Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "a2f036196cccc2d5c47e448cff86daf4b405250e"
Commit fdddc8f4 authored by Bartlomiej Kocot's avatar Bartlomiej Kocot
Browse files

Merge branch 'develop' of github.com:ROCmSoftwarePlatform/composable_kernel... 

Merge branch 'develop' of github.com:ROCmSoftwarePlatform/composable_kernel into barkocot/fix-cmake-tensor-op-instance
parents 8f48018d f7331c60
Show whitespace changes
Inline Side-by-side
Showing with 489 additions and 292 deletions
CHANGELOG.md
View file @ fdddc8f4
......@@ -14,10 +14,10 @@ None
### Additions
- Added an image to a column kernel (#867)
- Added a column to an image kernel (#930)
- Support for 3D grouped convolution forward on RDNA 3 GPUs (#935)
- Support for 3D grouped convolution on RDNA 3 GPUs (#935, #950, #985)
- Grouped convolution support for small K and C (#822 #879 #897)
- Support for NHWGC (2D and 3D) grouped convolution backward weight (#769 #804)
- Support for bf16/f32/f16 and NHWGC (2D and 3d) grouped convolution backward data (#757 #799)
- Support for bf16/f32/f16 and NHWGC (2D and 3D) grouped convolution backward data (#757 #799)
- Support for Batched Gemm DL (#732)
### Changes
......
CMakeLists.txt
View file @ fdddc8f4
......@@ -32,12 +32,10 @@ if (DTYPES)
if (DTYPES MATCHES "fp8")
add_definitions(-DCK_ENABLE_FP8)
set(CK_ENABLE_FP8 "ON")
add_compile_options(-Wno-bit-int-extension)
endif()
if (DTYPES MATCHES "bf8")
add_definitions(-DCK_ENABLE_BF8)
set(CK_ENABLE_BF8 "ON")
add_compile_options(-Wno-bit-int-extension)
endif()
if (DTYPES MATCHES "fp16")
add_definitions(-DCK_ENABLE_FP16)
......@@ -59,9 +57,11 @@ if (DTYPES)
else()
add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_BF8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
set(CK_ENABLE_ALL_DTYPES "ON")
add_compile_options(-Wno-bit-int-extension) # enable fp8 and bf8
endif()
#for f8/bf8_t type
add_compile_options(-Wno-bit-int-extension)
if(DL_KERNELS)
add_definitions(-DDL_KERNELS)
set(CK_ENABLE_DL_KERNELS "ON")
......
Dockerfile
View file @ fdddc8f4
......@@ -49,6 +49,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
vim \
nano \
zlib1g-dev \
zip \
openssh-server \
clang-format-12 \
kmod && \
......
Jenkinsfile
View file @ fdddc8f4
......@@ -526,6 +526,26 @@ def Build_CK(Map conf=[:]){
stash "ckprofiler_0.2.0_amd64.deb"
}
}
if (params.hipTensor_test && navi_node == 0 ){
//build and test hipTensor
sh """#!/bin/bash
rm -rf "${params.hipTensor_branch}".zip
rm -rf hipTensor-"${params.hipTensor_branch}"
wget https://github.com/ROCmSoftwarePlatform/hipTensor/archive/refs/heads/"${params.hipTensor_branch}".zip
unzip -o "${params.hipTensor_branch}".zip
"""
dir("hipTensor-${params.hipTensor_branch}"){
sh """#!/bin/bash
mkdir -p build
ls -ltr
CC=hipcc CXX=hipcc cmake -Bbuild . -D CMAKE_PREFIX_PATH="/opt/rocm;${env.WORKSPACE}/install"
cmake --build build -- -j
"""
}
dir("hipTensor-${params.hipTensor_branch}/build"){
sh 'ctest'
}
}
}
}
}
......@@ -654,6 +674,15 @@ pipeline {
name: "DL_KERNELS",
defaultValue: false,
description: "Select whether to build DL kernels (default: OFF)")
booleanParam(
name: "hipTensor_test",
defaultValue: true,
description: "Use the CK build to verify hipTensor build and tests (default: ON)")
string(
name: 'hipTensor_branch',
defaultValue: 'mainline',
description: 'Specify which branch of hipTensor to use (default: mainline)')
}
environment{
dbuser = "${dbuser}"
......@@ -761,8 +790,8 @@ pipeline {
}
agent{ label rocmnode("navi32") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1101" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1101" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1101" -DDL_KERNELS=ON """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1101" -DDL_KERNELS=ON -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
}
steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
......
client_example/05_layernorm/layernorm2d.cpp
View file @ fdddc8f4
......@@ -16,9 +16,11 @@ using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using SaveMeanInvStdDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
......@@ -50,12 +52,16 @@ int main(int argc, char* argv[])
SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
#ifdef SAVE_MEAN_INV_STD
SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * M);
SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * M);
#endif
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
Rank,
NumReduceDim>;
......@@ -84,14 +90,21 @@ int main(int argc, char* argv[])
{0, 1}, // gammaStrides
{0, 1}, // betaStrides
{Stride, 1}, // yStrides
{1}, // save_mean Strides
{1}, // save_inv_std Strides
{1}, // reduceDims
1e-4,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
......@@ -109,6 +122,10 @@ int main(int argc, char* argv[])
std::size_t num_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N +
sizeof(BetaDataType) * N + sizeof(YDataType) * M * N;
#ifdef SAVE_MEAN_INV_STD
num_byte += sizeof(SaveMeanInvStdDataType) * M * 2;
#endif
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
......@@ -140,17 +157,24 @@ int main(int argc, char* argv[])
auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // lengths
{Stride, 1}, // xStrides
{1}, // gammaStrides
{1}, // betaStrides
{0, 1}, // gammaStrides
{0, 1}, // betaStrides
{Stride, 1}, // yStrides
{1}, // save_mean Strides
{1}, // save_inv_std Strides
{1}, // reduceDims
1e-4,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
......
client_example/18_groupnorm/groupnorm_swish.cpp
View file @ fdddc8f4
......@@ -16,9 +16,11 @@ using XDataType = ck::half_t;
using GammaDataType = float;
using BetaDataType = float;
using YDataType = ck::half_t;
using ComputeDataType = float;
using SaveMeanInvStdDataType = float;
using Swish = ck::tensor_operation::element_wise::Swish;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
......@@ -51,17 +53,22 @@ int main(int argc, char* argv[])
std::vector<ck::index_t> xy_strides = {H * W * G * C, W * G * C, G * C, C, 1};
std::vector<ck::index_t> gamma_beta_strides = {0, 0, 0, C, 1};
std::vector<ck::index_t> save_mean_inv_std_strides = {G, 1};
SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size);
SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_beta_size);
SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * gamma_beta_size);
SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
#ifdef SAVE_MEAN_INV_STD
SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * N * G);
SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * N * G);
#endif
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
Swish,
Rank,
NumReduceDim>;
......@@ -80,14 +87,21 @@ int main(int argc, char* argv[])
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
save_mean_inv_std_strides, // save_mean Strides
save_mean_inv_std_strides, // save_inv_std Strides
{1, 2, 4}, // reduceDims
1e-6,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
Swish{});
if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get()))
......@@ -108,19 +122,27 @@ int main(int argc, char* argv[])
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
auto argument_ptr =
op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
save_mean_inv_std_strides, // save_mean Strides
save_mean_inv_std_strides, // save_inv_std Strides
{1, 2, 4}, // reduceDims
1e-6,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
Swish{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
......@@ -139,6 +161,10 @@ int main(int argc, char* argv[])
sizeof(XDataType) * xy_size + sizeof(GammaDataType) * gamma_beta_size +
sizeof(BetaDataType) * gamma_beta_size + sizeof(YDataType) * xy_size;
#ifdef SAVE_MEAN_INV_STD
num_byte += sizeof(SaveMeanInvStdDataType) * N * G * 2;
#endif
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
......@@ -169,19 +195,27 @@ int main(int argc, char* argv[])
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
auto argument_ptr =
op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
save_mean_inv_std_strides, // save_mean Strides
save_mean_inv_std_strides, // save_inv_std Strides
{1, 2, 4}, // reduceDims
1e-6,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
Swish{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
......
cmake/EnableCompilerWarnings.cmake
View file @ fdddc8f4
......@@ -70,6 +70,7 @@ else()
-Wno-option-ignored
-Wsign-compare
-Wno-extra-semi-stmt
-Wno-unused-template
)
if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "Clang")
list(APPEND CMAKE_COMPILER_WARNINGS
......
example/01_gemm/CMakeLists.txt
View file @ fdddc8f4
add_custom_target(example_gemm_dl)
add_example_executable(example_gemm_dl_fp32 gemm_dl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_dl example_gemm_dl_fp32)
endif()
add_example_dependencies(example_gemm_dl example_gemm_dl_fp32)
add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_dl example_gemm_dl_fp16)
endif()
add_example_dependencies(example_gemm_dl example_gemm_dl_fp16)
add_example_executable(example_gemm_dpp_fp16 gemm_dpp_fp16.cpp)
add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_dl example_gemm_dl_int8)
endif()
add_example_dependencies(example_gemm_dl example_gemm_dl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_dl_int4 gemm_dl_int4.cpp)
add_dependencies(example_gemm_dl example_gemm_dl_int4)
add_example_dependencies(example_gemm_dl example_gemm_dl_int4)
endif(USE_BITINT_EXTENSION_INT4)
add_custom_target(example_gemm_xdl)
add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
endif()
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
endif()
add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
endif()
add_example_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_custom_target(example_gemm_wmma)
add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
endif()
add_example_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
endif()
add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
add_example_executable(example_gemm_xdl_bf16_rtn gemm_xdl_bf16_rtn.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_bf16_rtn)
endif()
add_example_executable(example_gemm_xdl_bf16_rtn gemm_xdl_bf16_rtn.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_rtn)
add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_xdl example_gemm_xdl_int8)
endif()
add_example_dependencies(example_gemm_xdl example_gemm_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_xdl_int4 gemm_xdl_int4.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_int4)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4)
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
endif()
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
add_example_executable(example_gemm_xdl_streamk gemm_xdl_streamk.cpp)
add_example_executable(example_gemm_xdl_fp8 gemm_xdl_fp8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp8)
endif()
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8)
add_example_executable(example_gemm_xdl_fp8_bf8 gemm_xdl_fp8_bf8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp8_bf8)
endif()
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_bf8)
add_example_executable(example_gemm_xdl_fp16_fp8 gemm_xdl_fp16_fp8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8)
endif()
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8)
example/04_gemm_add_add_fastgelu/CMakeLists.txt
View file @ fdddc8f4
......@@ -4,25 +4,21 @@ foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_gemm_add_add_fastgelu_xdl)
add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16)
endif()
add_example_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp16)
endif()
add_example_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp16)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp32)
endif()
add_example_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp32)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_add_add_fastgelu_xdl_int4 gemm_add_add_fastgelu_xdl_int4.cpp)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
add_example_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
endif()
add_example_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
set(target 1)
endif()
endforeach()
example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
View file @ fdddc8f4
......@@ -3,25 +3,22 @@ set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_convnd_fwd_reduce_xdl)
add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
endif()
add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
endif()
add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
endif()
add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
endif()
add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)
add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4)
set(target 1)
endif()
......
example/12_reduce/README.md
View file @ fdddc8f4
......@@ -2,7 +2,7 @@
## Run ```example_reduce_blockwise```
```bash
# -D <xxx> : input 3d/4d/5d tensor lengths
# -D <xxx> : input 3D/4D/5D tensor lengths
# -R <xxx> : reduce dimension ids
# -v <x> : verification (0=no, 1=yes)
#arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64, 7: int4)
......@@ -22,7 +22,7 @@ Perf: 0.238063 ms, 264.285 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSr
## Run ```example_reduce_multiblock_atomic_add```
```bash
# -D <xxx> : input 3d/4d/5d tensor lengths
# -D <xxx> : input 3D/4D/5D tensor lengths
# -R <xxx> : reduce dimension ids
# -v <x> : verification (0=no, 1=yes)
#arg1: data type (0: fp32, 1: fp64)
......
example/15_grouped_gemm/CMakeLists.txt
View file @ fdddc8f4
add_custom_target(example_grouped_gemm_xdl)
add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
endif()
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp16)
endif()
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp16)
add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multiple_d_dl_fp16)
endif()
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multiple_d_dl_fp16)
add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_splitk_fp16)
endif()
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_splitk_fp16)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp16 grouped_gemm_xdl_fixed_nk_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp16)
endif()
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp16)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_bias_fp16 grouped_gemm_xdl_fixed_nk_bias_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_bias_fp16)
endif()
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_bias_fp16)
add_example_executable(example_grouped_gemm_xdl_bf16 grouped_gemm_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_bf16)
endif()
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_bf16)
add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
endif()
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp8 grouped_gemm_xdl_fixed_nk_fp8.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
endif()
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
endif()
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
endif()
example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
View file @ fdddc8f4
......@@ -6,56 +6,42 @@ foreach(gpu IN LISTS GPU_TARGETS)
add_custom_target(example_gemm_reduce_xdl_max)
add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
add_custom_target(example_gemm_add_add_mean_meansquare_xdl)
add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16)
endif()
add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16)
add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
endif()
add_example_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp16)
endif()
add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp16)
add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8)
endif()
add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8)
add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8)
endif()
add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8)
add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32)
endif()
add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32)
add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32)
endif()
add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32)
add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16)
endif()
add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16)
add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16)
endif()
add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16)
add_dependencies(example_gemm_reduce_xdl
add_example_dependencies(example_gemm_reduce_xdl
example_gemm_reduce_xdl_mean_meansquare
example_gemm_reduce_xdl_max
example_gemm_add_add_mean_meansquare_xdl)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
endif()
add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
endif()
set(target 1)
endif()
......
example/20_grouped_conv_bwd_weight/CMakeLists.txt
View file @ fdddc8f4
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list_wmma gfx1100 gfx1101 gfx1102)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
if(gpu IN_LIST gpu_list_xdl AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_weight)
add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16)
endif()
add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16)
add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16)
endif()
add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16)
if(GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR GPU_TARGETS MATCHES "gfx942")
add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8 grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8)
add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8)
endif()
set(target 1)
endif()
if(gpu IN_LIST gpu_list_wmma AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_weight)
add_example_executable(example_grouped_conv_bwd_weight_wmma_fp16 grouped_conv_bwd_weight_wmma_fp16.cpp)
add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_wmma_fp16)
set(target 1)
endif()
endforeach()
add_custom_target(example_grouped_conv_bwd_weight_dl)
add_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
endif()
add_example_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
example/20_grouped_conv_bwd_weight/common.hpp
View file @ fdddc8f4
......@@ -46,25 +46,21 @@ struct CommonLayoutSetting
using OutputLayout = OutputLay;
};
template <ck::index_t NDimSpatial>
struct CommonLayoutSettingSelector;
namespace ctl = ck::tensor_layout::convolution;
template <>
struct CommonLayoutSettingSelector<1> final : CommonLayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK>
{
};
template <>
struct CommonLayoutSettingSelector<2> final
: CommonLayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK>
{
};
template <>
struct CommonLayoutSettingSelector<3> final
: CommonLayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK>
template <ck::index_t NDimSpatial>
struct CommonLayoutSettingSelector
: CommonLayoutSetting<ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GNDHWC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GKZYXC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWK,
ck::tensor_layout::convolution::GNHWK,
ck::tensor_layout::convolution::GNDHWK>>>
{
};
......@@ -87,7 +83,7 @@ struct ExecutionConfig final
#define DefaultConvParam \
ck::utils::conv::ConvParam \
{ \
2, 4, 1, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, { 1, 1 } \
3, 4, 1, 128, 256, {3, 3, 3}, {14, 14, 14}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, { 1, 1, 1 } \
}
inline void print_help_msg()
......
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp
View file @ fdddc8f4
......@@ -76,4 +76,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
int main(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return 1;
}
switch(conv_param.num_dim_spatial_)
{
case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
default: break;
}
return 1;
}
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp 0 → 100644
View file @ fdddc8f4
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp"
using InDataType = F16;
using WeiDataType = F16;
using OutDataType = F16;
using AccDataType = F32;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = PassThrough;
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Wmma_CShuffle<
NDimSpatial,
ck::tensor_layout::convolution::GNDHWC,
ck::tensor_layout::convolution::GKZYXC,
ck::tensor_layout::convolution::GNDHWK,
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
16, // MPerWMMA
16, // NPerWMMA
4, // MRepeat
2, // NRepeat
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<0, 2, 1>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1>, // ABlockTransferSrcAccessOrder
1, // ABlockTransferSrcVectorDim
1, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
true, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<0, 2, 1>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1>, // BBlockTransferSrcAccessOrder
1, // BBlockTransferSrcVectorDim
1, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
true, // BBlockLdsExtraN
4,
2,
S<1, 32, 1, 8>,
1>;
template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return 1;
}
switch(conv_param.num_dim_spatial_)
{
case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
default: break;
}
return 1;
}
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
View file @ fdddc8f4
......@@ -78,4 +78,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
int main(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return 1;
}
switch(conv_param.num_dim_spatial_)
{
case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
default: break;
}
return 1;
}
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
View file @ fdddc8f4
......@@ -77,4 +77,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
int main(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return 1;
}
switch(conv_param.num_dim_spatial_)
{
case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
default: break;
}
return 1;
}
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp
View file @ fdddc8f4
......@@ -83,4 +83,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
int main(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return 1;
}
switch(conv_param.num_dim_spatial_)
{
case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
default: break;
}
return 1;
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment