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.h | 16 ++ .../core/CL/kernels/CLDepthwiseIm2ColKernel.h | 14 ++ .../CL/kernels/CLDepthwiseVectorToTensorKernel.h | 11 +- .../CL/kernels/CLDepthwiseWeightsReshapeKernel.h | 10 + .../CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h | 9 + .../CL/functions/CLDepthwiseConvolutionLayer.h | 30 +++ .../CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp | 273 ++++++++++++--------- src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp | 31 ++- .../CL/kernels/CLDepthwiseVectorToTensorKernel.cpp | 47 +++- .../CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp | 45 ++-- .../kernels/CLGEMMMatrixVectorMultiplyKernel.cpp | 61 +++-- .../CL/functions/CLDepthwiseConvolutionLayer.cpp | 65 ++++- tests/validation/CL/DepthwiseConvolutionLayer.cpp | 165 +++++++++++++ 13 files changed, 613 insertions(+), 164 deletions(-) diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h index f80985a936..59cdf339bd 100644 --- a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h +++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h @@ -50,6 +50,22 @@ public: */ void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info) override; + /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NCHWKernel + * + * @param[in] input Source tensor. DataType supported: F16/F32/QASYMM8. + * @param[in] weights Weights tensor. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input. + * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. + * Data type supported: Same as @p input. + * @param[in] output Destination tensor. Data type supported: Same as @p input. + * @param[in] conv_info Padding and stride information to use for the convolution. + * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported. + * @param[in] gpu_target (Optional) GPU target to validate the kernel for. Defaults to midgard. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, + ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD); void run(const Window &window, cl::CommandQueue &queue) override; BorderSize border_size() const override; diff --git a/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h index 3f3e36100a..00d9cb64e1 100644 --- a/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h +++ b/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h @@ -60,6 +60,20 @@ public: * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. */ void configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias = false, unsigned int depth_multiplier = 1); + /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseIm2ColKernel + * + * @param[in] input The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F32 + * @param[in] output The output tensor. First 3 lower dimensions represent a transform of each 3D input, + * while every dimension above 3 represents a batch. Data types supported: Same as @p input + * @param[in] kernel_dims The kernel dimensions (width and height). + * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. + * @param[in] has_bias Boolean that specifies if the depthwise convolution has bias. + * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; diff --git a/arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h index 36d7cadbec..c9ec8e13bf 100644 --- a/arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h +++ b/arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h @@ -58,7 +58,16 @@ public: * @param[in] conv_h The converted tensor's height. */ void configure(const ICLTensor *input, ICLTensor *output, size_t conv_w, size_t conv_h); - + /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseVectorToTensorKernel + * + * @param[in] input The input vector to convert. Data type supported: QASYMM8/S32/F16/F32. + * @param[in] output The output tensor. 3 lower dimensions represent a single input [width, height, IFM]. Data type supported: same as @p input. + * @param[in] conv_w The converted tensor's width. + * @param[in] conv_h The converted tensor's height. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *output, size_t conv_w, size_t conv_h); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; diff --git a/arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h index 1c1eaca474..34ffa17c2b 100644 --- a/arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h +++ b/arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h @@ -56,6 +56,16 @@ public: * @param[in] biases (Optional) The input biases to add. Shape [IFM]. Data type supported: same as @p input. */ void configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *biases = nullptr); + /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseWeightsReshapeKernel + * + * @param[in] input The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM]. + * Data type supported: QASYMM8/F32. + * @param[in] output The output tensor. Data type supported: same as @p input. + * @param[in] biases (Optional) The input biases to add. Shape [IFM]. Data type supported: same as @p input. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases = nullptr); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h index c27307e63b..6390f86d2c 100644 --- a/arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h +++ b/arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h @@ -51,6 +51,15 @@ public: * @param[out] output The output 2D tensor. Data types supported: Same as @p input */ void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output); + /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMMatrixVectorMultiplyKernel + * + * @param[in] input0 The reshaped input tensor. Data types supported: QASYMM8/F16/F32 + * @param[in] input1 The 2D reshaped weights tensor. Data type supported: Same as @p input, S32 for QASYMM8 input. + * @param[in] output The output 2D tensor. Data types supported: Same as @p input + * + * @return a status + */ + static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h index 6e5ce4cd48..b1eb4b9e04 100644 --- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h +++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h @@ -67,6 +67,22 @@ public: void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo()); + /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3 + * + * @param[in] input Source tensor. Data type supported: QASYMM8 for all layouts, F16/F32 for NCHW. + * @param[in] weights Weights tensor. A 3D tensor with shape [3, 3, IFM]. Data type supported: Same as @p input. + * @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed. + * Data type supported: Same as @p input, S32 when input is QASYMM8. + * @param[in] output Destination tensor. Data type supported: same as @p input. + * @param[in] conv_info Padding and stride information to use for the convolution. + * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported. + * @param[in] gpu_target (Optional) GPU target to validate the kernel for. Defaults to midgard. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, + ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD); // Inherited methods overriden: void run() override; @@ -108,6 +124,20 @@ public: */ void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1); + /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer + * + * @param[in] input Source tensor. Data type supported: QASYMM8/F32. + * @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input. + * @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed. + * Data type supported: Same as @p input, S32 when input is QASYMM8. + * @param[in] output Destination tensor. Data type supported: same as @p input. + * @param[in] conv_info Padding and stride information to use for the convolution. + * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier); + // Inherited methods overriden: void run() override; 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); diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp index 0aef52f791..f44f08b347 100644 --- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp @@ -42,14 +42,26 @@ CLDepthwiseIm2ColKernel::CLDepthwiseIm2ColKernel() { } +namespace +{ +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier) +{ + ARM_COMPUTE_UNUSED(conv_info); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && has_bias); + ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(2) * depth_multiplier) != output->dimension(2)); + ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0))); + + return Status{}; +} +} // namespace + void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier) { - 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, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); - ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input->info()->data_type()) && has_bias); - ARM_COMPUTE_ERROR_ON((input->info()->dimension(2) * depth_multiplier) != output->info()->dimension(2)); - ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0))); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, depth_multiplier)); _input = input; _output = output; @@ -93,6 +105,13 @@ void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *outpu ICLKernel::configure(win); } +Status CLDepthwiseIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, depth_multiplier)); + + return Status{}; +} + void CLDepthwiseIm2ColKernel::run(const Window &window, cl::CommandQueue &queue) { ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); diff --git a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp index 83fc168f45..26336ebf79 100644 --- a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp @@ -34,6 +34,34 @@ using namespace arm_compute; +namespace +{ +TensorShape compute_output_shape(const TensorShape &input, size_t conv_w, size_t conv_h) +{ + TensorShape output_shape(input); + output_shape.set(0, conv_w); + output_shape.set(1, conv_h); + output_shape.set(2, input.x() / (conv_w * conv_h)); + + return output_shape; +} + +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, size_t conv_w, size_t conv_h) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::S32, DataType::F16, DataType::F32); + + if(output->total_size() != 0) + { + TensorShape output_shape = compute_output_shape(input->tensor_shape(), conv_w, conv_h); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); + } + + return Status{}; +} +} // namespace + CLDepthwiseVectorToTensorKernel::CLDepthwiseVectorToTensorKernel() : _input(nullptr), _output(nullptr) { @@ -41,20 +69,13 @@ CLDepthwiseVectorToTensorKernel::CLDepthwiseVectorToTensorKernel() void CLDepthwiseVectorToTensorKernel::configure(const ICLTensor *input, ICLTensor *output, size_t conv_w, size_t conv_h) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::S32, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_NULLPTR(output); - - TensorShape output_shape = input->info()->tensor_shape(); - output_shape.set(0, conv_w); - output_shape.set(1, conv_h); - output_shape.set(2, input->info()->tensor_shape()[0] / (conv_w * conv_h)); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); // Output auto inizialitation if not yet initialized + TensorShape output_shape = compute_output_shape(input->info()->tensor_shape(), conv_w, conv_h); auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape)); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), conv_w, conv_h)); _input = input; _output = output; @@ -75,6 +96,12 @@ void CLDepthwiseVectorToTensorKernel::configure(const ICLTensor *input, ICLTenso ICLKernel::configure(win); } +Status CLDepthwiseVectorToTensorKernel::validate(const ITensorInfo *input, const ITensorInfo *output, size_t conv_w, size_t conv_h) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, conv_w, conv_h)); + return Status{}; +} + void CLDepthwiseVectorToTensorKernel::run(const Window &window, cl::CommandQueue &queue) { ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); diff --git a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp index 26da96f9ba..b5a607d92e 100644 --- a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp @@ -34,6 +34,29 @@ using namespace arm_compute; +namespace +{ +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && (biases != nullptr)); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) != output->dimension(1)); + ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (input->dimension(0) * input->dimension(1) + ((biases != nullptr) ? 1 : 0))); + + if(biases != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != input->dimension(2)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + + return Status{}; +} +} // namespace + CLDepthwiseWeightsReshapeKernel::CLDepthwiseWeightsReshapeKernel() : _input(nullptr), _biases(nullptr), _output(nullptr) { @@ -41,20 +64,8 @@ CLDepthwiseWeightsReshapeKernel::CLDepthwiseWeightsReshapeKernel() void CLDepthwiseWeightsReshapeKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *biases) { - 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, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); - ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input->info()->data_type()) && (biases != nullptr)); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(1)); - ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (input->info()->dimension(0) * input->info()->dimension(1) + ((biases != nullptr) ? 1 : 0))); - - if(biases != nullptr) - { - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); - ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != input->info()->dimension(2)); - ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); - } + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), (biases != nullptr) ? biases->info() : nullptr)); _input = input; _biases = biases; @@ -80,6 +91,12 @@ void CLDepthwiseWeightsReshapeKernel::configure(const ICLTensor *input, ICLTenso ICLKernel::configure(win); } +Status CLDepthwiseWeightsReshapeKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, biases)); + return Status{}; +} + void CLDepthwiseWeightsReshapeKernel::run(const Window &window, cl::CommandQueue &queue) { ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); diff --git a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp index a1e47f28d4..b2ea95be1a 100644 --- a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp @@ -34,6 +34,42 @@ using namespace arm_compute; +namespace +{ +Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); + ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input0->data_type()) && (output->data_type() != DataType::S32)); + ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(2) != input1->dimension(1)); + + return Status{}; +} + +std::pair validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output) +{ + constexpr unsigned int num_elems_read_per_iteration = 4; + constexpr unsigned int num_rows_read_per_iteration = 4; + + const unsigned int border_x = ceil_to_multiple(input0->dimension(0), num_elems_read_per_iteration) - input0->dimension(0); + const unsigned int border_y = ceil_to_multiple(input0->dimension(1), num_rows_read_per_iteration) - input0->dimension(1); + + Window win = calculate_max_window(*input0, Steps(num_elems_read_per_iteration)); + + AccessWindowRectangle input0_access(input0, 0, 0, num_elems_read_per_iteration, num_rows_read_per_iteration); + AccessWindowHorizontal input1_access(input1, 0, num_elems_read_per_iteration); + AccessWindowStatic output_access(output, 0, 0, output->dimension(0) + border_x, output->dimension(1) + border_y); + + bool window_changed = update_window_and_padding(win, input0_access, input1_access, output_access); + + output->set_valid_region(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 + CLGEMMMatrixVectorMultiplyKernel::CLGEMMMatrixVectorMultiplyKernel() : _input0(nullptr), _input1(nullptr), _output(nullptr), _num_rows_read_per_iteration(0), _border_size(0) { @@ -45,11 +81,8 @@ BorderSize CLGEMMMatrixVectorMultiplyKernel::border_size() const void CLGEMMMatrixVectorMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); - ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input0->info()->data_type()) && (output->info()->data_type() != DataType::S32)); - ARM_COMPUTE_ERROR_ON(input0->info()->dimension(2) != input1->info()->dimension(1)); + ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info())); _input0 = input0; _input1 = input1; @@ -93,17 +126,17 @@ void CLGEMMMatrixVectorMultiplyKernel::configure(const ICLTensor *input0, const _border_size = BorderSize(border_y, border_x); - Window win = calculate_max_window(*input0->info(), Steps(num_elems_read_per_iteration)); - - AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_read_per_iteration, _num_rows_read_per_iteration); - AccessWindowHorizontal input1_access(input1->info(), 0, num_elems_read_per_iteration); - AccessWindowStatic output_access(_output->info(), 0, 0, _output->info()->dimension(0) + border_x, _output->info()->dimension(1) + border_y); - - update_window_and_padding(win, input0_access, input1_access, output_access); + auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info()); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + ICLKernel::configure(win_config.second); +} - _output->info()->set_valid_region(ValidRegion(Coordinates(), _output->info()->tensor_shape())); +Status CLGEMMMatrixVectorMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), input1->clone().get(), output->clone().get()).first); - ICLKernel::configure(win); + return Status{}; } void CLGEMMMatrixVectorMultiplyKernel::run(const Window &window, cl::CommandQueue &queue) diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp index 88bb0c417d..676a121a76 100644 --- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp @@ -24,6 +24,8 @@ #include "arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h" #include "arm_compute/core/CL/ICLTensor.h" +#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h" +#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h" #include "arm_compute/core/PixelValue.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" @@ -66,6 +68,21 @@ void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor _border_handler.configure(input, _kernel->border_size(), BorderMode::CONSTANT, zero_value); } +Status CLDepthwiseConvolutionLayer3x3::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) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); + ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() != DataLayout::NCHW && input->data_layout() != DataLayout::NHWC); + + if(input->data_layout() == DataLayout::NCHW) + { + return CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target); + } + + return CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info); +} + void CLDepthwiseConvolutionLayer3x3::run() { CLScheduler::get().enqueue(_border_handler); @@ -82,7 +99,6 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w { 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((input->info()->dimension(2) * depth_multiplier) != weights->info()->dimension(2)); const size_t weights_w = weights->info()->dimension(0); const size_t weights_h = weights->info()->dimension(1); @@ -168,6 +184,53 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w _v2mm_output.allocator()->allocate(); } +Status CLDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + unsigned int depth_multiplier) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); + ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(2) * depth_multiplier) != weights->dimension(2)); + + const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type()); + const bool append_bias = (biases != nullptr) && !is_quantized; + const TensorShape output_shape = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier); + const size_t weights_w = weights->dimension(0); + const size_t weights_h = weights->dimension(1); + const size_t weights_z = weights->dimension(2); + const unsigned int conv_w = output_shape.x(); + const unsigned int conv_h = output_shape.y(); + const size_t patch_size = weights_w * weights_h + ((append_bias) ? 1 : 0); + const size_t conv_size = conv_w * conv_h; + + TensorShape shape_im2col = input->tensor_shape(); + shape_im2col.set(0, patch_size); + shape_im2col.set(1, conv_size); + shape_im2col.set(2, weights_z); + TensorInfo input_reshaped(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col)); + ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseIm2ColKernel::validate(input, &input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier)); + + const TensorShape shape_weights_reshape(patch_size, weights_z); + TensorInfo weights_reshaped(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape)); + ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseWeightsReshapeKernel::validate(weights, &weights_reshaped, append_bias ? biases : nullptr)); + + DataType v2mm_dt = (input->data_type() == DataType::QASYMM8) ? DataType::S32 : input->data_type(); + TensorShape shape_v2mm_out = input->tensor_shape(); + shape_v2mm_out.set(0, conv_size * weights_z); + shape_v2mm_out.set(1, 1); + shape_v2mm_out.set(2, 1); + TensorInfo v2mm_output(input->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out)); + ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixVectorMultiplyKernel::validate(&input_reshaped, &weights_reshaped, &v2mm_output)); + + TensorInfo output_reshaped(v2mm_output.clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape)); + ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseVectorToTensorKernel::validate(&v2mm_output, (is_quantized) ? &output_reshaped : output, conv_w, conv_h)); + + if(is_quantized) + { + ARM_COMPUTE_RETURN_ON_ERROR(CLDirectConvolutionLayerOutputStageKernel::validate(&output_reshaped, biases, output)); + } + + return Status{}; +} + void CLDepthwiseConvolutionLayer::run() { // Run weights reshaping (Runs once for every configure) diff --git a/tests/validation/CL/DepthwiseConvolutionLayer.cpp b/tests/validation/CL/DepthwiseConvolutionLayer.cpp index 54b7925a09..093d342ce1 100644 --- a/tests/validation/CL/DepthwiseConvolutionLayer.cpp +++ b/tests/validation/CL/DepthwiseConvolutionLayer.cpp @@ -53,6 +53,171 @@ const auto depth_multipliers = framework::dataset::make("DepthMultiplier", { 1, TEST_SUITE(CL) TEST_SUITE(DepthwiseConvolutionLayer) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32, 0), // Mismatching data type input/weights + TensorInfo(TensorShape(32U, 18U, 3U), 1, DataType::F32, 0), // Mismatching input feature maps + TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32, 0), // Unsupported weights dimensions + TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::QASYMM8, 0), // Unsupported activation + TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32, 0), // Mismatching depth multiplier + TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32, 0), // Invalid stride + TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32, 0), // Invalid biases size + TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32, 0), // Invalid biases dimensions + TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32, 0), // Invalid output size + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0), // Window shrink + TensorInfo(TensorShape(32U, 18U, 8U), 1, DataType::F32, 0), + TensorInfo(TensorShape(50U, 32U, 8U), 1, DataType::QASYMM8, 0), + }), + framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F16, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(5U, 5U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::QASYMM8, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 24U), 1, DataType::QASYMM8, 0), + })), + framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U), 1, DataType::S32, 0), + TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(4U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(16U), 1, DataType::F32, 0), + TensorInfo(TensorShape(24U), 1, DataType::S32, 0), + })), + framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::QASYMM8, 0), + TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(30U, 16U, 16U), 1, DataType::F32, 0), + TensorInfo(TensorShape(48U, 30U, 24U), 1, DataType::QASYMM8, 0), + })), + framework::dataset::make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(4, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + })), + framework::dataset::make("DepthMultiplier", { 1, + 1, + 1, + 1, + 3, + 1, + 1, + 1, + 1, + 1, + 2, + 3, + })), + framework::dataset::make("ActivationInfo", { ActivationLayerInfo(), + ActivationLayerInfo(), + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR), + ActivationLayerInfo(), + ActivationLayerInfo(), + ActivationLayerInfo(), + ActivationLayerInfo(), + ActivationLayerInfo(), + ActivationLayerInfo(), + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + })), + framework::dataset::make("Expected", { false, false, false, false, false, false, false, false, false, false, true, true })), + input_info, weights_info, biases_info, output_info, conv_info, depth_multiplier, act_info, expected) +{ + bool is_valid = bool(CLDepthwiseConvolutionLayer3x3::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, depth_multiplier, act_info)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} + +DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0), // Mismatching data type input/weights + TensorInfo(TensorShape(27U, 13U, 3U), 1, DataType::F32, 0), // Mismatching input feature maps + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0), // Mismatching depth multiplier + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0), // Invalid biases size + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0), // Invalid biases dimensions + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0), // Invalid output size + TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32, 0), + TensorInfo(TensorShape(32U, 13U, 8U), 1, DataType::QASYMM8, 0), + }), + framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F16, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 24U), 1, DataType::QASYMM8, 0), + })), + framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(4U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(16U), 1, DataType::F32, 0), + TensorInfo(TensorShape(24U), 1, DataType::S32, 0), + })), + framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32, 0), + TensorInfo(TensorShape(32U, 11U, 24U), 1, DataType::QASYMM8, 0), + })), + framework::dataset::make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 1, 0), + })), + framework::dataset::make("DepthMultiplier", { 1, + 1, + 3, + 1, + 1, + 1, + 2, + 3, + })), + framework::dataset::make("Expected", { false, false, false, false, false, false, true, true })), + input_info, weights_info, biases_info, output_info, conv_info, depth_multiplier, expected) +{ + bool is_valid = bool(CLDepthwiseConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, depth_multiplier)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + template using CLDepthwiseConvolutionLayerFixture = DepthwiseConvolutionLayerValidationFixture; -- cgit v1.2.1