From 1246b63ca04cb067f26ae860688647224d6ba24e Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Wed, 16 Aug 2017 18:38:32 +0100 Subject: COMPMID-477 - Optimized Direct Convolution 3x3 and 5x5 (f32) for Bifrost. Each work-item computes 4x3 output elements in case of 3x3 convolution and 4x2 in case of 5x5 convolution Change-Id: I6ebbaff8b7e971c1f90d5845c0b58d2a40f39df5 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/84345 Reviewed-by: Anthony Barbier Tested-by: Kaizen --- .../CL/kernels/CLDirectConvolutionLayerKernel.cpp | 139 +++++++++++++++------ 1 file changed, 100 insertions(+), 39 deletions(-) (limited to 'src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp') diff --git a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp index 1620d545c7..b012f59bd2 100644 --- a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp +++ b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp @@ -99,68 +99,129 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL _biases = biases; _border_size = BorderSize(_conv_pad_y, _conv_pad_x); - std::stringstream kernel_name; std::set options; - kernel_name << "direct_convolution" << kernel_size << "x" << kernel_size; - DataType promoted_type = input->info()->data_type(); - options.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); - options.emplace("-DDATA_SIZE=" + get_data_size_from_data_type(input->info()->data_type())); - options.emplace("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(_weights->info()->dimension(2))); - options.emplace("-DSTRIDE_X=" + support::cpp11::to_string(_conv_stride_x)); + const GPUTarget gpu_target = get_arch_from_target(get_target()); - if(is_data_type_fixed_point(input->info()->data_type())) + if(_biases != nullptr) + { + options.emplace("-DHAS_BIAS"); + } + + if((gpu_target == GPUTarget::BIFROST) && (kernel_size <= 5) && (_conv_stride_x == 1) && (_conv_stride_y == 1) && (input->info()->data_type() == DataType::F32)) { - options.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); + options.emplace("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(_weights->info()->dimension(2))); + + std::string kernel_name = "direct_convolution" + support::cpp11::to_string(kernel_size) + "x" + support::cpp11::to_string(kernel_size) + "_f32_bifrost"; + _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name, options)); - switch(input->info()->data_type()) + // Configure kernel window + Window win = calculate_max_window(*output->info()); + + 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; + + switch(kernel_size) { - case DataType::QS8: - promoted_type = DataType::QS16; + case 3: + { + num_elems_read_per_iteration_x = 6; + num_elems_read_per_iteration_y = 5; + num_elems_written_per_iteration_x = 4; + num_elems_written_per_iteration_y = 3; break; - case DataType::QS16: - promoted_type = DataType::QS32; + } + case 5: + { + num_elems_read_per_iteration_x = 8; + num_elems_read_per_iteration_y = 6; + num_elems_written_per_iteration_x = 4; + num_elems_written_per_iteration_y = 2; break; + } default: - ARM_COMPUTE_ERROR("Datatype not supported"); + { + ARM_COMPUTE_ERROR("Kernel size not optimized for Bifrost"); + } } - } - options.emplace("-DDATA_TYPE_PROMOTED=" + get_cl_type_from_data_type(promoted_type)); + // Calculate right and bottom border + const int input_width = input->info()->dimension(0) - kernel_size / 2 + _conv_pad_x; + const int input_height = input->info()->dimension(1) - kernel_size / 2 + _conv_pad_y; - if(_biases != nullptr) - { - options.emplace("-DHAS_BIAS"); + // Create window and update padding + 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_x, -_conv_pad_y, input_width + num_elems_read_per_iteration_x, input_height + num_elems_read_per_iteration_y); + AccessWindowStatic weights_access(weights->info(), 0, 0, kernel_size, kernel_size); + 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); + + output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); + + ICLKernel::configure(win); } + else + { + std::stringstream kernel_name; + kernel_name << "direct_convolution" << kernel_size << "x" << kernel_size; + DataType promoted_type = input->info()->data_type(); - _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name.str(), options)); + options.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); + options.emplace("-DDATA_SIZE=" + get_data_size_from_data_type(input->info()->data_type())); + options.emplace("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(_weights->info()->dimension(2))); + options.emplace("-DSTRIDE_X=" + support::cpp11::to_string(_conv_stride_x)); - // Configure kernel window - Window win = calculate_max_window(*output->info()); + if(is_data_type_fixed_point(input->info()->data_type())) + { + options.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); + + switch(input->info()->data_type()) + { + case DataType::QS8: + promoted_type = DataType::QS16; + break; + case DataType::QS16: + promoted_type = DataType::QS32; + break; + default: + ARM_COMPUTE_ERROR("Datatype not supported"); + } + } + + options.emplace("-DDATA_TYPE_PROMOTED=" + get_cl_type_from_data_type(promoted_type)); + + _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name.str(), options)); - bool is_stride2 = ((kernel_size != 1) && (_conv_stride_x == 2)); + // Configure kernel window - const unsigned int num_elems_read_per_iteration_x = 8 + 2 * (kernel_size / 2) + (is_stride2 ? 6 + kernel_size / 2 : 0); - const unsigned int num_elems_read_per_iteration_y = kernel_size; - const unsigned int num_elems_written_per_iteration_x = 8; - const unsigned int num_elems_written_per_iteration_y = 1; + bool is_stride2 = ((kernel_size != 1) && (_conv_stride_x == 2)); - // Calculate right and bottom border - const int input_width = input->info()->dimension(0) - kernel_size / 2 + _conv_pad_x; - const int input_height = input->info()->dimension(1) - kernel_size / 2 + _conv_pad_y; + const unsigned int num_elems_read_per_iteration_x = 8 + 2 * (kernel_size / 2) + (is_stride2 ? 6 + kernel_size / 2 : 0); + const unsigned int num_elems_read_per_iteration_y = kernel_size; + const unsigned int num_elems_written_per_iteration_x = 8; + const unsigned int num_elems_written_per_iteration_y = 1; - // Create window and update padding - win = calculate_max_window(*output->info(), Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y)); + // Calculate right and bottom border + const int input_width = input->info()->dimension(0) - kernel_size / 2 + _conv_pad_x; + const int input_height = input->info()->dimension(1) - kernel_size / 2 + _conv_pad_y; - AccessWindowStatic input_access(input->info(), -_conv_pad_x, -_conv_pad_y, input_width + num_elems_read_per_iteration_x, input_height + num_elems_read_per_iteration_y); - AccessWindowStatic weights_access(weights->info(), 0, 0, kernel_size, kernel_size); - AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_written_per_iteration_x, num_elems_written_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)); - update_window_and_padding(win, input_access, weights_access, output_access); + AccessWindowStatic input_access(input->info(), -_conv_pad_x, -_conv_pad_y, input_width + num_elems_read_per_iteration_x, input_height + num_elems_read_per_iteration_y); + AccessWindowStatic weights_access(weights->info(), 0, 0, kernel_size, kernel_size); + AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y); - output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); + update_window_and_padding(win, input_access, weights_access, output_access); - ICLKernel::configure(win); + output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); + + ICLKernel::configure(win); + } } void CLDirectConvolutionLayerKernel::run(const Window &window, cl::CommandQueue &queue) -- cgit v1.2.1