Skip to content

GitLab

Commit 92a0945d authored by Chao Liu's avatar Chao Liu
Browse files

convnd_fwd fp16 example

parent 63914743
Hide whitespace changes
Inline Side-by-side
Showing with 327 additions and 822 deletions
README.md
View file @ 92a0945d
...@@ -10,7 +10,7 @@ rocm/tensorflow:rocm5.1-tf2.6-dev \ ...@@ -10,7 +10,7 @@ rocm/tensorflow:rocm5.1-tf2.6-dev \
/bin/bash /bin/bash
``` ```
# Install the new rocm-cmake version # Install newer version of rocm-cmake
https://github.com/RadeonOpenCompute/rocm-cmake https://github.com/RadeonOpenCompute/rocm-cmake
## Build ## Build
...@@ -54,6 +54,7 @@ make install ...@@ -54,6 +54,7 @@ make install
``` ```
## Using CK as pre-built kernel library ## Using CK as pre-built kernel library
Instructions for using CK as a pre-built kernel library are under ```client_example/```
## Caveat ## Caveat
### Kernel Timing and Verification ### Kernel Timing and Verification
......
example/09_convnd_fwd/CMakeLists.txt
View file @ 92a0945d
...@@ -3,7 +3,3 @@ add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp) ...@@ -3,7 +3,3 @@ add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp) add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp) add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
target_link_libraries(example_convnd_fwd_xdl_fp64 PRIVATE conv_util)
target_link_libraries(example_convnd_fwd_xdl_fp32 PRIVATE conv_util)
target_link_libraries(example_convnd_fwd_xdl_int8 PRIVATE conv_util)
target_link_libraries(example_convnd_fwd_xdl_fp16 PRIVATE conv_util)
example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
View file @ 92a0945d
...@@ -8,17 +8,17 @@ ...@@ -8,17 +8,17 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp" #include "ck/tensor_operation/gpu/device/device_convnd_fwd_nwc_kxc_nwk_xdl.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/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp" #include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/host_tensor/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
namespace { #include "parse_conv_parameter.hpp"
using InDataType = ck::half_t; using InDataType = ck::half_t;
using WeiDataType = ck::half_t; using WeiDataType = ck::half_t;
...@@ -39,47 +39,42 @@ using OutElementOp = ck::tensor_operation::element_wise::PassThrough; ...@@ -39,47 +39,42 @@ using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvFwdDefault = static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
using DeviceConvFwdBasePtr =
ck::tensor_operation::device::DeviceConvFwdPtr<InElementOp, WeiElementOp, OutElementOp>;
template <ck::index_t NumDimSpatial> template <ck::index_t NumDimSpatial>
using DeviceConvNDFwdInstance = ck::tensor_operation::device:: using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwcKxcNwk_Xdl<
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< InDataType, //
// clang-format off WeiDataType, //
InDataType, // OutDataType, //
WeiDataType, // AccDataType, //
OutDataType, // InElementOp, // Input Elementwise Operation
AccDataType, // WeiElementOp, // Weights Elementwise Operation
InElementOp, // Input Elementwise Operation OutElementOp, // Output Elementwise Operation
WeiElementOp, // Weights Elementwise Operation ConvFwdDefault, // ConvForwardSpecialization
OutElementOp, // Output Elementwise Operation NumDimSpatial, // NumDimSpatial
ConvFwdDefault, // ConvForwardSpecialization 256, // BlockSize
NumDimSpatial, // NumDimSpatial 128, // MPerBlock
256, // BlockSize 256, // NPerBlock
128, // MPerBlock 4, // K0PerBlock
256, // NPerBlock 8, // K1
4, // K0PerBlock 32, // MPerXdl
8, // K1 32, // NPerXdl
32, // MPerXdl 2, // MXdlPerWave
32, // NPerXdl 4, // NXdlPerWave
2, // MXdlPerWave S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
4, // NXdlPerWave S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1 S<1, 0, 2>, // ABlockTransferSrcAccessOrder
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder 2, // ABlockTransferSrcVectorDim
S<1, 0, 2>, // ABlockTransferSrcAccessOrder 8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferSrcVectorDim 8, // ABlockTransferDstScalarPerVector_K1
8, // ABlockTransferSrcScalarPerVector true, // ABlockLdsAddExtraM
8, // ABlockTransferDstScalarPerVector_K1 S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
true, // ABlockLdsAddExtraM S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1 S<1, 0, 2>, // BBlockTransferSrcAccessOrder
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder 2, // BBlockTransferSrcVectorDim
S<1, 0, 2>, // BBlockTransferSrcAccessOrder 8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferSrcVectorDim 8, // BBlockTransferDstScalarPerVector_K1
8, // BBlockTransferSrcScalarPerVector true, // BBlockLdsAddExtraN
8, // BBlockTransferDstScalarPerVector_K1 7, // CThreadTransferSrcDstVectorDim
true, // BBlockLdsAddExtraN 1>; // CThreadTransferDstScalarPerVector
7, // CThreadTransferSrcDstVectorDim
1>; // CThreadTransferDstScalarPerVector
template <ck::index_t NumDimSpatial> template <ck::index_t NumDimSpatial>
using ReferenceConvNDFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType, using ReferenceConvNDFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
...@@ -90,107 +85,14 @@ using ReferenceConvNDFwdInstance = ck::tensor_operation::host::ReferenceConvFwd< ...@@ -90,107 +85,14 @@ using ReferenceConvNDFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<
OutElementOp, OutElementOp,
NumDimSpatial>; NumDimSpatial>;
DeviceConvFwdBasePtr get_conv_instance(int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 3: {
return std::make_unique<DeviceConvNDFwdInstance<3>>();
}
case 2: {
return std::make_unique<DeviceConvNDFwdInstance<2>>();
}
case 1: {
return std::make_unique<DeviceConvNDFwdInstance<1>>();
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
void print_use_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=n0, 1=yes)\n"
<< "arg4: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
<< " <filter spatial dimensions>, (ie Y, X for 2D)\n"
<< " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
<< " <strides>, (ie Sy, Sx for 2D)\n"
<< " <dilations>, (ie Dy, Dx for 2D)\n"
<< " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
<< " <right padding>, (ie RightPy, RightPx for 2D)\n"
<< std::endl;
}
ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, int argc, char* argv[])
{
// (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
int conv_args = 3 + num_dim_spatial * 6;
int cmdline_nargs = conv_args + 5;
if(cmdline_nargs != argc)
{
print_use_msg();
exit(0);
}
ck::utils::conv::ConvParams params;
int arg_idx = 5;
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
}
} // anonymous namespace
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
using namespace ck::utils::conv;
bool do_verification = true; bool do_verification = true;
int init_method = 1; int init_method = 1;
bool time_kernel = false; bool time_kernel = true;
int num_dim_spatial = 2; int num_dim_spatial = 2;
ck::utils::conv::ConvParams params; ck::tensor_operation::device::ConvParams params;
if(argc >= 5) if(argc >= 5)
{ {
...@@ -205,29 +107,25 @@ int main(int argc, char* argv[]) ...@@ -205,29 +107,25 @@ int main(int argc, char* argv[])
params = parse_conv_params(num_dim_spatial, argc, argv); params = parse_conv_params(num_dim_spatial, argc, argv);
} }
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_), auto f_nchw_host_tensor_descriptor =
static_cast<std::size_t>(params.C_)}; [](ck::index_t n, ck::index_t c, std::vector<ck::index_t> spatial_lengths) {
input_dims.insert(std::end(input_dims), std::vector<std::size_t> nhwc_lengths{static_cast<std::size_t>(n),
std::begin(params.input_spatial_lengths_), static_cast<std::size_t>(c)};
std::end(params.input_spatial_lengths_)); nhwc_lengths.insert(
nhwc_lengths.begin() + 1, spatial_lengths.begin(), spatial_lengths.end());
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_), return transpose_host_tensor_descriptor_given_new2old(
static_cast<std::size_t>(params.C_)}; HostTensorDescriptor(nhwc_lengths), std::vector<std::size_t>({0, 3, 1, 2}));
filter_dims.insert(std::end(filter_dims), };
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
Tensor<InDataType> input(get_input_host_tensor_descriptor(input_dims, num_dim_spatial)); Tensor<InDataType> input(
Tensor<WeiDataType> weights(get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial)); f_nchw_host_tensor_descriptor(params.N_, params.C_, params.input_spatial_lengths_));
Tensor<OutDataType> host_output(get_output_host_tensor_descriptor(output_dims, num_dim_spatial)); Tensor<InDataType> weights(
Tensor<OutDataType> device_output(get_output_host_tensor_descriptor(output_dims, num_dim_spatial)); f_nchw_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_));
Tensor<InDataType> host_output(
f_nchw_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
Tensor<InDataType> device_output(
f_nchw_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
std::cout << "input: " << input.mDesc << std::endl; std::cout << "input: " << input.mDesc << std::endl;
std::cout << "weights: " << weights.mDesc << std::endl; std::cout << "weights: " << weights.mDesc << std::endl;
...@@ -253,49 +151,41 @@ int main(int argc, char* argv[]) ...@@ -253,49 +151,41 @@ int main(int argc, char* argv[])
wei_device_buf.ToDevice(weights.mData.data()); wei_device_buf.ToDevice(weights.mData.data());
// do GEMM // do GEMM
auto conv = get_conv_instance(num_dim_spatial); auto conv = DeviceConvNDFwdInstance<2>{};
auto invoker = conv->MakeInvokerPointer(); auto invoker = conv.MakeInvoker();
auto argument = auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()), static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()), static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()), params.N_,
params.N_, params.K_,
params.K_, params.C_,
params.C_, params.input_spatial_lengths_,
params.input_spatial_lengths_, params.filter_spatial_lengths_,
params.filter_spatial_lengths_, params.GetOutputSpatialLengths(),
output_spatial_lengths, params.conv_filter_strides_,
params.conv_filter_strides_, params.conv_filter_dilations_,
params.conv_filter_dilations_, params.input_left_pads_,
params.input_left_pads_, params.input_right_pads_,
params.input_right_pads_, InElementOp{},
InElementOp{}, WeiElementOp{},
WeiElementOp{}, OutElementOp{});
OutElementOp{});
if(!conv.IsSupportedArgument(argument))
if(!conv->IsSupportedArgument(argument.get()))
{ {
throw std::runtime_error( throw std::runtime_error(
"wrong! device_conv with the specified compilation parameters does " "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"); "not support this Conv problem");
} }
float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = get_flops( std::size_t flop = params.GetFlops();
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths); std::size_t num_btype = params.GetByte<InDataType, WeiDataType, OutDataType>();
std::size_t num_btype = get_btype<InDataType, WeiDataType, OutDataType>(
params.N_,
params.C_,
params.K_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time; float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< conv->GetTypeString() << std::endl; << conv.GetTypeString() << std::endl;
if(do_verification) if(do_verification)
{ {
...@@ -315,22 +205,27 @@ int main(int argc, char* argv[]) ...@@ -315,22 +205,27 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(device_output.mData.data()); out_device_buf.FromDevice(device_output.mData.data());
return ck::utils::check_err( return ck::utils::check_err(host_output.mData,
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f) ? 0 : 1; device_output.mData,
"Error: incorrect results!",
1e-5f,
1e-4f)
? 0
: 1;
}; };
switch(num_dim_spatial) switch(num_dim_spatial)
{ {
case 3: { case 1: {
auto ref_conv = ReferenceConvNDFwdInstance<3>(); auto ref_conv = ReferenceConvNDFwdInstance<1>();
return verify_f(ref_conv); return verify_f(ref_conv);
} }
case 2: { case 2: {
auto ref_conv = ReferenceConvNDFwdInstance<2>(); auto ref_conv = ReferenceConvNDFwdInstance<2>();
return verify_f(ref_conv); return verify_f(ref_conv);
} }
case 1: { case 3: {
auto ref_conv = ReferenceConvNDFwdInstance<1>(); auto ref_conv = ReferenceConvNDFwdInstance<3>();
return verify_f(ref_conv); return verify_f(ref_conv);
} }
default: { default: {
......
example/09_convnd_fwd/parse_conv_parameter.hpp 0 → 100644
View file @ 92a0945d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
ck::tensor_operation::device::ConvParams
parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[])
{
ck::tensor_operation::device::ConvParams params;
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
}
void print_use_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=n0, 1=yes)\n"
<< "arg4: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
<< " <filter spatial dimensions>, (ie Y, X for 2D)\n"
<< " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
<< " <strides>, (ie Sy, Sx for 2D)\n"
<< " <dilations>, (ie Dy, Dx for 2D)\n"
<< " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
<< " <right padding>, (ie RightPy, RightPx for 2D)\n"
<< std::endl;
}
example/CMakeLists.txt
View file @ 92a0945d
...@@ -8,7 +8,7 @@ add_custom_target(examples) ...@@ -8,7 +8,7 @@ add_custom_target(examples)
function(add_example_executable EXAMPLE_NAME FILE_NAME) function(add_example_executable EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}") message("adding example ${EXAMPLE_NAME}")
add_executable(${EXAMPLE_NAME} ${FILE_NAME}) add_executable(${EXAMPLE_NAME} ${FILE_NAME})
target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor) target_link_libraries(${EXAMPLE_NAME} PRIVATE utility)
add_test(NAME ${EXAMPLE_NAME} COMMAND $<TARGET_FILE:${EXAMPLE_NAME}> ${ARGN}) add_test(NAME ${EXAMPLE_NAME} COMMAND $<TARGET_FILE:${EXAMPLE_NAME}> ${ARGN})
add_dependencies(examples ${EXAMPLE_NAME}) add_dependencies(examples ${EXAMPLE_NAME})
add_dependencies(check ${EXAMPLE_NAME}) add_dependencies(check ${EXAMPLE_NAME})
...@@ -17,7 +17,7 @@ endfunction(add_example_executable EXAMPLE_NAME) ...@@ -17,7 +17,7 @@ endfunction(add_example_executable EXAMPLE_NAME)
function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME) function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}") message("adding example ${EXAMPLE_NAME}")
add_executable(${EXAMPLE_NAME} ${FILE_NAME}) add_executable(${EXAMPLE_NAME} ${FILE_NAME})
target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor) target_link_libraries(${EXAMPLE_NAME} PRIVATE utility)
add_dependencies(examples ${EXAMPLE_NAME}) add_dependencies(examples ${EXAMPLE_NAME})
endfunction(add_example_executable_no_testing EXAMPLE_NAME) endfunction(add_example_executable_no_testing EXAMPLE_NAME)
......
include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
View file @ 92a0945d
...@@ -12,7 +12,14 @@ namespace ck { ...@@ -12,7 +12,14 @@ namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
template <typename InElementwiseOperation, template <ck::index_t NumDimSpatial,
typename InLayout,
typename WeiLayout,
typename OutLayout,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InElementwiseOperation,
typename WeiElementwiseOperation, typename WeiElementwiseOperation,
typename OutElementwiseOperation> typename OutElementwiseOperation>
struct DeviceConvFwd : public BaseOperator struct DeviceConvFwd : public BaseOperator
...@@ -38,11 +45,28 @@ struct DeviceConvFwd : public BaseOperator ...@@ -38,11 +45,28 @@ struct DeviceConvFwd : public BaseOperator
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0; virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
}; };
template <typename InElementwiseOperation, #if 0
template <ck::index_t NumDimSpatial,
typename InLayout,
typename WeiLayout,
typename OutLayout,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InElementwiseOperation,
typename WeiElementwiseOperation, typename WeiElementwiseOperation,
typename OutElementwiseOperation> typename OutElementwiseOperation>
using DeviceConvFwdPtr = std::unique_ptr< using DeviceConvFwdPtr = std::unique_ptr<DeviceConvFwd<NumDimSpatial,
DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>>; InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation>>;
#endif
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp → include/ck/tensor_operation/gpu/device/device_convnd_fwd_nwc_kxc_nwk_xdl.hpp
View file @ 92a0945d
...@@ -27,10 +27,10 @@ namespace device { ...@@ -27,10 +27,10 @@ namespace device {
// @brief Device Convolution operation. // @brief Device Convolution operation.
// //
// Supports: // Supports:
// @li Inputs with up to 3 spatial dimentions // @li Forward convolution with up to 3 spatial dimentions
// @li Input tensor in NHWC data format // @li Input tensor in NWC data format
// @li Weight tensor in KYXC data format // @li Weight tensor in KXC data format
// @li Output tensor in NHWK data format // @li Output tensor in NWK data format
// //
// 1D: // 1D:
// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C] // out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
...@@ -73,10 +73,28 @@ template <typename InDataType, ...@@ -73,10 +73,28 @@ template <typename InDataType,
bool BBlockLdsAddExtraN, bool BBlockLdsAddExtraN,
ck::index_t CThreadTransferSrcDstVectorDim, ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector> ck::index_t CThreadTransferDstScalarPerVector>
struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K struct DeviceConvNdFwdNwcKxcNwk_Xdl
: public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation> : public DeviceConvFwd<NumDimSpatial,
ck::tuple_element_t<NumDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::NDHWC>>,
ck::tuple_element_t<NumDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::KZYXC>>,
ck::tuple_element_t<NumDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::NDHWC>>,
InDataType,
WeiDataType,
OutDataType,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation>
{ {
using DeviceOp = DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K; using DeviceOp = DeviceConvNdFwdNwcKxcNwk_Xdl;
using ADataType = InDataType; using ADataType = InDataType;
using BDataType = WeiDataType; using BDataType = WeiDataType;
...@@ -1027,7 +1045,7 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K ...@@ -1027,7 +1045,7 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
auto str = std::stringstream(); auto str = std::stringstream();
// clang-format off // clang-format off
str << "DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K" str << "DeviceConvNdFwdNwcKxcNwk_Xdl"
<< "<" << "<"
<< BlockSize << ", " << BlockSize << ", "
<< MPerBlock << ", " << MPerBlock << ", "
......
include/ck/tensor_operation/gpu/device/device_gemm.hpp
View file @ 92a0945d
...@@ -46,25 +46,6 @@ struct DeviceGemm : public BaseOperator ...@@ -46,25 +46,6 @@ struct DeviceGemm : public BaseOperator
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0; virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
}; };
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
using DeviceGemmPtr = std::unique_ptr<DeviceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>>;
template <typename AElementwiseOperation, template <typename AElementwiseOperation,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CElementwiseOperation> typename CElementwiseOperation>
......
library/CMakeLists.txt
View file @ 92a0945d
add_subdirectory(src/tensor_operation_instance/gpu) add_subdirectory(src/tensor_operation_instance/gpu)
add_subdirectory(src/host_tensor)
add_subdirectory(src/utility) add_subdirectory(src/utility)
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View file @ 92a0945d
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include <sstream> #include <sstream>
#include "ck/tensor_operation/gpu/device/device_base.hpp" #include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/host_tensor/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
...@@ -267,7 +267,10 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -267,7 +267,10 @@ struct ReferenceConvFwd : public device::BaseOperator
return true; return true;
} }
bool IsSupportedArgument(const device::BaseArgument*) override { return true; } bool IsSupportedArgument(const device::BaseArgument*) override
{
return NumDimSpatial >= 1 && NumDimSpatial <= 3;
}
static auto MakeArgument(const Tensor<InDataType>& input, static auto MakeArgument(const Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight, const Tensor<WeiDataType>& weight,
......
library/include/ck/library/utility/check_err.hpp
View file @ 92a0945d
...@@ -13,8 +13,11 @@ ...@@ -13,8 +13,11 @@
#include <type_traits> #include <type_traits>
#include <vector> #include <vector>
#include "ck/ck.hpp"
#include "ck/utility/data_type.hpp" #include "ck/utility/data_type.hpp"
#include "ck/library/utility/io.hpp"
namespace ck { namespace ck {
namespace utils { namespace utils {
...@@ -201,10 +204,3 @@ check_err(const std::vector<T>& out, ...@@ -201,10 +204,3 @@ check_err(const std::vector<T>& out,
} // namespace utils } // namespace utils
} // namespace ck } // namespace ck
template <typename T>
std::ostream& operator<<(std::ostream& os, const std::vector<T>& v)
{
std::copy(std::begin(v), std::end(v), std::ostream_iterator<T>(os, " "));
return os;
}
library/include/ck/library/utility/conv_util.hpp deleted 100644 → 0
View file @ 63914743
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <functional>
#include <iterator>
#include <numeric>
#include <sstream>
#include <tuple>
#include <type_traits>
#include <vector>
#include "ck/ck.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"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/op_instance_engine.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
using DeviceConvFwdNoOpPtr = DeviceConvFwdPtr<element_wise::PassThrough,
element_wise::PassThrough,
element_wise::PassThrough>;
namespace instance {
void add_device_conv1d_fwd_xdl_nwc_kxc_nwk_bf16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv1d_fwd_xdl_nwc_kxc_nwk_f16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv1d_fwd_xdl_nwc_kxc_nwk_f32_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv1d_fwd_xdl_nwc_kxc_nwk_int8_instances(std::vector<DeviceConvFwdNoOpPtr>&);
} // namespace instance
namespace instance {
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(std::vector<DeviceConvFwdNoOpPtr>&);
} // namespace instance
namespace instance {
void add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_f16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_f32_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_int8_instances(std::vector<DeviceConvFwdNoOpPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace utils {
namespace conv {
using DeviceConvFwdNoOpPtr =
ck::tensor_operation::device::DeviceConvFwdPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
/**
* @brief Calculate number of FLOPs for Convolution
*
* @param[in] N Batch size.
* @param[in] C Number of input channels.
* @param[in] K Number of output channels.
* @param[in] filter_spatial_lengths Filter spatial dimensions lengths.
* @param[in] output_spatial_lengths Convolution output spatial dimensions
* lengths.
*
* @return The number of flops.
*/
std::size_t get_flops(ck::index_t N,
ck::index_t C,
ck::index_t K,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths);
/**
* @brief Calculate number of bytes read/write by convolution algorithm.
*
* @param[in] N Batch size.
* @param[in] C Number of input channels.
* @param[in] K Number of output channels.
* @param[in] input_spatial_lengths Input spatial dimensions lengths.
* @param[in] filter_spatial_lengths Filter spatial dimensions lengths.
* @param[in] output_spatial_lengths Output spatial dimensions lengths
*
* @tparam InDataType Input tensor data type.
* @tparam WeiDataType Weights tensor data type.
* @tparam OutDataType Output tensor data type.
*
* @return The number of used bytes.
*/
template <typename InDataType = float,
typename WeiDataType = InDataType,
typename OutDataType = InDataType>
std::size_t get_btype(ck::index_t N,
ck::index_t C,
ck::index_t K,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths)
{
// sizeof(InDataType) * (N * C * <input spatial lengths product>) +
// sizeof(WeiDataType) * (K * C * <filter spatial lengths product>) +
// sizeof(OutDataType) * (N * K * <output spatial lengths product>);
return sizeof(InDataType) * (N * C *
std::accumulate(std::begin(input_spatial_lengths),
std::end(input_spatial_lengths),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>())) +
sizeof(WeiDataType) * (K * C *
std::accumulate(std::begin(filter_spatial_lengths),
std::end(filter_spatial_lengths),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>())) +
sizeof(OutDataType) * (N * K *
std::accumulate(std::begin(output_spatial_lengths),
std::end(output_spatial_lengths),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>()));
}
struct ConvParams
{
ConvParams();
ConvParams(ck::index_t n_dim,
ck::index_t n_batch,
ck::index_t n_out_channels,
ck::index_t n_in_channels,
const std::vector<ck::index_t>& filters_len,
const std::vector<ck::index_t>& input_len,
const std::vector<ck::index_t>& strides,
const std::vector<ck::index_t>& dilations,
const std::vector<ck::index_t>& left_pads,
const std::vector<ck::index_t>& right_pads);
ck::index_t num_dim_spatial_;
ck::index_t N_;
ck::index_t K_;
ck::index_t C_;
std::vector<ck::index_t> filter_spatial_lengths_;
std::vector<ck::index_t> input_spatial_lengths_;
std::vector<ck::index_t> conv_filter_strides_;
std::vector<ck::index_t> conv_filter_dilations_;
std::vector<ck::index_t> input_left_pads_;
std::vector<ck::index_t> input_right_pads_;
std::vector<ck::index_t> GetOutputSpatialLengths() const;
};
ConvParams parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[]);
/**
* @brief Gets the host tensor descriptor.
*
* @param[in] dims The tensor dimensions lengths. Always in NCHW format.
* @param[in] layout The tensor data layout.
*
* @tparam TensorLayout Layout type.
*
* @return The host tensor descriptor object.
*/
template <typename TensorLayout>
HostTensorDescriptor get_host_tensor_descriptor(const std::vector<std::size_t>& dims,
const TensorLayout& layout)
{
std::size_t C = dims[1];
// 1D
if constexpr(std::is_same<TensorLayout, ck::tensor_layout::convolution::NCW>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::KCX>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::NKW>::value)
{
return HostTensorDescriptor(dims, std::vector<std::size_t>{C * dims[2], dims[2], 1});
}
else if constexpr(std::is_same<TensorLayout, ck::tensor_layout::convolution::NWC>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::KXC>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::NWK>::value)
{
return HostTensorDescriptor(dims, std::vector<std::size_t>{C * dims[2], 1, C});
}
// 2D
else if constexpr(std::is_same<TensorLayout, ck::tensor_layout::convolution::NCHW>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::KCYX>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(
dims, std::vector<std::size_t>{C * dims[2] * dims[3], dims[2] * dims[3], dims[3], 1});
}
else if constexpr(std::is_same<TensorLayout, ck::tensor_layout::convolution::NHWC>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::KYXC>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(
dims, std::vector<std::size_t>{C * dims[2] * dims[3], 1, dims[3] * C, C});
}
// 3D
else if constexpr(std::is_same<TensorLayout, ck::tensor_layout::convolution::NCDHW>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::KCZYX>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::NKDHW>::value)
{
return HostTensorDescriptor(dims,
std::vector<std::size_t>{C * dims[2] * dims[3] * dims[4],
dims[2] * dims[3] * dims[4],
dims[3] * dims[4],
dims[4],
1});
}
else if constexpr(std::is_same<TensorLayout, ck::tensor_layout::convolution::NDHWC>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::KZYXC>::value ||
std::is_same<TensorLayout, ck::tensor_layout::convolution::NDHWK>::value)
{
return HostTensorDescriptor(
dims,
std::vector<std::size_t>{
C * dims[2] * dims[3] * dims[4], 1, C * dims[3] * dims[4], C * dims[4], C});
}
std::stringstream err_msg;
err_msg << "Unsupported data layout provided: " << layout << "!";
throw std::runtime_error(err_msg.str());
}
HostTensorDescriptor get_output_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2);
HostTensorDescriptor get_filters_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2);
HostTensorDescriptor get_input_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2);
template <ck::index_t NDim,
typename InDataType = float,
typename WeiDataType = float,
typename OutDataType = float>
void run_reference_convolution_forward(const ConvParams& params,
const Tensor<InDataType>& input,
const Tensor<WeiDataType>& weights,
Tensor<OutDataType>& output)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
WeiDataType,
OutDataType,
PassThrough,
PassThrough,
PassThrough,
NDim>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(input,
weights,
output,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
PassThrough{},
PassThrough{},
PassThrough{});
ref_invoker.Run(ref_argument);
}
template <typename InDataType, typename WeiDataType, typename OutDataType>
struct ConvolutionFwdInstances;
template <>
struct ConvolutionFwdInstances<float, float, float>
{
template <int NumDimSpatial,
typename std::enable_if<NumDimSpatial >= 1 && NumDimSpatial <= 3, bool>::type = false>
static std::vector<DeviceConvFwdNoOpPtr> Get()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if constexpr(NumDimSpatial == 1)
{
ck::tensor_operation::device::instance::
add_device_conv1d_fwd_xdl_nwc_kxc_nwk_f32_instances(conv_ptrs);
}
else if constexpr(NumDimSpatial == 2)
{
ck::tensor_operation::device::instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
else if constexpr(NumDimSpatial == 3)
{
ck::tensor_operation::device::instance::
add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_f32_instances(conv_ptrs);
}
return conv_ptrs;
}
};
template <>
struct ConvolutionFwdInstances<half_t, half_t, half_t>
{
template <int NumDimSpatial,
typename std::enable_if<NumDimSpatial >= 1 && NumDimSpatial <= 3, bool>::type = false>
static std::vector<DeviceConvFwdNoOpPtr> Get()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if constexpr(NumDimSpatial == 1)
{
ck::tensor_operation::device::instance::
add_device_conv1d_fwd_xdl_nwc_kxc_nwk_f16_instances(conv_ptrs);
return conv_ptrs;
}
else if constexpr(NumDimSpatial == 2)
{
ck::tensor_operation::device::instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
ck::tensor_operation::device::instance::
add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
else if constexpr(NumDimSpatial == 3)
{
ck::tensor_operation::device::instance::
add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_f16_instances(conv_ptrs);
}
return conv_ptrs;
}
};
template <>
struct ConvolutionFwdInstances<bhalf_t, bhalf_t, bhalf_t>
{
template <int NumDimSpatial,
typename std::enable_if<NumDimSpatial >= 1 && NumDimSpatial <= 3, bool>::type = false>
static std::vector<DeviceConvFwdNoOpPtr> Get()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if constexpr(NumDimSpatial == 1)
{
ck::tensor_operation::device::instance::
add_device_conv1d_fwd_xdl_nwc_kxc_nwk_bf16_instances(conv_ptrs);
}
else if constexpr(NumDimSpatial == 2)
{
ck::tensor_operation::device::instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
}
else if constexpr(NumDimSpatial == 3)
{
ck::tensor_operation::device::instance::
add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(conv_ptrs);
}
return conv_ptrs;
}
};
template <>
struct ConvolutionFwdInstances<int8_t, int8_t, int8_t>
{
template <int NumDimSpatial,
typename std::enable_if<NumDimSpatial >= 1 && NumDimSpatial <= 3, bool>::type = false>
static std::vector<DeviceConvFwdNoOpPtr> Get()
{
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if constexpr(NumDimSpatial == 1)
{
ck::tensor_operation::device::instance::
add_device_conv1d_fwd_xdl_nwc_kxc_nwk_int8_instances(conv_ptrs);
}
else if constexpr(NumDimSpatial == 2)
{
ck::tensor_operation::device::instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
else if constexpr(NumDimSpatial == 3)
{
ck::tensor_operation::device::instance::
add_device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk_int8_instances(conv_ptrs);
}
return conv_ptrs;
}
};
template <typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InLayout = ck::tensor_layout::convolution::NHWC,
typename WeiLayout = ck::tensor_layout::convolution::KYXC,
typename OutLayout = ck::tensor_layout::convolution::NHWK,
typename InElementwiseOp = ck::tensor_operation::element_wise::PassThrough,
typename WeiElementwiseOp = ck::tensor_operation::element_wise::PassThrough,
typename OutElementwiseOp = ck::tensor_operation::element_wise::PassThrough,
typename InputInitFun = FillUniformDistribution<InDataType>,
typename WeightsInitFun = FillUniformDistribution<WeiDataType>>
class ConvFwdOpInstance : public ck::utils::OpInstance<OutDataType, InDataType, WeiDataType>
{
using DeviceConvFwdOp = tensor_operation::device::
DeviceConvFwd<InElementwiseOp, WeiElementwiseOp, OutElementwiseOp>;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
using DeviceBuffers = std::vector<DeviceMemPtr>;
using BaseType = ck::utils::OpInstance<OutDataType, InDataType, WeiDataType>;
template <typename T>
using TensorPtr = std::unique_ptr<Tensor<T>>;
using InTensorsTuple = std::tuple<TensorPtr<InDataType>, TensorPtr<WeiDataType>>;
public:
ConvFwdOpInstance() = delete;
ConvFwdOpInstance(const ConvFwdOpInstance&) = default;
ConvFwdOpInstance& operator=(const ConvFwdOpInstance&) = default;
ConvFwdOpInstance(const ConvParams& params,
bool do_init = true,
const InputInitFun& input_init_f = InputInitFun(),
const WeightsInitFun& weights_init_f = WeightsInitFun())
: BaseType(),
params_{params},
output_spatial_lengths_{params.GetOutputSpatialLengths()},
do_init_{do_init},
input_init_f_{input_init_f},
weights_init_f_{weights_init_f}
{
}
virtual ~ConvFwdOpInstance() override{};
virtual InTensorsTuple GetInputTensors() const override
{
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params_.N_),
static_cast<std::size_t>(params_.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params_.input_spatial_lengths_),
std::end(params_.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params_.K_),
static_cast<std::size_t>(params_.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params_.filter_spatial_lengths_),
std::end(params_.filter_spatial_lengths_));
auto input = std::make_unique<Tensor<InDataType>>(
get_host_tensor_descriptor(input_dims, InLayout{}));
auto weights = std::make_unique<Tensor<WeiDataType>>(
get_host_tensor_descriptor(filter_dims, WeiLayout{}));
if(do_init_)
{
input_init_f_(input->begin(), input->end());
weights_init_f_(weights->begin(), weights->end());
}
return std::make_tuple(std::move(input), std::move(weights));
}
virtual TensorPtr<OutDataType> GetOutputTensor() const override
{
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params_.N_),
static_cast<std::size_t>(params_.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths_),
std::end(output_spatial_lengths_));
auto output = std::make_unique<Tensor<OutDataType>>(
get_host_tensor_descriptor(output_dims, OutLayout{}));
if(do_init_)
{
std::fill(output->begin(), output->end(), OutDataType(0.f));
}
return output;
}
virtual std::unique_ptr<tensor_operation::device::BaseInvoker>
MakeInvokerPointer(tensor_operation::device::BaseOperator* op_ptr) const override
{
static_assert(
std::is_same_v<InElementwiseOp, ck::tensor_operation::element_wise::PassThrough>);
static_assert(
std::is_same_v<OutElementwiseOp, ck::tensor_operation::element_wise::PassThrough>);
static_assert(
std::is_same_v<WeiElementwiseOp, ck::tensor_operation::element_wise::PassThrough>);
auto conv_ptr = dynamic_cast<DeviceConvFwdOp*>(op_ptr);
if(!conv_ptr)
{
throw std::runtime_error(
"[ConvFwdOpInstance]: couldn't cast op_ptr to DeviceConvFwdNoOpPtr type!");
}
return conv_ptr->MakeInvokerPointer();
}
virtual std::unique_ptr<tensor_operation::device::BaseArgument>
MakeArgumentPointer(tensor_operation::device::BaseOperator* op_ptr,
const DeviceBuffers& in_device_buffers,
const DeviceMemPtr& out_device_buffer) const override
{
static_assert(
std::is_same_v<InElementwiseOp, ck::tensor_operation::element_wise::PassThrough>);
static_assert(
std::is_same_v<OutElementwiseOp, ck::tensor_operation::element_wise::PassThrough>);
static_assert(
std::is_same_v<WeiElementwiseOp, ck::tensor_operation::element_wise::PassThrough>);
auto conv_ptr = dynamic_cast<DeviceConvFwdOp*>(op_ptr);
if(!conv_ptr)
{
throw std::runtime_error(
"[ConvFwdOpInstance]: couldn't cast op_ptr to DeviceConvFwdNoOpPtr type!");
}
return conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buffers[0]->GetDeviceBuffer()),
static_cast<WeiDataType*>(in_device_buffers[1]->GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buffer->GetDeviceBuffer()),
params_.N_,
params_.K_,
params_.C_,
params_.input_spatial_lengths_,
params_.filter_spatial_lengths_,
output_spatial_lengths_,
params_.conv_filter_strides_,
params_.conv_filter_dilations_,
params_.input_left_pads_,
params_.input_right_pads_,
InElementwiseOp{},
WeiElementwiseOp{},
OutElementwiseOp{});
}
virtual std::size_t GetFlops() const override
{
return get_flops(params_.N_,
params_.C_,
params_.K_,
params_.filter_spatial_lengths_,
output_spatial_lengths_);
}
virtual std::size_t GetBtype() const override
{
return get_btype<InDataType, WeiDataType, OutDataType>(params_.N_,
params_.C_,
params_.K_,
params_.input_spatial_lengths_,
params_.filter_spatial_lengths_,
output_spatial_lengths_);
}
private:
const ConvParams& params_;
const std::vector<ck::index_t> output_spatial_lengths_;
const bool do_init_;
InputInitFun input_init_f_;
WeightsInitFun weights_init_f_;
};
} // namespace conv
} // namespace utils
} // namespace ck
std::ostream& operator<<(std::ostream& os, const ck::utils::conv::ConvParams& p);
library/include/ck/library/utility/convolution_parameter.hpp 0 → 100644
View file @ 92a0945d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <numeric>
#include <iterator>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/utility/io.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
struct ConvParams
{
ConvParams();
ConvParams(ck::index_t n_dim,
ck::index_t n_batch,
ck::index_t n_out_channels,
ck::index_t n_in_channels,
const std::vector<ck::index_t>& filters_len,
const std::vector<ck::index_t>& input_len,
const std::vector<ck::index_t>& strides,
const std::vector<ck::index_t>& dilations,
const std::vector<ck::index_t>& left_pads,
const std::vector<ck::index_t>& right_pads);
ck::index_t num_dim_spatial_;
ck::index_t N_;
ck::index_t K_;
ck::index_t C_;
std::vector<ck::index_t> filter_spatial_lengths_;
std::vector<ck::index_t> input_spatial_lengths_;
std::vector<ck::index_t> output_spatial_lengths_;
std::vector<ck::index_t> conv_filter_strides_;
std::vector<ck::index_t> conv_filter_dilations_;
std::vector<ck::index_t> input_left_pads_;
std::vector<ck::index_t> input_right_pads_;
std::vector<ck::index_t> GetOutputSpatialLengths() const;
std::size_t GetFlops() const;
template <typename InDataType, typename WeiDataType, typename OutDataType>
std::size_t GetByte() const
{
// sizeof(InDataType) * (N * C * <input spatial lengths product>) +
// sizeof(WeiDataType) * (K * C * <filter spatial lengths product>) +
// sizeof(OutDataType) * (N * K * <output spatial lengths product>);
return sizeof(InDataType) * (N_ * C_ *
std::accumulate(std::begin(input_spatial_lengths_),
std::end(input_spatial_lengths_),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>())) +
sizeof(WeiDataType) * (K_ * C_ *
std::accumulate(std::begin(filter_spatial_lengths_),
std::end(filter_spatial_lengths_),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>())) +
sizeof(OutDataType) * (N_ * K_ *
std::accumulate(std::begin(output_spatial_lengths_),
std::end(output_spatial_lengths_),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>()));
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
std::ostream& operator<<(std::ostream& os, const ck::tensor_operation::device::ConvParams& p);
library/include/ck/library/host_tensor/host_tensor.hpp → library/include/ck/library/utility/host_tensor.hpp
View file @ 92a0945d
...@@ -122,6 +122,22 @@ struct HostTensorDescriptor ...@@ -122,6 +122,22 @@ struct HostTensorDescriptor
std::vector<std::size_t> mStrides; std::vector<std::size_t> mStrides;
}; };
template <typename New2Old>
HostTensorDescriptor transpose_host_tensor_descriptor_given_new2old(const HostTensorDescriptor& a,
const New2Old& new2old)
{
std::vector<std::size_t> new_lengths(a.GetNumOfDimension());
std::vector<std::size_t> new_strides(a.GetNumOfDimension());
for(std::size_t i = 0; i < a.GetNumOfDimension(); i++)
{
new_lengths[i] = a.GetLengths()[new2old[i]];
new_strides[i] = a.GetStrides()[new2old[i]];
}
return HostTensorDescriptor(new_lengths, new_strides);
}
struct joinable_thread : std::thread struct joinable_thread : std::thread
{ {
template <typename... Xs> template <typename... Xs>
......
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