Skip to content

GitLab

Commit 0be1cf14 authored by Chao Liu's avatar Chao Liu
Browse files

update conv bwd weight

parent b054669b
Hide whitespace changes
Inline Side-by-side
Showing with 1520 additions and 1637 deletions
example/09_convnd_fwd/convnd_fwd_common.hpp
View file @ 0be1cf14
...@@ -8,7 +8,6 @@ ...@@ -8,7 +8,6 @@
#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_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"
...@@ -18,7 +17,24 @@ ...@@ -18,7 +17,24 @@
#include "ck/library/utility/convolution_parameter.hpp" #include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
void print_helper_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< "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"
<< " <in_n_hi_wi_c 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::tensor_operation::device::ConvParams ck::tensor_operation::device::ConvParams
parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[]) parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[])
{ {
const ck::index_t N = std::stoi(argv[arg_idx++]); const ck::index_t N = std::stoi(argv[arg_idx++]);
...@@ -74,23 +90,6 @@ parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[]) ...@@ -74,23 +90,6 @@ parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[])
input_right_pads}; input_right_pads};
} }
void print_helper_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< "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;
}
template <ck::index_t NDimSpatial, template <ck::index_t NDimSpatial,
typename InDataType, typename InDataType,
typename WeiDataType, typename WeiDataType,
...@@ -99,12 +98,14 @@ template <ck::index_t NDimSpatial, ...@@ -99,12 +98,14 @@ template <ck::index_t NDimSpatial,
typename InElementOp, typename InElementOp,
typename WeiElementOp, typename WeiElementOp,
typename OutElementOp, typename OutElementOp,
typename DeviceConvNDFwdInstance, typename DeviceConvNDFwdInstance>
typename ReferenceConvNDFwdInstance> int run_conv_fwd_nhwc(bool do_verification,
int run_conv_fwd_nhwc(const ck::tensor_operation::device::ConvParams& params,
bool do_verification,
int init_method, int init_method,
bool time_kernel) bool time_kernel,
const ck::tensor_operation::device::ConvParams& params,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{ {
auto f_nhwc_host_tensor_descriptor = auto f_nhwc_host_tensor_descriptor =
[](ck::index_t n, ck::index_t c, std::vector<ck::index_t> spatial_lengths) { [](ck::index_t n, ck::index_t c, std::vector<ck::index_t> spatial_lengths) {
...@@ -116,37 +117,37 @@ int run_conv_fwd_nhwc(const ck::tensor_operation::device::ConvParams& params, ...@@ -116,37 +117,37 @@ int run_conv_fwd_nhwc(const ck::tensor_operation::device::ConvParams& params,
return HostTensorDescriptor(nhwc_lengths); return HostTensorDescriptor(nhwc_lengths);
}; };
Tensor<InDataType> input( Tensor<InDataType> in_n_hi_wi_c(
f_nhwc_host_tensor_descriptor(params.N_, params.C_, params.input_spatial_lengths_)); f_nhwc_host_tensor_descriptor(params.N_, params.C_, params.input_spatial_lengths_));
Tensor<WeiDataType> weight( Tensor<WeiDataType> wei_k_y_x_c(
f_nhwc_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_)); f_nhwc_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_));
Tensor<OutDataType> host_output( Tensor<OutDataType> out_n_ho_wo_k_host(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths())); f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
Tensor<OutDataType> device_output( Tensor<OutDataType> out_n_ho_wo_k_device(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths())); f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
std::cout << "input: " << input.mDesc << std::endl; std::cout << "in_n_hi_wi_c: " << in_n_hi_wi_c.mDesc << std::endl;
std::cout << "weight: " << weight.mDesc << std::endl; std::cout << "wei_k_y_x_c: " << wei_k_y_x_c.mDesc << std::endl;
std::cout << "output: " << host_output.mDesc << std::endl; std::cout << "output: " << out_n_ho_wo_k_host.mDesc << std::endl;
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
case 1: case 1:
input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); in_n_hi_wi_c.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5}); wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break; break;
default: default:
input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0}); in_n_hi_wi_c.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5}); wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
} }
DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpace()); DeviceMem in_device_buf(sizeof(InDataType) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpace()); DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpace()); DeviceMem out_device_buf(sizeof(OutDataType) * out_n_ho_wo_k_device.mDesc.GetElementSpace());
in_device_buf.ToDevice(input.mData.data()); in_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_device_buf.ToDevice(weight.mData.data()); wei_device_buf.ToDevice(wei_k_y_x_c.mData.data());
// do GEMM // do GEMM
auto conv = DeviceConvNDFwdInstance{}; auto conv = DeviceConvNDFwdInstance{};
...@@ -164,9 +165,9 @@ int run_conv_fwd_nhwc(const ck::tensor_operation::device::ConvParams& params, ...@@ -164,9 +165,9 @@ int run_conv_fwd_nhwc(const ck::tensor_operation::device::ConvParams& params,
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{}, in_element_op,
WeiElementOp{}, wei_element_op,
OutElementOp{}); out_element_op);
if(!conv.IsSupportedArgument(argument)) if(!conv.IsSupportedArgument(argument))
{ {
...@@ -175,38 +176,48 @@ int run_conv_fwd_nhwc(const ck::tensor_operation::device::ConvParams& params, ...@@ -175,38 +176,48 @@ int run_conv_fwd_nhwc(const ck::tensor_operation::device::ConvParams& params,
"not support this Conv problem"); "not support this Conv problem");
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = params.GetFlops(); std::size_t flop = params.GetFlops();
std::size_t num_btype = params.GetByte<InDataType, WeiDataType, OutDataType>(); std::size_t num_btype = params.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / ave_time; float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< conv.GetTypeString() << std::endl; << conv.GetTypeString() << std::endl;
if(do_verification) if(do_verification)
{ {
auto ref_conv = ReferenceConvNDFwdInstance(); auto ref_conv =
ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>();
auto ref_invoker = ref_conv.MakeInvoker(); auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(input, auto ref_argument = ref_conv.MakeArgument(in_n_hi_wi_c,
weight, wei_k_y_x_c,
host_output, out_n_ho_wo_k_host,
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{}, in_element_op,
WeiElementOp{}, wei_element_op,
OutElementOp{}); out_element_op);
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(device_output.mData.data()); out_device_buf.FromDevice(out_n_ho_wo_k_device.mData.data());
return ck::utils::check_err(host_output.mData, return ck::utils::check_err(out_n_ho_wo_k_host.mData,
device_output.mData, out_n_ho_wo_k_device.mData,
"Error: incorrect results!", "Error: incorrect results!",
1e-5f, 1e-5f,
1e-4f) 1e-4f)
......
example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
View file @ 0be1cf14
...@@ -3,6 +3,8 @@ ...@@ -3,6 +3,8 @@
#include "convnd_fwd_common.hpp" #include "convnd_fwd_common.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_fwd_nwc_kxc_nwk_xdl.hpp"
using InDataType = ck::bhalf_t; using InDataType = ck::bhalf_t;
using WeiDataType = ck::bhalf_t; using WeiDataType = ck::bhalf_t;
using OutDataType = ck::bhalf_t; using OutDataType = ck::bhalf_t;
...@@ -55,15 +57,6 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwc ...@@ -55,15 +57,6 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwc
7, // CThreadTransferSrcDstVectorDim 7, // CThreadTransferSrcDstVectorDim
1>; // CThreadTransferDstScalarPerVector 1>; // CThreadTransferDstScalarPerVector
template <ck::index_t NumDimSpatial>
using ReferenceConvNDFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NumDimSpatial>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
print_helper_msg(); print_helper_msg();
...@@ -73,7 +66,8 @@ int main(int argc, char* argv[]) ...@@ -73,7 +66,8 @@ int main(int argc, char* argv[])
bool time_kernel = false; bool time_kernel = false;
int num_dim_spatial = 2; int num_dim_spatial = 2;
ck::tensor_operation::device::ConvParams params; ck::tensor_operation::device::ConvParams params{
2, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1) if(argc == 1)
{ {
...@@ -92,50 +86,66 @@ int main(int argc, char* argv[]) ...@@ -92,50 +86,66 @@ int main(int argc, char* argv[])
time_kernel = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]); num_dim_spatial = std::stoi(argv[4]);
params = parse_conv_params(num_dim_spatial, argc, argv); params = parse_conv_params(num_dim_spatial, 5, argv);
} }
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(num_dim_spatial == 1) if(num_dim_spatial == 1)
{ {
return run_conv_fwd<1, return run_conv_fwd_nhwc<1,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<1>, DeviceConvNDFwdInstance<1>>(do_verification,
ReferenceConvNDFwdInstance<1>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
else if(num_dim_spatial == 2) else if(num_dim_spatial == 2)
{ {
return run_conv_fwd<2, return run_conv_fwd_nhwc<2,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<2>, DeviceConvNDFwdInstance<2>>(do_verification,
ReferenceConvNDFwdInstance<2>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
else if(num_dim_spatial == 3) else if(num_dim_spatial == 3)
{ {
return run_conv_fwd<3, return run_conv_fwd_nhwc<3,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<3>, DeviceConvNDFwdInstance<3>>(do_verification,
ReferenceConvNDFwdInstance<3>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
return 0; return 0;
......
example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
View file @ 0be1cf14
...@@ -3,6 +3,8 @@ ...@@ -3,6 +3,8 @@
#include "convnd_fwd_common.hpp" #include "convnd_fwd_common.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_fwd_nwc_kxc_nwk_xdl.hpp"
using InDataType = ck::half_t; using InDataType = ck::half_t;
using WeiDataType = ck::half_t; using WeiDataType = ck::half_t;
using OutDataType = ck::half_t; using OutDataType = ck::half_t;
...@@ -18,15 +20,15 @@ using OutElementOp = ck::tensor_operation::element_wise::PassThrough; ...@@ -18,15 +20,15 @@ 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;
template <ck::index_t NumDimSpatial> template <ck::index_t NDimSpatial>
using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwcKxcNwk_Xdl< using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwcKxcNwk_Xdl<
NumDimSpatial, // NumDimSpatial NDimSpatial, //
InDataType, // InDataType, //
WeiDataType, // WeiDataType, //
OutDataType, // OutDataType, //
AccDataType, // AccDataType, //
InElementOp, // Input Elementwise Operation InElementOp, // Input Elementwise Operation
WeiElementOp, // Weights Elementwise Operation = WeiElementOp, // Weights Elementwise Operation
OutElementOp, // Output Elementwise Operation OutElementOp, // Output Elementwise Operation
ConvFwdDefault, // ConvForwardSpecialization ConvFwdDefault, // ConvForwardSpecialization
256, // BlockSize 256, // BlockSize
...@@ -55,19 +57,6 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwc ...@@ -55,19 +57,6 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwc
7, // CThreadTransferSrcDstVectorDim 7, // CThreadTransferSrcDstVectorDim
1>; // CThreadTransferDstScalarPerVector 1>; // CThreadTransferDstScalarPerVector
template <ck::index_t NumDimSpatial>
using ReferenceConvNDFwdInstance =
ck::tensor_operation::host::ReferenceConvFwd<NumDimSpatial,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
print_helper_msg(); print_helper_msg();
...@@ -100,6 +89,10 @@ int main(int argc, char* argv[]) ...@@ -100,6 +89,10 @@ int main(int argc, char* argv[])
params = parse_conv_params(num_dim_spatial, 5, argv); params = parse_conv_params(num_dim_spatial, 5, argv);
} }
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(num_dim_spatial == 1) if(num_dim_spatial == 1)
{ {
return run_conv_fwd_nhwc<1, return run_conv_fwd_nhwc<1,
...@@ -110,9 +103,13 @@ int main(int argc, char* argv[]) ...@@ -110,9 +103,13 @@ int main(int argc, char* argv[])
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<1>, DeviceConvNDFwdInstance<1>>(do_verification,
ReferenceConvNDFwdInstance<1>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
else if(num_dim_spatial == 2) else if(num_dim_spatial == 2)
{ {
...@@ -124,9 +121,13 @@ int main(int argc, char* argv[]) ...@@ -124,9 +121,13 @@ int main(int argc, char* argv[])
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<2>, DeviceConvNDFwdInstance<2>>(do_verification,
ReferenceConvNDFwdInstance<2>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
else if(num_dim_spatial == 3) else if(num_dim_spatial == 3)
{ {
...@@ -138,9 +139,13 @@ int main(int argc, char* argv[]) ...@@ -138,9 +139,13 @@ int main(int argc, char* argv[])
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<3>, DeviceConvNDFwdInstance<3>>(do_verification,
ReferenceConvNDFwdInstance<3>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
return 0; return 0;
......
example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
View file @ 0be1cf14
...@@ -3,6 +3,8 @@ ...@@ -3,6 +3,8 @@
#include "convnd_fwd_common.hpp" #include "convnd_fwd_common.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_fwd_nwc_kxc_nwk_xdl.hpp"
using InDataType = float; using InDataType = float;
using WeiDataType = float; using WeiDataType = float;
using OutDataType = float; using OutDataType = float;
...@@ -55,15 +57,6 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwc ...@@ -55,15 +57,6 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwc
7, // CThreadTransferSrcDstVectorDim 7, // CThreadTransferSrcDstVectorDim
1>; // CThreadTransferDstScalarPerVector 1>; // CThreadTransferDstScalarPerVector
template <ck::index_t NumDimSpatial>
using ReferenceConvNDFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NumDimSpatial>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
print_helper_msg(); print_helper_msg();
...@@ -73,7 +66,8 @@ int main(int argc, char* argv[]) ...@@ -73,7 +66,8 @@ int main(int argc, char* argv[])
bool time_kernel = false; bool time_kernel = false;
int num_dim_spatial = 2; int num_dim_spatial = 2;
ck::tensor_operation::device::ConvParams params; ck::tensor_operation::device::ConvParams params{
2, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1) if(argc == 1)
{ {
...@@ -92,50 +86,66 @@ int main(int argc, char* argv[]) ...@@ -92,50 +86,66 @@ int main(int argc, char* argv[])
time_kernel = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]); num_dim_spatial = std::stoi(argv[4]);
params = parse_conv_params(num_dim_spatial, argc, argv); params = parse_conv_params(num_dim_spatial, 5, argv);
} }
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(num_dim_spatial == 1) if(num_dim_spatial == 1)
{ {
return run_conv_fwd<1, return run_conv_fwd_nhwc<1,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<1>, DeviceConvNDFwdInstance<1>>(do_verification,
ReferenceConvNDFwdInstance<1>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
else if(num_dim_spatial == 2) else if(num_dim_spatial == 2)
{ {
return run_conv_fwd<2, return run_conv_fwd_nhwc<2,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<2>, DeviceConvNDFwdInstance<2>>(do_verification,
ReferenceConvNDFwdInstance<2>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
else if(num_dim_spatial == 3) else if(num_dim_spatial == 3)
{ {
return run_conv_fwd<3, return run_conv_fwd_nhwc<3,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<3>, DeviceConvNDFwdInstance<3>>(do_verification,
ReferenceConvNDFwdInstance<3>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
return 0; return 0;
......
example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
View file @ 0be1cf14
...@@ -3,6 +3,8 @@ ...@@ -3,6 +3,8 @@
#include "convnd_fwd_common.hpp" #include "convnd_fwd_common.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_fwd_nwc_kxc_nwk_xdl.hpp"
using InDataType = double; using InDataType = double;
using WeiDataType = double; using WeiDataType = double;
using OutDataType = double; using OutDataType = double;
...@@ -55,15 +57,6 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwc ...@@ -55,15 +57,6 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwc
7, // CThreadTransferSrcDstVectorDim 7, // CThreadTransferSrcDstVectorDim
1>; // CThreadTransferDstScalarPerVector 1>; // CThreadTransferDstScalarPerVector
template <ck::index_t NumDimSpatial>
using ReferenceConvNDFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NumDimSpatial>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
print_helper_msg(); print_helper_msg();
...@@ -73,7 +66,8 @@ int main(int argc, char* argv[]) ...@@ -73,7 +66,8 @@ int main(int argc, char* argv[])
bool time_kernel = false; bool time_kernel = false;
int num_dim_spatial = 2; int num_dim_spatial = 2;
ck::tensor_operation::device::ConvParams params; ck::tensor_operation::device::ConvParams params{
2, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1) if(argc == 1)
{ {
...@@ -92,50 +86,66 @@ int main(int argc, char* argv[]) ...@@ -92,50 +86,66 @@ int main(int argc, char* argv[])
time_kernel = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]); num_dim_spatial = std::stoi(argv[4]);
params = parse_conv_params(num_dim_spatial, argc, argv); params = parse_conv_params(num_dim_spatial, 5, argv);
} }
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(num_dim_spatial == 1) if(num_dim_spatial == 1)
{ {
return run_conv_fwd<1, return run_conv_fwd_nhwc<1,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<1>, DeviceConvNDFwdInstance<1>>(do_verification,
ReferenceConvNDFwdInstance<1>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
else if(num_dim_spatial == 2) else if(num_dim_spatial == 2)
{ {
return run_conv_fwd<2, return run_conv_fwd_nhwc<2,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<2>, DeviceConvNDFwdInstance<2>>(do_verification,
ReferenceConvNDFwdInstance<2>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
else if(num_dim_spatial == 3) else if(num_dim_spatial == 3)
{ {
return run_conv_fwd<3, return run_conv_fwd_nhwc<3,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<3>, DeviceConvNDFwdInstance<3>>(do_verification,
ReferenceConvNDFwdInstance<3>>( init_method,
params, do_verification, init_method, time_kernel); time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
} }
return 0; return 0;
......
example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
View file @ 0be1cf14
...@@ -3,6 +3,8 @@ ...@@ -3,6 +3,8 @@
#include "convnd_fwd_common.hpp" #include "convnd_fwd_common.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_fwd_nwc_kxc_nwk_xdl.hpp"
using InDataType = int8_t; using InDataType = int8_t;
using WeiDataType = int8_t; using WeiDataType = int8_t;
using OutDataType = int8_t; using OutDataType = int8_t;
...@@ -55,15 +57,6 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwc ...@@ -55,15 +57,6 @@ using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvNdFwdNwc
7, // CThreadTransferSrcDstVectorDim 7, // CThreadTransferSrcDstVectorDim
1>; // CThreadTransferDstScalarPerVector 1>; // CThreadTransferDstScalarPerVector
template <ck::index_t NumDimSpatial>
using ReferenceConvNDFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NumDimSpatial>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
print_helper_msg(); print_helper_msg();
...@@ -73,7 +66,8 @@ int main(int argc, char* argv[]) ...@@ -73,7 +66,8 @@ int main(int argc, char* argv[])
bool time_kernel = false; bool time_kernel = false;
int num_dim_spatial = 2; int num_dim_spatial = 2;
ck::tensor_operation::device::ConvParams params; ck::tensor_operation::device::ConvParams params{
2, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1) if(argc == 1)
{ {
...@@ -92,49 +86,49 @@ int main(int argc, char* argv[]) ...@@ -92,49 +86,49 @@ int main(int argc, char* argv[])
time_kernel = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]); num_dim_spatial = std::stoi(argv[4]);
params = parse_conv_params(num_dim_spatial, argc, argv); params = parse_conv_params(num_dim_spatial, 5, argv);
} }
if(num_dim_spatial == 1) if(num_dim_spatial == 1)
{ {
return run_conv_fwd<1, return run_conv_fwd_nhwc<1,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<1>, DeviceConvNDFwdInstance<1>,
ReferenceConvNDFwdInstance<1>>( ReferenceConvNDFwdInstance<1>>(
params, do_verification, init_method, time_kernel); params, do_verification, init_method, time_kernel);
} }
else if(num_dim_spatial == 2) else if(num_dim_spatial == 2)
{ {
return run_conv_fwd<2, return run_conv_fwd_nhwc<2,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<2>, DeviceConvNDFwdInstance<2>,
ReferenceConvNDFwdInstance<2>>( ReferenceConvNDFwdInstance<2>>(
params, do_verification, init_method, time_kernel); params, do_verification, init_method, time_kernel);
} }
else if(num_dim_spatial == 3) else if(num_dim_spatial == 3)
{ {
return run_conv_fwd<3, return run_conv_fwd_nhwc<3,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
AccDataType, AccDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<3>, DeviceConvNDFwdInstance<3>,
ReferenceConvNDFwdInstance<3>>( ReferenceConvNDFwdInstance<3>>(
params, do_verification, init_method, time_kernel); params, do_verification, init_method, time_kernel);
} }
......
example/10_conv2d_bwd_data/CMakeLists.txt
View file @ 0be1cf14
add_example_executable(example_conv2d_bwd_data_xdl conv2d_bwd_data_xdl.cpp) add_example_executable(example_conv2d_bwd_data_xdl conv2d_bwd_data_xdl.cpp)
target_link_libraries(example_conv2d_bwd_data_xdl PRIVATE conv_util) target_link_libraries(example_conv2d_bwd_data_xdl PRIVATE utility)
example/10_conv2d_bwd_data/conv2d_bwd_data_xdl.cpp
View file @ 0be1cf14
...@@ -29,6 +29,7 @@ using S = ck::Sequence<Is...>; ...@@ -29,6 +29,7 @@ using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough; using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough; using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdDefault = static constexpr auto ConvBwdDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default; ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
...@@ -68,125 +69,168 @@ using DeviceConvBwdDataInstance = ck::tensor_operation::device:: ...@@ -68,125 +69,168 @@ using DeviceConvBwdDataInstance = ck::tensor_operation::device::
7, 7,
1>; // GemmCThreadTransferDstScalarPerVector 1>; // GemmCThreadTransferDstScalarPerVector
using ReferenceConvBwdInstance = ck::tensor_operation::host::ReferenceConvBwdData<InDataType, using ReferenceConvBwdInstance =
WeiDataType, ck::tensor_operation::host::ReferenceConvBwdData<2,
OutDataType, ck::tensor_layout::convolution::NHWC,
AccDataType, ck::tensor_layout::convolution::KYXC,
InElementOp, ck::tensor_layout::convolution::NHWK,
WeiElementOp, InDataType,
OutElementOp>; WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
ck::tensor_operation::device::ConvParams
parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[])
{
const ck::index_t N = std::stoi(argv[arg_idx++]);
const ck::index_t K = std::stoi(argv[arg_idx++]);
const ck::index_t C = std::stoi(argv[arg_idx++]);
std::vector<ck::index_t> filter_spatial_lengths(num_dim_spatial);
std::vector<ck::index_t> input_spatial_lengths(num_dim_spatial);
std::vector<ck::index_t> conv_filter_strides(num_dim_spatial);
std::vector<ck::index_t> conv_filter_dilations(num_dim_spatial);
std::vector<ck::index_t> input_left_pads(num_dim_spatial);
std::vector<ck::index_t> input_right_pads(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_left_pads[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_right_pads[i] = std::stoi(argv[arg_idx++]);
}
return ck::tensor_operation::device::ConvParams{num_dim_spatial,
N,
K,
C,
filter_spatial_lengths,
input_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads};
}
void print_helper_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< "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;
}
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
print_helper_msg();
bool do_verification = true; bool do_verification = true;
int init_method = 1; int init_method = 1;
bool time_kernel = false; bool time_kernel = false;
int num_dim_spatial = 2;
// Conv shape ck::tensor_operation::device::ConvParams params{
ck::index_t N = 128; 2, 128, 256, 256, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
ck::index_t K = 256;
ck::index_t C = 256; if(argc == 1)
ck::index_t Y = 3; {
ck::index_t X = 3; // use default
ck::index_t Hi = 71; }
ck::index_t Wi = 71; else if(argc == 4)
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
if(argc == 4)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
else if(argc == 19) else
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
N = std::stoi(argv[4]); params = parse_conv_params(num_dim_spatial, 5, argv);
K = std::stoi(argv[5]);
C = std::stoi(argv[6]);
Y = std::stoi(argv[7]);
X = std::stoi(argv[8]);
Hi = std::stoi(argv[9]);
Wi = std::stoi(argv[10]);
conv_stride_h = std::stoi(argv[11]);
conv_stride_w = std::stoi(argv[12]);
conv_dilation_h = std::stoi(argv[13]);
conv_dilation_w = std::stoi(argv[14]);
in_left_pad_h = std::stoi(argv[15]);
in_left_pad_w = std::stoi(argv[16]);
in_right_pad_h = std::stoi(argv[17]);
in_right_pad_w = std::stoi(argv[18]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
} }
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1; auto f_nhwc_host_tensor_descriptor =
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1; [](ck::index_t n, ck::index_t c, std::vector<ck::index_t> spatial_lengths) {
std::vector<std::size_t> nhwc_lengths{static_cast<std::size_t>(n),
static_cast<std::size_t>(c)};
nhwc_lengths.insert(
nhwc_lengths.begin() + 1, spatial_lengths.begin(), spatial_lengths.end());
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1; return HostTensorDescriptor(nhwc_lengths);
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
// tensor layout
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W) {
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}; };
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo)); Tensor<InDataType> in_n_hi_wi_c_host_result(
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X)); f_nhwc_host_tensor_descriptor(params.N_, params.C_, params.input_spatial_lengths_));
Tensor<InDataType> in_n_c_hi_wi_host_result(f_host_tensor_descriptor(N, C, Hi, Wi)); Tensor<InDataType> in_n_hi_wi_c_device_result(
Tensor<InDataType> in_n_c_hi_wi_device_result(f_host_tensor_descriptor(N, C, Hi, Wi)); f_nhwc_host_tensor_descriptor(params.N_, params.C_, params.input_spatial_lengths_));
Tensor<WeiDataType> wei_k_y_x_c(
f_nhwc_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_));
Tensor<OutDataType> out_n_ho_wo_k(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi_host_result.mDesc << std::endl; std::cout << "in_n_hi_wi_c: " << in_n_hi_wi_c_host_result.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl; std::cout << "wei_k_y_x_c: " << wei_k_y_x_c.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl; std::cout << "out_n_ho_wo_k: " << out_n_ho_wo_k.mDesc << std::endl;
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
case 1: case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5}); out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5}); wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break; break;
default: default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1}); out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1}); wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
} }
DeviceMem in_device_buf(sizeof(InDataType) * DeviceMem in_device_buf(sizeof(InDataType) *
in_n_c_hi_wi_device_result.mDesc.GetElementSpace()); in_n_hi_wi_c_device_result.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace()); DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace()); DeviceMem out_device_buf(sizeof(OutDataType) * out_n_ho_wo_k.mDesc.GetElementSpace());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data()); out_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data()); wei_device_buf.ToDevice(wei_k_y_x_c.mData.data());
// reset input to zero // reset input to zero
in_n_c_hi_wi_device_result.GenerateTensorValue(GeneratorTensor_1<InDataType>{0}); in_device_buf.SetZero();
in_device_buf.ToDevice(in_n_c_hi_wi_device_result.mData.data());
// do GEMM // do GEMM
auto conv = DeviceConvBwdDataInstance{}; auto conv = DeviceConvBwdDataInstance{};
...@@ -194,16 +238,16 @@ int main(int argc, char* argv[]) ...@@ -194,16 +238,16 @@ int main(int argc, char* argv[])
auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()), auto argument = conv.MakeArgument(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()),
N, params.N_,
K, params.K_,
C, params.C_,
std::vector<ck::index_t>{{Hi, Wi}}, params.input_spatial_lengths_,
std::vector<ck::index_t>{{Y, X}}, params.filter_spatial_lengths_,
std::vector<ck::index_t>{{Ho, Wo}}, params.GetOutputSpatialLengths(),
conv_filter_strides, params.conv_filter_strides_,
conv_filter_dilations, params.conv_filter_dilations_,
input_left_pads, params.input_left_pads_,
input_right_pads, params.input_right_pads_,
InElementOp{}, InElementOp{},
WeiElementOp{}, WeiElementOp{},
OutElementOp{}); OutElementOp{});
...@@ -217,11 +261,8 @@ int main(int argc, char* argv[]) ...@@ -217,11 +261,8 @@ int main(int argc, char* argv[])
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X; std::size_t flop = params.GetFlops();
std::size_t num_btype = params.GetByte<InDataType, WeiDataType, OutDataType>();
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
...@@ -235,25 +276,26 @@ int main(int argc, char* argv[]) ...@@ -235,25 +276,26 @@ int main(int argc, char* argv[])
auto ref_conv = ReferenceConvBwdInstance{}; auto ref_conv = ReferenceConvBwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker(); auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result, auto ref_argument = ref_conv.MakeArgument(in_n_hi_wi_c_host_result,
wei_k_c_y_x, wei_k_y_x_c,
out_n_k_ho_wo, out_n_ho_wo_k,
conv_filter_strides, params.conv_filter_strides_,
conv_filter_dilations, params.conv_filter_dilations_,
input_left_pads, params.input_left_pads_,
input_right_pads, params.input_right_pads_,
InElementOp{}, InElementOp{},
WeiElementOp{}, WeiElementOp{},
OutElementOp{}); OutElementOp{});
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data()); in_device_buf.FromDevice(in_n_hi_wi_c_device_result.mData.data());
return ck::utils::check_err(in_n_c_hi_wi_device_result.mData, return ck::utils::check_err(in_n_hi_wi_c_device_result.mData,
in_n_c_hi_wi_host_result.mData) in_n_hi_wi_c_host_result.mData)
? 0 ? 0
: 1; : 1;
} }
return 0; return 0;
} }
example/11_conv2d_bwd_weight/CMakeLists.txt
View file @ 0be1cf14
add_example_executable(example_conv2d_bwd_weight_xdl conv2d_bwd_weight_xdl.cpp) add_example_executable(example_conv2d_bwd_weight_xdl conv2d_bwd_weight_xdl.cpp)
target_link_libraries(example_conv2d_bwd_weight_xdl PRIVATE conv_util) target_link_libraries(example_conv2d_bwd_weight_xdl PRIVATE utility)
example/11_conv2d_bwd_weight/conv2d_bwd_weight_xdl.cpp
View file @ 0be1cf14
...@@ -74,149 +74,171 @@ using DeviceConvBwdWeightInstance = ck::tensor_operation::device:: ...@@ -74,149 +74,171 @@ using DeviceConvBwdWeightInstance = ck::tensor_operation::device::
// clang-format on // clang-format on
using ReferenceConvBwdWeightInstance = using ReferenceConvBwdWeightInstance =
ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType, ck::tensor_operation::host::ReferenceConvBwdWeight<2,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp>; OutElementOp>;
void print_helper_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< "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"
<< "split_k" << std::endl;
}
ck::tensor_operation::device::ConvParams
parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[])
{
const ck::index_t N = std::stoi(argv[arg_idx++]);
const ck::index_t K = std::stoi(argv[arg_idx++]);
const ck::index_t C = std::stoi(argv[arg_idx++]);
std::vector<ck::index_t> filter_spatial_lengths(num_dim_spatial);
std::vector<ck::index_t> input_spatial_lengths(num_dim_spatial);
std::vector<ck::index_t> conv_filter_strides(num_dim_spatial);
std::vector<ck::index_t> conv_filter_dilations(num_dim_spatial);
std::vector<ck::index_t> input_left_pads(num_dim_spatial);
std::vector<ck::index_t> input_right_pads(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_left_pads[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_right_pads[i] = std::stoi(argv[arg_idx++]);
}
return ck::tensor_operation::device::ConvParams{num_dim_spatial,
N,
K,
C,
filter_spatial_lengths,
input_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads};
}
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
print_helper_msg();
bool do_verification = true; bool do_verification = true;
int init_method = 1; int init_method = 1;
bool time_kernel = false; bool time_kernel = false;
int do_log = 0; int num_dim_spatial = 2;
int split_k = 4;
ck::tensor_operation::device::ConvParams params{
// Conv shape 2, 32, 256, 1024, {3, 3}, {14, 14}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
ck::index_t N = 128;
ck::index_t K = 256; ck::index_t split_k = 4;
ck::index_t C = 1024;
ck::index_t Y = 3; if(argc == 1)
ck::index_t X = 3;
ck::index_t Hi = 14;
ck::index_t Wi = 14;
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 0;
ck::index_t in_left_pad_w = 0;
ck::index_t in_right_pad_h = 0;
ck::index_t in_right_pad_w = 0;
if(argc == 6)
{ {
do_verification = std::stoi(argv[1]); // use default
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
} }
else if(argc == 21) else if(argc == 4)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
N = std::stoi(argv[6]);
K = std::stoi(argv[7]);
C = std::stoi(argv[8]);
Y = std::stoi(argv[9]);
X = std::stoi(argv[10]);
Hi = std::stoi(argv[11]);
Wi = std::stoi(argv[12]);
conv_stride_h = std::stoi(argv[13]);
conv_stride_w = std::stoi(argv[14]);
conv_dilation_h = std::stoi(argv[15]);
conv_dilation_w = std::stoi(argv[16]);
in_left_pad_h = std::stoi(argv[17]);
in_left_pad_w = std::stoi(argv[18]);
in_right_pad_h = std::stoi(argv[19]);
in_right_pad_w = std::stoi(argv[20]);
} }
else else
{ {
printf("arg1: verification (0=no, 1=yes)\n"); do_verification = std::stoi(argv[1]);
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); init_method = std::stoi(argv[2]);
printf("arg3: time kernel (0=n0, 1=yes)\n"); time_kernel = std::stoi(argv[3]);
printf("arg4: is show log (0=no, 1=yes)\n"); num_dim_spatial = std::stoi(argv[4]);
printf("arg5: split-k \n");
printf("arg6 to 19: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, " params = parse_conv_params(num_dim_spatial, 5, argv);
"RightPx\n");
exit(0); split_k = std::stoi(argv[5 + 3 + 6 * num_dim_spatial - 1]);
split_k = std::max(1, split_k);
} }
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1; auto f_nhwc_host_tensor_descriptor =
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1; [](ck::index_t n, ck::index_t c, std::vector<ck::index_t> spatial_lengths) {
std::vector<std::size_t> nhwc_lengths{static_cast<std::size_t>(n),
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1; static_cast<std::size_t>(c)};
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1; nhwc_lengths.insert(
nhwc_lengths.begin() + 1, spatial_lengths.begin(), spatial_lengths.end());
const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}}; return HostTensorDescriptor(nhwc_lengths);
const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}}; };
const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
Tensor<InDataType> in_n_hi_wi_c(
// tensor layout f_nhwc_host_tensor_descriptor(params.N_, params.C_, params.input_spatial_lengths_));
auto f_host_tensor_descriptor = [](std::size_t N_, Tensor<WeiDataType> wei_k_y_x_c_host_result(
std::size_t C_, f_nhwc_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_));
std::size_t H, Tensor<WeiDataType> wei_k_y_x_c_device_result(
std::size_t W, f_nhwc_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_));
auto layout) { Tensor<OutDataType> out_n_ho_wo_k(
if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value || f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
ck::is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
ck::is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value) std::cout << "input: " << in_n_hi_wi_c.mDesc << std::endl;
{ std::cout << "weight: " << wei_k_y_x_c_host_result.mDesc << std::endl;
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}), std::cout << "output: " << out_n_ho_wo_k.mDesc << std::endl;
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWC>::value ||
ck::is_same<decltype(layout),
ck::tensor_layout::convolution::KYXC>::value ||
ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x_host_result(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<WeiDataType> wei_k_c_y_x_device_result(
f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
case 1: case 1:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); in_n_hi_wi_c.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5}); out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break; break;
default: default:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}); in_n_hi_wi_c.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1}); out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
} }
wei_k_c_y_x_device_result.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{0});
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace()); DeviceMem in_device_buf(sizeof(InDataType) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * DeviceMem wei_device_buf(sizeof(WeiDataType) *
wei_k_c_y_x_device_result.mDesc.GetElementSpace()); wei_k_y_x_c_device_result.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace()); DeviceMem out_device_buf(sizeof(OutDataType) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
out_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data()); // wei_device_buf.SetZero();
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x_device_result.mData.data());
// do GEMM // do GEMM
auto conv = DeviceConvBwdWeightInstance{}; auto conv = DeviceConvBwdWeightInstance{};
...@@ -224,16 +246,16 @@ int main(int argc, char* argv[]) ...@@ -224,16 +246,16 @@ int main(int argc, char* argv[])
auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()), auto argument = conv.MakeArgument(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()),
N, params.N_,
K, params.K_,
C, params.C_,
std::vector<ck::index_t>{{Hi, Wi}}, params.input_spatial_lengths_,
std::vector<ck::index_t>{{Y, X}}, params.filter_spatial_lengths_,
std::vector<ck::index_t>{{Ho, Wo}}, params.output_spatial_lengths_,
conv_filter_strides, params.conv_filter_strides_,
conv_filter_dilations, params.conv_filter_dilations_,
input_left_pads, params.input_left_pads_,
input_right_pads, params.input_right_pads_,
InElementOp{}, InElementOp{},
WeiElementOp{}, WeiElementOp{},
OutElementOp{}, OutElementOp{},
...@@ -247,19 +269,16 @@ int main(int argc, char* argv[]) ...@@ -247,19 +269,16 @@ int main(int argc, char* argv[])
return 1; return 1;
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) + std::size_t flop = params.GetFlops();
sizeof(WeiDataType) * (K * C * Y * X) + std::size_t num_btype = params.GetByte<InDataType, WeiDataType, OutDataType>();
sizeof(OutDataType) * (N * K * Ho * Wo);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / ave_time; float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s" std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl; << std::endl;
if(do_verification) if(do_verification)
...@@ -267,34 +286,25 @@ int main(int argc, char* argv[]) ...@@ -267,34 +286,25 @@ int main(int argc, char* argv[])
auto ref_conv = ReferenceConvBwdWeightInstance{}; auto ref_conv = ReferenceConvBwdWeightInstance{};
auto ref_invoker = ref_conv.MakeInvoker(); auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi, auto ref_argument = ref_conv.MakeArgument(in_n_hi_wi_c,
wei_k_c_y_x_host_result, wei_k_y_x_c_host_result,
out_n_k_ho_wo, out_n_ho_wo_k,
conv_filter_strides, params.conv_filter_strides_,
conv_filter_dilations, params.conv_filter_dilations_,
input_left_pads, params.input_left_pads_,
input_right_pads, params.input_right_pads_,
InElementOp{}, InElementOp{},
WeiElementOp{}, WeiElementOp{},
OutElementOp{}); OutElementOp{});
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data()); wei_device_buf.FromDevice(wei_k_y_x_c_device_result.mData.data());
if(do_log) return ck::utils::check_err(wei_k_y_x_c_device_result.mData, wei_k_y_x_c_host_result.mData)
{
LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "wei_device(after): ", wei_k_c_y_x_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "wei_host : ", wei_k_c_y_x_host_result.mData, ",")
<< std::endl;
}
return ck::utils::check_err(wei_k_c_y_x_device_result.mData, wei_k_c_y_x_host_result.mData)
? 0 ? 0
: 1; : 1;
} }
return 0; return 0;
} }
example/17_convnd_bwd_data_xdl/CMakeLists.txt
View file @ 0be1cf14
add_example_executable(example_convnd_bwd_data_xdl convnd_bwd_data_xdl.cpp) add_example_executable(example_convnd_bwd_data_xdl convnd_bwd_data_xdl.cpp)
target_link_libraries(example_convnd_bwd_data_xdl PRIVATE conv_util) target_link_libraries(example_convnd_bwd_data_xdl PRIVATE utility)
example/17_convnd_bwd_data_xdl/convnd_bwd_data_common.hpp 0 → 100644
View file @ 0be1cf14
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
ck::tensor_operation::device::ConvParams
parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[])
{
const ck::index_t N = std::stoi(argv[arg_idx++]);
const ck::index_t K = std::stoi(argv[arg_idx++]);
const ck::index_t C = std::stoi(argv[arg_idx++]);
std::vector<ck::index_t> filter_spatial_lengths(num_dim_spatial);
std::vector<ck::index_t> input_spatial_lengths(num_dim_spatial);
std::vector<ck::index_t> conv_filter_strides(num_dim_spatial);
std::vector<ck::index_t> conv_filter_dilations(num_dim_spatial);
std::vector<ck::index_t> input_left_pads(num_dim_spatial);
std::vector<ck::index_t> input_right_pads(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_left_pads[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_right_pads[i] = std::stoi(argv[arg_idx++]);
}
return ck::tensor_operation::device::ConvParams{num_dim_spatial,
N,
K,
C,
filter_spatial_lengths,
input_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads};
}
void print_helper_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< "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;
}
template <ck::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp,
typename DeviceConvNdBwdDataInstance>
int run_conv_bwd_data_nhwc(bool do_verification,
int init_method,
bool time_kernel,
const ck::tensor_operation::device::ConvParams& params,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{
auto f_nhwc_host_tensor_descriptor =
[](ck::index_t n, ck::index_t c, std::vector<ck::index_t> spatial_lengths) {
std::vector<std::size_t> nhwc_lengths{static_cast<std::size_t>(n),
static_cast<std::size_t>(c)};
nhwc_lengths.insert(
nhwc_lengths.begin() + 1, spatial_lengths.begin(), spatial_lengths.end());
return HostTensorDescriptor(nhwc_lengths);
};
Tensor<InDataType> in_n_hi_wi_c_host(
f_nhwc_host_tensor_descriptor(params.N_, params.C_, params.input_spatial_lengths_));
Tensor<InDataType> in_n_hi_wi_c_device(
f_nhwc_host_tensor_descriptor(params.N_, params.C_, params.input_spatial_lengths_));
Tensor<WeiDataType> wei_k_y_x_c(
f_nhwc_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_));
Tensor<OutDataType> out_n_ho_wo_k(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
std::cout << "in_n_hi_wi_c: " << in_n_hi_wi_c_host.mDesc << std::endl;
std::cout << "wei_k_y_x_c: " << wei_k_y_x_c.mDesc << std::endl;
std::cout << "out_n_ho_wo_k: " << out_n_ho_wo_k.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_hi_wi_c_device.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_ho_wo_k.mDesc.GetElementSpace());
out_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
wei_device_buf.ToDevice(wei_k_y_x_c.mData.data());
// reset input to zero
in_device_buf.SetZero();
// do GEMM
auto conv = DeviceConvNdBwdDataInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.GetOutputSpatialLengths(),
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
in_element_op,
wei_element_op,
out_element_op);
if(!conv.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = params.GetFlops();
std::size_t num_btype = params.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
if(do_verification)
{
auto ref_conv =
ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_hi_wi_c_host,
wei_k_y_x_c,
out_n_ho_wo_k,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
in_device_buf.FromDevice(in_n_hi_wi_c_device.mData.data());
return ck::utils::check_err(in_n_hi_wi_c_device.mData, in_n_hi_wi_c_host.mData) ? 0 : 1;
}
return 0;
}
example/17_convnd_bwd_data_xdl/convnd_bwd_data_xdl.cpp
View file @ 0be1cf14
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include "convnd_bwd_data_common.hpp"
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp" #include "ck/tensor_operation/gpu/device/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_bwd_data_xdl_ndhwc_kzyxc_ndhwk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
using InDataType = ck::half_t; using InDataType = ck::half_t;
using WeiDataType = ck::half_t; using WeiDataType = ck::half_t;
...@@ -29,324 +16,137 @@ using S = ck::Sequence<Is...>; ...@@ -29,324 +16,137 @@ using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough; using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough; using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdDefault = static constexpr auto ConvBwdDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default; ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
using DeviceConvBwdDataBasePtr = template <ck::index_t NDimSpatial>
ck::tensor_operation::device::DeviceConvBwdDataPtr<InElementOp, WeiElementOp, OutElementOp>; using DeviceConvNdBwdDataInstance = ck::tensor_operation::device::DeviceConvNdBwdDataNwcKxcNwk_Xdl<
NDimSpatial, // NDimSpatial
template <ck::index_t NumDimSpatial> InDataType, // InDataType
using DeviceConvNDBwdDataInstance = ck::tensor_operation::device:: WeiDataType, // WeiDataType
DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K< OutDataType, // OutDataType
InDataType, // InDataType AccDataType, // AccDataType
WeiDataType, // WeiDataType InElementOp, // InElementwiseOperation
OutDataType, // OutDataType WeiElementOp, // WeiElementwiseOperation
AccDataType, // AccDataType OutElementOp, // OutElementwiseOperation
InElementOp, // InElementwiseOperation ConvBwdDefault, // ConvolutionBackwardDataSpecialization
WeiElementOp, // WeiElementwiseOperation 256, // BlockSize
OutElementOp, // OutElementwiseOperation 128, // MPerBlock
ConvBwdDefault, // ConvolutionBackwardDataSpecialization 128, // NPerBlock
NumDimSpatial, // NumDimSpatial 4, // K0PerBlock
256, // BlockSize 8, // K1
128, // MPerBlock 32, // MPerXdl
128, // NPerBlock 32, // NPerXdl
4, // K0PerBlock 2, // MXdlPerWave
8, // K1 2, // NXdlPerWave
32, // MPerXdl S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
32, // NPerXdl S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
2, // MXdlPerWave S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // NXdlPerWave 2, // ABlockTransferSrcVectorDim
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1 8, // ABlockTransferSrcScalarPerVector
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder 8, // ABlockTransferDstScalarPerVector_K1
S<1, 0, 2>, // ABlockTransferSrcAccessOrder true, // ABlockLdsAddExtraM
2, // ABlockTransferSrcVectorDim S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
8, // ABlockTransferSrcScalarPerVector S<2, 0, 1>, // BBlockTransferThreadClusterArrangeOrder
8, // ABlockTransferDstScalarPerVector_K1 S<0, 2, 1>, // BBlockTransferSrcAccessOrder
true, // ABlockLdsAddExtraM 1, // BBlockTransferSrcVectorDim
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1 2, // BBlockTransferSrcScalarPerVector
S<2, 0, 1>, // BBlockTransferThreadClusterArrangeOrder 8, // BBlockTransferDstScalarPerVector_K1
S<0, 2, 1>, // BBlockTransferSrcAccessOrder true, // BBlockLdsAddExtraN
1, // BBlockTransferSrcVectorDim 7,
2, // BBlockTransferSrcScalarPerVector 1>; // GemmCThreadTransferDstScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
7,
1>; // GemmCThreadTransferDstScalarPerVector
template <ck::index_t NumDimSpatial>
using ReferenceConvBwdDataInstance =
ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NumDimSpatial>;
void print_use_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\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::tensor_operation::device::ConvParams parse_conv_params(int num_dim_spatial, char* argv[])
{
// (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
ck::tensor_operation::device::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;
}
DeviceConvBwdDataBasePtr get_conv_instance(int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 3: {
return std::make_unique<DeviceConvNDBwdDataInstance<3>>();
}
case 2: {
return std::make_unique<DeviceConvNDBwdDataInstance<2>>();
}
case 1: {
return std::make_unique<DeviceConvNDBwdDataInstance<1>>();
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
print_helper_msg();
bool do_verification = true; bool do_verification = true;
int init_method = 1; int init_method = 1;
bool time_kernel = false; bool time_kernel = false;
int num_dim_spatial = 2; int num_dim_spatial = 2;
ck::tensor_operation::device::ConvParams params; ck::tensor_operation::device::ConvParams params{
params.C_ = 128; 2, 128, 256, 256, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 4) if(argc == 1)
{
// use default
}
else if(argc == 4)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
else if(argc > 4) else
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]); num_dim_spatial = std::stoi(argv[4]);
// check args number
int conv_args = 3 + num_dim_spatial * 6;
int cmdline_nargs = conv_args + 5;
if(cmdline_nargs != argc)
{
print_use_msg();
exit(1);
}
params = parse_conv_params(num_dim_spatial, argv); params = parse_conv_params(num_dim_spatial, 5, argv);
}
else if(argc != 1)
{
print_use_msg();
exit(1);
} }
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_), const auto in_element_op = InElementOp{};
static_cast<std::size_t>(params.C_)}; const auto wei_element_op = WeiElementOp{};
input_dims.insert(std::end(input_dims), const auto out_element_op = OutElementOp{};
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_), if(num_dim_spatial == 1)
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_));
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> in_n_c_hi_wi_host_result(
ck::utils::conv::get_input_host_tensor_descriptor(input_dims, num_dim_spatial));
Tensor<InDataType> in_n_c_hi_wi_device_result(
ck::utils::conv::get_input_host_tensor_descriptor(input_dims, num_dim_spatial));
Tensor<WeiDataType> wei_k_c_y_x(
ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
Tensor<OutDataType> out_n_k_ho_wo(
ck::utils::conv::get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi_host_result.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{ {
case 0: break; return run_conv_bwd_data_nhwc<1,
case 1: InDataType,
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.2, 0.2}); WeiDataType,
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.2, 0.2}); OutDataType,
break; AccDataType,
default: InElementOp,
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1}); WeiElementOp,
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1}); OutElementOp,
} DeviceConvNdBwdDataInstance<1>>(do_verification,
init_method,
DeviceMem in_device_buf(sizeof(InDataType) * time_kernel,
in_n_c_hi_wi_device_result.mDesc.GetElementSpace()); params,
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace()); in_element_op,
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace()); wei_element_op,
out_element_op);
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data()); }
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data()); else if(num_dim_spatial == 2)
// reset input to zero
in_device_buf.SetZero();
// do GEMM
auto conv = get_conv_instance(num_dim_spatial);
auto invoker = conv->MakeInvokerPointer();
auto argument =
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.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_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
if(!conv->IsSupportedArgument(argument.get()))
{ {
throw std::runtime_error( return run_conv_bwd_data_nhwc<2,
"wrong! device_conv with the specified compilation parameters does " InDataType,
"not support this Conv problem"); WeiDataType,
} OutDataType,
AccDataType,
float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel}); InElementOp,
WeiElementOp,
std::size_t flop = ck::utils::conv::get_flops( OutElementOp,
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths); DeviceConvNdBwdDataInstance<2>>(do_verification,
std::size_t num_btype = ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>( init_method,
params.N_, time_kernel,
params.C_, params,
params.K_, in_element_op,
params.input_spatial_lengths_, wei_element_op,
params.filter_spatial_lengths_, out_element_op);
output_spatial_lengths); }
else if(num_dim_spatial == 3)
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
if(do_verification)
{ {
auto verify_f = [&](const auto& ref_conv) { return run_conv_bwd_data_nhwc<3,
auto ref_invoker = ref_conv.MakeInvoker(); InDataType,
WeiDataType,
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result, OutDataType,
wei_k_c_y_x, AccDataType,
out_n_k_ho_wo, InElementOp,
params.conv_filter_strides_, WeiElementOp,
params.conv_filter_dilations_, OutElementOp,
params.input_left_pads_, DeviceConvNdBwdDataInstance<3>>(do_verification,
params.input_right_pads_, init_method,
InElementOp{}, time_kernel,
WeiElementOp{}, params,
OutElementOp{}); in_element_op,
wei_element_op,
ref_invoker.Run(ref_argument); out_element_op);
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
return ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
in_n_c_hi_wi_host_result.mData)
? 0
: 1;
};
switch(num_dim_spatial)
{
case 3: {
auto ref_conv = ReferenceConvBwdDataInstance<3>();
return verify_f(ref_conv);
}
case 2: {
auto ref_conv = ReferenceConvBwdDataInstance<2>();
return verify_f(ref_conv);
}
case 1: {
auto ref_conv = ReferenceConvBwdDataInstance<1>();
return verify_f(ref_conv);
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
} }
return 0; return 0;
} }
example/20_convnd_bwd_weight_xdl/CMakeLists.txt
View file @ 0be1cf14
add_example_executable(example_convnd_bwd_weight_xdl convnd_bwd_weight_xdl.cpp) add_example_executable(example_convnd_bwd_weight_xdl_fp16 convnd_bwd_weight_xdl_fp16.cpp)
add_example_executable(example_convnd_bwd_weight_xdl_bf16_splitk convnd_bwd_weight_xdl_bf16_splitk.cpp) add_example_executable(example_convnd_bwd_weight_xdl_bf16 convnd_bwd_weight_xdl_bf16.cpp)
target_link_libraries(example_convnd_bwd_weight_xdl PRIVATE conv_util)
target_link_libraries(example_convnd_bwd_weight_xdl_bf16_splitk PRIVATE conv_util) target_link_libraries(example_convnd_bwd_weight_xdl_fp16 PRIVATE utility)
\ No newline at end of file target_link_libraries(example_convnd_bwd_weight_xdl_bf16 PRIVATE utility)
example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_common.hpp 0 → 100644
View file @ 0be1cf14
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp"
void print_helper_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< "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"
<< "split_k" << std::endl;
}
ck::tensor_operation::device::ConvParams
parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[])
{
const ck::index_t N = std::stoi(argv[arg_idx++]);
const ck::index_t K = std::stoi(argv[arg_idx++]);
const ck::index_t C = std::stoi(argv[arg_idx++]);
std::vector<ck::index_t> filter_spatial_lengths(num_dim_spatial);
std::vector<ck::index_t> input_spatial_lengths(num_dim_spatial);
std::vector<ck::index_t> conv_filter_strides(num_dim_spatial);
std::vector<ck::index_t> conv_filter_dilations(num_dim_spatial);
std::vector<ck::index_t> input_left_pads(num_dim_spatial);
std::vector<ck::index_t> input_right_pads(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_left_pads[i] = std::stoi(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_right_pads[i] = std::stoi(argv[arg_idx++]);
}
return ck::tensor_operation::device::ConvParams{num_dim_spatial,
N,
K,
C,
filter_spatial_lengths,
input_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads};
}
template <ck::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp,
typename DeviceConvBwdWeightInstance>
int run_conv_bwd_weight_nhwc(bool do_verification,
int init_method,
bool time_kernel,
const ck::tensor_operation::device::ConvParams& params,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op,
ck::index_t split_k)
{
auto f_nhwc_host_tensor_descriptor =
[](ck::index_t n, ck::index_t c, std::vector<ck::index_t> spatial_lengths) {
std::vector<std::size_t> nhwc_lengths{static_cast<std::size_t>(n),
static_cast<std::size_t>(c)};
nhwc_lengths.insert(
nhwc_lengths.begin() + 1, spatial_lengths.begin(), spatial_lengths.end());
return HostTensorDescriptor(nhwc_lengths);
};
Tensor<InDataType> in_n_hi_wi_c(
f_nhwc_host_tensor_descriptor(params.N_, params.C_, params.input_spatial_lengths_));
Tensor<WeiDataType> wei_k_y_x_c_host_result(
f_nhwc_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_));
Tensor<WeiDataType> wei_k_y_x_c_device_result(
f_nhwc_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_));
Tensor<OutDataType> out_n_ho_wo_k(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
std::cout << "input: " << in_n_hi_wi_c.mDesc << std::endl;
std::cout << "weight: " << wei_k_y_x_c_host_result.mDesc << std::endl;
std::cout << "output: " << out_n_ho_wo_k.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in_n_hi_wi_c.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break;
default:
in_n_hi_wi_c.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) *
wei_k_y_x_c_device_result.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
out_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
// init to 0
wei_device_buf.SetZero();
// do GEMM
auto conv = DeviceConvBwdWeightInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.output_spatial_lengths_,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
in_element_op,
wei_element_op,
out_element_op,
split_k);
if(!conv.IsSupportedArgument(argument))
{
std::cout << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return 1;
}
float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = params.GetFlops();
std::size_t num_btype = params.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< conv.GetTypeString() << std::endl;
if(do_verification)
{
auto ref_conv =
ck::tensor_operation::host::ReferenceConvBwdWeight<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_hi_wi_c,
wei_k_y_x_c_host_result,
out_n_ho_wo_k,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
wei_device_buf.FromDevice(wei_k_y_x_c_device_result.mData.data());
return ck::utils::check_err(wei_k_y_x_c_device_result.mData, wei_k_y_x_c_host_result.mData)
? 0
: 1;
}
return 0;
}
example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_xdl.cpp deleted 100644 → 0
View file @ b054669b
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdWeightDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
using DeviceConvBwdWeightBasePtr =
ck::tensor_operation::device::DeviceConvBwdWeightPtr<InElementOp, WeiElementOp, OutElementOp>;
// clang-format off
template <ck::index_t NumDimSpatial>
using DeviceConvndBwdWeightInstance = ck::tensor_operation::device::
DeviceConvndBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
NumDimSpatial, // NumDimSpatial
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
template <ck::index_t NumDimSpatial>
using ReferenceConvBwdWeightInstance =
ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NumDimSpatial>;
void print_use_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "arg4: is show log (0=no, 1=yes)\n"
<< "arg5: split-k \n"
<< "arg6: 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::tensor_operation::device::ConvParams parse_conv_params(int num_dim_spatial, char* argv[])
{
// (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
ck::tensor_operation::device::ConvParams params;
int arg_idx = 7;
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;
}
DeviceConvBwdWeightBasePtr get_conv_instance(int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 3: {
return std::make_unique<DeviceConvndBwdWeightInstance<3>>();
}
case 2: {
return std::make_unique<DeviceConvndBwdWeightInstance<2>>();
}
case 1: {
return std::make_unique<DeviceConvndBwdWeightInstance<1>>();
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
int do_log = 0;
int split_k = 1;
ck::tensor_operation::device::ConvParams params;
params.C_ = 128;
if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
}
else if(argc > 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
num_dim_spatial = std::stoi(argv[6]);
// check args number
int conv_args = 3 + num_dim_spatial * 6;
int cmdline_nargs = conv_args + 7;
if(cmdline_nargs != argc)
{
print_use_msg();
exit(1);
}
params = parse_conv_params(num_dim_spatial, argv);
}
else if(argc != 1)
{
print_use_msg();
exit(1);
}
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_));
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> in_n_c_hi_wi(
ck::utils::conv::get_input_host_tensor_descriptor(input_dims, num_dim_spatial));
Tensor<WeiDataType> wei_k_c_y_x_host_result(
ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
Tensor<WeiDataType> wei_k_c_y_x_device_result(
ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
Tensor<OutDataType> out_n_k_ho_wo(
ck::utils::conv::get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_device_result.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
break;
default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) *
wei_k_c_y_x_device_result.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
// reset input to zero
wei_device_buf.SetZero();
// do GEMM
auto conv = get_conv_instance(num_dim_spatial);
auto invoker = conv->MakeInvokerPointer();
auto argument =
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.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_,
InElementOp{},
WeiElementOp{},
OutElementOp{},
split_k);
// alloc work space
float ave_time = 0.f;
if(!conv->IsSupportedArgument(argument.get()))
{
std::cout << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return 1;
}
ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = ck::utils::conv::get_flops(
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype = ck::utils::conv::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 gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
if(do_verification)
{
auto verify_f = [&](const auto& ref_conv) {
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x_host_result,
out_n_k_ho_wo,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
if(do_log)
{
LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "wei_device(after): ", wei_k_c_y_x_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "wei_host : ", wei_k_c_y_x_host_result.mData, ",")
<< std::endl;
}
return ck::utils::check_err(wei_k_c_y_x_device_result.mData,
wei_k_c_y_x_host_result.mData)
? 0
: 1;
};
switch(num_dim_spatial)
{
case 3: {
auto ref_conv = ReferenceConvBwdWeightInstance<3>();
return verify_f(ref_conv);
}
case 2: {
auto ref_conv = ReferenceConvBwdWeightInstance<2>();
return verify_f(ref_conv);
}
case 1: {
auto ref_conv = ReferenceConvBwdWeightInstance<1>();
return verify_f(ref_conv);
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
return 0;
}
example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_xdl_bf16.cpp 0 → 100644
View file @ 0be1cf14
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_bwd_weight_common.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_bwd_weight_nwc_kxc_nwk_xdl_cshuffle.hpp"
using InDataType = ck::bhalf_t;
using WeiDataType =
float; // bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
using OutDataType = ck::bhalf_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdWeightDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
template <ck::index_t NDimSpatial>
using DeviceConvndBwdWeightInstance =
ck::tensor_operation::device::DeviceConvNdBwdWeightNwcKxcNwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
4>; // CBlockTransferScalarPerVector_NWaveNPerXdl
int main(int argc, char* argv[])
{
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
ck::tensor_operation::device::ConvParams params{
2, 32, 256, 1024, {3, 3}, {14, 14}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
ck::index_t split_k = 4;
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
params = parse_conv_params(num_dim_spatial, 5, argv);
split_k = std::stoi(argv[5 + 3 + 6 * num_dim_spatial - 1]);
split_k = std::max(1, split_k);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(num_dim_spatial == 1)
{
return run_conv_bwd_weight_nhwc<1,
InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<1>>(do_verification,
init_method,
time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
else if(num_dim_spatial == 2)
{
return run_conv_bwd_weight_nhwc<2,
InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<2>>(do_verification,
init_method,
time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
else if(num_dim_spatial == 3)
{
return run_conv_bwd_weight_nhwc<3,
InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<3>>(do_verification,
init_method,
time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
return 0;
}
example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_xdl_bf16_splitk.cpp deleted 100644 → 0
View file @ b054669b
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/device_unary_elementwise.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp"
using InDataType = ck::bhalf_t;
using WeiDataType = ck::bhalf_t;
using OutDataType = ck::bhalf_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnaryTypeConvert = ck::tensor_operation::element_wise::UnaryTypeConvert<ck::bhalf_t, float>;
using DeviceUnaryElementwiseTypeConvertInstance = ck::tensor_operation::device::
DeviceUnaryElementwise<AccDataType, WeiDataType, UnaryTypeConvert, 1, 4>;
static constexpr auto ConvBwdWeightDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
using DeviceConvBwdWeightBasePtr =
ck::tensor_operation::device::DeviceConvBwdWeightPtr<InElementOp, WeiElementOp, OutElementOp>;
// clang-format off
template <ck::index_t NumDimSpatial>
using DeviceConvndBwdWeightInstance_bf16_splitk = ck::tensor_operation::device::
DeviceConvndBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
InDataType, // InDataType
AccDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
NumDimSpatial, // NumDimSpatial
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
4>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
template <ck::index_t NumDimSpatial>
using ReferenceConvBwdWeightInstance =
ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NumDimSpatial>;
template <typename HostTensorB, typename HostTensorA, typename Functor>
void host_elementwise(HostTensorB& B,
const HostTensorA& A,
const std::vector<std::size_t>& shape,
Functor functor)
{
size_t tensor_size = std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>{});
std::cout << __LINE__ << ":" << tensor_size << ", " << A.mData[0] << std::endl;
for(std::size_t n = 0; n < tensor_size; ++n)
{
B.mData[n] = functor(A.mData[n]);
}
}
void print_use_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "arg4: is show log (0=no, 1=yes)\n"
<< "arg5: split-k : in this example split-k must be larger than 1\n"
<< "arg6: 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::tensor_operation::device::ConvParams parse_conv_params(int num_dim_spatial, char* argv[])
{
// (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
ck::tensor_operation::device::ConvParams params;
int arg_idx = 7;
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;
}
DeviceConvBwdWeightBasePtr get_conv_instance(int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 3: {
return std::make_unique<DeviceConvndBwdWeightInstance_bf16_splitk<3>>();
}
case 2: {
return std::make_unique<DeviceConvndBwdWeightInstance_bf16_splitk<2>>();
}
case 1: {
return std::make_unique<DeviceConvndBwdWeightInstance_bf16_splitk<1>>();
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
int do_log = 0;
int split_k = 2;
ck::tensor_operation::device::ConvParams params;
params.C_ = 128;
if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
}
else if(argc > 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
num_dim_spatial = std::stoi(argv[6]);
// check args number
int conv_args = 3 + num_dim_spatial * 6;
int cmdline_nargs = conv_args + 7;
if(cmdline_nargs != argc)
{
print_use_msg();
exit(1);
}
params = parse_conv_params(num_dim_spatial, argv);
}
else if(argc != 1)
{
print_use_msg();
exit(1);
}
if(split_k <= 1)
{
print_use_msg();
exit(1);
}
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_));
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> in_n_c_hi_wi(
ck::utils::conv::get_input_host_tensor_descriptor(input_dims, num_dim_spatial));
Tensor<WeiDataType> wei_k_c_y_x_host_result(
ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
Tensor<WeiDataType> wei_k_c_y_x_device_result(
ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
Tensor<OutDataType> out_n_k_ho_wo(
ck::utils::conv::get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_device_result.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
break;
default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) *
wei_k_c_y_x_device_result.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
// reset input to zero
wei_device_buf.SetZero();
// do GEMM
auto conv = get_conv_instance(num_dim_spatial);
auto invoker = conv->MakeInvokerPointer();
auto argument =
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.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_,
InElementOp{},
WeiElementOp{},
OutElementOp{},
split_k);
// alloc work space
size_t bwd_weight_workspace_size = conv->GetWorkSpaceSize(argument.get());
if(bwd_weight_workspace_size <= 0)
{
print_use_msg();
exit(1);
}
float conv_ave_time = 0.f;
DeviceMem wei_work_space_device_buf(bwd_weight_workspace_size);
wei_work_space_device_buf.SetZero();
conv->SetWorkSpacePointer(argument.get(), wei_work_space_device_buf.GetDeviceBuffer());
if(!conv->IsSupportedArgument(argument.get()))
{
std::cout << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return 1;
}
conv_ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = ck::utils::conv::get_flops(
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype = ck::utils::conv::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 / conv_ave_time;
float gb_per_sec = num_btype / 1.E6 / conv_ave_time;
std::cout << "Perf: conv: " << conv_ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s" << std::endl;
if(do_verification)
{
auto verify_f = [&](const auto& ref_conv) {
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x_host_result,
out_n_k_ho_wo,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
if(do_log)
{
LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "wei_device(after): ", wei_k_c_y_x_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "wei_host : ", wei_k_c_y_x_host_result.mData, ",")
<< std::endl;
}
return ck::utils::check_err(wei_k_c_y_x_device_result.mData,
wei_k_c_y_x_host_result.mData)
? 0
: 1;
};
switch(num_dim_spatial)
{
case 3: {
auto ref_conv = ReferenceConvBwdWeightInstance<3>();
verify_f(ref_conv);
break;
}
case 2: {
auto ref_conv = ReferenceConvBwdWeightInstance<2>();
verify_f(ref_conv);
break;
}
case 1: {
auto ref_conv = ReferenceConvBwdWeightInstance<1>();
verify_f(ref_conv);
break;
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
return 0;
}
example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_xdl_fp16.cpp 0 → 100644
View file @ 0be1cf14
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_bwd_weight_common.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_bwd_weight_nwc_kxc_nwk_xdl_cshuffle.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdWeightDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
template <ck::index_t NDimSpatial>
using DeviceConvndBwdWeightInstance =
ck::tensor_operation::device::DeviceConvNdBwdWeightNwcKxcNwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
int main(int argc, char* argv[])
{
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
ck::tensor_operation::device::ConvParams params{
2, 32, 256, 1024, {3, 3}, {14, 14}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
ck::index_t split_k = 4;
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
params = parse_conv_params(num_dim_spatial, 5, argv);
split_k = std::stoi(argv[5 + 3 + 6 * num_dim_spatial - 1]);
split_k = std::max(1, split_k);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(num_dim_spatial == 1)
{
return run_conv_bwd_weight_nhwc<1,
InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<1>>(do_verification,
init_method,
time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
else if(num_dim_spatial == 2)
{
return run_conv_bwd_weight_nhwc<2,
InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<2>>(do_verification,
init_method,
time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
else if(num_dim_spatial == 3)
{
return run_conv_bwd_weight_nhwc<3,
InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<3>>(do_verification,
init_method,
time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
return 0;
}
include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
View file @ 0be1cf14
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CONVOLUTION_BACKWARD_DATA_SPECIALIZATION #pragma once
#define CONVOLUTION_BACKWARD_DATA_SPECIALIZATION
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
...@@ -14,7 +13,18 @@ enum struct ConvolutionBackwardDataSpecialization ...@@ -14,7 +13,18 @@ enum struct ConvolutionBackwardDataSpecialization
Filter1x1Stride1Pad0, Filter1x1Stride1Pad0,
}; };
inline std::string
getConvBackwardDataSpecializationString(const ConvolutionBackwardDataSpecialization& s)
{
switch(s)
{
case ConvolutionBackwardDataSpecialization::Default: return "Default";
case ConvolutionBackwardDataSpecialization::FFilter1x1Stride1Pad0:
return "FFilter1x1Stride1Pad0";
default: return "Unrecognized specialization!";
}
}
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
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