Skip to content

GitLab

Commit 48fe8532 authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'amd-develop' into amd-master

parents d88411d1 39002e9e
Hide whitespace changes
Inline Side-by-side
Showing with 454 additions and 243 deletions
CHANGELOG.md
View file @ 48fe8532
...@@ -25,8 +25,8 @@ Full documentation for Composable Kernel is not yet available. ...@@ -25,8 +25,8 @@ Full documentation for Composable Kernel is not yet available.
- Added multi-embeddings support (#542). - Added multi-embeddings support (#542).
- Added Navi3x blockwise GEMM and real GEMM support (#541). - Added Navi3x blockwise GEMM and real GEMM support (#541).
- Added Navi grouped ConvBwdWeight support (#505). - Added Navi grouped ConvBwdWeight support (#505).
- Added pool3d forward (#697). - Added MaxPool, AvgPool forward (#815).
- Added maxpool backward (#750). - Added MaxPool backward (#750).
### Changed ### Changed
- Changed ... - Changed ...
CMakeLists.txt
View file @ 48fe8532
...@@ -93,7 +93,7 @@ find_package(hip) ...@@ -93,7 +93,7 @@ find_package(hip)
# SWDEV-413293 and https://reviews.llvm.org/D155213 # SWDEV-413293 and https://reviews.llvm.org/D155213
math(EXPR hip_VERSION_FLAT "(${hip_VERSION_MAJOR} * 1000 + ${hip_VERSION_MINOR}) * 100000 + ${hip_VERSION_PATCH}") math(EXPR hip_VERSION_FLAT "(${hip_VERSION_MAJOR} * 1000 + ${hip_VERSION_MINOR}) * 100000 + ${hip_VERSION_PATCH}")
message("hip_version_flat=${hip_VERSION_FLAT}") message("hip_version_flat=${hip_VERSION_FLAT}")
if(${hip_VERSION_FLAT} GREATER 500700000) if(${hip_VERSION_FLAT} GREATER 500723302)
message("Adding the fno-offload-uniform-block compiler flag") message("Adding the fno-offload-uniform-block compiler flag")
add_compile_options(-fno-offload-uniform-block) add_compile_options(-fno-offload-uniform-block)
endif() endif()
......
CONTRIBUTORS.md
View file @ 48fe8532
...@@ -6,9 +6,11 @@ This is the list of developers and contributors to Composable Kernel library ...@@ -6,9 +6,11 @@ This is the list of developers and contributors to Composable Kernel library
## Developers ## Developers
[Chao Liu](https://github.com/asroy), [Jing Zhang](https://github.com/zjing14), 2018-2023 [Chao Liu](https://github.com/asroy), [Jing Zhang](https://github.com/zjing14), 2018-2023
[Letao Qin](https://github.com/ltqin), [Qianfeng Zhang](https://github.com/qianfengz), [Liang Huang](https://github.com/carlushuang), [Shaojie Wang](https://github.com/shaojiewang), 2019-2022 [Letao Qin](https://github.com/ltqin), [Qianfeng Zhang](https://github.com/qianfengz), [Liang Huang](https://github.com/carlushuang), [Shaojie Wang](https://github.com/shaojiewang), 2019-2023
[Anthony Chang](https://github.com/rosenrodt), [Chunyu Lai](https://github.com/rocking5566), [Illia Silin](https://github.com/illsilin), [Adam Osewski](https://github.com/aosewski), [Poyen Chen](https://github.com/poyenc), [Rosty Geyyer](https://github.com/geyyer), 2022 [Anthony Chang](https://github.com/rosenrodt), [Chunyu Lai](https://github.com/rocking5566), [Illia Silin](https://github.com/illsilin), [Adam Osewski](https://github.com/aosewski), [Poyen Chen](https://github.com/poyenc), [Rosty Geyyer](https://github.com/geyyer), [Astha Rai](https://github.com/arai713), [Shi YanXing](https://github.com/Yanxing-Shi), 2022-2023
[Hari Sadasivan](https://github.com/hsadasiv), [Bartlomiej Kocot](https://github.com/bartekxk), [Bartlomiej Wroblewski](https://github.com/bwroblew), 2023
Hanwen Chang, 2019-2021, Hanwen Chang, 2019-2021,
......
Dockerfile
View file @ 48fe8532
...@@ -26,9 +26,14 @@ RUN if [ "$ROCMVERSION" != "5.7" ]; then \ ...@@ -26,9 +26,14 @@ RUN if [ "$ROCMVERSION" != "5.7" ]; then \
sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \ sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \ sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \
elif [ "$ROCMVERSION" = "5.7" ] && [ "$compiler_version" = "" ] || [ "$compiler_version" = "amd-stg-open" ]; then \ elif [ "$ROCMVERSION" = "5.7" ] && [ "$compiler_version" = "" ] || [ "$compiler_version" = "amd-stg-open" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.7-20.04-1_all.deb" && \ sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.7-20.04-1_all.deb" && \
apt update && apt-get install -y ./amdgpu-install-internal_5.7-20.04-1_all.deb && \ apt update && apt-get install -y ./amdgpu-install-internal_5.7-20.04-1_all.deb && \
amdgpu-repo --amdgpu-build=1609671 --rocm-build=compute-rocm-npi-mi300/1354; \ amdgpu-repo --amdgpu-build=1609671 --rocm-build=compute-rocm-npi-mi300/1354; \
elif [ "$ROCMVERSION" = "5.7" ] && [ "$compiler_version" = "rc1" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.7-20.04-1_all.deb" && \
apt update && apt-get install -y ./amdgpu-install-internal_5.7-20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 5.7 rel-19 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=1637781; \
fi fi
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list" RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
......
Jenkinsfile
View file @ 48fe8532
...@@ -612,7 +612,7 @@ def process_results(Map conf=[:]){ ...@@ -612,7 +612,7 @@ def process_results(Map conf=[:]){
} }
//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=5.7;COMPILER_VERSION=rc1
0 21 * * * % ROCMVERSION=5.6;COMPILER_VERSION=;COMPILER_COMMIT= 0 21 * * * % ROCMVERSION=5.6;COMPILER_VERSION=;COMPILER_COMMIT=
0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : "" 0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
...@@ -710,8 +710,8 @@ pipeline { ...@@ -710,8 +710,8 @@ pipeline {
} }
agent{ label rocmnode("gfx908 || gfx90a") } agent{ label rocmnode("gfx908 || gfx90a") }
environment{ environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940" """ setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941" """
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="${build_compiler()}" .. && make -j """ execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941" -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')
......
client_example/19_pool_fwd/avg_pool3d_fwd.cpp
View file @ 48fe8532
...@@ -16,6 +16,9 @@ using InDataType = ck::half_t; ...@@ -16,6 +16,9 @@ using InDataType = ck::half_t;
using OutDataType = ck::half_t; using OutDataType = ck::half_t;
using IndexDataType = int32_t; using IndexDataType = int32_t;
using InLayout = ck::tensor_layout::convolution::NDHWC;
using OutLayout = ck::tensor_layout::convolution::NDHWC;
constexpr ck::index_t InOutRank = 5; constexpr ck::index_t InOutRank = 5;
constexpr ck::index_t WindowRank = 3; constexpr ck::index_t WindowRank = 3;
#if 0 #if 0
...@@ -44,33 +47,41 @@ struct SimpleDeviceMem ...@@ -44,33 +47,41 @@ struct SimpleDeviceMem
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
ck::index_t N = 2; ck::index_t N = 2;
ck::index_t C = 32; ck::index_t C = 32;
ck::index_t Z = 2; ck::index_t Z = 2;
ck::index_t Y = 2; ck::index_t Y = 2;
ck::index_t X = 2; ck::index_t X = 2;
ck::index_t Di = 30; ck::index_t Di = 30;
ck::index_t Hi = 30; ck::index_t Hi = 30;
ck::index_t Wi = 30; ck::index_t Wi = 30;
ck::index_t window_stride_d = 2; ck::index_t window_stride_d = 2;
ck::index_t window_stride_h = 2; ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2; ck::index_t window_stride_w = 2;
ck::index_t in_left_pad_d = 1; ck::index_t window_dilation_d = 1;
ck::index_t in_left_pad_h = 1; ck::index_t window_dilation_h = 1;
ck::index_t in_left_pad_w = 1; ck::index_t window_dilation_w = 1;
ck::index_t in_right_pad_d = 1; ck::index_t in_left_pad_d = 1;
ck::index_t in_right_pad_h = 1; ck::index_t in_left_pad_h = 1;
ck::index_t in_right_pad_w = 1; ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_d = 1;
ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Z) / window_stride_d + 1; ck::index_t in_right_pad_h = 1;
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1; ck::index_t in_right_pad_w = 1;
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
const ck::index_t Zs = (Z - 1) * window_dilation_d + 1;
const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Zs) / window_stride_d + 1;
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
// Pool API only support the order of NCDHW // 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> in_length = {N, C, Di, Hi, Wi};
std::vector<ck::index_t> out_length = {N, C, Do, Ho, Wo}; 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_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> window_strides = {window_stride_d, window_stride_h, window_stride_w};
std::vector<ck::index_t> window_dilations{
window_dilation_d, window_dilation_h, window_dilation_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_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::vector<ck::index_t> input_right_pads = {in_right_pad_d, in_right_pad_h, in_right_pad_w};
...@@ -90,6 +101,8 @@ int main(int argc, char* argv[]) ...@@ -90,6 +101,8 @@ int main(int argc, char* argv[])
InDataType, InDataType,
OutDataType, OutDataType,
IndexDataType, IndexDataType,
InLayout,
OutLayout,
ReduceOpId, ReduceOpId,
OutputIndex>; OutputIndex>;
...@@ -122,6 +135,7 @@ int main(int argc, char* argv[]) ...@@ -122,6 +135,7 @@ int main(int argc, char* argv[])
out_tensor_stride, out_tensor_stride,
out_tensor_stride, out_tensor_stride,
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads, input_right_pads,
{2, 3, 4}); {2, 3, 4});
...@@ -181,6 +195,7 @@ int main(int argc, char* argv[]) ...@@ -181,6 +195,7 @@ int main(int argc, char* argv[])
out_tensor_stride, out_tensor_stride,
out_tensor_stride, out_tensor_stride,
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads, input_right_pads,
{2, 3, 4}); {2, 3, 4});
......
client_example/19_pool_fwd/max_pool2d_fwd.cpp
View file @ 48fe8532
...@@ -10,14 +10,18 @@ ...@@ -10,14 +10,18 @@
#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp" #include "ck/tensor_operation/gpu/device/device_pool_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/pool2d_fwd.hpp" #include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
using InDataType = ck::half_t; using InDataType = ck::half_t;
using OutDataType = ck::half_t; using OutDataType = ck::half_t;
using IndexDataType = int32_t; using IndexDataType = int32_t;
constexpr ck::index_t InOutRank = 4; // We use pool3d to implement pool2d in this example
constexpr ck::index_t WindowRank = 2; using InLayout = ck::tensor_layout::convolution::NDHWC;
using OutLayout = ck::tensor_layout::convolution::NDHWC;
constexpr ck::index_t InOutRank = 5;
constexpr ck::index_t WindowRank = 3;
#if 1 #if 1
constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX; constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
constexpr bool OutputIndex = true; constexpr bool OutputIndex = true;
...@@ -42,31 +46,66 @@ struct SimpleDeviceMem ...@@ -42,31 +46,66 @@ struct SimpleDeviceMem
void* p_mem_; void* p_mem_;
}; };
void TransformPool2dparamToPool3d(std::vector<ck::index_t>& input_lengths,
std::vector<ck::index_t>& window_lengths,
std::vector<ck::index_t>& output_lengths,
std::vector<ck::index_t>& input_stride,
std::vector<ck::index_t>& output_stride,
std::vector<ck::index_t>& indices_stride,
std::vector<ck::index_t>& window_strides,
std::vector<ck::index_t>& window_dilations,
std::vector<ck::index_t>& input_left_pads,
std::vector<ck::index_t>& input_right_pads,
std::vector<ck::index_t>& pooling_dims)
{
// NCHW to NCDHW
input_lengths.insert(input_lengths.begin() + 2, 1);
output_lengths.insert(output_lengths.begin() + 2, 1);
input_stride.insert(input_stride.begin() + 2, 0);
output_stride.insert(output_stride.begin() + 2, 0);
indices_stride.insert(indices_stride.begin() + 2, 0);
// YX to ZYX
window_lengths.insert(window_lengths.begin(), 1);
window_strides.insert(window_strides.begin(), 0);
window_dilations.insert(window_dilations.begin(), 0);
input_left_pads.insert(input_left_pads.begin(), 0);
input_right_pads.insert(input_right_pads.begin(), 0);
pooling_dims = {2, 3, 4};
}
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
ck::index_t N = 2; ck::index_t N = 2;
ck::index_t C = 32; ck::index_t C = 32;
ck::index_t Y = 2; ck::index_t Y = 2;
ck::index_t X = 2; ck::index_t X = 2;
ck::index_t Hi = 30; ck::index_t Hi = 30;
ck::index_t Wi = 30; ck::index_t Wi = 30;
ck::index_t window_stride_h = 2; ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2; ck::index_t window_stride_w = 2;
ck::index_t in_left_pad_h = 1; ck::index_t window_dilation_h = 1;
ck::index_t in_left_pad_w = 1; ck::index_t window_dilation_w = 1;
ck::index_t in_right_pad_h = 1; ck::index_t in_left_pad_h = 1;
ck::index_t in_right_pad_w = 1; ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1; ck::index_t in_right_pad_w = 1;
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
// Pool API only support the order of NCHW // Pool API only support the order of NCHW
std::vector<ck::index_t> in_length = {N, C, Hi, Wi}; 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> out_length = {N, C, Ho, Wo};
std::vector<ck::index_t> window_spatial_lengths = {Y, X}; 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> window_strides = {window_stride_h, window_stride_w};
std::vector<ck::index_t> window_dilations = {window_dilation_h, window_dilation_w};
std::vector<ck::index_t> input_left_pads = {in_left_pad_h, in_left_pad_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::vector<ck::index_t> input_right_pads = {in_right_pad_h, in_right_pad_w};
std::vector<ck::index_t> pooling_dims = {2, 3};
std::size_t in_tensor_size = N * C * Hi * Wi; std::size_t in_tensor_size = N * C * Hi * Wi;
std::size_t out_tensor_size = N * C * Ho * Wo; std::size_t out_tensor_size = N * C * Ho * Wo;
...@@ -75,6 +114,18 @@ int main(int argc, char* argv[]) ...@@ -75,6 +114,18 @@ int main(int argc, char* argv[])
std::vector<ck::index_t> in_tensor_stride = {C * Hi * Wi, 1, Wi * C, C}; 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}; std::vector<ck::index_t> out_tensor_stride = {C * Ho * Wo, 1, Wo * C, C};
TransformPool2dparamToPool3d(in_length,
window_spatial_lengths,
out_length,
in_tensor_stride,
out_tensor_stride,
out_tensor_stride,
window_strides,
window_dilations,
input_left_pads,
input_right_pads,
pooling_dims);
SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size); SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size); SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
SimpleDeviceMem out_indices_device_buf(sizeof(IndexDataType) * out_tensor_size); SimpleDeviceMem out_indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
...@@ -84,6 +135,8 @@ int main(int argc, char* argv[]) ...@@ -84,6 +135,8 @@ int main(int argc, char* argv[])
InDataType, InDataType,
OutDataType, OutDataType,
IndexDataType, IndexDataType,
InLayout,
OutLayout,
ReduceOpId, ReduceOpId,
OutputIndex>; OutputIndex>;
...@@ -116,9 +169,10 @@ int main(int argc, char* argv[]) ...@@ -116,9 +169,10 @@ int main(int argc, char* argv[])
out_tensor_stride, out_tensor_stride,
out_tensor_stride, out_tensor_stride,
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads, input_right_pads,
{2, 3}); pooling_dims);
auto invoker_ptr = op_ptr->MakeInvokerPointer(); auto invoker_ptr = op_ptr->MakeInvokerPointer();
...@@ -175,9 +229,10 @@ int main(int argc, char* argv[]) ...@@ -175,9 +229,10 @@ int main(int argc, char* argv[])
out_tensor_stride, out_tensor_stride,
out_tensor_stride, out_tensor_stride,
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads, input_right_pads,
{2, 3}); pooling_dims);
auto invoker_ptr = op_ptr->MakeInvokerPointer(); auto invoker_ptr = op_ptr->MakeInvokerPointer();
......
docs/sphinx/requirements.in
View file @ 48fe8532
rocm-docs-core==0.10.3 rocm-docs-core>=0.20.0
sphinxcontrib-bibtex==2.5.0 sphinxcontrib-bibtex==2.5.0
docs/sphinx/requirements.txt
View file @ 48fe8532
...@@ -38,6 +38,8 @@ docutils==0.16 ...@@ -38,6 +38,8 @@ docutils==0.16
# pydata-sphinx-theme # pydata-sphinx-theme
# sphinx # sphinx
# sphinxcontrib-bibtex # sphinxcontrib-bibtex
fastjsonschema==2.18.0
# via rocm-docs-core
gitdb==4.0.10 gitdb==4.0.10
# via gitpython # via gitpython
gitpython==3.1.31 gitpython==3.1.31
...@@ -46,20 +48,12 @@ idna==3.4 ...@@ -46,20 +48,12 @@ idna==3.4
# via requests # via requests
imagesize==1.4.1 imagesize==1.4.1
# via sphinx # via sphinx
importlib-metadata==6.0.0
# via
# sphinx
# sphinxcontrib-bibtex
importlib-resources==5.12.0
# via rocm-docs-core
jinja2==3.1.2 jinja2==3.1.2
# via # via
# myst-parser # myst-parser
# sphinx # sphinx
latexcodec==2.0.1 latexcodec==2.0.1
# via pybtex # via pybtex
linkify-it-py==1.0.3
# via myst-parser
markdown-it-py==2.2.0 markdown-it-py==2.2.0
# via # via
# mdit-py-plugins # mdit-py-plugins
...@@ -70,7 +64,7 @@ mdit-py-plugins==0.3.5 ...@@ -70,7 +64,7 @@ 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-parser[linkify]==1.0.0 myst-parser==1.0.0
# via rocm-docs-core # via rocm-docs-core
packaging==23.0 packaging==23.0
# via # via
...@@ -99,18 +93,17 @@ pyjwt[crypto]==2.6.0 ...@@ -99,18 +93,17 @@ pyjwt[crypto]==2.6.0
# via pygithub # via pygithub
pynacl==1.5.0 pynacl==1.5.0
# via pygithub # via pygithub
pytz==2023.3
# via babel
pyyaml==6.0 pyyaml==6.0
# via # via
# myst-parser # myst-parser
# pybtex # pybtex
# rocm-docs-core
# sphinx-external-toc # sphinx-external-toc
requests==2.28.2 requests==2.28.2
# via # via
# pygithub # pygithub
# sphinx # sphinx
rocm-docs-core==0.10.3 rocm-docs-core>=0.20.0
# via -r requirements.in # via -r requirements.in
six==1.16.0 six==1.16.0
# via # via
...@@ -160,13 +153,7 @@ sphinxcontrib-serializinghtml==1.1.5 ...@@ -160,13 +153,7 @@ sphinxcontrib-serializinghtml==1.1.5
# via sphinx # via sphinx
typing-extensions==4.5.0 typing-extensions==4.5.0
# via pydata-sphinx-theme # via pydata-sphinx-theme
uc-micro-py==1.0.1
# via linkify-it-py
urllib3==1.26.15 urllib3==1.26.15
# via requests # via requests
wrapt==1.15.0 wrapt==1.15.0
# via deprecated # via deprecated
zipp==3.15.0
# via
# importlib-metadata
# importlib-resources
example/01_gemm/CMakeLists.txt
View file @ 48fe8532
...@@ -6,6 +6,8 @@ if(DL_KERNELS) ...@@ -6,6 +6,8 @@ if(DL_KERNELS)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp) add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
add_dependencies(example_gemm_dl example_gemm_dl_fp16) add_dependencies(example_gemm_dl example_gemm_dl_fp16)
add_example_executable(example_gemm_dl_dpp8_fp16 gemm_dl_dpp8_fp16.cpp)
add_dependencies(example_gemm_dl example_gemm_dl_dpp8_fp16)
endif() endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES) if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp) add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
...@@ -64,3 +66,6 @@ if(DTYPES MATCHES "fp8" OR NOT DEFINED DTYPES) ...@@ -64,3 +66,6 @@ if(DTYPES MATCHES "fp8" OR NOT DEFINED DTYPES)
add_dependencies(example_gemm_xdl example_gemm_xdl_f8) add_dependencies(example_gemm_xdl example_gemm_xdl_f8)
endif() endif()
endif() endif()
add_example_executable(example_gemm_xdl_fp16_f8 gemm_xdl_fp16_f8.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp16_f8)
example/01_gemm/gemm_dl_dpp8_fp16.cpp 0 → 100644
View file @ 48fe8532
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using ALayout = Col;
using BLayout = Row;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDlDpp8
// ######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ######| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 128, 128, 16, 2, 1, 8, 8, S<8, 8>, S<4, 1>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
#include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/01_gemm/gemm_xdl_fp16_f8.cpp 0 → 100644
View file @ 48fe8532
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
using ADataType = ck::f8_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto LoopSched = ck::make_default_loop_scheduler();
static constexpr auto PipelineVer = ck::PipelineVersion::v1;
using ComputeType = ck::half_t;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// ######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Loop| Pipeline| ComputeType|
// ######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Scheduler| Version| |
// ######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | |
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopSched, PipelineVer, ComputeType>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
#include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/13_pool2d_fwd/pool2d_fwd_common.hpp
View file @ 48fe8532
...@@ -39,31 +39,35 @@ bool pool_test(bool do_verification, ...@@ -39,31 +39,35 @@ bool pool_test(bool do_verification,
ck::index_t Wi, ck::index_t Wi,
ck::index_t window_stride_h, ck::index_t window_stride_h,
ck::index_t window_stride_w, ck::index_t window_stride_w,
ck::index_t window_dilation_h,
ck::index_t window_dilation_w,
ck::index_t in_left_pad_h, ck::index_t in_left_pad_h,
ck::index_t in_left_pad_w, ck::index_t in_left_pad_w,
ck::index_t in_right_pad_h, ck::index_t in_right_pad_h,
ck::index_t in_right_pad_w) ck::index_t in_right_pad_w)
{ {
using DevicePoolFwdInstance = using DevicePoolFwdInstance =
ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C< ck::tensor_operation::device::DevicePool2dFwd_NHWC_NHWC<InDataType,
InDataType, // InDataType OutDataType,
OutDataType, // OutDataType IndexDataType,
IndexDataType, // IndexDataType ComputeDataType,
ComputeDataType, // ComputeDataType ReduceOpId,
ReduceOpId, OutputIndex,
OutputIndex, 64, // BlockSize
64, // BlockSize 64, // ReduceMThreadClusterSize
64, // ReduceMThreadClusterSize 1, // ReduceKThreadClusterSize
1, // ReduceKThreadClusterSize 4, // ReduceMThreadSliceSize
4, // ReduceMThreadSliceSize 1, // ReduceKThreadSliceSize
1, // ReduceKThreadSliceSize 1>; // InSrcOutDstVectorSize
4>; // InSrcOutDstVectorSize
const ck::index_t Ys = (Y - 1) * window_dilation_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 Xs = (X - 1) * window_dilation_w + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1; const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
const std::vector<ck::index_t> window_spatial_lengths{Y, X}; const std::vector<ck::index_t> window_spatial_lengths{Y, X};
const std::vector<ck::index_t> window_strides{window_stride_h, window_stride_w}; const std::vector<ck::index_t> window_strides{window_stride_h, window_stride_w};
const std::vector<ck::index_t> window_dilations{window_dilation_h, window_dilation_w};
const std::vector<ck::index_t> 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::vector<ck::index_t> 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};
...@@ -123,6 +127,7 @@ bool pool_test(bool do_verification, ...@@ -123,6 +127,7 @@ bool pool_test(bool do_verification,
{C * Ho * Wo, 1, Wo * C, C}, {C * Ho * Wo, 1, Wo * C, C},
{C * Ho * Wo, 1, Wo * C, C}, {C * Ho * Wo, 1, Wo * C, C},
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads, input_right_pads,
{2, 3}); {2, 3});
...@@ -144,8 +149,8 @@ bool pool_test(bool do_verification, ...@@ -144,8 +149,8 @@ bool pool_test(bool do_verification,
float gb_per_sec = num_btype / 1.E6 / ave_time; float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s" std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< std::endl; << " GB / s " << std::endl;
bool pass = true; bool pass = true;
...@@ -169,6 +174,7 @@ bool pool_test(bool do_verification, ...@@ -169,6 +174,7 @@ bool pool_test(bool do_verification,
out_indices_n_c_ho_wo_host, out_indices_n_c_ho_wo_host,
window_spatial_lengths, window_spatial_lengths,
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads); input_right_pads);
......
example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
View file @ 48fe8532
...@@ -34,18 +34,20 @@ int main(int argc, char* argv[]) ...@@ -34,18 +34,20 @@ int main(int argc, char* argv[])
bool time_kernel; bool time_kernel;
// Pool shape // Pool shape
ck::index_t N = 128; ck::index_t N = 128;
ck::index_t C = 192; ck::index_t C = 192;
ck::index_t Y = 3; ck::index_t Y = 3;
ck::index_t X = 3; ck::index_t X = 3;
ck::index_t Hi = 71; ck::index_t Hi = 71;
ck::index_t Wi = 71; ck::index_t Wi = 71;
ck::index_t window_stride_h = 2; ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2; ck::index_t window_stride_w = 2;
ck::index_t in_left_pad_h = 1; ck::index_t window_dilation_h = 1;
ck::index_t in_left_pad_w = 1; ck::index_t window_dilation_w = 1;
ck::index_t in_right_pad_h = 1; ck::index_t in_left_pad_h = 1;
ck::index_t in_right_pad_w = 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;
if(argc == 1) if(argc == 1)
{ {
...@@ -59,31 +61,33 @@ int main(int argc, char* argv[]) ...@@ -59,31 +61,33 @@ int main(int argc, char* argv[])
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
time_kernel = static_cast<bool>(std::stoi(argv[3])); time_kernel = static_cast<bool>(std::stoi(argv[3]));
} }
else if(argc == 16) else if(argc == 18)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
time_kernel = static_cast<bool>(std::stoi(argv[3])); time_kernel = static_cast<bool>(std::stoi(argv[3]));
N = std::stoi(argv[4]); N = std::stoi(argv[4]);
C = std::stoi(argv[5]); C = std::stoi(argv[5]);
Y = std::stoi(argv[6]); Y = std::stoi(argv[6]);
X = std::stoi(argv[7]); X = std::stoi(argv[7]);
Hi = std::stoi(argv[8]); Hi = std::stoi(argv[8]);
Wi = std::stoi(argv[9]); Wi = std::stoi(argv[9]);
window_stride_h = std::stoi(argv[10]); window_stride_h = std::stoi(argv[10]);
window_stride_w = std::stoi(argv[11]); window_stride_w = std::stoi(argv[11]);
in_left_pad_h = std::stoi(argv[12]); window_dilation_h = std::stoi(argv[12]);
in_left_pad_w = std::stoi(argv[13]); window_dilation_w = std::stoi(argv[13]);
in_right_pad_h = std::stoi(argv[14]); in_left_pad_h = std::stoi(argv[14]);
in_right_pad_w = std::stoi(argv[15]); in_left_pad_w = std::stoi(argv[15]);
in_right_pad_h = std::stoi(argv[16]);
in_right_pad_w = std::stoi(argv[17]);
} }
else else
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n"); printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, " printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n"); "RightPx\n");
exit(0); exit(0);
} }
...@@ -107,6 +111,8 @@ int main(int argc, char* argv[]) ...@@ -107,6 +111,8 @@ int main(int argc, char* argv[])
Wi, Wi,
window_stride_h, window_stride_h,
window_stride_w, window_stride_w,
window_dilation_h,
window_dilation_w,
in_left_pad_h, in_left_pad_h,
in_left_pad_w, in_left_pad_w,
in_right_pad_h, in_right_pad_h,
......
example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
View file @ 48fe8532
...@@ -34,18 +34,20 @@ int main(int argc, char* argv[]) ...@@ -34,18 +34,20 @@ int main(int argc, char* argv[])
bool time_kernel; bool time_kernel;
// Pool shape // Pool shape
ck::index_t N = 128; ck::index_t N = 128;
ck::index_t C = 192; ck::index_t C = 192;
ck::index_t Y = 3; ck::index_t Y = 3;
ck::index_t X = 3; ck::index_t X = 3;
ck::index_t Hi = 71; ck::index_t Hi = 71;
ck::index_t Wi = 71; ck::index_t Wi = 71;
ck::index_t window_stride_h = 2; ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2; ck::index_t window_stride_w = 2;
ck::index_t in_left_pad_h = 1; ck::index_t window_dilation_h = 1;
ck::index_t in_left_pad_w = 1; ck::index_t window_dilation_w = 1;
ck::index_t in_right_pad_h = 1; ck::index_t in_left_pad_h = 1;
ck::index_t in_right_pad_w = 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;
if(argc == 1) if(argc == 1)
{ {
...@@ -59,31 +61,33 @@ int main(int argc, char* argv[]) ...@@ -59,31 +61,33 @@ int main(int argc, char* argv[])
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
time_kernel = static_cast<bool>(std::stoi(argv[3])); time_kernel = static_cast<bool>(std::stoi(argv[3]));
} }
else if(argc == 16) else if(argc == 18)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
time_kernel = static_cast<bool>(std::stoi(argv[3])); time_kernel = static_cast<bool>(std::stoi(argv[3]));
N = std::stoi(argv[4]); N = std::stoi(argv[4]);
C = std::stoi(argv[5]); C = std::stoi(argv[5]);
Y = std::stoi(argv[6]); Y = std::stoi(argv[6]);
X = std::stoi(argv[7]); X = std::stoi(argv[7]);
Hi = std::stoi(argv[8]); Hi = std::stoi(argv[8]);
Wi = std::stoi(argv[9]); Wi = std::stoi(argv[9]);
window_stride_h = std::stoi(argv[10]); window_stride_h = std::stoi(argv[10]);
window_stride_w = std::stoi(argv[11]); window_stride_w = std::stoi(argv[11]);
in_left_pad_h = std::stoi(argv[12]); window_dilation_h = std::stoi(argv[12]);
in_left_pad_w = std::stoi(argv[13]); window_dilation_w = std::stoi(argv[13]);
in_right_pad_h = std::stoi(argv[14]); in_left_pad_h = std::stoi(argv[14]);
in_right_pad_w = std::stoi(argv[15]); in_left_pad_w = std::stoi(argv[15]);
in_right_pad_h = std::stoi(argv[16]);
in_right_pad_w = std::stoi(argv[17]);
} }
else else
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n"); printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, " printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n"); "RightPx\n");
exit(0); exit(0);
} }
...@@ -107,6 +111,8 @@ int main(int argc, char* argv[]) ...@@ -107,6 +111,8 @@ int main(int argc, char* argv[])
Wi, Wi,
window_stride_h, window_stride_h,
window_stride_w, window_stride_w,
window_dilation_h,
window_dilation_w,
in_left_pad_h, in_left_pad_h,
in_left_pad_w, in_left_pad_w,
in_right_pad_h, in_right_pad_h,
......
example/48_pool3d_fwd/pool3d_fwd_common.hpp
View file @ 48fe8532
...@@ -18,7 +18,45 @@ ...@@ -18,7 +18,45 @@
#include "ck/library/utility/literals.hpp" #include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
template <typename InDataType, template <typename TensorLayout>
std::vector<ck::index_t> f_tensor_strides_ncdhw(ck::index_t N_,
ck::index_t C_,
ck::index_t D,
ck::index_t H,
ck::index_t W,
TensorLayout layout)
{
using namespace ck::literals;
(void)N_;
if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
return {C_ * D * H * W, D * H * W, H * W, W, 1_uz};
else if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
return {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_};
};
template <typename TensorLayout>
HostTensorDescriptor f_host_tensor_descriptor(std::size_t N_,
std::size_t C_,
std::size_t D,
std::size_t H,
std::size_t W,
TensorLayout layout)
{
using namespace ck::literals;
if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
{
return HostTensorDescriptor({N_, C_, D, H, W}, {C_ * D * H * W, D * H * W, H * W, W, 1_uz});
}
else if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
{
return HostTensorDescriptor({N_, C_, D, H, W},
{D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
}
};
template <typename DevicePoolFwdInstance,
typename InDataType,
typename OutDataType, typename OutDataType,
typename ComputeDataType, typename ComputeDataType,
typename IndexDataType, typename IndexDataType,
...@@ -40,6 +78,9 @@ bool pool3d_test(bool do_verification, ...@@ -40,6 +78,9 @@ bool pool3d_test(bool do_verification,
ck::index_t window_stride_d, ck::index_t window_stride_d,
ck::index_t window_stride_h, ck::index_t window_stride_h,
ck::index_t window_stride_w, ck::index_t window_stride_w,
ck::index_t window_dilation_d,
ck::index_t window_dilation_h,
ck::index_t window_dilation_w,
ck::index_t in_left_pad_d, ck::index_t in_left_pad_d,
ck::index_t in_left_pad_h, ck::index_t in_left_pad_h,
ck::index_t in_left_pad_w, ck::index_t in_left_pad_w,
...@@ -47,53 +88,21 @@ bool pool3d_test(bool do_verification, ...@@ -47,53 +88,21 @@ bool pool3d_test(bool do_verification,
ck::index_t in_right_pad_h, ck::index_t in_right_pad_h,
ck::index_t in_right_pad_w) ck::index_t in_right_pad_w)
{ {
using DevicePoolFwdInstance = const ck::index_t Zs = (Z - 1) * window_dilation_d + 1;
ck::tensor_operation::device::DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C< const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
InDataType, // InDataType const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
OutDataType, // OutDataType const ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Zs) / window_stride_d + 1;
IndexDataType, // IndexDataType const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
ComputeDataType, // ComputeDataType const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
ReduceOpId,
OutputIndex,
64, // BlockSize
64, // ReduceMThreadClusterSize
1, // ReduceKThreadClusterSize
4, // ReduceMThreadSliceSize
1, // ReduceKThreadSliceSize
4>; // InSrcOutDstVectorSize
const ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Z) / window_stride_d + 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 std::vector<ck::index_t> window_spatial_lengths{Z, Y, X}; const std::vector<ck::index_t> window_spatial_lengths{Z, Y, X};
const std::vector<ck::index_t> window_strides{ const std::vector<ck::index_t> window_strides{
window_stride_d, window_stride_h, window_stride_w}; window_stride_d, window_stride_h, window_stride_w};
const std::vector<ck::index_t> window_dilations{
window_dilation_d, window_dilation_h, window_dilation_w};
const std::vector<ck::index_t> input_left_pads{in_left_pad_d, in_left_pad_h, in_left_pad_w}; const std::vector<ck::index_t> input_left_pads{in_left_pad_d, in_left_pad_h, in_left_pad_w};
const std::vector<ck::index_t> input_right_pads{in_right_pad_d, in_right_pad_h, in_right_pad_w}; const std::vector<ck::index_t> input_right_pads{in_right_pad_d, in_right_pad_h, in_right_pad_w};
// tensor layout
auto f_host_tensor_descriptor = [](std::size_t N_,
std::size_t C_,
std::size_t D,
std::size_t H,
std::size_t W,
auto layout) {
using namespace ck::literals;
if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
{
return HostTensorDescriptor({N_, C_, D, H, W},
{C_ * D * H * W, D * H * W, H * W, W, 1_uz});
}
else if constexpr(ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NDHWC>::value)
{
return HostTensorDescriptor({N_, C_, D, H, W},
{D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
}
};
Tensor<InDataType> in_n_c_di_hi_wi(f_host_tensor_descriptor(N, C, Di, Hi, Wi, InLayout{})); Tensor<InDataType> in_n_c_di_hi_wi(f_host_tensor_descriptor(N, C, Di, Hi, Wi, InLayout{}));
Tensor<OutDataType> out_n_c_do_ho_wo_host( Tensor<OutDataType> out_n_c_do_ho_wo_host(
f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{})); f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
...@@ -126,10 +135,11 @@ bool pool3d_test(bool do_verification, ...@@ -126,10 +135,11 @@ bool pool3d_test(bool do_verification,
{N, C, Di, Hi, Wi}, {N, C, Di, Hi, Wi},
{Z, Y, X}, {Z, Y, X},
{N, C, Do, Ho, Wo}, {N, C, Do, Ho, Wo},
{Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C}, f_tensor_strides_ncdhw(N, C, Di, Hi, Wi, InLayout{}),
{Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C}, f_tensor_strides_ncdhw(N, C, Do, Ho, Wo, OutLayout{}),
{Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C}, f_tensor_strides_ncdhw(N, C, Do, Ho, Wo, OutLayout{}),
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads, input_right_pads,
{2, 3, 4}); {2, 3, 4});
...@@ -165,6 +175,7 @@ bool pool3d_test(bool do_verification, ...@@ -165,6 +175,7 @@ bool pool3d_test(bool do_verification,
out_indices_n_c_do_ho_wo_host, out_indices_n_c_do_ho_wo_host,
window_spatial_lengths, window_spatial_lengths,
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads); input_right_pads);
......
example/48_pool3d_fwd/pool3d_fwd_fp16.cpp
View file @ 48fe8532
...@@ -27,31 +27,49 @@ static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG; ...@@ -27,31 +27,49 @@ static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
static constexpr bool OutputIndex = false; static constexpr bool OutputIndex = false;
static constexpr bool PropagateNan = false; static constexpr bool PropagateNan = false;
using DevicePoolFwdInstance =
ck::tensor_operation::device::DevicePool3dFwd_NDHWC_NDHWC<InDataType,
OutDataType,
IndexDataType,
ComputeDataType,
ReduceOpId,
OutputIndex,
64, // BlockSize
64, // ReduceMThreadClusterSize
1, // ReduceKThreadClusterSize
1, // ReduceMThreadSliceSize
1, // ReduceKThreadSliceSize
1>; // InSrcOutDstVectorSize
int main() int main()
{ {
bool do_verification = true; bool do_verification = true;
bool time_kernel = false; bool time_kernel = false;
// Pool shape // Pool shape
ck::index_t N = 2; ck::index_t N = 2;
ck::index_t C = 32; ck::index_t C = 32;
ck::index_t Z = 2; ck::index_t Z = 2;
ck::index_t Y = 2; ck::index_t Y = 2;
ck::index_t X = 2; ck::index_t X = 2;
ck::index_t Di = 30; ck::index_t Di = 30;
ck::index_t Hi = 30; ck::index_t Hi = 30;
ck::index_t Wi = 30; ck::index_t Wi = 30;
ck::index_t window_stride_d = 2; ck::index_t window_stride_d = 2;
ck::index_t window_stride_h = 2; ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2; ck::index_t window_stride_w = 2;
ck::index_t in_left_pad_d = 1; ck::index_t window_dilation_d = 1;
ck::index_t in_left_pad_h = 1; ck::index_t window_dilation_h = 1;
ck::index_t in_left_pad_w = 1; ck::index_t window_dilation_w = 1;
ck::index_t in_right_pad_d = 1; ck::index_t in_left_pad_d = 1;
ck::index_t in_right_pad_h = 1; ck::index_t in_left_pad_h = 1;
ck::index_t in_right_pad_w = 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;
bool pass = pool3d_test<InDataType, bool pass = pool3d_test<DevicePoolFwdInstance,
InDataType,
OutDataType, OutDataType,
ComputeDataType, ComputeDataType,
IndexDataType, IndexDataType,
...@@ -72,6 +90,9 @@ int main() ...@@ -72,6 +90,9 @@ int main()
window_stride_d, window_stride_d,
window_stride_h, window_stride_h,
window_stride_w, window_stride_w,
window_dilation_d,
window_dilation_h,
window_dilation_w,
in_left_pad_d, in_left_pad_d,
in_left_pad_h, in_left_pad_h,
in_left_pad_w, in_left_pad_w,
......
example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp
View file @ 48fe8532
...@@ -24,18 +24,20 @@ int main() ...@@ -24,18 +24,20 @@ int main()
bool time_kernel = false; bool time_kernel = false;
// Pool shape // Pool shape
ck::index_t N = 1; ck::index_t N = 1;
ck::index_t C = 1; ck::index_t C = 1;
ck::index_t Y = 3; ck::index_t Y = 3;
ck::index_t X = 3; ck::index_t X = 3;
ck::index_t Hi = 32; ck::index_t Hi = 32;
ck::index_t Wi = 32; ck::index_t Wi = 32;
ck::index_t window_stride_h = 1; ck::index_t window_stride_h = 1;
ck::index_t window_stride_w = 1; ck::index_t window_stride_w = 1;
ck::index_t in_left_pad_h = 0; ck::index_t window_dilation_h = 1;
ck::index_t in_left_pad_w = 0; ck::index_t window_dilation_w = 1;
ck::index_t in_right_pad_h = 0; ck::index_t in_left_pad_h = 0;
ck::index_t in_right_pad_w = 0; ck::index_t in_left_pad_w = 0;
ck::index_t in_right_pad_h = 0;
ck::index_t in_right_pad_w = 0;
bool pass = maxpool_bwd_test<InDataType, bool pass = maxpool_bwd_test<InDataType,
OutDataType, OutDataType,
...@@ -53,6 +55,8 @@ int main() ...@@ -53,6 +55,8 @@ int main()
Wi, Wi,
window_stride_h, window_stride_h,
window_stride_w, window_stride_w,
window_dilation_h,
window_dilation_w,
in_left_pad_h, in_left_pad_h,
in_left_pad_w, in_left_pad_w,
in_right_pad_h, in_right_pad_h,
......
example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp
View file @ 48fe8532
...@@ -36,6 +36,8 @@ bool maxpool_bwd_test(bool do_verification, ...@@ -36,6 +36,8 @@ bool maxpool_bwd_test(bool do_verification,
ck::index_t Wi, ck::index_t Wi,
ck::index_t window_stride_h, ck::index_t window_stride_h,
ck::index_t window_stride_w, ck::index_t window_stride_w,
ck::index_t window_dilation_h,
ck::index_t window_dilation_w,
ck::index_t in_left_pad_h, ck::index_t in_left_pad_h,
ck::index_t in_left_pad_w, ck::index_t in_left_pad_w,
ck::index_t in_right_pad_h, ck::index_t in_right_pad_h,
...@@ -44,28 +46,30 @@ bool maxpool_bwd_test(bool do_verification, ...@@ -44,28 +46,30 @@ bool maxpool_bwd_test(bool do_verification,
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DevicePoolFwdInstance = using DevicePoolFwdInstance =
ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C< ck::tensor_operation::device::DevicePool2dFwd_NHWC_NHWC<InDataType, // InDataType
InDataType, // InDataType OutDataType, // OutDataType
OutDataType, // OutDataType IndexDataType, // IndexDataType
IndexDataType, // IndexDataType ComputeDataType, // ComputeDataType
ComputeDataType, // ComputeDataType ck::ReduceTensorOp::MAX,
ck::ReduceTensorOp::MAX, true,
true, // OutputIndex 64, // BlockSize
64, // BlockSize 64, // ReduceMThreadClusterSize
64, // ReduceMThreadClusterSize 1, // ReduceKThreadClusterSize
1, // ReduceKThreadClusterSize 4, // ReduceMThreadSliceSize
4, // ReduceMThreadSliceSize 1, // ReduceKThreadSliceSize
1, // ReduceKThreadSliceSize 1>; // InSrcOutDstVectorSize
1>; // InSrcOutDstVectorSize
using DeviceMaxPoolBwdInstance = ck::tensor_operation::device:: using DeviceMaxPoolBwdInstance = ck::tensor_operation::device::
DeviceIndexPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 4>; DeviceIndexPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 4>;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1; const ck::index_t Ys = (Y - 1) * window_dilation_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 Xs = (X - 1) * window_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
const std::vector<ck::index_t> window_spatial_lengths{Y, X}; const std::vector<ck::index_t> window_spatial_lengths{Y, X};
const std::vector<ck::index_t> window_strides{window_stride_h, window_stride_w}; const std::vector<ck::index_t> window_strides{window_stride_h, window_stride_w};
const std::vector<ck::index_t> window_dilations{window_dilation_h, window_dilation_w};
const std::vector<ck::index_t> 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::vector<ck::index_t> 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};
...@@ -128,6 +132,7 @@ bool maxpool_bwd_test(bool do_verification, ...@@ -128,6 +132,7 @@ bool maxpool_bwd_test(bool do_verification,
{C * Ho * Wo, 1, Wo * C, C}, {C * Ho * Wo, 1, Wo * C, C},
{C * Ho * Wo, 1, Wo * C, C}, {C * Ho * Wo, 1, Wo * C, C},
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads, input_right_pads,
{2, 3}); {2, 3});
...@@ -191,6 +196,7 @@ bool maxpool_bwd_test(bool do_verification, ...@@ -191,6 +196,7 @@ bool maxpool_bwd_test(bool do_verification,
indices_n_c_ho_wo_host, indices_n_c_ho_wo_host,
window_spatial_lengths, window_spatial_lengths,
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads); input_right_pads);
ref_pooling_fwd_invoker.Run(ref_pooling_fwd_argument); ref_pooling_fwd_invoker.Run(ref_pooling_fwd_argument);
......
example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp
View file @ 48fe8532
...@@ -24,18 +24,20 @@ int main() ...@@ -24,18 +24,20 @@ int main()
bool time_kernel = false; bool time_kernel = false;
// Pool shape // Pool shape
ck::index_t N = 1; ck::index_t N = 1;
ck::index_t C = 1; ck::index_t C = 1;
ck::index_t Y = 3; ck::index_t Y = 3;
ck::index_t X = 3; ck::index_t X = 3;
ck::index_t Hi = 32; ck::index_t Hi = 32;
ck::index_t Wi = 32; ck::index_t Wi = 32;
ck::index_t window_stride_h = 1; ck::index_t window_stride_h = 1;
ck::index_t window_stride_w = 1; ck::index_t window_stride_w = 1;
ck::index_t in_left_pad_h = 0; ck::index_t window_dilation_h = 1;
ck::index_t in_left_pad_w = 0; ck::index_t window_dilation_w = 1;
ck::index_t in_right_pad_h = 0; ck::index_t in_left_pad_h = 0;
ck::index_t in_right_pad_w = 0; ck::index_t in_left_pad_w = 0;
ck::index_t in_right_pad_h = 0;
ck::index_t in_right_pad_w = 0;
bool pass = maxpool_bwd_test<InDataType, bool pass = maxpool_bwd_test<InDataType,
OutDataType, OutDataType,
...@@ -53,6 +55,8 @@ int main() ...@@ -53,6 +55,8 @@ int main()
Wi, Wi,
window_stride_h, window_stride_h,
window_stride_w, window_stride_w,
window_dilation_h,
window_dilation_w,
in_left_pad_h, in_left_pad_h,
in_left_pad_w, in_left_pad_w,
in_right_pad_h, in_right_pad_h,
......
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