Skip to content

GitLab

Unverified Commit ae8b307a authored by Po Yen Chen's avatar Po Yen Chen Committed by GitHub
Browse files

Merge branch 'develop' into feature/support-readfirstlane-for-object-types

parents ad8bc60b ac9e01e2
Hide whitespace changes
Inline Side-by-side
Showing with 569 additions and 227 deletions
Jenkinsfile
View file @ ae8b307a
...@@ -493,10 +493,11 @@ def Build_CK(Map conf=[:]){ ...@@ -493,10 +493,11 @@ def Build_CK(Map conf=[:]){
{ {
cmake_build(conf) cmake_build(conf)
dir("build"){ dir("build"){
//run tests and examples
sh 'make -j\$(( \$(nproc) / 2 )) check'
if (navi_node == 0 ){ if (navi_node == 0 ){
//run tests and examples on all nodes except Navi //we only need the ckProfiler to run the performance tests, so we pack and stash it
sh 'make -j check' //do not stash profiler on Navi nodes
//we only need the ckProfiler to run the performance tests, so we pack and stash it
sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler' sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler'
stash "ckProfiler.tar.gz" stash "ckProfiler.tar.gz"
} }
...@@ -686,12 +687,31 @@ pipeline { ...@@ -686,12 +687,31 @@ pipeline {
{ {
parallel parallel
{ {
stage("Build CK and run Tests on MI100/MI200/MI300")
{
when {
beforeAgent true
expression { params.RUN_FULL_QA.toBoolean() }
}
agent{ label rocmnode("gfx908 || gfx90a") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940" -D CMAKE_CXX_COMPILER="${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')
}
}
stage("Build CK and run Tests on MI100/MI200") stage("Build CK and run Tests on MI100/MI200")
{ {
when {
beforeAgent true
expression { !params.RUN_FULL_QA.toBoolean() }
}
agent{ label rocmnode("gfx908 || gfx90a") } agent{ label rocmnode("gfx908 || gfx90a") }
environment{ environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" """ setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908,gfx90a" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
} }
steps{ 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') Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
...@@ -705,8 +725,8 @@ pipeline { ...@@ -705,8 +725,8 @@ pipeline {
} }
agent{ label rocmnode("navi21") } agent{ label rocmnode("navi21") }
environment{ environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install """ setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1030" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1030;gfx1100;gfx1101;gfx1102" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1030" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
} }
steps{ steps{
......
client_example/18_groupnorm/groupnorm_swish.cpp
View file @ ae8b307a
...@@ -131,11 +131,12 @@ int main(int argc, char* argv[]) ...@@ -131,11 +131,12 @@ int main(int argc, char* argv[])
} }
} }
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
// run the best intance // run the best intance
if(found)
{ {
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
auto& op_ptr = op_ptrs[best_op_id]; auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString() std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl; << std::endl;
......
client_example/19_pool_fwd/CMakeLists.txt 0 → 100644
View file @ ae8b307a
add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp)
target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_operations)
add_executable(client_avg_pool3d_fwd avg_pool3d_fwd.cpp)
target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_operations)
\ No newline at end of file
client_example/19_pool_fwd/avg_pool3d_fwd.cpp 0 → 100644
View file @ ae8b307a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
using InDataType = ck::half_t;
using OutDataType = ck::half_t;
using IndexDataType = int32_t;
constexpr ck::index_t InOutRank = 5;
constexpr ck::index_t WindowRank = 3;
#if 0
constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
constexpr bool OutputIndex = false;
#else
constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
constexpr bool OutputIndex = false;
#endif
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main(int argc, char* argv[])
{
ck::index_t N = 2;
ck::index_t C = 32;
ck::index_t Z = 2;
ck::index_t Y = 2;
ck::index_t X = 2;
ck::index_t Di = 30;
ck::index_t Hi = 30;
ck::index_t Wi = 30;
ck::index_t window_stride_d = 2;
ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2;
ck::index_t in_left_pad_d = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_d = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Z) / window_stride_d + 1;
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
// Pool API only support the order of NCDHW
std::vector<ck::index_t> in_length = {N, C, Di, Hi, Wi};
std::vector<ck::index_t> out_length = {N, C, Do, Ho, Wo};
std::vector<ck::index_t> window_spatial_lengths = {Z, Y, X};
std::vector<ck::index_t> window_strides = {window_stride_d, window_stride_h, window_stride_w};
std::vector<ck::index_t> input_left_pads = {in_left_pad_d, in_left_pad_h, in_left_pad_w};
std::vector<ck::index_t> input_right_pads = {in_right_pad_d, in_right_pad_h, in_right_pad_w};
std::size_t in_tensor_size = N * C * Di * Hi * Wi;
std::size_t out_tensor_size = N * C * Do * Ho * Wo;
// tensor layout = NDHWC
std::vector<ck::index_t> in_tensor_stride = {Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C};
std::vector<ck::index_t> out_tensor_stride = {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C};
SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
SimpleDeviceMem out_indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
ReduceOpId,
OutputIndex>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
in_length,
window_spatial_lengths,
out_length,
in_tensor_stride,
out_tensor_stride,
out_tensor_stride,
window_strides,
input_left_pads,
input_right_pads,
{2, 3, 4});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_bytes =
in_tensor_size * sizeof(InDataType) + out_tensor_size * sizeof(OutDataType);
if constexpr(OutputIndex)
num_bytes += out_tensor_size * sizeof(IndexDataType);
float gb_per_sec = num_bytes / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
<< op_name << std::endl;
if(ave_time < best_ave_time)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
// run the best intance
if(found)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
in_length,
window_spatial_lengths,
out_length,
in_tensor_stride,
out_tensor_stride,
out_tensor_stride,
window_strides,
input_left_pads,
input_right_pads,
{2, 3, 4});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
client_example/19_pool_fwd/max_pool2d_fwd.cpp 0 → 100644
View file @ ae8b307a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/pool2d_fwd.hpp"
using InDataType = ck::half_t;
using OutDataType = ck::half_t;
using IndexDataType = int32_t;
constexpr ck::index_t InOutRank = 4;
constexpr ck::index_t WindowRank = 2;
#if 1
constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
constexpr bool OutputIndex = true;
#else
constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
constexpr bool OutputIndex = false;
#endif
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main(int argc, char* argv[])
{
ck::index_t N = 2;
ck::index_t C = 32;
ck::index_t Y = 2;
ck::index_t X = 2;
ck::index_t Hi = 30;
ck::index_t Wi = 30;
ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
// Pool API only support the order of NCHW
std::vector<ck::index_t> in_length = {N, C, Hi, Wi};
std::vector<ck::index_t> out_length = {N, C, Ho, Wo};
std::vector<ck::index_t> window_spatial_lengths = {Y, X};
std::vector<ck::index_t> window_strides = {window_stride_h, window_stride_w};
std::vector<ck::index_t> input_left_pads = {in_left_pad_h, in_left_pad_w};
std::vector<ck::index_t> input_right_pads = {in_right_pad_h, in_right_pad_w};
std::size_t in_tensor_size = N * C * Hi * Wi;
std::size_t out_tensor_size = N * C * Ho * Wo;
// tensor layout = NHWC
std::vector<ck::index_t> in_tensor_stride = {C * Hi * Wi, 1, Wi * C, C};
std::vector<ck::index_t> out_tensor_stride = {C * Ho * Wo, 1, Wo * C, C};
SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
SimpleDeviceMem out_indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
ReduceOpId,
OutputIndex>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
in_length,
window_spatial_lengths,
out_length,
in_tensor_stride,
out_tensor_stride,
out_tensor_stride,
window_strides,
input_left_pads,
input_right_pads,
{2, 3});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_bytes =
in_tensor_size * sizeof(InDataType) + out_tensor_size * sizeof(OutDataType);
if constexpr(OutputIndex)
num_bytes += out_tensor_size * sizeof(IndexDataType);
float gb_per_sec = num_bytes / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
<< op_name << std::endl;
if(ave_time < best_ave_time)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
// run the best intance
if(found)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
in_length,
window_spatial_lengths,
out_length,
in_tensor_stride,
out_tensor_stride,
out_tensor_stride,
window_strides,
input_left_pads,
input_right_pads,
{2, 3});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
example/02_gemm_bilinear/CMakeLists.txt
View file @ ae8b307a
add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102") if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp) add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
endif() endif()
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
endif()
example/03_gemm_bias_relu/CMakeLists.txt
View file @ ae8b307a
add_example_executable(example_gemm_bias_relu_xdl_fp16 gemm_bias_relu_xdl_fp16.cpp) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_example_executable(example_gemm_bias_relu_xdl_fp16 gemm_bias_relu_xdl_fp16.cpp)
endif()
\ No newline at end of file
example/04_gemm_add_add_fastgelu/CMakeLists.txt
View file @ ae8b307a
add_custom_target(example_gemm_add_add_fastgelu_xdl) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
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) add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp) add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp) add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp)
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_add_add_fastgelu_xdl_int4 gemm_add_add_fastgelu_xdl_int4.cpp) add_example_executable(example_gemm_add_add_fastgelu_xdl_int4 gemm_add_add_fastgelu_xdl_int4.cpp)
endif(USE_BITINT_EXTENSION_INT4) endif(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp) add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16) add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp16) add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp16)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp32) add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp32)
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4) add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4) endif(USE_BITINT_EXTENSION_INT4)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8) add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
endif()
\ No newline at end of file
example/09_convnd_fwd/CMakeLists.txt
View file @ ae8b307a
add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp) add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp) add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp) add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp) # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
endif()
add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp) add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp) add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp) add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
......
example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
View file @ ae8b307a
add_custom_target(example_convnd_fwd_reduce_xdl) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_custom_target(example_convnd_fwd_reduce_xdl)
add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp) add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp) add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp) add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp) add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8) add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16) add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16) add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
add_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)
if(USE_BITINT_EXTENSION_INT4) add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)
add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp) endif(USE_BITINT_EXTENSION_INT4)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4) endif()
endif(USE_BITINT_EXTENSION_INT4) \ No newline at end of file
example/13_pool2d_fwd/pool2d_fwd_common.hpp
View file @ ae8b307a
...@@ -17,115 +17,11 @@ ...@@ -17,115 +17,11 @@
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp" #include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
template <typename InDataType, template <typename InDataType,
typename OutDataType, typename OutDataType,
typename AccDataType, typename ComputeDataType,
typename IndexDataType,
ck::ReduceTensorOp ReduceOpId,
bool PropagateNan,
bool OutputIndex>
static void pool_host_verify(const Tensor<InDataType>& in,
Tensor<OutDataType>& out,
Tensor<IndexDataType>& out_indices,
const std::array<ck::index_t, 2>& window_spatial_lengths,
const std::array<ck::index_t, 2>& window_strides,
const std::array<ck::index_t, 2>& in_left_pads,
const std::array<ck::index_t, 2>& /*in_right_pads*/)
{
const int32_t reduceLength = window_spatial_lengths[0] * window_spatial_lengths[1];
using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
auto elementwise_ops =
ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);
auto in_elementwise_op = std::get<0>(elementwise_ops);
auto acc_elementwise_op = std::get<1>(elementwise_ops);
if constexpr(!OutputIndex)
{
using Accumulation =
ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
{
ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
{
ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < static_cast<ck::index_t>(in.mDesc.GetLengths()[2]) &&
wi >= 0 && wi < static_cast<ck::index_t>(in.mDesc.GetLengths()[3]))
{
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
in_elementwise_op(currVal, currVal);
Accumulation::Calculate(accuVal, currVal);
}
}
}
acc_elementwise_op(accuVal, accuVal);
out(n, c, ho, wo) = accuVal;
};
make_ParallelTensorFunctor(f_nchw,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
}
else
{
using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
ReduceOperation,
AccDataType,
IndexDataType>;
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
IndexDataType accuIndex = 0;
for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
{
ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
{
ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
{
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
IndexDataType currIndex = y * window_spatial_lengths[1] + x;
in_elementwise_op(currVal, currVal);
Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
}
}
}
acc_elementwise_op(accuVal, accuVal);
out(n, c, ho, wo) = accuVal;
out_indices(n, c, ho, wo) = accuIndex;
};
make_ParallelTensorFunctor(f_nchw,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
};
}
template <typename InDataType,
typename OutDataType,
typename AccDataType,
typename IndexDataType, typename IndexDataType,
typename InLayout, typename InLayout,
typename OutLayout, typename OutLayout,
...@@ -150,9 +46,10 @@ bool pool_test(bool do_verification, ...@@ -150,9 +46,10 @@ bool pool_test(bool do_verification,
{ {
using DevicePoolFwdInstance = using DevicePoolFwdInstance =
ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C< ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<
InDataType, // InDataType InDataType, // InDataType
OutDataType, // OutDataType OutDataType, // OutDataType
AccDataType, // AccDataType IndexDataType, // IndexDataType
ComputeDataType, // ComputeDataType
ReduceOpId, ReduceOpId,
OutputIndex, OutputIndex,
64, // BlockSize 64, // BlockSize
...@@ -165,10 +62,10 @@ bool pool_test(bool do_verification, ...@@ -165,10 +62,10 @@ bool pool_test(bool do_verification,
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1; const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1; const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
const std::array<ck::index_t, 2> window_spatial_lengths{{Y, X}}; const std::vector<ck::index_t> window_spatial_lengths{Y, X};
const std::array<ck::index_t, 2> window_strides{{window_stride_h, window_stride_w}}; const std::vector<ck::index_t> window_strides{window_stride_h, window_stride_w};
const std::array<ck::index_t, 2> input_left_pads{{in_left_pad_h, in_left_pad_w}}; const std::vector<ck::index_t> input_left_pads{in_left_pad_h, in_left_pad_w};
const std::array<ck::index_t, 2> input_right_pads{{in_right_pad_h, in_right_pad_w}}; const std::vector<ck::index_t> input_right_pads{in_right_pad_h, in_right_pad_w};
// tensor layout // tensor layout
auto f_host_tensor_descriptor = auto f_host_tensor_descriptor =
...@@ -219,14 +116,16 @@ bool pool_test(bool do_verification, ...@@ -219,14 +116,16 @@ bool pool_test(bool do_verification,
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()), static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()), static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()), static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
N, {N, C, Hi, Wi},
C, {Y, X},
std::array<ck::index_t, 2>{{Hi, Wi}}, {N, C, Ho, Wo},
std::array<ck::index_t, 2>{{Y, X}}, {C * Hi * Wi, 1, Wi * C, C},
std::array<ck::index_t, 2>{{Ho, Wo}}, {C * Ho * Wo, 1, Wo * C, C},
{C * Ho * Wo, 1, Wo * C, C},
window_strides, window_strides,
input_left_pads, input_left_pads,
input_right_pads); input_right_pads,
{2, 3});
if(!pool.IsSupportedArgument(argument_ptr.get())) if(!pool.IsSupportedArgument(argument_ptr.get()))
{ {
...@@ -252,19 +151,28 @@ bool pool_test(bool do_verification, ...@@ -252,19 +151,28 @@ bool pool_test(bool do_verification,
if(do_verification) if(do_verification)
{ {
pool_host_verify<InDataType, using ReferencePoolingFwdInstance =
OutDataType, ck::tensor_operation::host::ReferencePoolingFwd<4,
AccDataType, 2,
IndexDataType, InDataType,
ReduceOpId, OutDataType,
PropagateNan, ComputeDataType,
OutputIndex>(in_n_c_hi_wi, IndexDataType,
out_n_c_ho_wo_host, ReduceOpId,
out_indices_n_c_ho_wo_host, PropagateNan,
window_spatial_lengths, OutputIndex>;
window_strides,
input_left_pads, auto ref_pooling = ReferencePoolingFwdInstance{};
input_right_pads); auto ref_pooling_invoker = ref_pooling.MakeInvoker();
auto ref_pooling_argument = ref_pooling.MakeArgument(in_n_c_hi_wi,
out_n_c_ho_wo_host,
out_indices_n_c_ho_wo_host,
window_spatial_lengths,
window_strides,
input_left_pads,
input_right_pads);
ref_pooling_invoker.Run(ref_pooling_argument);
out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data()); out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
......
example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
View file @ ae8b307a
...@@ -2,7 +2,6 @@ ...@@ -2,7 +2,6 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <cstdlib>
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
...@@ -10,9 +9,9 @@ ...@@ -10,9 +9,9 @@
#include "pool2d_fwd_common.hpp" #include "pool2d_fwd_common.hpp"
using InDataType = ck::half_t; using InDataType = ck::half_t;
using OutDataType = ck::half_t; using OutDataType = ck::half_t;
using AccDataType = float; using ComputeDataType = float;
using IndexDataType = int32_t; using IndexDataType = int32_t;
...@@ -91,7 +90,7 @@ int main(int argc, char* argv[]) ...@@ -91,7 +90,7 @@ int main(int argc, char* argv[])
bool pass = pool_test<InDataType, bool pass = pool_test<InDataType,
OutDataType, OutDataType,
AccDataType, ComputeDataType,
IndexDataType, IndexDataType,
InLayout, InLayout,
OutLayout, OutLayout,
......
example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
View file @ ae8b307a
...@@ -2,7 +2,6 @@ ...@@ -2,7 +2,6 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <cstdlib>
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp" #include "ck/utility/reduction_enums.hpp"
...@@ -10,9 +9,9 @@ ...@@ -10,9 +9,9 @@
#include "pool2d_fwd_common.hpp" #include "pool2d_fwd_common.hpp"
using InDataType = float; using InDataType = float;
using OutDataType = float; using OutDataType = float;
using AccDataType = float; using ComputeDataType = float;
using IndexDataType = int32_t; using IndexDataType = int32_t;
...@@ -91,7 +90,7 @@ int main(int argc, char* argv[]) ...@@ -91,7 +90,7 @@ int main(int argc, char* argv[])
bool pass = pool_test<InDataType, bool pass = pool_test<InDataType,
OutDataType, OutDataType,
AccDataType, ComputeDataType,
IndexDataType, IndexDataType,
InLayout, InLayout,
OutLayout, OutLayout,
......
example/14_gemm_quantization/CMakeLists.txt
View file @ ae8b307a
...@@ -2,5 +2,7 @@ ...@@ -2,5 +2,7 @@
add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp) add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
# xdlops # xdlops
add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp) add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)
\ No newline at end of file add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp)
endif()
\ No newline at end of file
example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
View file @ ae8b307a
add_custom_target(example_gemm_reduce_xdl) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_custom_target(example_gemm_reduce_xdl_max) add_custom_target(example_gemm_reduce_xdl)
add_custom_target(example_gemm_reduce_xdl_mean_meansquare) add_custom_target(example_gemm_reduce_xdl_max)
add_custom_target(example_gemm_add_add_mean_meansquare_xdl) 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) add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp) add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp) add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp) add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp) add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp) add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp) add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp) add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp) add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
add_dependencies(example_gemm_reduce_xdl_max add_dependencies(example_gemm_reduce_xdl_max
example_gemm_max_xdl_bf16 example_gemm_max_xdl_bf16
example_gemm_max_xdl_fp16 example_gemm_max_xdl_fp16
example_gemm_max_xdl_fp32 example_gemm_max_xdl_fp32
example_gemm_max_xdl_int8) example_gemm_max_xdl_int8)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare add_dependencies(example_gemm_reduce_xdl_mean_meansquare
example_gemm_mean_meansquare_xdl_fp16 example_gemm_mean_meansquare_xdl_fp16
example_gemm_mean_meansquare_xdl_fp32 example_gemm_mean_meansquare_xdl_fp32
example_gemm_mean_meansquare_xdl_bf16 example_gemm_mean_meansquare_xdl_bf16
example_gemm_add_addsquare_xdl_int8) example_gemm_add_addsquare_xdl_int8)
add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16) add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
add_dependencies(example_gemm_reduce_xdl add_dependencies(example_gemm_reduce_xdl
example_gemm_reduce_xdl_mean_meansquare example_gemm_reduce_xdl_mean_meansquare
example_gemm_reduce_xdl_max example_gemm_reduce_xdl_max
example_gemm_add_add_mean_meansquare_xdl) example_gemm_add_add_mean_meansquare_xdl)
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp) add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4) add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
endif()
endif() endif()
example/17_convnd_bwd_data/CMakeLists.txt
View file @ ae8b307a
add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility) add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
endif()
add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp) add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility) target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
example/18_batched_gemm_reduce/CMakeLists.txt
View file @ ae8b307a
add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp)
endif()
example/20_grouped_conv_bwd_weight/CMakeLists.txt
View file @ ae8b307a
add_custom_target(example_grouped_conv_bwd_weight) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
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) add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp) add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16
example_grouped_conv_bwd_weight_xdl_bf16) example_grouped_conv_bwd_weight_xdl_bf16)
endif()
add_custom_target(example_grouped_conv_bwd_weight_dl) add_custom_target(example_grouped_conv_bwd_weight_dl)
......
example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
View file @ ae8b307a
...@@ -18,7 +18,9 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config, ...@@ -18,7 +18,9 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
// Set split_k = 2 for xdl op, split_k = 1 for dl // Set split_k = 2 for xdl op, split_k = 1 for dl
// Dl op doesn't support split_k > 1 // Dl op doesn't support split_k > 1
// TODO: Add Dl op split_k > 1 support // TODO: Add Dl op split_k > 1 support
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030")) if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102"))
{ {
split_k = 2; split_k = 2;
} }
......
example/21_gemm_layernorm/CMakeLists.txt
View file @ ae8b307a
add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_welford_fp16 gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_naive_fp16 gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp) add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_welford_fp16 gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp)
add_example_executable(example_gemm_layernorm_xdl_naive_fp16 gemm_layernorm_xdl_naive_fp16.cpp) add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_naive_fp16 gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp)
add_example_executable(example_gemm_xdl_layernorm_naive_single_kernel_fp16 gemm_xdl_layernorm_naive_single_kernel_fp16.cpp) add_example_executable(example_gemm_layernorm_xdl_naive_fp16 gemm_layernorm_xdl_naive_fp16.cpp)
add_example_executable(example_gemm_xdl_layernorm_naive_single_kernel_fp16 gemm_xdl_layernorm_naive_single_kernel_fp16.cpp)
endif()
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