adding implicit GEMM v4r2

a4b52461 · Chao Liu · e87aa851 · a4b52461
Commit a4b52461 authored Jul 13, 2019 by Chao Liu
Hide whitespace changes
Inline Side-by-side

Showing with 7 additions and 3 deletions

driver/include/device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw.hpp .../device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw.hpp +7 -3

No files found.
--- a/driver/include/device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw.hpp
@@ -55,13 +55,13 @@ void device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw(InDesc,

 #if 1
    // 1x1 filter, 8x8 image
-    constexpr index_t N1 = 2;
+    constexpr index_t N0 = 1;
    constexpr index_t N2 = 1;

-    constexpr index_t Ho1 = 8;
+    constexpr index_t Ho0 = 1;
    constexpr index_t Ho2 = 1;

-    constexpr index_t Wo1 = 1;
+    constexpr index_t Wo0 = 2;
    constexpr index_t Wo2 = 4;

    constexpr index_t BlockSize = 256;
@@ -105,6 +105,10 @@ void device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw(InDesc,
    constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
 #endif

+    constexpr index_t N1  = N / (N0 * N2);
+    constexpr index_t Ho1 = Ho / (Ho0 * Ho2);
+    constexpr index_t Wo1 = Wo / (Wo0 * Wo2);
+
    constexpr index_t B = N1 * Ho1 * Wo1;

    constexpr index_t GridSize =