COMPMID-1201 - Implementing Winograd Convolution Layer 1x3 and 3x1 kernels on OpenCL

Change-Id: I39667bab49daa4da009694163274a59fd3574c73
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/137595
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/tests/datasets/WinogradOutputTransformDataset.h b/tests/datasets/WinogradOutputTransformDataset.h
index c7ba3b2..a4689c6 100644
--- a/tests/datasets/WinogradOutputTransformDataset.h
+++ b/tests/datasets/WinogradOutputTransformDataset.h
@@ -99,12 +99,11 @@
     std::vector<WinogradInfo> _info{};
 };
 
-class SmallWinogradOutputTransformDataset final : public WinogradOutputTransformDataset
+class SmallWinogradOutputTransformDatasetNCHW final : public WinogradOutputTransformDataset
 {
 public:
-    SmallWinogradOutputTransformDataset()
+    SmallWinogradOutputTransformDatasetNCHW()
     {
-        // NCHW
         // (2x2, 3x3)
         add_config(TensorShape(13U, 6U, 16U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(7U, 6U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(7U, 20U, 16U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(10U, 11U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
@@ -120,6 +119,34 @@
         add_config(TensorShape(24U, 16U, 36U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(14U, 14U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(7U, 12U, 16U, 5U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
 
+        // (2x1, 3x1)
+        add_config(TensorShape(13U, 18U, 4U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(7U, 6U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(7U, 44U, 4U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(10U, 11U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(1U, 891U, 4U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(53U, 33U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
+        add_config(TensorShape(7U, 30U, 4U, 3U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(24U, 98U, 4U, 2U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(14U, 14U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
+
+        // (1x2, 1x3)
+        add_config(TensorShape(13U, 14U, 4U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(7U, 6U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(7U, 50U, 4U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(10U, 11U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(1U, 901U, 4U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(53U, 33U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+        add_config(TensorShape(7U, 32U, 4U, 3U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(24U, 98U, 4U, 2U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(14U, 14U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+
+        // (4x1, 3x1)
+        add_config(TensorShape(13U, 12U, 6U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(7U, 6U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(7U, 22U, 6U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(10U, 11U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(1U, 462U, 6U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(53U, 33U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
+        add_config(TensorShape(7U, 20U, 6U, 3U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(24U, 56U, 6U, 2U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(14U, 14U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
+
+        // (1x4, 1x3)
+        add_config(TensorShape(13U, 7U, 6U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(7U, 6U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(7U, 30U, 6U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(10U, 11U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(1U, 477U, 6U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(53U, 33U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+        add_config(TensorShape(7U, 16U, 6U, 3U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(24U, 56U, 6U, 2U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(14U, 14U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+
         // (4x4, 5x5)
         add_config(TensorShape(13U, 1U, 64U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(7U, 6U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(7U, 4U, 64U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(10U, 11U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
@@ -127,8 +154,14 @@
         add_config(TensorShape(7U, 2U, 64U, 3U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(24U, 9U, 64U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(14U, 14U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(7U, 2U, 64U, 5U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+    }
+};
 
-        // NHWC
+class SmallWinogradOutputTransformDatasetNHWC final : public WinogradOutputTransformDataset
+{
+public:
+    SmallWinogradOutputTransformDatasetNHWC()
+    {
         // (4x4, 3x3)
         add_config(TensorShape(13U, 4U, 36U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(10U, 9U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC));
         add_config(TensorShape(13U, 6U, 36U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(10U, 11U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC));
@@ -146,10 +179,10 @@
     }
 };
 
-class LargeWinogradOutputTransformDataset final : public WinogradOutputTransformDataset
+class LargeWinogradOutputTransformDatasetNCHW final : public WinogradOutputTransformDataset
 {
 public:
-    LargeWinogradOutputTransformDataset()
+    LargeWinogradOutputTransformDatasetNCHW()
     {
         // NCHW
         // (2x2, 3x3)
@@ -168,13 +201,51 @@
         add_config(TensorShape(32U, 784U, 36U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(112U, 112U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
         add_config(TensorShape(13U, 196U, 36U, 5U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
 
+        // (2x1, 3x1)
+        add_config(TensorShape(64U, 24976U, 4U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(224U, 223U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
+        add_config(TensorShape(32U, 6160U, 4U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(112U, 110U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
+        add_config(TensorShape(13U, 1568U, 4U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(56U, 56U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
+        add_config(TensorShape(64U, 24753U, 4U, 3U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(224U, 223U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(32U, 6050U, 4U, 2U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(112U, 110U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(13U, 1512U, 4U, 5U), WinogradInfo(Size2D(2U, 1U), Size2D(3U, 1U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+
+        // (1x2, 1x3)
+        add_config(TensorShape(64U, 25088U, 4U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(224U, 223U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+        add_config(TensorShape(32U, 6160U, 4U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(112U, 110U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+        add_config(TensorShape(13U, 1568U, 4U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+        add_config(TensorShape(64U, 24864U, 4U, 3U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(224U, 223U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(32U, 6048U, 4U, 2U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(112U, 110U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(13U, 1512U, 4U, 5U), WinogradInfo(Size2D(1U, 2U), Size2D(1U, 3U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+
+        // (4x1, 3x1)
+        add_config(TensorShape(64U, 12488U, 6U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(224U, 223U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
+        add_config(TensorShape(32U, 3080U, 6U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(112U, 110U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
+        add_config(TensorShape(13U, 784U, 6U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(56U, 56U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
+        add_config(TensorShape(64U, 12488U, 6U, 3U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(224U, 223U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(32U, 3080U, 6U, 2U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(112U, 110U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(13U, 784U, 6U, 5U), WinogradInfo(Size2D(4U, 1U), Size2D(3U, 1U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+
+        // (1x4, 1x3)
+        add_config(TensorShape(64U, 12544U, 6U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(224U, 223U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+        add_config(TensorShape(32U, 3136U, 6U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(112U, 110U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+        add_config(TensorShape(13U, 784U, 6U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+        add_config(TensorShape(64U, 12544U, 6U, 3U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(224U, 223U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(32U, 3024U, 6U, 2U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(112U, 110U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        add_config(TensorShape(13U, 784U, 6U, 5U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 3U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+
         // (4x4, 5x5)
         add_config(TensorShape(32U, 756U, 64U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(112U, 112U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
         add_config(TensorShape(13U, 182U, 64U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
         add_config(TensorShape(32U, 756U, 64U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(112U, 112U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
         add_config(TensorShape(13U, 182U, 64U, 5U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+    }
+};
 
-        // NHWC
+class LargeWinogradOutputTransformDatasetNHWC final : public WinogradOutputTransformDataset
+{
+public:
+    LargeWinogradOutputTransformDatasetNHWC()
+    {
         // (4x4, 3x3)
         add_config(TensorShape(64U, 3136U, 36U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(224U, 224U), PadStrideInfo(1, 1, 1, 1), DataLayout::NHWC));
         add_config(TensorShape(32U, 784U, 36U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(112U, 112U), PadStrideInfo(1, 1, 1, 0), DataLayout::NHWC));