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, 11 insertions, 6 deletions

diff --git a/src/core/GLES_COMPUTE/cs_shaders/convolution_layer.cs b/src/core/GLES_COMPUTE/cs_shaders/convolution_layer.cs
index 2701f5b262..ad3f14d442 100644
--- a/src/core/GLES_COMPUTE/cs_shaders/convolution_layer.cs
+++ b/src/core/GLES_COMPUTE/cs_shaders/convolution_layer.cs
@@ -164,6 +164,7 @@ void main()
  * @note STRIDE_X/STRIDE_Y must be passed for stride info, e.g. "#define STRIDE_X xxx"
  * @note CONVOLVED_WIDTH/CONVOLVED_HEIGHT must be passed for convolved dimension, e.g. "#define CONVOLVED_WIDTH xxx"
  * @note SRC_WIDTH/SRC_HEIGHT must be passed for input dimension, e.g. "#define SRC_WIDTH xxx"
+ * @note DILATION_X/DILATION_Y must be passed for dilation sizes, e.g. "#define DILATION_X xxx"
  * @note In case biases will be added to the convolution "#define HAS_BIAS" has to be passed to append the final matrix with 1 in each row.
  *
  * @param[in]  src_ptr      Pointer to the source tensor. Supported data types: F16/F32
@@ -210,9 +211,9 @@ void main(void)
     uint src_pos = 0u;
 
     // Linearize convolution elements
-    for(uint y = yi, y_e = yi + uint(KERNEL_HEIGHT); y < y_e; ++y)
+    for(uint y = yi, y_e = yi + uint(KERNEL_HEIGHT) * uint(DILATION_Y); y < y_e; y += uint(DILATION_Y))
     {
-        for(uint x = xi, x_e = xi + uint(KERNEL_WIDTH); x < x_e; ++x, TENSOR_OFFSET_ADVANCE(dst_iter, 1u))
+        for(uint x = xi, x_e = xi + uint(KERNEL_WIDTH) * uint(DILATION_X); x < x_e; x += uint(DILATION_X), TENSOR_OFFSET_ADVANCE(dst_iter, 1u))
         {
 #if PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0
             src_pos = TENSOR_OFFSET_ADVANCE_IN_BYTES(src_iter, x * src_attrs.stride_x + y * src_attrs.stride_y);
diff --git a/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp
index 47bfebcc09..eb790471fb 100644
--- a/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp
@@ -65,7 +65,7 @@ GCIm2ColKernel::GCIm2ColKernel()
 {
 }
 
-void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
@@ -98,7 +98,8 @@ void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, const
                                      && (std::equal(input->info()->tensor_shape().cbegin() + 3,
                                                     input->info()->tensor_shape().cend(),
                                                     output->info()->tensor_shape().cbegin() + 1))
-                                     && ((stride_x == 1) && (stride_y == 1) && !conv_info.has_padding());
+                                     && ((stride_x == 1) && (stride_y == 1) && !conv_info.has_padding())
+                                     && (dilation == Size2D(1U, 1U));
 
     std::string kernel_name = "im2col_generic";
     if(!run_img2col_reduced)
@@ -111,7 +112,7 @@ void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, const
         build_opts.emplace("#define IM2COL_GENERIC");
         _convolved_dims = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1),
                                             kernel_dims.width, kernel_dims.height,
-                                            conv_info);
+                                            conv_info, dilation);
         _num_elems_processed_per_iteration = 2;
 
         build_opts.emplace("#define KERNEL_WIDTH " + support::cpp11::to_string(kernel_dims.width));
@@ -127,6 +128,8 @@ void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, const
         build_opts.emplace("#define PAD_BOTTOM " + support::cpp11::to_string(conv_info.pad_bottom()));
         build_opts.emplace("#define SRC_WIDTH " + support::cpp11::to_string(input->info()->dimension(0)));
         build_opts.emplace("#define SRC_HEIGHT " + support::cpp11::to_string(input->info()->dimension(1)));
+        build_opts.emplace("#define DILATION_X " + support::cpp11::to_string(dilation.x()));
+        build_opts.emplace("#define DILATION_Y " + support::cpp11::to_string(dilation.y()));
 
         _run_func = &GCIm2ColKernel::run_generic;
     }
@@ -206,11 +209,12 @@ void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, const
     IGCKernel::configure(win);
 }
 
-Status GCIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+Status GCIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation)
 {
     ARM_COMPUTE_UNUSED(kernel_dims);
     ARM_COMPUTE_UNUSED(conv_info);
     ARM_COMPUTE_UNUSED(has_bias);
+    ARM_COMPUTE_UNUSED(dilation);
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
     return Status{};
 }