): - if executor == "vm": - log.info("Build TVM virtual machine") - lib = vm.compile( - copy.deepcopy(irmod), - target, - params=params, - ) - elif executor == "graph": - log.info("Build TVM graph executor") - lib = relay.build(irmod, target=target, params=params) - else: - log.error(f'Executor type {executor} is unsupported. Only "vm" and "graph" types are supported') - return None - return lib - - model = onnx.load_model_from_string(bytes(model_string)) - if model_path: - base_dir = os.path.dirname(os.path.abspath(model_path)) - onnx.load_external_data_for_model(model, base_dir) - - # Collect only feed input names from all input names - all_input_names = [node.name for node in model.graph.input] - all_initializer = [node.name for node in model.graph.initializer] - net_feed_input_names = list(set(all_input_names) - set(all_initializer)) - - # Match names and input shapes - all_input_mapping = [(name, shape) for (name, shape) in zip(all_input_names, input_shapes)] - # Using an ordereddict maintains input ordering. - shape_dict = collections.OrderedDict(all_input_mapping) - # Get only feed input pairs - feed_shape_dict = {} - for name in net_feed_input_names: - feed_shape_dict[name] = shape_dict[name] - - irmod, params = relay.frontend.from_onnx(model, feed_shape_dict, opset=opset, freeze_params=freeze_params) - irmod = relay.transform.DynamicToStatic()(irmod) - - # Tuning file can be set by client through ep options - if not tuning_logfile: - tuning_logfile = os.getenv("AUTOTVM_TUNING_LOG") - lib = None - tvm_target = tvm.target.Target(target, host=target_host) - if tuning_logfile: - if tuning_type == ANSOR_TYPE: - desired_layouts = { - "nn.conv2d": ["NHWC", "default"], - "nn.conv2d_transpose": ["NHWC", "default"], - "nn.upsampling": ["NHWC", "default"], - "vision.roi_align": ["NHWC", "default"], - } - log.info("Use tuning file from %s: %s", ANSOR_TYPE, tuning_logfile) - with auto_scheduler.ApplyHistoryBest(tuning_logfile): # noqa: SIM117 - with tvm.transform.PassContext( - opt_level=opt_level, - config={ - "relay.backend.use_auto_scheduler": True, - "relay.FuseOps.max_depth": 30, - }, - ): - if nhwc: - seq = tvm.transform.Sequential( - [ - relay.transform.InferType(), - relay.transform.ConvertLayout(desired_layouts), - relay.transform.EliminateCommonSubexpr(), - relay.transform.FoldConstant(), - ] - ) - irmod = seq(irmod) - lib = get_tvm_executor(irmod, executor, tvm_target, params) - elif tuning_type == AUTO_TVM_TYPE: - with relay.build_config(opt_level=opt_level): - log.info("Use tuning file from %s: %s", AUTO_TVM_TYPE, tuning_logfile) - with autotvm.apply_history_best(tuning_logfile): - lib = get_tvm_executor(irmod, executor, tvm_target, params) - else: - log.error( - f"Tuning log type {tuning_type} is unsupported. " - f"Only {ANSOR_TYPE} and {AUTO_TVM_TYPE} types are supported" - ) - return None - else: - with tvm.transform.PassContext(opt_level=opt_level): - lib = get_tvm_executor(irmod, executor, tvm_target, params) - - if lib is None: - return None - - ctx = tvm.device(target, 0) - if executor == "vm": - m = tvm.runtime.vm.VirtualMachine(lib, ctx) - elif executor == "graph": - m = graph_executor.GraphModule(lib["default"](ctx)) - else: - print( - f"ERROR: Executor type {executor} is unsupported. ", - 'Only "vm" and "graph" types are supported', - ) - return None - - return m.module diff --git a/onnxruntime/test/framework/function_test.cc b/onnxruntime/test/framework/function_test.cc index fa3545ef27d72..180a75a64c10e 100644 --- a/onnxruntime/test/framework/function_test.cc +++ b/onnxruntime/test/framework/function_test.cc @@ -580,13 +580,7 @@ TEST(FunctionTest, TestInlinedLocalFunctionNotRemoved) { // myfun is not removed because it was claimed by InternalTestingEP model_proto = session_object.GetModel().ToProto(); -#ifdef USE_TVM - // TVM EP takes the whole graph and optimizes it within its own framework. - // It does not retain the original graph. - ASSERT_EQ(0, model_proto.functions_size()); -#else ASSERT_EQ(1, model_proto.functions_size()); -#endif } TEST(FunctionTest, TestInlinedFunctionDoesNotReserrectNonExistingArgs) { diff --git a/onnxruntime/test/platform/windows/stacktrace_test.cc b/onnxruntime/test/platform/windows/stacktrace_test.cc index de09dbcf270a9..9b1840f4b5d65 100644 --- a/onnxruntime/test/platform/windows/stacktrace_test.cc +++ b/onnxruntime/test/platform/windows/stacktrace_test.cc @@ -14,7 +14,6 @@ namespace onnxruntime { namespace test { using namespace ::testing; -// TVM is not working with StackTrace now. #if !defined(ORT_NO_EXCEPTIONS) TEST(StacktraceTests, BasicTests) { auto result = ::onnxruntime::GetStackTrace(); diff --git a/onnxruntime/test/python/onnxruntime_test_python_tvm.py b/onnxruntime/test/python/onnxruntime_test_python_tvm.py deleted file mode 100644 index 0080bf53520f2..0000000000000 --- a/onnxruntime/test/python/onnxruntime_test_python_tvm.py +++ /dev/null @@ -1,242 +0,0 @@ -# ------------------------------------------------------------------------- -# Copyright (c) Microsoft Corporation. All rights reserved. -# Licensed under the MIT License. See License.txt in the project root for -# license information. -# -------------------------------------------------------------------------- -""" -Module for unit testing of TVM EP -""" - -import os -import sys -import tempfile -import unittest -from typing import Any, AnyStr, Dict, List, Tuple - -import numpy -import tvm -from numpy.testing import assert_almost_equal -from onnx import ModelProto, TensorProto, mapping -from onnx.helper import make_graph, make_model, make_node, make_tensor_value_info - -import onnxruntime - -numpy.random.seed(32) - - -def is_windows(): - """ - Function to determine the Windows system - """ - return sys.platform.startswith("win") - - -def get_model_with_dynamic_shapes() -> ModelProto: - """ - Create model with Dynamic Shapes - """ - x = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) # pylint: disable=invalid-name, no-member - a = make_tensor_value_info("A", TensorProto.FLOAT, [None, None]) # pylint: disable=invalid-name, no-member - b = make_tensor_value_info("B", TensorProto.FLOAT, [None, None]) # pylint: disable=invalid-name, no-member - y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None]) # pylint: disable=invalid-name, no-member - node1 = make_node("MatMul", ["X", "A"], ["XA"]) - node2 = make_node("Add", ["XA", "B"], ["Y"]) - graph = make_graph([node1, node2], "lr", [x, a, b], [y]) - onnx_model = make_model(graph) - return onnx_model - - -def get_model_with_fixed_shapes() -> ModelProto: - """ - Create model with Static Shapes - """ - - def change_input_shape(model: ModelProto, ind: int, shape: Tuple) -> None: - """ - Function to change the input form - """ - dims = model.graph.input[ind].type.tensor_type.shape.dim - assert len(dims) == len(shape), "Input rank and new shape rank do not match." - for i, new_dim in enumerate(shape): - model.graph.input[ind].type.tensor_type.shape.dim[i].dim_value = new_dim - - dynamic_model = get_model_with_dynamic_shapes() - change_input_shape(dynamic_model, 0, (1, 2)) # X - change_input_shape(dynamic_model, 1, (2, 2)) # A - change_input_shape(dynamic_model, 2, (1, 2)) # B - return dynamic_model - - -def get_input_data_for_model_with_dynamic_shapes() -> Dict[AnyStr, numpy.ndarray]: - """ - Create input data for model with dynamic shapes - """ - a = numpy.random.randn(2, 2).astype(numpy.float32) # pylint: disable=invalid-name - b = numpy.random.randn(1, 2).astype(numpy.float32) # pylint: disable=invalid-name - x = numpy.random.randn(1, 2).astype(numpy.float32) # pylint: disable=invalid-name - data = {"A": a, "B": b, "X": x} - return data - - -def get_input_data_for_model_with_fixed_shapes(onnx_model: ModelProto) -> Dict[AnyStr, numpy.ndarray]: - """ - Create input data for model with static shapes - """ - - def get_onnx_input_names(model: ModelProto) -> List[AnyStr]: - inputs = [node.name for node in model.graph.input] - initializer = [node.name for node in model.graph.initializer] - inputs = list(set(inputs) - set(initializer)) - return sorted(inputs) - - def get_onnx_input_types(model: ModelProto) -> List[numpy.dtype]: - input_names = get_onnx_input_names(model) - return [ - mapping.TENSOR_TYPE_TO_NP_TYPE[node.type.tensor_type.elem_type] - for node in sorted(model.graph.input, key=lambda node: node.name) - if node.name in input_names - ] - - def get_onnx_input_shapes(model: ModelProto) -> List[List[int]]: - input_names = get_onnx_input_names(model) - return [ - [dv.dim_value for dv in node.type.tensor_type.shape.dim] - for node in sorted(model.graph.input, key=lambda node: node.name) - if node.name in input_names - ] - - input_names = get_onnx_input_names(onnx_model) - input_shapes = get_onnx_input_shapes(onnx_model) - input_types = get_onnx_input_types(onnx_model) - assert len(input_names) == len(input_types) == len(input_shapes) - random_inputs = [numpy.random.uniform(size=shape).astype(dtype) for shape, dtype in zip(input_shapes, input_types)] - return dict(zip(input_names, random_inputs)) - - -def get_input_names_and_shapes(data: Dict[AnyStr, numpy.ndarray]) -> Tuple[List[AnyStr], List[AnyStr]]: - """ - Create text representations for model input names and shapes - """ - keys = list(data.keys()) - values = [data[key] for key in keys] - return ( - list(data.keys()), - [str(value.shape).replace(",", "").replace("(", "[").replace(")", "]") for value in values], - ) - - -def get_cpu_output(onnx_model: ModelProto, data: Dict[AnyStr, numpy.ndarray]) -> List[numpy.ndarray]: - """ - Run inference with CPUExecutionProvider - """ - # pylint: disable=no-member - sess = onnxruntime.InferenceSession( - onnx_model.SerializeToString(), - providers=["CPUExecutionProvider"], - ) - output = sess.run(None, data) - return output - - -def get_tvm_output( - onnx_model: ModelProto, data: Dict[AnyStr, numpy.ndarray], provider_options: Dict[AnyStr, Any] -) -> List[numpy.ndarray]: - """ - Run inference with TVMExecutionProvider - """ - session_options = onnxruntime.SessionOptions() # pylint: disable=no-member - session_options.log_severity_level = 0 - session_options.log_verbosity_level = 0 - # pylint: disable=no-member - session_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_DISABLE_ALL - - sess = onnxruntime.InferenceSession( - onnx_model.SerializeToString(), - session_options, - providers=["TvmExecutionProvider"], - provider_options=[provider_options], - ) - - output = sess.run(None, data) - return output - - -# pylint: disable=no-member -def compile_virtual_machine(model: ModelProto, target_str: AnyStr) -> tvm.runtime.vm.Executable: - """ - Compile ONNX model using VirtualMachine - """ - ir_mod, _ = tvm.relay.frontend.from_onnx( - model, - opset=model.opset_import[0].version, - freeze_params=True, - ) - target = tvm.target.Target(target=target_str, host=target_str) - return tvm.relay.backend.vm.compile(ir_mod, target) - - -def serialize_virtual_machine(vm_exec: tvm.runtime.vm.Executable) -> AnyStr: - """ - Serialize VirtualMachine - """ - temp_directory = tempfile.mkdtemp() - path_consts = os.path.join(temp_directory, "consts") - vm_exec.move_late_bound_consts(path_consts, byte_limit=256) - lib_path = os.path.join(temp_directory, f"model.{'dll' if is_windows() else 'so'}") - code_path = os.path.join(temp_directory, "model.ro") - code, lib = vm_exec.save() - lib.export_library(lib_path) - with open(code_path, "wb") as code_file: - code_file.write(code) - return temp_directory - - -class TestTVM(unittest.TestCase): - """ - Unit tests for TVM EP - """ - - @staticmethod - def test_accuracy_for_model_with_dynamic_shapes(): - """ - Accuracy test for model with dynamic shapes - """ - onnx_model = get_model_with_dynamic_shapes() - data = get_input_data_for_model_with_dynamic_shapes() - - cpu_output = get_cpu_output(onnx_model, data) - names, shapes = get_input_names_and_shapes(data) - provider_options = dict( - target="llvm", - input_names=" ".join(names), - input_shapes=" ".join(shapes), - ) - tvm_output = get_tvm_output(onnx_model, data, provider_options) - - assert_almost_equal(cpu_output, tvm_output, decimal=5) - - @staticmethod - def test_accuracy_for_tvm_so(): - """ - Accuracy test for TVMso Ep - """ - onnx_model = get_model_with_fixed_shapes() - data = get_input_data_for_model_with_fixed_shapes(onnx_model) - - cpu_output = get_cpu_output(onnx_model, data) - - compiled_vm_exec = compile_virtual_machine(onnx_model, target_str="llvm") - so_folder = serialize_virtual_machine(compiled_vm_exec) - provider_options = dict( - target="llvm", - so_folder=so_folder, - ) - tvm_output = get_tvm_output(onnx_model, data, provider_options) - - assert_almost_equal(cpu_output, tvm_output, decimal=5) - - -if __name__ == "__main__": - if "TvmExecutionProvider" not in onnxruntime.get_available_providers(): - raise AssertionError(f"Unable to find 'TvmExecutionProvider' in {onnxruntime.get_available_providers()}") - unittest.main() diff --git a/onnxruntime/test/util/default_providers.cc b/onnxruntime/test/util/default_providers.cc index 62bdedd833025..821f32d462d16 100644 --- a/onnxruntime/test/util/default_providers.cc +++ b/onnxruntime/test/util/default_providers.cc @@ -182,14 +182,6 @@ std::unique_ptr DnnlExecutionProviderWithOptions(const OrtDn return nullptr; } -// std::unique_ptr DefaultTvmExecutionProvider() { -// #ifdef USE_TVM -// return TVMProviderFactoryCreator::Create("")->CreateProvider(); -// #else -// return nullptr; -// #endif -// } - std::unique_ptr DefaultNnapiExecutionProvider() { // The NNAPI EP uses a stub implementation on non-Android platforms so cannot be used to execute a model. // Manually append an NNAPI EP instance to the session to unit test the GetCapability and Compile implementation. diff --git a/onnxruntime/test/util/include/default_providers.h b/onnxruntime/test/util/include/default_providers.h index ed95bf67f1ffb..9b44150d972db 100644 --- a/onnxruntime/test/util/include/default_providers.h +++ b/onnxruntime/test/util/include/default_providers.h @@ -20,7 +20,6 @@ std::shared_ptr CreateExecutionProviderFactory_Dnnl(c std::shared_ptr CreateExecutionProviderFactory_MIGraphX(const OrtMIGraphXProviderOptions* params); std::shared_ptr CreateExecutionProviderFactory_Nnapi( uint32_t flags, const optional& partitioning_stop_ops_list); -// std::shared_ptr CreateExecutionProviderFactory_Tvm(const char*); std::shared_ptr CreateExecutionProviderFactory_VSINPU(); std::shared_ptr CreateExecutionProviderFactory_Rknpu(); std::shared_ptr CreateExecutionProviderFactory_Rocm(const OrtROCMProviderOptions* provider_options); diff --git a/onnxruntime/test/util/include/providers.h b/onnxruntime/test/util/include/providers.h index a73b237ae10df..01be1a444646b 100644 --- a/onnxruntime/test/util/include/providers.h +++ b/onnxruntime/test/util/include/providers.h @@ -7,9 +7,6 @@ #ifdef USE_DNNL #include "core/providers/dnnl/dnnl_provider_factory.h" #endif -#ifdef USE_TVM -#include "core/providers/tvm/tvm_provider_factory.h" -#endif #ifdef USE_OPENVINO #include "core/providers/openvino/openvino_provider_factory.h" #endif diff --git a/tools/ci_build/build.py b/tools/ci_build/build.py index 24dc6124d4a89..aa1198102f978 100644 --- a/tools/ci_build/build.py +++ b/tools/ci_build/build.py @@ -584,13 +584,7 @@ def convert_arg_line_to_args(self, arg_line): parser.add_argument("--use_preinstalled_eigen", action="store_true", help="Use pre-installed Eigen.") parser.add_argument("--eigen_path", help="Path to pre-installed Eigen.") parser.add_argument("--enable_msinternal", action="store_true", help="Enable for Microsoft internal builds only.") - parser.add_argument("--llvm_path", help="Path to llvm dir") parser.add_argument("--use_vitisai", action="store_true", help="Build with Vitis-AI") - parser.add_argument("--use_tvm", action="store_true", help="Build with TVM") - parser.add_argument("--tvm_cuda_runtime", action="store_true", default=False, help="Build TVM with CUDA support") - parser.add_argument( - "--use_tvm_hash", action="store_true", help="Build ipp-crypto for hash generation. It is used by TVM EP only" - ) parser.add_argument("--use_tensorrt", action="store_true", help="Build with TensorRT") parser.add_argument( "--use_tensorrt_builtin_parser", action="store_true", default=True, help="Use TensorRT builtin parser" @@ -602,12 +596,6 @@ def convert_arg_line_to_args(self, arg_line): parser.add_argument("--migraphx_home", help="Path to MIGraphX installation dir") parser.add_argument("--use_full_protobuf", action="store_true", help="Use the full protobuf library") - parser.add_argument( - "--llvm_config", - type=str, - default="", - help="Path to llvm-config.exe for LLVM built from sources. It is strongly needed for build on Windows", - ) parser.add_argument( "--skip_onnx_tests", action="store_true", @@ -1031,16 +1019,11 @@ def generate_build_tree( "-Donnxruntime_USE_NNAPI_BUILTIN=" + ("ON" if args.use_nnapi else "OFF"), "-Donnxruntime_USE_VSINPU=" + ("ON" if args.use_vsinpu else "OFF"), "-Donnxruntime_USE_RKNPU=" + ("ON" if args.use_rknpu else "OFF"), - "-Donnxruntime_USE_LLVM=" + ("ON" if args.use_tvm else "OFF"), "-Donnxruntime_ENABLE_MICROSOFT_INTERNAL=" + ("ON" if args.enable_msinternal else "OFF"), "-Donnxruntime_USE_VITISAI=" + ("ON" if args.use_vitisai else "OFF"), "-Donnxruntime_USE_TENSORRT=" + ("ON" if args.use_tensorrt else "OFF"), "-Donnxruntime_USE_TENSORRT_BUILTIN_PARSER=" + ("ON" if args.use_tensorrt_builtin_parser and not args.use_tensorrt_oss_parser else "OFF"), - # set vars for TVM - "-Donnxruntime_USE_TVM=" + ("ON" if args.use_tvm else "OFF"), - "-Donnxruntime_TVM_CUDA_RUNTIME=" + ("ON" if args.use_tvm and args.tvm_cuda_runtime else "OFF"), - "-Donnxruntime_TVM_USE_HASH=" + ("ON" if args.use_tvm_hash else "OFF"), # set vars for migraphx "-Donnxruntime_USE_MIGRAPHX=" + ("ON" if args.use_migraphx else "OFF"), "-Donnxruntime_DISABLE_CONTRIB_OPS=" + ("ON" if args.disable_contrib_ops else "OFF"), @@ -1172,8 +1155,6 @@ def generate_build_tree( cmake_args.append("-Donnxruntime_ROCM_VERSION=" + args.rocm_version) if args.use_tensorrt: cmake_args.append("-Donnxruntime_TENSORRT_HOME=" + tensorrt_home) - if args.llvm_config: - cmake_args.append("-Donnxruntime_TVM_USE_LLVM=" + args.llvm_config) if args.use_cuda: add_default_definition(cmake_extra_defines, "onnxruntime_USE_CUDA", "ON") @@ -1256,9 +1237,6 @@ def generate_build_tree( if args.use_full_protobuf or args.use_openvino or args.use_vitisai or args.gen_doc: cmake_args += ["-Donnxruntime_USE_FULL_PROTOBUF=ON", "-DProtobuf_USE_STATIC_LIBS=ON"] - if args.use_tvm and args.llvm_path is not None: - cmake_args += [f"-DLLVM_DIR={args.llvm_path}"] - if args.use_cuda and not is_windows(): nvml_stub_path = cuda_home + "/lib64/stubs" cmake_args += ["-DCUDA_CUDA_LIBRARY=" + nvml_stub_path] @@ -1659,16 +1637,6 @@ def generate_build_tree( cxxflags = cflags.copy() config_build_dir = get_config_build_dir(build_dir, config) os.makedirs(config_build_dir, exist_ok=True) - if args.use_tvm: - os.environ["PATH"] = ( - os.path.join(config_build_dir, "_deps", "tvm-build") - + os.pathsep - + os.path.join(config_build_dir, "_deps", "tvm-src") - + os.pathsep - + os.path.dirname(sys.executable) - + os.pathsep - + os.environ["PATH"] - ) preinstalled_dir = Path(build_dir) / config temp_cmake_args = cmake_args.copy() if cflags is not None and cxxflags is not None and len(cflags) != 0 and len(cxxflags) != 0: @@ -2097,8 +2065,6 @@ def run_onnxruntime_tests(args, source_dir, ctest_path, build_dir, configs): if args.enable_pybind: python_path = None - if args.use_tvm: - python_path = str((Path(build_dir) / config / "_deps" / "tvm-src" / "python").resolve()) # Disable python tests in a reduced build as we don't know which ops have been included and which # models can run. @@ -2221,17 +2187,6 @@ def run_onnxruntime_tests(args, source_dir, ctest_path, build_dir, configs): run_subprocess([sys.executable, "onnxruntime_test_python_keras.py"], cwd=cwd, dll_path=dll_path) -def tvm_run_python_tests(build_dir, configs): - for config in configs: - cwd = get_config_build_dir(build_dir, config) - if is_windows(): - cwd = os.path.join(cwd, config) - python_path = os.path.join(build_dir, config, "_deps", "tvm-src", "python") - run_subprocess( - [sys.executable, "onnxruntime_test_python_tvm.py"], cwd=cwd, python_path=os.path.abspath(python_path) - ) - - def run_nodejs_tests(nodejs_binding_dir): args = ["npm", "test", "--", "--timeout=90000"] if is_windows(): @@ -2251,7 +2206,6 @@ def build_python_wheel( use_dnnl, use_tensorrt, use_openvino, - use_tvm, use_vitisai, use_acl, use_armnn, @@ -2304,8 +2258,6 @@ def build_python_wheel( args.append("--use_openvino") elif use_dnnl: args.append("--use_dnnl") - elif use_tvm: - args.append("--use_tvm") elif use_vitisai: args.append("--use_vitisai") elif use_acl: @@ -2334,7 +2286,6 @@ def build_nuget_package( use_openvino, use_tensorrt, use_dnnl, - use_tvm, use_winml, use_qnn, enable_training_apis, @@ -2381,9 +2332,6 @@ def build_nuget_package( package_name = "/p:OrtPackageId=Microsoft.ML.OnnxRuntime.Gpu" elif use_rocm: package_name = "/p:OrtPackageId=Microsoft.ML.OnnxRuntime.ROCm" - elif use_tvm: - execution_provider = "/p:ExecutionProvider=tvm" - package_name = "/p:OrtPackageId=Microsoft.ML.OnnxRuntime.Tvm" elif use_qnn: execution_provider = "/p:ExecutionProvider=qnn" package_name = "/p:OrtPackageId=Microsoft.ML.OnnxRuntime.QNN" @@ -2625,7 +2573,7 @@ def main(): if args.use_tensorrt: args.use_cuda = True - if args.build_wheel or args.gen_doc or args.use_tvm or args.enable_training: + if args.build_wheel or args.gen_doc or args.enable_training: args.enable_pybind = True if ( @@ -2907,12 +2855,6 @@ def main(): run_onnxruntime_tests(args, source_dir, ctest_path, build_dir, configs) - # TODO(agladyshev): - # to support Windows, we need to update .github/workflows/windows.yml - # and add to the PATH variable the following value: C:Program Files\LLVM\bin - if args.enable_pybind and args.use_tvm and not is_windows(): - tvm_run_python_tests(build_dir, configs) - # run node.js binding tests if args.build_nodejs and not args.skip_nodejs_tests: nodejs_binding_dir = os.path.normpath(os.path.join(source_dir, "js", "node")) @@ -2940,7 +2882,6 @@ def main(): args.use_dnnl, args.use_tensorrt, args.use_openvino, - args.use_tvm, args.use_vitisai, args.use_acl, args.use_armnn, @@ -2968,7 +2909,6 @@ def main(): args.use_openvino, args.use_tensorrt, args.use_dnnl, - args.use_tvm, args.use_winml, args.use_qnn, args.enable_training_apis, diff --git a/tools/ci_build/github/azure-pipelines/templates/orttraining-linux-gpu-test-ci-pipeline.yml b/tools/ci_build/github/azure-pipelines/templates/orttraining-linux-gpu-test-ci-pipeline.yml deleted file mode 100644 index 5f073433265fa..0000000000000 --- a/tools/ci_build/github/azure-pipelines/templates/orttraining-linux-gpu-test-ci-pipeline.yml +++ /dev/null @@ -1,41 +0,0 @@ -parameters: -- name: DockerImageTag - type: string -- name: BuildConfig - type: string - -steps: - -- template: jobs/download_training_test_data.yml - - # Entry point for all ORTModule tests - # The onnxruntime folder is deleted in the build directory - # to enforce use of the onnxruntime wheel - # Uninstall orttraining requirements.txt and install ortmodule requirements.txt before running tests. -- script: | - docker run \ - --gpus all \ - --shm-size=1024m \ - --rm \ - --volume $(Build.SourcesDirectory):/onnxruntime_src \ - --volume $(Build.BinariesDirectory)/${{ parameters.BuildConfig }}:/build \ - --volume $(Agent.TempDirectory)/mnist:/mnist \ - ${{ parameters.DockerImageTag }} \ - bash -c "rm -rf /build/onnxruntime/ && python3 -m pip show torch && python3 -m pip install torch==2.3.1+cu118 --index-url https://download.pytorch.org/whl/cu118 && python3 -m pip install /build/dist/onnxruntime*.whl && python3 -m onnxruntime.training.ortmodule.torch_cpp_extensions.install && /build/launch_test.py --cmd_line_with_args 'python orttraining_ortmodule_tests.py --mnist /mnist --bert_data /bert_data/hf_data/glue_data/CoLA/original/raw' --cwd /build" \ - displayName: 'Run orttraining_ortmodule_tests.py' - condition: succeededOrFailed() - timeoutInMinutes: 60 - -# Entry point for all ort training api tests -- script: | - docker run \ - --gpus all \ - --shm-size=1024m \ - --rm \ - --volume $(Build.SourcesDirectory):/onnxruntime_src \ - --volume $(Build.BinariesDirectory)/${{ parameters.BuildConfig }}:/build \ - ${{ parameters.DockerImageTag }} \ - bash -c "rm -rf /build/onnxruntime/ && python3 -m pip install /build/dist/onnxruntime*.whl && python3 -m pip install torch==2.3.1+cu118 --index-url https://download.pytorch.org/whl/cu118 && /build/launch_test.py --cmd_line_with_args 'python orttraining_test_ort_apis.py --cwd /build' --cwd /build" \ - displayName: 'Run ORT Training APIs Tests' - condition: succeededOrFailed() - timeoutInMinutes: 120 diff --git a/tools/ci_build/github/azure-pipelines/templates/py-packaging-training-cuda-stage-steps.yml b/tools/ci_build/github/azure-pipelines/templates/py-packaging-training-cuda-stage-steps.yml deleted file mode 100644 index fc163d17e44a9..0000000000000 --- a/tools/ci_build/github/azure-pipelines/templates/py-packaging-training-cuda-stage-steps.yml +++ /dev/null @@ -1,209 +0,0 @@ -parameters: - build_py_parameters: '' - torch_version: '' - opset_version: '' - cuda_version: '' - cmake_cuda_architectures: '' - docker_file: '' - upload_wheel: '' - debug_build: '' - python_version: '' - stage_name: '' - SpecificArtifact: false - BuildId: '0' - build_pool_name: '' - -stages: - - stage: Build_${{ parameters.stage_name }} - variables: - - name: isMain - value: ${{ or(eq(variables['Build.SourceBranch'], 'refs/heads/main'), startsWith(variables['Build.SourceBranch'], 'refs/heads/rel-')) }} - - name: finalStorage - ${{ if eq(variables['isMain'], 'true') }}: - value: '--final_storage' - ${{ else }}: - value: '' - - name: buildConfig - ${{ if eq(parameters['debug_build'], 'true') }}: - value: 'Debug' - ${{ else }}: - value: 'Release' - - name: PythonVersion - value: ${{ parameters.python_version }} - - name: Repository - value: onnxruntimetraininggpubuild_cu${{ replace(parameters.cuda_version, '.', '') }}_py${{ replace(parameters.python_version, '.', '') }} - dependsOn: [] - - jobs: - - job: Build - pool: ${{ parameters.build_pool_name }} - timeoutInMinutes: 180 - steps: - - task: mspremier.PostBuildCleanup.PostBuildCleanup-task.PostBuildCleanup@3 - displayName: 'Clean Agent Directories' - condition: always() - - - task: CmdLine@2 - displayName: 'check variables' - inputs: - script: | - echo "Branch is "${{ variables['Build.SourceBranch'] }} && \ - echo "isMain is "${{ variables['isMain'] }} && \ - echo "final_storage is "${{ variables['finalStorage'] }} - - - checkout: self - clean: true - submodules: recursive - - - template: set-python-manylinux-variables-step.yml - - - template: get-docker-image-steps.yml - parameters: - Dockerfile: tools/ci_build/github/linux/docker/${{ parameters.docker_file }} - Context: tools/ci_build/github/linux/docker - DockerBuildArgs: >- - --build-arg TORCH_VERSION=${{ parameters.torch_version }} - --build-arg OPSET_VERSION=${{ parameters.opset_version }} - --build-arg PYTHON_VERSION=${{ parameters.python_version }} - --build-arg INSTALL_DEPS_EXTRA_ARGS=-tu - --build-arg BUILD_UID=$(id -u) - Repository: $(Repository) - - - task: CmdLine@2 - displayName: 'build onnxruntime' - inputs: - script: | - set -e -x - mkdir -p $HOME/.onnx - docker run --rm \ - --volume /data/onnx:/data/onnx:ro \ - --volume $(Build.SourcesDirectory):/onnxruntime_src \ - --volume $(Build.BinariesDirectory):/build \ - --volume /data/models:/build/models:ro \ - --volume $HOME/.onnx:/home/onnxruntimedev/.onnx \ - -e NIGHTLY_BUILD \ - -e DEFAULT_TRAINING_PACKAGE_DEVICE \ - -e BUILD_BUILDNUMBER \ - -e ORT_DISABLE_PYTHON_PACKAGE_LOCAL_VERSION \ - $(Repository) \ - $(PythonManylinuxDir)/bin/python3 /onnxruntime_src/tools/ci_build/build.py \ - --build_dir /build \ - --config ${{ variables['buildConfig'] }} \ - --skip_submodule_sync \ - --parallel --use_binskim_compliant_compile_flags \ - --build_wheel \ - --enable_onnx_tests \ - ${{ parameters.build_py_parameters }} \ - --cmake_extra_defines 'CMAKE_CUDA_ARCHITECTURES=${{ parameters.cmake_cuda_architectures }}' onnxruntime_BUILD_UNIT_TESTS=OFF \ - --use_cuda --cuda_version=${{ parameters.cuda_version }} --cuda_home=/usr/local/cuda-${{ parameters.cuda_version }} --cudnn_home=/usr/local/cuda-${{ parameters.cuda_version }}; - workingDirectory: $(Build.SourcesDirectory) - - - task: CopyFiles@2 - displayName: 'Copy Python Wheel to: $(Build.ArtifactStagingDirectory)' - inputs: - SourceFolder: '$(Build.BinariesDirectory)' - Contents: "${{ variables['buildConfig'] }}/dist/*.whl" - TargetFolder: '$(Build.ArtifactStagingDirectory)' - - - task: PublishBuildArtifacts@1 - displayName: 'Publish Artifact: ONNXRuntime python wheel and documentation' - inputs: - ArtifactName: "onnxruntime_gpu_${{ variables['buildConfig'] }}_${{ parameters.python_version }}" - - - template: component-governance-component-detection-steps.yml - parameters: - condition: 'succeeded' - - - template: clean-agent-build-directory-step.yml - - - stage: Test_${{ parameters.stage_name }} - variables: - - name: isMain - value: ${{ or(eq(variables['Build.SourceBranch'], 'refs/heads/main'), startsWith(variables['Build.SourceBranch'], 'refs/heads/rel-')) }} - - name: finalStorage - ${{ if eq(variables['isMain'], 'true') }}: - value: '--final_storage' - ${{ else }}: - value: '' - - name: buildConfig - ${{ if eq(parameters['debug_build'], 'true') }}: - value: 'Debug' - ${{ else }}: - value: 'Release' - - name: PythonVersion - value: ${{ parameters.python_version }} - - name: Repository - value: onnxruntimetraininggpubuild_cu${{ replace(parameters.cuda_version, '.', '') }}_py${{ replace(parameters.python_version, '.', '') }} - - name: UploadWheel - value: ${{ parameters.upload_wheel }} - dependsOn: Build_${{ parameters.stage_name }} - jobs: - - job: Test_GPU - pool: Onnxruntime-Linux-GPU - steps: - - task: mspremier.PostBuildCleanup.PostBuildCleanup-task.PostBuildCleanup@3 - displayName: 'Clean Agent Directories' - condition: always() - - - checkout: self - clean: true - submodules: none - - - template: jobs/download_training_test_data.yml - - - template: set-python-manylinux-variables-step.yml - - - template: flex-downloadPipelineArtifact.yml - parameters: - ArtifactName: "onnxruntime_gpu_${{ variables['buildConfig'] }}_${{ parameters.python_version }}" - StepName: 'Download Pipeline Artifact - Linux Training Build' - TargetPath: '$(Build.ArtifactStagingDirectory)' - SpecificArtifact: ${{ parameters.SpecificArtifact }} - BuildId: ${{ parameters.BuildId }} - - - script: | - set -e -x - whlfilename=$(ls $(Build.ArtifactStagingDirectory)/Release/dist/*.whl | head -n 1) ; \ - echo $whlfilename ; du -sh $whlfilename ; \ - (( $(wc -c < "$whlfilename") - 400*1024*1024 < 0 )) || ( echo 'Wheel size bigger than 400M'; exit 1) - displayName: 'Check wheel size' - continueOnError: true - - - template: get-docker-image-steps.yml - parameters: - Dockerfile: tools/ci_build/github/linux/docker/${{ parameters.docker_file }} - Context: tools/ci_build/github/linux/docker - DockerBuildArgs: >- - --build-arg TORCH_VERSION=${{ parameters.torch_version }} - --build-arg OPSET_VERSION=${{ parameters.opset_version }} - --build-arg PYTHON_VERSION=${{ parameters.python_version }} - --build-arg INSTALL_DEPS_EXTRA_ARGS=-tu - --build-arg BUILD_UID=$(id -u) - Repository: $(Repository) - - - task: CmdLine@2 - displayName: 'test ortmodule' - inputs: - script: | - set -ex ; \ - whlfilename=$(ls $(Build.ArtifactStagingDirectory)/Release/dist/*.whl | head -n 1) ; \ - echo $whlfilename ; \ - basefilename=$(basename $whlfilename) ; \ - docker run --rm \ - --gpus all \ - -e NVIDIA_VISIBLE_DEVICES=all \ - --volume $(Build.ArtifactStagingDirectory):/build \ - --volume $(Agent.TempDirectory)/MNIST:/mnist \ - $(Repository) \ - bash -c " $(PythonManylinuxDir)/bin/python3 -m pip install /build/Release/dist/$basefilename && $(PythonManylinuxDir)/bin/python3 -m onnxruntime.training.ortmodule.torch_cpp_extensions.install " ; - workingDirectory: $(Build.SourcesDirectory) - - - task: CmdLine@2 - displayName: 'Upload wheel' - condition: and(succeeded(), and(eq(variables['UploadWheel'], 'yes'), ne(variables['ORT_DISABLE_PYTHON_PACKAGE_LOCAL_VERSION'], 'true'))) - inputs: - script: | - set -e -x - whlfilename=$(ls $(Build.ArtifactStagingDirectory)/Release/dist/*.whl | head -n 1) ; \ - python3 tools/ci_build/upload_python_package_to_azure_storage.py \ - --python_wheel_path $whlfilename ${{ variables['finalStorage'] }} diff --git a/tools/ci_build/github/linux/docker/scripts/install_os_deps.sh b/tools/ci_build/github/linux/docker/scripts/install_os_deps.sh index 7f3160371aa24..87b9b960b7ebc 100755 --- a/tools/ci_build/github/linux/docker/scripts/install_os_deps.sh +++ b/tools/ci_build/github/linux/docker/scripts/install_os_deps.sh @@ -12,7 +12,6 @@ d) DEVICE_TYPE=${OPTARG};; v) echo "Cuda version is no longer accepted as an input to this script. Ignoring the input argument -v.";; t) echo "Installing python training dependencies argument is no longer accepted as an input to this script. Ignoring the input argument -t.";; m) INSTALL_DEPS_DISTRIBUTED_SETUP=true;; -u) echo "Installing ortmodule python dependencies argument is no longer accepted as an input to this script. Ignoring the input argument -u.";; r) echo "Installing ROCM python dependencies argument is no longer accepted as an input to this script. Ignoring the input argument -r.";; esac done diff --git a/tools/ci_build/github/linux/docker/scripts/install_python_deps.sh b/tools/ci_build/github/linux/docker/scripts/install_python_deps.sh index 1ac1d226deec6..2d7acd1f701ff 100755 --- a/tools/ci_build/github/linux/docker/scripts/install_python_deps.sh +++ b/tools/ci_build/github/linux/docker/scripts/install_python_deps.sh @@ -3,7 +3,6 @@ set -e -x INSTALL_DEPS_TRAINING=false INSTALL_DEPS_DISTRIBUTED_SETUP=false -ORTMODULE_BUILD=false TARGET_ROCM=false CU_VER="11.8" TORCH_VERSION='2.0.0' @@ -18,7 +17,6 @@ d) DEVICE_TYPE=${OPTARG};; v) CU_VER=${OPTARG};; t) INSTALL_DEPS_TRAINING=true;; m) INSTALL_DEPS_DISTRIBUTED_SETUP=true;; -u) ORTMODULE_BUILD=true;; r) TARGET_ROCM=true;; c) USE_CONDA=true;; esac @@ -55,17 +53,3 @@ fi export ONNX_ML=1 export CMAKE_ARGS="-DONNX_GEN_PB_TYPE_STUBS=OFF -DONNX_WERROR=OFF" ${PYTHON_EXE} -m pip install -r ${0/%install_python_deps\.sh/requirements\.txt} -if [ $DEVICE_TYPE = "gpu" ]; then - if [[ $INSTALL_DEPS_TRAINING = true ]]; then - if [[ $ORTMODULE_BUILD = false ]]; then - ${PYTHON_EXE} -m pip install -r ${0/%install_python_deps.sh/training\/requirements.txt} - else - if [[ $TARGET_ROCM = false ]]; then - ${PYTHON_EXE} -m pip install -r ${0/%install_python_deps.sh/training\/ortmodule\/stage1\/requirements_torch${TORCH_VERSION}_cu${CU_VER}\/requirements.txt} - ${PYTHON_EXE} -m pip install -r ${0/%install_python_deps.sh/training\/ortmodule\/stage2\/requirements.txt} - else - ${PYTHON_EXE} -m pip install -r ${0/%install_python_deps.sh/training\/ortmodule\/stage1\/requirements_rocm\/requirements.txt} - fi - fi - fi -fi diff --git a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_rocm/requirements.txt b/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_rocm/requirements.txt deleted file mode 100644 index 051f42dac335d..0000000000000 --- a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_rocm/requirements.txt +++ /dev/null @@ -1,2 +0,0 @@ -numpy==1.21.6 ; python_version < '3.9' -numpy==2.1.2 ; python_version >= '3.9' diff --git a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch2.0.0_cu11.8/requirements.txt b/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch2.0.0_cu11.8/requirements.txt deleted file mode 100644 index b3b2651c8d26d..0000000000000 --- a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch2.0.0_cu11.8/requirements.txt +++ /dev/null @@ -1,8 +0,0 @@ ---pre --f https://download.pytorch.org/whl/torch_stable.html -torch==2.0.0+cu118 -torchvision==0.15.1+cu118 -torchtext==0.15.1 -# TODO(bmeswani): packaging 22.0 removes support for LegacyVersion leading to errors because transformers 4.4.2 uses LegacyVersion -packaging==21.3 -setuptools>=68.2.2 diff --git a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch2.1.0_cu12.2/requirements.txt b/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch2.1.0_cu12.2/requirements.txt deleted file mode 100644 index 152a17db90366..0000000000000 --- a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch2.1.0_cu12.2/requirements.txt +++ /dev/null @@ -1,7 +0,0 @@ ---pre --f https://download.pytorch.org/whl/torch_stable.html -torch==2.1.0+cu121 -torchvision==0.16.0+cu121 -torchtext==0.16.0 -packaging==23.1 -setuptools>=68.2.2 diff --git a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch_cpu/requirements.txt b/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch_cpu/requirements.txt deleted file mode 100644 index 846f8c15b257d..0000000000000 --- a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch_cpu/requirements.txt +++ /dev/null @@ -1,3 +0,0 @@ --f https://download.pytorch.org/whl/torch_stable.html -torch==2.3.0+cpu -setuptools>=68.2.2 diff --git a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch_nightly/requirements.txt b/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch_nightly/requirements.txt deleted file mode 100644 index 01fa7b0ff956e..0000000000000 --- a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch_nightly/requirements.txt +++ /dev/null @@ -1,5 +0,0 @@ -scikit-learn -packaging==21.3 -transformers==v4.36.0 -accelerate==0.25.0 -wget diff --git a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/torch_eager_cpu/requirements.txt b/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/torch_eager_cpu/requirements.txt deleted file mode 100644 index 6346c54decf9c..0000000000000 --- a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/torch_eager_cpu/requirements.txt +++ /dev/null @@ -1,11 +0,0 @@ ---pre --f https://download.pytorch.org/whl/torch_stable.html -torch==2.2.0 -setuptools>=68.2.2 -cerberus -h5py -scikit-learn -numpy==1.21.6 ; python_version < '3.9' -numpy==2.1.2 ; python_version >= '3.9' -pandas -parameterized diff --git a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage2/requirements.txt b/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage2/requirements.txt deleted file mode 100644 index dd86b32f88c76..0000000000000 --- a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage2/requirements.txt +++ /dev/null @@ -1,15 +0,0 @@ -pandas -scikit-learn -numpy==1.21.6 ; python_version < '3.9' -numpy==2.1.2 ; python_version >= '3.9' -transformers==v4.36.0 -accelerate==0.25.0 -rsa==4.9 -tensorboard==2.13.0 -h5py -wget -pytorch-lightning==2.3.3 -deepspeed==0.9.0 -fairscale==0.4.6 -parameterized>=0.8.1 -pydantic<2.0.0 diff --git a/tools/ci_build/github/linux/run_dockerbuild.sh b/tools/ci_build/github/linux/run_dockerbuild.sh index 9944861f519f4..6618810c77f6d 100755 --- a/tools/ci_build/github/linux/run_dockerbuild.sh +++ b/tools/ci_build/github/linux/run_dockerbuild.sh @@ -15,10 +15,6 @@ BUILD_DIR=$BUILD_BINARIESDIRECTORY YOCTO_VERSION="4.19" #Training only INSTALL_DEPS_DISTRIBUTED_SETUP=false -#Training only -ORTMODULE_BUILD=false -#Training only -USE_CONDA=false ALLOW_RELEASED_ONNX_OPSET_ONLY_ENV="ALLOW_RELEASED_ONNX_OPSET_ONLY="$ALLOW_RELEASED_ONNX_OPSET_ONLY echo "ALLOW_RELEASED_ONNX_OPSET_ONLY environment variable is set as $ALLOW_RELEASED_ONNX_OPSET_ONLY_ENV" @@ -44,10 +40,6 @@ t) EXTRA_IMAGE_TAG=${OPTARG};; i) IMAGE_CACHE_CONTAINER_REGISTRY_NAME=${OPTARG};; # install distributed setup dependencies m) INSTALL_DEPS_DISTRIBUTED_SETUP=true;; -# install ortmodule specific dependencies -u) ORTMODULE_BUILD=true;; -# install and use conda -e) USE_CONDA=true;; *) echo "Invalid option";; esac done @@ -82,24 +74,6 @@ if [ $BUILD_OS = "yocto" ]; then $GET_DOCKER_IMAGE_CMD --repository "onnxruntime-$IMAGE" \ --docker-build-args="--build-arg TOOL_CHAIN=$TOOL_CHAIN_SCRIPT --build-arg BUILD_USER=onnxruntimedev --build-arg BUILD_UID=$(id -u) --build-arg PYTHON_VERSION=${PYTHON_VER}" \ --dockerfile $DOCKER_FILE --context . -elif [ $BUILD_DEVICE = "gpu" ]; then - # This code path is only for training. Inferecing pipeline uses CentOS - IMAGE="$BUILD_OS-gpu_training" - # Current build script doesn't support building shared lib with Python dependency. To enable building with PythonOp, - # We need to avoid `--no-undefined` when building shared lib (Otherwise, CIs will report `undefined symbols`), but removing that would bring some other concerns. - # Plus the fact training did not need build shared library, we disable the --build_shared_lib for training CIs. - NEED_BUILD_SHARED_LIB=false - INSTALL_DEPS_EXTRA_ARGS="${INSTALL_DEPS_EXTRA_ARGS} -t" - if [[ $INSTALL_DEPS_DISTRIBUTED_SETUP = true ]]; then - INSTALL_DEPS_EXTRA_ARGS="${INSTALL_DEPS_EXTRA_ARGS} -m" - fi - if [[ $ORTMODULE_BUILD = true ]]; then - INSTALL_DEPS_EXTRA_ARGS="${INSTALL_DEPS_EXTRA_ARGS} -u" - fi - INSTALL_DEPS_EXTRA_ARGS="${INSTALL_DEPS_EXTRA_ARGS} -v 11.8" - $GET_DOCKER_IMAGE_CMD --repository "onnxruntime-$IMAGE" \ - --docker-build-args="--build-arg BASEIMAGE=nvcr.io/nvidia/cuda:11.8.0-cudnn8-devel-${BUILD_OS} --build-arg BUILD_USER=onnxruntimedev --build-arg BUILD_UID=$(id -u) --build-arg PYTHON_VERSION=${PYTHON_VER} --build-arg INSTALL_DEPS_EXTRA_ARGS=\"${INSTALL_DEPS_EXTRA_ARGS}\" --build-arg USE_CONDA=${USE_CONDA} --network=host" \ - --dockerfile Dockerfile.ubuntu_gpu_training --context . elif [[ $BUILD_DEVICE = "openvino"* ]]; then BUILD_ARGS="--build-arg BUILD_USER=onnxruntimedev --build-arg BUILD_UID=$(id -u) --build-arg PYTHON_VERSION=${PYTHON_VER} --build-arg OPENVINO_VERSION=${OPENVINO_VERSION} --build-arg UBUNTU_VERSION=${UBUNTU_VERSION}" IMAGE="$BUILD_OS-openvino" diff --git a/tools/ci_build/github/pai/pai_huggingface_bert_large_test.sh b/tools/ci_build/github/pai/pai_huggingface_bert_large_test.sh deleted file mode 100755 index fb4dbeb2e73d3..0000000000000 --- a/tools/ci_build/github/pai/pai_huggingface_bert_large_test.sh +++ /dev/null @@ -1,43 +0,0 @@ -#!/bin/bash - -set -ex - -usage() { echo "Usage: $0 [-v ]" 1>&2; exit 1; } - -while getopts "v:" parameter_Option -do case "${parameter_Option}" -in -v) ROCM_VERSION=${OPTARG};; -*) usage ;; -esac -done - -MI200_DEVICE_NUMBERS=$(rocm-smi --showproductname | grep -c "MI250" | xargs) - -if [ "$MI200_DEVICE_NUMBERS" -gt "0" ]; then - RESULT_FILE=ci-mi200.huggingface.bert-large-rocm${ROCM_VERSION}.json -else - RESULT_FILE=ci-mi100.huggingface.bert-large-rocm${ROCM_VERSION}.json -fi - -python \ - /stage/huggingface-transformers/examples/pytorch/language-modeling/run_mlm.py \ - --model_name_or_path bert-large-uncased \ - --dataset_name wikitext \ - --dataset_config_name wikitext-2-raw-v1 \ - --do_train \ - --max_steps 260 \ - --logging_steps 20 \ - --output_dir ./test-mlm-bbu \ - --overwrite_output_dir \ - --per_device_train_batch_size 8 \ - --fp16 \ - --dataloader_num_workers 1 \ - --ort \ - --skip_memory_metrics - -cat ci-pipeline-actual.json - -python /onnxruntime_src/orttraining/tools/ci_test/compare_huggingface.py \ - ci-pipeline-actual.json \ - /onnxruntime_src/orttraining/tools/ci_test/results/"$RESULT_FILE" diff --git a/tools/ci_build/set-trigger-rules.py b/tools/ci_build/set-trigger-rules.py index 0e9cd514d8aa5..b46d1e2559e46 100644 --- a/tools/ci_build/set-trigger-rules.py +++ b/tools/ci_build/set-trigger-rules.py @@ -30,14 +30,9 @@ "mac-ios-ci-pipeline.yml", "mac-ios-packaging-pipeline.yml", "mac-react-native-ci-pipeline.yml", - "orttraining-linux-ci-pipeline.yml", - "orttraining-linux-gpu-ci-pipeline.yml", - "orttraining-linux-gpu-ortmodule-distributed-test-ci-pipeline.yml", - "orttraining-mac-ci-pipeline.yml", "win-ci-pipeline.yml", "win-gpu-dml-ci-pipeline.yml", "win-gpu-cuda-ci-pipeline.yml", - "win-gpu-training-ci-pipeline.yml", "win-gpu-doc-gen-ci-pipeline.yml", "win-gpu-tensorrt-ci-pipeline.yml", "win-gpu-webgpu-ci-pipeline.yml", diff --git a/tools/nuget/generate_nuspec_for_native_nuget.py b/tools/nuget/generate_nuspec_for_native_nuget.py index ff6556b1fd31a..ba125f4e2d980 100644 --- a/tools/nuget/generate_nuspec_for_native_nuget.py +++ b/tools/nuget/generate_nuspec_for_native_nuget.py @@ -138,7 +138,7 @@ def parse_arguments(): required=False, default="None", type=str, - choices=["cuda", "dnnl", "openvino", "tensorrt", "snpe", "tvm", "qnn", "None"], + choices=["cuda", "dnnl", "openvino", "tensorrt", "snpe", "qnn", "None"], help="The selected execution provider for this build.", ) parser.add_argument("--sdk_info", required=False, default="", type=str, help="dependency SDK information.") @@ -375,13 +375,11 @@ def generate_files(line_list, args): "mklml": "mklml.dll", "openmp": "libiomp5md.dll", "dnnl": "dnnl.dll", - "tvm": "tvm.dll", "providers_shared_lib": "onnxruntime_providers_shared.dll", "dnnl_ep_shared_lib": "onnxruntime_providers_dnnl.dll", "tensorrt_ep_shared_lib": "onnxruntime_providers_tensorrt.dll", "openvino_ep_shared_lib": "onnxruntime_providers_openvino.dll", "cuda_ep_shared_lib": "onnxruntime_providers_cuda.dll", - "tvm_ep_shared_lib": "onnxruntime_providers_tvm.lib", "onnxruntime_perf_test": "onnxruntime_perf_test.exe", "onnx_test_runner": "onnx_test_runner.exe", } @@ -394,7 +392,6 @@ def generate_files(line_list, args): "mklml_1": "libmklml_gnu.so", "openmp": "libiomp5.so", "dnnl": "libdnnl.so.1", - "tvm": "libtvm.so.0.5.1", "providers_shared_lib": "libonnxruntime_providers_shared.so", "dnnl_ep_shared_lib": "libonnxruntime_providers_dnnl.so", "tensorrt_ep_shared_lib": "libonnxruntime_providers_tensorrt.so", @@ -456,14 +453,6 @@ def generate_files(line_list, args): + '" target="build\\native\\include" />' ) - if args.execution_provider == "tvm": - files_list.append( - "' - ) - if args.execution_provider == "openvino": files_list.append( "' ) - if args.execution_provider == "tvm": - files_list.append( - "' - ) - files_list.append( - "' - ) - - tvm_build_path = os.path.join(args.ort_build_path, args.build_config, "_deps", "tvm-build") - if is_windows(): - files_list.append( - "' - ) - else: - # TODO(agladyshev): Add support for Linux. - raise RuntimeError("Now only Windows is supported for TVM EP.") - if args.execution_provider == "rocm" or is_rocm_gpu_package and not is_ado_packaging_build: files_list.append( "" ) - # Process tvm dependency - if os.path.exists(os.path.join(args.native_build_path, nuget_dependencies["tvm"])): - files_list.append( - "" - ) - # Some tools to be packaged in nightly debug build only, should not be released # These are copied to the runtimes folder for convenience of loading with the dlls # NOTE: nuget gives a spurious error on linux if these aren't in a separate directory to the library so diff --git a/tools/scripts/python_test.sh b/tools/scripts/python_test.sh index d12f6e6d33772..53d350cf30611 100755 --- a/tools/scripts/python_test.sh +++ b/tools/scripts/python_test.sh @@ -13,9 +13,6 @@ echo Install Python Deps cp $src_dir/tools/ci_build/github/linux/docker/scripts/manylinux/requirements.txt $build_dir/requirements.txt python3 -m pip install -r $build_dir/requirements.txt -mkdir -p $build_dir/requirements_torch_cpu/ -cp $src_dir/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch_cpu/requirements.txt $build_dir/requirements_torch_cpu/requirements.txt -python3 -m pip install -r $build_dir/requirements_torch_cpu/requirements.txt python3 -m pip list | grep onnx echo Install $config python package @@ -23,6 +20,5 @@ rm -rf $build_dir/$config/onnxruntime $build_dir/$config/pybind11 python3 -m pip install $build_dir/$config/dist/*.whl echo Run $config unit tests -pushd $build_dir/$config/ -python3 $src_dir/tools/ci_build/build.py --build_dir $build_dir --cmake_generator Ninja --config $config --test --skip_submodule_sync --build_shared_lib --parallel --use_binskim_compliant_compile_flags --build_wheel --enable_onnx_tests --enable_transformers_tool_test --ctest_path "" -popd +cd $build_dir/$config/ +python3 $src_dir/tools/ci_build/build.py --build_dir $build_dir --cmake_generator Ninja --config $config --test --skip_submodule_sync --build_shared_lib --parallel --use_binskim_compliant_compile_flags --build_wheel --enable_onnx_tests --enable_transformers_tool_test