From ad0c7388f6261989a268ffb2d042f2bd80736e3f Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Mon, 23 Apr 2018 16:16:21 +0100 Subject: COMPMID-1068 Create validate method to CLDepthWiseConvolution Change-Id: I3301b66a8a072c6ecd0d7f2dabef350017b55ac4 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/128677 Tested-by: Jenkins Reviewed-by: Anthony Barbier --- .../CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp | 273 ++++++++++++--------- 1 file changed, 155 insertions(+), 118 deletions(-) (limited to 'src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp') diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp index 1997a901fe..e4ad97faca 100644 --- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp @@ -39,116 +39,57 @@ using namespace arm_compute; using namespace arm_compute::misc::shape_calculator; -CLDepthwiseConvolutionLayer3x3NCHWKernel::CLDepthwiseConvolutionLayer3x3NCHWKernel() - : _conv_stride_x(0), _conv_pad_top(0) +namespace { -} - -BorderSize CLDepthwiseConvolutionLayer3x3NCHWKernel::border_size() const +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, + const ActivationLayerInfo &act_info) { - return _border_size; -} - -void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, - unsigned int depth_multiplier, - ActivationLayerInfo act_info) -{ - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); - ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 3 || weights->info()->dimension(1) != 3); - - bool is_qasymm = is_data_type_quantized_asymmetric(input->info()->data_type()); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(act_info.enabled() && ((input->data_type() != DataType::QASYMM8) || ((act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + && (act_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU))), + "For QASYMM8 only relu, lower bounded relu and lower-upper bounded relu are supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) != 3 || weights->dimension(1) != 3); + ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(2) * depth_multiplier) != output->dimension(2)); + ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1 || conv_info.stride().first > 3); + + const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type()); if(biases != nullptr) { if(is_qasymm) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); } else { - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); } - ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(2)); - ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(2)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } - // Get convolved dimensions - const TensorShape output_shape = compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info, depth_multiplier); - - // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), - output_shape, - 1, - input->info()->data_type(), - input->info()->fixed_point_position(), - input->info()->quantization_info()); - - ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); - ARM_COMPUTE_ERROR_ON(output->info()->dimension(2) != weights->info()->dimension(2)); - - _input = input; - _output = output; - _weights = weights; - _biases = biases; - _conv_stride_x = conv_info.stride().first; - _conv_stride_y = conv_info.stride().second; - _conv_pad_left = conv_info.pad_left(); - _conv_pad_top = conv_info.pad_top(); - _border_size = BorderSize(_conv_pad_top, conv_info.pad_right(), conv_info.pad_bottom(), _conv_pad_left); - - // Set build options - ARM_COMPUTE_ERROR_ON(_conv_stride_x < 1 || _conv_stride_x > 3); - CLBuildOptions build_opts; - build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier)); - build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x)); - build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS"); - - if(is_qasymm) + if(output->total_size() != 0) { - 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); - - build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(_conv_stride_y)); - build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-_input->info()->quantization_info().offset)); - build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-_weights->info()->quantization_info().offset)); - build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(_output->info()->quantization_info().offset)); - build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * input->info()->quantization_info().offset * weights->info()->quantization_info().offset)); - build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier)); - build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift)); - - if(act_info.enabled()) - { - const int a_val = input->info()->quantization_info().quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP); - const int b_val = input->info()->quantization_info().quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP); - const int o1 = input->info()->quantization_info().offset; - - build_opts.add_option("-DFUSED_ACTIVATION=" + lower_string(string_from_activation_func(act_info.activation()))); - build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val)); - build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val)); - build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1)); + const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); + } - if(output != nullptr) - { - const float s1 = input->info()->quantization_info().scale; - const float s2 = output->info()->quantization_info().scale; - const int o2 = output->info()->quantization_info().offset; + return Status{}; +} - if(o1 != o2 || s1 != s2) - { - build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1)); - build_opts.add_option("-DS2_VAL=" + float_to_string_with_full_precision(s2)); - build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1)); - build_opts.add_option("-DO2_VAL=" + support::cpp11::to_string(o2)); - } - } - } - } +std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, + GPUTarget gpu_target, std::string &kernel_name) +{ + // Output auto inizialitation if not yet initialized + const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier); + auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape)); - const GPUTarget gpu_target = get_target(); - const bool is_bifrost = gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX); + const unsigned int conv_stride_x = conv_info.stride().first; + const unsigned int conv_stride_y = conv_info.stride().second; + const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type()); + const bool is_bifrost = get_arch_from_target(gpu_target) == GPUTarget::BIFROST; // Configure kernel window unsigned int num_elems_read_per_iteration_x = 0; @@ -156,16 +97,13 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, unsigned int num_elems_written_per_iteration_x = 0; unsigned int num_elems_written_per_iteration_y = 0; - // Create kernel - std::string kernel_name; - - if(input->info()->data_type() == DataType::F16) + if(input->data_type() == DataType::F16) { 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_x = 8 / data_size_from_type(input->data_type()); num_elems_written_per_iteration_y = 1; num_elems_read_per_iteration_y = 3; - switch(_conv_stride_x) + switch(conv_stride_x) { case 1: num_elems_read_per_iteration_x = 8; @@ -177,12 +115,12 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, 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; + num_elems_read_per_iteration_x = 3 + (num_elems_written_per_iteration_x - 1) * conv_stride_x; break; } if(is_bifrost) { - if(_conv_stride_x == 1 && _conv_stride_y == 1) + if(conv_stride_x == 1 && conv_stride_y == 1) { kernel_name = "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16"; num_elems_read_per_iteration_x = 8; @@ -190,7 +128,7 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, num_elems_read_per_iteration_y = 6; num_elems_written_per_iteration_y = 4; } - else if(_conv_stride_x == 2 && _conv_stride_y == 2) + else if(conv_stride_x == 2 && conv_stride_y == 2) { kernel_name = "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16"; num_elems_read_per_iteration_x = 10; @@ -200,9 +138,9 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, } } } - else if(input->info()->data_type() == DataType::F32 && is_bifrost) + else if(input->data_type() == DataType::F32 && is_bifrost) { - if(_conv_stride_x == 1 && _conv_stride_y == 1) + if(conv_stride_x == 1 && conv_stride_y == 1) { kernel_name = "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32"; num_elems_read_per_iteration_x = 4; @@ -210,7 +148,7 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, num_elems_written_per_iteration_x = 2; num_elems_written_per_iteration_y = 4; } - else if(_conv_stride_x == 2 && _conv_stride_y == 2) + else if(conv_stride_x == 2 && conv_stride_y == 2) { kernel_name = "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32"; num_elems_read_per_iteration_x = 6; @@ -221,35 +159,123 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, 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_x = 8 / data_size_from_type(input->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_x = 3 + (num_elems_written_per_iteration_x - 1) * conv_stride_x; num_elems_read_per_iteration_y = 3; } } else { kernel_name = is_qasymm ? "depthwise_convolution_3x3_quantized_nchw" : "depthwise_convolution_3x3"; - num_elems_written_per_iteration_x = 8 / data_size_from_type(input->info()->data_type()); - num_elems_written_per_iteration_y = (is_qasymm && _conv_stride_y < 3) ? (2 / _conv_stride_y) : 1; - num_elems_read_per_iteration_x = 3 + (num_elems_written_per_iteration_x - 1) * _conv_stride_x; + num_elems_written_per_iteration_x = 8 / data_size_from_type(input->data_type()); + num_elems_written_per_iteration_y = (is_qasymm && conv_stride_y < 3) ? (2 / conv_stride_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 = num_elems_written_per_iteration_y + 2; } // 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)); + Window win = calculate_max_window(*output, Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y)); - AccessWindowRectangle input_access(input->info(), -_conv_pad_left, -_conv_pad_top, + AccessWindowRectangle input_access(input, -conv_info.pad_left(), -conv_info.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); + conv_stride_x, conv_stride_y); + AccessWindowStatic weights_access(weights, 0, 0, 3, 3); + AccessWindowRectangle output_access(output, 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y); + + bool window_changed = update_window_and_padding(win, input_access, weights_access, output_access); + + output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} +} // namespace + +CLDepthwiseConvolutionLayer3x3NCHWKernel::CLDepthwiseConvolutionLayer3x3NCHWKernel() + : _conv_stride_x(0), _conv_pad_top(0) +{ +} + +BorderSize CLDepthwiseConvolutionLayer3x3NCHWKernel::border_size() const +{ + return _border_size; +} - update_window_and_padding(win, input_access, weights_access, output_access); +void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, + unsigned int depth_multiplier, + ActivationLayerInfo act_info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); + + bool is_qasymm = is_data_type_quantized_asymmetric(input->info()->data_type()); + + _input = input; + _output = output; + _weights = weights; + _biases = biases; + _conv_stride_x = conv_info.stride().first; + _conv_stride_y = conv_info.stride().second; + _conv_pad_left = conv_info.pad_left(); + _conv_pad_top = conv_info.pad_top(); + _border_size = BorderSize(_conv_pad_top, conv_info.pad_right(), conv_info.pad_bottom(), _conv_pad_left); - output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); + // Set build options + CLBuildOptions build_opts; + build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier)); + build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x)); + build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS"); - ICLKernel::configure(win); + if(is_qasymm) + { + 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); + + build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(_conv_stride_y)); + build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-_input->info()->quantization_info().offset)); + build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-_weights->info()->quantization_info().offset)); + build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(_output->info()->quantization_info().offset)); + build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * input->info()->quantization_info().offset * weights->info()->quantization_info().offset)); + build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier)); + build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift)); + + if(act_info.enabled()) + { + const int a_val = input->info()->quantization_info().quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP); + const int b_val = input->info()->quantization_info().quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP); + const int o1 = input->info()->quantization_info().offset; + + build_opts.add_option("-DFUSED_ACTIVATION=" + lower_string(string_from_activation_func(act_info.activation()))); + build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val)); + build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val)); + build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1)); + + if(output != nullptr) + { + const float s1 = input->info()->quantization_info().scale; + const float s2 = output->info()->quantization_info().scale; + const int o2 = output->info()->quantization_info().offset; + + if(o1 != o2 || s1 != s2) + { + build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1)); + build_opts.add_option("-DS2_VAL=" + float_to_string_with_full_precision(s2)); + build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1)); + build_opts.add_option("-DO2_VAL=" + support::cpp11::to_string(o2)); + } + } + } + } + + // Configure kernel window + std::string kernel_name; + const GPUTarget gpu_target = get_target(); + + auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info, depth_multiplier, gpu_target, kernel_name); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + ICLKernel::configure(win_config.second); _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); @@ -269,6 +295,17 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, _config_id += support::cpp11::to_string(output->info()->dimension(1)); } +Status CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + unsigned int depth_multiplier, + ActivationLayerInfo act_info, GPUTarget gpu_target) +{ + std::string kernel_name; + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(), conv_info, depth_multiplier, gpu_target, kernel_name).first); + + return Status{}; +} + void CLDepthwiseConvolutionLayer3x3NCHWKernel::run(const Window &window, cl::CommandQueue &queue) { ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); -- cgit v1.2.1