COMPMID-1586: Add support for NHWC CLDeconvolutionLayer

COMPMID-1651: Fix QASYMM8 CLDeconvolutionLayer

This patch also extends the range of values used for testing Convolution and
Deconvolution to cover quantized [-1.0f, 1.0f].

Change-Id: I8b280669db67bb3ec25bf5d411c8f5954f5b0dab
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/149869
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Tested-by: bsgcomp <bsgcomp@arm.com>

5 files changed, 123 insertions, 82 deletions

diff --git a/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp b/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp
index be3a926b96..dd7d79002d 100644
--- a/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp
+++ b/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp
@@ -45,11 +45,19 @@ Status CLDeconvolutionLayerUpsampleKernel::validate(const ITensorInfo *input, co
 
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) == 0);
-    ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) == 0);
+
+    const DataLayout data_layout = input->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 size_t idx_c = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
+    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());
 
-    for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_c) != output->dimension(idx_c));
+    for(size_t i = 3; i < Coordinates::num_max_dimensions; ++i)
     {
         ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(i) != output->dimension(i));
     }
@@ -93,28 +101,61 @@ void CLDeconvolutionLayerUpsampleKernel::run(const Window &window, cl::CommandQu
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
 
+    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 int out_start_x = _info.pad().first;
-    const int out_end_x   = _output->info()->dimension(0) - _inner_border.right - _info.pad().first + _info.stride().first - 1;
+    const int out_end_x   = _output->info()->dimension(idx_w) - _inner_border.right - _info.pad().first + _info.stride().first - 1;
     const int out_step_x  = _info.stride().first;
 
     const int out_start_y = _inner_border.top + _info.pad().second;
-    const int out_end_y   = _output->info()->dimension(1) - _info.pad().second + _info.stride().second - 1;
+    const int out_end_y   = _output->info()->dimension(idx_h) - _info.pad().second + _info.stride().second - 1;
     const int out_step_y  = _info.stride().second;
 
-    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
-
-    Window slice_out = collapsed.first_slice_window_3D();
-    slice_out.set(Window::DimX, Window::Dimension(out_start_x, out_end_x, out_step_x));
-    slice_out.set(Window::DimY, Window::Dimension(out_start_y, out_end_y, out_step_y));
-
-    Window slice_in = collapsed.first_slice_window_3D();
-
-    do
+    switch(data_layout)
     {
-        unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, slice_in);
-        add_3D_tensor_argument(idx, _output, slice_out);
-        enqueue(queue, *this, slice_out);
+        case DataLayout::NCHW:
+        {
+            Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+
+            Window slice_out = collapsed.first_slice_window_3D();
+            slice_out.set(Window::DimX, Window::Dimension(out_start_x, out_end_x, out_step_x));
+            slice_out.set(Window::DimY, Window::Dimension(out_start_y, out_end_y, out_step_y));
+
+            Window slice_in = collapsed.first_slice_window_3D();
+
+            do
+            {
+                unsigned int idx = 0;
+                add_3D_tensor_argument(idx, _input, slice_in);
+                add_3D_tensor_argument(idx, _output, slice_out);
+                enqueue(queue, *this, slice_out);
+            }
+            while(collapsed.slide_window_slice_3D(slice_in) && collapsed.slide_window_slice_3D(slice_out));
+            break;
+        }
+        case DataLayout::NHWC:
+        {
+            // NOTE: not collapsing in NHWC
+            Window slice_out = window.first_slice_window_3D();
+            slice_out.set(Window::DimY, Window::Dimension(out_start_x, out_end_x, out_step_x));
+            slice_out.set(Window::DimZ, Window::Dimension(out_start_y, out_end_y, out_step_y));
+
+            Window slice_in = window.first_slice_window_3D();
+
+            do
+            {
+                unsigned int idx = 0;
+                add_3D_tensor_argument(idx, _input, slice_in);
+                add_3D_tensor_argument(idx, _output, slice_out);
+                enqueue(queue, *this, slice_out);
+            }
+            while(window.slide_window_slice_3D(slice_in) && window.slide_window_slice_3D(slice_out));
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Unsupported data layout");
     }
-    while(collapsed.slide_window_slice_3D(slice_in) && collapsed.slide_window_slice_3D(slice_out));
 }
diff --git a/src/core/CPP/kernels/CPPFlipWeightsKernel.cpp b/src/core/CPP/kernels/CPPFlipWeightsKernel.cpp
index 741218e4f7..2d4c0ce5c8 100644
--- a/src/core/CPP/kernels/CPPFlipWeightsKernel.cpp
+++ b/src/core/CPP/kernels/CPPFlipWeightsKernel.cpp
@@ -42,25 +42,36 @@ CPPFlipWeightsKernel::CPPFlipWeightsKernel()
 }
 
 template <typename T>
-void CPPFlipWeightsKernel::flip_weights(const Window &window_input, const Window &window)
+void CPPFlipWeightsKernel::flip_weights(const Window &window_input)
 {
     // Create iterators
     Iterator in(_input, window_input);
 
-    Iterator out(_output, window);
+    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 int kernel_size = _input->info()->dimension(0);
+    const int kernel_width  = _input->info()->dimension(idx_w);
+    const int kernel_height = _input->info()->dimension(idx_h);
 
     execute_window_loop(window_input, [&](const Coordinates & id)
     {
-        *((reinterpret_cast<T *>(out.ptr()) + kernel_size * (kernel_size - id.y() - 1) + (kernel_size - id.x() - 1))) = *(reinterpret_cast<const T *>(in.ptr()));
+        const unsigned int x = kernel_width - id[idx_w] - 1;
+        const unsigned int y = kernel_height - id[idx_h] - 1;
+        Coordinates        output_coord(id);
+        output_coord.set(idx_w, x);
+        output_coord.set(idx_h, y);
+        *(reinterpret_cast<T *>(_output->ptr_to_element(output_coord))) = *(reinterpret_cast<const T *>(in.ptr()));
     },
-    in, out);
+    in);
 }
 
 void CPPFlipWeightsKernel::configure(const ITensor *input, ITensor *output)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 
     _input  = input;
     _output = output;
@@ -98,9 +109,5 @@ void CPPFlipWeightsKernel::run(const Window &window, const ThreadInfo &info)
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window);
     ARM_COMPUTE_ERROR_ON(_func == nullptr);
 
-    Window out_window{ window };
-    out_window.set(Window::DimX, Window::Dimension(0, 0, 0));
-    out_window.set(Window::DimY, Window::Dimension(0, 0, 0));
-
-    (this->*_func)(window, out_window);
+    (this->*_func)(window);
 }
diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp
index a6a5771ec1..41fc87e87a 100644
--- a/src/core/Utils.cpp
+++ b/src/core/Utils.cpp
@@ -323,15 +323,6 @@ PadStrideInfo arm_compute::calculate_same_pad(TensorShape input_shape, TensorSha
     return PadStrideInfo(strides.first, strides.second, same_pad_left, same_pad_right, same_pad_top, same_pad_bottom, DimensionRoundingType::CEIL);
 }
 
-TensorShape arm_compute::deconvolution_output_shape(const std::pair<unsigned int, unsigned int> &out_dims, TensorShape input, TensorShape weights)
-{
-    TensorShape out_shape(input);
-    out_shape.set(0, out_dims.first);
-    out_shape.set(1, out_dims.second);
-    out_shape.set(2, weights[3]);
-    return out_shape;
-}
-
 const 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)
diff --git a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
index 26d44e9c96..951d1ec4f0 100644
--- a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
@@ -53,8 +53,16 @@ Status CLDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) != weights->dimension(1));
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) < 1);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights);
+
+    const DataLayout data_layout = input->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 size_t idx_c = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
+    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;
@@ -63,10 +71,10 @@ Status CLDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(inner_border_right > stride_x - 1, "inner_border_right must be smaller than stride_x");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(inner_border_top > stride_y - 1, "inner_border_top must be smaller than stride_y");
 
-    auto out_dims = deconvolution_output_dimensions(input->dimension(0), input->dimension(1), weights->dimension(0), weights->dimension(1),
+    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);
 
-    const TensorShape output_shape = deconvolution_output_shape(out_dims, input->tensor_shape(), weights->tensor_shape());
+    const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input, *weights);
 
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights);
 
@@ -80,15 +88,17 @@ Status CLDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
         {
             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
         }
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, bias);
     }
 
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid.");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid.");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid.");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_w) != output_shape[idx_w], "Output's width is invalid.");
+    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.");
 
-    TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_deconvolution_shape(*input, *weights, stride_x, stride_y, inner_border_right,
-                                                                                                      inner_border_top,
-                                                                                                      out_dims)));
+    unsigned int        padx            = 0;
+    unsigned int        pady            = 0;
+    const TensorShape   scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, inner_border_right, inner_border_top, out_dims, padx, pady);
+    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);
 
     ARM_COMPUTE_RETURN_ON_ERROR(CLDeconvolutionLayerUpsample::validate(input, &scale_out_info, BorderSize(inner_border_right, inner_border_top), info));
@@ -105,17 +115,22 @@ void CLDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights, const
     const unsigned int stride_x = info.stride().first;
     const unsigned int stride_y = info.stride().second;
 
+    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);
+
     _weights = weights;
-    _weights_flipped.allocator()->init(TensorInfo(weights->info()->tensor_shape(), 1, weights->info()->data_type()));
+    _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout));
     _flip_weights.configure(weights, &_weights_flipped);
 
-    auto out_dims = deconvolution_output_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1),
+    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);
 
-    const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape());
+    const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input->info(), *weights->info());
 
     // Output auto initialization if not yet initialized
-    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->quantization_info());
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).set_data_layout(data_layout));
 
     // Perform validation step
     ARM_COMPUTE_ERROR_THROW_ON(CLDeconvolutionLayer::validate(input->info(), weights->info(), bias == nullptr ? nullptr : bias->info(), output->info(), info, inner_border_right, inner_border_top));
@@ -125,20 +140,13 @@ void CLDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights, const
     _memory_group.manage(&_scaled_output);
     _memory_group.manage(&_weights_flipped);
 
-    // Find the upsampled dimensions
-    unsigned int out_x = (input->info()->dimension(0) - 1) * stride_x + inner_border_right + 1;
-    unsigned int out_y = (input->info()->dimension(1) - 1) * stride_y + inner_border_top + 1;
-
-    // Find the padding needed for the convolution with stride 1 in order to match output shape
-    unsigned int padx = out_dims.first - (out_x - weights->info()->dimension(0) + 1);
-    unsigned int pady = out_dims.second - (out_y - weights->info()->dimension(1) + 1);
-    out_x += padx;
-    out_y += pady;
+    // 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, inner_border_right, inner_border_top, out_dims, padx, pady);
 
-    TensorShape scale_out_shape(input->info()->tensor_shape());
-    scale_out_shape.set(0, out_x);
-    scale_out_shape.set(1, out_y);
     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
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
index 6ca60c66a4..cbe7c51662 100644
--- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
@@ -76,15 +76,17 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 
-        const TensorShape output_shape = deconvolution_output_shape(out_dims, input->tensor_shape(), weights->tensor_shape());
+        const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input, *weights);
+
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid.");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid.");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid.");
     }
 
-    TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_deconvolution_shape(*input, *weights, stride_x, stride_y, inner_border_right,
-                                                                                                      inner_border_top,
-                                                                                                      out_dims)));
+    unsigned int        padx            = 0;
+    unsigned int        pady            = 0;
+    const TensorShape   scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, inner_border_right, inner_border_top, out_dims, padx, pady);
+    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);
 
     for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
@@ -116,7 +118,7 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con
     auto out_dims = deconvolution_output_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1),
                                                     info.pad().first, info.pad().second, stride_x, stride_y);
 
-    const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape());
+    const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input->info(), *weights->info());
     // Output auto initialization if not yet initialized
     auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->quantization_info());
 
@@ -125,19 +127,11 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con
 
     _memory_group.manage(&_scaled_output);
 
-    // Find the upsampled dimensions
-    unsigned int out_x = (input->info()->dimension(0) - 1) * stride_x + inner_border_right + 1;
-    unsigned int out_y = (input->info()->dimension(1) - 1) * stride_y + inner_border_top + 1;
-
-    // Find the padding needed for the convolution with stride 1 in order to match output shape
-    unsigned int padx = out_dims.first - (out_x - weights->info()->dimension(0) + 1);
-    unsigned int pady = out_dims.second - (out_y - weights->info()->dimension(1) + 1);
-    out_x += padx;
-    out_y += pady;
+    // 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, inner_border_right, inner_border_top, out_dims, padx, pady);
 
-    TensorShape scale_out_shape(input->info()->tensor_shape());
-    scale_out_shape.set(0, out_x);
-    scale_out_shape.set(1, out_y);
     TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
     _scaled_output.allocator()->init(scale_out_info);
 
@@ -171,4 +165,4 @@ void NEDeconvolutionLayer::prepare()
         _conv_f.prepare();
         _is_prepared = true;
     }
-}
\ No newline at end of file
+}