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

tuned implicit gemm v1 for 3x3 on AMD to 82%. Fixed a bug in 4d tensor blockwise copy.

Browse Source
This commit is contained in:
Chao Liu
2019-04-10 18:10:18 -05:00
parent edc89778c3
commit 96ee9571e2
6 changed files with 163 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp
View File

@@ -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
// for 3x3, 34x34, Pascal
constexpr index_t NPerBlock = 16;
constexpr index_t KPerBlock = 64;
constexpr index_t CPerBlock = 4;
@@ -111,6 +111,39 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr index_t OutThreadCopyDataPerWrite = 2;
constexpr index_t BlockSize = 128;
#elif 1
// for 3x3, 34x34, Vega 20
constexpr index_t NPerBlock = 16;
constexpr index_t KPerBlock = 128;
constexpr index_t CPerBlock = 4;
constexpr index_t HoPerBlock = 2;
constexpr index_t WoPerBlock = 4;
constexpr index_t NPerThread = 4;
constexpr index_t KPerThread = 8;
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 2;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 4;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 2;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
constexpr index_t InBlockCopy_ThreadPerDimC = 4;
constexpr index_t InBlockCopy_ThreadPerDimH = 4;
constexpr index_t InBlockCopy_ThreadPerDimW = 2;
constexpr index_t InBlockCopy_ThreadPerDimN = 8;
constexpr index_t InBlockCopyDataPerRead = 2;
constexpr index_t WeiBlockCopyDataPerRead = 2;
constexpr index_t OutThreadCopyDataPerWrite = 4;
constexpr index_t BlockSize = 256;
#elif 0
// for 5x5, 36x36
constexpr index_t NPerBlock = 16;
@@ -264,7 +297,7 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr index_t OutThreadCopyDataPerWrite = 4;
constexpr index_t BlockSize = 128;
#elif 1
#elif 0
// for 3x3, 28x28, v1, Pacal
constexpr index_t NPerBlock = 32;
constexpr index_t KPerBlock = 64;

8
driver/driver.hip.cpp
View File

@@ -409,13 +409,13 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
#elif 1
// 3x3, 34x34
constexpr index_t N = 64;
constexpr index_t C = 256;
constexpr index_t HI = 34;
constexpr index_t WI = 34;
constexpr index_t K = 64;
constexpr index_t K = 128;
constexpr index_t Y = 3;
constexpr index_t X = 3;
@@ -511,7 +511,7 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 1;
constexpr index_t WPad = 1;
#elif 1
#elif 0
// 3x3 filter, 28x28 image
constexpr index_t N = 128;
constexpr index_t C = 256;
@@ -681,7 +681,7 @@ 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 0
#elif 1
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);