diff options
Diffstat (limited to 'src/core/CL/kernels')
10 files changed, 132 insertions, 96 deletions
diff --git a/src/core/CL/kernels/CLActivationLayerKernel.cpp b/src/core/CL/kernels/CLActivationLayerKernel.cpp index 5062fd1801..270eb78dcb 100644 --- a/src/core/CL/kernels/CLActivationLayerKernel.cpp +++ b/src/core/CL/kernels/CLActivationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2019 ARM Limited. + * Copyright (c) 2016-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -39,14 +39,14 @@ #include <cmath> #include <set> -using namespace arm_compute; - +namespace arm_compute +{ namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QASYMM8, DataType::QSYMM16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16, DataType::F16, DataType::F32); static std::set<ActivationLayerInfo::ActivationFunction> quantized_supported_activations = { @@ -63,12 +63,15 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized(data_type) && (quantized_supported_activations.count(f_act) == 0), "For Quantized data type only tanh, logistic, relu and lower/upper bounded relu are supported"); - ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 128))); - ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, 0))); + ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8 && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 128))); + ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8 && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, 0))); ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_symmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 32768.f, 0))); ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_symmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 32768.f, 0))); + ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 0))); + ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, -128))); + // Checks performed when output is configured if((output != nullptr) && (output->total_size() != 0)) { @@ -135,27 +138,11 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act const DataType dt = input->info()->data_type(); float a_const = act_info.a(); float b_const = act_info.b(); - int a_const_int = 0; - int b_const_int = 0; const ActivationLayerInfo::ActivationFunction f_act = act_info.activation(); const bool is_quantized = is_data_type_quantized(dt); const bool perform_activation_in_float = (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) || (f_act == ActivationLayerInfo::ActivationFunction::TANH); - // Create quantized version of constants a, b if needed - if(dt == DataType::QASYMM8) - { - const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform(); - a_const_int = quantize_qasymm8(a_const, iq_info); - b_const_int = quantize_qasymm8(b_const, iq_info); - } - else if(dt == DataType::QSYMM16) - { - const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform(); - a_const_int = quantize_qsymm16(a_const, iq_info); - b_const_int = quantize_qsymm16(b_const, iq_info); - } - // Set build options CLBuildOptions build_opts; build_opts.add_option_if(perform_activation_in_float, "-DFLOAT_DOMAIN"); @@ -164,28 +151,59 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(dt))); build_opts.add_option(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration))); - // Set A, B constants in build options - if(is_quantized && !perform_activation_in_float) - { - build_opts.add_option(("-DA_VAL=" + support::cpp11::to_string(a_const_int))); - build_opts.add_option(("-DB_VAL=" + support::cpp11::to_string(b_const_int))); - } - else - { - build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(a_const))); - build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(b_const))); - } + std::string kernel_name = std::string("activation_layer"); // Set quantization info build options if(is_quantized) { const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform(); + if(!perform_activation_in_float) + { + int a_const_int = 0; + int b_const_int = 0; + + // Create quantized version of constants a, b if needed + switch(dt) + { + case DataType::QASYMM8: + { + a_const_int = quantize_qasymm8(a_const, iq_info); + b_const_int = quantize_qasymm8(b_const, iq_info); + } + break; + case DataType::QASYMM8_SIGNED: + { + a_const_int = quantize_qasymm8_signed(a_const, iq_info); + b_const_int = quantize_qasymm8_signed(b_const, iq_info); + } + break; + case DataType::QSYMM16: + { + a_const_int = quantize_qsymm16(a_const, iq_info); + b_const_int = quantize_qsymm16(b_const, iq_info); + } + break; + default: + break; + } + build_opts.add_option(("-DA_VAL=" + support::cpp11::to_string(a_const_int))); + build_opts.add_option(("-DB_VAL=" + support::cpp11::to_string(b_const_int))); + } + else + { + build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(a_const))); + build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(b_const))); + } + // Quantized value of 0 corresponds to the offset o1 build_opts.add_option(("-DCONST_0=" + (is_data_type_quantized_asymmetric(dt) ? support::cpp11::to_string(iq_info.offset) : "0"))); build_opts.add_option(("-DS1_VAL=" + float_to_string_with_full_precision(iq_info.scale))); build_opts.add_option_if(is_data_type_quantized_asymmetric(dt), "-DO1_VAL=" + support::cpp11::to_string(iq_info.offset)); + // Set correct kernel name + kernel_name += perform_activation_in_float ? std::string("_quant_f32") : std::string("_quant"); + // Set scale and offset of the input and output if they have different quantization info if(output != nullptr) { @@ -198,14 +216,14 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act } } } - - // Create kernel - std::string kernel_name = std::string("activation_layer"); - if(is_quantized) + else { - kernel_name += perform_activation_in_float ? std::string("_quant_f32") : std::string("_quant"); + // Set A, B constants in build options for float types + build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(a_const))); + build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(b_const))); } + // Create kernel _kernel = create_opencl_kernel(_ctx, kernel_name, build_opts); // Make sure _kernel is initialized before calling the parent's configure _input = input; @@ -254,3 +272,4 @@ void CLActivationLayerKernel::run(const Window &window, cl::CommandQueue &queue) } while(collapsed.slide_window_slice_3D(slice)); } +} // namespace arm_compute diff --git a/src/core/CL/kernels/CLComparisonKernel.cpp b/src/core/CL/kernels/CLComparisonKernel.cpp index 5570ecfc79..afee429219 100644 --- a/src/core/CL/kernels/CLComparisonKernel.cpp +++ b/src/core/CL/kernels/CLComparisonKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 ARM Limited. + * Copyright (c) 2018-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -52,12 +52,7 @@ int calculate_num_elems_processed_per_iteration(const ITensorInfo &input) Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output, ComparisonOperation operation) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input1); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, - 1, - DataType::U8, DataType::S8, DataType::QASYMM8, - DataType::U16, DataType::S16, - DataType::U32, DataType::S32, - DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(input1.data_type() == DataType::UNKNOWN); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2); ARM_COMPUTE_RETURN_ERROR_ON(supported_comparison_ops.count(operation) == 0); @@ -132,7 +127,7 @@ void CLComparisonKernel::configure(const ICLTensor *input1, const ICLTensor *inp build_opts.emplace("-DVEC_SIZE=" + support::cpp11::to_string(calculate_num_elems_processed_per_iteration(*input1->info()))); build_opts.emplace("-DOP=" + operation_name); build_opts.emplace("-DOP_NAME=" + lower_string(operation_name)); - if(is_data_type_quantized_asymmetric(input1->info()->data_type())) + if(is_data_type_quantized(input1->info()->data_type())) { const UniformQuantizationInfo iq1_info = input1->info()->quantization_info().uniform(); const UniformQuantizationInfo iq2_info = input2->info()->quantization_info().uniform(); diff --git a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp b/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp index 81856769b2..7ec6841149 100644 --- a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp +++ b/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 ARM Limited. + * Copyright (c) 2018-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -64,7 +64,7 @@ void CLConvertFullyConnectedWeightsKernel::configure(const ICLTensor *input, ICL // Set build options CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); + build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size())); build_opts.add_option("-DFACTOR_1=" + support::cpp11::to_string(factor_1)); build_opts.add_option("-DFACTOR_2=" + support::cpp11::to_string(factor_2)); @@ -79,18 +79,15 @@ void CLConvertFullyConnectedWeightsKernel::configure(const ICLTensor *input, ICL Status CLConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout) { + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, - DataType::U8, DataType::S8, DataType::QASYMM8, - DataType::U16, DataType::S16, - DataType::U32, DataType::S32, - DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN); ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() != 2); ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != original_input_shape.total_size_lower(3)); ARM_COMPUTE_RETURN_ERROR_ON(data_layout == DataLayout::UNKNOWN); // Checks performed when output is configured - if((output != nullptr) && (output->total_size() != 0)) + if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); diff --git a/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp b/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp index cd9552f149..ee392032ca 100644 --- a/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp +++ b/src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp @@ -32,8 +32,8 @@ #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" -using namespace arm_compute; - +namespace arm_compute +{ CLDeconvolutionLayerUpsampleKernel::CLDeconvolutionLayerUpsampleKernel() : _input(nullptr), _output(nullptr), _info(), _data_layout(DataLayout::UNKNOWN) { @@ -45,7 +45,7 @@ Status CLDeconvolutionLayerUpsampleKernel::validate(const ITensorInfo *input, co ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); @@ -82,7 +82,7 @@ void CLDeconvolutionLayerUpsampleKernel::configure(const ICLTensor *input, ICLTe // Create kernel CLBuildOptions build_opts; - build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()))); + build_opts.add_option(("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()))); _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("deconvolution_upsample", build_opts.options())); constexpr unsigned int num_elems_processed_per_iteration = 1; @@ -156,3 +156,4 @@ void CLDeconvolutionLayerUpsampleKernel::run(const Window &window, cl::CommandQu ARM_COMPUTE_ERROR("Unsupported data layout"); } } +} // namespace arm_compute diff --git a/src/core/CL/kernels/CLDepthToSpaceLayerKernel.cpp b/src/core/CL/kernels/CLDepthToSpaceLayerKernel.cpp index 49a5590231..f23f7ce542 100644 --- a/src/core/CL/kernels/CLDepthToSpaceLayerKernel.cpp +++ b/src/core/CL/kernels/CLDepthToSpaceLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019 ARM Limited. + * Copyright (c) 2019-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -36,6 +36,7 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, int32_t block_shape) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN); ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4); ARM_COMPUTE_RETURN_ERROR_ON(block_shape < 2); @@ -81,7 +82,7 @@ void CLDepthToSpaceLayerKernel::configure(const ICLTensor *input, ICLTensor *out // Create kernel CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); + build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size())); build_opts.add_option("-DCHANNEL_SIZE=" + support::cpp11::to_string(input->info()->dimension(idx_channel))); build_opts.add_option("-DBLOCK_SHAPE=" + support::cpp11::to_string(block_shape)); build_opts.add_option("-DWIDTH_IN=" + support::cpp11::to_string(input->info()->dimension(idx_width))); diff --git a/src/core/CL/kernels/CLDequantizationLayerKernel.cpp b/src/core/CL/kernels/CLDequantizationLayerKernel.cpp index 60659faaaf..f85cb7636a 100644 --- a/src/core/CL/kernels/CLDequantizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLDequantizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 ARM Limited. + * Copyright (c) 2017-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -40,7 +40,7 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8, DataType::QSYMM16); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8, DataType::QSYMM16); if(output->tensor_shape().total_size() > 0) { diff --git a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp index 9e5d677e89..c158937839 100644 --- a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 ARM Limited. + * Copyright (c) 2017-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -90,11 +90,11 @@ void CLGEMMMatrixVectorMultiplyKernel::configure(const ICLTensor *input0, const _output = output; // Check if is a quantized operation - bool is_quantized = is_data_type_quantized_asymmetric(_input0->info()->data_type()); + const bool is_quantized = is_data_type_quantized_asymmetric(_input0->info()->data_type()); // Create kernel CLBuildOptions build_opts; - build_opts.add_option_if(!is_quantized, "-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type())); + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type())); build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input0->info()->dimension(0))); build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input0->info()->dimension(1))); diff --git a/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp b/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp index b255ba346f..220c2cd576 100644 --- a/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp +++ b/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 ARM Limited. + * Copyright (c) 2018-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -35,15 +35,15 @@ #include "support/ToolchainSupport.h" -using namespace arm_compute; - +namespace arm_compute +{ namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *std) { + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, mean, std); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, std); @@ -97,12 +97,8 @@ CLNormalizePlanarYUVLayerKernel::CLNormalizePlanarYUVLayerKernel() void CLNormalizePlanarYUVLayerKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *mean, const ICLTensor *std) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, mean, std); - - // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output->info(), *input->info()->clone()); - // Perform validation step + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, mean, std); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), mean->info(), std->info())); _input = input; @@ -183,3 +179,4 @@ void CLNormalizePlanarYUVLayerKernel::run(const Window &window, cl::CommandQueue } while(collapsed.slide_window_slice_3D(slice)); } +} // namespace arm_compute diff --git a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp index 50cdc9c7f4..6bdb1242a6 100644 --- a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp +++ b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2019 ARM Limited. + * Copyright (c) 2016-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -50,10 +50,18 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, ARM_COMPUTE_UNUSED(overflow_policy); ARM_COMPUTE_UNUSED(rounding_policy); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input2); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, + 1, + DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::S16, DataType::QSYMM16, DataType::F16, + DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, + 1, + DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::S16, DataType::QSYMM16, DataType::F16, + DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(scale < 0, "Scale cannot be negative."); const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape()); @@ -63,12 +71,17 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, // Validate in case of configured output if(output->total_size() > 0) { - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, + 1, + DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::S16, DataType::QSYMM16, DataType::F16, + DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::U8 && (input1->data_type() != DataType::U8 || input2->data_type() != DataType::U8), "Output can only be U8 if both inputs are U8"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::QASYMM8 && (input1->data_type() != DataType::QASYMM8 || input2->data_type() != DataType::QASYMM8), "Output can only be QASYMM8 if both inputs are QASYMM8"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::QASYMM8_SIGNED && (input1->data_type() != DataType::QASYMM8_SIGNED || input2->data_type() != DataType::QASYMM8_SIGNED), + "Output can only be QASYMM8_SIGNED if both inputs are QASYMM8_SIGNED"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::QSYMM16 && (input1->data_type() != DataType::QSYMM16 || input2->data_type() != DataType::QSYMM16), "Output can only be QSYMM16 if both inputs are QSYMM16"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output"); @@ -99,6 +112,10 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input1, ITe { set_data_type_if_unknown(*output, DataType::QASYMM8); } + else if(input1->data_type() == DataType::QASYMM8_SIGNED) + { + set_data_type_if_unknown(*output, DataType::QASYMM8_SIGNED); + } else if(input1->data_type() == DataType::QSYMM16) { set_data_type_if_unknown(*output, DataType::QSYMM16); diff --git a/src/core/CL/kernels/CLPoolingLayerKernel.cpp b/src/core/CL/kernels/CLPoolingLayerKernel.cpp index e3f1114f21..2d75e5f969 100644 --- a/src/core/CL/kernels/CLPoolingLayerKernel.cpp +++ b/src/core/CL/kernels/CLPoolingLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 ARM Limited. + * Copyright (c) 2017-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -40,7 +40,8 @@ #include <string> #include <tuple> -using namespace arm_compute; +namespace arm_compute +{ using namespace arm_compute::misc::shape_calculator; namespace @@ -57,19 +58,8 @@ void auto_init(const ITensorInfo *input, ITensorInfo *output, PoolingLayerInfo p Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - DataLayout data_layout = input->data_layout(); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - switch(data_layout) - { - case DataLayout::NCHW: - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); - break; - case DataLayout::NHWC: - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); - break; - default: - ARM_COMPUTE_ERROR("Data layout not supported"); - } + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG((is_data_type_quantized_asymmetric(input->data_type()) && pool_info.pool_type() == PoolingType::L2), "Unsupported combination of parameters!"); @@ -234,7 +224,25 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output, build_opts.add_option("-DPOOL_SIZE_X=" + support::cpp11::to_string(pool_size_x)); build_opts.add_option("-DPOOL_SIZE_Y=" + support::cpp11::to_string(pool_size_y)); - build_opts.add_option_if(data_type == DataType::F16, "-DFP16"); + // Set the initial value for the pooling operation accordingly with the data type + if(pool_type == PoolingType::MAX) + { + if(is_data_type_quantized(data_type)) + { + PixelValue type_min{}; + std::tie(type_min, std::ignore) = get_min_max(data_type); + build_opts.add_option("-DINITIAL_VALUE=" + support::cpp11::to_string(type_min.get<int32_t>())); + } + else + { + build_opts.add_option("-DINITIAL_VALUE=" + float_to_string_with_full_precision(std::numeric_limits<float>::lowest())); + } + } + else + { + // Pool AVG and Pool L2 initial value + build_opts.add_option("-DINITIAL_VALUE=0"); + } const auto use_fp_mixed_precision = (data_type == DataType::F16) && pool_info.fp_mixed_precision(); const auto use_wider_accumulator = use_fp_mixed_precision && (pool_type != PoolingType::MAX); @@ -389,3 +397,4 @@ void CLPoolingLayerKernel::run(const Window &window, cl::CommandQueue &queue) ARM_COMPUTE_ERROR("Not implemented"); } } +} // namespace arm_compute |