COMPMID-471 Implement Deconvolution on OpenCL
Change-Id: Ie00c6b08a51d30c5ce2637d40ee3d165b8a68686
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110311
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
diff --git a/tests/validation/reference/DeconvolutionLayer.cpp b/tests/validation/reference/DeconvolutionLayer.cpp
index 82c2188..0cf1087 100644
--- a/tests/validation/reference/DeconvolutionLayer.cpp
+++ b/tests/validation/reference/DeconvolutionLayer.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017, 2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -39,26 +39,27 @@
const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &a)
{
// Create reference
+ const int stride_x = info.stride().first;
+ const int stride_y = info.stride().second;
TensorShape scaled_shape = src.shape();
- scaled_shape.set(0, output_shape.x());
- scaled_shape.set(1, output_shape.y());
+ int out_x = src.shape().x() + (src.shape().x() - 1) * (stride_x - 1) + a.first + 2 * info.pad().first;
+ int out_y = src.shape().y() + (src.shape().y() - 1) * (stride_y - 1) + a.second + 2 * info.pad().second;
+ scaled_shape.set(0, out_x);
+ scaled_shape.set(1, out_y);
SimpleTensor<T> scaled{ scaled_shape, src.data_type(), 1, src.fixed_point_position() };
- const int width_in = src.shape().x();
- const int height_in = src.shape().y();
- const int width_scaled = scaled.shape().x();
- const int height_scaled = scaled.shape().y();
- const int num_2d_slices = src.shape().total_size() / (width_in * height_in);
- const float width_ratio = static_cast<float>(width_in) / static_cast<float>(width_scaled);
- const float height_ratio = static_cast<float>(height_in) / static_cast<float>(height_scaled);
- const int ax = a.first; // The number of zeros added to right edge of the input.
- const int ay = a.second; // The number of zeros added to bottom edge of the input.
- const unsigned int kernel_size = weights.shape().x();
- ARM_COMPUTE_ERROR_ON(info.pad().first > (kernel_size - 1));
- const int transposed_convolution_padx = kernel_size - info.pad().first - 1;
- const int transposed_convolution_pady = kernel_size - info.pad().second - 1;
- const int stridex = info.stride().first;
- const int stridey = info.stride().second;
+ const int width_in = src.shape().x();
+ const int height_in = src.shape().y();
+ const int width_scaled = scaled.shape().x();
+ const int height_scaled = scaled.shape().y();
+ const int num_2d_slices = src.shape().total_size() / (width_in * height_in);
+ const int ax = a.first; // The number of zeros added to right edge of the input.
+ const int ay = a.second; // The number of zeros added to top edge of the input.
+ ARM_COMPUTE_ERROR_ON(info.pad().first > (weights.shape().x() - 1));
+
+ ARM_COMPUTE_ERROR_ON_MSG(ax > stride_x - 1, "ax must be smaller than stride_x");
+ ARM_COMPUTE_ERROR_ON_MSG(ay > stride_y - 1, "ay must be smaller than stride_y");
+
for(int j = 0; j < scaled.num_elements(); ++j)
{
scaled[j] = T(0);
@@ -68,34 +69,23 @@
{
const int offset_slice_in = slice * width_in * height_in;
const int offset_slice_out = slice * width_scaled * height_scaled;
- for(int yi = ay; yi < height_scaled; yi += stridey)
+ const int start_x = info.pad().first;
+ const int start_y = ay + info.pad().second;
+ const int end_y = height_scaled - info.pad().second;
+ const int end_x = width_scaled - ax - info.pad().first;
+
+ for(int yi = start_y, in_y = 0; yi < end_y; yi += stride_y, in_y++)
{
- for(int xi = transposed_convolution_padx; xi < width_scaled; xi += stridex)
+ for(int xi = start_x, in_x = 0; xi < end_x; xi += stride_x, in_x++)
{
- const float x_src = (xi + 0.5f) * width_ratio - 0.5f;
- const float y_src = (yi + 0.5f) * height_ratio - 0.5f;
- T *out = scaled.data() + offset_slice_out + xi + yi * width_scaled;
- const bool in_bounds = x_src > -1 && y_src > -1 && x_src < width_in && y_src < height_in;
- const bool in_axy = xi < transposed_convolution_padx || xi >= (width_scaled - ax) // this is checking if the x coordinate is in the padded left/right area
- || yi < ay || yi >= (height_scaled - transposed_convolution_pady); // like above but top and bottom padding in the upscaled XY plane
- if(!in_axy)
- {
- if(in_bounds)
- {
- const int in_scaled_x = (x_src < 0.f) ? static_cast<int>(x_src - 0.5f) : static_cast<int>(x_src + 0.5f);
- const int in_scaled_y = (y_src < 0.f) ? static_cast<int>(y_src - 0.5f) : static_cast<int>(y_src + 0.5f);
- const T *in = src.data() + offset_slice_in + in_scaled_x + in_scaled_y * width_in;
- *out = *in;
- }
- else
- {
- *out = T(0);
- }
- }
+ const T *in = src.data() + offset_slice_in + in_y * width_in + in_x;
+ T *out = scaled.data() + offset_slice_out + xi + yi * width_scaled;
+ *out = *in;
}
}
}
- const PadStrideInfo conv_info(1, 1, 1, 1, DimensionRoundingType::CEIL);
+
+ const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
return convolution_layer(scaled, weights, bias, output_shape, conv_info);
}