hand tuned params

driver/conv.cu

@@ -170,7 +170,7 @@ int main()
 
     int num_thread = std::thread::hardware_concurrency();
 
-#if 1
+#if 0
     in.GenerateTensorValue(GeneratorTensor<float>{}, num_thread);
     wei.GenerateTensorValue(GeneratorTensor<float>{}, num_thread);
 #endif
@@ -180,7 +180,7 @@ int main()
         device_convolution(in_desc, in, wei_desc, wei, out_desc, out_device);
     }
 
-#if 1
+#if 0
     host_convolution(in, wei, out_host);
 
     float error      = 0;

driver/device_direct_convolution_3.cuh

@@ -27,9 +27,9 @@ void device_convolution(
     constexpr unsigned OutTileSizeH = 2;
     constexpr unsigned OutTileSizeW = 2;
     constexpr unsigned NPerBlock    = 2;
-    constexpr unsigned KPerBlock    = 8;
+    constexpr unsigned KPerBlock    = 32;
     constexpr unsigned CPerBlock    = 2;
-    constexpr unsigned YPerBlock    = 4;
+    constexpr unsigned YPerBlock    = 1;
     constexpr unsigned XPerBlock    = 16;
 
     constexpr unsigned NPerThread = 2;