This PR introduces the @auto_pipeline decorator that enables automatic multi-level pipelining optimization for Triton kernels, achieving up to 2.18x speedup on GEMM operations compared to non-pipelined kernels.

Performance Results (2048x2048x2048 GEMM on Nvidia)

AutoPipeline vs Default Pipeline: 1.53x faster

Features

• Global-to-Shared (G2S) Pipelining: Multi-stage async data prefetching from global memory to shared memory with configurable pipeline depth
• Shared-to-Register (S2R) Pipelining: Double-buffering optimization for shared memory to register transfers, reducing memory access latency
• Warp Specialization Support: Producer-consumer pattern configuration with dedicated prefetch and compute warps
• Async Copy: Automatic async copy instruction generation for supported hardware
• Swizzle Optimization: Memory access pattern optimization to reduce bank conflicts

Usage

  import triton
  import triton.language as tl
  from triton.language import auto_pipeline, PipelineConfig, WarpSpecConfig

  @triton.jit
  @auto_pipeline(PipelineConfig(
      global_to_shared_stages=4,      # G2S pipeline depth
      shared_to_register_stages=2,    # S2R double-buffering
      enable_async_copy=True,         # Use async copy instructions
      enable_swizzle=True,            # Optimize memory access patterns
      enable_warp_specialization=True,
      warp_spec_config=WarpSpecConfig(
          num_producer_warps=1,
          num_consumer_warps=3,
      )
  ))
  def matmul_kernel(A, B, C, M, N, K, ...):
      # Standard GEMM implementation - no manual changes needed!
      ...

Files Changed

Python API (4 files, ~1,700 lines)

• python/triton/language/pipeline.py - Core @auto_pipeline decorator and PipelineConfig class
• python/triton/language/autotune_config.py - Smart autotuning utilities
• python/triton/compiler/pipeline_config.py - Compiler integration hook
• python/triton/language/init.py - Public exports

C++ MLIR Passes (8 files, ~4,500 lines)

• lib/Dialect/TritonGPU/Transforms/AdvancedPipeliner.cpp - Main S2R pipelining pass
• lib/Dialect/TritonGPU/Transforms/BufferAccessAnalysis.cpp - Memory access pattern analysis
• lib/Dialect/TritonGPU/Transforms/CircularBufferTransform.cpp - Circular buffer implementation
• lib/Dialect/TritonGPU/Transforms/MultiBufferFusion.cpp - Buffer fusion optimization
• lib/Dialect/TritonGPU/Transforms/PipelineOpportunityDetector.cpp - Pipeline opportunity detection
• lib/Dialect/TritonGPU/Transforms/SynchronizationInsertion.cpp - Barrier insertion
• lib/Dialect/TritonGPU/Transforms/WarpSpecialization.cpp - Warp specialization transform
• lib/Dialect/TritonGPU/Transforms/TMASupport.cpp - TMA (Tensor Memory Access) support

Headers (8 files)

• include/triton/Dialect/TritonGPU/Transforms/*.h - Pass declarations

Tests & Examples (1 file)

• python/test/benchmark_autopipeline.py - Benchmark demonstrating speedup

How It Works

1. Decorator Application: @auto_pipeline(config) attaches pipeline configuration to the kernel function
2. Compiler Hook: During compilation, PipelineCompilerHook detects the configuration and injects optimization passes
3. Pass Execution: The AdvancedPipeliner pass analyzes the kernel IR and applies:
   - Circular buffer allocation for multi-stage pipelining
   - Double-buffering for S2R optimization
   - Async copy instruction insertion
   - Synchronization barrier placement

Test Plan

• Run python python/test/benchmark_autopipeline.py - Verifies 2.18x speedup
• Correctness validation via torch.allclose() comparison with reference implementation
• Build with pip install -e . --no-build-isolation
• Run existing unit tests

Breaking Changes

None. This is a purely additive feature that doesn't modify existing APIs.

Dependencies

• Triton 3.1.x base