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

fixed LDS alignment bug

Browse Source
This commit is contained in:
Chao Liu
2019-02-06 01:54:13 -06:00
parent 079d63a788
commit 5e77650415
5 changed files with 63 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
driver/conv.cu
View File

@@ -391,7 +391,7 @@ int main()
constexpr unsigned HPad = 0;
constexpr unsigned WPad = 0;
#elif 0
#elif 1
// 3x3, 34x34
constexpr unsigned N = 64;
constexpr unsigned C = 256;
@@ -484,7 +484,7 @@ int main()
constexpr unsigned HPad = 1;
constexpr unsigned WPad = 1;
#elif 1
#elif 0
// 1x1 filter, 28x28 image
constexpr unsigned N = 16;
constexpr unsigned C = 256;
@@ -608,7 +608,7 @@ int main()
nrepeat);
#endif
#if 0
#if 1
if(S == 3 && R == 3)
{
host_winograd_3x3_convolution(in_nchw, wei_kcsr, out_nkhw_host, lower_pads, upper_pads);

29
driver/device_implicit_gemm_convolution_2_cnhw_csrk_knhw.cuh
View File

@@ -67,7 +67,7 @@ void device_implicit_gemm_convolution_2_cnhw_csrk_knhw(InDesc,
Tensor<T> out_knhw(make_TensorDescriptor(out_knhw_desc));
#if 0
#if 1
// 3x3, 34x34
constexpr unsigned BPerBlock = 128;
constexpr unsigned KPerBlock = 64;
@@ -86,11 +86,11 @@ void device_implicit_gemm_convolution_2_cnhw_csrk_knhw(InDesc,
constexpr unsigned WeiBlockCopyThreadPerDim0 = 4;
constexpr unsigned WeiBlockCopyThreadPerDim1 = 16;
constexpr unsigned InBlockCopyDataPerRead = 2;
constexpr unsigned InBlockCopyDataPerRead = 4;
constexpr unsigned WeiBlockCopyDataPerRead = 4;
constexpr unsigned BlockSize = 128;
#elif 1
#elif 0
// 1x1, 28x28
constexpr unsigned BPerBlock = 64;
constexpr unsigned KPerBlock = 64;
@@ -112,6 +112,29 @@ void device_implicit_gemm_convolution_2_cnhw_csrk_knhw(InDesc,
constexpr unsigned InBlockCopyDataPerRead = 4;
constexpr unsigned WeiBlockCopyDataPerRead = 4;
constexpr unsigned BlockSize = 64;
#elif 1
// 1x1, 28x28 try
constexpr unsigned BPerBlock = 64;
constexpr unsigned KPerBlock = 64;
constexpr unsigned CPerBlock = 8;
constexpr unsigned BPerThread = 4;
constexpr unsigned KPerThread = 16;
constexpr unsigned CPerThread = 1;
constexpr unsigned GemmThreadPerColumnPerCluster = 4;
constexpr unsigned GemmThreadPerRowPerCluster = 8;
constexpr unsigned InBlockCopyThreadPerDim0 = 4;
constexpr unsigned InBlockCopyThreadPerDim1 = 16;
constexpr unsigned WeiBlockCopyThreadPerDim0 = 4;
constexpr unsigned WeiBlockCopyThreadPerDim1 = 16;
constexpr unsigned InBlockCopyDataPerRead = 4;
constexpr unsigned WeiBlockCopyDataPerRead = 4;
constexpr unsigned BlockSize = 64;
#endif