Skip to content

Module not executed with option --lower-composite #490

@sqPoseidon

Description

@sqPoseidon

In some cases, the kernels are not executed by the main module. Not sure which part affects the inference.
I have two kernels: bn1.mlir and relu.mlir. They are called in the main.mlir. I delete the execution part and only print an array.

bn1.mlir:

module {
  memref.global "private" @gv0 : memref<4x4xf32> = dense<[[1.0, 2.0, 3.0, 4.0], [1.0, 2.0, 3.0, 4.0], [1.0, 2.0, 3.0, 4.0], [1.0, 2.0, 3.0, 4.0]]>
  func.func private @printMemrefF32(%ptr : memref<*xf32>)
  memref.global "private" constant @weight : memref<64xi8> = dense<1>
  memref.global "private" constant @bias : memref<64xi8> = dense<0>
  func.func @bn1(%arg0: memref<1x64x112x112xf32>) -> memref<1x64x112x112xi8> attributes {itypes = "_", llvm.emit_c_interface, otypes = "s", top} {
    %c0 = arith.constant 0 : index
    %0 = memref.get_global @weight : memref<64xi8>
    %1 = memref.get_global @bias : memref<64xi8>
    %2 = memref.alloc() {name = "BatchNorm2d"} : memref<1x64x112x112xi8>
     %temp = memref.get_global @gv0 : memref<4x4xf32>
     // hcl.print(%0) : memref<4x4xf32>
     %casted = memref.cast %temp : memref<4x4xf32> to memref<*xf32>
     func.call @printMemrefF32(%casted) : (memref<*xf32>) -> ()
    return %2 : memref<1x64x112x112xi8>
  }
}

relu.mlir

module {
  memref.global "private" @gv0 : memref<4x4xf32> = dense<[[1.0, 2.0, 3.0, 4.0], [1.0, 2.0, 3.0, 4.0], [1.0, 2.0, 3.0, 4.0], [1.0, 2.0, 3.0, 4.0]]>
  func.func private @printMemrefF32(%ptr : memref<*xf32>)
  func.func @relu(%arg0: memref<1x64x112x112xi8>) -> memref<1x64x112x112xi8> attributes {itypes = "s", llvm.emit_c_interface, otypes = "s", top} {
    %c0 = arith.constant 0 : index
    %0 = memref.alloc() {name = "ReLU"} : memref<1x64x112x112xi8>
     %temp = memref.get_global @gv0 : memref<4x4xf32>
     // hcl.print(%0) : memref<4x4xf32>
     %casted = memref.cast %temp : memref<4x4xf32> to memref<*xf32>
     func.call @printMemrefF32(%casted) : (memref<*xf32>) -> ()
    return %0 : memref<1x64x112x112xi8>
  }
}

main.mlir

module {
  func.func private @bn1(memref<1x64x112x112xf32>) -> memref<1x64x112x112xi8>
  func.func private @relu(memref<1x64x112x112xi8>) -> memref<1x64x112x112xi8>
  func.func @main() {
    %0 = memref.alloc() : memref<1x64x112x112xf32>
    // affine.for %arg = 0 to 20 {
        %1 = func.call @bn1(%0) : (memref<1x64x112x112xf32>) -> memref<1x64x112x112xi8>
        %2 = func.call @relu(%1) : (memref<1x64x112x112xi8>) -> memref<1x64x112x112xi8>
    // }
    return
  }
}

compile.sh

set -x
# export HCL_OPT_FLAGS=" --fixed-to-integer --verify-each --verify-diagnostics --allow-unregistered-dialect --preload-dialects-in-context --lower-bitops --lower-to-llvm"
export HCL_OPT_FLAGS=" --fixed-to-integer --verify-each --verify-diagnostics --allow-unregistered-dialect --preload-dialects-in-context --lower-composite --lower-bitops --lower-to-llvm"
hcl-opt bn1.mlir $HCL_OPT_FLAGS | mlir-translate -mlir-to-llvmir | opt -O3 -S | llc -O3 -filetype=obj -o bn1.o
hcl-opt relu.mlir $HCL_OPT_FLAGS | mlir-translate -mlir-to-llvmir | opt -O3 -S | llc -O3 -filetype=obj -o relu.o
hcl-opt main.mlir $HCL_OPT_FLAGS | mlir-translate -mlir-to-llvmir | opt -O3 -S | llc -O3 -filetype=obj -o main.o
clang main.o bn1.o relu.o --target=x86-64 -fPIE \
    -L/work/shared/users/staff/qs228/MLIR/heterocl-mlir/hcl-dialect/llvm-project/build/lib/ \
    -lmlir_c_runner_utils \
    -lmlir_runner_utils \
    -o x86-app

echo "Compilation Done."
start=$(date +%s%6N)
./x86-app # 2>&1 | tee temp.log
end=$(date +%s%6N)
runtime=`expr "scale=3; $( echo "$end - $start" | bc -l ) / 1000.0" | bc `
echo "X86 Execution Done."
echo Simulation Time: $runtime ms.

If these two modules are executed, you'll see the printed arrays. With option --lower-composite, it doesn't print anything.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions