From 574775c7fa78a094bbeb7f9f87aca832936884e2 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 18 Feb 2019 20:08:02 +0000 Subject: COMPMID-1937: Adds support for DequantizationLayer for NEON/CL. Change-Id: I4b73edd176a277294e0e42e642460bc61210778a Signed-off-by: Georgios Pinitas Reviewed-on: https://review.mlplatform.org/c/744 Tested-by: Arm Jenkins Reviewed-by: Giuseppe Rossini --- src/core/CL/cl_kernels/dequantization_layer.cl | 85 ++++++---- .../CL/kernels/CLDequantizationLayerKernel.cpp | 93 +++++----- .../NEON/kernels/NEDequantizationLayerKernel.cpp | 187 ++++++++++++--------- 3 files changed, 202 insertions(+), 163 deletions(-) (limited to 'src/core') diff --git a/src/core/CL/cl_kernels/dequantization_layer.cl b/src/core/CL/cl_kernels/dequantization_layer.cl index 4908bb0b31..7307700473 100644 --- a/src/core/CL/cl_kernels/dequantization_layer.cl +++ b/src/core/CL/cl_kernels/dequantization_layer.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,51 +23,68 @@ */ #include "helpers.h" +#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(SCALE) && defined(OFFSET) + /** This performs the dequantization of 8-bit unsigned integers to floating point. * - * @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32 - * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image - * @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr - * @param[in] output_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image - * @param[in] min_max_ptr Pointer to the min/max vector. Minimum value in position 0, maximum value in position 1. Suppported data types: F32. - * @param[in] min_max_stride_x Stride of the min/max vector in X dimension (in bytes) - * @param[in] min_max_step_x min_max_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] min_max_offset_first_element_in_bytes The offset of the first element in the min/max vector + * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Quantization scale of input tensor is passed in with -DSCALE=scale. + * @note Quantization offset of input tensor is passed in with -DOFFSET=offset. + * + * @param[in] input_ptr Pointer to the source tensor. Supported data types: QASYMM8 + * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] output_ptr Pointer to the destination tensor. Supported data types: F16/F32 + * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor */ __kernel void dequantization_layer( TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(output), - VECTOR_DECLARATION(min_max)) + TENSOR3D_DECLARATION(output)) { // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); - Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); - Vector min_max = CONVERT_TO_VECTOR_STRUCT(min_max); - - // min_max_value.s0 = min, min_max_value.s1 = max - const float2 min_max_value = vload2(0, (__global float *)min_max.ptr); + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); + Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); - const float4 vmin = (float4)min_max_value.s0; - const float4 scale = (float4)((min_max_value.s1 - min_max_value.s0) / 255.0f); +#if defined(VEC_SIZE) && defined(LAST_ACCESSED_X) + // Check if access on width gets out of bounds + // If it does shift access vector to access elements within bounds + const int xi = (int)(get_global_id(0) * VEC_SIZE); + input.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * input_stride_x; + output.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * output_stride_x; // Load data - const uchar4 data = vload4(0, (__global uchar *)input.ptr); + VEC_DATA_TYPE(int, VEC_SIZE) + val = CONVERT(VLOAD(VEC_SIZE)(0, (__global uchar *)input.ptr), VEC_DATA_TYPE(int, VEC_SIZE)); + + // Create scale and offset vectors + const VEC_DATA_TYPE(float, VEC_SIZE) + vscale = SCALE; + + const VEC_DATA_TYPE(int, VEC_SIZE) + voffset = OFFSET; // Dequantize - const float4 res = convert_float4(data) * scale + vmin; + VEC_DATA_TYPE(float, VEC_SIZE) + res = vscale * CONVERT((val - voffset), VEC_DATA_TYPE(float, VEC_SIZE)); // Store result - vstore4(res, 0, (__global float *)output.ptr); + VSTORE(VEC_SIZE) + (CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), 0, (__global DATA_TYPE *)output.ptr); +#else // !defined(VEC_SIZE) || !defined(LAST_ACCESSED_X) + *((__global DATA_TYPE *)(output.ptr)) = (DATA_TYPE)((float)((int)(*((__global uchar *)(input.ptr))) - (int)(OFFSET)) * (float)(SCALE)); +#endif // defined(VEC_SIZE) && defined(LAST_ACCESSED_X) } + +#endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(SCALE) && defined(OFFSET) \ No newline at end of file diff --git a/src/core/CL/kernels/CLDequantizationLayerKernel.cpp b/src/core/CL/kernels/CLDequantizationLayerKernel.cpp index d4c1bec5f4..78cc5596dd 100644 --- a/src/core/CL/kernels/CLDequantizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLDequantizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -26,6 +26,7 @@ #include "arm_compute/core/AccessWindowStatic.h" #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/CL/CLValidate.h" #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" @@ -36,74 +37,78 @@ using namespace arm_compute; namespace { -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *min_max) +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) { - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output, min_max); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); - ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() < 3); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8); if(output->tensor_shape().total_size() > 0) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); } return Status{}; } -std::tuple validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *min_max) +std::tuple validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) { + // Configure kernel window + Window win = calculate_max_window(*input, Steps()); + // Output tensor auto initialization if not yet initialized auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::F32); - constexpr unsigned int num_elems_processed_per_iteration = 4; - - // Configure window - Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); - AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); - AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); - AccessWindowStatic min_max_access(min_max, 0, 0, 2, min_max->dimension(1)); - - // Update window and padding - bool window_changed = update_window_and_padding(win, input_access, output_access, min_max_access); + // CLDequantizationLayerKernel doesn't need padding so update_window_and_padding() can be skipped + Coordinates coord; + coord.set_num_dimensions(output->num_dimensions()); + output->set_valid_region(ValidRegion(coord, output->tensor_shape())); - output_access.set_valid_region(win, input->valid_region()); - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_tuple(err, win); + return std::make_tuple(Status{}, win); } } // namespace CLDequantizationLayerKernel::CLDequantizationLayerKernel() - : _input(nullptr), _output(nullptr), _min_max(nullptr) + : _input(nullptr), _output(nullptr) { } -void CLDequantizationLayerKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *min_max) +void CLDequantizationLayerKernel::configure(const ICLTensor *input, ICLTensor *output) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, min_max); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), min_max->info())); - - _input = input; - _output = output; - _min_max = min_max; + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info())); - // Create kernel - _kernel = static_cast(CLKernelLibrary::get().create_kernel("dequantization_layer")); + _input = input; + _output = output; - // Configure kernel window - auto win_config = validate_and_configure_window(input->info(), output->info(), min_max->info()); + const int vec_size_x = 16 / output->info()->element_size(); + const int output_width_x = output->info()->tensor_shape().x(); + const bool multi_access_x = (output_width_x / vec_size_x > 0); - ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); + // Create and update the window (if needed) + Window win = calculate_max_window(*output->info()); + if(multi_access_x) + { + win.set(Window::DimX, + Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x)); + } + ICLKernel::configure_internal(win); - ICLKernel::configure_internal(std::get<1>(win_config)); + // Create kernel + CLBuildOptions build_opts; + build_opts.add_option("-DSCALE=" + float_to_string_with_full_precision(input->info()->quantization_info().scale)); + build_opts.add_option("-DOFFSET=" + support::cpp11::to_string(input->info()->quantization_info().offset)); + build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(output->info()->data_type())); + build_opts.add_option_if(multi_access_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max(output_width_x - vec_size_x, 0))); + _kernel = static_cast(CLKernelLibrary::get().create_kernel("dequantization_layer", build_opts.options())); } -Status CLDequantizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *min_max) +Status CLDequantizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, min_max)); - ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get(), min_max->clone().get()))); - + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output)); + ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get()))); return Status{}; } @@ -115,20 +120,12 @@ void CLDequantizationLayerKernel::run(const Window &window, cl::CommandQueue &qu Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), 3); Window slice = window_collapsed.first_slice_window_3D(); - Window min_max_window = window; - min_max_window.set(Window::DimX, Window::Dimension(0, 0, 0)); - min_max_window.set(Window::DimY, Window::Dimension(0, _min_max->info()->dimension(1), 1)); - min_max_window.set(Window::DimZ, Window::Dimension(0, 0, 0)); - - Window min_max_slice = min_max_window.first_slice_window_1D(); - do { unsigned int idx = 0; add_3D_tensor_argument(idx, _input, slice); add_3D_tensor_argument(idx, _output, slice); - add_1D_tensor_argument(idx, _min_max, min_max_slice); enqueue(queue, *this, slice); } - while(window_collapsed.slide_window_slice_3D(slice) && min_max_window.slide_window_slice_1D(min_max_slice)); + while(window_collapsed.slide_window_slice_3D(slice)); } diff --git a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp index 47c895c594..119aa4ad9a 100644 --- a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -24,83 +24,143 @@ #include "arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h" #include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/CPP/Validate.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" +#include "arm_compute/core/NEON/NEAsymm.h" +#include "arm_compute/core/NEON/wrapper/wrapper.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" #include -using namespace arm_compute; - +namespace arm_compute +{ namespace { -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *min_max) +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) { - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output, min_max); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); - ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() < 3); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8); if(output->tensor_shape().total_size() > 0) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); } return Status{}; } -std::tuple validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *min_max) +std::tuple validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) { + // Configure kernel window + Window win = calculate_max_window(*input, Steps()); + // Output tensor auto initialization if not yet initialized auto_init_if_empty(*output, input->tensor_shape(), 1, DataType::F32); - constexpr unsigned int num_elems_processed_per_iteration = 8; + // NEDequantizationLayerKernel doesn't need padding so update_window_and_padding() can be skipped + Coordinates coord; + coord.set_num_dimensions(output->num_dimensions()); + output->set_valid_region(ValidRegion(coord, output->tensor_shape())); + + return std::make_tuple(Status{}, win); +} + +template +inline void store_result(T *ptr, const float32x4x4_t &v) +{ + ARM_COMPUTE_UNUSED(ptr, v); +} + +template <> +inline void store_result(float *ptr, const float32x4x4_t &v) +{ + wrapper::vstore(ptr, v.val[0]); + wrapper::vstore(ptr + 4, v.val[1]); + wrapper::vstore(ptr + 8, v.val[2]); + wrapper::vstore(ptr + 12, v.val[3]); +} + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +template <> +inline void store_result(float16_t *ptr, const float32x4x4_t &v) +{ + wrapper::vstore(ptr, vcombine_f16(vcvt_f16_f32(v.val[0]), vcvt_f16_f32(v.val[1]))); + wrapper::vstore(ptr + 8, vcombine_f16(vcvt_f16_f32(v.val[2]), vcvt_f16_f32(v.val[3]))); +} +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + +template +void run_dequantization(const ITensor *input, ITensor *output, const Window &window) +{ + const QuantizationInfo &qinfo = input->info()->quantization_info(); + + const int window_step_x = 16; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + // Collapse window and reset first dimension to handle tail calculations manually + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - // Configure window - Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); - AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); - AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); - AccessWindowStatic min_max_access(min_max, 0, 0, 2, min_max->dimension(1)); + // Create iterators + Iterator in(input, win_collapsed); + Iterator out(output, win_collapsed); - // Update window and padding - bool window_changed = update_window_and_padding(win, input_access, output_access, min_max_access); + execute_window_loop(win_collapsed, [&](const Coordinates & id) + { + const auto in_ptr = reinterpret_cast(in.ptr()); + const auto out_ptr = reinterpret_cast(out.ptr()); + + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin = wrapper::vloadq(in_ptr + x); + const auto vdeq = vdequantize(vin, qinfo); - output_access.set_valid_region(win, input->valid_region()); + store_result(reinterpret_cast(out_ptr + x), vdeq); + } - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_tuple(err, win); + // Compute left-over elements + for(; x < window_end_x; ++x) + { + uint8_t val = *(in_ptr + x); + *(out_ptr + x) = static_cast(qinfo.dequantize(val)); + } + }, + in, out); } } // namespace NEDequantizationLayerKernel::NEDequantizationLayerKernel() - : _input(nullptr), _output(nullptr), _min_max(nullptr) + : _input(nullptr), _output(nullptr) { } -void NEDequantizationLayerKernel::configure(const ITensor *input, ITensor *output, const ITensor *min_max) +void NEDequantizationLayerKernel::configure(const ITensor *input, ITensor *output) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, min_max); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), min_max->info())); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info())); - _input = input; - _output = output; - _min_max = min_max; + _input = input; + _output = output; // Configure kernel window - auto win_config = validate_and_configure_window(input->info(), output->info(), min_max->info()); + auto win_config = validate_and_configure_window(input->info(), output->info()); ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); INEKernel::configure(std::get<1>(win_config)); } -Status NEDequantizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *min_max) +Status NEDequantizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, min_max)); - ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get(), min_max->clone().get()))); - + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output)); + ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get()))); return Status{}; } @@ -110,53 +170,18 @@ void NEDequantizationLayerKernel::run(const Window &window, const ThreadInfo &in ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - Window window_input_output(window); - window_input_output.set(3, Window::Dimension(0, 1, 1)); - - Window window_min_max; - window_min_max.use_tensor_dimensions(_min_max->info()->tensor_shape()); - window_min_max.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(_input, window_input_output); - Iterator output(_output, window_input_output); - Iterator min_max(_min_max, window_min_max); - - execute_window_loop(window_min_max, [&](const Coordinates & id_batch) + switch(_output->info()->data_type()) { - // Get the min and max - const float min = *(reinterpret_cast(min_max.ptr()) + 0); - const float max = *(reinterpret_cast(min_max.ptr()) + 1); - - const float32x4_t vmin = vdupq_n_f32(min); - const float range = max - min; - const float32x4_t scaling = vdupq_n_f32(range / 255.0f); - - // Uniformly map values to range 8bit integers, i.e. [min, max] -> [0, 255] - execute_window_loop(window_input_output, [&](const Coordinates & id) - { - // Get the input values - const auto input_ptr = reinterpret_cast(input.ptr() + id_batch[1] * _input->info()->strides_in_bytes()[3]); - - const uint8x8_t val_u8 = vld1_u8(input_ptr); - const uint16x8_t val_u16 = vmovl_u8(val_u8); - const uint32x4_t val_u32_low = vmovl_u16(vget_low_u16(val_u16)); - const uint32x4_t val_u32_high = vmovl_u16(vget_high_u16(val_u16)); - float32x4_t val_low = vcvtq_f32_u32(val_u32_low); - float32x4_t val_high = vcvtq_f32_u32(val_u32_high); - - // Dequantize -> (q / 255.0 * range) + min - val_low = vmulq_f32(val_low, scaling); - val_high = vmulq_f32(val_high, scaling); - val_low = vaddq_f32(val_low, vmin); - val_high = vaddq_f32(val_high, vmin); - - const float32x4x2_t dequantized = vuzpq_f32(val_low, val_high); - - // Store the dequantized values - auto output_ptr = reinterpret_cast(output.ptr() + id_batch[1] * _output->info()->strides_in_bytes()[3]); - vst2q_f32(output_ptr, dequantized); - }, - input, output); - }, - min_max); -} \ No newline at end of file + case DataType::F32: + run_dequantization(_input, _output, window); + break; +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + run_dequantization(_input, _output, window); + break; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + default: + ARM_COMPUTE_ERROR("Unsupported data type."); + } +} +} // namespace arm_compute \ No newline at end of file -- cgit v1.2.1