Skip to content

GitLab

"...pipelines/stable_diffusion/test_onnx_stable_diffusion.py" did not exist on "2d35f6733a2d698e8917896071444a5923993ae7"
Commit 24672339 authored by Po-Yen, Chen's avatar Po-Yen, Chen
Browse files

Merge branch 'feature/integrage-karg-simplification-pr' into feature/test

parents f2c5ca5a 853e797e
Show whitespace changes
Inline Side-by-side
Showing with 537 additions and 183 deletions
.github/dependabot.yml
View file @ 24672339
...@@ -6,7 +6,7 @@ ...@@ -6,7 +6,7 @@
version: 2 version: 2
updates: updates:
- package-ecosystem: "pip" # See documentation for possible values - package-ecosystem: "pip" # See documentation for possible values
directory: "/docs/.sphinx" # Location of package manifests directory: "/docs/sphinx" # Location of package manifests
open-pull-requests-limit: 10 open-pull-requests-limit: 10
schedule: schedule:
interval: "daily" interval: "daily"
.gitignore
View file @ 24672339
...@@ -49,10 +49,10 @@ build* ...@@ -49,10 +49,10 @@ build*
install.dir* install.dir*
# documentation artifacts # documentation artifacts
build/
_build/ _build/
_images/ _images/
_static/ _static/
_templates/ _templates/
_toc.yml _toc.yml
docBin/ docBin/
_doxygen/
.readthedocs.yaml
View file @ 24672339
...@@ -11,8 +11,8 @@ build: ...@@ -11,8 +11,8 @@ build:
sphinx: sphinx:
configuration: docs/conf.py configuration: docs/conf.py
formats: [htmlzip] formats: [htmlzip, pdf, epub]
python: python:
install: install:
- requirements: docs/.sphinx/requirements.txt - requirements: docs/sphinx/requirements.txt
Jenkinsfile
View file @ 24672339
...@@ -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
sh 'make -j check'
//we only need the ckProfiler to run the performance tests, so we pack and stash it //we only need the ckProfiler to run the performance tests, so we pack and stash it
//do not stash profiler on Navi nodes
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{
......
README.md
View file @ 24672339
# Composable Kernel # Composable Kernel
## Methodology ## Methodology
Composable Kernel (CK) library aims to provide a programming model for writing performance critical kernels for machine learning workloads across multiple architectures including GPUs, CPUs, etc, through general purpose kernel languages, like HIP C++. Composable Kernel (CK) library aims to provide a programming model for writing performance critical kernels for machine learning workloads across multiple architectures including GPUs, CPUs, etc, through general purpose kernel languages, like HIP C++.
CK utilizes two concepts to achieve performance portability and code maintainability: CK utilizes two concepts to achieve performance portability and code maintainability:
...@@ -10,6 +11,7 @@ CK utilizes two concepts to achieve performance portability and code maintainabi ...@@ -10,6 +11,7 @@ CK utilizes two concepts to achieve performance portability and code maintainabi
![ALT](/docs/data/ck_component.png "CK Components") ![ALT](/docs/data/ck_component.png "CK Components")
## Code Structure ## Code Structure
Current CK library are structured into 4 layers: Current CK library are structured into 4 layers:
* "Templated Tile Operators" layer * "Templated Tile Operators" layer
* "Templated Kernel and Invoker" layer * "Templated Kernel and Invoker" layer
...@@ -24,30 +26,35 @@ Run the steps below to build documentation locally. ...@@ -24,30 +26,35 @@ Run the steps below to build documentation locally.
``` ```
cd docs cd docs
pip3 install -r .sphinx/requirements.txt pip3 install -r sphinx/requirements.txt
python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html
``` ```
## Contributors ## Contributors
The list of developers and contributors is here: [Contributors](/CONTRIBUTORS.md) The list of developers and contributors is here: [Contributors](/CONTRIBUTORS.md)
## Citation ## Citation
If you use CK, please use following citations: If you use CK, please use following citations:
* CK paper will be freely available on arXiv soon: [Realizing Tensor Operators Using Coordinate Transformations and Tile Based Programming](???) * CK paper will be freely available on arXiv soon: [Realizing Tensor Operators Using Coordinate Transformations and Tile Based Programming](???)
* [CITATION.cff](/CITATION.cff) * [CITATION.cff](/CITATION.cff)
## License ## License
CK is released under the MIT license. [License File](/LICENSE) CK is released under the MIT license. [License File](/LICENSE)
# Build CK # Build CK
## Build docker image ## Build docker image
```bash ```bash
DOCKER_BUILDKIT=1 docker build -t ck:latest -f Dockerfile . DOCKER_BUILDKIT=1 docker build -t ck:latest -f Dockerfile .
``` ```
## Launch docker ## Launch docker
```bash ```bash
docker run \ docker run \
-it \ -it \
...@@ -60,10 +67,12 @@ ck:latest \ ...@@ -60,10 +67,12 @@ ck:latest \
``` ```
## Build CK ## Build CK
```bash ```bash
mkdir build && cd build mkdir build && cd build
# Need to specify target ID, example below is for gfx908 and gfx90a # Need to specify target ID, example below is for gfx908 and gfx90a
cmake \ cmake \
-D CMAKE_PREFIX_PATH=/opt/rocm \ -D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \ -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
...@@ -74,6 +83,7 @@ cmake ...@@ -74,6 +83,7 @@ cmake
``` ```
### Build examples and tests ### Build examples and tests
```bash ```bash
make -j examples tests make -j examples tests
make test make test
...@@ -83,21 +93,25 @@ Instructions for running each individual examples are under [example](/example) ...@@ -83,21 +93,25 @@ Instructions for running each individual examples are under [example](/example)
## Build ckProfiler ## Build ckProfiler
```bash ```bash
make -j ckProfiler make -j ckProfiler
``` ```
Instructions for running ckProfiler are under [profiler](/profiler) Instructions for running ckProfiler are under [profiler](/profiler)
## Install CK ## Install CK
```bash ```bash
make install make install
``` ```
## Using CK as pre-built kernel library ## Using CK as pre-built kernel library
Instructions for using CK as a pre-built kernel library are under [client_example](/client_example) Instructions for using CK as a pre-built kernel library are under [client_example](/client_example)
## Caveat ## Caveat
### Kernel Timing and Verification ### Kernel Timing and Verification
CK's own kernel timer will warn up kernel once, and then run it multiple times CK's own kernel timer will warn up kernel once, and then run it multiple times
to get average kernel time. For some kernels that use atomic add, this will cause to get average kernel time. For some kernels that use atomic add, this will cause
output buffer to be accumulated multiple times, causing verification failure. output buffer to be accumulated multiple times, causing verification failure.
......
client_example/18_groupnorm/groupnorm_swish.cpp
View file @ 24672339
...@@ -131,11 +131,12 @@ int main(int argc, char* argv[]) ...@@ -131,11 +131,12 @@ int main(int argc, char* argv[])
} }
} }
// run the best intance
if(found)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, " std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl; << best_op_name << std::endl;
// run the best intance
{
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 @ 24672339
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 @ 24672339
// 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 @ 24672339
// 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;
}
docs/conf.py
View file @ 24672339
...@@ -4,10 +4,21 @@ ...@@ -4,10 +4,21 @@
# list see the documentation: # list see the documentation:
# https://www.sphinx-doc.org/en/master/usage/configuration.html # https://www.sphinx-doc.org/en/master/usage/configuration.html
import subprocess
from rocm_docs import ROCmDocs from rocm_docs import ROCmDocs
docs_core = ROCmDocs("Composable Kernel Documentation")
docs_core.run_doxygen() name = "Composable Kernel"
get_version = r'sed -n -e "s/^rocm_setup_version(.* \([0-9\.]\{1,\}\).*/\1/p" ../CMakeLists.txt'
version = subprocess.getoutput(get_version)
if len(version) > 0:
name = f"{name} {version}"
external_toc_path = "./sphinx/_toc.yml"
docs_core = ROCmDocs(f"{name} Documentation")
docs_core.run_doxygen(doxygen_root="doxygen", doxygen_path="doxygen/docBin/xml")
docs_core.setup() docs_core.setup()
mathjax3_config = { mathjax3_config = {
......
docs/license.rst 0 → 100644
View file @ 24672339
=======
License
=======
.. include:: ../LICENSE
:literal:
docs/sphinx/_toc.yml.in 0 → 100644
View file @ 24672339
# Anywhere {branch} is used, the branch name will be substituted.
# These comments will also be removed.
defaults:
numbered: False
maxdepth: 6
root: index
subtrees:
- caption: About
entries:
- file: license
docs/.sphinx/requirements.in → docs/sphinx/requirements.in
View file @ 24672339
rocm-docs-core==0.2.0 rocm-docs-core==0.10.3
sphinxcontrib-bibtex==2.5.0 sphinxcontrib-bibtex==2.5.0
docs/.sphinx/requirements.txt → docs/sphinx/requirements.txt
View file @ 24672339
# #
# This file is autogenerated by pip-compile with Python 3.10 # This file is autogenerated by pip-compile with Python 3.8
# by the following command: # by the following command:
# #
# pip-compile .sphinx/requirements.in # pip-compile requirements.in
# #
accessible-pygments==0.0.3 accessible-pygments==0.0.3
# via pydata-sphinx-theme # via pydata-sphinx-theme
alabaster==0.7.13 alabaster==0.7.13
# via sphinx # via sphinx
asttokens==2.2.1
# via stack-data
attrs==22.2.0
# via
# jsonschema
# jupyter-cache
babel==2.12.1 babel==2.12.1
# via # via
# pydata-sphinx-theme # pydata-sphinx-theme
# sphinx # sphinx
backcall==0.2.0
# via ipython
beautifulsoup4==4.11.2 beautifulsoup4==4.11.2
# via pydata-sphinx-theme # via pydata-sphinx-theme
breathe==4.34.0 breathe==4.34.0
...@@ -27,19 +19,15 @@ breathe==4.34.0 ...@@ -27,19 +19,15 @@ breathe==4.34.0
certifi==2022.12.7 certifi==2022.12.7
# via requests # via requests
cffi==1.15.1 cffi==1.15.1
# via pynacl # via
# cryptography
# pynacl
charset-normalizer==3.1.0 charset-normalizer==3.1.0
# via requests # via requests
click==8.1.3 click==8.1.3
# via # via sphinx-external-toc
# jupyter-cache cryptography==40.0.2
# sphinx-external-toc # via pyjwt
comm==0.1.2
# via ipykernel
debugpy==1.6.6
# via ipykernel
decorator==5.1.1
# via ipython
deprecated==1.2.13 deprecated==1.2.13
# via pygithub # via pygithub
docutils==0.16 docutils==0.16
...@@ -48,52 +36,26 @@ docutils==0.16 ...@@ -48,52 +36,26 @@ docutils==0.16
# myst-parser # myst-parser
# pybtex-docutils # pybtex-docutils
# pydata-sphinx-theme # pydata-sphinx-theme
# rocm-docs-core
# sphinx # sphinx
# sphinxcontrib-bibtex # sphinxcontrib-bibtex
executing==1.2.0
# via stack-data
fastjsonschema==2.16.3
# via nbformat
gitdb==4.0.10 gitdb==4.0.10
# via gitpython # via gitpython
gitpython==3.1.31 gitpython==3.1.31
# via rocm-docs-core # via rocm-docs-core
greenlet==2.0.2
# via sqlalchemy
idna==3.4 idna==3.4
# via requests # via requests
imagesize==1.4.1 imagesize==1.4.1
# via sphinx # via sphinx
importlib-metadata==6.0.0 importlib-metadata==6.0.0
# via # via
# jupyter-cache # sphinx
# myst-nb # sphinxcontrib-bibtex
ipykernel==6.21.3 importlib-resources==5.12.0
# via myst-nb # via rocm-docs-core
ipython==8.11.0
# via
# ipykernel
# myst-nb
jedi==0.18.2
# via ipython
jinja2==3.1.2 jinja2==3.1.2
# via # via
# myst-parser # myst-parser
# sphinx # sphinx
jsonschema==4.17.3
# via nbformat
jupyter-cache==0.5.0
# via myst-nb
jupyter-client==8.0.3
# via
# ipykernel
# nbclient
jupyter-core==5.3.0
# via
# ipykernel
# jupyter-client
# nbformat
latexcodec==2.0.1 latexcodec==2.0.1
# via pybtex # via pybtex
linkify-it-py==1.0.3 linkify-it-py==1.0.3
...@@ -104,54 +66,16 @@ markdown-it-py==2.2.0 ...@@ -104,54 +66,16 @@ markdown-it-py==2.2.0
# myst-parser # myst-parser
markupsafe==2.1.2 markupsafe==2.1.2
# via jinja2 # via jinja2
matplotlib-inline==0.1.6
# via
# ipykernel
# ipython
mdit-py-plugins==0.3.5 mdit-py-plugins==0.3.5
# via myst-parser # via myst-parser
mdurl==0.1.2 mdurl==0.1.2
# via markdown-it-py # via markdown-it-py
myst-nb==0.17.1 myst-parser[linkify]==1.0.0
# via rocm-docs-core # via rocm-docs-core
myst-parser[linkify]==0.18.1
# via
# myst-nb
# rocm-docs-core
nbclient==0.5.13
# via
# jupyter-cache
# myst-nb
nbformat==5.7.3
# via
# jupyter-cache
# myst-nb
# nbclient
nest-asyncio==1.5.6
# via
# ipykernel
# nbclient
packaging==23.0 packaging==23.0
# via # via
# ipykernel
# pydata-sphinx-theme # pydata-sphinx-theme
# sphinx # sphinx
parso==0.8.3
# via jedi
pexpect==4.8.0
# via ipython
pickleshare==0.7.5
# via ipython
platformdirs==3.1.1
# via jupyter-core
prompt-toolkit==3.0.38
# via ipython
psutil==5.9.4
# via ipykernel
ptyprocess==0.7.0
# via pexpect
pure-eval==0.2.2
# via stack-data
pybtex==0.24.0 pybtex==0.24.0
# via # via
# pybtex-docutils # pybtex-docutils
...@@ -160,57 +84,47 @@ pybtex-docutils==1.0.2 ...@@ -160,57 +84,47 @@ pybtex-docutils==1.0.2
# via sphinxcontrib-bibtex # via sphinxcontrib-bibtex
pycparser==2.21 pycparser==2.21
# via cffi # via cffi
pydata-sphinx-theme==0.13.1 pydata-sphinx-theme==0.13.3
# via sphinx-book-theme # via
pygithub==1.57 # rocm-docs-core
# sphinx-book-theme
pygithub==1.58.2
# via rocm-docs-core # via rocm-docs-core
pygments==2.14.0 pygments==2.14.0
# via # via
# accessible-pygments # accessible-pygments
# ipython
# pydata-sphinx-theme # pydata-sphinx-theme
# sphinx # sphinx
pyjwt==2.6.0 pyjwt[crypto]==2.6.0
# via pygithub # via pygithub
pynacl==1.5.0 pynacl==1.5.0
# via pygithub # via pygithub
pyrsistent==0.19.3 pytz==2023.3
# via jsonschema # via babel
python-dateutil==2.8.2
# via jupyter-client
pyyaml==6.0 pyyaml==6.0
# via # via
# jupyter-cache
# myst-nb
# myst-parser # myst-parser
# pybtex # pybtex
# sphinx-external-toc # sphinx-external-toc
pyzmq==25.0.1
# via
# ipykernel
# jupyter-client
requests==2.28.2 requests==2.28.2
# via # via
# pygithub # pygithub
# sphinx # sphinx
rocm-docs-core==0.2.0 rocm-docs-core==0.10.3
# via -r .sphinx/requirements.in # via -r requirements.in
six==1.16.0 six==1.16.0
# via # via
# asttokens
# latexcodec # latexcodec
# pybtex # pybtex
# python-dateutil
smmap==5.0.0 smmap==5.0.0
# via gitdb # via gitdb
snowballstemmer==2.2.0 snowballstemmer==2.2.0
# via sphinx # via sphinx
soupsieve==2.4 soupsieve==2.4
# via beautifulsoup4 # via beautifulsoup4
sphinx==4.3.1 sphinx==5.3.0
# via # via
# breathe # breathe
# myst-nb
# myst-parser # myst-parser
# pydata-sphinx-theme # pydata-sphinx-theme
# rocm-docs-core # rocm-docs-core
...@@ -220,7 +134,7 @@ sphinx==4.3.1 ...@@ -220,7 +134,7 @@ sphinx==4.3.1
# sphinx-external-toc # sphinx-external-toc
# sphinx-notfound-page # sphinx-notfound-page
# sphinxcontrib-bibtex # sphinxcontrib-bibtex
sphinx-book-theme==1.0.0rc2 sphinx-book-theme==1.0.1
# via rocm-docs-core # via rocm-docs-core
sphinx-copybutton==0.5.1 sphinx-copybutton==0.5.1
# via rocm-docs-core # via rocm-docs-core
...@@ -233,7 +147,7 @@ sphinx-notfound-page==0.8.3 ...@@ -233,7 +147,7 @@ sphinx-notfound-page==0.8.3
sphinxcontrib-applehelp==1.0.4 sphinxcontrib-applehelp==1.0.4
# via sphinx # via sphinx
sphinxcontrib-bibtex==2.5.0 sphinxcontrib-bibtex==2.5.0
# via -r .sphinx/requirements.in # via -r requirements.in
sphinxcontrib-devhelp==1.0.2 sphinxcontrib-devhelp==1.0.2
# via sphinx # via sphinx
sphinxcontrib-htmlhelp==2.0.1 sphinxcontrib-htmlhelp==2.0.1
...@@ -244,40 +158,15 @@ sphinxcontrib-qthelp==1.0.3 ...@@ -244,40 +158,15 @@ sphinxcontrib-qthelp==1.0.3
# via sphinx # via sphinx
sphinxcontrib-serializinghtml==1.1.5 sphinxcontrib-serializinghtml==1.1.5
# via sphinx # via sphinx
sqlalchemy==1.4.46
# via jupyter-cache
stack-data==0.6.2
# via ipython
tabulate==0.9.0
# via jupyter-cache
tornado==6.2
# via
# ipykernel
# jupyter-client
traitlets==5.9.0
# via
# comm
# ipykernel
# ipython
# jupyter-client
# jupyter-core
# matplotlib-inline
# nbclient
# nbformat
typing-extensions==4.5.0 typing-extensions==4.5.0
# via # via pydata-sphinx-theme
# myst-nb
# myst-parser
uc-micro-py==1.0.1 uc-micro-py==1.0.1
# via linkify-it-py # via linkify-it-py
urllib3==1.26.15 urllib3==1.26.15
# via requests # via requests
wcwidth==0.2.6
# via prompt-toolkit
wrapt==1.15.0 wrapt==1.15.0
# via deprecated # via deprecated
zipp==3.15.0 zipp==3.15.0
# via importlib-metadata # via
# importlib-metadata
# The following packages are considered to be unsafe in a requirements file: # importlib-resources
# setuptools
example/02_gemm_bilinear/CMakeLists.txt
View file @ 24672339
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 @ 24672339
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 @ 24672339
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 @ 24672339
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)
......
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