From c799ed85bcf0b18269e65a64f34382073a68bd93 Mon Sep 17 00:00:00 2001 From: Gian Marco Date: Thu, 1 Feb 2018 16:57:48 +0000 Subject: COMPMID-895 - Optimizing CLDepthwiseConvolution3x3Kernel This patch brings the MACs utilisation up to 25 % when both stride_x and stride_y are equal to 1 Performance reported in the following confluence page: https://confluence.arm.com/display/MLENG/Depthwise+convolution+3x3+FP32+performance%3A+ACL+18.02 Change-Id: Ida1b64be9a88805902a3d90194559b58eb1224a3 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/119068 Reviewed-by: Michalis Spyrou Tested-by: Jenkins --- .../CLDepthwiseConvolutionLayer3x3Kernel.cpp | 134 ++++++++++++--------- 1 file changed, 76 insertions(+), 58 deletions(-) (limited to 'src/core/CL/kernels') diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp index a9167ee859..2a60f60723 100644 --- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp @@ -98,76 +98,74 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x)); build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS"); - // Create kernel - std::string kernel_name = is_data_type_quantized_asymmetric(_input->info()->data_type()) ? "depthwise_convolution_3x3_quantized" : "depthwise_convolution_3x3"; - _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); + // Configure the local work size for Bifrost with a value obtained + // via exhaustive autotuning for the MobileNets tensor shapes. + const GPUTarget gpu_target = get_arch_from_target(get_target()); - // Set static arguments - if(is_data_type_quantized_asymmetric(_input->info()->data_type())) - { - float multiplier = _input->info()->quantization_info().scale * _weights->info()->quantization_info().scale / _output->info()->quantization_info().scale; - int output_multiplier = 0; - int output_shift = 0; - quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); + // Configure kernel window + const unsigned int conv_pad_left = std::max(conv_info.pad_left(), 1U); + const unsigned int conv_pad_top = std::max(conv_info.pad_top(), 1U); + const unsigned int conv_pad_right = std::max(conv_info.pad_right(), 1U); + const unsigned int conv_pad_bottom = std::max(conv_info.pad_bottom(), 1U); - unsigned int idx = 3 * num_arguments_per_3D_tensor() + ((_biases != nullptr) ? num_arguments_per_1D_tensor() : 0); + 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; + unsigned int num_elems_written_per_iteration_y = 0; - _kernel.setArg(idx++, -_input->info()->quantization_info().offset); - _kernel.setArg(idx++, -_weights->info()->quantization_info().offset); - _kernel.setArg(idx++, _output->info()->quantization_info().offset); - _kernel.setArg(idx++, output_multiplier); - _kernel.setArg(idx++, output_shift); - } + // Create kernel + std::string kernel_name; - // Configure the local work size for Bifrost with a value obtained - // via exhaustive autotuning for the MobileNets tensor shapes. - const GPUTarget gpu_target = get_arch_from_target(get_target()); - if(gpu_target == GPUTarget::BIFROST) + if(input->info()->data_type() == DataType::F32 && gpu_target == GPUTarget::BIFROST) { - // Assume uniform padding and striding. - const size_t pad = _conv_pad_left; - const size_t stride = _conv_stride_x; - const size_t width = input->info()->dimension(0); - if(pad == 1) + if(_conv_stride_x == 1 && _conv_stride_y == 1) { - const size_t width_by_stride = width / stride; - if(width_by_stride == 28) // 56/2 or 28/1 - { - _lws_hint = cl::NDRange(7, 4, 3); - } - else if(width_by_stride == 14) // 28/2 or 14/1 - { - _lws_hint = cl::NDRange(7, 7, 4); - } + kernel_name = "depthwise_convolution_3x3_stridex1_stridey1_bifrost"; + num_elems_read_per_iteration_x = 4; + num_elems_read_per_iteration_y = 6; + num_elems_written_per_iteration_x = 2; + num_elems_written_per_iteration_y = 4; } - else if(pad == 0) + else if(_conv_stride_x == 2 && _conv_stride_y == 2) { - if(width >= 56) // 56 or 112 - { - _lws_hint = cl::NDRange(8, 5, 2); - } - else if(width >= 14) // 14 or 28 - { - _lws_hint = cl::NDRange(1, 5, 2); - } - else // 7 - { - _lws_hint = cl::NDRange(1, 1, 2); - } + kernel_name = "depthwise_convolution_3x3_stridex2_stridey2_bifrost"; + num_elems_read_per_iteration_x = 6; + num_elems_read_per_iteration_y = 5; + num_elems_written_per_iteration_x = 2; + num_elems_written_per_iteration_y = 2; } + else + { + kernel_name = "depthwise_convolution_3x3"; + 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; + } + } + else + { + kernel_name = is_data_type_quantized_asymmetric(_input->info()->data_type()) ? "depthwise_convolution_3x3_quantized" : "depthwise_convolution_3x3"; + 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; } - // Configure kernel window - const unsigned int num_elems_processed_per_iteration = 8 / data_size_from_type(input->info()->data_type()); - const unsigned int num_elems_written_per_iteration = num_elems_processed_per_iteration; - const unsigned int num_elems_read_per_iteration = 3 + (num_elems_processed_per_iteration - 1) * _conv_stride_x; - const unsigned int num_rows_read_per_iteration = 3; + // 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); - Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); + // 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)); - AccessWindowRectangle input_access(input->info(), -border_size().left, -border_size().top, num_elems_read_per_iteration, num_rows_read_per_iteration, _conv_stride_x, _conv_stride_y); - AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); - AccessWindowStatic weights_access(weights->info(), 0, 0, weights->info()->dimension(0), weights->info()->dimension(1)); + AccessWindowStatic input_access(input->info(), -conv_pad_left, -conv_pad_top, input_width, input_height); + 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); update_window_and_padding(win, input_access, weights_access, output_access); @@ -175,8 +173,28 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con ICLKernel::configure(win); + _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); + + // Set static arguments + if(is_data_type_quantized_asymmetric(_input->info()->data_type())) + { + float multiplier = _input->info()->quantization_info().scale * _weights->info()->quantization_info().scale / _output->info()->quantization_info().scale; + int output_multiplier = 0; + int output_shift = 0; + quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); + + unsigned int idx = 3 * num_arguments_per_3D_tensor() + ((_biases != nullptr) ? num_arguments_per_1D_tensor() : 0); + + _kernel.setArg(idx++, -_input->info()->quantization_info().offset); + _kernel.setArg(idx++, -_weights->info()->quantization_info().offset); + _kernel.setArg(idx++, _output->info()->quantization_info().offset); + _kernel.setArg(idx++, output_multiplier); + _kernel.setArg(idx++, output_shift); + } + // Set config_id for enabling LWS tuning - _config_id = "depthwise_convolution3x3_"; + _config_id = kernel_name; + _config_id += "_"; _config_id += lower_string(string_from_data_type(input->info()->data_type())); _config_id += "_"; _config_id += support::cpp11::to_string(input->info()->dimension(0)); -- cgit v1.2.1