COMPMID-2605: Add asymmetric padding support for Deconvolution layer

Change-Id: I63b773bdce25f1342ccd3a08ded623a1508f70fe
Signed-off-by: Matthew Jackson <matthew.jackson@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1797
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>

13 files changed, 218 insertions, 109 deletions

diff --git a/arm_compute/core/Utils.h b/arm_compute/core/Utils.h
index 6ffab4b9ea..0ce2ee0161 100644
--- a/arm_compute/core/Utils.h
+++ b/arm_compute/core/Utils.h
@@ -853,21 +853,17 @@ PadStrideInfo calculate_same_pad(TensorShape input_shape, TensorShape weights_sh
 
 /** Returns expected width and height of the deconvolution's output tensor.
  *
- * @param[in] in_width      Width of input tensor (Number of columns)
- * @param[in] in_height     Height of input tensor (Number of rows)
- * @param[in] kernel_width  Kernel width.
- * @param[in] kernel_height Kernel height.
- * @param[in] padx          X axis padding.
- * @param[in] pady          Y axis padding.
- * @param[in] stride_x      X axis input stride.
- * @param[in] stride_y      Y axis input stride.
+ * @param[in] in_width        Width of input tensor (Number of columns)
+ * @param[in] in_height       Height of input tensor (Number of rows)
+ * @param[in] kernel_width    Kernel width.
+ * @param[in] kernel_height   Kernel height.
+ * @param[in] pad_stride_info Pad and stride information.
  *
  * @return A pair with the new width in the first position and the new height in the second.
  */
 std::pair<unsigned int, unsigned int> deconvolution_output_dimensions(unsigned int in_width, unsigned int in_height,
                                                                       unsigned int kernel_width, unsigned int kernel_height,
-                                                                      unsigned int padx, unsigned int pady,
-                                                                      unsigned int stride_x, unsigned int stride_y);
+                                                                      const PadStrideInfo &pad_stride_info);
 
 /** Returns expected width and height of output scaled tensor depending on dimensions rounding mode.
  *
diff --git a/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp b/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp
index 50f654680c..4ae9cabd1f 100644
--- a/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp
+++ b/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp
@@ -56,7 +56,6 @@ Status CLDeconvolutionLayerUpsampleKernel::validate(const ITensorInfo *input, co
 
     ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(idx_w) == 0);
     ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(idx_h) == 0);
-    ARM_COMPUTE_RETURN_ERROR_ON(!info.padding_is_symmetric());
 
     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_c) != output->dimension(idx_c));
     for(size_t i = 3; i < Coordinates::num_max_dimensions; ++i)
@@ -104,12 +103,12 @@ void CLDeconvolutionLayerUpsampleKernel::run(const Window &window, cl::CommandQu
     const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
     const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
 
-    const int out_start_x = _info.pad().first;
-    const int out_end_x   = _output->info()->dimension(idx_w) - _info.pad().first + _info.stride().first - 1;
+    const int out_start_x = _info.pad_left();
+    const int out_end_x   = _output->info()->dimension(idx_w) - _info.pad_right() + _info.stride().first - 1;
     const int out_step_x  = _info.stride().first;
 
-    const int out_start_y = _info.pad().second;
-    const int out_end_y   = _output->info()->dimension(idx_h) - _info.pad().second + _info.stride().second - 1;
+    const int out_start_y = _info.pad_top();
+    const int out_end_y   = _output->info()->dimension(idx_h) - _info.pad_bottom() + _info.stride().second - 1;
     const int out_step_y  = _info.stride().second;
 
     switch(data_layout)
diff --git a/src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.cpp b/src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.cpp
index 71218f5b52..69e5eff213 100644
--- a/src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.cpp
+++ b/src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.cpp
@@ -40,8 +40,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output, input_info, weights_info);
     const DataLayout   data_layout = input_info->data_layout();
-    const unsigned int stride_x    = deconv_info.stride().first;
-    const unsigned int stride_y    = deconv_info.stride().second;
 
     const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
     const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
@@ -77,8 +75,8 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
 
     if(output->total_size() != 0)
     {
-        auto out_dims = deconvolution_output_dimensions(input_info->dimension(idx_w), input_info->dimension(idx_h), weights_info->dimension(idx_w), weights_info->dimension(idx_h),
-                                                        0, 0, stride_x, stride_y);
+        const PadStrideInfo stride_info(deconv_info.stride().first, deconv_info.stride().second);
+        auto out_dims = deconvolution_output_dimensions(input_info->dimension(idx_w), input_info->dimension(idx_h), weights_info->dimension(idx_w), weights_info->dimension(idx_h), stride_info);
 
         const TensorShape output_shape = misc::shape_calculator::compute_deconvolution_output_shape(out_dims, *input_info, *weights_info);
 
@@ -92,14 +90,11 @@ std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *input
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
     const DataLayout data_layout = input_info->data_layout();
+    const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+    const PadStrideInfo stride_info(deconv_info.stride().first, deconv_info.stride().second);
 
-    const unsigned int stride_x = deconv_info.stride().first;
-    const unsigned int stride_y = deconv_info.stride().second;
-    const size_t       idx_w    = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const size_t       idx_h    = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-
-    auto out_dims = deconvolution_output_dimensions(input_info->dimension(idx_w), input_info->dimension(idx_h), weights_info->dimension(idx_w), weights_info->dimension(idx_h),
-                                                    0, 0, stride_x, stride_y);
+    auto out_dims = deconvolution_output_dimensions(input_info->dimension(idx_w), input_info->dimension(idx_h), weights_info->dimension(idx_w), weights_info->dimension(idx_h), stride_info);
 
     const TensorShape output_shape = misc::shape_calculator::compute_deconvolution_output_shape(out_dims, *input_info, *weights_info);
 
diff --git a/src/core/CPP/kernels/CPPUpsampleKernel.cpp b/src/core/CPP/kernels/CPPUpsampleKernel.cpp
index ad2d54a0f8..e63808f80e 100644
--- a/src/core/CPP/kernels/CPPUpsampleKernel.cpp
+++ b/src/core/CPP/kernels/CPPUpsampleKernel.cpp
@@ -76,10 +76,10 @@ void CPPUpsampleKernel::run(const Window &window, const ThreadInfo &info)
     const int    height_scaled = _output->info()->dimension(1);
     const int    stride_x      = _info.stride().first;
     const int    stride_y      = _info.stride().second;
-    const int    start_x       = _info.pad().first;
-    const int    start_y       = _info.pad().second;
-    const int    end_y         = height_scaled - _info.pad().second;
-    const int    end_x         = width_scaled - _info.pad().first;
+    const int    start_x       = _info.pad_left();
+    const int    start_y       = _info.pad_top();
+    const int    end_x         = width_scaled - _info.pad_right();
+    const int    end_y         = height_scaled - _info.pad_bottom();
     const size_t element_size  = _input->info()->element_size();
 
     //The fill value is normally 0, but for QASYMM8 the '0' corresponds to the offset
diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp
index 122373f5c6..d11788acd3 100644
--- a/src/core/Utils.cpp
+++ b/src/core/Utils.cpp
@@ -374,15 +374,22 @@ PadStrideInfo arm_compute::calculate_same_pad(TensorShape input_shape, TensorSha
     return same_info;
 }
 
-std::pair<unsigned int, unsigned int> arm_compute::deconvolution_output_dimensions(
-    unsigned int in_width, unsigned int in_height, unsigned int kernel_width, unsigned int kernel_height, unsigned int padx, unsigned int pady,
-    unsigned int stride_x, unsigned int stride_y)
+std::pair<unsigned int, unsigned int> arm_compute::deconvolution_output_dimensions(unsigned int in_width, unsigned int in_height,
+                                                                                   unsigned int kernel_width, unsigned int kernel_height,
+                                                                                   const PadStrideInfo &pad_stride_info)
 {
+    const unsigned int pad_left   = pad_stride_info.pad_left();
+    const unsigned int pad_top    = pad_stride_info.pad_top();
+    const unsigned int pad_right  = pad_stride_info.pad_right();
+    const unsigned int pad_bottom = pad_stride_info.pad_bottom();
+    const unsigned int stride_x   = pad_stride_info.stride().first;
+    const unsigned int stride_y   = pad_stride_info.stride().second;
+
     ARM_COMPUTE_ERROR_ON(in_width < 1 || in_height < 1);
-    ARM_COMPUTE_ERROR_ON(((in_width - 1) * stride_x + kernel_width) < 2 * padx);
-    ARM_COMPUTE_ERROR_ON(((in_height - 1) * stride_y + kernel_height) < 2 * pady);
-    const int w = stride_x * (in_width - 1) + kernel_width - 2 * padx;
-    const int h = stride_y * (in_height - 1) + kernel_height - 2 * pady;
+    ARM_COMPUTE_ERROR_ON(((in_width - 1) * stride_x + kernel_width) < (pad_left + pad_right));
+    ARM_COMPUTE_ERROR_ON(((in_height - 1) * stride_y + kernel_height) < (pad_top + pad_bottom));
+    const int w = stride_x * (in_width - 1) + kernel_width - (pad_left + pad_right);
+    const int h = stride_y * (in_height - 1) + kernel_height - (pad_top + pad_bottom);
 
     return std::make_pair<unsigned int, unsigned int>(w, h);
 }
diff --git a/src/graph/nodes/DeconvolutionLayerNode.cpp b/src/graph/nodes/DeconvolutionLayerNode.cpp
index 28c75297a5..d4a5b769e7 100644
--- a/src/graph/nodes/DeconvolutionLayerNode.cpp
+++ b/src/graph/nodes/DeconvolutionLayerNode.cpp
@@ -56,10 +56,7 @@ TensorDescriptor DeconvolutionLayerNode::compute_output_descriptor(const TensorD
     const unsigned int kernel_width  = get_dimension_size(weights_descriptor, DataLayoutDimension::WIDTH);
     const unsigned int kernel_height = get_dimension_size(weights_descriptor, DataLayoutDimension::HEIGHT);
 
-    std::tie(output_width, output_height) = deconvolution_output_dimensions(input_width, input_height,
-                                                                            kernel_width, kernel_height,
-                                                                            info.pad().first, info.pad().second,
-                                                                            info.stride().first, info.stride().second);
+    std::tie(output_width, output_height) = deconvolution_output_dimensions(input_width, input_height, kernel_width, kernel_height, info);
 
     const DataLayout data_layout       = input_descriptor.layout;
     TensorDescriptor output_descriptor = input_descriptor;
diff --git a/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp
index c1a39ef26a..b8089d8229 100644
--- a/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp
@@ -63,13 +63,8 @@ Status CLDirectDeconvolutionLayer::validate(const ITensorInfo *input, const ITen
 
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) != weights->dimension(idx_h));
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) < 1);
-    ARM_COMPUTE_RETURN_ERROR_ON(!info.padding_is_symmetric());
 
-    const unsigned int stride_x = info.stride().first;
-    const unsigned int stride_y = info.stride().second;
-
-    auto out_dims = deconvolution_output_dimensions(input->dimension(idx_w), input->dimension(idx_h), weights->dimension(idx_w), weights->dimension(idx_h),
-                                                    info.pad().first, info.pad().second, stride_x, stride_y);
+    auto out_dims = deconvolution_output_dimensions(input->dimension(idx_w), input->dimension(idx_h), weights->dimension(idx_w), weights->dimension(idx_h), info);
 
     const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input, *weights);
 
@@ -92,9 +87,11 @@ Status CLDirectDeconvolutionLayer::validate(const ITensorInfo *input, const ITen
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_h) != output_shape[idx_h], "Output's height is invalid.");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_c) != output_shape[idx_c], "Output's depth is invalid.");
 
-    unsigned int        padx            = 0;
-    unsigned int        pady            = 0;
-    const TensorShape   scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, out_dims, padx, pady);
+    unsigned int        deconv_pad_x = 0;
+    unsigned int        deconv_pad_y = 0;
+    const unsigned int  stride_x = info.stride().first;
+    const unsigned int  stride_y = info.stride().second;
+    const TensorShape   scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y);
     TensorInfo          scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(scale_out_shape).set_data_layout(data_layout));
     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
 
@@ -109,6 +106,10 @@ void CLDirectDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights,
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 
+    const unsigned int pad_left   = info.pad_left();
+    const unsigned int pad_right  = info.pad_right();
+    const unsigned int pad_top    = info.pad_top();
+    const unsigned int pad_bottom = info.pad_bottom();
     const unsigned int stride_x = info.stride().first;
     const unsigned int stride_y = info.stride().second;
 
@@ -122,8 +123,7 @@ void CLDirectDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights,
     _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout));
     _flip_weights.configure(weights, &_weights_flipped, &_flip_axis);
 
-    auto out_dims = deconvolution_output_dimensions(input->info()->dimension(idx_w), input->info()->dimension(idx_h), weights->info()->dimension(idx_w), weights->info()->dimension(idx_h),
-                                                    info.pad().first, info.pad().second, stride_x, stride_y);
+    auto out_dims = deconvolution_output_dimensions(input->info()->dimension(idx_w), input->info()->dimension(idx_h), weights->info()->dimension(idx_w), weights->info()->dimension(idx_h), info);
 
     const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input->info(), *weights->info());
 
@@ -138,16 +138,30 @@ void CLDirectDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights,
     _memory_group.manage(&_scaled_output);
 
     // Find the upsampled dimensions and the padding needed for the convolution with stride 1 in order to match output shape
-    unsigned int      padx            = 0;
-    unsigned int      pady            = 0;
-    const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(), stride_x, stride_y, out_dims, padx, pady);
+    unsigned int      deconv_pad_x    = 0;
+    unsigned int      deconv_pad_y    = 0;
+    const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(), stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y);
+
+    unsigned int deconv_pad_left  = pad_right > pad_left ? pad_right - pad_left : 0;
+    unsigned int deconv_pad_right = pad_left > pad_right ? pad_left - pad_right : 0;
+    deconv_pad_x -= deconv_pad_left + deconv_pad_right;
+    ARM_COMPUTE_ERROR_ON((deconv_pad_x % 2) != 0);
+    deconv_pad_left  += deconv_pad_x / 2;
+    deconv_pad_right += deconv_pad_x / 2;
+
+    unsigned int deconv_pad_top    = pad_bottom > pad_top ? pad_bottom - pad_top : 0;
+    unsigned int deconv_pad_bottom = pad_top > pad_bottom ? pad_top - pad_bottom : 0;
+    deconv_pad_y -= deconv_pad_top + deconv_pad_bottom;
+    ARM_COMPUTE_ERROR_ON((deconv_pad_y % 2) != 0);
+    deconv_pad_top    += deconv_pad_y / 2;
+    deconv_pad_bottom += deconv_pad_y / 2;
 
     TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
     scale_out_info.set_data_layout(data_layout);
     _scaled_output.allocator()->init(scale_out_info);
 
     // configure scale function
-    const PadStrideInfo upsample_info(stride_x, stride_y, padx / 2, pady / 2);
+    const PadStrideInfo upsample_info(stride_x, stride_y, deconv_pad_left, deconv_pad_right, deconv_pad_top, deconv_pad_bottom, DimensionRoundingType::FLOOR);
     _scale_f.configure(input, &_scaled_output, upsample_info);
 
     // Setup the function to convolve the upscaled output
diff --git a/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
index 36a120e4ef..78e1ae7f05 100644
--- a/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
@@ -151,8 +151,8 @@ Status CLGEMMDeconvolutionLayer::validate(const ITensorInfo *input, const ITenso
         ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(&input->clone()->set_tensor_shape(nhwc_input_shape).set_is_resizable(true), &reshaped_t_info, nullptr, &gemm_output_info, 1.0f, 0.0f, gemm_info));
     }
 
-    auto out_dims = deconvolution_output_dimensions(input->dimension(idx_w), input->dimension(idx_h), weights->dimension(idx_w), weights->dimension(idx_h),
-                                                    0, 0, deconv_info.stride().first, deconv_info.stride().second);
+    const PadStrideInfo stride_info(deconv_info.stride().first, deconv_info.stride().second);
+    auto out_dims = deconvolution_output_dimensions(input->dimension(idx_w), input->dimension(idx_h), weights->dimension(idx_w), weights->dimension(idx_h), stride_info);
     const TensorShape deconv_shape       = misc::shape_calculator::compute_deconvolution_output_shape(out_dims, *input, *weights);
     TensorInfo        col2im_output_info = gemm_output_info.clone()->set_tensor_shape(deconv_shape).set_is_resizable(true);
 
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
index 1f2cc3d73b..bbb91b4651 100644
--- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
@@ -64,13 +64,8 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
     const unsigned int height_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::HEIGHT);
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(width_idx) != weights->dimension(height_idx));
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(width_idx) < 1);
-    ARM_COMPUTE_RETURN_ERROR_ON(!info.padding_is_symmetric());
 
-    const unsigned int stride_x = info.stride().first;
-    const unsigned int stride_y = info.stride().second;
-
-    auto out_dims = deconvolution_output_dimensions(input->dimension(width_idx), input->dimension(height_idx), weights->dimension(width_idx), weights->dimension(height_idx),
-                                                    info.pad().first, info.pad().second, stride_x, stride_y);
+    auto out_dims = deconvolution_output_dimensions(input->dimension(width_idx), input->dimension(height_idx), weights->dimension(width_idx), weights->dimension(height_idx), info);
 
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     if(bias != nullptr)
@@ -96,9 +91,11 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid.");
     }
 
-    unsigned int        padx            = 0;
-    unsigned int        pady            = 0;
-    const TensorShape   scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, out_dims, padx, pady);
+    unsigned int        deconv_pad_x = 0;
+    unsigned int        deconv_pad_y = 0;
+    const unsigned int  stride_x = info.stride().first;
+    const unsigned int  stride_y = info.stride().second;
+    const TensorShape   scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y);
     TensorInfo          scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(scale_out_shape));
     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
 
@@ -126,14 +123,17 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con
     _is_prepared      = false;
     _is_nchw          = data_layout == DataLayout::NCHW;
 
+    const unsigned int pad_left   = info.pad_left();
+    const unsigned int pad_right  = info.pad_right();
+    const unsigned int pad_top    = info.pad_top();
+    const unsigned int pad_bottom = info.pad_bottom();
     const unsigned int stride_x = info.stride().first;
     const unsigned int stride_y = info.stride().second;
 
     const unsigned int width_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
     const unsigned int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    auto               out_dims   = deconvolution_output_dimensions(input->info()->dimension(width_idx), input->info()->dimension(height_idx), weights->info()->dimension(width_idx),
-                                                                    weights->info()->dimension(height_idx),
-                                                                    info.pad().first, info.pad().second, stride_x, stride_y);
+    auto               out_dims   = deconvolution_output_dimensions(input->info()->dimension(width_idx), input->info()->dimension(height_idx),
+                                                                    weights->info()->dimension(width_idx), weights->info()->dimension(height_idx), info);
 
     const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input->info(), *weights->info());
     // Output auto initialization if not yet initialized
@@ -157,16 +157,30 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con
         _permuted_weights.info()->set_data_layout(DataLayout::NCHW);
 
         // Find the upsampled dimensions and the padding needed for the convolution with stride 1 in order to match output shape
-        unsigned int      padx            = 0;
-        unsigned int      pady            = 0;
-        const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*_permuted_input.info(), *_permuted_weights.info(), stride_x, stride_y, out_dims, padx,
-                                                                                  pady);
+        unsigned int      deconv_pad_x    = 0;
+        unsigned int      deconv_pad_y    = 0;
+        const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*_permuted_input.info(), *_permuted_weights.info(), stride_x, stride_y, out_dims,
+                                                                                  deconv_pad_x, deconv_pad_y);
+
+        unsigned int deconv_pad_left  = pad_right > pad_left ? pad_right - pad_left : 0;
+        unsigned int deconv_pad_right = pad_left > pad_right ? pad_left - pad_right : 0;
+        deconv_pad_x -= deconv_pad_left + deconv_pad_right;
+        ARM_COMPUTE_ERROR_ON((deconv_pad_x % 2) != 0);
+        deconv_pad_left  += deconv_pad_x / 2;
+        deconv_pad_right += deconv_pad_x / 2;
+
+        unsigned int deconv_pad_top    = pad_bottom > pad_top ? pad_bottom - pad_top : 0;
+        unsigned int deconv_pad_bottom = pad_top > pad_bottom ? pad_top - pad_bottom : 0;
+        deconv_pad_y -= deconv_pad_top + deconv_pad_bottom;
+        ARM_COMPUTE_ERROR_ON((deconv_pad_y % 2) != 0);
+        deconv_pad_top    += deconv_pad_y / 2;
+        deconv_pad_bottom += deconv_pad_y / 2;
 
         TensorInfo scale_out_info(scale_out_shape, 1, _permuted_input.info()->data_type(), _permuted_input.info()->quantization_info());
         scale_out_info.set_data_layout(DataLayout::NCHW);
         _scaled_output.allocator()->init(scale_out_info);
 
-        const PadStrideInfo upsample_info(stride_x, stride_y, padx / 2, pady / 2);
+        const PadStrideInfo upsample_info(stride_x, stride_y, deconv_pad_left, deconv_pad_right, deconv_pad_top, deconv_pad_bottom, DimensionRoundingType::FLOOR);
         _upsample_f.configure(&_permuted_input, &_scaled_output, upsample_info);
 
         _weights_flipped.allocator()->init(*_permuted_weights.info()->clone());
@@ -189,14 +203,30 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con
     else
     {
         // Find the upsampled dimensions and the padding needed for the convolution with stride 1 in order to match output shape
-        unsigned int      padx            = 0;
-        unsigned int      pady            = 0;
-        const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(), stride_x, stride_y, out_dims, padx, pady);
+        unsigned int      deconv_pad_x    = 0;
+        unsigned int      deconv_pad_y    = 0;
+        const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(), stride_x, stride_y,
+                                                                                  out_dims, deconv_pad_x, deconv_pad_y);
+
+        unsigned int deconv_pad_left  = pad_right > pad_left ? pad_right - pad_left : 0;
+        unsigned int deconv_pad_right = pad_left > pad_right ? pad_left - pad_right : 0;
+        deconv_pad_x -= deconv_pad_left + deconv_pad_right;
+        ARM_COMPUTE_ERROR_ON((deconv_pad_x % 2) != 0);
+        deconv_pad_left  += deconv_pad_x / 2;
+        deconv_pad_right += deconv_pad_x / 2;
+
+        unsigned int deconv_pad_top    = pad_bottom > pad_top ? pad_bottom - pad_top : 0;
+        unsigned int deconv_pad_bottom = pad_top > pad_bottom ? pad_top - pad_bottom : 0;
+        deconv_pad_y -= deconv_pad_top + deconv_pad_bottom;
+        ARM_COMPUTE_ERROR_ON((deconv_pad_y % 2) != 0);
+        deconv_pad_top    += deconv_pad_y / 2;
+        deconv_pad_bottom += deconv_pad_y / 2;
 
         TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
         scale_out_info.set_data_layout(data_layout);
         _scaled_output.allocator()->init(scale_out_info);
-        const PadStrideInfo upsample_info(stride_x, stride_y, padx / 2, pady / 2);
+
+        const PadStrideInfo upsample_info(stride_x, stride_y, deconv_pad_left, deconv_pad_right, deconv_pad_top, deconv_pad_bottom, DimensionRoundingType::FLOOR);
         _upsample_f.configure(input, &_scaled_output, upsample_info);
 
         _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout));
diff --git a/tests/validation/CL/DeconvolutionLayer.cpp b/tests/validation/CL/DeconvolutionLayer.cpp
index 44b3428c52..9dafd1ef89 100644
--- a/tests/validation/CL/DeconvolutionLayer.cpp
+++ b/tests/validation/CL/DeconvolutionLayer.cpp
@@ -55,6 +55,9 @@ const auto data4x4 = datasets::SmallDeconvolutionShapes() * framework::dataset::
 const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 2)
                      * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("NumKernels", { 3 });
 
+const auto data3x3_asymm = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 2) * framework::dataset::make("StrideY", 1, 2) * framework::dataset::make("PadLeft", 0, 1)
+                           * framework::dataset::make("PadRight", 0, 1) * framework::dataset::make("PadTop", 0, 1) * framework::dataset::make("PadBottom", 0, 1) * framework::dataset::make("NumKernels", { 3 });
+
 const auto data3x3_precommit = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 2) * framework::dataset::make("StrideY", 1, 2) * framework::dataset::make("PadX", 0, 2)
                                * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("NumKernels", { 3 });
 
@@ -120,16 +123,19 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
 // *INDENT-ON*
 
 template <typename T>
-using CLDeconvolutionLayerFixture4x4 = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 4, 4>;
+using CLDeconvolutionLayerFixture4x4      = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 4, 4>;
+
+template <typename T>
+using CLDeconvolutionLayerFixture3x3      = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 3, 3>;
 
 template <typename T>
-using CLDeconvolutionLayerFixture3x3 = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 3, 3>;
+using CLDeconvolutionLayerAsymmFixture3x3 = DeconvolutionValidationAsymmFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 3, 3>;
 
 template <typename T>
-using CLDeconvolutionLayerFixture2x2 = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 2, 2>;
+using CLDeconvolutionLayerFixture2x2      = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 2, 2>;
 
 template <typename T>
-using CLDeconvolutionLayerFixture1x1 = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 1, 1>;
+using CLDeconvolutionLayerFixture1x1      = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 1, 1>;
 
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
@@ -153,6 +159,14 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerFixture3x3<float>, framewor
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_fp32);
 }
+FIXTURE_DATA_TEST_CASE(RunAsymm, CLDeconvolutionLayerAsymmFixture3x3<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data3x3_asymm, framework::dataset::make("DataType",
+                                                                                                                      DataType::F32)),
+                                                                                                                      data_layouts_dataset),
+                                                                                                                      add_bias_dataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_fp32);
+}
 FIXTURE_DATA_TEST_CASE(RunLarge, CLDeconvolutionLayerFixture3x3<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data3x3, framework::dataset::make("DataType", DataType::F32)),
                                                                                                                  data_layouts_dataset),
                                                                                                                  add_bias_dataset))
diff --git a/tests/validation/NEON/DeconvolutionLayer.cpp b/tests/validation/NEON/DeconvolutionLayer.cpp
index 727f501393..500ef10661 100644
--- a/tests/validation/NEON/DeconvolutionLayer.cpp
+++ b/tests/validation/NEON/DeconvolutionLayer.cpp
@@ -56,6 +56,9 @@ const auto data4x4 = datasets::SmallDeconvolutionShapes() * framework::dataset::
 const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 2)
                      * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("NumKernels", { 3 });
 
+const auto data3x3_asymm = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 2) * framework::dataset::make("StrideY", 1, 2) * framework::dataset::make("PadLeft", 0, 1)
+                           * framework::dataset::make("PadRight", 0, 1) * framework::dataset::make("PadTop", 0, 1) * framework::dataset::make("PadBottom", 0, 1) * framework::dataset::make("NumKernels", { 3 });
+
 const auto data3x3_precommit = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 2) * framework::dataset::make("StrideY", 1, 2) * framework::dataset::make("PadX", 0, 2)
                                * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("NumKernels", { 3 });
 
@@ -74,13 +77,14 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, (combine(datasets::Sm
                input_shape, data_type)
 {
     // Create shapes
-    const unsigned int kernel_size_x = 3;
-    const unsigned int kernel_size_y = 3;
-    const unsigned int num_kernels   = 1;
-    const TensorShape  weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels);
-    const TensorShape  bias_shape(num_kernels);
-    auto               out_dim      = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, 1, 1, 1, 1);
-    TensorShape        output_shape = compute_deconvolution_output_shape(out_dim, TensorInfo(input_shape, 1, data_type), TensorInfo(weights_shape, 1, data_type));
+    const unsigned int  kernel_size_x = 3;
+    const unsigned int  kernel_size_y = 3;
+    const unsigned int  num_kernels   = 1;
+    const TensorShape   weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels);
+    const TensorShape   bias_shape(num_kernels);
+    const PadStrideInfo info(1, 1, 1, 1);
+    auto                out_dim      = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, info);
+    TensorShape         output_shape = compute_deconvolution_output_shape(out_dim, TensorInfo(input_shape, 1, data_type), TensorInfo(weights_shape, 1, data_type));
 
     // Create tensors
     Tensor src     = create_tensor<Tensor>(input_shape, data_type, 1);
@@ -157,13 +161,16 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
 // *INDENT-ON*
 
 template <typename T>
-using NEDeconvolutionLayerFixture4x4 = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 4, 4>;
+using NEDeconvolutionLayerFixture4x4      = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 4, 4>;
+
+template <typename T>
+using NEDeconvolutionLayerFixture3x3      = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 3, 3>;
 
 template <typename T>
-using NEDeconvolutionLayerFixture3x3 = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 3, 3>;
+using NEDeconvolutionLayerAsymmFixture3x3 = DeconvolutionValidationAsymmFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 3, 3>;
 
 template <typename T>
-using NEDeconvolutionLayerFixture1x1 = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 1, 1>;
+using NEDeconvolutionLayerFixture1x1      = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 1, 1>;
 
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
@@ -185,6 +192,14 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEDeconvolutionLayerFixture3x3<float>, framewor
     // Validate output
     validate(Accessor(_target), _reference, tolerance_fp32);
 }
+FIXTURE_DATA_TEST_CASE(RunAsymm, NEDeconvolutionLayerAsymmFixture3x3<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data3x3_asymm, framework::dataset::make("DataType",
+                                                                                                                      DataType::F32)),
+                                                                                                                      data_layouts_dataset),
+                                                                                                                      add_bias_dataset))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fp32);
+}
 FIXTURE_DATA_TEST_CASE(RunLarge, NEDeconvolutionLayerFixture3x3<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data3x3, framework::dataset::make("DataType", DataType::F32)),
                                                                                                                  data_layouts_dataset),
                                                                                                                  add_bias_dataset))
diff --git a/tests/validation/fixtures/DeconvolutionLayerFixture.h b/tests/validation/fixtures/DeconvolutionLayerFixture.h
index 9f90f07c97..a25a65f997 100644
--- a/tests/validation/fixtures/DeconvolutionLayerFixture.h
+++ b/tests/validation/fixtures/DeconvolutionLayerFixture.h
@@ -218,7 +218,27 @@ public:
         const TensorShape   weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels);
         const TensorShape   bias_shape(num_kernels);
         const PadStrideInfo info(sx, sy, padx, pady, DimensionRoundingType::CEIL);
-        auto                out_dim = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, padx, pady, sx, sy);
+        auto                out_dim = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, info);
+        TensorInfo          input_info(input_shape, 1, data_type);
+        TensorInfo          weights_info(weights_shape, 1, data_type);
+        TensorShape         output_shape = compute_deconvolution_output_shape(out_dim, input_info, weights_info);
+        DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, data_type, data_layout, QuantizationInfo(), add_bias);
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T, unsigned int kernel_size_x, unsigned int kernel_size_y>
+class DeconvolutionValidationAsymmFixture : public DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, unsigned int sx, unsigned int sy, unsigned int pad_left, unsigned int pad_right, unsigned int pad_top,
+               unsigned int pad_bottom, unsigned int num_kernels, DataType data_type, DataLayout data_layout, bool add_bias)
+    {
+        ARM_COMPUTE_ERROR_ON_MSG(kernel_size_x != kernel_size_y, "Only square kernels supported");
+        const TensorShape   weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels);
+        const TensorShape   bias_shape(num_kernels);
+        const PadStrideInfo info(sx, sy, pad_left, pad_right, pad_top, pad_bottom, DimensionRoundingType::CEIL);
+        auto                out_dim = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, info);
         TensorInfo          input_info(input_shape, 1, data_type);
         TensorInfo          weights_info(weights_shape, 1, data_type);
         TensorShape         output_shape = compute_deconvolution_output_shape(out_dim, input_info, weights_info);
@@ -238,7 +258,7 @@ public:
         const TensorShape   weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels);
         const TensorShape   bias_shape(num_kernels);
         const PadStrideInfo info(sx, sy, padx, pady, DimensionRoundingType::CEIL);
-        auto                out_dim = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, padx, pady, sx, sy);
+        auto                out_dim = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, info);
         TensorInfo          input_info(input_shape, 1, data_type, quantization_info);
         TensorInfo          weights_info(weights_shape, 1, data_type, quantization_info);
         TensorShape         output_shape = compute_deconvolution_output_shape(out_dim, input_info, weights_info);
diff --git a/tests/validation/reference/DeconvolutionLayer.cpp b/tests/validation/reference/DeconvolutionLayer.cpp
index af59830722..343ea5e725 100644
--- a/tests/validation/reference/DeconvolutionLayer.cpp
+++ b/tests/validation/reference/DeconvolutionLayer.cpp
@@ -38,21 +38,44 @@ SimpleTensor<T> deconvolution_layer(const SimpleTensor<T> &src, const SimpleTens
                                     const PadStrideInfo &info)
 {
     // Create reference
-    const int stride_x           = info.stride().first;
-    const int stride_y           = info.stride().second;
-    const int weights_width      = weights.shape().x();
-    const int weights_height     = weights.shape().y();
-    const int weights_upper_dims = weights.shape().total_size() / (weights_width * weights_height);
+    const unsigned int pad_left   = info.pad_left();
+    const unsigned int pad_right  = info.pad_right();
+    const unsigned int pad_top    = info.pad_top();
+    const unsigned int pad_bottom = info.pad_bottom();
+    const int stride_x            = info.stride().first;
+    const int stride_y            = info.stride().second;
+    const int weights_width       = weights.shape().x();
+    const int weights_height      = weights.shape().y();
+    const int weights_upper_dims  = weights.shape().total_size() / (weights_width * weights_height);
+
+    ARM_COMPUTE_ERROR_ON(pad_left   > (weights.shape().x() - 1));
+    ARM_COMPUTE_ERROR_ON(pad_right  > (weights.shape().x() - 1));
+    ARM_COMPUTE_ERROR_ON(pad_top    > (weights.shape().y() - 1));
+    ARM_COMPUTE_ERROR_ON(pad_bottom > (weights.shape().y() - 1));
 
     // Find the upsampled dimensions
     unsigned int out_x = (src.shape().x() - 1) * stride_x + 1;
     unsigned int out_y = (src.shape().y() - 1) * stride_y + 1;
 
     // Find the padding needed for the convolution with stride 1 in order to match output shape
-    unsigned int padx = output_shape.x() - (out_x - weights_width + 1);
-    unsigned int pady = output_shape.y() - (out_y - weights_height + 1);
-    out_x += padx;
-    out_y += pady;
+    unsigned int deconv_pad_x = output_shape.x() - (out_x - weights_width + 1);
+    unsigned int deconv_pad_y = output_shape.y() - (out_y - weights_height + 1);
+    out_x += deconv_pad_x;
+    out_y += deconv_pad_y;
+
+    unsigned int deconv_pad_left  = pad_right > pad_left ? pad_right - pad_left : 0;
+    unsigned int deconv_pad_right = pad_left > pad_right ? pad_left - pad_right : 0;
+    deconv_pad_x -= deconv_pad_left + deconv_pad_right;
+    ARM_COMPUTE_ERROR_ON((deconv_pad_x % 2) != 0);
+    deconv_pad_left  += deconv_pad_x / 2;
+    deconv_pad_right += deconv_pad_x / 2;
+
+    unsigned int deconv_pad_top    = pad_bottom > pad_top ? pad_bottom - pad_top : 0;
+    unsigned int deconv_pad_bottom = pad_top > pad_bottom ? pad_top - pad_bottom : 0;
+    deconv_pad_y -= deconv_pad_top + deconv_pad_bottom;
+    ARM_COMPUTE_ERROR_ON((deconv_pad_y % 2) != 0);
+    deconv_pad_top    += deconv_pad_y / 2;
+    deconv_pad_bottom += deconv_pad_y / 2;
 
     TensorShape scaled_shape = src.shape();
     scaled_shape.set(0, out_x);
@@ -64,7 +87,6 @@ SimpleTensor<T> deconvolution_layer(const SimpleTensor<T> &src, const SimpleTens
     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);
-    ARM_COMPUTE_ERROR_ON(info.pad().first > (weights.shape().x() - 1));
 
     if(src.data_type() == DataType::QASYMM8)
     {
@@ -94,10 +116,10 @@ SimpleTensor<T> deconvolution_layer(const SimpleTensor<T> &src, const SimpleTens
     {
         const int offset_slice_in  = slice * width_in * height_in;
         const int offset_slice_out = slice * width_scaled * height_scaled;
-        const int start_x          = padx / 2;
-        const int start_y          = pady / 2;
-        const int end_y            = height_scaled - pady / 2;
-        const int end_x            = width_scaled - padx / 2;
+        const int start_x          = deconv_pad_left;
+        const int start_y          = deconv_pad_top;
+        const int end_x            = width_scaled - deconv_pad_right;
+        const int end_y            = height_scaled - deconv_pad_bottom;
 
         for(int yi = start_y, in_y = 0; yi < end_y; yi += stride_y, in_y++)
         {