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

added lds double buffer (on C dimension) for implicit gemm v1r3, as a result, it should achieve 90% of peak for all filter sizes, on CHWN format

Browse Source
This commit is contained in:
Chao Liu
2019-04-19 17:49:25 -05:00
parent 6d066ede00
commit 87d8740bf5
3 changed files with 111 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
driver/driver.hip.cpp
View File

@@ -451,7 +451,7 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 1
#elif 0
// 3x3 filter, 28x28 image
constexpr index_t N = 128;
constexpr index_t C = 256;
@@ -499,6 +499,18 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 1;
constexpr index_t WPad = 1;
#elif 1
// 5x5 filter, 20x86 image
constexpr index_t N = 16;
constexpr index_t C = 256;
constexpr index_t HI = 20;
constexpr index_t WI = 86;
constexpr index_t K = 512;
constexpr index_t Y = 5;
constexpr index_t X = 5;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 5x5 filter, 20x86 image, 1x1 padding
constexpr index_t N = 16;
@@ -535,7 +547,7 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
#elif 10
// 1x1 filter, 14x14 image
constexpr index_t N = 128;
constexpr index_t C = 512;