vector-reshapes.md

Virtual Vector -> SPIR-V

SPIRVInitialVectorLowering

Upstream unrolling patterns + canonicalization

• Vectorizing tensor.pad

  • Apply vectorization to generate vectors of the original tensor shape for tensor.pad ops.
  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L306

• Peephole optimizations

  • Shuffle broadcast/transpose ops to cancel them or embed into contract ops.
  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L318

• Trimming contract leading unit dims

  • Drop the leading unit batch dimension for 1-D convolution ops
  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L342

• Folding tensor extract/insert slice ops

  • Fold tensor.extract_slice and tensor.insert_slice ops into transfer ops
  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L362

• Lowering multi reduction ops

  • Lower vector.multi_dimension early if any operand is a transpose op. The lowering itself generates transpose ops.
  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L381

• Preparing for SPIR-V dot product lowering

  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L401

• Unrolling vector ops

  • Unroll vectors to native vector size
  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L414

• Lowering size-1 reduction contract ops

  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L427

• Lowering transpose ops

  • Lower vector transpose by converting them into extract and insert pairs, in preparation of futher transformations to canonicalize/cancel
  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L450

• Trimming leading unit dims

  • Run canonicalization to cast away leading size-1 dimensions which can be generated from vector unrolling
  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L450

• Lowering to SPIR-V dot product

  • Lower vector reduction to SPIR-V integer dot product
  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVInitialVectorLowering.cpp#L501

Other one-off patterns for breaking down large vectors

• Breaking down loop carried variables

  • Convert scf.for loop-carried values < 128 total bits, but more than 4 elements
  • E.g. insert vector.bitcast ops to cast vector<8xf16> to vector<4xf32>
  • The loop-carried values will be lowered into SPIR-V local variables
  • This pattern allows packing i4/i8/f16 values into i32 variables tightly so that we can generate shader conformant SPIR-V
  • https://github.com/iree-org/iree/blob/7e06b6e5ece2adcda91352331db817adaa058e85/compiler/src/iree/compiler/Codegen/Common/ForOpCanonicalizationPass.cpp#L186

• Breaking down other large vectors

  • Convert vector.extract_strided_slice into a chain of vector.extract and then a chain of vector.insert ops
  • Helps to cancel with previous vector.extract and vector.insert ops
  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVBreakDownLargeVector.cpp#L138

• Convert to memref<vector> + bitcasts to get 128 bit read/store width on < 32 bit data types

  • https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/SPIRV/SPIRVVectorizeLoadStore.cpp#L1034