mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-19 20:40:07 +00:00
4a7d8df85a
* [What] Separate fixpoint gemm from gemm example
[Why] let example of gemm_int8 be pure gemm.
[What]
1. Add gemm_requant_relu_requant,
2. Let CDataType be int32 in pure gemm, because no one use int8 CDataType. It is also part of gemm_requant_relu_requant
* Fix path
* Revise cmakelist due to merge develop
* Add gemm fp16 test
* Extract PrepareGemmTensor
* Extract TestGemm
* Add test for different layout
* Add 4 layouts of shuffle version of fp32
* Add 4 layouts of shuffle version of int8
* Add 4 layouts of shuffle version of bf16
* replace all DeviceGemmPtr_ with DeviceGemmNoOpPtr to fit naming convension
* Add test for non-shuffle verstion of gemm
* Fix typo
* Print kernel information
* Add rest of the fp32 kernel to the test
* 1. Add rest of the fp16 device iop.
2. Mark the invalid device operation
Co-authored-by: rocking <chunylai@amd.com>
[ROCm/composable_kernel commit: 485ea46a40]
Instructions for gemm_xdl_bias_relu_add Example
Docker script
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
Build gemm_xdl_bias_relu_add
mkdir build && cd build
# Need to specify target ID, example below is gfx908
cmake \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
make -j gemm_xdl_bias_relu_add
Run gemm_xdl_bias_relu_add
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
#arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC
./example/gemm_xdl_bias_relu_add 0 1 5 3840 4096 4096 4096 4096 4096
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
c1_m_n: dim 2, lengths {3840, 4096}, strides {1, 0}
arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
arg.c_grid_desc_m_n_{ 3840, 4096}
arg.c0_grid_desc_m_n_{ 3840, 4096}
arg.c1_grid_desc_m_n_{ 3840, 4096}
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 1.27583 ms, 100.992 TFlops, 73.9688 GB/s