Files
composable_kernel/tile_engine/ops/gemm_multi_d/README.md
2025-08-19 00:23:07 -07:00

6.1 KiB

CK Tile Engine for GEMM Multi D is used to generate and run GEMM kernels with different combinations of BlockTile sizes, WarpTile sizes, WarpTile mapping for all valid pipelines, schedulers and epilogues while able to give custom datatype and Layout selections

Kernel Configurations

User Specific

Users can specify custom kernel configurations such as tile size, warp size, padding, pipeline, scheduler, and epilogue in the config file. This allows building only for selected configurations, significantly reducing build time. For reference please see ./configs/user_provided_config.json.

Default

The Tile engine also has a default kernel configuration for providing range of configuration parameter values, which helps users who lack kernel development experience to benchmark. For reference please see in ./configs/default_config.json

If user does not provide kernel configuration, the tile engine uses default kernel configuration to generate kernel instances and benchmark.

Build Instructions

# in the root of composable kernel create build directory
mkdir build && cd build
# build composable kernel
# replace [Arch] with the appropriate architecture or leave blank and 
# replace [Datatype] in comma separated datatypes string (possible datatypes are [fp16])
# replace [Layout1;Layout2;...] in comma separated datatypes string (possible layouts are [rcr, rrr, crr, ccr])
# replace "mul" with either of mul,add,passthrough for Elementwise function as Multiply, Add or Passthrough respectively. If this is not specified it is considered as mul by default.
../script/cmake-ck-dev.sh  ../ [Arch] -DGEMM_MULTI_D_DATATYPE="[Datatype]" -DGEMM_MULTI_D_LAYOUT="[Layout1;Layout2]" -DGEMM_MULTI_D_ELEMENTWISE_FUNCTION="mul"
# generate different executable for each passed datatype
make benchmark_gemm_multi_d_[Datatype]_[Layout1] -j
make benchmark_gemm_multi_d_[Datatype]_[Layout2] -j

benchmark_gemm_multi_d_[Datatype]_[Layout] will be located in the ./bin/ directory.

benchmark_gemm_multi_d_[Datatype]_[Layout] must be rebuilt everytime if configuration file is modified.

rm -rf tile_engine/ && make benchmark_gemm_multi_d_[Datatype]_[Layout] -j  # rebuild

For eaxmple build for gfx942 for datatype with rcr layout

mkdir build && cd build
../script/cmake-ck-dev.sh  ../ gfx942 -DGEMM_MULTI_D_DATATYPE="fp16" -DGEMM_MULTI_D_LAYOUT="rcrr" 
make benchmark_gemm_multi_d_fp16_rcrr -j

## benchmark_gemm inputs
                  -m    The value for m dimension. Default is 3840.
                  -n    The value for n dimension. Default is 4096.
                  -k    The value for k dimension. Default is 2048.
           -stride_a    The stride value for tensor A. Default is 0.
           -stride_b    The stride value for tensor B. Default is 0.
          -stride_ds    The stride value for tensor Ds. Default is 0.
           -stride_e    The stride value for tensor E. Default is 0.
            -split_k    The split value for k dimension. Default is 1.
              -verify    The type of validation. Set to 0 for no validation, 1 for validation on CPU, or 2 for validation on GPU. Default is 1, validation on CPU, as validation on GPU is not supported.
                -log    Wether output kernel instance information or not. Possible values are true or false. Default is false.
             -warmup    The number of iterations before benchmark the kernel. Default is 50.
             -repeat    The number of iterations to benchmark the kernel. Default is 100.
              -timer    Whether if the timer is gpu timer or not. Possible values are false or true. Default is true.
               -init    The method of tensor initialization. Set to 0 for random, to 1 for linear, or 2 for constant(1). Default is 0, random.
        -flush_cache    To flush cache, possible values are true or false. Default is false.
     -rotating_count    Number of iterations to rotate the cache. Default is 5.
             -metric    Metric with which to measure kernel performance. Set to 0 for latency, 1 for tflops, or 2 for bandwidth. Default is 0, latency.
       -csv_filename    The filename of benchmark result. Default is gemm_multi_d_kernel.
           -pipeline    The type of pipeline. Possible values are compv3, compv4 or mem. Default is compv3.
          -scheduler    The type of scheduler. Possible values are intrawave. Default is intrawave.
           -epilogue    The type of epilogue. Possible values are cshuffle or default. Default is cshuffle.
              -pad_m    Whether pad or not in m direction. Possible values are true or false. Default is false.
              -pad_n    Whether pad or not in n direction. Possible values are true or false. Default is false.
              -pad_k    Whether pad or not in k direction. Possible values are true or false. Default is false.

Note: pipeline, scheduler, epilogue, pad_m, pad_n, pad_k should be one of the options specified in user_provided_config.json

Note: In `./configs/user_provided_config.json` pipeline, scheduler, epilogue, pad_m, pad_n, pad_k should be from one of the values specified above.

## Example

The following JSON file specifies parameters used to generate and build GEMM kernels across all possible combinations of pipelines, schedulers, epilogues with different tile and warp sizes.

```json
{     
    /// other parameters ///
    
    "tile_m": {
      "values": [256]
    },
    "tile_n": {
      "values": [256]
    },
    "tile_k": {
      "values": [64, 32]
    },

    /// other parameters ///

    "pipeline": {
      "values": ["compv3", "compv4", "mem"]
    },
    "scheduler": {
      "values": ["intrawave", "interwave"]
    },
    "epilogue": {
      "values": ["cshuffle"]
    }
}

At runtime, a specific subset of the generated kernels can be selected using command-line arguments.

./bin/benchmark_gemm_multi_d_[Datatype]_[Layout] -pipeline=compv3 -scheduler=intrawave -epilogue=cshuffle 

The above command runs kernels configured with the compv3 pipeline, intrawave scheduler, and cshuffle epilogue, while sweeping over different BlockTile sizes, WarpTile sizes, and WarpTile mappings.