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

adding implicit gemm

Browse Source
This commit is contained in:
Chao Liu
2019-01-14 11:13:36 -06:00
parent dc60d16962
commit aa0199a31c
3 changed files with 70 additions and 109 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
driver/conv.cu
View File

@@ -8,7 +8,7 @@
#include "conv_common.cuh"
#include "device_direct_convolution_1.cuh"
#include "device_direct_convolution_2.cuh"
//#include "device_implicit_gemm_convolution.cuh"
#include "device_implicit_gemm_convolution.cuh"
//#include "device_winograd_convolution.cuh"
struct GeneratorTensor_1
@@ -393,9 +393,9 @@ int main()
{
#if 0
device_direct_convolution_1(in_desc, in, wei_desc, wei, out_desc, out_device);
#elif 1
device_direct_convolution_2(in_desc, in, wei_desc, wei, out_desc, out_device);
#elif 0
device_direct_convolution_2(in_desc, in, wei_desc, wei, out_desc, out_device);
#elif 1
device_implicit_gemm_convolution(in_desc, in, wei_desc, wei, out_desc, out_device);
#elif 0
device_winograd_convolution(in_desc, in, wei_desc, wei, out_desc, out_device);

82
driver/device_implicit_gemm_convolution.cuh
View File

@@ -26,53 +26,24 @@ void device_implicit_gemm_convolution(
constexpr auto out_desc = OutDesc{};
#if 1
constexpr unsigned OutTileSizeH = 2;
constexpr unsigned OutTileSizeW = 2;
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
constexpr unsigned CPerBlock = 4;
constexpr unsigned YPerBlock = 1;
constexpr unsigned XPerBlock = 16;
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 128;
constexpr unsigned CPerBlock = 4;
constexpr unsigned HoPerBlock = 2;
constexpr unsigned WoPerBlock = 32;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned CPerThread = 2;
constexpr unsigned BlockSize = 128;
#elif 0
constexpr unsigned OutTileSizeH = 2;
constexpr unsigned OutTileSizeW = 2;
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
constexpr unsigned CPerBlock = 4;
constexpr unsigned YPerBlock = 1;
constexpr unsigned XPerBlock = 27;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned CPerThread = 2;
constexpr unsigned BlockSize = 216;
#elif 0
constexpr unsigned OutTileSizeH = 2;
constexpr unsigned OutTileSizeW = 2;
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
constexpr unsigned CPerBlock = 4;
constexpr unsigned YPerBlock = 1;
constexpr unsigned XPerBlock = 32;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned KPerThread = 8;
constexpr unsigned CPerThread = 2;
constexpr unsigned HoPerThread = 1;
constexpr unsigned WoPerThread = 4;
constexpr unsigned BlockSize = 256;
#endif
constexpr unsigned GridSize = (out_desc.GetLength(I0) / NPerBlock) *
(out_desc.GetLength(I1) / KPerBlock) *
(out_desc.GetLength(I2) / (OutTileSizeH * YPerBlock)) *
(out_desc.GetLength(I3) / (OutTileSizeW * XPerBlock));
constexpr unsigned GridSize =
(out_desc.GetLength(I0) / NPerBlock) * (out_desc.GetLength(I1) / KPerBlock) *
(out_desc.GetLength(I2) / HoPerBlock) * (out_desc.GetLength(I3) / WoPerBlock);
dim3 block_dim(BlockSize);
dim3 grid_dim(GridSize);
@@ -85,22 +56,21 @@ void device_implicit_gemm_convolution(
cudaEventCreate(&start);
cudaEventRecord(start, 0);
gridwise_implicit_gemm_convolution<T,
InDesc,
WeiDesc,
OutDesc,
OutTileSizeH,
OutTileSizeW,
NPerBlock,
KPerBlock,
CPerBlock,
YPerBlock,
XPerBlock,
NPerThread,
KPerThread,
CPerThread,
BlockSize,
GridSize>
gridwise_implicit_gemm_convolution_nchw_kcsr<GridSize,
BlockSize,
T,
InDesc,
WeiDesc,
OutDesc,
NPerBlock,
KPerBlock,
CPerBlock,
HoPerBlock,
WoPerBlock,
KPerThread,
CPerThread,
HoPerThread,
WoPerThread>
<<<grid_dim, block_dim>>>(InDesc{},
static_cast<T*>(in_device_buf.GetDeviceBuffer()),
WeiDesc{},