COMPMID-935 - Implementing Convolution with Winograd on OpenCL (part 4)

Implemented Winograd Output Transform (2x2,3x3) on OpenCL
Implemented CLWinogradConvolutionLayer on OpenCL

Change-Id: I6a113fc5f052ca07f878d2b800d2ab003f84af65
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/125148
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

1 files changed, 171 insertions, 47 deletions

diff --git a/tests/validation/reference/Winograd.cpp b/tests/validation/reference/Winograd.cpp
index 3ed55fb9fc..c760663b22 100644
--- a/tests/validation/reference/Winograd.cpp
+++ b/tests/validation/reference/Winograd.cpp
@@ -39,6 +39,87 @@ namespace reference
 namespace
 {
 template <typename T>
+void winograd_filter_transform3x3(const SimpleTensor<T> &in, SimpleTensor<T> &out)
+{
+    // Simple tensor for the 3x3 input tile
+    SimpleTensor<T> input_tile{ TensorShape(3u, 3u), in.data_type(), 1 };
+
+    // Simple tensor for the transformation matrix
+    SimpleTensor<T> trans_matrix{ TensorShape(3u, 4u), in.data_type(), 1 };
+
+    // Simple tensor for the transformation matrix transpose
+    SimpleTensor<T> trans_matrix_transposed{ TensorShape(4u, 3u), in.data_type(), 1 };
+
+    // Simple tensor for the 4x3 temporary tile
+    SimpleTensor<T> tmp_tile{ TensorShape(3u, 4u), in.data_type(), 1 };
+
+    // Simple tensor for the 4x4 output tile
+    SimpleTensor<T> output_tile{ TensorShape(4u, 4u), in.data_type(), 1 };
+
+    // Initialize transformation matrix
+    // 1   | 0   | 0
+    // 0.5 | 0.5 | 0.5
+    // 0.5 |-0.5 | 0.5
+    // 0   | 0   | 1
+    trans_matrix[0 + 0 * 3] = 1.0f;
+    trans_matrix[1 + 0 * 3] = 0.0f;
+    trans_matrix[2 + 0 * 3] = 0.0f;
+    trans_matrix[0 + 1 * 3] = 0.5f;
+    trans_matrix[1 + 1 * 3] = 0.5f;
+    trans_matrix[2 + 1 * 3] = 0.5f;
+    trans_matrix[0 + 2 * 3] = 0.5f;
+    trans_matrix[1 + 2 * 3] = -0.5f;
+    trans_matrix[2 + 2 * 3] = 0.5f;
+    trans_matrix[0 + 3 * 3] = 0.0f;
+    trans_matrix[1 + 3 * 3] = 0.0f;
+    trans_matrix[2 + 3 * 3] = 1.0f;
+
+    // Transpose the transformation matrix
+    transpose_matrix(trans_matrix, trans_matrix_transposed);
+
+    const int num_channels = in.shape()[2];
+    const int num_filters  = in.shape()[3];
+    const int num_batches  = in.shape().total_size() / (9 * num_channels * num_filters);
+
+    for(int n = 0; n < num_batches; ++n)
+    {
+        for(int w = 0; w < num_filters; ++w)
+        {
+            for(int z = 0; z < num_channels; ++z)
+            {
+                // Load the 3x3 tile from the input tensor
+                get_tile(in, input_tile, Coordinates(0, 0, z, w, n));
+
+                // First transformation
+                matrix_multiply(trans_matrix, input_tile, tmp_tile);
+
+                // Second transformation
+                matrix_multiply(tmp_tile, trans_matrix_transposed, output_tile);
+
+                // Store the 4x4 output tile across the 16 channels
+                const int output_offset                              = w + z * num_filters;
+                out[output_offset + 0 * num_filters * num_channels]  = output_tile[0 + 0 * 4];
+                out[output_offset + 1 * num_filters * num_channels]  = output_tile[1 + 0 * 4];
+                out[output_offset + 2 * num_filters * num_channels]  = output_tile[2 + 0 * 4];
+                out[output_offset + 3 * num_filters * num_channels]  = output_tile[3 + 0 * 4];
+                out[output_offset + 4 * num_filters * num_channels]  = output_tile[0 + 1 * 4];
+                out[output_offset + 5 * num_filters * num_channels]  = output_tile[1 + 1 * 4];
+                out[output_offset + 6 * num_filters * num_channels]  = output_tile[2 + 1 * 4];
+                out[output_offset + 7 * num_filters * num_channels]  = output_tile[3 + 1 * 4];
+                out[output_offset + 8 * num_filters * num_channels]  = output_tile[0 + 2 * 4];
+                out[output_offset + 9 * num_filters * num_channels]  = output_tile[1 + 2 * 4];
+                out[output_offset + 10 * num_filters * num_channels] = output_tile[2 + 2 * 4];
+                out[output_offset + 11 * num_filters * num_channels] = output_tile[3 + 2 * 4];
+                out[output_offset + 12 * num_filters * num_channels] = output_tile[0 + 3 * 4];
+                out[output_offset + 13 * num_filters * num_channels] = output_tile[1 + 3 * 4];
+                out[output_offset + 14 * num_filters * num_channels] = output_tile[2 + 3 * 4];
+                out[output_offset + 15 * num_filters * num_channels] = output_tile[3 + 3 * 4];
+            }
+        }
+    }
+}
+
+template <typename T>
 void winograd_input_transform3x3(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const PadStrideInfo &conv_info)
 {
     TensorShape shape4x4(4u, 4u);
@@ -112,56 +193,70 @@ void winograd_input_transform3x3(const SimpleTensor<T> &src, SimpleTensor<T> &ds
 }
 
 template <typename T>
-void winograd_filter_transform3x3(const SimpleTensor<T> &in, SimpleTensor<T> &out)
+void winograd_output_transform3x3(const SimpleTensor<T> &in, SimpleTensor<T> &out, int num_tiles_x)
 {
+    ARM_COMPUTE_ERROR_ON(in.shape()[2] != 16);
+    ARM_COMPUTE_ERROR_ON(in.shape()[0] != out.shape()[2]);
+
     // Simple tensor for the 3x3 input tile
-    SimpleTensor<T> input_tile{ TensorShape(3u, 3u), in.data_type(), 1 };
+    SimpleTensor<T> input_tile{ TensorShape(4u, 4u), in.data_type(), 1 };
 
     // Simple tensor for the transformation matrix
-    SimpleTensor<T> trans_matrix{ TensorShape(3u, 4u), in.data_type(), 1 };
+    SimpleTensor<T> trans_matrix{ TensorShape(4u, 2u), in.data_type(), 1 };
 
     // Simple tensor for the transformation matrix transpose
-    SimpleTensor<T> trans_matrix_transposed{ TensorShape(4u, 3u), in.data_type(), 1 };
+    SimpleTensor<T> trans_matrix_transposed{ TensorShape(2u, 4u), in.data_type(), 1 };
 
     // Simple tensor for the 4x3 temporary tile
-    SimpleTensor<T> tmp_tile{ TensorShape(3u, 4u), in.data_type(), 1 };
+    SimpleTensor<T> tmp_tile{ TensorShape(4u, 2u), in.data_type(), 1 };
 
     // Simple tensor for the 4x4 output tile
-    SimpleTensor<T> output_tile{ TensorShape(4u, 4u), in.data_type(), 1 };
+    SimpleTensor<T> output_tile{ TensorShape(2u, 2u), in.data_type(), 1 };
 
     // Initialize transformation matrix
-    // 1   | 0   | 0
-    // 0.5 | 0.5 | 0.5
-    // 0.5 |-0.5 | 0.5
-    // 0   | 0   | 1
-    trans_matrix[0 + 0 * 3] = 1.0f;
-    trans_matrix[1 + 0 * 3] = 0.0f;
-    trans_matrix[2 + 0 * 3] = 0.0f;
-    trans_matrix[0 + 1 * 3] = 0.5f;
-    trans_matrix[1 + 1 * 3] = 0.5f;
-    trans_matrix[2 + 1 * 3] = 0.5f;
-    trans_matrix[0 + 2 * 3] = 0.5f;
-    trans_matrix[1 + 2 * 3] = -0.5f;
-    trans_matrix[2 + 2 * 3] = 0.5f;
-    trans_matrix[0 + 3 * 3] = 0.0f;
-    trans_matrix[1 + 3 * 3] = 0.0f;
-    trans_matrix[2 + 3 * 3] = 1.0f;
+    // 1   | 1   | 1   | 1
+    // 0   | 1   | -1  | -1
+    trans_matrix[0 + 0 * 4] = 1.0f;
+    trans_matrix[1 + 0 * 4] = 1.0f;
+    trans_matrix[2 + 0 * 4] = 1.0f;
+    trans_matrix[3 + 0 * 4] = 0.0f;
+    trans_matrix[0 + 1 * 4] = 0.0f;
+    trans_matrix[1 + 1 * 4] = 1.0f;
+    trans_matrix[2 + 1 * 4] = -1.0f;
+    trans_matrix[3 + 1 * 4] = -1.0f;
 
     // Transpose the transformation matrix
     transpose_matrix(trans_matrix, trans_matrix_transposed);
 
-    const int num_channels = in.shape()[2];
-    const int num_filters  = in.shape()[3];
-    const int num_batches  = in.shape().total_size() / (9 * num_channels * num_filters);
+    const int w_in        = in.shape()[0];
+    const int h_in        = in.shape()[1];
+    const int c_in        = in.shape()[2];
+    const int w_out       = out.shape()[0];
+    const int h_out       = out.shape()[1];
+    const int c_out       = out.shape()[2];
+    const int num_batches = in.shape().total_size() / (w_in * h_in * c_in);
+
+    // Input strides
+    const int stridey_in = w_in;
+    const int stridez_in = stridey_in * h_in;
+    const int stridew_in = stridez_in * c_in;
+
+    // Output strides
+    const int stridey_out = w_out;
+    const int stridez_out = stridey_out * h_out;
+    const int stridew_out = stridez_out * c_out;
 
     for(int n = 0; n < num_batches; ++n)
     {
-        for(int w = 0; w < num_filters; ++w)
+        for(int y = 0; y < h_in; ++y)
         {
-            for(int z = 0; z < num_channels; ++z)
+            for(int x = 0; x < w_in; ++x)
             {
-                // Load the 3x3 tile from the input tensor
-                get_tile(in, input_tile, Coordinates(0, 0, z, w, n));
+                // Load the 4x4 tile across the 16 channels of the input tensor
+                for(int z = 0; z < c_in; ++z)
+                {
+                    input_tile[z] = in[x + (y * stridey_in) + (z * stridez_in) + (n * stridew_in)];
+                }
 
                 // First transformation
                 matrix_multiply(trans_matrix, input_tile, tmp_tile);
@@ -169,24 +264,29 @@ void winograd_filter_transform3x3(const SimpleTensor<T> &in, SimpleTensor<T> &ou
                 // Second transformation
                 matrix_multiply(tmp_tile, trans_matrix_transposed, output_tile);
 
-                // Store the 4x4 output tile across the 16 channels
-                const int output_offset                              = w + z * num_filters;
-                out[output_offset + 0 * num_filters * num_channels]  = output_tile[0 + 0 * 4];
-                out[output_offset + 1 * num_filters * num_channels]  = output_tile[1 + 0 * 4];
-                out[output_offset + 2 * num_filters * num_channels]  = output_tile[2 + 0 * 4];
-                out[output_offset + 3 * num_filters * num_channels]  = output_tile[3 + 0 * 4];
-                out[output_offset + 4 * num_filters * num_channels]  = output_tile[0 + 1 * 4];
-                out[output_offset + 5 * num_filters * num_channels]  = output_tile[1 + 1 * 4];
-                out[output_offset + 6 * num_filters * num_channels]  = output_tile[2 + 1 * 4];
-                out[output_offset + 7 * num_filters * num_channels]  = output_tile[3 + 1 * 4];
-                out[output_offset + 8 * num_filters * num_channels]  = output_tile[0 + 2 * 4];
-                out[output_offset + 9 * num_filters * num_channels]  = output_tile[1 + 2 * 4];
-                out[output_offset + 10 * num_filters * num_channels] = output_tile[2 + 2 * 4];
-                out[output_offset + 11 * num_filters * num_channels] = output_tile[3 + 2 * 4];
-                out[output_offset + 12 * num_filters * num_channels] = output_tile[0 + 3 * 4];
-                out[output_offset + 13 * num_filters * num_channels] = output_tile[1 + 3 * 4];
-                out[output_offset + 14 * num_filters * num_channels] = output_tile[2 + 3 * 4];
-                out[output_offset + 15 * num_filters * num_channels] = output_tile[3 + 3 * 4];
+                // Store the 2x2 output tile
+                const int xo = (y % num_tiles_x) * 2;
+                const int yo = (y / num_tiles_x) * 2;
+                const int zo = x;
+
+                const int output_offset                  = xo + (yo * stridey_out) + (zo * stridez_out) + (n * stridew_out);
+                out[output_offset + 0 * stridey_out + 0] = output_tile[0 + 0 * 2];
+
+                // Check out-of-bound writes
+                if(xo + 1 < w_out)
+                {
+                    out[output_offset + 0 * stridey_out + 1] = output_tile[1 + 0 * 2];
+                }
+
+                if(yo + 1 < h_out)
+                {
+                    out[output_offset + 1 * stridey_out + 0] = output_tile[0 + 1 * 2];
+                }
+
+                if((yo + 1 < h_out) && (xo + 1 < w_out))
+                {
+                    out[output_offset + 1 * stridey_out + 1] = output_tile[1 + 1 * 2];
+                }
             }
         }
     }
@@ -234,8 +334,32 @@ SimpleTensor<T> winograd_filter_transform(const SimpleTensor<T> &in, const Tenso
     return out;
 }
 
+template <typename T>
+SimpleTensor<T> winograd_output_transform(const SimpleTensor<T> &in, const TensorShape &output_shape, const Size2D &kernel_dims, const Size2D &num_tiles)
+{
+    ARM_COMPUTE_ERROR_ON_MSG(in.data_layout() != DataLayout::NCHW, "Only supported NCHW data format");
+    ARM_COMPUTE_ERROR_ON(kernel_dims.width != kernel_dims.height);
+    ARM_COMPUTE_ERROR_ON(in.shape()[1] != num_tiles.area());
+
+    // Create reference
+    SimpleTensor<T> out{ output_shape, in.data_type(), 1 };
+
+    switch(kernel_dims.width)
+    {
+        case 3:
+            winograd_output_transform3x3(in, out, num_tiles.width);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Only supported 3x3 kernel");
+            break;
+    }
+
+    return out;
+}
+
 template SimpleTensor<float> winograd_input_transform(const SimpleTensor<float> &src, const TensorShape &dst_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
 template SimpleTensor<float> winograd_filter_transform(const SimpleTensor<float> &in, const TensorShape &output_shape);
+template SimpleTensor<float> winograd_output_transform(const SimpleTensor<float> &in, const TensorShape &output_shape, const Size2D &kernel_dims, const Size2D &num_tiles);
 } // namespace reference
 } // namespace validation
 } // namespace test