ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 17:00:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

implicit gemm v1r2: only load 1d filter

Browse Source
This commit is contained in:
Chao Liu
2019-04-13 11:19:17 -05:00
parent 96ee9571e2
commit 00899f191b
17 changed files with 426 additions and 142 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp
View File

@@ -2,8 +2,8 @@
#include <unistd.h>
#include "device.hpp"
#include "gridwise_convolution_wrapper.hip.hpp"
#include "gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hip.hpp"
#include "gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn_lds_double_buffer.hip.hpp"
#include "gridwise_convolution_implicit_gemm_v1r1_chwn_cyxk_khwn.hip.hpp"
#include "gridwise_convolution_implicit_gemm_v1r1_chwn_cyxk_khwn_lds_double_buffer.hip.hpp"
#include "gridwise_convolution_implicit_gemm_v1r2_chwn_cyxk_khwn.hip.hpp"
template <class T, class InDesc, class WeiDesc, class OutDesc>
@@ -78,7 +78,7 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
out_khwn_device_buf.ToDevice(out_khwn.mData.data());
#if 0
// for 3x3, 34x34, Pascal
// for 3x3, 34x34, v1r1, Pascal
constexpr index_t NPerBlock = 16;
constexpr index_t KPerBlock = 64;
constexpr index_t CPerBlock = 4;
@@ -112,6 +112,40 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr index_t BlockSize = 128;
#elif 1
// for 3x3, 34x34, v1r2, Pascal
constexpr index_t NPerBlock = 4;
constexpr index_t KPerBlock = 64;
constexpr index_t CPerBlock = 8;
constexpr index_t HoPerBlock = 4;
constexpr index_t WoPerBlock = 8;
constexpr index_t NPerThread = 4;
constexpr index_t KPerThread = 8;
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 2;
constexpr index_t InBlockCopy_ThreadPerDimC = 4;
constexpr index_t InBlockCopy_ThreadPerDimH = 4;
constexpr index_t InBlockCopy_ThreadPerDimW = 2;
constexpr index_t InBlockCopy_ThreadPerDimN = 1;
constexpr index_t InBlockCopyDataPerRead = 4;
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 2;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
constexpr index_t OutThreadCopyDataPerWrite = 2;
constexpr index_t BlockSize = 128;
#elif 0
// for 3x3, 34x34, Vega 20
constexpr index_t NPerBlock = 16;
constexpr index_t KPerBlock = 128;
@@ -406,12 +440,12 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
for(index_t i = 0; i < nrepeat; ++i)
{
constexpr auto gridwise_conv =
#if 1
GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn
#if 0
GridwiseConvolutionImplicitGemm_v1r1_chwn_cyxk_khwn
#elif 0
GridwiseConvolutionImplicitGemm_v1r1_chwn_cyxk_khwn_lds_double_buffer
#elif 1
GridwiseConvolutionImplicitGemm_v1r2_chwn_cyxk_khwn
#elif 0
GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn_lds_double_buffer
#endif
<GridSize,
BlockSize,

43
driver/driver.hip.cpp
View File

@@ -35,6 +35,29 @@ struct GeneratorTensor_2
}
};
struct GeneratorTensor_3
{
int min_value = 0;
int max_value = 9;
template <class... Is>
double operator()(Is... is)
{
std::array<index_t, sizeof...(Is)> dims = {{static_cast<index_t>(is)...}};
#if 0
auto f_acc = std::plus<index_t>{};
#else
auto f_acc = [](auto a, auto b){ return 10*a + b;};
#endif
return std::accumulate(dims.begin(),
dims.end(),
index_t(0),
f_acc);
}
};
struct GeneratorTensor_Checkboard
{
template <class... Ts>
@@ -398,18 +421,7 @@ void check_error(const Tensor<T>& ref, const Tensor<T>& result)
int main(int argc, char* argv[])
{
#if 0
constexpr index_t N = 1;
constexpr index_t C = 1;
constexpr index_t HI = 28;
constexpr index_t WI = 28;
constexpr index_t K = 1;
constexpr index_t Y = 3;
constexpr index_t X = 3;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 1
#if 1
// 3x3, 34x34
constexpr index_t N = 64;
constexpr index_t C = 256;
@@ -656,13 +668,10 @@ int main(int argc, char* argv[])
#if 0
in_nchw.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#elif 0
in_nchw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#elif 1
in_nchw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
#elif 0
#elif 1
in_nchw.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
auto gen_wei = [](auto... is) {
@@ -681,7 +690,7 @@ int main(int argc, char* argv[])
device_direct_convolution_2_vectorized_nchw_kcyx_nkhw
#elif 1
device_implicit_gemm_convolution_1_chwn_cyxk_khwn
#elif 1
#elif 0
device_implicit_gemm_convolution_2_chwn_cyxk_khwn
#endif
(in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);