GEMM Benchmark supporting multiple GPU BLAS implementaitons

Goals:

• Use the familiar interface of rocblas-bench/hipblaslt-bench
• Implement native support for cuBLAS, cuBLASLT, and other GPU BLAS libraries
• Expose implementation specific tuneables to the user

Cloning the Repository

To get started, clone the repository from GitHub using:

git clone git@github.com:ROCm/mblas-bench.git
cd mblas-bench

Building for ROCm only

cmake -S src -B build -DWITH_ROCM=true -DWITH_CUDA=false  
cmake --build build -j 

Conditional ROCm backends

ROCm builds include both hipBLASLt and rocBLAS backends in a single mblaship library. Disable backends individually if needed:

cmake -S src -B build -DWITH_CUDA=false -DWITH_HIPBLASLT=true -DWITH_ROCBLAS=false  # hipBLASLt only
cmake -S src -B build -DWITH_CUDA=false -DWITH_HIPBLASLT=false -DWITH_ROCBLAS=true  # rocBLAS only

Building for CUDA only

cmake -S src -B build -DWITH_ROCM=false -DWITH_CUDA=true 
cmake --build build -j

Building for both CUDA and ROCm

cmake -S src -B build -DWITH_ROCM=true -DWITH_CUDA=true
cmake --build build -j

Running on ROCm or CUDA

First, select a driver based on which backend you'd like to use

ROCm backends

• hipBLASLt: export DRIVER="hipblaslt"
• rocBLAS: export DRIVER="rocblas"

CUDA backends

• cuBLAS: export DRIVER="cublas"
• cuBLASLt: export DRIVER="cublaslt"

Run 4k gemms on the command line

Use the below commands to run "4k" gemms on ROCm or CUDA. You'll need to make sure you've set your driver as shown above

Precision	Base Command
FP64	build/mblas-bench -m 4096 -n 4096 -k 4096 --alpha 1 --beta 0 --transposeA N --transposeB T --initialization trig_float --iters 10 --cold_iters 2 --function matmul --precision d --rotating 512 --driver ${DRIVER}
FP32	build/mblas-bench -m 4096 -n 4096 -k 4096 --alpha 1 --beta 0 --transposeA N --transposeB T --initialization trig_float --iters 10 --cold_iters 2 --a_type f32_r --b_type f32_r --c_type f32_r --d_type f32_r --compute_type f32_r --function matmul --rotating 512 --driver ${DRIVER}
TF32	build/mblas-bench -m 4096 -n 4096 -k 4096 --alpha 1 --beta 0 --transposeA T --transposeB N --initialization trig_float --iters 10 --cold_iters 2 --compute_type CUBLAS_COMPUTE_32F_FAST_TF32 --function gemm_ex --precision s --rotating 512 --driver ${DRIVER}
FP16	build/mblas-bench -m 4096 -n 4096 -k 4096 --alpha 1 --beta 0 --transposeA N --transposeB T --initialization trig_float --iters 10 --cold_iters 2 --a_type f16_r --b_type f16_r --c_type f16_r --d_type f16_r --compute_type f32_r --function matmul --rotating 512 --driver ${DRIVER}
BF16	build/mblas-bench -m 4096 -n 4096 -k 4096 --alpha 1 --beta 0 --transposeA N --transposeB T --initialization trig_float --iters 10 --cold_iters 2 --a_type bf16_r --b_type bf16_r --c_type bf16_r --d_type bf16_r --compute_type f32_r --function matmul --rotating 512 --driver ${DRIVER}
FP8	build/mblas-bench -m 4096 -n 4096 -k 4096 --alpha 1 --beta 0 --transposeA T --transposeB N --initialization trig_float --iters 10 --cold_iters 2 --a_type f8_r --b_type f8_r --c_type bf16_r --d_type bf16_r --compute_type f32_r --function matmul --rotating 512 --driver ${DRIVER}
INT8	build/mblas-bench -m 4096 -n 4096 -k 4096 --alpha 1 --beta 0 --transposeA T --transposeB N --initialization rand_int --iters 10 --cold_iters 2 --a_type i8_r --b_type i8_r --c_type i32_r --d_type i32_r --compute_type i32_r --function matmul --rotating 512 --driver ${DRIVER}

Running gemms from a YAML file

Instead of specifying all parameters on the command line, you can use the --yaml argument to provide a configuration file that defines one or more GEMM operations. This is especially useful for running benchmarks across multiple matrix sizes, data types, or configurations.

The YAML file should contain a list of GEMM configurations. Each configuration can specify parameters such as matrix dimensions (m, n, k), transpose operations (transA, transB), data types (a_type, b_type, c_type, d_type), compute type, and other options.

For example, to run all the 4k gemms from the table above using the provided example YAML file:

build/mblas-bench --yaml examples/4k_gemms.yaml --driver ${DRIVER}