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/validation/reference/Winograd.cpp b/tests/validation/reference/Winograd.cpp
index 197d218..5be4fe2 100644
--- a/tests/validation/reference/Winograd.cpp
+++ b/tests/validation/reference/Winograd.cpp
@@ -29,6 +29,7 @@
#include "arm_compute/core/Types.h"
#include <algorithm>
+#include <cmath>
namespace arm_compute
{
@@ -142,12 +143,24 @@
{
{ WinogradKey(std::pair<int, int>(2, 2), std::pair<int, int>(3, 3), WinogradTransformType::INPUT), imatrix2x2_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3), WinogradTransformType::INPUT), imatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(2, 1), std::pair<int, int>(3, 1), WinogradTransformType::INPUT), imatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(4, 1), std::pair<int, int>(3, 1), WinogradTransformType::INPUT), imatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 2), std::pair<int, int>(1, 3), WinogradTransformType::INPUT), imatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 4), std::pair<int, int>(1, 3), WinogradTransformType::INPUT), imatrix4x4_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5), WinogradTransformType::INPUT), imatrix4x4_5x5 },
{ WinogradKey(std::pair<int, int>(2, 2), std::pair<int, int>(3, 3), WinogradTransformType::FILTER), fmatrix2x2_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3), WinogradTransformType::FILTER), fmatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(2, 1), std::pair<int, int>(3, 1), WinogradTransformType::FILTER), fmatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(4, 1), std::pair<int, int>(3, 1), WinogradTransformType::FILTER), fmatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 2), std::pair<int, int>(1, 3), WinogradTransformType::FILTER), fmatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 4), std::pair<int, int>(1, 3), WinogradTransformType::FILTER), fmatrix4x4_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5), WinogradTransformType::FILTER), fmatrix4x4_5x5 },
{ WinogradKey(std::pair<int, int>(2, 2), std::pair<int, int>(3, 3), WinogradTransformType::OUTPUT), omatrix2x2_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3), WinogradTransformType::OUTPUT), omatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(2, 1), std::pair<int, int>(3, 1), WinogradTransformType::OUTPUT), omatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(4, 1), std::pair<int, int>(3, 1), WinogradTransformType::OUTPUT), omatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 2), std::pair<int, int>(1, 3), WinogradTransformType::OUTPUT), omatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 4), std::pair<int, int>(1, 3), WinogradTransformType::OUTPUT), omatrix4x4_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5), WinogradTransformType::OUTPUT), omatrix4x4_5x5 },
};
@@ -175,6 +188,20 @@
} // namespace
template <typename T>
+void print_tile(SimpleTensor<T> &in)
+{
+ for(int y = 0; y < in.shape()[1]; y++)
+ {
+ for(int x = 0; x < in.shape()[0]; x++)
+ {
+ std::cout << in[x + y * in.shape()[0]] << " ";
+ }
+
+ std::cout << std::endl;
+ }
+}
+
+template <typename T>
SimpleTensor<T> winograd_input_transform(const SimpleTensor<T> &in, const TensorShape &output_shape, const WinogradInfo &winograd_info)
{
ARM_COMPUTE_ERROR_ON(in.data_layout() != DataLayout::NCHW);
@@ -189,7 +216,10 @@
const unsigned int tile_w = output_tile_size.width + kernel_size.width - 1;
const unsigned int tile_h = output_tile_size.height + kernel_size.height - 1;
- TensorShape tile_dims(tile_w, tile_h);
+ // Get the maximum dimension from the tile size
+ const unsigned int tile_max_dim = std::max(tile_w, tile_h);
+
+ TensorShape tile_dims(tile_max_dim, tile_max_dim);
// Simple tensor for the input tile
SimpleTensor<T> src_tile{ tile_dims, in.data_type() };
@@ -217,11 +247,46 @@
const int in_d = in.shape().z();
const int out_d = out.shape().z();
const int num_batches = in.shape().total_size() / (in_w * in_h * in_d);
- const int num_tiles_x = std::ceil((in_w - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right()) / static_cast<float>(output_tile_size.width));
- const int num_tiles_y = std::ceil((in_h - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom()) / static_cast<float>(output_tile_size.height));
const int step_x = output_tile_size.width;
const int step_y = output_tile_size.height;
+ // Compute the number of output tiles along the x and y direction of size "output_tile_size"
+ const Size2D num_tiles = compute_winograd_convolution_tiles(Size2D(in_w, in_h),
+ kernel_size,
+ output_tile_size,
+ conv_info);
+
+ const int num_tiles_x = num_tiles.width;
+ const int num_tiles_y = num_tiles.height;
+
+ // In case of 1D convolution, the input tile has to be partially filled with zeros
+ int start_x_zero = 0;
+ int start_y_zero = 0;
+ int end_x_zero = 0;
+ int end_y_zero = 0;
+
+ if(output_tile_size.width == 1)
+ {
+ start_x_zero = 1;
+ start_y_zero = 0;
+ end_x_zero = tile_max_dim - 1;
+ end_y_zero = tile_max_dim;
+ }
+ else if(output_tile_size.height == 1)
+ {
+ start_x_zero = 0;
+ start_y_zero = 1;
+ end_x_zero = tile_max_dim;
+ end_y_zero = tile_max_dim - 1;
+ }
+
+ // Set the anchor and shape of the zeros area
+ const Coordinates anchor_zeros(start_x_zero, start_y_zero);
+ const TensorShape shape_zeros(end_x_zero, end_y_zero);
+
+ // If we have a vertical filter (i.e. 1x3, 1x5,..), we need to take the elements along the y direction (step = width of the output tile)
+ const int step_y_transf_tile = kernel_size.width == 1 ? tile_max_dim : 1;
+
ARM_COMPUTE_ERROR_ON((num_tiles_x * num_tiles_y) != static_cast<int>(out.shape().y()));
for(int b = 0; b < num_batches; ++b)
@@ -238,6 +303,9 @@
// Get the tile from the input tensor
get_tile(in, src_tile, Coordinates(xi, yi, z, b));
+ // Fill partially with zeros in case of 1D convolution
+ zeros(src_tile, anchor_zeros, shape_zeros);
+
// Compute the transformation
matrix_multiply(matrix, src_tile, tmp_tile);
matrix_multiply(tmp_tile, matrix_transposed, dst_tile);
@@ -247,7 +315,7 @@
{
int xo = z;
int yo = x + y * num_tiles_x;
- out[coords2index(out.shape(), Coordinates(xo, yo, i, b))] = dst_tile[i];
+ out[coords2index(out.shape(), Coordinates(xo, yo, i, b))] = dst_tile[i * step_y_transf_tile];
}
}
}
@@ -268,27 +336,31 @@
const Size2D output_tile_size = winograd_info.output_tile_size;
const Size2D kernel_size = winograd_info.kernel_size;
- TensorShape kernel_tile_dims(kernel_size.width, kernel_size.height);
-
// Calculate dimensions for the tile
const unsigned int input_tile_w = output_tile_size.width + kernel_size.width - 1;
const unsigned int input_tile_h = output_tile_size.height + kernel_size.height - 1;
const unsigned int input_tile_area = input_tile_w * input_tile_h;
+ // Get the maximum dimension from the filter size
+ const unsigned int kernel_max_dim = std::max(kernel_size.width, kernel_size.height);
+
+ // Get the maximum dimension from the input tile
+ const unsigned int input_tile_max_dim = std::max(input_tile_w, input_tile_h);
+
// Simple tensor for the input tile
- SimpleTensor<T> input_tile{ kernel_tile_dims, in.data_type(), 1 };
+ SimpleTensor<T> input_tile{ TensorShape(kernel_max_dim, kernel_max_dim), in.data_type(), 1 };
// Simple tensor for the transformation matrix
- SimpleTensor<T> trans_matrix{ TensorShape(kernel_tile_dims[0], input_tile_w), in.data_type(), 1 };
+ SimpleTensor<T> trans_matrix{ TensorShape(kernel_max_dim, input_tile_max_dim), in.data_type(), 1 };
// Simple tensor for the transformation matrix transpose
- SimpleTensor<T> trans_matrix_transposed{ TensorShape(input_tile_w, kernel_tile_dims[0]), in.data_type(), 1 };
+ SimpleTensor<T> trans_matrix_transposed{ TensorShape(input_tile_max_dim, kernel_max_dim), in.data_type(), 1 };
// Simple tensor for the temporary tile
- SimpleTensor<T> tmp_tile{ TensorShape(kernel_tile_dims[0], input_tile_w), in.data_type(), 1 };
+ SimpleTensor<T> tmp_tile{ TensorShape(kernel_max_dim, input_tile_max_dim), in.data_type(), 1 };
// Simple tensor for the output tile
- SimpleTensor<T> transf_tile{ TensorShape(input_tile_w, input_tile_w), in.data_type(), 1 };
+ SimpleTensor<T> transf_tile{ TensorShape(input_tile_max_dim, input_tile_max_dim), in.data_type(), 1 };
// Initialize matrix for the filter transform
initialize_matrix_transform(trans_matrix, output_tile_size, kernel_size, WinogradTransformType::FILTER);
@@ -300,6 +372,9 @@
const int num_filters = in.shape()[3];
const int num_batches = in.shape().total_size() / (kernel_size.area() * num_channels * num_filters);
+ // If we have a vertical filter (i.e. 1x3, 1x5,..), we need to take the elements along the y direction (step_y_transf_tile = width of the output tile)
+ const int step_y_transf_tile = kernel_size.width == 1 ? input_tile_max_dim : 1;
+
for(int n = 0; n < num_batches; ++n)
{
for(int w = 0; w < num_filters; ++w)
@@ -321,7 +396,7 @@
// Store the values across the channels
for(unsigned int i = 0; i < input_tile_area; ++i)
{
- out[output_offset + i * num_filters * num_channels] = transf_tile[i];
+ out[output_offset + i * num_filters * num_channels] = transf_tile[i * step_y_transf_tile];
}
}
}
@@ -350,15 +425,19 @@
ARM_COMPUTE_ERROR_ON(in.shape()[2] != (in_tile_w * in_tile_h));
ARM_COMPUTE_ERROR_ON(in.shape()[0] != out.shape()[get_data_layout_dimension_index(winograd_info.output_data_layout, DataLayoutDimension::CHANNEL)]);
+ // Get the maximum dimension from the tile size
+ const unsigned int in_tile_max_dim = std::max(in_tile_w, in_tile_h);
+ const unsigned int out_tile_max_dim = std::max(output_tile_size.width, output_tile_size.height);
+
// Compute tile dimensions
// Input tile dimensions
- TensorShape in_tile_dims(in_tile_w, in_tile_h);
+ TensorShape in_tile_dims(in_tile_max_dim, in_tile_max_dim);
// Output tile dimensions
- TensorShape out_tile_dims(output_tile_size.width, output_tile_size.height);
+ TensorShape out_tile_dims(out_tile_max_dim, out_tile_max_dim);
// Transformation matrix dimensions
- TensorShape tr_tile_dims(in_tile_w, output_tile_size.width);
+ TensorShape tr_tile_dims(in_tile_max_dim, out_tile_max_dim);
// Create tensors
// Simple tensor for the input tile
@@ -400,15 +479,24 @@
const int stridez_out = stridey_out * h_out;
const int stridew_out = stridez_out * c_out;
- // Compute number of elements to process in the X and Y direction
- const int num_elements_x = input_dimensions.width - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right();
- const int num_elements_y = input_dimensions.height - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom();
- const int num_tiles_x = std::ceil(num_elements_x / static_cast<float>(output_tile_size.width));
- const int num_tiles_y = std::ceil(num_elements_y / static_cast<float>(output_tile_size.height));
+ // Compute the number of output tiles along the x and y direction of size "output_tile_size"
+ const Size2D num_tiles = compute_winograd_convolution_tiles(Size2D(input_dimensions.width, input_dimensions.height),
+ kernel_size,
+ output_tile_size,
+ conv_info);
+
+ const int num_tiles_x = num_tiles.width;
+ const int num_tiles_y = num_tiles.height;
ARM_COMPUTE_UNUSED(num_tiles_y);
ARM_COMPUTE_ERROR_ON(in.shape()[1] != static_cast<unsigned int>(num_tiles_x * num_tiles_y));
+ // If we have a vertical filter (i.e. 1x3, 1x5,..), we still need to take the elements along the x direction (step_y_transf_tile = 1)
+ const int step_y_transf_tile = kernel_size.width == 1 ? 1 : output_tile.shape()[0];
+
+ // Initialize with zeros the input tile
+ zeros(input_tile, Coordinates(0, 0), input_tile.shape());
+
for(int n = 0; n < num_batches; ++n)
{
for(int y = 0; y < h_in; ++y)
@@ -441,7 +529,7 @@
// Check out-of-bound writes
if((xo + xi < w_out) && (yo + yi < h_out))
{
- out[output_offset + yi * stridey_out + xi] = output_tile[xi + yi * out_tile_w];
+ out[output_offset + yi * stridey_out + xi] = output_tile[xi + yi * step_y_transf_tile];
// Add bias
out[output_offset + yi * stridey_out + xi] += b[zo];