double index buffer

driver/include/device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp

@@ -189,11 +189,12 @@ void device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc,
 
     constexpr index_t GemmMPerThreadSubC = 4;
     constexpr index_t GemmNPerThreadSubC = 4;
+    constexpr index_t GemmKPerThreadLoop = 1;
+
     constexpr index_t GemmMLevel0Cluster     = 4;
     constexpr index_t GemmNLevel0Cluster     = 4;
     constexpr index_t GemmMLevel1Cluster     = 4;
     constexpr index_t GemmNLevel1Cluster     = 4;
-    constexpr index_t GemmKPerThreadLoop     = 1;
     constexpr index_t ThreadGemmDataPerReadM = 4;
     constexpr index_t ThreadGemmDataPerReadN = 4;
 
@@ -237,11 +238,11 @@ void device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc,
         GemmKPerBlock,
         GemmMPerThreadSubC,
         GemmNPerThreadSubC,
+        GemmKPerThreadLoop,
         GemmMLevel0Cluster,
         GemmNLevel0Cluster,
         GemmMLevel1Cluster,
         GemmNLevel1Cluster,
-        GemmKPerThreadLoop,
         ThreadGemmDataPerReadM,
         ThreadGemmDataPerReadN,
         GemmABlockCopyThreadSliceLengths_GemmK_GemmM,

driver/src/conv_driver.cpp

@@ -29,13 +29,13 @@ int main(int argc, char* argv[])
 {
     using namespace ck;
 
-#if 1
+#if 0
     // 1x1
     constexpr index_t N  = 64;
-    constexpr index_t C  = 64;
-    constexpr index_t HI = 56;
-    constexpr index_t WI = 56;
-    constexpr index_t K  = 256;
+    constexpr index_t C  = 512;
+    constexpr index_t HI = 28;
+    constexpr index_t WI = 28;
+    constexpr index_t K  = 1024;
     constexpr index_t Y  = 1;
     constexpr index_t X  = 1;
 
@@ -44,13 +44,13 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 0>;
     using RightPads = Sequence<0, 0>;
-#elif 0
+#elif 1
     // 1x7
     constexpr index_t N  = 128;
-    constexpr index_t C  = 1024;
+    constexpr index_t C  = 128;
     constexpr index_t HI = 17;
     constexpr index_t WI = 17;
-    constexpr index_t K  = 1024;
+    constexpr index_t K  = 128;
     constexpr index_t Y  = 1;
     constexpr index_t X  = 7;
 
@@ -281,7 +281,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 0>;
     using RightPads = Sequence<0, 0>;
-#elif 0
+#elif 1
     // 3x3 filter, 2x2 stride, 35x35 input, 17x17 output
     // cudnn@V100 90%, ck@V100 93%, ck@P100 83%, ck@VII 81%
     constexpr index_t N  = 128;
@@ -327,7 +327,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 3>;
     using RightPads = Sequence<0, 3>;
-#elif 1
+#elif 0
     // 7x1 filter, 3x0 pad, 17x17 input
     constexpr index_t N  = 128;
     constexpr index_t C  = 128;
@@ -434,7 +434,7 @@ int main(int argc, char* argv[])
                                                                     ConvStrides{},
                                                                     ConvDilations{},
                                                                     nrepeat);
-#elif 0
+#elif 1
     device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(in_nchw_desc,
                                                          in_nchw,
                                                          wei_kcyx_desc,