COMPMID-1017: Implement dilated convolution in NEON, OpenCL, and GC

Change-Id: If4626ec9e215e14dffe22e80812da5bac84a52e2
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/125734
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

2 files changed, 19 insertions, 15 deletions

diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp
index 1a486ce5c7..64bda93ff0 100644
--- a/src/runtime/CL/functions/CLConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLConvolutionLayer.cpp
@@ -42,13 +42,14 @@ CLConvolutionLayer::CLConvolutionLayer(std::shared_ptr<IMemoryManager> memory_ma
 {
 }
 
-void CLConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
+void CLConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
+                                   const Size2D &dilation)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(CLConvolutionLayer::validate(input->info(), weights->info(), ((biases != nullptr) ? biases->info() : nullptr), output->info(), conv_info, weights_info));
+    ARM_COMPUTE_ERROR_THROW_ON(CLConvolutionLayer::validate(input->info(), weights->info(), ((biases != nullptr) ? biases->info() : nullptr), output->info(), conv_info, weights_info, dilation));
 
     switch(CLConvolutionLayer::get_convolution_method(input->info(), weights->info(), ((biases != nullptr) ? biases->info() : nullptr), output->info(), conv_info,
-                                                      weights_info, CLScheduler::get().target()))
+                                                      weights_info, CLScheduler::get().target(), dilation))
     {
         case ConvolutionMethod::DIRECT:
         {
@@ -60,7 +61,7 @@ void CLConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, c
         case ConvolutionMethod::GEMM:
         {
             auto f = arm_compute::support::cpp14::make_unique<CLGEMMConvolutionLayer>(_memory_manager);
-            f->configure(input, weights, biases, output, conv_info, weights_info);
+            f->configure(input, weights, biases, output, conv_info, weights_info, dilation);
             _function = std::move(f);
             break;
         }
@@ -71,14 +72,14 @@ void CLConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, c
 }
 
 Status CLConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                    const WeightsInfo &weights_info)
+                                    const WeightsInfo &weights_info, const Size2D &dilation)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
 
     //Configure if the parameters match the direct convolution or the gemm-based
     const GPUTarget gpu_target = CLScheduler::get().target();
 
-    switch(CLConvolutionLayer::get_convolution_method(input, weights, biases, output, conv_info, weights_info, gpu_target))
+    switch(CLConvolutionLayer::get_convolution_method(input, weights, biases, output, conv_info, weights_info, gpu_target, dilation))
     {
         case ConvolutionMethod::DIRECT:
         {
@@ -89,7 +90,7 @@ Status CLConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo
         case ConvolutionMethod::GEMM:
         {
             // Validate gemm-based convolution layer
-            CLGEMMConvolutionLayer::validate(input, weights, biases, output, conv_info, weights_info);
+            CLGEMMConvolutionLayer::validate(input, weights, biases, output, conv_info, weights_info, dilation);
             break;
         }
         default:
@@ -101,7 +102,7 @@ Status CLConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo
 }
 
 ConvolutionMethod CLConvolutionLayer::get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                                             const WeightsInfo &weights_info, const GPUTarget gpu_target)
+                                                             const WeightsInfo &weights_info, const GPUTarget gpu_target, const Size2D &dilation)
 {
     ARM_COMPUTE_UNUSED(input);
     ARM_COMPUTE_UNUSED(weights);
@@ -110,6 +111,7 @@ ConvolutionMethod CLConvolutionLayer::get_convolution_method(const ITensorInfo *
     ARM_COMPUTE_UNUSED(conv_info);
     ARM_COMPUTE_UNUSED(weights_info);
     ARM_COMPUTE_UNUSED(gpu_target);
+    ARM_COMPUTE_UNUSED(dilation);
 
     return ConvolutionMethod::GEMM;
 }
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index bc339f176f..e7ad62f5ff 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -151,7 +151,8 @@ Status CLGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens
     return Status{};
 }
 
-void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
+void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
+                                       const Size2D &dilation)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 
@@ -160,7 +161,8 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
                                                                 biases != nullptr ? biases->info() : nullptr,
                                                                 output->info(),
                                                                 conv_info,
-                                                                weights_info));
+                                                                weights_info,
+                                                                dilation));
 
     _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
 
@@ -187,7 +189,7 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
     const unsigned int kernel_width  = weights->info()->dimension(0);
     const unsigned int kernel_height = weights->info()->dimension(1);
     std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
-                                                 conv_info);
+                                                 conv_info, dilation);
 
     unsigned int mat_weights_cols = weights->info()->dimension(3);
     unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element;
@@ -224,7 +226,7 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
     _memory_group.manage(&_gemm_output);
 
     // Configure im2col
-    _im2col_kernel.configure(input, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, append_bias);
+    _im2col_kernel.configure(input, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation);
 
     // Configure GEMM
     configure_mm(&_im2col_output, weights, &_gemm_output);
@@ -260,7 +262,7 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
 }
 
 Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                        const WeightsInfo &weights_info)
+                                        const WeightsInfo &weights_info, const Size2D &dilation)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!");
@@ -282,7 +284,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     const unsigned int kernel_width  = weights->dimension(0);
     const unsigned int kernel_height = weights->dimension(1);
 
-    std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height, conv_info);
+    std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height, conv_info, dilation);
 
     unsigned int mat_weights_cols = weights->dimension(3);
     unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + bias_element;
@@ -298,7 +300,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     shape_im2col.set(2, 1);
     TensorInfo im2col_reshaped_info(shape_im2col, 1, dt, input->fixed_point_position());
     im2col_reshaped_info.set_quantization_info(input->quantization_info());
-    CLIm2ColKernel::validate(input, &im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, append_bias);
+    CLIm2ColKernel::validate(input, &im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation);
 
     // Create GEMM output tensor
     TensorShape shape_gemm = im2col_reshaped_info.tensor_shape();