TensileLite

TensileLite is a GEMM assembly generator in hipBLASLt.

Quick Start

This is a basic config yaml to generate an assembly kernel.

GlobalParameters:
  SleepPercent: 50
  Device: 0
  MaxFileName: 256
  DataInitTypeAlpha: 1
  DataInitTypeBeta: 0
  PrintSolutionRejectionReason: True
  NumElementsToValidate: -1
  NumWarmups: 0
  EnqueuesPerSync: 1
  SyncsPerBenchmark: 1
  NumBenchmarks: 1
  MinKForGSU: 1
  KeepBuildTmp: 1
BenchmarkProblems:

  #######################################
  # NN - standard
  #######################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: H
      DestDataType: H
      ComputeDataType: S
      HighPrecisionAccumulate:  True
      TransposeA: 1
      TransposeB: 0
      UseBeta: True
      Sparse: 0
      Batched: True
      Activation: True
      ActivationType: all

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 16, 1, 1, 2, 2, 2, 2]
        - GlobalReadVectorWidthA: [4]
        - WorkGroupMappingXCC: [8]
        - DepthU: [32]
        - TransposeLDS: [-1]
        - ScheduleIterAlg: [3]
        - PrefetchLocalRead: [1]
        - PrefetchGlobalRead: [1]
        - GlobalSplitU: [1, 2]
        - 1LDSBuffer: [1]
        - DirectToVgprSparseMetadata: [0]
        - WorkGroupMapping: [1]
        - StoreVectorWidth: [-1]
        - UseSgprForGRO: [1]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [512, 512, 1, 512]
        - ActivationArgs:
          - [Enum: none]

Follow the README.md here to run cmake first.

After running the cmake, navigate to the build folder and make a yaml file with the content above.

cd build
Tensile.sh <path to the yaml file> .

Then you'll see these folders inside the build folder.

The assembly file is located inside the assembly folder.