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

debugging implicit gemm v1: use 10d tensor output

Browse Source
This commit is contained in:
Chao Liu
2019-04-08 10:27:32 -05:00
parent 90abf42799
commit c9fa46af0b
17 changed files with 324 additions and 178 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
driver/device_direct_convolution_2_vectorized_nchw_kcyx_nkhw.hpp
View File

@@ -52,7 +52,7 @@ void device_direct_convolution_2_vectorized_nchw_kcyx_nkhw(InDesc,
in_nchw_vec(n, c, h, w) =
vector_t::Pack(in_nchw(n, 2 * c, h, w), in_nchw(n, 2 * c + 1, h, w));
#elif 1
in_nchw_vec(n, c, h, w) = vector_t::Pack(in_nchw(n, 4 * c, h, w),
in_nchw_vec(n, c, h, w) = vector_t::Pack(in_nchw(n, 4 * c, h, w),
in_nchw(n, 4 * c + 1, h, w),
in_nchw(n, 4 * c + 2, h, w),
in_nchw(n, 4 * c + 3, h, w));
@@ -114,37 +114,37 @@ void device_direct_convolution_2_vectorized_nchw_kcyx_nkhw(InDesc,
constexpr index_t BlockSize = 128;
#elif 0
// 3x3, 34x34, 128 thread, fp32, vector = 2
constexpr index_t NPerBlock = 2;
constexpr index_t KPerBlock = 32;
constexpr index_t CPerBlock = 2;
constexpr index_t NPerBlock = 2;
constexpr index_t KPerBlock = 32;
constexpr index_t CPerBlock = 2;
constexpr index_t HoPerBlock = 2;
constexpr index_t WoPerBlock = 32;
constexpr index_t NPerThread = 2;
constexpr index_t KPerThread = 4;
constexpr index_t CPerThread = 1;
constexpr index_t NPerThread = 2;
constexpr index_t KPerThread = 4;
constexpr index_t CPerThread = 1;
constexpr index_t HoPerThread = 2;
constexpr index_t WoPerThread = 2;
constexpr index_t InBlockCopyDataPerRead = 2;
constexpr index_t InBlockCopyDataPerRead = 2;
constexpr index_t WeiBlockCopyDataPerRead = 2;
constexpr index_t BlockSize = 128;
#elif 0
// 3x3, 34x34, 128 thread, int8, vector = 4
constexpr index_t NPerBlock = 2;
constexpr index_t KPerBlock = 32;
constexpr index_t CPerBlock = 8;
constexpr index_t NPerBlock = 2;
constexpr index_t KPerBlock = 32;
constexpr index_t CPerBlock = 8;
constexpr index_t HoPerBlock = 4;
constexpr index_t WoPerBlock = 32;
constexpr index_t NPerThread = 1;
constexpr index_t KPerThread = 8;
constexpr index_t CPerThread = 2;
constexpr index_t NPerThread = 1;
constexpr index_t KPerThread = 8;
constexpr index_t CPerThread = 2;
constexpr index_t HoPerThread = 4;
constexpr index_t WoPerThread = 2;
constexpr index_t InBlockCopyDataPerRead = 2;
constexpr index_t InBlockCopyDataPerRead = 2;
constexpr index_t WeiBlockCopyDataPerRead = 2;
constexpr index_t BlockSize = 128;

42
driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp
View File

@@ -248,16 +248,15 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr index_t BlockSize = 128;
#elif 1
// for 1x1, 14x14
// for 1x1, 14x14, Pascal
constexpr index_t NPerBlock = 16;
constexpr index_t KPerBlock = 128;
constexpr index_t CPerBlock = 8;
constexpr index_t HoPerBlock = 2;
constexpr index_t WoPerBlock = 2;
constexpr index_t NPerThread = 4;
constexpr index_t KPerThread = 16;
constexpr index_t CPerThread = 1;
constexpr index_t NPerThread = 8;
constexpr index_t KPerThread = 8;
constexpr index_t HoPerThread = 1;
constexpr index_t WoPerThread = 1;
@@ -265,8 +264,8 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
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 GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 2;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t InBlockCopy_ThreadPerDimC = 8;
@@ -278,6 +277,37 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t OutThreadCopyDataPerWrite = 2;
constexpr index_t BlockSize = 128;
#elif 1
// for 1x1, 14x14, Pascal, try
constexpr index_t NPerBlock = 16;
constexpr index_t KPerBlock = 128;
constexpr index_t CPerBlock = 8;
constexpr index_t HoPerBlock = 1;
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 = 2;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t InBlockCopy_ThreadPerDimC = 8;
constexpr index_t InBlockCopy_ThreadPerDimH = 1;
constexpr index_t InBlockCopy_ThreadPerDimW = 4;
constexpr index_t InBlockCopy_ThreadPerDimN = 4;
constexpr index_t InBlockCopyDataPerRead = 4;
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t OutThreadCopyDataPerWrite = 4;
constexpr index_t BlockSize = 128;
#endif

14
driver/device_implicit_gemm_convolution_2_chwn_cyxk_khwn.hpp
View File

@@ -69,7 +69,7 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
Tensor<T> out_khwn(make_TensorDescriptor(out_khwn_desc));
#if 0
#if 1
// 3x3, 34x34
// need to use register double buffer for GEMM
constexpr index_t BPerBlock = 128;
@@ -87,9 +87,6 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
constexpr index_t GemmNLevel1Cluster = 8;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmThreadPerColumnPerCluster = 8;
constexpr index_t GemmThreadPerRowPerCluster = 8;
constexpr index_t InBlockCopyThreadPerDim0 = 4;
constexpr index_t InBlockCopyThreadPerDim1 = 16;
@@ -98,6 +95,7 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
constexpr index_t InBlockCopyDataPerRead = 4;
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t OutThreadCopyDataPerWrite = 4;
constexpr index_t BlockSize = 128;
#elif 0
@@ -214,8 +212,8 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
constexpr index_t WeiBlockCopyThreadPerDim0 = 4;
constexpr index_t WeiBlockCopyThreadPerDim1 = 16;
constexpr index_t InBlockCopyDataPerRead = 4;
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t InBlockCopyDataPerRead = 4;
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t OutThreadCopyDataPerWrite = 4;
constexpr index_t BlockSize = 128;
@@ -242,8 +240,8 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
constexpr index_t WeiBlockCopyThreadPerDim0 = 4;
constexpr index_t WeiBlockCopyThreadPerDim1 = 16;
constexpr index_t InBlockCopyDataPerRead = 4;
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t InBlockCopyDataPerRead = 4;
constexpr index_t WeiBlockCopyDataPerRead = 4;
constexpr index_t OutThreadCopyDataPerWrite = 4;
constexpr index_t BlockSize = 256;

10
driver/driver.hip.cpp
View File

@@ -353,7 +353,7 @@ void host_winograd_3x3_convolution(const Tensor<TIn>& in_nchw,
std::size_t ho = HoPerTile * htile + j;
for(int i = 0; i < WoPerTile; ++i)
{
std::size_t wo = WoPerTile * wtile + i;
std::size_t wo = WoPerTile * wtile + i;
out_nkhw(n, k, ho, wo) = out_hold(n, k, htile, wtile, j, i);
}
}
@@ -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;
@@ -580,7 +580,7 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 1
#elif 0
// 1x1 filter, 14x14 image, C = 2048
constexpr index_t N = 128;
constexpr index_t C = 2048;
@@ -592,7 +592,7 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
#elif 1
// 1x1 filter, 14x14 image, C = 512
constexpr index_t N = 128;
constexpr index_t C = 512;
@@ -663,7 +663,7 @@ 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);