ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 02:02:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

WMMA support for GEMM reduce (#2823)

Browse Source 
Added gemm + reduce instance library for RDNA4. This includes:

- New device implementation running GEMM and reduction kernel
- instances for wmma (xdl parity)
- examples for wmma (xdl parity)
- tests for existing xdl and wmma

[ROCm/composable_kernel commit: b25d4d684a]
This commit is contained in:
Wojciech Laskowski
2025-09-12 21:36:43 +02:00
committed by GitHub
parent 8c0cdebe63
commit f2edb06bb0
27 changed files with 1911 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
profiler/include/profiler/profile_gemm_universal_reduce_impl.hpp
View File

@@ -10,6 +10,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3r1.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3r1.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_universal_reduce.hpp"
@@ -86,10 +87,21 @@ bool profile_gemm_universal_reduce_impl(int do_verification,
switch(init_method)
{
case 0: break;
case 0:
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
break;
case 3:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});

4
profiler/src/CMakeLists.txt
View File

@@ -68,7 +68,6 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_OPS profile_gemm_splitk.cpp)
list(APPEND PROFILER_OPS profile_batched_gemm_b_scale.cpp)
list(APPEND PROFILER_OPS profile_gemm_universal_batched.cpp)
list(APPEND PROFILER_OPS profile_gemm_universal_reduce.cpp)
list(APPEND PROFILER_OPS profile_gemm_universal_streamk.cpp)
list(APPEND PROFILER_OPS profile_conv_fwd_bias_relu.cpp)
list(APPEND PROFILER_OPS profile_conv_fwd_bias_relu_add.cpp)
@@ -90,6 +89,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1[1
list(APPEND PROFILER_OPS profile_gemm_universal.cpp)
list(APPEND PROFILER_OPS profile_batched_gemm.cpp)
list(APPEND PROFILER_OPS profile_gemm_b_scale.cpp)
list(APPEND PROFILER_OPS profile_gemm_universal_reduce.cpp)
list(APPEND PROFILER_OPS profile_grouped_conv_fwd.cpp)
list(APPEND PROFILER_OPS profile_grouped_conv_fwd_bias_clamp.cpp)
list(APPEND PROFILER_OPS profile_grouped_conv_fwd_clamp.cpp)
@@ -185,7 +185,6 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND DEVICE_INSTANCES device_gemm_splitk_instance)
list(APPEND DEVICE_INSTANCES device_batched_gemm_b_scale_instance)
list(APPEND DEVICE_INSTANCES device_gemm_universal_batched_instance)
list(APPEND DEVICE_INSTANCES device_gemm_universal_reduce_instance)
list(APPEND DEVICE_INSTANCES device_gemm_universal_streamk_instance)
list(APPEND DEVICE_INSTANCES device_gemm_add_multiply_instance)
list(APPEND DEVICE_INSTANCES device_gemm_add_instance)
@@ -221,6 +220,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1[1
list(APPEND DEVICE_INSTANCES device_gemm_universal_instance)
list(APPEND DEVICE_INSTANCES device_batched_gemm_instance)
list(APPEND DEVICE_INSTANCES device_gemm_b_scale_instance)
list(APPEND DEVICE_INSTANCES device_gemm_universal_reduce_instance)
list(APPEND DEVICE_INSTANCES device_grouped_conv3d_fwd_instance)
list(APPEND DEVICE_INSTANCES device_grouped_conv2d_bwd_data_instance)
list(APPEND DEVICE_INSTANCES device_grouped_conv3d_bwd_data_instance)