Skip to content

GitLab

"docs/vscode:/vscode.git/clone" did not exist on "9e8d2c7f7490d553c98abbfc84ecb60a8789f89c"
Commit 349552e2 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into stream-k-initial-impl

parents daa98dae 87f2bbcf
Hide whitespace changes
Inline Side-by-side
Showing with 738 additions and 38 deletions
.pre-commit-config.yaml 0 → 100644
View file @ 349552e2
repos:
- repo: local
hooks:
- id: clang-format
name: clang-format
entry: clang-format-10 -i --style=file
language: system
types_or: [c++, inc]
- id: copyright-year-checker
name: copyright-year-checker
entry: script/check_copyright_year.sh
verbose: false
language: script
types: [c++]
Dockerfile
View file @ 349552e2
......@@ -12,24 +12,27 @@ RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins
RUN chmod 1777 /tmp
RUN apt-get update
RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl
RUN if [ "$ROCMVERSION" != "5.6" ]; then \
ENV APT_KEY_DONT_WARN_ON_DANGEROUS_USAGE=DontWarn
RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
RUN wget https://repo.radeon.com/amdgpu-install/5.6/ubuntu/focal/amdgpu-install_5.6.50600-1_all.deb --no-check-certificate
RUN apt-get update && \
DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
./amdgpu-install_5.6.50600-1_all.deb
RUN if [ "$ROCMVERSION" != "5.7" ]; then \
wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"; \
elif [ "$ROCMVERSION" = "5.6" ] && [ "$compiler_version" = "" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amd-nonfree-radeon_20.04-1_all.deb" && \
apt update && apt-get install -y ./amd-nonfree-radeon_20.04-1_all.deb && \
amdgpu-repo --amdgpu-build=1567752 --rocm-build=compute-rocm-dkms-no-npi-hipclang/11914 && \
amdgpu-install -y --usecase=rocm --no-dkms; \
elif [ "$ROCMVERSION" = "5.6" ] && [ "$compiler_version" = "rc3" ] || [ "$compiler_version" = "amd-stg-open" ]; then \
sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.6-20.04-1_all.deb" && \
apt update && apt-get install -y ./amdgpu-install-internal_5.6-20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 5.6 rel-45 > /etc/apt/sources.list.d/rocm-build.list' && \
amdgpu-repo --amdgpu-build=1602498 && amdgpu-install -y --usecase=rocm --no-dkms; \
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'; \
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" && \
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; \
fi
RUN wget --no-check-certificate -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | apt-key add -
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
RUN amdgpu-install -y --usecase=rocm --no-dkms
# Install dependencies
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
......
Jenkinsfile
View file @ 349552e2
......@@ -11,6 +11,20 @@ def show_node_info() {
"""
}
def nthreads() {
def nproc = sh(returnStdout: true, script: 'nproc')
echo "Number of cores: ${nproc}"
def n = nproc.toInteger()
if (n > 32){
n /= 2
}
if (n > 64){
n = 64
}
echo "Number of threads used for building: ${n}"
return n
}
def runShell(String command){
def responseCode = sh returnStatus: true, script: "${command} > tmp.txt"
def output = readFile(file: "tmp.txt")
......@@ -19,7 +33,7 @@ def runShell(String command){
def getDockerImageName(){
def img
if (params.ROCMVERSION != "5.6"){
if (params.ROCMVERSION != "5.7"){
if (params.COMPILER_VERSION == "") {
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}"
}
......@@ -219,7 +233,8 @@ def cmake_build(Map conf=[:]){
"""
def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ")
// reduce parallelism when compiling, clang uses too much memory
def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j\$(( \$(nproc) / 2 )) ${config_targets}")
def nt = nthreads()
def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j${nt} ${config_targets}")
def execute_cmd = conf.get("execute_cmd", "")
def cmd = conf.get("cmd", """
......@@ -461,7 +476,7 @@ def Build_CK(Map conf=[:]){
else{
echo "GPU is OK"
}
if ( runShell('grep -n "gfx1030" clinfo.log') ){
if ( runShell('grep -n "gfx1030" clinfo.log') || runShell('grep -n "gfx1101" clinfo.log') ){
navi_node = 1
}
}
......@@ -482,7 +497,7 @@ def Build_CK(Map conf=[:]){
else{
echo "GPU is OK"
}
if ( runShell('grep -n "gfx1030" clinfo.log') ){
if ( runShell('grep -n "gfx1030" clinfo.log') || runShell('grep -n "gfx1101" clinfo.log') ){
navi_node = 1
}
}
......@@ -493,8 +508,9 @@ def Build_CK(Map conf=[:]){
{
cmake_build(conf)
dir("build"){
//run tests and examples
sh 'make -j\$(( \$(nproc) / 2 )) check'
//run tests and examples
def nt = nthreads()
sh 'make -j${nt} check'
if (navi_node == 0 ){
//we only need the ckProfiler to run the performance tests, so we pack and stash it
//do not stash profiler on Navi nodes
......@@ -598,7 +614,7 @@ 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
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true
0 21 * * * % ROCMVERSION=5.5;COMPILER_VERSION=release;COMPILER_COMMIT=
0 21 * * * % ROCMVERSION=5.6;COMPILER_VERSION=;COMPILER_COMMIT=
0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
pipeline {
......@@ -717,7 +733,7 @@ pipeline {
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 Navi")
stage("Build CK and run Tests on Navi21")
{
when {
beforeAgent true
......
README.md
View file @ 349552e2
......@@ -109,6 +109,24 @@ make install
Instructions for using CK as a pre-built kernel library are under [client_example](/client_example)
## Contributing
When you contribute to Composable Kernel, make sure to run `clang-format` on all the changed files. We highly recommend using git hooks that are managed by the `pre-commit` framework. To install hooks, run:
```bash
sudo script/install_precommit.sh
```
This way, `pre-commit` will add the appropriate hooks to your local repository and automatically run `clang-format` (and possibly additional checks) before any commit is created.
If you need to uninstall hooks from the repository, you can do so by running the following command:
```bash
script/uninstall_precommit.sh
```
If for any reason, you need to temporarily disable precommit hooks, you can add the `--no-verify` option to the `git commit` command.
## Caveat
### Kernel Timing and Verification
......
example/01_gemm/CMakeLists.txt
View file @ 349552e2
......@@ -45,3 +45,8 @@ if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS
endif()
add_example_executable(example_gemm_xdl_streamk gemm_xdl_streamk.cpp)
if(GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR GPU_TARGETS MATCHES "gfx942")
add_example_executable(example_gemm_xdl_f8 gemm_xdl_f8.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_f8)
endif()
example/01_gemm/gemm_xdl_f8.cpp 0 → 100644
View file @ 349552e2
// 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::f8_t;
using CDataType = ck::f8_t;
using AccDataType = float;
using CShuffleDataType = ck::f8_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;
// 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|
// ######| | | | 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|
// ######| | | | | | | | | 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, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 16>;
// 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_wavelet_fp16.cpp
View file @ 349552e2
......@@ -3,7 +3,7 @@
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_xdl_waveletmodel_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp"
using ADataType = ck::half_t;
using BDataType = ck::half_t;
......
example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
View file @ 349552e2
......@@ -3,7 +3,7 @@
#include "convnd_fwd_dl_common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
......
example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
View file @ 349552e2
......@@ -3,7 +3,7 @@
#include "convnd_fwd_dl_common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
......
example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
View file @ 349552e2
......@@ -3,7 +3,7 @@
#include "convnd_fwd_dl_common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
......
example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
View file @ 349552e2
......@@ -17,7 +17,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......
example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
View file @ 349552e2
......@@ -17,7 +17,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......
example/34_batchnorm/CMakeLists.txt
View file @ 349552e2
add_example_executable(example_batchnorm_forward_training batchnorm_forward_training_nhwc.cpp)
add_example_executable(example_batchnorm_forward_training_obsolete batchnorm_forward_training_nhwc_obsolete.cpp)
add_example_executable(example_batchnorm_forward_inferring batchnorm_forward_inferring_nhwc.cpp)
add_example_executable(example_batchnorm_backward batchnorm_backward_nhwc.cpp)
example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
View file @ 349552e2
......@@ -414,7 +414,7 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
y_dev.FromDevice(y.mData.data());
pass = pass && ck::utils::check_err(y, y_ref);
pass = pass && ck::utils::check_err(y, y_ref, "Incorrect normalized output values");
if(updateMovingAverage)
{
......@@ -424,8 +424,12 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
pass = pass && ck::utils::check_err(resultRunningMean, resultRunningMean_ref);
pass = pass && ck::utils::check_err(resultRunningVariance, resultRunningVariance_ref);
pass = pass && ck::utils::check_err(resultRunningMean,
resultRunningMean_ref,
"Incorrect running mean values");
pass = pass && ck::utils::check_err(resultRunningVariance,
resultRunningVariance_ref,
"Incorrect running variance values");
};
if(saveMeanAndInvVariance)
......@@ -438,8 +442,11 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
pass = pass && ck::utils::check_err(resultSaveMean, resultSaveMean_ref);
pass = pass && ck::utils::check_err(resultSaveInvVariance, resultSaveInvVariance_ref);
pass = pass && ck::utils::check_err(
resultSaveMean, resultSaveMean_ref, "Incorrect saved mean values");
pass = pass && ck::utils::check_err(resultSaveInvVariance,
resultSaveInvVariance_ref,
"Incorrect saved invvariance values");
};
};
......
example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp 0 → 100644
View file @ 349552e2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <limits>
#include <iostream>
#include <vector>
#include <array>
#include <algorithm>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
{"verify", required_argument, nullptr, 'v'},
{"help", no_argument, nullptr, '?'},
{nullptr, 0, nullptr, 0}};
class BatchNormFwdArg
{
private:
int option_index = 0;
public:
std::vector<size_t> inOutLengths;
bool do_verification = false;
bool updateMovingAverage;
bool saveMeanAndInvVariance;
int data_type = 0;
int init_method = 2;
bool time_kernel = false;
bool use_multiblock_welford = false;
public:
void show_usage(const char* cmd)
{
std::cout << "Usage of " << cmd << std::endl;
std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
"lengths, must have 4 integers for nhwc"
<< std::endl;
std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
"result by "
"comparing with the host-based batch-normalization"
<< std::endl;
std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
std::cout << "Arg2: 1/0 to indicate whether to update the moving average and variance "
"(0=no, 1=yes)"
<< std::endl;
std::cout << "Arg3: 1/0 to indicate whether to save the calculated mean and invVariance "
"(0=no, 1=yes)"
<< std::endl;
std::cout << "Arg4: init method used for bnScale and bnBias (0=no init, 1=single integer "
"value, 2=scope integer "
"value, 3=decimal value)"
<< std::endl;
std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
std::cout << "Arg6: use multi-block welford (0=n0, 1=yes)" << std::endl;
};
int processArgs(int argc, char* argv[])
{
using ck::host_common::getTypeValuesFromString;
int ch;
while(1)
{
ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
{
case 'D':
if(!optarg)
throw std::runtime_error("Invalid option format!");
inOutLengths = getTypeValuesFromString<size_t>(optarg);
if(inOutLengths.size() != 4)
throw std::runtime_error(
"NHWC tensor layout should have 4 length values specified!");
break;
case 'v':
if(!optarg)
throw std::runtime_error("Invalid option format!");
do_verification = static_cast<bool>(std::atoi(optarg));
break;
case '?':
if(std::string(long_options[option_index].name) == "help")
{
show_usage(argv[0]);
return (-1);
};
break;
default: show_usage(argv[0]); return (-1);
};
};
if(optind + 6 > argc)
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
data_type = std::atoi(argv[optind++]);
updateMovingAverage = std::atoi(argv[optind++]);
saveMeanAndInvVariance = std::atoi(argv[optind++]);
init_method = std::atoi(argv[optind++]);
time_kernel = static_cast<bool>(std::atoi(argv[optind++]));
use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
return (-1);
return (0);
};
};
using namespace ck;
template <typename InOutDataType, typename AccDataType, bool UseMultiblockInK>
bool bnorm_fwd_nhwc_test(bool do_verification,
int init_method,
bool time_kernel,
const std::vector<size_t> inOutLengths,
bool updateMovingAverage,
bool saveMeanAndInvVariance,
double averageFactor,
double epsilon)
{
// for NHWC BatchNorm calculation of mean and meansquare
constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
// when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
// eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
// input data of the batchnorm forward algorithm
Tensor<InOutDataType> x(inOutLengths);
Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
// output data of the batchnorm forward algorithm
Tensor<InOutDataType> y_ref(inOutLengths);
Tensor<InOutDataType> y(inOutLengths);
Tensor<AccDataType> resultSaveMean_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultSaveInvVariance_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultRunningMean_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultRunningVariance_ref(scaleBiasMeanVarLengths);
auto inOutStrides = x.mDesc.GetStrides();
auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
std::size_t num_thread = std::thread::hardware_concurrency();
if(updateMovingAverage)
{
if constexpr(std::is_same<InOutDataType, int8_t>::value)
{
x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
const float x_mean = 0.0f;
const float x_stddev = 2.5f;
const float noise_stddev = 0.04f;
resultRunningMean_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
resultRunningVariance_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
}
else
{
const float x_mean = 0.0f;
const float x_stddev = 1.0f;
const float noise_stddev = 0.04f;
// input data in normal distribution
x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
// initialize the runningMean to be values with tiny variation to the mean of the x
// values
resultRunningMean_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
// initialize the runningVariance to be values with tiny variation to the variance of
// the x values
resultRunningVariance_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
};
}
else
{
if constexpr(std::is_same<InOutDataType, int8_t>::value)
x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
else
x.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0f, 5.0f}, num_thread);
};
if(do_verification)
{
switch(init_method)
{
case 0:
bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
break;
case 1:
bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
break;
case 2:
bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
break;
default:
bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
}
};
// these buffers are usually provided by the user application
DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
// mean_dev or resultSaveMean_dev
DeviceMem resultSaveMean_dev(sizeof(AccDataType) *
resultSaveMean_ref.mDesc.GetElementSpaceSize());
// meansquare_dev or resultSaveInvVariance_dev
DeviceMem resultSaveInvVariance_dev(sizeof(AccDataType) *
resultSaveInvVariance_ref.mDesc.GetElementSpaceSize());
// resultRunningMean_dev
DeviceMem resultRunningMean_dev(sizeof(AccDataType) *
resultRunningMean_ref.mDesc.GetElementSpaceSize());
// resultRunningVariance_dev
DeviceMem resultRunningVariance_dev(sizeof(AccDataType) *
resultRunningVariance_ref.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data());
bnScale_dev.ToDevice(bnScale.mData.data());
bnBias_dev.ToDevice(bnBias.mData.data());
if(updateMovingAverage)
{
resultRunningMean_dev.ToDevice(resultRunningMean_ref.mData.data());
resultRunningVariance_dev.ToDevice(resultRunningVariance_ref.mData.data());
};
std::array<index_t, Rank> i_inOutLengths;
std::array<index_t, Rank> i_inOutStrides;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceBatchNormFwdInstance =
ck::tensor_operation::device::DeviceBatchNormFwdImpl<InOutDataType,
InOutDataType,
AccDataType,
AccDataType, // ScaleDataType
AccDataType, // BiasDataType
AccDataType, // MeanVarDataType
PassThroughOp, // YElementwiseOp
Rank,
NumReduceDim,
UseMultiblockInK,
256,
16,
16,
1,
2,
0,
1,
1,
1,
1,
1>;
auto batchnorm_fwd = DeviceBatchNormFwdInstance{};
auto argument_ptr = batchnorm_fwd.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x_dev.GetDeviceBuffer(),
bnScale_dev.GetDeviceBuffer(),
bnBias_dev.GetDeviceBuffer(),
epsilon,
PassThroughOp{},
y_dev.GetDeviceBuffer(),
saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
averageFactor,
updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
if(!batchnorm_fwd.IsSupportedArgument(argument_ptr.get()))
{
std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
"exiting!"
<< std::endl;
return (false);
};
size_t workspace_sz = batchnorm_fwd.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
batchnorm_fwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = batchnorm_fwd.MakeInvokerPointer();
if(time_kernel)
{
float avg_time = 0.0f;
size_t num_bytes = 0;
size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
size_t invariant_length = inOutLengths[3];
avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
// inputing of x, scale, bias, outputing of y
num_bytes +=
total_length * sizeof(InOutDataType) * 2 + invariant_length * sizeof(AccDataType) * 2;
// outputing of mean, inv-variance
num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(AccDataType) * 2 : 0;
// updating of moving mean, variance
num_bytes += updateMovingAverage ? invariant_length * sizeof(AccDataType) * 4 : 0;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
}
else
(void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
bool pass = true;
if(do_verification)
{
using ReferenceBatchNormFwdInstance =
ck::tensor_operation::host::ReferenceBatchNormFwd<InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType,
PassThroughOp,
Rank,
NumReduceDim>;
auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x.mData.data(),
bnScale.mData.data(),
bnBias.mData.data(),
epsilon,
PassThroughOp{},
y_ref.mData.data(),
saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
averageFactor,
updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
{
std::cout << "The runtime parameters seems not supported by the BatchNorm reference "
"instance, exiting!"
<< std::endl;
return (false);
};
auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
y_dev.FromDevice(y.mData.data());
pass = pass && ck::utils::check_err(y, y_ref, "Incorrect normalized output values");
if(updateMovingAverage)
{
Tensor<AccDataType> resultRunningMean(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultRunningVariance(scaleBiasMeanVarLengths);
resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
pass = pass && ck::utils::check_err(resultRunningMean,
resultRunningMean_ref,
"Incorrect running mean values");
pass = pass && ck::utils::check_err(resultRunningVariance,
resultRunningVariance_ref,
"Incorrect running variance values");
};
if(saveMeanAndInvVariance)
{
using ck::host_common::dumpBufferToFile;
Tensor<AccDataType> resultSaveMean(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
pass = pass && ck::utils::check_err(
resultSaveMean, resultSaveMean_ref, "Incorrect saved mean values");
pass = pass && ck::utils::check_err(resultSaveInvVariance,
resultSaveInvVariance_ref,
"Incorrect saved invvariance values");
};
};
return (pass);
};
const double epsilon = std::numeric_limits<float>::epsilon();
static const double averageFactor = 0.1;
int main(int argc, char* argv[])
{
bool pass = true;
if(argc > 1)
{
BatchNormFwdArg arg;
if(arg.processArgs(argc, argv) < 0)
return (-1);
if(arg.data_type == 0)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 1)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<float, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<float, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 3)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<int8_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<int8_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 5)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 6)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<double, double, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<double, double, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
}
else
{
pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(true,
2,
false, // don't time kernel
{128, 16, 6, 512},
true,
true,
averageFactor,
epsilon);
pass = pass && bnorm_fwd_nhwc_test<ck::half_t, float, false>(true,
2,
false, // don't time kernel
{128, 16, 3, 1024},
true,
true,
averageFactor,
epsilon);
};
return (pass ? 0 : 1);
}
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp
View file @ 349552e2
......@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
......
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp
View file @ 349552e2
......@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
......
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp
View file @ 349552e2
......@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
......
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp
View file @ 349552e2
......@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
......
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp
View file @ 349552e2
......@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
......
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