rename

driver/conv.cu

@@ -5,7 +5,8 @@
 #include "nvToolsExt.h"
 #include "tensor.hpp"
 #include "constant_tensor_descriptor.cuh"
-#include "device_direct_convolution_3.cuh"
+#include "device_direct_convolution_1.cuh"
+#include "device_direct_convolution_2.cuh"
 
 template <class T>
 struct GeneratorConstant
@@ -177,7 +178,7 @@ int main()
 
     for(int i = 0; i < 20; ++i)
     {
-        device_convolution(in_desc, in, wei_desc, wei, out_desc, out_device);
+        device_direct_convolution_2(in_desc, in, wei_desc, wei, out_desc, out_device);
     }
 
 #if 0

driver/device_direct_convolution_3.cuh → driver/device_direct_convolution_1.cuh

 #pragma once
-#include "direct_convolution_3.cuh"
+#include "gridwise_direct_convolution_1.cuh"
 
 template <class T, class InDesc, class WeiDesc, class OutDesc>
-void device_convolution(
+void device_direct_convolution_1(
     InDesc, const Tensor<T>& in, WeiDesc, const Tensor<T>& wei, OutDesc, Tensor<T>& out)
 {
     std::size_t data_sz = sizeof(T);
@@ -26,16 +26,12 @@ void device_convolution(
     constexpr auto out_desc         = OutDesc{};
     constexpr unsigned OutTileSizeH = 2;
     constexpr unsigned OutTileSizeW = 2;
-    constexpr unsigned NPerBlock    = 2;
-    constexpr unsigned KPerBlock    = 32;
+    constexpr unsigned NPerBlock    = 1;
+    constexpr unsigned KPerBlock    = 4;
     constexpr unsigned CPerBlock    = 2;
-    constexpr unsigned YPerBlock    = 1;
+    constexpr unsigned YPerBlock    = 8;
     constexpr unsigned XPerBlock    = 16;
 
-    constexpr unsigned NPerThread = 2;
-    constexpr unsigned KPerThread = 4;
-    constexpr unsigned CPerThread = 2;
-
     constexpr unsigned NBlockOpLen0 = 1;
     constexpr unsigned NBlockOpLen1 = 1;
     constexpr unsigned NBlockOpLen2 = 4;
@@ -70,9 +66,6 @@ void device_convolution(
                          CPerBlock,
                          YPerBlock,
                          XPerBlock,
-                         NPerThread,
-                         KPerThread,
-                         CPerThread,
                          NBlockOpLen0,
                          NBlockOpLen1,
                          NBlockOpLen2,

driver/device_direct_convolution_2.cuh

 #pragma once
-#include "direct_convolution_2.cuh"
+#include "gridwise_direct_convolution_2.cuh"
 
 template <class T, class InDesc, class WeiDesc, class OutDesc>
-void device_convolution(
+void device_direct_convolution_2(
     InDesc, const Tensor<T>& in, WeiDesc, const Tensor<T>& wei, OutDesc, Tensor<T>& out)
 {
     std::size_t data_sz = sizeof(T);
@@ -26,12 +26,16 @@ void device_convolution(
     constexpr auto out_desc         = OutDesc{};
     constexpr unsigned OutTileSizeH = 2;
     constexpr unsigned OutTileSizeW = 2;
-    constexpr unsigned NPerBlock    = 1;
-    constexpr unsigned KPerBlock    = 4;
+    constexpr unsigned NPerBlock    = 2;
+    constexpr unsigned KPerBlock    = 32;
     constexpr unsigned CPerBlock    = 2;
-    constexpr unsigned YPerBlock    = 8;
+    constexpr unsigned YPerBlock    = 1;
     constexpr unsigned XPerBlock    = 16;
 
+    constexpr unsigned NPerThread = 2;
+    constexpr unsigned KPerThread = 4;
+    constexpr unsigned CPerThread = 2;
+
     constexpr unsigned NBlockOpLen0 = 1;
     constexpr unsigned NBlockOpLen1 = 1;
     constexpr unsigned NBlockOpLen2 = 4;
@@ -66,6 +70,9 @@ void device_convolution(
                          CPerBlock,
                          YPerBlock,
                          XPerBlock,
+                         NPerThread,
+                         KPerThread,
+                         CPerThread,
                          NBlockOpLen0,
                          NBlockOpLen1,
                          NBlockOpLen2,

src/include/blockwise_convolution.cuh → src/include/blockwise_direct_convolution.cuh

 #pragma once
 #include "constant_tensor_descriptor.cuh"
 #include "threadwise_tensor_op.cuh"
-#include "threadwise_convolution.cuh"
+#include "threadwise_direct_convolution.cuh"
 
 template <class TFloat,
           class InBlockDesc,

src/include/direct_convolution_2.cuh → src/include/gridwise_direct_convolution_1.cuh

 #pragma once
 #include "constant_tensor_descriptor.cuh"
 #include "blockwise_tensor_op.cuh"
-#include "blockwise_convolution.cuh"
+#include "blockwise_direct_convolution.cuh"
 
 template <class TFloat,
           class InGlobalDesc,

src/include/direct_convolution_3.cuh → src/include/gridwise_direct_convolution_2.cuh

 #pragma once
 #include "constant_tensor_descriptor.cuh"
 #include "blockwise_tensor_op.cuh"
-#include "blockwise_convolution.cuh"
+#include "blockwise_direct_convolution.cuh"
 #include "threadwise_tensor_op.cuh"
-#include "threadwise_convolution.cuh"
+#include "threadwise_direct_convolution.cuh"
 
 template <class TFloat,
           class InGlobalDesc,