Skip to content

GitLab

Unverified Commit 4b70d68e authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Merge branch 'develop' into add_fp16_wmma_conv_instance

parents 212b9299 f82bd593
Hide whitespace changes
Inline Side-by-side
Showing with 693 additions and 34 deletions
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
View file @ 4b70d68e
......@@ -3,7 +3,7 @@
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
using InDataType = BF16;
// bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
......@@ -17,8 +17,20 @@ using OutElementOp = PassThrough;
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffle<
NDimSpatial,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GNDHWC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GKZYXC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWK,
ck::tensor_layout::convolution::GNHWK,
ck::tensor_layout::convolution::GNDHWK>>,
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
......
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
View file @ 4b70d68e
......@@ -3,7 +3,7 @@
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
using InDataType = F16;
using WeiDataType = F16;
......@@ -16,8 +16,20 @@ using OutElementOp = PassThrough;
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffle<
NDimSpatial,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GNDHWC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GKZYXC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWK,
ck::tensor_layout::convolution::GNHWK,
ck::tensor_layout::convolution::GNDHWK>>,
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
......
example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
View file @ 4b70d68e
......@@ -75,6 +75,8 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
std::array<ck::index_t, NDimSpatial + 3> input_strides{};
std::array<ck::index_t, NDimSpatial + 3> output_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
......@@ -85,6 +87,8 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
range_copy(conv_param.input_spatial_lengths_, begin(input_spatial_lengths));
range_copy(conv_param.filter_spatial_lengths_, begin(filter_spatial_lengths));
range_copy(conv_param.output_spatial_lengths_, begin(output_spatial_lengths));
range_copy(in_g_n_c_wis_desc.GetStrides(), begin(input_strides));
range_copy(out_g_n_k_wos_desc.GetStrides(), begin(output_strides));
range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
range_copy(conv_param.input_left_pads_, begin(input_left_pads));
......@@ -103,6 +107,8 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
input_strides,
output_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......
example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
View file @ 4b70d68e
......@@ -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 @ 4b70d68e
......@@ -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 @ 4b70d68e
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 @ 4b70d68e
......@@ -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 @ 4b70d68e
// 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/CMakeLists.txt
View file @ 4b70d68e
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
......@@ -25,4 +26,5 @@ add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_i
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
# Conv + bias + tanh perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
\ No newline at end of file
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
endif()
\ No newline at end of file
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp
View file @ 4b70d68e
......@@ -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 @ 4b70d68e
......@@ -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 @ 4b70d68e
......@@ -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 @ 4b70d68e
......@@ -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 @ 4b70d68e
......@@ -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_perlayer_quantization_int8.cpp
View file @ 4b70d68e
......@@ -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/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
View file @ 4b70d68e
......@@ -8,7 +8,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......
example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
View file @ 4b70d68e
......@@ -8,7 +8,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......
include/ck/ck.hpp
View file @ 4b70d68e
......@@ -27,6 +27,21 @@
#define CK_WAVELET_MIN_BLOCK_PER_CU 2
#endif
// kernel attribute: amdgpu_waves_per_eu()
#ifdef CK_USE_WAVES_PER_EU
// for 1-wave kernels, control arguments of amdgpu_waves_per_eu() attribute
#ifndef CK_MIN_WAVES_PER_EU
#define CK_MIN_WAVES_PER_EU 0
#endif
#ifndef CK_MAX_WAVES_PER_EU
#define CK_MAX_WAVES_PER_EU 0
#endif
#else
#define CK_USE_WAVES_PER_EU 0
#endif
// buffer resource
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_BUFFER_RESOURCE_3RD_DWORD -1
......@@ -148,6 +163,10 @@
#define CK_EXPERIMENTAL_INTER_WAVE_INSTANCES 1
// experimental feature: add instances using pipeline v2
#define CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES 1
// experimental feature: optimize pipeline v2 by IGLP strategy (value=ID of strategy)
#ifndef CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT
#define CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT 0
#endif
// hack: have underlying assumption that need to be satsified, otherwise it's a bug
// hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be
......
include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
View file @ 4b70d68e
......@@ -4,7 +4,7 @@
#pragma once
#include "ck/tensor_description/cluster_descriptor.hpp"
#include "ck/utility/reduction_common.hpp"
#include "ck/utility/get_shift.hpp"
namespace ck {
......@@ -35,10 +35,11 @@ struct BlockwiseWelford
static constexpr auto thread_cluster_desc =
make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
template <typename CountDataType>
__device__ static inline void
Merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b)
Merge(T& mean_a, T& var_a, CountDataType& count_a, T mean_b, T var_b, CountDataType count_b)
{
int count = count_a + count_b;
CountDataType count = count_a + count_b;
T count_b_over_count = count == 0 ? type_convert<T>(0) : type_convert<T>(count_b) / count;
T delta = mean_b - mean_a;
mean_a += delta * count_b_over_count;
......@@ -46,11 +47,12 @@ struct BlockwiseWelford
count_a = count;
}
__device__ static void Run(T& mean_value, T& var_value, int& count)
template <typename CountDataType>
__device__ static void Run(T& mean_value, T& var_value, CountDataType& count)
{
__shared__ T mean_block_buf[BlockSize];
__shared__ T var_block_buf[BlockSize];
__shared__ int count_block_buf[BlockSize];
__shared__ CountDataType count_block_buf[BlockSize];
constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
......@@ -76,13 +78,13 @@ struct BlockwiseWelford
index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
make_tuple(0, indOffset));
T mean1 = mean_block_buf[offset1];
T var1 = var_block_buf[offset1];
int count1 = count_block_buf[offset1];
T mean1 = mean_block_buf[offset1];
T var1 = var_block_buf[offset1];
CountDataType count1 = count_block_buf[offset1];
T mean2 = mean_block_buf[offset2];
T var2 = var_block_buf[offset2];
int count2 = count_block_buf[offset2];
T mean2 = mean_block_buf[offset2];
T var2 = var_block_buf[offset2];
CountDataType count2 = count_block_buf[offset2];
Merge(mean1, var1, count1, mean2, var2, count2);
......
include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
View file @ 4b70d68e
......@@ -4,7 +4,7 @@
#pragma once
#include "ck/tensor_description/cluster_descriptor.hpp"
#include "ck/utility/reduction_common.hpp"
#include "ck/utility/get_shift.hpp"
#include "ck/utility/reduction_functions_accumulate.hpp"
namespace ck {
......
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