Skip to content

GitLab

Commit 2724c519 authored by Jing Zhang's avatar Jing Zhang
Browse files

merge develop

parents 1fb4a474 2eb74a9c
Hide whitespace changes
Inline Side-by-side
Showing with 700 additions and 155 deletions
client_example/08_fused_attention/CMakeLists.txt
View file @ 2724c519
add_executable(client_fused_attention fused_attention.cpp) add_executable(client_fused_attention fused_attention.cpp)
target_link_libraries(client_fused_attention PRIVATE composable_kernel::device_operations) target_link_libraries(client_fused_attention PRIVATE composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
add_executable(client_fused_attention_bias fused_attention_bias.cpp) add_executable(client_fused_attention_bias fused_attention_bias.cpp)
target_link_libraries(client_fused_attention_bias PRIVATE composable_kernel::device_operations) target_link_libraries(client_fused_attention_bias PRIVATE composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
client_example/09_quantization/CMakeLists.txt
View file @ 2724c519
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_executable(client_conv2d_fwd_bias_tanh_perchannel_quantization conv2d_fwd_bias_tanh_perchannel_quantization.cpp) add_executable(client_conv2d_fwd_bias_tanh_perchannel_quantization conv2d_fwd_bias_tanh_perchannel_quantization.cpp)
target_link_libraries(client_conv2d_fwd_bias_tanh_perchannel_quantization PRIVATE composable_kernel::device_operations) target_link_libraries(client_conv2d_fwd_bias_tanh_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
add_executable(client_conv2d_fwd_bias_relu_perchannel_quantization conv2d_fwd_bias_relu_perchannel_quantization.cpp) add_executable(client_conv2d_fwd_bias_relu_perchannel_quantization conv2d_fwd_bias_relu_perchannel_quantization.cpp)
target_link_libraries(client_conv2d_fwd_bias_relu_perchannel_quantization PRIVATE composable_kernel::device_operations) target_link_libraries(client_conv2d_fwd_bias_relu_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
add_executable(client_conv2d_fwd_bias_tanh_perlayer_quantization conv2d_fwd_bias_tanh_perlayer_quantization.cpp) add_executable(client_conv2d_fwd_bias_tanh_perlayer_quantization conv2d_fwd_bias_tanh_perlayer_quantization.cpp)
target_link_libraries(client_conv2d_fwd_bias_tanh_perlayer_quantization PRIVATE composable_kernel::device_operations) target_link_libraries(client_conv2d_fwd_bias_tanh_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
add_executable(client_conv2d_fwd_bias_relu_perlayer_quantization conv2d_fwd_bias_relu_perlayer_quantization.cpp) add_executable(client_conv2d_fwd_bias_relu_perlayer_quantization conv2d_fwd_bias_relu_perlayer_quantization.cpp)
target_link_libraries(client_conv2d_fwd_bias_relu_perlayer_quantization PRIVATE composable_kernel::device_operations) target_link_libraries(client_conv2d_fwd_bias_relu_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
add_executable(client_conv2d_fwd_perchannel_quantization conv2d_fwd_perchannel_quantization.cpp) add_executable(client_conv2d_fwd_perchannel_quantization conv2d_fwd_perchannel_quantization.cpp)
target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composable_kernel::device_operations) target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp) add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_operations) target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
add_executable(client_gemm_quantization gemm_quantization.cpp) add_executable(client_gemm_quantization gemm_quantization.cpp)
target_link_libraries(client_gemm_quantization PRIVATE composable_kernel::device_operations) target_link_libraries(client_gemm_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
endif()
client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
View file @ 2724c519
...@@ -80,7 +80,7 @@ int main(int argc, char* argv[]) ...@@ -80,7 +80,7 @@ int main(int argc, char* argv[])
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
View file @ 2724c519
...@@ -78,18 +78,18 @@ int main(int argc, char* argv[]) ...@@ -78,18 +78,18 @@ int main(int argc, char* argv[])
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<BiasLayout>, ck::Tuple<BiasLayout>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<BiasDataType>, ck::Tuple<BiasDataType>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
View file @ 2724c519
...@@ -83,7 +83,7 @@ int main(int argc, char* argv[]) ...@@ -83,7 +83,7 @@ int main(int argc, char* argv[])
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
......
client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
View file @ 2724c519
...@@ -79,18 +79,18 @@ int main(int argc, char* argv[]) ...@@ -79,18 +79,18 @@ int main(int argc, char* argv[])
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<BiasLayout>, ck::Tuple<BiasLayout>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<BiasDataType>, ck::Tuple<BiasDataType>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
View file @ 2724c519
...@@ -76,19 +76,19 @@ int main(int argc, char* argv[]) ...@@ -76,19 +76,19 @@ int main(int argc, char* argv[])
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<RequantScaleLayout>, ck::Tuple<RequantScaleLayout>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<RequantScaleDataType>, ck::Tuple<RequantScaleDataType>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
View file @ 2724c519
...@@ -72,18 +72,18 @@ int main(int argc, char* argv[]) ...@@ -72,18 +72,18 @@ int main(int argc, char* argv[])
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
ck::Tuple<>, ck::Tuple<>,
OutLayout, OutLayout,
InDataType, InDataType,
WeiDataType, WeiDataType,
ck::Tuple<>, ck::Tuple<>,
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
OutElementOp>; OutElementOp>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/10_grouped_conv2d_bwd_data/CMakeLists.txt deleted 100644 → 0
View file @ 1fb4a474
add_executable(client_grouped_conv2d_bwd_data grouped_conv2d_bwd_data.cpp)
target_link_libraries(client_grouped_conv2d_bwd_data PRIVATE composable_kernel::device_operations)
client_example/10_grouped_convnd_bwd_data/CMakeLists.txt 0 → 100644
View file @ 2724c519
add_executable(client_grouped_conv2d_bwd_data grouped_conv2d_bwd_data.cpp)
target_link_libraries(client_grouped_conv2d_bwd_data PRIVATE composable_kernel::device_conv_operations)
add_executable(client_grouped_conv3d_bwd_data grouped_conv3d_bwd_data.cpp)
target_link_libraries(client_grouped_conv3d_bwd_data PRIVATE composable_kernel::device_conv_operations)
add_executable(client_grouped_conv3d_bwd_data_input_fp16_comp_bf8f8 grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp)
target_link_libraries(client_grouped_conv3d_bwd_data_input_fp16_comp_bf8f8 PRIVATE composable_kernel::device_conv_operations)
client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 2;
static constexpr ck::index_t N = 16;
static constexpr ck::index_t K = 16;
static constexpr ck::index_t C = 16;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 14;
static constexpr ck::index_t Hi = 14;
static constexpr ck::index_t Wi = 14;
static constexpr ck::index_t Do = 14;
static constexpr ck::index_t Ho = 14;
static constexpr ck::index_t Wo = 14;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{
C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{
K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
SimpleDeviceMem in(sizeof(InDataType) * G * N * Di * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Do * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
PassThrough,
PassThrough,
PassThrough>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * G * N * K * C * Do * Ho * Wo * Y * X;
std::size_t num_bytes = sizeof(InDataType) * G * N * Di * Hi * Wi * C +
sizeof(WeiDataType) * G * K * Z * Y * X * C +
sizeof(OutDataType) * G * N * Do * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
}
client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 2;
static constexpr ck::index_t N = 16;
static constexpr ck::index_t K = 16;
static constexpr ck::index_t C = 16;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 14;
static constexpr ck::index_t Hi = 14;
static constexpr ck::index_t Wi = 14;
static constexpr ck::index_t Do = 14;
static constexpr ck::index_t Ho = 14;
static constexpr ck::index_t Wo = 14;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{
C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{
K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
SimpleDeviceMem in(sizeof(InDataType) * G * N * Di * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Do * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
PassThrough,
PassThrough,
PassThrough,
ck::bf8_t,
ck::f8_t>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * G * N * K * C * Do * Ho * Wo * Y * X;
std::size_t num_bytes = sizeof(InDataType) * G * N * Di * Hi * Wi * C +
sizeof(WeiDataType) * G * K * Z * Y * X * C +
sizeof(OutDataType) * G * N * Do * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
}
client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
View file @ 2724c519
...@@ -2,8 +2,10 @@ add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_f ...@@ -2,8 +2,10 @@ add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_f
add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp) add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp) add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp) add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_conv_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_conv_operations)
client_example/11_grouped_conv_bwd_weight/common.hpp
View file @ 2724c519
...@@ -32,63 +32,49 @@ struct SimpleDeviceMem ...@@ -32,63 +32,49 @@ struct SimpleDeviceMem
}; };
template <ck::index_t NumDimSpatial> template <ck::index_t NumDimSpatial>
std::size_t GetFlops(ck::index_t G, std::size_t GetFlops(const std::array<ck::index_t, NumDimSpatial>& output_lengths,
ck::index_t N, const std::array<ck::index_t, NumDimSpatial>& filter_lengths)
ck::index_t K,
ck::index_t C,
const std::array<ck::index_t, NumDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NumDimSpatial>& filter_spatial_lengths)
{ {
constexpr ck::index_t spatial_offset = 3;
const auto C = filter_lengths[2];
// 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product> // 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
return static_cast<std::size_t>(2) * G * N * K * C * return static_cast<std::size_t>(2) * C *
std::accumulate(std::begin(output_spatial_lengths), std::accumulate(std::begin(output_lengths),
std::end(output_spatial_lengths), std::end(output_lengths),
static_cast<std::size_t>(1), static_cast<std::size_t>(1),
std::multiplies<>()) * std::multiplies<>()) *
std::accumulate(std::begin(filter_spatial_lengths), std::accumulate(std::begin(filter_lengths) + spatial_offset,
std::end(filter_spatial_lengths), std::end(filter_lengths),
static_cast<std::size_t>(1), static_cast<std::size_t>(1),
std::multiplies<>()); std::multiplies<>());
} }
template <typename InDataType, ck::index_t NumDimSpatial> template <typename InDataType, ck::index_t NumDimSpatial>
std::size_t GetInputByte(ck::index_t G, std::size_t GetInputByte(const std::array<ck::index_t, NumDimSpatial>& input_lengths)
ck::index_t N,
ck::index_t C,
const std::array<ck::index_t, NumDimSpatial>& input_spatial_lengths)
{ {
// sizeof(InDataType) * (G * N * C * <input spatial lengths product>) + // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
return sizeof(InDataType) * (G * N * C * return sizeof(InDataType) * (std::accumulate(std::begin(input_lengths),
std::accumulate(std::begin(input_spatial_lengths), std::end(input_lengths),
std::end(input_spatial_lengths),
static_cast<std::size_t>(1), static_cast<std::size_t>(1),
std::multiplies<>())); std::multiplies<>()));
} }
template <typename WeiDataType, ck::index_t NumDimSpatial> template <typename WeiDataType, ck::index_t NumDimSpatial>
std::size_t GetWeightByte(ck::index_t G, std::size_t GetWeightByte(const std::array<ck::index_t, NumDimSpatial>& filter_lengths)
ck::index_t K,
ck::index_t C,
const std::array<ck::index_t, NumDimSpatial>& filter_spatial_lengths)
{ {
// sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) + // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
return sizeof(WeiDataType) * (G * K * C * return sizeof(WeiDataType) * (std::accumulate(std::begin(filter_lengths),
std::accumulate(std::begin(filter_spatial_lengths), std::end(filter_lengths),
std::end(filter_spatial_lengths),
static_cast<std::size_t>(1), static_cast<std::size_t>(1),
std::multiplies<>())); std::multiplies<>()));
} }
template <typename OutDataType, ck::index_t NumDimSpatial> template <typename OutDataType, ck::index_t NumDimSpatial>
std::size_t GetOutputByte(ck::index_t G, std::size_t GetOutputByte(const std::array<ck::index_t, NumDimSpatial>& output_lengths)
ck::index_t N,
ck::index_t K,
const std::array<ck::index_t, NumDimSpatial>& output_spatial_lengths)
{ {
// sizeof(OutDataType) * (G * N * K * <output spatial lengths product>); // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
return sizeof(OutDataType) * (G * N * K * return sizeof(OutDataType) * (std::accumulate(std::begin(output_lengths),
std::accumulate(std::begin(output_spatial_lengths), std::end(output_lengths),
std::end(output_spatial_lengths),
static_cast<std::size_t>(1), static_cast<std::size_t>(1),
std::multiplies<std::size_t>())); std::multiplies<std::size_t>()));
} }
...@@ -99,15 +85,16 @@ template <ck::index_t NumDimSpatial, ...@@ -99,15 +85,16 @@ template <ck::index_t NumDimSpatial,
typename OutDataType, typename OutDataType,
typename InLayout, typename InLayout,
typename WeiLayout, typename WeiLayout,
typename OutLayout> typename OutLayout,
typename AComputeType = InDataType,
typename BComputeType = AComputeType>
bool run_grouped_conv_bwd_weight( bool run_grouped_conv_bwd_weight(
ck::index_t G, const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths,
ck::index_t N, const std::array<ck::index_t, NumDimSpatial + 3>& input_strides,
ck::index_t K, const std::array<ck::index_t, NumDimSpatial + 3>& filter_lengths,
ck::index_t C, const std::array<ck::index_t, NumDimSpatial + 3>& weights_strides,
const std::array<ck::index_t, NumDimSpatial>& input_spatial_lengths, const std::array<ck::index_t, NumDimSpatial + 3>& output_lengths,
const std::array<ck::index_t, NumDimSpatial>& filter_spatial_lengths, const std::array<ck::index_t, NumDimSpatial + 3>& output_strides,
const std::array<ck::index_t, NumDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NumDimSpatial>& conv_filter_strides, const std::array<ck::index_t, NumDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NumDimSpatial>& conv_filter_dilations, const std::array<ck::index_t, NumDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NumDimSpatial>& input_left_pads, const std::array<ck::index_t, NumDimSpatial>& input_left_pads,
...@@ -115,9 +102,9 @@ bool run_grouped_conv_bwd_weight( ...@@ -115,9 +102,9 @@ bool run_grouped_conv_bwd_weight(
{ {
ck::index_t split_k = 2; ck::index_t split_k = 2;
SimpleDeviceMem in(GetInputByte<InDataType, NumDimSpatial>(G, N, C, input_spatial_lengths)); SimpleDeviceMem in(GetInputByte<InDataType, NumDimSpatial + 3>(input_lengths));
SimpleDeviceMem wei(GetWeightByte<WeiDataType, NumDimSpatial>(G, K, C, filter_spatial_lengths)); SimpleDeviceMem wei(GetWeightByte<WeiDataType, NumDimSpatial + 3>(filter_lengths));
SimpleDeviceMem out(GetOutputByte<OutDataType, NumDimSpatial>(G, N, K, output_spatial_lengths)); SimpleDeviceMem out(GetOutputByte<OutDataType, NumDimSpatial + 3>(output_lengths));
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NumDimSpatial,
InLayout, InLayout,
...@@ -128,7 +115,9 @@ bool run_grouped_conv_bwd_weight( ...@@ -128,7 +115,9 @@ bool run_grouped_conv_bwd_weight(
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>; PassThrough,
AComputeType,
BComputeType>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
...@@ -141,6 +130,10 @@ bool run_grouped_conv_bwd_weight( ...@@ -141,6 +130,10 @@ bool run_grouped_conv_bwd_weight(
float best_gb_per_sec = 0; float best_gb_per_sec = 0;
float best_tflops = 0; float best_tflops = 0;
std::array<ck::index_t, NumDimSpatial + 3> a_g_n_c_wis_lengths{};
std::array<ck::index_t, NumDimSpatial + 3> a_g_n_c_wis_strides{};
std::array<ck::index_t, NumDimSpatial + 3> b_g_k_c_xs_lengths{};
// profile device operation instances // profile device operation instances
std::cout << "Run all instances and do timing" << std::endl; std::cout << "Run all instances and do timing" << std::endl;
...@@ -150,13 +143,12 @@ bool run_grouped_conv_bwd_weight( ...@@ -150,13 +143,12 @@ bool run_grouped_conv_bwd_weight(
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(), auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(), wei.GetDeviceBuffer(),
out.GetDeviceBuffer(), out.GetDeviceBuffer(),
G, input_lengths,
N, input_strides,
K, filter_lengths,
C, weights_strides,
input_spatial_lengths, output_lengths,
filter_spatial_lengths, output_strides,
output_spatial_lengths,
conv_filter_strides, conv_filter_strides,
conv_filter_dilations, conv_filter_dilations,
input_left_pads, input_left_pads,
...@@ -172,12 +164,10 @@ bool run_grouped_conv_bwd_weight( ...@@ -172,12 +164,10 @@ bool run_grouped_conv_bwd_weight(
{ {
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t flop = GetFlops<NumDimSpatial + 3>(output_lengths, filter_lengths);
GetFlops<NumDimSpatial>(G, N, K, C, output_spatial_lengths, filter_spatial_lengths); std::size_t num_bytes = GetInputByte<InDataType, NumDimSpatial + 3>(input_lengths) +
std::size_t num_bytes = GetWeightByte<WeiDataType, NumDimSpatial + 3>(filter_lengths) +
GetInputByte<InDataType, NumDimSpatial>(G, N, C, input_spatial_lengths) + GetOutputByte<OutDataType, NumDimSpatial + 3>(output_lengths);
GetWeightByte<WeiDataType, NumDimSpatial>(G, K, C, filter_spatial_lengths) +
GetOutputByte<OutDataType, NumDimSpatial>(G, N, K, output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / avg_time; float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
...@@ -217,13 +207,12 @@ bool run_grouped_conv_bwd_weight( ...@@ -217,13 +207,12 @@ bool run_grouped_conv_bwd_weight(
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(), auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(), wei.GetDeviceBuffer(),
out.GetDeviceBuffer(), out.GetDeviceBuffer(),
G, input_lengths,
N, input_strides,
K, filter_lengths,
C, weights_strides,
input_spatial_lengths, output_lengths,
filter_spatial_lengths, output_strides,
output_spatial_lengths,
conv_filter_strides, conv_filter_strides,
conv_filter_dilations, conv_filter_dilations,
input_left_pads, input_left_pads,
......
client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp
View file @ 2724c519
...@@ -22,6 +22,16 @@ static constexpr ck::index_t C = 192; ...@@ -22,6 +22,16 @@ static constexpr ck::index_t C = 192;
static constexpr ck::index_t X = 3; static constexpr ck::index_t X = 3;
static constexpr ck::index_t Wi = 28; static constexpr ck::index_t Wi = 28;
static constexpr ck::index_t Wo = 28; static constexpr ck::index_t Wo = 28;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{N * Wi * C, Wi* C, 1, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{K * X * C, X* C, 1, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{N * Wo * K, Wo* K, 1, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1};
int main() int main()
{ {
...@@ -31,7 +41,16 @@ int main() ...@@ -31,7 +41,16 @@ int main()
OutDataType, OutDataType,
InLayout, InLayout,
WeiLayout, WeiLayout,
OutLayout>(G, N, K, C, {Wi}, {X}, {Wo}, {1}, {1}, {1}, {1}) OutLayout>(input_lengths,
input_strides,
filter_lengths,
weights_strides,
output_lengths,
output_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads)
? EXIT_SUCCESS ? EXIT_SUCCESS
: EXIT_FAILURE; : EXIT_FAILURE;
} }
client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp
View file @ 2724c519
...@@ -25,6 +25,19 @@ static constexpr ck::index_t Hi = 28; ...@@ -25,6 +25,19 @@ static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 28; static constexpr ck::index_t Wi = 28;
static constexpr ck::index_t Ho = 28; static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 28; static constexpr ck::index_t Wo = 28;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
N * Hi * Wi * C, Hi* Wi* C, 1, Wi* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Y * X * C, Y* X* C, 1, X* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
N * Ho * Wo * K, Ho* Wo* K, 1, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
int main() int main()
{ {
...@@ -34,8 +47,16 @@ int main() ...@@ -34,8 +47,16 @@ int main()
OutDataType, OutDataType,
InLayout, InLayout,
WeiLayout, WeiLayout,
OutLayout>( OutLayout>(input_lengths,
G, N, K, C, {Hi, Wi}, {Y, X}, {Ho, Wo}, {1, 1}, {1, 1}, {1, 1}, {1, 1}) input_strides,
filter_lengths,
weights_strides,
output_lengths,
output_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads)
? EXIT_SUCCESS ? EXIT_SUCCESS
: EXIT_FAILURE; : EXIT_FAILURE;
} }
client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp
View file @ 2724c519
...@@ -28,6 +28,19 @@ static constexpr ck::index_t Wi = 3; ...@@ -28,6 +28,19 @@ static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28; static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28; static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3; static constexpr ck::index_t Wo = 3;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
int main() int main()
{ {
...@@ -37,17 +50,16 @@ int main() ...@@ -37,17 +50,16 @@ int main()
OutDataType, OutDataType,
InLayout, InLayout,
WeiLayout, WeiLayout,
OutLayout>(G, OutLayout>(input_lengths,
N, input_strides,
K, filter_lengths,
C, weights_strides,
{Di, Hi, Wi}, output_lengths,
{Z, Y, X}, output_strides,
{Do, Ho, Wo}, conv_filter_strides,
{1, 1, 1}, conv_filter_dilations,
{1, 1, 1}, input_left_pads,
{1, 1, 1}, input_right_pads)
{1, 1, 1})
? EXIT_SUCCESS ? EXIT_SUCCESS
: EXIT_FAILURE; : EXIT_FAILURE;
} }
client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using AComputeType = ck::bf8_t;
using BComputeType = ck::f8_t;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 8;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 128;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 28;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
int main()
{
return run_grouped_conv_bwd_weight<NumDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InLayout,
WeiLayout,
OutLayout,
AComputeType,
BComputeType>(input_lengths,
input_strides,
filter_lengths,
weights_strides,
output_lengths,
output_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads)
? EXIT_SUCCESS
: EXIT_FAILURE;
}
client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp
View file @ 2724c519
...@@ -28,6 +28,19 @@ static constexpr ck::index_t Wi = 3; ...@@ -28,6 +28,19 @@ static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28; static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28; static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3; static constexpr ck::index_t Wo = 3;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
int main() int main()
{ {
...@@ -37,17 +50,16 @@ int main() ...@@ -37,17 +50,16 @@ int main()
OutDataType, OutDataType,
InLayout, InLayout,
WeiLayout, WeiLayout,
OutLayout>(G, OutLayout>(input_lengths,
N, input_strides,
K, filter_lengths,
C, weights_strides,
{Di, Hi, Wi}, output_lengths,
{Z, Y, X}, output_strides,
{Do, Ho, Wo}, conv_filter_strides,
{1, 1, 1}, conv_filter_dilations,
{1, 1, 1}, input_left_pads,
{1, 1, 1}, input_right_pads)
{1, 1, 1})
? EXIT_SUCCESS ? EXIT_SUCCESS
: EXIT_FAILURE; : EXIT_FAILURE;
} }
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