ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 20:40:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

refactor

Browse Source 
[ROCm/composable_kernel commit: 1bd880a674]
This commit is contained in:
Chao Liu
2019-04-09 16:13:14 -05:00
parent e99510c4db
commit 4fba215ede
11 changed files with 113 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp
View File

@@ -77,7 +77,7 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
wei_cyxk_device_buf.ToDevice(wei_cyxk.mData.data());
out_khwn_device_buf.ToDevice(out_khwn.mData.data());
#if 0
#if 1
// for 3x3, 34x34
constexpr index_t NPerBlock = 16;
constexpr index_t KPerBlock = 64;
@@ -105,6 +105,8 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
constexpr index_t OutThreadCopyDataPerWrite = 2;
@@ -145,7 +147,7 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr index_t BlockSize = 128;
#elif 0
// 3x3 58x58, NKC = 64, 64, 256
// 3x3 58x58
constexpr index_t NPerBlock = 16;
constexpr index_t KPerBlock = 64;
constexpr index_t CPerBlock = 4;
@@ -166,21 +168,6 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr index_t BlockSize = 128;
#elif 0
// 3x3 58x58, NKC = 16,256,128
constexpr index_t NPerBlock = 8;
constexpr index_t KPerBlock = 64;
constexpr index_t CPerBlock = 2;
constexpr index_t HoPerBlock = 4;
constexpr index_t WoPerBlock = 4;
constexpr index_t NPerThread = 4;
constexpr index_t KPerThread = 16;
constexpr index_t CPerThread = 1;
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 1;
constexpr index_t BlockSize = 128;
#elif 1
// for 7x7, 38x38
constexpr index_t NPerBlock = 16;
constexpr index_t KPerBlock = 128;
@@ -210,9 +197,42 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t OutThreadCopyDataPerWrite = 4;
constexpr index_t BlockSize = 128;
#elif 0
// for 3x3, 56x56, v1, Pacal
constexpr index_t NPerBlock = 32;
constexpr index_t KPerBlock = 64;
constexpr index_t CPerBlock = 4;
constexpr index_t HoPerBlock = 2;
constexpr index_t WoPerBlock = 2;
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 = 1;
constexpr index_t InBlockCopy_ThreadPerDimH = 4;
constexpr index_t InBlockCopy_ThreadPerDimW = 4;
constexpr index_t InBlockCopy_ThreadPerDimN = 8;
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 = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t OutThreadCopyDataPerWrite = 2;
constexpr index_t BlockSize = 128;
#elif 1
// for 3x3, 56x56
// for 3x3, 56x56, v1r2, Pascal
// for 3x3, 34x34, v1r2, Pascal
constexpr index_t NPerBlock = 16;
constexpr index_t KPerBlock = 128;
constexpr index_t CPerBlock = 8;
@@ -231,6 +251,8 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 2;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 1;
constexpr index_t GemmDataPerReadB = 1;
constexpr index_t InBlockCopy_ThreadPerDimC = 2;
constexpr index_t InBlockCopy_ThreadPerDimH = 4;
@@ -317,7 +339,7 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
for(index_t i = 0; i < nrepeat; ++i)
{
constexpr auto gridwise_conv =
#if 0
#if 1
GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn
#elif 1
GridwiseConvolutionImplicitGemm_v1r2_chwn_cyxk_khwn
@@ -346,6 +368,8 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB,
Sequence<InBlockCopy_ThreadPerDimC,
InBlockCopy_ThreadPerDimH,
InBlockCopy_ThreadPerDimW,

6
driver/device_implicit_gemm_convolution_2_chwn_cyxk_khwn.hpp
View File

@@ -205,6 +205,8 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
constexpr index_t InBlockCopyThreadPerDim0 = 4;
constexpr index_t InBlockCopyThreadPerDim1 = 16;
@@ -233,6 +235,8 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
constexpr index_t InBlockCopyThreadPerDim0 = 4;
constexpr index_t InBlockCopyThreadPerDim1 = 16;
@@ -289,6 +293,8 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB,
InBlockCopyThreadPerDim0,
InBlockCopyThreadPerDim1,
WeiBlockCopyThreadPerDim0,

4
driver/driver.hip.cpp
View File

@@ -409,7 +409,7 @@ 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;
@@ -454,7 +454,7 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 1
#elif 0
// 7x7, 38x38
constexpr index_t N = 64;
constexpr index_t C = 256;