Skip to content

GitLab

Commit 5ed51b71 authored by Jing Zhang's avatar Jing Zhang
Browse files

init commit of conv+add

parent b53926e9
Expand all Show whitespace changes
Inline Side-by-side
Showing with 2672 additions and 122 deletions
composable_kernel/include/driver/driver_static_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw_outpad.hpp
View file @ 5ed51b71
...@@ -226,7 +226,7 @@ struct DriverStaticConvolutionForwardImplicitGemm_v5r1_nchw_kcyx_nkhw_outpad ...@@ -226,7 +226,7 @@ struct DriverStaticConvolutionForwardImplicitGemm_v5r1_nchw_kcyx_nkhw_outpad
Sequence<0, 0, 0, 0, 0>{})); Sequence<0, 0, 0, 0, 0>{}));
// GEMM // GEMM
using gridwise_gemm = GridwiseStaticGemm_km_kn_mn_v3< using gridwise_gemm = GridwiseStaticGemm_km_kn_mn_v2<
BlockSize, BlockSize,
FloatAB, FloatAB,
FloatAcc, FloatAcc,
...@@ -273,11 +273,10 @@ struct DriverStaticConvolutionForwardImplicitGemm_v5r1_nchw_kcyx_nkhw_outpad ...@@ -273,11 +273,10 @@ struct DriverStaticConvolutionForwardImplicitGemm_v5r1_nchw_kcyx_nkhw_outpad
index_t nrepeat = 100; index_t nrepeat = 100;
std::cout << "conv_v5r1__NCHWc" << K1 << "_n" << N << "c" << C << "h" << Hi << "w" << Wi std::cout << "NCHWc" << K1 << "_n" << N << "c" << C << "h" << Hi << "w" << Wi << "-k" << K
<< "-k" << K << "c" << C << "y" << Y << "x" << X << "-u" << conv_strides[I0] << "c" << C << "y" << Y << "x" << X << "-u" << conv_strides[I0] << "v"
<< "v" << conv_strides[I1] << "l" << conv_dilations[I0] << "j" << conv_strides[I1] << "l" << conv_dilations[I0] << "j" << conv_dilations[I1]
<< conv_dilations[I1] << "q" << in_left_pads[I0] << "p" << in_right_pads[I0] << "q" << in_left_pads[I0] << "p" << in_right_pads[I0] << std::endl;
<< std::endl;
std::cout << "GridSize = " << GridSize << " BlockSize = " << BlockSize << std::endl; std::cout << "GridSize = " << GridSize << " BlockSize = " << BlockSize << std::endl;
...@@ -990,11 +989,10 @@ struct DriverStaticConvolutionForwardImplicitGemm_v5r1_nchw_kcyx_nkhw_1x1 ...@@ -990,11 +989,10 @@ struct DriverStaticConvolutionForwardImplicitGemm_v5r1_nchw_kcyx_nkhw_1x1
index_t nrepeat = 100; index_t nrepeat = 100;
std::cout << "conv_v5r1_NCHWc" << K1 << "_n" << N << "c" << C << "h" << Hi << "w" << Wi std::cout << "NCHWc" << K1 << "_n" << N << "c" << C << "h" << Hi << "w" << Wi << "-k" << K
<< "-k" << K << "c" << C << "y" << Y << "x" << X << "-u" << conv_strides[I0] << "c" << C << "y" << Y << "x" << X << "-u" << conv_strides[I0] << "v"
<< "v" << conv_strides[I1] << "l" << conv_dilations[I0] << "j" << conv_strides[I1] << "l" << conv_dilations[I0] << "j" << conv_dilations[I1]
<< conv_dilations[I1] << "q" << in_left_pads[I0] << "p" << in_right_pads[I0] << "q" << in_left_pads[I0] << "p" << in_right_pads[I0] << std::endl;
<< std::endl;
std::cout << "GridSize = " << GridSize << " BlockSize = " << BlockSize << std::endl; std::cout << "GridSize = " << GridSize << " BlockSize = " << BlockSize << std::endl;
......
composable_kernel/include/tensor_operation/gridwise_static_gemm_v2.hpp
View file @ 5ed51b71
...@@ -47,7 +47,7 @@ template <index_t BlockSize, ...@@ -47,7 +47,7 @@ template <index_t BlockSize,
typename CGlobalIteratorHacks, typename CGlobalIteratorHacks,
typename AGlobalMoveSliceWindowIteratorHacks, typename AGlobalMoveSliceWindowIteratorHacks,
typename BGlobalMoveSliceWindowIteratorHacks> typename BGlobalMoveSliceWindowIteratorHacks>
struct GridwiseStaticGemm_km_kn_mn_v3 struct GridwiseStaticGemm_km_kn_mn_v2
{ {
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte() __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{ {
...@@ -237,10 +237,12 @@ struct GridwiseStaticGemm_km_kn_mn_v3 ...@@ -237,10 +237,12 @@ struct GridwiseStaticGemm_km_kn_mn_v3
c_thread_buf; c_thread_buf;
// initialize output thread tensor // initialize output thread tensor
#if 0
ThreadwiseDynamicTensorSliceSet_v1<FloatAcc, ThreadwiseDynamicTensorSliceSet_v1<FloatAcc,
decltype(c_k_n_ho_wo_thread_desc), decltype(c_k_n_ho_wo_thread_desc),
Sequence<KPerThread, 1, HoPerThread, WoPerThread>>{} Sequence<KPerThread, 1, HoPerThread, WoPerThread>>{}
.Run(c_k_n_ho_wo_thread_desc, make_tuple(I0, I0, I0, I0), c_thread_buf, FloatAcc{0}); .Run(c_k_n_ho_wo_thread_desc, make_tuple(I0, I0, I0, I0), c_thread_buf, FloatAcc{0});
#endif
constexpr auto b_thread_slice_copy_step = make_multi_index(EPerBlock, 0, 0, 0); constexpr auto b_thread_slice_copy_step = make_multi_index(EPerBlock, 0, 0, 0);
......
host/driver_offline/CMakeLists.txt
View file @ 5ed51b71
...@@ -13,9 +13,15 @@ include_directories(BEFORE ...@@ -13,9 +13,15 @@ include_directories(BEFORE
set(CONV_FWD_DRIVER_OFFLINE_SOURCE conv_fwd_driver_offline.cpp) set(CONV_FWD_DRIVER_OFFLINE_SOURCE conv_fwd_driver_offline.cpp)
set(CONV_BWD_DRIVER_OFFLINE_SOURCE conv_bwd_driver_offline.cpp) set(CONV_BWD_DRIVER_OFFLINE_SOURCE conv_bwd_driver_offline.cpp)
set(CONV_ADD_FWD_DRIVER_OFFLINE_SOURCE conv_add_fwd_driver_offline.cpp)
set(CONV_ACTIV_FWD_DRIVER_OFFLINE_SOURCE conv_activ_fwd_driver_offline.cpp)
add_executable(conv_fwd_driver_offline ${CONV_FWD_DRIVER_OFFLINE_SOURCE}) add_executable(conv_fwd_driver_offline ${CONV_FWD_DRIVER_OFFLINE_SOURCE})
add_executable(conv_bwd_driver_offline ${CONV_BWD_DRIVER_OFFLINE_SOURCE}) add_executable(conv_bwd_driver_offline ${CONV_BWD_DRIVER_OFFLINE_SOURCE})
add_executable(conv_add_fwd_driver_offline ${CONV_ADD_FWD_DRIVER_OFFLINE_SOURCE})
add_executable(conv_activ_fwd_driver_offline ${CONV_ACTIV_FWD_DRIVER_OFFLINE_SOURCE})
target_link_libraries(conv_fwd_driver_offline PRIVATE host_tensor) target_link_libraries(conv_fwd_driver_offline PRIVATE host_tensor)
target_link_libraries(conv_bwd_driver_offline PRIVATE host_tensor) target_link_libraries(conv_bwd_driver_offline PRIVATE host_tensor)
target_link_libraries(conv_add_fwd_driver_offline PRIVATE host_tensor)
target_link_libraries(conv_activ_fwd_driver_offline PRIVATE host_tensor)
host/driver_offline/conv_activ_fwd_driver_offline.cpp 0 → 100644
View file @ 5ed51b71
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "conv_common.hpp"
#include "host_conv.hpp"
#include "device_tensor.hpp"
#include "device_static_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp"
#define USE_DYNAMIC_MODE 0
#define USE_CONV_FWD_V5R1_NCHW 1
enum ConvForwardAlgo
{
V5R1NCHW
};
int main(int argc, char* argv[])
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto I6 = Number<6>{};
// static mode
if(argc < 7)
{
printf("arg1 to 5: layout, algo, do_verification, init_method, do_log, nrepeat\n");
exit(1);
}
const ConvTensorLayout layout = static_cast<ConvTensorLayout>(atoi(argv[1]));
const ConvForwardAlgo algo = static_cast<ConvForwardAlgo>(atoi(argv[2]));
const bool do_verification = atoi(argv[3]);
const int init_method = atoi(argv[4]);
const bool do_log = atoi(argv[5]);
const int nrepeat = atoi(argv[6]);
#if 1
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 540;
constexpr index_t Wi = 960;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 480;
constexpr index_t Wi = 270;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 240;
constexpr index_t Wi = 135;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 540;
constexpr index_t Wi = 960;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 480;
constexpr index_t Wi = 270;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 8;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 4;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#endif
const index_t conv_stride_h = 1;
const index_t conv_stride_w = 1;
const index_t conv_dilation_h = 1;
const index_t conv_dilation_w = 1;
const index_t in_left_pad_h = 1;
const index_t in_left_pad_w = 1;
const index_t in_right_pad_h = 1;
const index_t in_right_pad_w = 1;
const index_t YEff = (Y - 1) * conv_dilation_h + 1;
const index_t XEff = (X - 1) * conv_dilation_w + 1;
const index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
#if 0
using in_data_t = float;
using acc_data_t = float;
using out_data_t = float;
#elif 1
using in_data_t = half_t;
using acc_data_t = float;
using out_data_t = half_t;
#elif 1
using in_data_t = int8_t;
using acc_data_t = int32_t;
using out_data_t = int8_t;
#endif
std::vector<std::size_t> in_lengths_host(4), wei_lengths_host(4), out_lengths_host(4);
switch(layout)
{
case ConvTensorLayout::NCHW:
// NCHW
in_lengths_host[0] = static_cast<std::size_t>(N);
in_lengths_host[1] = static_cast<std::size_t>(C);
in_lengths_host[2] = static_cast<std::size_t>(Hi);
in_lengths_host[3] = static_cast<std::size_t>(Wi);
wei_lengths_host[0] = static_cast<std::size_t>(K);
wei_lengths_host[1] = static_cast<std::size_t>(C);
wei_lengths_host[2] = static_cast<std::size_t>(Y);
wei_lengths_host[3] = static_cast<std::size_t>(X);
out_lengths_host[0] = static_cast<std::size_t>(N);
out_lengths_host[1] = static_cast<std::size_t>(K);
out_lengths_host[2] = static_cast<std::size_t>(Ho);
out_lengths_host[3] = static_cast<std::size_t>(Wo);
break;
case ConvTensorLayout::NHWC:
// NHWC
in_lengths_host[0] = static_cast<std::size_t>(N);
in_lengths_host[1] = static_cast<std::size_t>(Hi);
in_lengths_host[2] = static_cast<std::size_t>(Wi);
in_lengths_host[3] = static_cast<std::size_t>(C);
wei_lengths_host[0] = static_cast<std::size_t>(K);
wei_lengths_host[1] = static_cast<std::size_t>(Y);
wei_lengths_host[2] = static_cast<std::size_t>(X);
wei_lengths_host[3] = static_cast<std::size_t>(C);
out_lengths_host[0] = static_cast<std::size_t>(N);
out_lengths_host[1] = static_cast<std::size_t>(Ho);
out_lengths_host[2] = static_cast<std::size_t>(Wo);
out_lengths_host[3] = static_cast<std::size_t>(K);
break;
default: throw std::runtime_error("wrong! not implemented");
}
Tensor<in_data_t> in(in_lengths_host);
Tensor<in_data_t> wei(wei_lengths_host);
Tensor<out_data_t> out_host(out_lengths_host);
Tensor<out_data_t> out_device(out_lengths_host);
Tensor<out_data_t> add_device(out_lengths_host);
std::cout << "layout: " << layout << std::endl;
ostream_HostTensorDescriptor(in.mDesc, std::cout << "in: ");
ostream_HostTensorDescriptor(wei.mDesc, std::cout << "wei: ");
ostream_HostTensorDescriptor(out_host.mDesc, std::cout << "out: ");
print_array("InLeftPads", make_tuple(in_left_pad_h, in_left_pad_w));
print_array("InRightPads", make_tuple(in_right_pad_h, in_right_pad_w));
print_array("ConvStrides", make_tuple(conv_stride_h, conv_stride_w));
print_array("ConvDilations", make_tuple(conv_dilation_h, conv_dilation_w));
std::size_t num_thread = std::thread::hardware_concurrency();
switch(init_method)
{
case 0:
// no initialization
break;
case 1:
in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
break;
case 2:
in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
break;
case 3:
in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
break;
case 4:
in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
break;
case 5:
in.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_3<float>{-0.5, 0.5}, num_thread);
break;
default:
in.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
auto gen_wei = [](auto... is) {
return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
};
wei.GenerateTensorValue(gen_wei, num_thread);
}
auto f_make_for_device_nchw = [&]() {
const auto in_lengths_dev =
make_tuple(Number<N>{}, Number<C>{}, Number<Hi>{}, Number<Wi>{});
const auto wei_lengths_dev = make_tuple(Number<K>{}, Number<C>{}, Number<Y>{}, Number<X>{});
const auto out_lengths_dev =
make_tuple(Number<N>{}, Number<K>{}, Number<Ho>{}, Number<Wo>{});
const auto conv_strides_dev = make_tuple(Number<conv_stride_h>{}, Number<conv_stride_w>{});
const auto conv_dilations_dev =
make_tuple(Number<conv_dilation_h>{}, Number<conv_dilation_w>{});
const auto in_left_pads_dev = make_tuple(Number<in_left_pad_h>{}, Number<in_left_pad_w>{});
const auto in_right_pads_dev =
make_tuple(Number<in_right_pad_h>{}, Number<in_right_pad_w>{});
return make_tuple(in_lengths_dev,
wei_lengths_dev,
out_lengths_dev,
conv_strides_dev,
conv_dilations_dev,
in_left_pads_dev,
in_right_pads_dev);
};
auto f_make_for_device_nhwc = [&]() {
const auto in_lengths_dev =
make_tuple(Number<N>{}, Number<Hi>{}, Number<Wi>{}, Number<C>{});
const auto wei_lengths_dev = make_tuple(Number<K>{}, Number<Y>{}, Number<X>{}, Number<C>{});
const auto out_lengths_dev =
make_tuple(Number<N>{}, Number<Ho>{}, Number<Wo>{}, Number<K>{});
const auto conv_strides_dev = make_tuple(Number<conv_stride_h>{}, Number<conv_stride_w>{});
const auto conv_dilations_dev =
make_tuple(Number<conv_dilation_h>{}, Number<conv_dilation_w>{});
const auto in_left_pads_dev = make_tuple(Number<in_left_pad_h>{}, Number<in_left_pad_w>{});
const auto in_right_pads_dev =
make_tuple(Number<in_right_pad_h>{}, Number<in_right_pad_w>{});
return make_tuple(in_lengths_dev,
wei_lengths_dev,
out_lengths_dev,
conv_strides_dev,
conv_dilations_dev,
in_left_pads_dev,
in_right_pads_dev);
};
constexpr ck::index_t activ_type = 2;
#if USE_CONV_FWD_V5R1_NCHW
if(algo == ConvForwardAlgo::V5R1NCHW)
{
if(layout != ConvTensorLayout::NCHW)
{
throw std::runtime_error("wrong! layout");
}
const auto tmp = f_make_for_device_nchw();
device_static_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw<in_data_t,
8,
8,
activ_type,
acc_data_t,
out_data_t>(tmp[I0],
tmp[I1],
tmp[I2],
tmp[I3],
tmp[I4],
tmp[I5],
tmp[I6],
in,
wei,
out_device,
nrepeat);
}
#endif
if(do_verification)
{
host_direct_convolution_activ(in,
wei,
out_host,
make_tuple(conv_stride_h, conv_stride_w),
make_tuple(conv_dilation_h, conv_dilation_w),
make_tuple(in_left_pad_h, in_left_pad_w),
make_tuple(in_right_pad_h, in_right_pad_w),
activ_type,
layout);
check_error(out_host, out_device);
#if 0
if(do_log)
{
LogRangeAsType<float>(std::cout << "in : ", in.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "wei: ", wei.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "out_host : ", out_host.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "out_device: ", out_device.mData, ",") << std::endl;
}
#endif
}
}
host/driver_offline/conv_add_fwd_driver_offline.cpp 0 → 100644
View file @ 5ed51b71
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "conv_common.hpp"
#include "host_conv.hpp"
#include "device_tensor.hpp"
#include "device_static_convolution_add_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp"
#define USE_DYNAMIC_MODE 0
#define USE_CONV_FWD_V5R1_NCHW 1
enum ConvForwardAlgo
{
V5R1NCHW
};
int main(int argc, char* argv[])
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto I6 = Number<6>{};
constexpr auto I7 = Number<7>{};
// static mode
if(argc < 7)
{
printf("arg1 to 5: layout, algo, do_verification, init_method, do_log, nrepeat\n");
exit(1);
}
const ConvTensorLayout layout = static_cast<ConvTensorLayout>(atoi(argv[1]));
const ConvForwardAlgo algo = static_cast<ConvForwardAlgo>(atoi(argv[2]));
const bool do_verification = atoi(argv[3]);
const int init_method = atoi(argv[4]);
const bool do_log = atoi(argv[5]);
const int nrepeat = atoi(argv[6]);
#if 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 540;
constexpr index_t Wi = 960;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 480;
constexpr index_t Wi = 270;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 1
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 240;
constexpr index_t Wi = 135;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 540;
constexpr index_t Wi = 960;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 480;
constexpr index_t Wi = 270;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 8;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 4;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#endif
const index_t conv_stride_h = 1;
const index_t conv_stride_w = 1;
const index_t conv_dilation_h = 1;
const index_t conv_dilation_w = 1;
const index_t in_left_pad_h = 1;
const index_t in_left_pad_w = 1;
const index_t in_right_pad_h = 1;
const index_t in_right_pad_w = 1;
const index_t YEff = (Y - 1) * conv_dilation_h + 1;
const index_t XEff = (X - 1) * conv_dilation_w + 1;
const index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const index_t Hox2 = Ho * 2;
const index_t Wox2 = Wo * 2;
#if 0
using in_data_t = float;
using acc_data_t = float;
using out_data_t = float;
#elif 1
using in_data_t = half_t;
using acc_data_t = float;
using out_data_t = half_t;
#elif 1
using in_data_t = int8_t;
using acc_data_t = int32_t;
using out_data_t = int8_t;
#endif
std::vector<std::size_t> in_lengths_host(4), wei_lengths_host(4), out_lengths_host(4),
add_lengths_host(4);
switch(layout)
{
case ConvTensorLayout::NCHW:
// NCHW
in_lengths_host[0] = static_cast<std::size_t>(N);
in_lengths_host[1] = static_cast<std::size_t>(C);
in_lengths_host[2] = static_cast<std::size_t>(Hi);
in_lengths_host[3] = static_cast<std::size_t>(Wi);
wei_lengths_host[0] = static_cast<std::size_t>(K);
wei_lengths_host[1] = static_cast<std::size_t>(C);
wei_lengths_host[2] = static_cast<std::size_t>(Y);
wei_lengths_host[3] = static_cast<std::size_t>(X);
out_lengths_host[0] = static_cast<std::size_t>(N);
out_lengths_host[1] = static_cast<std::size_t>(K);
out_lengths_host[2] = static_cast<std::size_t>(Ho);
out_lengths_host[3] = static_cast<std::size_t>(Wo);
add_lengths_host[0] = static_cast<std::size_t>(N);
add_lengths_host[1] = static_cast<std::size_t>(K);
add_lengths_host[2] = static_cast<std::size_t>(Hox2);
add_lengths_host[3] = static_cast<std::size_t>(Wox2);
break;
case ConvTensorLayout::NHWC:
// NHWC
in_lengths_host[0] = static_cast<std::size_t>(N);
in_lengths_host[1] = static_cast<std::size_t>(Hi);
in_lengths_host[2] = static_cast<std::size_t>(Wi);
in_lengths_host[3] = static_cast<std::size_t>(C);
wei_lengths_host[0] = static_cast<std::size_t>(K);
wei_lengths_host[1] = static_cast<std::size_t>(Y);
wei_lengths_host[2] = static_cast<std::size_t>(X);
wei_lengths_host[3] = static_cast<std::size_t>(C);
out_lengths_host[0] = static_cast<std::size_t>(N);
out_lengths_host[1] = static_cast<std::size_t>(Ho);
out_lengths_host[2] = static_cast<std::size_t>(Wo);
out_lengths_host[3] = static_cast<std::size_t>(K);
add_lengths_host[0] = static_cast<std::size_t>(N);
add_lengths_host[1] = static_cast<std::size_t>(Hox2);
add_lengths_host[2] = static_cast<std::size_t>(Wox2);
add_lengths_host[3] = static_cast<std::size_t>(K);
break;
default: throw std::runtime_error("wrong! not implemented");
}
Tensor<in_data_t> in(in_lengths_host);
Tensor<in_data_t> wei(wei_lengths_host);
Tensor<in_data_t> add(add_lengths_host);
Tensor<out_data_t> out_host(add_lengths_host);
Tensor<out_data_t> out_device(add_lengths_host);
std::cout << "layout: " << layout << std::endl;
ostream_HostTensorDescriptor(in.mDesc, std::cout << "in: ");
ostream_HostTensorDescriptor(wei.mDesc, std::cout << "wei: ");
ostream_HostTensorDescriptor(add.mDesc, std::cout << "add: ");
ostream_HostTensorDescriptor(out_host.mDesc, std::cout << "out: ");
print_array("InLeftPads", make_tuple(in_left_pad_h, in_left_pad_w));
print_array("InRightPads", make_tuple(in_right_pad_h, in_right_pad_w));
print_array("ConvStrides", make_tuple(conv_stride_h, conv_stride_w));
print_array("ConvDilations", make_tuple(conv_dilation_h, conv_dilation_w));
std::size_t num_thread = std::thread::hardware_concurrency();
switch(init_method)
{
case 0:
// no initialization
break;
case 1:
in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
break;
case 2:
in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
break;
case 3:
in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
break;
case 4:
in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
break;
case 5:
in.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_3<float>{-0.5, 0.5}, num_thread);
break;
default:
in.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
auto gen_wei = [](auto... is) {
return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
};
wei.GenerateTensorValue(gen_wei, num_thread);
}
auto f_make_for_device_nchw = [&]() {
const auto in_lengths_dev =
make_tuple(Number<N>{}, Number<C>{}, Number<Hi>{}, Number<Wi>{});
const auto wei_lengths_dev = make_tuple(Number<K>{}, Number<C>{}, Number<Y>{}, Number<X>{});
const auto out_lengths_dev =
make_tuple(Number<N>{}, Number<K>{}, Number<Ho>{}, Number<Wo>{});
const auto add_lengths_dev =
make_tuple(Number<N>{}, Number<K>{}, Number<Hox2>{}, Number<Wox2>{});
const auto conv_strides_dev = make_tuple(Number<conv_stride_h>{}, Number<conv_stride_w>{});
const auto conv_dilations_dev =
make_tuple(Number<conv_dilation_h>{}, Number<conv_dilation_w>{});
const auto in_left_pads_dev = make_tuple(Number<in_left_pad_h>{}, Number<in_left_pad_w>{});
const auto in_right_pads_dev =
make_tuple(Number<in_right_pad_h>{}, Number<in_right_pad_w>{});
return make_tuple(in_lengths_dev,
wei_lengths_dev,
add_lengths_dev,
out_lengths_dev,
conv_strides_dev,
conv_dilations_dev,
in_left_pads_dev,
in_right_pads_dev);
};
auto f_make_for_device_nhwc = [&]() {
const auto in_lengths_dev =
make_tuple(Number<N>{}, Number<Hi>{}, Number<Wi>{}, Number<C>{});
const auto wei_lengths_dev = make_tuple(Number<K>{}, Number<Y>{}, Number<X>{}, Number<C>{});
const auto out_lengths_dev =
make_tuple(Number<N>{}, Number<Ho>{}, Number<Wo>{}, Number<K>{});
const auto add_lengths_dev =
make_tuple(Number<N>{}, Number<Hox2>{}, Number<Wox2>{}, Number<K>{});
const auto conv_strides_dev = make_tuple(Number<conv_stride_h>{}, Number<conv_stride_w>{});
const auto conv_dilations_dev =
make_tuple(Number<conv_dilation_h>{}, Number<conv_dilation_w>{});
const auto in_left_pads_dev = make_tuple(Number<in_left_pad_h>{}, Number<in_left_pad_w>{});
const auto in_right_pads_dev =
make_tuple(Number<in_right_pad_h>{}, Number<in_right_pad_w>{});
return make_tuple(in_lengths_dev,
wei_lengths_dev,
add_lengths_dev,
out_lengths_dev,
conv_strides_dev,
conv_dilations_dev,
in_left_pads_dev,
in_right_pads_dev);
};
constexpr ck::index_t activ_type = 2;
#if USE_CONV_FWD_V5R1_NCHW
if(algo == ConvForwardAlgo::V5R1NCHW)
{
if(layout != ConvTensorLayout::NCHW)
{
throw std::runtime_error("wrong! layout");
}
const auto tmp = f_make_for_device_nchw();
#if 1
device_static_convolution_add_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw<in_data_t,
8,
8,
activ_type,
acc_data_t,
out_data_t>(
tmp[I0], // in_lengths_dev
tmp[I1], // wei_lengths_dev
tmp[I2], // add_lengths_dev
tmp[I3], // out_lengths_dev
tmp[I4],
tmp[I5],
tmp[I6],
tmp[I7],
in,
wei,
add,
out_device,
nrepeat);
#endif
}
#endif
if(do_verification)
{
host_direct_convolution_add(in,
wei,
add,
out_host,
make_tuple(conv_stride_h, conv_stride_w),
make_tuple(conv_dilation_h, conv_dilation_w),
make_tuple(in_left_pad_h, in_left_pad_w),
make_tuple(in_right_pad_h, in_right_pad_w),
activ_type,
layout);
check_error(out_host, out_device);
#if 0
if(do_log)
{
LogRangeAsType<float>(std::cout << "in : ", in.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "wei: ", wei.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "out_host : ", out_host.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "out_device: ", out_device.mData, ",") << std::endl;
}
#endif
}
}
host/driver_offline/conv_fwd_driver_offline.cpp
View file @ 5ed51b71
...@@ -15,16 +15,15 @@ ...@@ -15,16 +15,15 @@
#include "device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp" #include "device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp"
#include "device_dynamic_convolution_forward_implicit_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp" #include "device_dynamic_convolution_forward_implicit_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp"
#include "device_dynamic_convolution_forward_implicit_gemm_v6r1_nchw_kcyx_nkhw.hpp" #include "device_dynamic_convolution_forward_implicit_gemm_v6r1_nchw_kcyx_nkhw.hpp"
#include "device_static_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp"
#include "device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp" #include "device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp"
#include "device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp" #include "device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp"
#include "device_dynamic_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp" #include "device_dynamic_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp"
#define USE_DYNAMIC_MODE 0 #define USE_DYNAMIC_MODE 1
#define USE_CONV_FWD_V4R4_NCHW 0 #define USE_CONV_FWD_V4R4_NCHW 1
#define USE_CONV_FWD_V4R4R2_NHWC 0 #define USE_CONV_FWD_V4R4R2_NHWC 0
#define USE_CONV_FWD_V6R1_NCHW 0 #define USE_CONV_FWD_V6R1_NCHW 0
#define USE_CONV_FWD_V5R1_NCHW 1 #define USE_CONV_FWD_V5R1_NCHW 0
#define USE_CONV_FWD_V4R4R2_XDL_NCHW 0 #define USE_CONV_FWD_V4R4R2_XDL_NCHW 0
#define USE_CONV_FWD_V4R4R4_XDL_NHWC 0 #define USE_CONV_FWD_V4R4R4_XDL_NHWC 0
...@@ -103,82 +102,16 @@ int main(int argc, char* argv[]) ...@@ -103,82 +102,16 @@ int main(int argc, char* argv[])
const bool do_log = atoi(argv[5]); const bool do_log = atoi(argv[5]);
const int nrepeat = atoi(argv[6]); const int nrepeat = atoi(argv[6]);
#if 1 constexpr index_t N = 128;
constexpr index_t N = 1; constexpr index_t C = 192;
constexpr index_t C = 16; constexpr index_t Hi = 71;
constexpr index_t Hi = 1080; constexpr index_t Wi = 71;
constexpr index_t Wi = 1920; constexpr index_t K = 256;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 540;
constexpr index_t Wi = 960;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 480;
constexpr index_t Wi = 270;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 240;
constexpr index_t Wi = 135;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 540;
constexpr index_t Wi = 960;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 480;
constexpr index_t Wi = 270;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 8;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 16;
constexpr index_t Hi = 1080;
constexpr index_t Wi = 1920;
constexpr index_t K = 4;
constexpr index_t Y = 3; constexpr index_t Y = 3;
constexpr index_t X = 3; constexpr index_t X = 3;
#endif
const index_t conv_stride_h = 1; const index_t conv_stride_h = 2;
const index_t conv_stride_w = 1; const index_t conv_stride_w = 2;
const index_t conv_dilation_h = 1; const index_t conv_dilation_h = 1;
const index_t conv_dilation_w = 1; const index_t conv_dilation_w = 1;
const index_t in_left_pad_h = 1; const index_t in_left_pad_h = 1;
...@@ -193,7 +126,7 @@ int main(int argc, char* argv[]) ...@@ -193,7 +126,7 @@ int main(int argc, char* argv[])
const index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1; const index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
#endif #endif
#if 0 #if 1
using in_data_t = float; using in_data_t = float;
using acc_data_t = float; using acc_data_t = float;
using out_data_t = float; using out_data_t = float;
...@@ -437,8 +370,6 @@ int main(int argc, char* argv[]) ...@@ -437,8 +370,6 @@ int main(int argc, char* argv[])
} }
#endif #endif
constexpr ck::index_t activ_type = 2;
#if USE_CONV_FWD_V5R1_NCHW #if USE_CONV_FWD_V5R1_NCHW
if(algo == ConvForwardAlgo::V5R1NCHW) if(algo == ConvForwardAlgo::V5R1NCHW)
{ {
...@@ -449,12 +380,10 @@ int main(int argc, char* argv[]) ...@@ -449,12 +380,10 @@ int main(int argc, char* argv[])
const auto tmp = f_make_for_device_nchw(); const auto tmp = f_make_for_device_nchw();
#if 1 device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw<in_data_t,
device_static_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw 16,
#else acc_data_t,
device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw out_data_t>(tmp[I0],
#endif
<in_data_t, 8, 8, activ_type, acc_data_t, out_data_t>(tmp[I0],
tmp[I1], tmp[I1],
tmp[I2], tmp[I2],
tmp[I3], tmp[I3],
...@@ -524,14 +453,13 @@ int main(int argc, char* argv[]) ...@@ -524,14 +453,13 @@ int main(int argc, char* argv[])
if(do_verification) if(do_verification)
{ {
host_direct_convolution_activ(in, host_direct_convolution(in,
wei, wei,
out_host, out_host,
make_tuple(conv_stride_h, conv_stride_w), make_tuple(conv_stride_h, conv_stride_w),
make_tuple(conv_dilation_h, conv_dilation_w), make_tuple(conv_dilation_h, conv_dilation_w),
make_tuple(in_left_pad_h, in_left_pad_w), make_tuple(in_left_pad_h, in_left_pad_w),
make_tuple(in_right_pad_h, in_right_pad_w), make_tuple(in_right_pad_h, in_right_pad_w),
activ_type,
layout); layout);
check_error(out_host, out_device); check_error(out_host, out_device);
......
host/driver_offline/include/device_static_convolution_add_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp 0 → 100644
View file @ 5ed51b71
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "driver_static_convolution_add_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw_outpad.hpp"
template <typename TInWei,
ck::index_t InWeiVectorSize,
ck::index_t OutVectorSize,
ck::index_t activ_type,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename AddLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_static_convolution_add_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw(
const InLengths& in_n_c_hi_wi_lengths,
const WeiLengths& wei_k_c_y_x_lengths,
const AddLengths& add_n_k_hox2_wox2_lengths,
const OutLengths& out_n_k_ho_wo_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_c_hi_wi,
const Tensor<TInWei>& wei_k_c_y_x,
const Tensor<TOut>& add_n_k_hox2_wox2,
Tensor<TOut>& out_n_k_hox2_wox2,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
const auto N = out_n_k_ho_wo_lengths[I0];
const auto K = out_n_k_ho_wo_lengths[I1];
const auto C = wei_k_c_y_x_lengths[I1];
const auto Hi = in_n_c_hi_wi_lengths[I2];
const auto Wi = in_n_c_hi_wi_lengths[I3];
const auto Ho = out_n_k_ho_wo_lengths[I2];
const auto Wo = out_n_k_ho_wo_lengths[I3];
const auto Hox2 = Ho * 2;
const auto Wox2 = Wo * 2;
const auto Y = wei_k_c_y_x_lengths[I2];
const auto X = wei_k_c_y_x_lengths[I3];
const auto C0 = C / Number<InWeiVectorSize>{};
const auto C1 = Number<InWeiVectorSize>{};
const auto K0 = K / Number<OutVectorSize>{};
const auto K1 = Number<OutVectorSize>{};
Tensor<TInWei> in_n_c0_hi_wi_c1(
HostTensorDescriptor(std::initializer_list<index_t>{N, C0, Hi, Wi, C1}));
Tensor<TInWei> wei_k_c0_y_x_c1(
HostTensorDescriptor(std::initializer_list<index_t>{K, C0, Y, X, C1}));
Tensor<TOut> out_n_k0_hox2_wox2_k1(
HostTensorDescriptor(std::initializer_list<index_t>{N, K0, Hox2, Wox2, K1}));
Tensor<TOut> add_n_k0_hox2_wox2_k1(
HostTensorDescriptor(std::initializer_list<index_t>{N, K0, Hox2, Wox2, K1}));
auto f_nchw2nc0hwc1 = [&](auto n, auto hi, auto wi, auto c) {
in_n_c0_hi_wi_c1(n, c / InWeiVectorSize, hi, wi, c % InWeiVectorSize) =
in_n_c_hi_wi(n, c, hi, wi);
};
auto f_kcyx2kc0yxc1 = [&](auto k, auto y, auto x, auto c) {
wei_k_c0_y_x_c1(k, c / InWeiVectorSize, y, x, c % InWeiVectorSize) =
wei_k_c_y_x(k, c, y, x);
};
auto f_nchx2wx2_to_nc0hx2wx2c1 = [&](auto n, auto ho, auto wo, auto c) {
add_n_k0_hox2_wox2_k1(n, c / InWeiVectorSize, ho, wo, c % InWeiVectorSize) =
add_n_k_hox2_wox2(n, c, ho, wo);
};
make_ParallelTensorFunctor(f_nchw2nc0hwc1, N, Hi, Wi, C)();
make_ParallelTensorFunctor(f_kcyx2kc0yxc1, K, Y, X, C)();
make_ParallelTensorFunctor(f_nchx2wx2_to_nc0hx2wx2c1, N, Hox2, Wox2, K)();
DeviceMem in_n_c0_hi_wi_c1_device_buf(sizeof(TInWei) *
in_n_c0_hi_wi_c1.mDesc.GetElementSpace());
DeviceMem wei_k_c0_y_x_c1_device_buf(sizeof(TInWei) * wei_k_c0_y_x_c1.mDesc.GetElementSpace());
DeviceMem add_n_k0_hox2_wox2_k1_device_buf(sizeof(TOut) *
add_n_k0_hox2_wox2_k1.mDesc.GetElementSpace());
DeviceMem out_n_k0_hox2_wox2_k1_device_buf(sizeof(TOut) *
out_n_k0_hox2_wox2_k1.mDesc.GetElementSpace());
in_n_c0_hi_wi_c1_device_buf.ToDevice(in_n_c0_hi_wi_c1.mData.data());
wei_k_c0_y_x_c1_device_buf.ToDevice(wei_k_c0_y_x_c1.mData.data());
const auto in_n_c0_hi_wi_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(make_tuple(N, C0, Hi, Wi));
const auto wei_k_c0_y_x_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(make_tuple(K, C0, Y, X));
const auto out_n_k0_ho_wo_k1_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(make_tuple(N, K0, Ho, Wo, K1));
const auto add_n_k0_hox2_wox2_k1_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(make_tuple(N, K0, Hox2, Wox2, K1));
// cdata = 64, BlockSize = 64, 16x8x32x4
constexpr index_t BlockSize = 64;
constexpr index_t KPerBlock = K;
constexpr index_t HoPerBlock = 8;
constexpr index_t WoPerBlock = 32;
constexpr index_t EPerBlock = C0;
constexpr index_t KPerThread = KPerBlock;
constexpr index_t HoPerThread = 2;
constexpr index_t WoPerThread = 2;
constexpr index_t EPerThread = EPerBlock;
using ABlockTransferThreadSliceLengths_E_K = Sequence<Y * X, 1>;
using ABlockTransferThreadClusterLengths_E_K = Sequence<EPerBlock, KPerBlock>;
constexpr index_t ABlockTransferSrcScalarPerVector_E = 1;
constexpr index_t ABlockTransferDstScalarPerVector_K = 1;
constexpr index_t BThreadTransferSrcScalarPerVector_W = 1;
constexpr index_t CThreadTransferDstScalarPerVector_W = K1;
static_assert(KPerThread % CThreadTransferDstScalarPerVector_W == 0, "");
constexpr auto conv_driver =
DriverStaticConvolutionAddForwardImplicitGemm_v5r1_nchw_kcyx_nkhw_outpad<
BlockSize,
typename vector_type<TInWei, InWeiVectorSize>::type,
TAcc,
TOut,
KPerBlock,
HoPerBlock,
WoPerBlock,
EPerBlock,
KPerThread,
HoPerThread,
WoPerThread,
EPerThread,
ABlockTransferThreadSliceLengths_E_K,
ABlockTransferThreadClusterLengths_E_K,
ABlockTransferSrcScalarPerVector_E,
ABlockTransferDstScalarPerVector_K,
BThreadTransferSrcScalarPerVector_W,
CThreadTransferDstScalarPerVector_W>{};
conv_driver.Run(wei_k_c0_y_x_desc,
in_n_c0_hi_wi_desc,
add_n_k0_hox2_wox2_k1_desc,
out_n_k0_ho_wo_k1_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<activ_type>{},
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
wei_k_c0_y_x_c1_device_buf.GetDeviceBuffer()),
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
in_n_c0_hi_wi_c1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(add_n_k0_hox2_wox2_k1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_k0_hox2_wox2_k1_device_buf.GetDeviceBuffer()));
out_n_k0_hox2_wox2_k1_device_buf.FromDevice(out_n_k0_hox2_wox2_k1.mData.data());
auto f_nk0hwk1_to_nkhw = [&](auto n, auto k, auto ho, auto wo) {
out_n_k_hox2_wox2(n, k, ho, wo) =
out_n_k0_hox2_wox2_k1(n, k / InWeiVectorSize, ho, wo, k % InWeiVectorSize);
};
make_ParallelTensorFunctor(f_nk0hwk1_to_nkhw, N, K, Hox2, Wox2)();
}
host/host_tensor/include/host_conv.hpp
View file @ 5ed51b71
...@@ -188,6 +188,115 @@ void host_direct_convolution_activ(const Tensor<TIn>& in, ...@@ -188,6 +188,115 @@ void host_direct_convolution_activ(const Tensor<TIn>& in,
} }
} }
template <typename TIn,
typename TWei,
typename TOut,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void host_direct_convolution_add(const Tensor<TIn>& in,
const Tensor<TWei>& wei,
const Tensor<TOut>& add,
Tensor<TOut>& out,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const ck::index_t activ_type,
const ConvTensorLayout layout = ConvTensorLayout::NCHW)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
double v = 0;
for(int c = 0; c < wei.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < wei.mDesc.GetLengths()[2]; ++y)
{
int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
for(int x = 0; x < wei.mDesc.GetLengths()[3]; ++x)
{
int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
{
v += static_cast<const double>(in(n, c, hi, wi)) *
static_cast<const double>(wei(k, c, y, x));
}
}
}
}
index_t hox2 = ho * 2;
index_t wox2 = wo * 2;
v = activ(v, activ_type);
out(n, k, hox2, wox2) = v + add(n, k, hox2, wox2);
out(n, k, hox2, wox2 + 1) = v + add(n, k, hox2, wox2 + 1);
out(n, k, hox2 + 1, wox2) = v + add(n, k, hox2 + 1, wox2);
out(n, k, hox2 + 1, wox2 + 1) = v + add(n, k, hox2 + 1, wox2 + 1);
};
auto f_nhwc = [&](auto n, auto ho, auto wo, auto k) {
double v = 0;
for(int c = 0; c < wei.mDesc.GetLengths()[3]; ++c)
{
for(int y = 0; y < wei.mDesc.GetLengths()[1]; ++y)
{
int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
for(int x = 0; x < wei.mDesc.GetLengths()[2]; ++x)
{
int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[1] && wi >= 0 &&
wi < in.mDesc.GetLengths()[2])
{
v += static_cast<const double>(in(n, hi, wi, c)) *
static_cast<const double>(wei(k, y, x, c));
}
}
}
}
index_t hox2 = ho * 2;
index_t wox2 = wo * 2;
v = activ(v, activ_type);
out(n, k, hox2, wox2) = v + add(n, k, hox2, wox2);
out(n, k, hox2, wox2 + 1) = v + add(n, k, hox2, wox2 + 1);
out(n, k, hox2 + 1, wox2) = v + add(n, k, hox2 + 1, wox2);
out(n, k, hox2 + 1, wox2 + 1) = v + add(n, k, hox2 + 1, wox2 + 1);
};
switch(layout)
{
case ConvTensorLayout::NCHW:
make_ParallelTensorFunctor(f_nchw,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2] / 2,
out.mDesc.GetLengths()[3] /
2)(std::thread::hardware_concurrency());
break;
case ConvTensorLayout::NHWC:
make_ParallelTensorFunctor(f_nhwc,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2] / 2,
out.mDesc.GetLengths()[3] /
2)(std::thread::hardware_concurrency());
break;
default: throw std::runtime_error("wrong! not supported layout");
}
}
template <typename TIn, typename TWei, typename TOut, typename InLeftPads, typename InRightPads> template <typename TIn, typename TWei, typename TOut, typename InLeftPads, typename InRightPads>
void host_winograd_3x3_convolution(const Tensor<TIn>& in_nchw, void host_winograd_3x3_convolution(const Tensor<TIn>& in_nchw,
const Tensor<TWei>& wei_kcyx, const Tensor<TWei>& wei_kcyx,
......
script/run.sh
View file @ 5ed51b71
...@@ -12,7 +12,7 @@ ...@@ -12,7 +12,7 @@
#export OLC_DEBUG_HIP_DUMP=1 #export OLC_DEBUG_HIP_DUMP=1
#export OLC_DEBUG_SAVE_TEMP_DIR=1 #export OLC_DEBUG_SAVE_TEMP_DIR=1
make -j conv_fwd_driver_offline make -j conv_add_fwd_driver_offline
#make -j conv_bwd_driver_offline #make -j conv_bwd_driver_offline
#make -j conv_fwd_driver_online #make -j conv_fwd_driver_online
...@@ -26,7 +26,7 @@ INIT=$4 ...@@ -26,7 +26,7 @@ INIT=$4
LOG=$5 LOG=$5
REPEAT=$6 REPEAT=$6
./host/driver_offline/conv_fwd_driver_offline $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT ./host/driver_offline/conv_add_fwd_driver_offline $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT
################################################ layout algo verify init log repeat N__ K___ C___ Y X Hi_ Wi__ Strides Dilations LeftPads RightPads ################################################ layout algo verify init log repeat N__ K___ C___ Y X Hi_ Wi__ Strides Dilations LeftPads RightPads
#./host/driver_offline/conv_fwd_driver_offline $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 128 192 3 3 71 71 2 2 1 1 1 1 1 1 #./host/driver_offline/conv_fwd_driver_offline $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 128 192 3 3 71 71 2 2 1 1 1 1 1 1
......
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