Skip to content

GitLab

"...composable_kernel_rocm.git" did not exist on "b03761009d453d42df8b20d0b5a3e32ba84b87d2"
Commit 61510f0a authored by Chao Liu's avatar Chao Liu
Browse files

clean

parent 65c56e56
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1214 additions and 3277 deletions
example/06_conv2d_fwd_bias_relu/CMakeLists.txt deleted 100644 → 0
View file @ 65c56e56
add_example_executable(example_conv2d_fwd_bias_relu_xdl_fp16 conv2d_fwd_bias_relu_xdl_fp16.cpp)
target_link_libraries(example_conv2d_fwd_bias_relu_xdl_fp16 PRIVATE utility)
add_example_executable(example_convnd_fwd_bias_relu_xdl_fp16 convnd_fwd_bias_relu_xdl_fp16.cpp)
target_link_libraries(example_convnd_fwd_bias_relu_xdl_fp16 PRIVATE utility)
example/06_conv2d_fwd_bias_relu/README.md deleted 100644 → 0
View file @ 65c56e56
# Instructions for ```example_conv_xdl_bias_relu```
## Run ```example_conv_xdl_bias_relu```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
#arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
./bin/example_conv_xdl_bias_relu 0 1 5
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
bias_k: dim 1, lengths {256}, strides {1}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 1.39009 ms, 105.581 TFlops, 239.981 GB/s
```
example/06_conv2d_fwd_bias_relu/conv2d_fwd_bias_relu_xdl_fp16.cpp deleted 100644 → 0
View file @ 65c56e56
// 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_conv2d_fwd_xdl_c_shuffle_bias_activation_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_fwd.hpp"
namespace {
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 InLayout = ck::tensor_layout::convolution::NHWC;
using WeiLayout = ck::tensor_layout::convolution::KYXC;
using OutLayout = ck::tensor_layout::convolution::NHWK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
static constexpr auto MemorySet = ck::InMemoryDataOperationEnum::Set;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
// clang-format off
using DeviceConvFwdInstance = ck::tensor_operation::device::
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_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
MemorySet, // OutGlobalMemoryDataOperation
ConvFwdDefault, // ConvForwardSpecialization
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
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;
}
ck::utils::conv::ConvParam parse_conv_param(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::utils::conv::ConvParam{num_dim_spatial,
N,
K,
C,
filter_spatial_lengths,
input_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads};
}
} // namespace
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::utils::conv::ConvParam params{
2, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
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{};
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(
f_nhwc_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_));
// bias: assume contiguous 1d vector
Tensor<OutDataType> bias_k(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(params.K_)})));
Tensor<OutDataType> out_n_ho_wo_k_host(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
Tensor<OutDataType> out_n_ho_wo_k_device(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
std::cout << "in_n_hi_wi_c: " << in_n_hi_wi_c.mDesc << std::endl;
std::cout << "wei_k_y_x_c: " << wei_k_y_x_c.mDesc << std::endl;
std::cout << "bias_k: " << bias_k.mDesc << std::endl;
std::cout << "output: " << out_n_ho_wo_k_host.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in_n_hi_wi_c.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
bias_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break;
default:
in_n_hi_wi_c.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
bias_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_hi_wi_c.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_y_x_c.mDesc.GetElementSpaceSize());
DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpaceSize());
DeviceMem out_device_buf(sizeof(OutDataType) *
out_n_ho_wo_k_device.mDesc.GetElementSpaceSize());
in_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_device_buf.ToDevice(wei_k_y_x_c.mData.data());
bias_device_buf.ToDevice(bias_k.mData.data());
// do GEMM
auto conv = DeviceConvFwdInstance{};
auto invoker = conv.MakeInvoker();
auto argument =
conv.MakeArgument(static_cast<const InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<const WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<const OutDataType*>(bias_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 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)
{
// use OutDataType for intermediate data
Tensor<OutDataType> tmp_n_ho_wo_k_host(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
auto ref_conv =
ck::tensor_operation::host::ReferenceConvFwd<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
PassThrough>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_hi_wi_c,
wei_k_y_x_c,
tmp_n_ho_wo_k_host,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
in_element_op,
wei_element_op,
PassThrough{});
ref_invoker.Run(ref_argument);
// FIXME: implement reference pointwise operation
for(int n = 0; n < params.N_; n++)
{
for(int ho = 0; ho < params.output_spatial_lengths_[0]; ho++)
{
for(int wo = 0; wo < params.output_spatial_lengths_[1]; wo++)
{
for(int k = 0; k < params.K_; k++)
{
out_element_op(out_n_ho_wo_k_host(n, ho, wo, k),
tmp_n_ho_wo_k_host(n, ho, wo, k),
bias_k(k));
}
}
}
}
out_device_buf.FromDevice(out_n_ho_wo_k_device.mData.data());
return ck::utils::check_err(out_n_ho_wo_k_host.mData,
out_n_ho_wo_k_device.mData,
"Error: incorrect results!",
1e-5f,
1e-4f)
? 0
: 1;
}
return 0;
}
example/07_conv2d_fwd_bias_relu_add/CMakeLists.txt deleted 100644 → 0
View file @ 65c56e56
add_example_executable(example_conv2d_fwd_bias_relu_add_xdl_fp16 conv2d_fwd_bias_relu_add_xdl_fp16.cpp)
target_link_libraries(example_conv2d_fwd_bias_relu_add_xdl_fp16 PRIVATE utility)
example/07_conv2d_fwd_bias_relu_add/README.md deleted 100644 → 0
View file @ 65c56e56
# Instructions for ```example_conv_xdl_bias_relu_add```
## Run ```example_conv_xdl_bias_relu_add```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
#arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
./bin/example_conv_xdl_bias_relu_add 0 1 5
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
bias_k: dim 1, lengths {256}, strides {1}
resi_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 1.44711 ms, 101.421 TFlops, 289.218 GB/s
```
example/07_conv2d_fwd_bias_relu_add/conv2d_fwd_bias_relu_add_xdl_fp16.cpp deleted 100644 → 0
View file @ 65c56e56
// 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_conv2d_fwd_xdl_c_shuffle_bias_activation_add_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_fwd.hpp"
namespace {
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 InLayout = ck::tensor_layout::convolution::NHWC;
using WeiLayout = ck::tensor_layout::convolution::KYXC;
using OutLayout = ck::tensor_layout::convolution::NHWK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
// clang-format off
using DeviceConvFwdInstance = ck::tensor_operation::device::
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_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
ConvFwdDefault, // ConvForwardSpecialization
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
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::utils::conv::ConvParam 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::utils::conv::ConvParam{num_dim_spatial,
N,
K,
C,
filter_spatial_lengths,
input_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads};
}
} // anonymous namespace
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::utils::conv::ConvParam params{
2, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
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{};
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(
f_nhwc_host_tensor_descriptor(params.K_, params.C_, params.filter_spatial_lengths_));
// bias: assume contiguous 1d vector
Tensor<OutDataType> bias_k(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(params.K_)})));
// resi: assume same layout as output tensor
Tensor<OutDataType> resi_n_ho_wo_k(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
Tensor<OutDataType> out_n_ho_wo_k_host(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
Tensor<OutDataType> out_n_ho_wo_k_device(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
std::cout << "in_n_hi_wi_c: " << in_n_hi_wi_c.mDesc << std::endl;
std::cout << "wei_k_y_x_c: " << wei_k_y_x_c.mDesc << std::endl;
std::cout << "bias_k: " << bias_k.mDesc << std::endl;
std::cout << "resi_n_ho_wo_k: " << resi_n_ho_wo_k.mDesc << std::endl;
std::cout << "out_n_ho_wo_k: " << out_n_ho_wo_k_host.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in_n_hi_wi_c.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
bias_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
resi_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});
wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
bias_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
resi_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_hi_wi_c.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_y_x_c.mDesc.GetElementSpaceSize());
DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpaceSize());
DeviceMem resi_device_buf(sizeof(OutDataType) * resi_n_ho_wo_k.mDesc.GetElementSpaceSize());
DeviceMem out_device_buf(sizeof(OutDataType) *
out_n_ho_wo_k_device.mDesc.GetElementSpaceSize());
in_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_device_buf.ToDevice(wei_k_y_x_c.mData.data());
bias_device_buf.ToDevice(bias_k.mData.data());
resi_device_buf.ToDevice(resi_n_ho_wo_k.mData.data());
auto conv = DeviceConvFwdInstance{};
auto invoker = conv.MakeInvoker();
auto argument =
conv.MakeArgument(static_cast<const InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<const WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
static_cast<const OutDataType*>(resi_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);
if(!conv.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device operator with the specified compilation parameters does "
"not support this problem");
}
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)
{
// use OutDataType for intermediate data
Tensor<OutDataType> tmp_n_ho_wo_k_host(
f_nhwc_host_tensor_descriptor(params.N_, params.K_, params.GetOutputSpatialLengths()));
auto ref_conv =
ck::tensor_operation::host::ReferenceConvFwd<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
PassThrough>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_hi_wi_c,
wei_k_y_x_c,
tmp_n_ho_wo_k_host,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
in_element_op,
wei_element_op,
PassThrough{});
ref_invoker.Run(ref_argument);
// FIXME: implement reference pointwise operation
for(int n = 0; n < params.N_; n++)
{
for(int ho = 0; ho < params.output_spatial_lengths_[0]; ho++)
{
for(int wo = 0; wo < params.output_spatial_lengths_[1]; wo++)
{
for(int k = 0; k < params.K_; k++)
{
out_element_op(out_n_ho_wo_k_host(n, ho, wo, k),
tmp_n_ho_wo_k_host(n, ho, wo, k),
bias_k(k),
resi_n_ho_wo_k(n, ho, wo, k));
}
}
}
}
out_device_buf.FromDevice(out_n_ho_wo_k_device.mData.data());
return ck::utils::check_err(out_n_ho_wo_k_host.mData,
out_n_ho_wo_k_device.mData,
"Error: incorrect results!",
1e-5f,
1e-4f)
? 0
: 1;
}
return 0;
}
example/28_group_convnd_fwd_bias_relu/CMakeLists.txt
View file @ 61510f0a
add_example_executable(example_group_convnd_fwd_bias_relu_xdl_fp16 group_convnd_fwd_bias_relu_xdl_fp16.cpp) add_example_executable(example_grouped_convnd_fwd_bias_relu_xdl_fp16 grouped_convnd_fwd_bias_relu_xdl_fp16.cpp)
target_link_libraries(example_grouped_convnd_fwd_bias_relu_xdl_fp16 PRIVATE utility)
example/28_group_convnd_fwd_bias_relu/README.md 0 → 100644
View file @ 61510f0a
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: time kernel (0=no, 1=yes)
#Following arguments (depending on number of spatial dims):
# N spatial dimensions
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
# <strides>, (ie Sy, Sx for 2D)
# <dilations>, (ie Dy, Dx for 2D)
# <left padding>, (ie LeftPy, LeftPx for 2D)
# <right padding>, (ie RightPy, RightPx for 2D)
bin/example_grouped_convnd_fwd_bias_relu_xdl_fp16 1 1 1
```
Result (MI100)
```
in: dim 5, lengths {1, 128, 192, 71, 71}, strides {6912, 967872, 1, 13632, 192}
wei: dim 5, lengths {1, 256, 192, 3, 3}, strides {192, 1728, 1, 576, 192}
bias: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
out: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 331776, 1, 9216, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 1.19215 ms, 123.112 TFlops, 279.827 GB/s, DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<256, 128, 256, 32, Default>
```
example/28_group_convnd_fwd_bias_relu/group_convnd_fwd_bias_relu_xdl_fp16.cpp deleted 100644 → 0
View file @ 65c56e56
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "../09_convnd_fwd/convnd_fwd_common.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_fwd_nwc_kxc_nwk_xdl.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_fwd_multiple_d_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::UnaryConvert;
using CShuffleDataType = ck::half_t;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial>
using DeviceConvNDFwdInstance =
ck::tensor_operation::device::DeviceConvNdFwdMultipleD_NwcKxcNwk_Xdl_CShuffle<
NDimSpatial, //
InDataType, //
WeiDataType, //
AccDataType, //
CShuffleDataType, //
ck::Tuple<>, //
OutDataType, //
InElementOp, // Input Elementwise Operation
WeiElementOp, // Weights Elementwise Operation
OutElementOp, // Output Elementwise Operation
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
32, // KPerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
1, // BBlockLdsExtraN
1,
1,
S<1, 32, 1, 8>,
8>;
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::utils::conv::ConvParam params{
2, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
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)
{
return run_conv_fwd<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvNDFwdInstance<1>>(do_verification,
init_method,
time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
}
else if(num_dim_spatial == 2)
{
return run_conv_fwd<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvNDFwdInstance<2>>(do_verification,
init_method,
time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
}
else if(num_dim_spatial == 3)
{
return run_conv_fwd<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvNDFwdInstance<3>>(do_verification,
init_method,
time_kernel,
params,
in_element_op,
wei_element_op,
out_element_op);
}
return 0;
}
example/06_conv2d_fwd_bias_relu/convnd_fwd_bias_common.hpp → example/28_group_convnd_fwd_bias_relu/grouped_convnd_fwd_bias_common.hpp
View file @ 61510f0a
...@@ -23,8 +23,8 @@ void print_helper_msg() ...@@ -23,8 +23,8 @@ void print_helper_msg()
std::cout << "arg1: verification (0=no, 1=yes)\n" std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n" << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\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" << "Following arguments (depending on number of spatial dims):\n"
<< " N spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)\n"
<< " G, N, K, C, \n" << " G, N, K, C, \n"
<< " <filter spatial dimensions>, (ie Y, X for 2D)\n" << " <filter spatial dimensions>, (ie Y, X for 2D)\n"
<< " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n" << " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
...@@ -92,7 +92,6 @@ ck::utils::conv::ConvParam parse_conv_param(int num_dim_spatial, int arg_idx, ch ...@@ -92,7 +92,6 @@ ck::utils::conv::ConvParam parse_conv_param(int num_dim_spatial, int arg_idx, ch
input_right_pads}; input_right_pads};
} }
// FIXME: current implementation only support NCHW/NHWC layout
template <ck::index_t NDimSpatial, template <ck::index_t NDimSpatial,
typename InDataType, typename InDataType,
typename WeiDataType, typename WeiDataType,
...@@ -101,17 +100,17 @@ template <ck::index_t NDimSpatial, ...@@ -101,17 +100,17 @@ template <ck::index_t NDimSpatial,
typename WeiElementOp, typename WeiElementOp,
typename OutElementOp, typename OutElementOp,
typename DeviceConvNDFwdInstance> typename DeviceConvNDFwdInstance>
int run_conv_fwd_bias(bool do_verification, int run_grouped_conv_fwd_bias(bool do_verification,
int init_method, int init_method,
bool time_kernel, bool time_kernel,
const ck::utils::conv::ConvParam& conv_param, const ck::utils::conv::ConvParam& conv_param,
const HostTensorDescriptor& in_g_n_c_wis_desc, const HostTensorDescriptor& in_g_n_c_wis_desc,
const HostTensorDescriptor& wei_g_k_c_xs_desc, const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& bias_g_n_k_wos_desc, const HostTensorDescriptor& bias_g_n_k_wos_desc,
const HostTensorDescriptor& out_g_n_k_wos_desc, const HostTensorDescriptor& out_g_n_k_wos_desc,
const InElementOp& in_element_op, const InElementOp& in_element_op,
const WeiElementOp& wei_element_op, const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op) const OutElementOp& out_element_op)
{ {
Tensor<InDataType> in(in_g_n_c_wis_desc); Tensor<InDataType> in(in_g_n_c_wis_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc); Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
...@@ -175,7 +174,7 @@ int run_conv_fwd_bias(bool do_verification, ...@@ -175,7 +174,7 @@ int run_conv_fwd_bias(bool do_verification,
copy(conv_param.input_left_pads_, input_left_pads); copy(conv_param.input_left_pads_, input_left_pads);
copy(conv_param.input_right_pads_, input_right_pads); copy(conv_param.input_right_pads_, input_right_pads);
// do GEMM // do Conv
auto conv = DeviceConvNDFwdInstance{}; auto conv = DeviceConvNDFwdInstance{};
auto invoker = conv.MakeInvoker(); auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument( auto argument = conv.MakeArgument(
......
example/06_conv2d_fwd_bias_relu/convnd_fwd_bias_relu_xdl_fp16.cpp → example/28_group_convnd_fwd_bias_relu/grouped_convnd_fwd_bias_relu_xdl_fp16.cpp
View file @ 61510f0a
// 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 "convnd_fwd_bias_common.hpp" #include "grouped_convnd_fwd_bias_common.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
using InDataType = ck::half_t; using InDataType = ck::half_t;
using WeiDataType = ck::half_t; using WeiDataType = ck::half_t;
...@@ -29,51 +29,52 @@ template <ck::index_t NDimSpatial, ...@@ -29,51 +29,52 @@ template <ck::index_t NDimSpatial,
typename WeiLayout, typename WeiLayout,
typename BiasLayout, typename BiasLayout,
typename OutLayout> typename OutLayout>
using DeviceConvNDFwdInstance = ck::tensor_operation::device::DeviceConvFwdMultipleD_Xdl_CShuffle< using DeviceGroupledConvNDFwdInstance =
NDimSpatial, ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
InLayout, NDimSpatial,
WeiLayout, InLayout,
ck::Tuple<BiasLayout>, WeiLayout,
OutLayout, ck::Tuple<BiasLayout>,
InDataType, OutLayout,
WeiDataType, InDataType,
AccDataType, WeiDataType,
CShuffleDataType, AccDataType,
ck::Tuple<BiasDataType>, CShuffleDataType,
OutDataType, ck::Tuple<BiasDataType>,
InElementOp, OutDataType,
WeiElementOp, InElementOp,
OutElementOp, WeiElementOp,
ConvSpec, // ConvForwardSpecialization OutElementOp,
GemmSpec, // GemmSpecialization ConvSpec, // ConvForwardSpecialization
1, // GemmSpec, // GemmSpecialization
256, // BlockSize 1, //
128, // MPerBlock 256, // BlockSize
256, // NPerBlock 128, // MPerBlock
32, // KPerBlock 256, // NPerBlock
8, // K1 32, // KPerBlock
32, // MPerXdl 8, // K1
32, // NPerXdl 32, // MPerXdl
2, // MXdlPerWave 32, // NPerXdl
4, // NXdlPerWave 2, // MXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1 4, // NXdlPerWave
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferSrcAccessOrder S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
2, // ABlockTransferSrcVectorDim S<1, 0, 2>, // ABlockTransferSrcAccessOrder
8, // ABlockTransferSrcScalarPerVector 2, // ABlockTransferSrcVectorDim
8, // ABlockTransferDstScalarPerVector_K1 8, // ABlockTransferSrcScalarPerVector
1, // ABlockLdsExtraM 8, // ABlockTransferDstScalarPerVector_K1
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1 1, // ABlockLdsExtraM
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferSrcAccessOrder S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
2, // BBlockTransferSrcVectorDim S<1, 0, 2>, // BBlockTransferSrcAccessOrder
8, // BBlockTransferSrcScalarPerVector 2, // BBlockTransferSrcVectorDim
8, // BBlockTransferDstScalarPerVector_K1 8, // BBlockTransferSrcScalarPerVector
1, // BBlockLdsExtraN 8, // BBlockTransferDstScalarPerVector_K1
1, 1, // BBlockLdsExtraN
1, 1,
S<1, 32, 1, 8>, 1,
8>; S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
...@@ -155,7 +156,7 @@ int main(int argc, char* argv[]) ...@@ -155,7 +156,7 @@ int main(int argc, char* argv[])
conv_param.G_ * conv_param.K_ // wo conv_param.G_ * conv_param.K_ // wo
}); });
return run_conv_fwd_bias< return run_grouped_conv_fwd_bias<
1, 1,
InDataType, InDataType,
WeiDataType, WeiDataType,
...@@ -163,7 +164,7 @@ int main(int argc, char* argv[]) ...@@ -163,7 +164,7 @@ int main(int argc, char* argv[])
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<1, InLayout, WeiLayout, BiasLayout, OutLayout>>( DeviceGroupledConvNDFwdInstance<1, InLayout, WeiLayout, BiasLayout, OutLayout>>(
do_verification, do_verification,
init_method, init_method,
time_kernel, time_kernel,
...@@ -242,7 +243,7 @@ int main(int argc, char* argv[]) ...@@ -242,7 +243,7 @@ int main(int argc, char* argv[])
conv_param.G_ * conv_param.K_ // wo conv_param.G_ * conv_param.K_ // wo
}); });
return run_conv_fwd_bias< return run_grouped_conv_fwd_bias<
2, 2,
InDataType, InDataType,
WeiDataType, WeiDataType,
...@@ -250,7 +251,7 @@ int main(int argc, char* argv[]) ...@@ -250,7 +251,7 @@ int main(int argc, char* argv[])
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<2, InLayout, WeiLayout, BiasLayout, OutLayout>>( DeviceGroupledConvNDFwdInstance<2, InLayout, WeiLayout, BiasLayout, OutLayout>>(
do_verification, do_verification,
init_method, init_method,
time_kernel, time_kernel,
...@@ -340,7 +341,7 @@ int main(int argc, char* argv[]) ...@@ -340,7 +341,7 @@ int main(int argc, char* argv[])
conv_param.G_ * conv_param.K_ // wo conv_param.G_ * conv_param.K_ // wo
}); });
return run_conv_fwd_bias< return run_grouped_conv_fwd_bias<
3, 3,
InDataType, InDataType,
WeiDataType, WeiDataType,
...@@ -348,7 +349,7 @@ int main(int argc, char* argv[]) ...@@ -348,7 +349,7 @@ int main(int argc, char* argv[])
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvNDFwdInstance<3, InLayout, WeiLayout, BiasLayout, OutLayout>>( DeviceGroupledConvNDFwdInstance<3, InLayout, WeiLayout, BiasLayout, OutLayout>>(
do_verification, do_verification,
init_method, init_method,
time_kernel, time_kernel,
......
example/CMakeLists.txt
View file @ 61510f0a
...@@ -25,8 +25,6 @@ add_subdirectory(01_gemm) ...@@ -25,8 +25,6 @@ add_subdirectory(01_gemm)
add_subdirectory(02_gemm_bilinear) add_subdirectory(02_gemm_bilinear)
add_subdirectory(03_gemm_bias_relu) add_subdirectory(03_gemm_bias_relu)
add_subdirectory(04_gemm_add_add_fastgelu) add_subdirectory(04_gemm_add_add_fastgelu)
add_subdirectory(06_conv2d_fwd_bias_relu)
add_subdirectory(07_conv2d_fwd_bias_relu_add)
add_subdirectory(09_convnd_fwd) add_subdirectory(09_convnd_fwd)
add_subdirectory(12_reduce) add_subdirectory(12_reduce)
add_subdirectory(13_pool2d_fwd) add_subdirectory(13_pool2d_fwd)
......
include/ck/tensor_operation/gpu/device/device_conv_fwd_multiple_d.hpp deleted 100644 → 0
View file @ 65c56e56
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// Convolution Forward:
// input : input image A[N, C, Hi, Wi],
// input : weight B[K, C, Y, X],
// input : D0[N, K, Ho, Wo], D1[N, K, Ho, Wo], ...
// output : output image E[N, K, Ho, Wo]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceConvFwdMultipleD : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
const void* p_a,
const void* p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
const std::array<index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
View file @ 61510f0a
...@@ -11,14 +11,14 @@ namespace ck { ...@@ -11,14 +11,14 @@ namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
// Grouped Convolution Forword // Convolution Forward:
// input : input image A[G, C, N, Hi, Wi], // input : input image A[G, N, C, Hi, Wi],
// input : weight B[G, K, C, Y, X], // input : weight B[G, K, C, Y, X],
// input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ... // input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
// output : output image E[G, N, K, Ho, Wo] // output : output image E[G, N, K, Ho, Wo]
// C = a_op(A) * b_op(B) // C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...) // E = cde_op(C, D0, D1, ...)
template <ck::index_t NDimSpatial, template <index_t NDimSpatial,
typename ALayout, typename ALayout,
typename BLayout, typename BLayout,
typename DsLayout, typename DsLayout,
...@@ -34,26 +34,26 @@ struct DeviceGroupedConvFwdMultipleD : public BaseOperator ...@@ -34,26 +34,26 @@ struct DeviceGroupedConvFwdMultipleD : public BaseOperator
{ {
static constexpr index_t NumDTensor = DsDataType::Size(); static constexpr index_t NumDTensor = DsDataType::Size();
virtual std::unique_ptr<BaseArgument> virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
MakeArgumentPointer(const void* p_a, const void* p_a,
const void* p_b, const void* p_b,
std::array<const void*, NumDTensor> p_ds, const std::array<const void*, NumDTensor>& p_ds,
void* p_e, void* p_e,
const std::vector<ck::index_t>& a_g_n_c_wis_lengths, const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::vector<ck::index_t>& a_g_n_c_wis_strides, const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::vector<ck::index_t>& b_g_k_c_xs_lengths, const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::vector<ck::index_t>& b_g_k_c_xs_strides, const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
std::array<std::vector<ck::index_t>, NumDTensor> ds_g_n_k_wos_lengths; const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
std::array<std::vector<ck::index_t>, NumDTensor> ds_g_n_k_wos_strides; const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
const std::vector<ck::index_t>& e_g_n_k_wos_lengths, const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::vector<ck::index_t>& e_g_n_k_wos_strides, const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::vector<ck::index_t>& conv_filter_strides, const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations, const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads, const std::array<index_t, NDimSpatial>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads, const std::array<index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op, const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op, const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) = 0; const CDEElementwiseOperation& cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0; virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
}; };
......
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