load a_block as whole

1014e6c9 · root · c23de07d · 1014e6c9 · 1014e6c9 · 1014e6c9
Commit 1014e6c9 authored Mar 17, 2021 by root
3 changed files
--- a/composable_kernel/include/driver/driver_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp
+++ b/composable_kernel/include/driver/driver_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp
@@ -165,7 +165,7 @@ struct DriverDynamicConvolutionForwardImplicitGemm_v5r1_nchw_kcyx_nkhw_pad

 #if 1
        // GEMM
-        using gridwise_gemm = GridwiseDynamicGemm_km_kn_mn_v2<
+        using gridwise_gemm = GridwiseDynamicGemm_km_kn_mn_v3<
            BlockSize,
            Float,
            AccFloat,

--- a/composable_kernel/include/tensor_operation/gridwise_dynamic_gemm_v2.hpp
+++ b/composable_kernel/include/tensor_operation/gridwise_dynamic_gemm_v2.hpp
--- a/driver/include/device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp
@@ -78,7 +78,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw(InDesc
    constexpr index_t KPerThread   = 8;
    constexpr index_t HPerThread   = 1;
    constexpr index_t WPerThread   = 1;
-    constexpr index_t CYXPerThread = 1;
+    constexpr index_t CYXPerThread = 4;

    using GemmABlockTransferThreadSliceLengths_GemmK_GemmM   = Sequence<1, 1>;
    using GemmABlockTransferThreadClusterLengths_GemmK_GemmM = Sequence<4, 16>;