From 1a03d76786a59a7d20deabb02f047516e98680d4 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Wed, 21 Feb 2018 14:47:09 +0000 Subject: COMPMID-765: Fix windows in DirectConvLayer and DepthwiseConvLayer Change-Id: I6c68733c8a2ada12aa3994e3e5213d20222df861 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/121637 Reviewed-by: Anthony Barbier Tested-by: Jenkins --- .../CLDepthwiseConvolutionLayer3x3Kernel.cpp | 33 ++++++++++++---------- .../CL/kernels/CLDirectConvolutionLayerKernel.cpp | 4 ++- 2 files changed, 21 insertions(+), 16 deletions(-) (limited to 'src/core/CL/kernels') diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp index c24420a7e3..29564b36c9 100644 --- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp @@ -121,11 +121,6 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con const GPUTarget gpu_target = get_arch_from_target(get_target()); // Configure kernel window - const unsigned int conv_pad_left = conv_info.pad_left(); - const unsigned int conv_pad_top = conv_info.pad_top(); - const unsigned int conv_pad_right = conv_info.pad_right(); - const unsigned int conv_pad_bottom = conv_info.pad_bottom(); - unsigned int num_elems_read_per_iteration_x = 0; unsigned int num_elems_read_per_iteration_y = 0; unsigned int num_elems_written_per_iteration_x = 0; @@ -139,8 +134,22 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con kernel_name = "depthwise_convolution_3x3_f16"; num_elems_written_per_iteration_x = 8 / data_size_from_type(input->info()->data_type()); num_elems_written_per_iteration_y = 1; - num_elems_read_per_iteration_x = 3 + (num_elems_written_per_iteration_x - 1) * _conv_stride_x; num_elems_read_per_iteration_y = 3; + switch(_conv_stride_x) + { + case 1: + num_elems_read_per_iteration_x = 8; + break; + case 2: + num_elems_read_per_iteration_x = 9; + break; + case 3: + num_elems_read_per_iteration_x = 16; + break; + default: + num_elems_read_per_iteration_x = 3 + (num_elems_written_per_iteration_x - 1) * _conv_stride_x; + break; + } } else if(input->info()->data_type() == DataType::F32 && gpu_target == GPUTarget::BIFROST) { @@ -178,18 +187,12 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con num_elems_read_per_iteration_y = num_elems_written_per_iteration_y + 2; } - // Calculate right and bottom border - int input_width = input->info()->dimension(0) + conv_pad_left + conv_pad_right; - int input_height = input->info()->dimension(1) + conv_pad_top + conv_pad_bottom; - - // Add padding only if necessary or it would always result in a window_changed - input_width = ceil_to_multiple(input_width, num_elems_read_per_iteration_x); - input_height = ceil_to_multiple(input_height, num_elems_read_per_iteration_y); - // Create window and update padding Window win = calculate_max_window(*output->info(), Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y)); - AccessWindowStatic input_access(input->info(), -conv_pad_left, -conv_pad_top, input_width, input_height); + AccessWindowRectangle input_access(input->info(), -_conv_pad_left, -_conv_pad_top, + num_elems_read_per_iteration_x, num_elems_read_per_iteration_y, + _conv_stride_x, _conv_stride_y); AccessWindowStatic weights_access(weights->info(), 0, 0, 3, 3); AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y); diff --git a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp index 6f5c7a35f9..c01a6660a7 100644 --- a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp +++ b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp @@ -241,7 +241,9 @@ std::pair validate_and_configure_window(ITensorInfo *input, ITen bool window_changed = false; Window win = calculate_max_window(*output, Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y)); - AccessWindowRectangle input_access(input, -conv_pad_left, -conv_pad_top, num_elems_read_per_iteration_x, num_elems_read_per_iteration_y, conv_stride_x, conv_stride_y); + AccessWindowRectangle input_access(input, -conv_pad_left, -conv_pad_top, + num_elems_read_per_iteration_x, num_elems_read_per_iteration_y, + conv_stride_x, conv_stride_y); AccessWindowStatic weights_access(weights, 0, 0, kernel_size, kernel_size); AccessWindowRectangle output_access(output, 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y); -- cgit v1.2.1