Skip to content

GitLab

Commit 261d3267 authored by Bartlomiej Wroblewski's avatar Bartlomiej Wroblewski
Browse files

Merge remote-tracking branch 'origin/develop' into bwroblew/direct_loads

parents 2d5b22fe f2398f61
Hide whitespace changes
Inline Side-by-side
Showing with 578 additions and 130 deletions
Jenkinsfile
View file @ 261d3267
...@@ -139,7 +139,7 @@ def buildDocker(install_prefix){ ...@@ -139,7 +139,7 @@ def buildDocker(install_prefix){
else{ else{
echo "Checking for image: ${image_name}" echo "Checking for image: ${image_name}"
sh "docker manifest inspect --insecure ${image_name}" sh "docker manifest inspect --insecure ${image_name}"
echo "Image: ${image_name} found!! Skipping building image" echo "Image: ${image_name} found! Skipping building image"
} }
} }
catch(Exception ex){ catch(Exception ex){
...@@ -213,8 +213,10 @@ def cmake_build(Map conf=[:]){ ...@@ -213,8 +213,10 @@ def cmake_build(Map conf=[:]){
if (setup_args.contains("gfx94")){ if (setup_args.contains("gfx94")){
invocation_tag="gfx94" invocation_tag="gfx94"
} }
echo "invocation tag: ${invocation_tag}"
def redis_pre_setup_cmd = pre_setup_cmd
if(check_host() && params.USE_SCCACHE && "${env.CK_SCCACHE}" != "null" && "${invocation_tag}" != "") { if(check_host() && params.USE_SCCACHE && "${env.CK_SCCACHE}" != "null" && "${invocation_tag}" != "") {
pre_setup_cmd = pre_setup_cmd + """ redis_pre_setup_cmd = pre_setup_cmd + """
#!/bin/bash #!/bin/bash
export ROCM_PATH=/opt/rocm export ROCM_PATH=/opt/rocm
export SCCACHE_ENABLED=true export SCCACHE_ENABLED=true
...@@ -228,18 +230,30 @@ def cmake_build(Map conf=[:]){ ...@@ -228,18 +230,30 @@ def cmake_build(Map conf=[:]){
export SCCACHE_C_CUSTOM_CACHE_BUSTER="${invocation_tag}" export SCCACHE_C_CUSTOM_CACHE_BUSTER="${invocation_tag}"
echo \$SCCACHE_C_CUSTOM_CACHE_BUSTER echo \$SCCACHE_C_CUSTOM_CACHE_BUSTER
stunnel ../script/redis-cli.conf stunnel ../script/redis-cli.conf
( ../script/sccache_wrapper.sh --enforce_redis
set -e
../script/sccache_wrapper.sh --enforce_redis
)
error_code=\$?
if [ \$error_code -ne 0 ]; then
echo "could not connect to the redis server. using sccache locally."
../script/sccache_wrapper.sh
fi
""" """
setup_args = " -DCMAKE_CXX_COMPILER_LAUNCHER=sccache -DCMAKE_C_COMPILER_LAUNCHER=sccache " + setup_args try {
def cmd1 = conf.get("cmd1", """
${redis_pre_setup_cmd}
""")
sh cmd1
setup_args = " -DCMAKE_CXX_COMPILER_LAUNCHER=sccache -DCMAKE_C_COMPILER_LAUNCHER=sccache " + setup_args
}
catch(Exception err){
echo "could not connect to redis server: ${err.getMessage()}. will not use sccache."
def cmd2 = conf.get("cmd2", """
${pre_setup_cmd}
""")
sh cmd2
}
}
else{
def cmd3 = conf.get("cmd3", """
${pre_setup_cmd}
""")
sh cmd3
} }
def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ") def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ")
// reduce parallelism when compiling, clang uses too much memory // reduce parallelism when compiling, clang uses too much memory
def nt = nthreads() def nt = nthreads()
...@@ -247,14 +261,16 @@ def cmake_build(Map conf=[:]){ ...@@ -247,14 +261,16 @@ def cmake_build(Map conf=[:]){
def execute_cmd = conf.get("execute_cmd", "") def execute_cmd = conf.get("execute_cmd", "")
def cmd = conf.get("cmd", """ def cmd = conf.get("cmd", """
${pre_setup_cmd}
${setup_cmd} ${setup_cmd}
${build_cmd} ${build_cmd}
${execute_cmd} ${execute_cmd}
""") """)
echo cmd echo cmd
sh cmd
dir("build"){
sh cmd
}
// Only archive from master or develop // Only archive from master or develop
if (package_build == true && (env.BRANCH_NAME == "develop" || env.BRANCH_NAME == "amd-master")) { if (package_build == true && (env.BRANCH_NAME == "develop" || env.BRANCH_NAME == "amd-master")) {
...@@ -686,8 +702,8 @@ pipeline { ...@@ -686,8 +702,8 @@ pipeline {
description: "Use the CK build to verify hipTensor build and tests (default: ON)") description: "Use the CK build to verify hipTensor build and tests (default: ON)")
string( string(
name: 'hipTensor_branch', name: 'hipTensor_branch',
defaultValue: 'mainline', defaultValue: 'develop',
description: 'Specify which branch of hipTensor to use (default: mainline)') description: 'Specify which branch of hipTensor to use (default: develop)')
booleanParam( booleanParam(
name: "USE_SCCACHE", name: "USE_SCCACHE",
defaultValue: true, defaultValue: true,
...@@ -751,7 +767,7 @@ pipeline { ...@@ -751,7 +767,7 @@ pipeline {
} }
agent{ label rocmnode("gfx908 || gfx90a") } agent{ label rocmnode("gfx908 || gfx90a") }
environment{ environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" """ setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" -DCMAKE_EXE_LINKER_FLAGS=" -L ${env.WORKSPACE}/script -T hip_fatbin_insert " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" -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;gfx940;gfx941;gfx942" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
} }
steps{ steps{
......
client_example/05_layernorm/CMakeLists.txt
View file @ 261d3267
add_executable(client_layernorm2d layernorm2d.cpp) add_executable(client_layernorm2d_fwd layernorm2d_fwd.cpp)
target_link_libraries(client_layernorm2d PRIVATE composable_kernel::device_operations) target_link_libraries(client_layernorm2d_fwd PRIVATE composable_kernel::device_operations)
add_executable(client_layernorm4d_fwd layernorm4d_fwd.cpp)
target_link_libraries(client_layernorm4d_fwd PRIVATE composable_kernel::device_operations)
client_example/05_layernorm/layernorm2d.cpp → client_example/05_layernorm/layernorm2d_fwd.cpp
View file @ 261d3267
...@@ -7,10 +7,10 @@ ...@@ -7,10 +7,10 @@
#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"
#include "ck/tensor_operation/gpu/device/device_normalization.hpp" #include "ck/tensor_operation/gpu/device/device_normalization_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/normalization.hpp" #include "ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp"
using XDataType = ck::half_t; using XDataType = ck::half_t;
using GammaDataType = ck::half_t; using GammaDataType = ck::half_t;
...@@ -57,14 +57,14 @@ int main(int argc, char* argv[]) ...@@ -57,14 +57,14 @@ int main(int argc, char* argv[])
SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * M); SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * M);
#endif #endif
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType, using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
SaveMeanInvStdDataType, SaveMeanInvStdDataType,
PassThrough, PassThrough,
Rank, Rank,
NumReduceDim>; NumReduceDim>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......
client_example/05_layernorm/layernorm4d_fwd.cpp 0 → 100644
View file @ 261d3267
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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_normalization_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
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 = 256;
ck::index_t H = 16;
ck::index_t W = 16;
ck::index_t C = 8;
std::vector<ck::index_t> strideXY = {H * W * C, W * C, C, 1};
std::vector<ck::index_t> strideGammaBeta = {0, W * C, C, 1};
std::vector<ck::index_t> strideSaveMeanInvStd = {1};
SimpleDeviceMem x_device_buf(sizeof(XDataType) * N * H * W * C);
SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * H * W * C);
SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * H * W * C);
SimpleDeviceMem y_device_buf(sizeof(YDataType) * N * H * W * C);
#ifdef SAVE_MEAN_INV_STD
SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * N);
SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * N);
#endif
using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
Rank,
NumReduceDim>;
// 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({N, H, W, C}, // lengths
strideXY, // xStrides
strideGammaBeta, // gammaStrides
strideGammaBeta, // betaStrides
strideXY, // yStrides
strideSaveMeanInvStd, // save_mean Strides
strideSaveMeanInvStd, // save_inv_std Strides
{1, 2, 3}, // 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();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_byte =
sizeof(XDataType) * N * H * W * C + sizeof(GammaDataType) * H * W * C +
sizeof(BetaDataType) * H * W * C + sizeof(YDataType) * N * H * W * C;
#ifdef SAVE_MEAN_INV_STD
num_byte += sizeof(SaveMeanInvStdDataType) * N * 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, "
<< 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;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
// run the best intance
{
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({N, H, W, C}, // lengths
strideXY, // xStrides
strideGammaBeta, // gammaStrides
strideGammaBeta, // betaStrides
strideXY, // yStrides
strideSaveMeanInvStd, // save_mean Strides
strideSaveMeanInvStd, // save_inv_std Strides
{1, 2, 3}, // 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();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp
View file @ 261d3267
...@@ -100,18 +100,18 @@ int main() ...@@ -100,18 +100,18 @@ int main()
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Wo * K); SimpleDeviceMem out(sizeof(OutDataType) * G * N * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<>, ck::Tuple<>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<>, ck::Tuple<>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>; PassThrough>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......
client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
View file @ 261d3267
...@@ -71,18 +71,18 @@ int main() ...@@ -71,18 +71,18 @@ int main()
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<>, ck::Tuple<>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<>, ck::Tuple<>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>; PassThrough>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......
client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
View file @ 261d3267
...@@ -80,7 +80,7 @@ int main(int argc, char* argv[]) ...@@ -80,7 +80,7 @@ int main(int argc, char* argv[])
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
View file @ 261d3267
...@@ -78,18 +78,18 @@ int main(int argc, char* argv[]) ...@@ -78,18 +78,18 @@ int main(int argc, char* argv[])
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<BiasLayout>, ck::Tuple<BiasLayout>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<BiasDataType>, ck::Tuple<BiasDataType>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
View file @ 261d3267
...@@ -83,7 +83,7 @@ int main(int argc, char* argv[]) ...@@ -83,7 +83,7 @@ int main(int argc, char* argv[])
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
View file @ 261d3267
...@@ -79,18 +79,18 @@ int main(int argc, char* argv[]) ...@@ -79,18 +79,18 @@ int main(int argc, char* argv[])
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<BiasLayout>, ck::Tuple<BiasLayout>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<BiasDataType>, ck::Tuple<BiasDataType>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
View file @ 261d3267
...@@ -76,19 +76,19 @@ int main(int argc, char* argv[]) ...@@ -76,19 +76,19 @@ int main(int argc, char* argv[])
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<RequantScaleLayout>, ck::Tuple<RequantScaleLayout>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<RequantScaleDataType>, ck::Tuple<RequantScaleDataType>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
View file @ 261d3267
...@@ -72,18 +72,18 @@ int main(int argc, char* argv[]) ...@@ -72,18 +72,18 @@ int main(int argc, char* argv[])
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<>, ck::Tuple<>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<>, ck::Tuple<>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
View file @ 261d3267
...@@ -2,8 +2,10 @@ add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_f ...@@ -2,8 +2,10 @@ add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_f
add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp) add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp) add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp) add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_operations)
client_example/11_grouped_conv_bwd_weight/common.hpp
View file @ 261d3267
...@@ -85,7 +85,9 @@ template <ck::index_t NumDimSpatial, ...@@ -85,7 +85,9 @@ template <ck::index_t NumDimSpatial,
typename OutDataType, typename OutDataType,
typename InLayout, typename InLayout,
typename WeiLayout, typename WeiLayout,
typename OutLayout> typename OutLayout,
typename AComputeType = InDataType,
typename BComputeType = AComputeType>
bool run_grouped_conv_bwd_weight( bool run_grouped_conv_bwd_weight(
const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths, const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths,
const std::array<ck::index_t, NumDimSpatial + 3>& input_strides, const std::array<ck::index_t, NumDimSpatial + 3>& input_strides,
...@@ -113,7 +115,9 @@ bool run_grouped_conv_bwd_weight( ...@@ -113,7 +115,9 @@ bool run_grouped_conv_bwd_weight(
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>; PassThrough,
AComputeType,
BComputeType>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp 0 → 100644
View file @ 261d3267
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using AComputeType = ck::bf8_t;
using BComputeType = ck::f8_t;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 8;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 128;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 28;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
int main()
{
return run_grouped_conv_bwd_weight<NumDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InLayout,
WeiLayout,
OutLayout,
AComputeType,
BComputeType>(input_lengths,
input_strides,
filter_lengths,
weights_strides,
output_lengths,
output_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads)
? EXIT_SUCCESS
: EXIT_FAILURE;
}
client_example/16_convnd_fwd/common.hpp
View file @ 261d3267
...@@ -11,7 +11,7 @@ ...@@ -11,7 +11,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp" #include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
...@@ -174,19 +174,19 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD ...@@ -174,19 +174,19 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD
std::size_t flop = GetFlops<NumDimSpatial>(out_lengths, wei_lengths); std::size_t flop = GetFlops<NumDimSpatial>(out_lengths, wei_lengths);
std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size; std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size;
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<>, ck::Tuple<>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<>, ck::Tuple<>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough, PassThrough,
ComputeType>; ComputeType>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/18_groupnorm/groupnorm_swish.cpp
View file @ 261d3267
...@@ -7,10 +7,10 @@ ...@@ -7,10 +7,10 @@
#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"
#include "ck/tensor_operation/gpu/device/device_normalization.hpp" #include "ck/tensor_operation/gpu/device/device_normalization_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/normalization_swish.hpp" #include "ck/library/tensor_operation_instance/gpu/normalization_fwd_swish.hpp"
using XDataType = ck::half_t; using XDataType = ck::half_t;
using GammaDataType = float; using GammaDataType = float;
...@@ -64,14 +64,14 @@ int main(int argc, char* argv[]) ...@@ -64,14 +64,14 @@ int main(int argc, char* argv[])
SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * N * G); SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * N * G);
#endif #endif
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType, using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
SaveMeanInvStdDataType, SaveMeanInvStdDataType,
Swish, Swish,
Rank, Rank,
NumReduceDim>; NumReduceDim>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......
client_example/23_elementwise_transpose/CMakeLists.txt 0 → 100644
View file @ 261d3267
add_executable(client_elementwise_transpose3d elementwise_transpose_3d.cpp)
target_link_libraries(client_elementwise_transpose3d PRIVATE composable_kernel::device_operations)
client_example/23_elementwise_transpose/elementwise_transpose_3d.cpp 0 → 100644
View file @ 261d3267
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/transpose_3d.hpp"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
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()
{
const int N = 16;
const int C = 8;
const int D = 8;
const int H = 8;
const int W = 8;
std::vector<std::size_t> ncdhw = {N, C, D, H, W};
std::vector<std::size_t> nchwd = {N, C, H, W, D};
auto size = N * C * D * H * W;
std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, C, H, W, D
SimpleDeviceMem a_dev_buf(sizeof(ADataType) * size);
SimpleDeviceMem b_dev_buf(sizeof(BDataType) * size);
std::array<const void*, 1> input = {a_dev_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_dev_buf.GetDeviceBuffer()};
using DeviceElementwisePermuteInstance = ck::tensor_operation::device::
DeviceElementwise<ck::Tuple<ADataType>, ck::Tuple<BDataType>, PassThrough, 5>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceElementwisePermuteInstance>::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(
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
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_byte =
sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
float gb_per_sec = num_byte / 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;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
// run the best intance
{
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(
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
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/23_grouped_convnd_fwd_scaleadd_scaleadd_relu/CMakeLists.txt 0 → 100644
View file @ 261d3267
add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 grouped_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp)
target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 PRIVATE composable_kernel::device_operations)
add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp16 grouped_conv_fwd_scaleadd_scaleadd_relu_fp16.cpp)
target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp16 PRIVATE composable_kernel::device_operations)
add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_bf16 grouped_conv_fwd_scaleadd_scaleadd_relu_bf16.cpp)
target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_bf16 PRIVATE composable_kernel::device_operations)
add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_int8 grouped_conv_fwd_scaleadd_scaleadd_relu_int8.cpp)
target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_int8 PRIVATE composable_kernel::device_operations)
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