added lds doubl buffer for implicit gemm v4 (nchw, kcyx)

driver/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp

@@ -3,6 +3,7 @@
 #include "device.hpp"
 #include "gridwise_convolution_wrapper.hip.hpp"
 #include "gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hip.hpp"
+#include "gridwise_convolution_implicit_gemm_v4_lds_double_buffer_nchw_kcyx_nkhw.hip.hpp"
 
 template <class T, class InDesc, class WeiDesc, class OutDesc>
 void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
@@ -86,8 +87,10 @@ void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
     for(index_t i = 0; i < nrepeat; ++i)
     {
         constexpr auto gridwise_conv =
-#if 1
+#if 0
             GridwiseConvolutionImplicitGemm_v4_nchw_kcyx_nkhw
+#else
+            GridwiseConvolutionImplicitGemm_v4_lds_double_buffer_nchw_kcyx_nkhw
 #endif
             <GridSize,
              BlockSize,

driver/driver.hip.cpp

@@ -455,10 +455,10 @@ int main(int argc, char* argv[])
 
     constexpr index_t HPad = 0;
     constexpr index_t WPad = 0;
-#elif 0
+#elif 1
     // 1x1 filter, 28x28 image
-    constexpr index_t N  = 16;
-    constexpr index_t C  = 256;
+    constexpr index_t N  = 128;
+    constexpr index_t C  = 512;
     constexpr index_t HI = 28;
     constexpr index_t WI = 28;
     constexpr index_t K  = 512;