Skip to content

GitLab

Commit 95a83c6e authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge remote-tracking branch 'origin/develop' into wavelet_model

parents 5b7c2432 892a8d76
Hide whitespace changes
Inline Side-by-side
Showing with 551 additions and 325 deletions
example/34_batchnorm/batchnorm_infer_impl.hpp
View file @ 95a83c6e
...@@ -14,8 +14,12 @@ ...@@ -14,8 +14,12 @@
#include "batchnorm_common.hpp" #include "batchnorm_common.hpp"
template <typename InOutDataType, template <typename XDataType,
typename YDataType,
typename AccDataType, typename AccDataType,
typename ScaleDataType,
typename BiasDataType,
typename MeanVarDataType,
ck::index_t Rank, ck::index_t Rank,
ck::index_t NumBatchNormReduceDim, ck::index_t NumBatchNormReduceDim,
bool fastest_dim_is_reduced = false> bool fastest_dim_is_reduced = false>
...@@ -26,7 +30,9 @@ int bnorm_infer( ...@@ -26,7 +30,9 @@ int bnorm_infer(
const std::array<ck::index_t, Rank> xStrides, const std::array<ck::index_t, Rank> xStrides,
const std::array<ck::index_t, Rank> yStrides, const std::array<ck::index_t, Rank> yStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths, const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarStrides, const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
const void* p_x, const void* p_x,
const void* p_scale, const void* p_scale,
const void* p_bias, const void* p_bias,
...@@ -41,11 +47,11 @@ int bnorm_infer( ...@@ -41,11 +47,11 @@ int bnorm_infer(
"Invalid number of reduced dimensions for batchnorm!"); "Invalid number of reduced dimensions for batchnorm!");
using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise< using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<InOutDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean, ck::Tuple<XDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
// variance, // variance,
// scale, // scale,
// bias, // bias,
ck::Tuple<InOutDataType>, // y ck::Tuple<YDataType>, // y
NormalizeInInfer, NormalizeInInfer,
Rank, Rank,
2, // MPerthread 2, // MPerthread
...@@ -53,14 +59,18 @@ int bnorm_infer( ...@@ -53,14 +59,18 @@ int bnorm_infer(
ck::Sequence<1>>; // scalarPerVector: y ck::Sequence<1>>; // scalarPerVector: y
auto invariantDims = get_invariant_dims<Rank, NumBatchNormReduceDim>(reduceDims); auto invariantDims = get_invariant_dims<Rank, NumBatchNormReduceDim>(reduceDims);
std::array<ck::index_t, Rank> aligned_scaleBiasMeanVarStrides{0}; std::array<ck::index_t, Rank> aligned_bnScaleStrides{0};
std::array<ck::index_t, Rank> aligned_bnBiasStrides{0};
std::array<ck::index_t, Rank> aligned_bnMeanVarStrides{0};
int i = 0; int i = 0;
for(auto dim : invariantDims) for(auto dim : invariantDims)
{ {
assert(xyLengths[dim] == bnScaleBiasMeanVarLengths[i]); assert(xyLengths[dim] == bnScaleBiasMeanVarLengths[i]);
aligned_scaleBiasMeanVarStrides[dim] = bnScaleBiasMeanVarStrides[i]; aligned_bnScaleStrides[dim] = bnScaleStrides[i];
aligned_bnBiasStrides[dim] = bnBiasStrides[i];
aligned_bnMeanVarStrides[dim] = bnMeanVarStrides[i];
i++; i++;
}; };
...@@ -84,10 +94,10 @@ int bnorm_infer( ...@@ -84,10 +94,10 @@ int bnorm_infer(
auto argument_ptr1 = dev_normalize.MakeArgumentPointer( auto argument_ptr1 = dev_normalize.MakeArgumentPointer(
xyLengths, xyLengths,
{xStrides, {xStrides,
aligned_scaleBiasMeanVarStrides, aligned_bnMeanVarStrides,
aligned_scaleBiasMeanVarStrides, aligned_bnMeanVarStrides,
aligned_scaleBiasMeanVarStrides, aligned_bnScaleStrides,
aligned_scaleBiasMeanVarStrides}, aligned_bnBiasStrides},
{yStrides}, {yStrides},
{p_x, p_estimatedMean, p_estimatedVariance, p_scale, p_bias}, {p_x, p_estimatedMean, p_estimatedVariance, p_scale, p_bias},
{p_y}, {p_y},
...@@ -105,8 +115,10 @@ int bnorm_infer( ...@@ -105,8 +115,10 @@ int bnorm_infer(
avg_time += invoker_ptr1->Run(argument_ptr1.get(), StreamConfig{nullptr, time_kernel}); avg_time += invoker_ptr1->Run(argument_ptr1.get(), StreamConfig{nullptr, time_kernel});
num_bytes += (total_length * (1 * sizeof(InOutDataType) + 4 * sizeof(AccDataType)) + num_bytes += total_length * sizeof(XDataType) +
total_length * sizeof(InOutDataType)); invariantLength *
(sizeof(ScaleDataType) + sizeof(BiasDataType) + 2 * sizeof(MeanVarDataType)) +
total_length * sizeof(YDataType);
if(time_kernel) if(time_kernel)
{ {
......
example/34_batchnorm/batchnorm_infer_nhwc.cpp
View file @ 95a83c6e
...@@ -9,6 +9,7 @@ ...@@ -9,6 +9,7 @@
#include <getopt.h> #include <getopt.h>
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
...@@ -18,11 +19,6 @@ ...@@ -18,11 +19,6 @@
#include "batchnorm_infer_impl.hpp" #include "batchnorm_infer_impl.hpp"
template <typename InOutDataType, typename AccDataType>
using ReferenceBatchNormInferInstance =
ck::tensor_operation::host::ReferenceBatchNormInfer_Input_N_H_W_C_Output_C<InOutDataType,
AccDataType>;
static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'}, static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
{"verify", required_argument, nullptr, 'v'}, {"verify", required_argument, nullptr, 'v'},
{"help", no_argument, nullptr, '?'}, {"help", no_argument, nullptr, '?'},
...@@ -225,32 +221,37 @@ bool bnorm_infer_nhwc_test(bool do_verification, ...@@ -225,32 +221,37 @@ bool bnorm_infer_nhwc_test(bool do_verification,
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths; std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides; std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
std::copy(inOutLengths.begin(), inOutLengths.end(), i_inOutLengths.begin()); ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
std::copy(inOutStrides.begin(), inOutStrides.end(), i_inOutStrides.begin()); ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
std::copy(scaleBiasMeanVarLengths.begin(), ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
scaleBiasMeanVarLengths.end(), ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
i_scaleBiasMeanVarLengths.begin());
std::copy(scaleBiasMeanVarStrides.begin(),
scaleBiasMeanVarStrides.end(),
i_scaleBiasMeanVarStrides.begin());
int result = 0; int result = 0;
result = bnorm_infer<InOutDataType, AccDataType, Rank, NumReduceDim, false>( result = bnorm_infer<InOutDataType,
time_kernel, InOutDataType,
{0, 1, 2}, AccDataType,
i_inOutLengths, AccDataType,
i_inOutStrides, AccDataType,
i_inOutStrides, AccDataType,
i_scaleBiasMeanVarLengths, Rank,
i_scaleBiasMeanVarStrides, NumReduceDim,
x_dev.GetDeviceBuffer(), false>(time_kernel,
bnScale_dev.GetDeviceBuffer(), {0, 1, 2},
bnBias_dev.GetDeviceBuffer(), i_inOutLengths,
epsilon, i_inOutStrides,
estimatedMean_dev.GetDeviceBuffer(), i_inOutStrides,
estimatedVariance_dev.GetDeviceBuffer(), i_scaleBiasMeanVarLengths,
y_dev.GetDeviceBuffer()); i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x_dev.GetDeviceBuffer(),
bnScale_dev.GetDeviceBuffer(),
bnBias_dev.GetDeviceBuffer(),
epsilon,
estimatedMean_dev.GetDeviceBuffer(),
estimatedVariance_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer());
if(result < 0) if(result < 0)
return (false); return (false);
...@@ -259,7 +260,15 @@ bool bnorm_infer_nhwc_test(bool do_verification, ...@@ -259,7 +260,15 @@ bool bnorm_infer_nhwc_test(bool do_verification,
if(do_verification) if(do_verification)
{ {
auto batchNormInfer_ref = ReferenceBatchNormInferInstance<InOutDataType, AccDataType>{}; using ReferenceBatchNormInferInstance =
ck::tensor_operation::host::ReferenceBatchNormInfer_Input_N_H_W_C_Output_C<
InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType>;
auto batchNormInfer_ref = ReferenceBatchNormInferInstance{};
auto argument_ptr_ref = auto argument_ptr_ref =
batchNormInfer_ref.MakeArgumentPointer(i_inOutLengths, batchNormInfer_ref.MakeArgumentPointer(i_inOutLengths,
...@@ -267,6 +276,8 @@ bool bnorm_infer_nhwc_test(bool do_verification, ...@@ -267,6 +276,8 @@ bool bnorm_infer_nhwc_test(bool do_verification,
i_inOutStrides, i_inOutStrides,
i_scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x.mData.data(), x.mData.data(),
bnScale.mData.data(), bnScale.mData.data(),
bnBias.mData.data(), bnBias.mData.data(),
...@@ -288,7 +299,7 @@ bool bnorm_infer_nhwc_test(bool do_verification, ...@@ -288,7 +299,7 @@ bool bnorm_infer_nhwc_test(bool do_verification,
(void)invoker_ptr_ref->Run(argument_ptr_ref.get()); (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
y_dev.FromDevice(y.mData.data()); y_dev.FromDevice(y.mData.data());
pass = pass && ck::utils::check_err(y.mData, y_ref.mData); pass = pass && ck::utils::check_err(y, y_ref);
}; };
return (pass); return (pass);
......
example/35_splitK_gemm/run_splitK_gemm_example.inc
View file @ 95a83c6e
...@@ -34,15 +34,15 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con ...@@ -34,15 +34,15 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con
auto f_host_tensor_descriptor = auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) { [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value) if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {stride, 1_uz});
std::vector<std::size_t>({stride, 1}));
} }
else else
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {1_uz, stride});
std::vector<std::size_t>({1, stride}));
} }
}; };
...@@ -146,15 +146,12 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con ...@@ -146,15 +146,12 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con
if(std::is_same<CDataType, ck::half_t>::value) if(std::is_same<CDataType, ck::half_t>::value)
{ {
pass &= ck::utils::check_err(c_m_n_device_result.mData, pass &= ck::utils::check_err(
c_m_n_host_result.mData, c_m_n_device_result, c_m_n_host_result, "fp16 incorrect result", 3e-3, 1e-3);
"fp16 incorrect result",
3e-3,
1e-3);
} }
else else
{ {
pass &= ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData); pass &= ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
} }
} }
......
example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
View file @ 95a83c6e
...@@ -86,12 +86,10 @@ int main() ...@@ -86,12 +86,10 @@ int main()
constexpr auto index_length = 2048; constexpr auto index_length = 2048;
constexpr AccDataType epsilon = 1e-4; constexpr AccDataType epsilon = 1e-4;
auto f_host_tensor_desc_1d = [](std::size_t len_) { auto f_host_tensor_desc_1d = [](std::size_t len_) { return HostTensorDescriptor({len_}); };
return HostTensorDescriptor(std::vector<std::size_t>({len_}));
};
auto f_host_tensor_desc_2d = [](std::size_t rows_, std::size_t cols_) { auto f_host_tensor_desc_2d = [](std::size_t rows_, std::size_t cols_) {
return HostTensorDescriptor(std::vector<std::size_t>({rows_, cols_})); return HostTensorDescriptor({rows_, cols_});
}; };
using ReferenceInstance = using ReferenceInstance =
...@@ -203,8 +201,7 @@ int main() ...@@ -203,8 +201,7 @@ int main()
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
out_dev.FromDevice(out_from_dev.mData.data()); out_dev.FromDevice(out_from_dev.mData.data());
pass &= ck::utils::check_err( pass &= ck::utils::check_err(out_from_dev, out, "Error: Incorrect results", 1e-3, 1e-3);
out_from_dev.mData, out.mData, "Error: Incorrect results", 1e-3, 1e-3);
} }
double total_read = current_dim * index_length * 3 * sizeof(EmbType) + double total_read = current_dim * index_length * 3 * sizeof(EmbType) +
......
example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
View file @ 95a83c6e
...@@ -19,6 +19,7 @@ Computes C_m_o = Relu(A0[m, k] * B0[n, k] + D00[m, n] + D01[mn]) * B1[n, o] + D1 ...@@ -19,6 +19,7 @@ Computes C_m_o = Relu(A0[m, k] * B0[n, k] + D00[m, n] + D01[mn]) * B1[n, o] + D1
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
...@@ -314,15 +315,15 @@ int main(int argc, char* argv[]) ...@@ -314,15 +315,15 @@ int main(int argc, char* argv[])
std::size_t stride, std::size_t stride,
std::size_t batch_stride, std::size_t batch_stride,
auto layout) { auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), Row>::value) if(std::is_same<decltype(layout), Row>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}), return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
std::vector<std::size_t>({batch_stride, stride, 1}));
} }
else else
{ {
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}), return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
std::vector<std::size_t>({batch_stride, 1, stride}));
} }
}; };
...@@ -511,8 +512,7 @@ int main(int argc, char* argv[]) ...@@ -511,8 +512,7 @@ int main(int argc, char* argv[])
cde1_element_op(e1_g_m_o_host_result(idx), c1_g_m_o(idx), d1_g_m_o(idx)); cde1_element_op(e1_g_m_o_host_result(idx), c1_g_m_o(idx), d1_g_m_o(idx));
}); });
return ck::utils::check_err(e1_g_m_o_device_result.mData, e1_g_m_o_host_result.mData) ? 0 return ck::utils::check_err(e1_g_m_o_device_result, e1_g_m_o_host_result) ? 0 : 1;
: 1;
} }
return 0; return 0;
......
example/38_grouped_conv_bwd_data_bias_relu/CMakeLists.txt deleted 100644 → 0
View file @ 5b7c2432
add_example_executable(example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp)
example/38_grouped_conv_bwd_data_bias_relu/grouped_conv_bwd_data_bias_relu_fp16.cpp deleted 100644 → 0
View file @ 5b7c2432
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "grouped_conv_bwd_data_bias_relu_common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using OutDataType = ck::half_t;
using WeiDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using BiasDataType = ck::half_t; // bias
using InDataType = ck::half_t;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using BiasLayout = ck::tensor_layout::convolution::G_C;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using CBiasInElementOp = ck::tensor_operation::element_wise::AddRelu;
static constexpr auto ConvBwdDataDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
template <ck::index_t NDimSpatial>
using DeviceConvNdBwdDataInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<
NDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<BiasLayout>,
InLayout,
OutDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<BiasDataType>,
InDataType,
OutElementOp,
WeiElementOp,
CBiasInElementOp,
ConvBwdDataDefault,
true, // DoPadGemmM
true, // DoPadGemmN
1,
256,
128,
256,
32,
8,
2,
32,
32,
2,
4,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
S<4, 64, 1>,
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
0,
1,
1,
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 2, 128, 256, 256, {3, 3}, {14, 14}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
}
const auto in_element_op = CBiasInElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_param.num_dim_spatial_ == 2)
{
// output image: GNHWK
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
// weight: GKYXC
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
// input image bias: G_C
const auto bias_g_n_c_wis_desc =
HostTensorDescriptor({conv_param.G_,
conv_param.N_,
conv_param.C_,
conv_param.input_spatial_lengths_[0],
conv_param.input_spatial_lengths_[1]},
{
conv_param.C_, // g
0, // n
1, // c
0, // hi
0 // wi
});
// input image: GNHWC
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
using DeviceInstance = DeviceConvNdBwdDataInstance<2>;
run_conv_bwd_data_bias_relu<2,
OutDataType,
WeiDataType,
BiasDataType,
InDataType,
OutElementOp,
WeiElementOp,
CBiasInElementOp,
DeviceInstance>(do_verification,
init_method,
time_kernel,
conv_param,
out_g_n_k_wos_desc,
wei_g_k_c_xs_desc,
bias_g_n_c_wis_desc,
in_g_n_c_wis_desc,
wei_element_op,
out_element_op,
in_element_op);
}
return 0;
}
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt 0 → 100644
View file @ 95a83c6e
add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_fp16 grouped_conv_bwd_data_fp16.cpp)
add_example_executable(example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_fp16)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16)
example/38_grouped_conv_bwd_data_multiple_d/common.hpp 0 → 100644
View file @ 95a83c6e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <initializer_list>
#include <iostream>
#include <numeric>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static inline constexpr ck::index_t NDimSpatial = 2;
static constexpr auto ConvBwdDataDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
using FP16 = ck::half_t;
using FP32 = float;
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
};
#define DefaultConvParams \
ck::utils::conv::ConvParam \
{ \
NDimSpatial, 32, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, { 1, 1 } \
}
inline void print_help_msg()
{
std::cerr << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
}
inline bool parse_cmd_args(int argc,
char* argv[],
ExecutionConfig& config,
ck::utils::conv::ConvParam& conv_params)
{
constexpr int num_execution_config_args =
3; // arguments for do_verification, init_method, time_kernel
constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
constexpr int threshold_to_catch_all_args =
threshold_to_catch_partial_args + num_conv_param_leading_args;
if(argc == 1)
{
// use default
config = ExecutionConfig{};
}
// catch only ExecutionConfig arguments
else if(argc == threshold_to_catch_partial_args)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
// catch both ExecutionConfig & ConvParam arguments
else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_params = ck::utils::conv::parse_conv_param(
num_dim_spatial, threshold_to_catch_partial_args, argv);
}
else
{
print_help_msg();
return false;
}
return true;
}
example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_fp16.cpp 0 → 100644
View file @ 95a83c6e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using OutDataType = FP16;
using WeiDataType = FP16;
using AccDataType = FP32;
using CShuffleDataType = FP16;
using BiasDataType = FP16; // bias
using InDataType = FP16;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using BiasLayout = ck::Tuple<ck::tensor_layout::convolution::G_C>;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using OutElementOp = PassThrough;
using WeiElementOp = PassThrough;
using InElementOp = ck::tensor_operation::element_wise::AddRelu;
// clang-format off
using DeviceConvInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
// ######| NDimSpatial| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer|
// ######| | | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< NDimSpatial, OutLayout, WeiLayout, BiasLayout, InLayout, OutDataType, WeiDataType, AccDataType, CShuffleDataType, ck::Tuple<BiasDataType>, InDataType, OutElementOp, WeiElementOp, InElementOp, ConvBwdDataDefault, true, true, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
#include "run_grouped_conv_bwd_data_bias_relu_example.inc"
int main(int argc, char* argv[]) { return run_grouped_conv_bwd_data_bias_relu_example(argc, argv); }
example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_fp16.cpp 0 → 100644
View file @ 95a83c6e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using OutDataType = FP16;
using WeiDataType = FP16;
using AccDataType = FP32;
using CShuffleDataType = FP16;
using DsDataType = ck::Tuple<>;
using InDataType = FP16;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using DsLayout = ck::Tuple<>;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using OutElementOp = PassThrough;
using WeiElementOp = PassThrough;
using InElementOp = PassThrough;
// clang-format off
using DeviceConvInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
// ######| NDimSpatial| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer|
// ######| | | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< NDimSpatial, OutLayout, WeiLayout, DsLayout, InLayout, OutDataType, WeiDataType, AccDataType, CShuffleDataType, DsDataType, InDataType, OutElementOp, WeiElementOp, InElementOp, ConvBwdDataDefault, true, true, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
#include "run_grouped_conv_bwd_data_example.inc"
int main(int argc, char* argv[]) { return run_grouped_conv_bwd_data_example(argc, argv); }
example/38_grouped_conv_bwd_data_bias_relu/grouped_conv_bwd_data_bias_relu_common.hpp → example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
View file @ 95a83c6e
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> bool run_conv_bwd_data_bias_relu(const ExecutionConfig& config,
#include <numeric> const ck::utils::conv::ConvParam& conv_params,
#include <initializer_list> const HostTensorDescriptor& out_g_n_k_wos_desc,
#include <cstdlib> const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& bias_g_n_c_wis_desc,
#include "ck/ck.hpp" const HostTensorDescriptor& in_g_n_c_wis_desc,
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" const OutElementOp& out_element_op,
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" const WeiElementOp& wei_element_op,
const InElementOp& in_element_op)
#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/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
void print_helper_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
}
template <ck::index_t NDimSpatial,
typename OutDataType,
typename WeiDataType,
typename BiasDataType,
typename InDataType,
typename OutElementOp,
typename WeiElementOp,
typename InElementOp,
typename DeviceInstance>
int run_conv_bwd_data_bias_relu(bool do_verification,
int init_method,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param,
const HostTensorDescriptor& out_g_n_k_wos_desc,
const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& bias_g_n_c_wis_desc,
const HostTensorDescriptor& in_g_n_c_wis_desc,
const OutElementOp& out_element_op,
const WeiElementOp& wei_element_op,
const InElementOp& in_element_op)
{ {
Tensor<OutDataType> out(out_g_n_k_wos_desc); Tensor<OutDataType> out(out_g_n_k_wos_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc); Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
...@@ -58,7 +22,7 @@ int run_conv_bwd_data_bias_relu(bool do_verification, ...@@ -58,7 +22,7 @@ int run_conv_bwd_data_bias_relu(bool do_verification,
std::cout << "bias: " << bias.mDesc << std::endl; std::cout << "bias: " << bias.mDesc << std::endl;
std::cout << "in: " << in_host.mDesc << std::endl; std::cout << "in: " << in_host.mDesc << std::endl;
switch(init_method) switch(config.init_method)
{ {
case 0: break; case 0: break;
case 1: case 1:
...@@ -97,7 +61,7 @@ int run_conv_bwd_data_bias_relu(bool do_verification, ...@@ -97,7 +61,7 @@ int run_conv_bwd_data_bias_relu(bool do_verification,
std::array<ck::index_t, NDimSpatial> input_left_pads{}; std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{}; std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); }; auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths); copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides); copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides);
...@@ -107,13 +71,15 @@ int run_conv_bwd_data_bias_relu(bool do_verification, ...@@ -107,13 +71,15 @@ int run_conv_bwd_data_bias_relu(bool do_verification,
copy(bias_g_n_c_wis_desc.GetStrides(), d0_g_n_c_wis_strides); copy(bias_g_n_c_wis_desc.GetStrides(), d0_g_n_c_wis_strides);
copy(in_g_n_c_wis_desc.GetLengths(), e_g_n_c_wis_lengths); copy(in_g_n_c_wis_desc.GetLengths(), e_g_n_c_wis_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), e_g_n_c_wis_strides); copy(in_g_n_c_wis_desc.GetStrides(), e_g_n_c_wis_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides); copy(conv_params.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations); copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads); copy(conv_params.input_left_pads_, input_left_pads);
copy(conv_param.input_right_pads_, input_right_pads); copy(conv_params.input_right_pads_, input_right_pads);
static_assert(std::is_default_constructible_v<DeviceConvInstance>);
// do conv // do conv
auto conv = DeviceInstance{}; auto conv = DeviceConvInstance{};
auto invoker = conv.MakeInvoker(); auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument( auto argument = conv.MakeArgument(
out_device_buf.GetDeviceBuffer(), out_device_buf.GetDeviceBuffer(),
...@@ -138,16 +104,17 @@ int run_conv_bwd_data_bias_relu(bool do_verification, ...@@ -138,16 +104,17 @@ int run_conv_bwd_data_bias_relu(bool do_verification,
if(!conv.IsSupportedArgument(argument)) if(!conv.IsSupportedArgument(argument))
{ {
printf("wrong! device_conv with the specified compilation parameters does " std::cerr << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem\n"); "not support this Conv problem"
<< std::endl;
return 1; return false;
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t flop = conv_param.GetFlops(); std::size_t flop = conv_params.GetFlops();
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>(); std::size_t num_btype = conv_params.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
...@@ -156,10 +123,8 @@ int run_conv_bwd_data_bias_relu(bool do_verification, ...@@ -156,10 +123,8 @@ int run_conv_bwd_data_bias_relu(bool do_verification,
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s" std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl; << std::endl;
if(do_verification) if(config.do_verification)
{ {
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// c doesn't physically exist, any layout is fine // c doesn't physically exist, any layout is fine
Tensor<float> c_host(in_g_n_c_wis_desc); Tensor<float> c_host(in_g_n_c_wis_desc);
...@@ -176,10 +141,10 @@ int run_conv_bwd_data_bias_relu(bool do_verification, ...@@ -176,10 +141,10 @@ int run_conv_bwd_data_bias_relu(bool do_verification,
auto ref_argument = ref_conv.MakeArgument(c_host, auto ref_argument = ref_conv.MakeArgument(c_host,
wei, wei,
out, out,
conv_param.conv_filter_strides_, conv_params.conv_filter_strides_,
conv_param.conv_filter_dilations_, conv_params.conv_filter_dilations_,
conv_param.input_left_pads_, conv_params.input_left_pads_,
conv_param.input_right_pads_, conv_params.input_right_pads_,
PassThrough{}, PassThrough{},
wei_element_op, wei_element_op,
out_element_op); out_element_op);
...@@ -192,8 +157,68 @@ int run_conv_bwd_data_bias_relu(bool do_verification, ...@@ -192,8 +157,68 @@ int run_conv_bwd_data_bias_relu(bool do_verification,
in_device_buf.FromDevice(in_device.mData.data()); in_device_buf.FromDevice(in_device.mData.data());
return ck::utils::check_err(in_device.mData, in_host.mData) ? 0 : 1; return ck::utils::check_err(in_device, in_host);
}
return true;
}
int run_grouped_conv_bwd_data_bias_relu_example(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
ExecutionConfig config;
ck::utils::conv::ConvParam conv_params = DefaultConvParams;
if(!parse_cmd_args(argc, argv, config, conv_params))
{
return EXIT_FAILURE;
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_params.num_dim_spatial_ != NDimSpatial)
{
std::cerr << "unsupported # of spatials dimensions" << std::endl;
return EXIT_FAILURE;
} }
return 0; // output image: GNHWK
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_params);
// weight: GKYXC
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_params);
// input image bias: G_C
const auto bias_g_n_c_wis_desc = HostTensorDescriptor({conv_params.G_,
conv_params.N_,
conv_params.C_,
conv_params.input_spatial_lengths_[0],
conv_params.input_spatial_lengths_[1]},
{
conv_params.C_, // g
0, // n
1, // c
0, // hi
0 // wi
});
// input image: GNHWC
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_params);
return !run_conv_bwd_data_bias_relu(config,
conv_params,
out_g_n_k_wos_desc,
wei_g_k_c_xs_desc,
bias_g_n_c_wis_desc,
in_g_n_c_wis_desc,
wei_element_op,
out_element_op,
in_element_op);
} }
example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc 0 → 100644
View file @ 95a83c6e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
bool run_conv_bwd_data(const ExecutionConfig& config,
const ck::utils::conv::ConvParam& conv_params,
const HostTensorDescriptor& out_g_n_k_wos_desc,
const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& in_g_n_c_wis_desc,
const OutElementOp& out_element_op,
const WeiElementOp& wei_element_op,
const InElementOp& in_element_op)
{
Tensor<OutDataType> out(out_g_n_k_wos_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
Tensor<InDataType> in_host(in_g_n_c_wis_desc);
Tensor<InDataType> in_device(in_g_n_c_wis_desc);
std::cout << "out: " << out.mDesc << std::endl;
std::cout << "wei: " << wei.mDesc << std::endl;
std::cout << "in: " << in_host.mDesc << std::endl;
switch(config.init_method)
{
case 0: break;
case 1:
out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
}
DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
out_device_buf.ToDevice(out.mData.data());
wei_device_buf.ToDevice(wei.mData.data());
// reset input to zero
in_device_buf.SetZero();
std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_lengths{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_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{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides);
copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
copy(in_g_n_c_wis_desc.GetLengths(), e_g_n_c_wis_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), e_g_n_c_wis_strides);
copy(conv_params.conv_filter_strides_, conv_filter_strides);
copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
copy(conv_params.input_left_pads_, input_left_pads);
copy(conv_params.input_right_pads_, input_right_pads);
static_assert(std::is_default_constructible_v<DeviceConvInstance>);
// do conv
auto conv = DeviceConvInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(out_device_buf.GetDeviceBuffer(),
wei_device_buf.GetDeviceBuffer(),
std::array<const void*, 0>{},
in_device_buf.GetDeviceBuffer(),
a_g_n_k_wos_lengths,
a_g_n_k_wos_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
e_g_n_c_wis_lengths,
e_g_n_c_wis_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
out_element_op,
wei_element_op,
in_element_op);
if(!conv.IsSupportedArgument(argument))
{
std::cerr << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return false;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t flop = conv_params.GetFlops();
std::size_t num_btype = conv_params.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / 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::endl;
if(config.do_verification)
{
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
PassThrough,
WeiElementOp,
OutElementOp>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_host,
wei,
out,
conv_params.conv_filter_strides_,
conv_params.conv_filter_dilations_,
conv_params.input_left_pads_,
conv_params.input_right_pads_,
PassThrough{},
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
in_device_buf.FromDevice(in_device.mData.data());
return ck::utils::check_err(in_device.mData, in_host.mData);
}
return true;
}
int run_grouped_conv_bwd_data_example(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
ExecutionConfig config;
ck::utils::conv::ConvParam conv_params = DefaultConvParams;
if(!parse_cmd_args(argc, argv, config, conv_params))
{
return EXIT_FAILURE;
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_params.num_dim_spatial_ != NDimSpatial)
{
std::cerr << "unsupported # of spatials dimensions" << std::endl;
return EXIT_FAILURE;
}
// output image: GNHWK
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_params);
// weight: GKYXC
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_params);
// input image: GNHWC
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_params);
return !run_conv_bwd_data(config,
conv_params,
out_g_n_k_wos_desc,
wei_g_k_c_xs_desc,
in_g_n_c_wis_desc,
wei_element_op,
out_element_op,
in_element_op);
}
example/39_permute/common.hpp
View file @ 95a83c6e
...@@ -19,6 +19,7 @@ ...@@ -19,6 +19,7 @@
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/utility/type.hpp" #include "ck/utility/type.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/fill.hpp" #include "ck/library/utility/fill.hpp"
...@@ -247,19 +248,6 @@ inline auto to_array(Range& range) noexcept ...@@ -247,19 +248,6 @@ inline auto to_array(Range& range) noexcept
return detail::to_array_proxy<ck::remove_cvref_t<Range>>{range}; return detail::to_array_proxy<ck::remove_cvref_t<Range>>{range};
} }
namespace ranges {
template <typename InputRange, typename OutputIterator>
inline auto copy(InputRange&& range, OutputIterator iter)
-> decltype(std::copy(std::begin(std::forward<InputRange>(range)),
std::end(std::forward<InputRange>(range)),
iter))
{
return std::copy(std::begin(std::forward<InputRange>(range)),
std::end(std::forward<InputRange>(range)),
iter);
}
} // namespace ranges
template <typename Axes> template <typename Axes>
inline auto is_valid_axes(const Axes& axes) inline auto is_valid_axes(const Axes& axes)
-> std::enable_if_t<detail::is_random_access_range_v<Axes>, bool> -> std::enable_if_t<detail::is_random_access_range_v<Axes>, bool>
...@@ -350,7 +338,7 @@ auto extend_shape(const Problem::Shape& shape, std::size_t new_dim) ...@@ -350,7 +338,7 @@ auto extend_shape(const Problem::Shape& shape, std::size_t new_dim)
using std::begin, std::end; using std::begin, std::end;
std::copy(begin(shape), end(shape), begin(extended_shape)); ck::ranges::copy(shape, begin(extended_shape));
extended_shape.back() = new_dim; extended_shape.back() = new_dim;
return extended_shape; return extended_shape;
...@@ -362,7 +350,7 @@ auto extend_axes(const Problem::Axes& axes) ...@@ -362,7 +350,7 @@ auto extend_axes(const Problem::Axes& axes)
using std::begin, std::end; using std::begin, std::end;
std::copy(begin(axes), end(axes), begin(extended_axes)); ck::ranges::copy(axes, begin(extended_axes));
extended_axes.back() = detail::get_array_size_v<Problem::Axes>; extended_axes.back() = detail::get_array_size_v<Problem::Axes>;
return extended_axes; return extended_axes;
......
example/39_permute/run_permute_bundle_example.inc
View file @ 95a83c6e
...@@ -57,7 +57,7 @@ bool run_permute_bundle(const Problem& problem) ...@@ -57,7 +57,7 @@ bool run_permute_bundle(const Problem& problem)
using std::begin; using std::begin;
Tensor<DataType> input_tensor(input_shape); Tensor<DataType> input_tensor(input_shape);
ranges::copy(input_bundle_tensor.AsSpan<const DataType>(), begin(input_tensor)); ck::ranges::copy(input_bundle_tensor.AsSpan<const DataType>(), begin(input_tensor));
Tensor<DataType> output_tensor(transpose(input_shape, input_axes)); Tensor<DataType> output_tensor(transpose(input_shape, input_axes));
if(!host_permute(input_tensor, input_axes, PassThrough{}, output_tensor)) if(!host_permute(input_tensor, input_axes, PassThrough{}, output_tensor))
......
example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
View file @ 95a83c6e
...@@ -11,6 +11,7 @@ ...@@ -11,6 +11,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
......
example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
View file @ 95a83c6e
...@@ -11,6 +11,7 @@ ...@@ -11,6 +11,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
......
example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
View file @ 95a83c6e
...@@ -11,6 +11,7 @@ ...@@ -11,6 +11,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
......
example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
View file @ 95a83c6e
...@@ -15,6 +15,7 @@ ...@@ -15,6 +15,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
......
example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
View file @ 95a83c6e
...@@ -11,6 +11,7 @@ ...@@ -11,6 +11,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
......
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