From 7485d5a62685cb745ab50e970adb722cb71557ac Mon Sep 17 00:00:00 2001 From: Vidhya Sudhan Loganathan Date: Wed, 4 Jul 2018 09:34:00 +0100 Subject: COMPMID-970 : Remove QS8 / QS16 support Removed fixed point related code. Change-Id: I487acf138dace3b0450e0d72ca7071eaec254566 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/137678 Tested-by: Jenkins Reviewed-by: Anthony Barbier --- .../NEON/kernels/NEDepthConvertLayerKernel.cpp | 187 +-------------------- 1 file changed, 5 insertions(+), 182 deletions(-) (limited to 'src/core/NEON/kernels/NEDepthConvertLayerKernel.cpp') diff --git a/src/core/NEON/kernels/NEDepthConvertLayerKernel.cpp b/src/core/NEON/kernels/NEDepthConvertLayerKernel.cpp index c29cb57513..8280b52fcb 100644 --- a/src/core/NEON/kernels/NEDepthConvertLayerKernel.cpp +++ b/src/core/NEON/kernels/NEDepthConvertLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -40,13 +40,13 @@ class Coordinates; } // namespace arm_compute NEDepthConvertLayerKernel::NEDepthConvertLayerKernel() - : _input(nullptr), _output(nullptr), _policy(), _shift(0), _fixed_point_position_input(0), _fixed_point_position_output(0) + : _input(nullptr), _output(nullptr), _policy(), _shift(0) { } void NEDepthConvertLayerKernel::configure(ITensor *input, ITensor *output, ConvertPolicy policy, uint32_t shift) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QS8, DataType::S16, DataType::U16, DataType::QS16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16, DataType::U16); _input = input; _output = input; @@ -58,48 +58,26 @@ void NEDepthConvertLayerKernel::configure(ITensor *input, ITensor *output, Conve // Auto initialize output shape if not initialized (We can only auto-configure the shape, datatype must be given) set_shape_if_empty(*output->info(), input->info()->tensor_shape()); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::S16, DataType::U16, DataType::QS16, DataType::U32, DataType::S32, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); // Set output _output = output; } - // Set initial fixed point position of input and output - _fixed_point_position_input = input->info()->fixed_point_position(); - _fixed_point_position_output = _output->info()->fixed_point_position(); - - // Set the fixed point position to the output tensor if needed - if(is_data_type_fixed_point(input->info()->data_type()) && is_data_type_fixed_point(_output->info()->data_type())) - { - // If in-place set the fixed point position of the output tensor to be equal to shift - _fixed_point_position_output = (_input == _output) ? static_cast(_shift) : _fixed_point_position_output; - // Set fixed point position to output tensor - _output->info()->set_fixed_point_position(_fixed_point_position_output); - } - - ARM_COMPUTE_ERROR_ON(shift >= 8 && (!is_data_type_fixed_point(input->info()->data_type()) && !is_data_type_fixed_point(output->info()->data_type()))); + ARM_COMPUTE_ERROR_ON(shift >= 8); ARM_COMPUTE_ERROR_ON(input == output && (data_size_from_type(input->info()->data_type()) != data_size_from_type(output->info()->data_type()))); ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U8 && (output->info()->data_type() != DataType::S16 && output->info()->data_type() != DataType::U16 && output->info()->data_type() != DataType::S32), "Only data_types supported [in] U8 -> [out] U16, S16, S32"); - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::QS8 && (output->info()->data_type() != DataType::QS8 && output->info()->data_type() != DataType::F32), - "Only data_types supported [in] QS8 -> [out] QS8, F32"); - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U16 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::U32), "Only data_types supported [in] U16 -> [out] U8, U32"); ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::S16 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::S32), "Only data_types supported [in] S16 -> [out] U8, S32"); - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::QS16 && (output->info()->data_type() != DataType::QS16 && output->info()->data_type() != DataType::F32), - "Only data_types supported [in] QS16 -> [out] QS16, F32"); - - ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::F32 && (output->info()->data_type() != DataType::QS8 && output->info()->data_type() != DataType::QS16), - "Only data_types supported [in] F32 -> [out] QS8, QS16"); - constexpr unsigned int num_elems_processed_per_iteration = 16; // Configure kernel window @@ -132,8 +110,6 @@ void NEDepthConvertLayerKernel::run(const Window &window, const ThreadInfo &info Iterator input(_input, window); Iterator output(_output, window); - bool in_place = (_input == _output); - switch(_input->info()->data_type()) { case DataType::U8: @@ -212,49 +188,6 @@ void NEDepthConvertLayerKernel::run(const Window &window, const ThreadInfo &info } break; } - case DataType::QS8: - { - switch(_output->info()->data_type()) - { - case DataType::QS8: - { - const int relative_shift = _fixed_point_position_output - _fixed_point_position_input; - /* Fixed point position conversion QS8 -> QS8 */ - if(relative_shift != 0 || !in_place) - { - const auto relative_shift_vec = vdupq_n_qs8(relative_shift); - execute_window_loop(window, [&](const Coordinates & id) - { - const qint8x16_t texels_qs8 = vld1q_qs8(reinterpret_cast(input.ptr())); - vst1q_qs8(reinterpret_cast(output.ptr()), vqrshlq_s8(texels_qs8, relative_shift_vec)); - }, - input, output); - } - break; - } - case DataType::F32: - { - /* Up-conversion QS8 -> F32 */ - execute_window_loop(window, [&](const Coordinates & id) - { - const qint8x16_t texels_qs8 = vld1q_qs8(reinterpret_cast(input.ptr())); - - float32x4x2_t texels_low = vcvt_f32_qs8(vget_low_s8(texels_qs8), _fixed_point_position_input); - float32x4x2_t texels_high = vcvt_f32_qs8(vget_high_s8(texels_qs8), _fixed_point_position_input); - - vst1q_f32(reinterpret_cast(output.ptr()), texels_low.val[0]); - vst1q_f32(reinterpret_cast(output.ptr()) + 4, texels_low.val[1]); - vst1q_f32(reinterpret_cast(output.ptr()) + 8, texels_high.val[0]); - vst1q_f32(reinterpret_cast(output.ptr()) + 12, texels_high.val[1]); - }, - input, output); - break; - } - default: - ARM_COMPUTE_ERROR("Output data type not supported"); - } - break; - } case DataType::S16: { switch(_output->info()->data_type()) @@ -408,116 +341,6 @@ void NEDepthConvertLayerKernel::run(const Window &window, const ThreadInfo &info } break; } - case DataType::QS16: - { - switch(_output->info()->data_type()) - { - case DataType::QS16: - { - const int relative_shift = _fixed_point_position_output - _fixed_point_position_input; - /* Fixed point position conversion QS16 -> QS16 */ - if(relative_shift != 0 || !in_place) - { - const auto relative_shift_vec = vdupq_n_qs16(relative_shift); - execute_window_loop(window, [&](const Coordinates & id) - { - const qint16x8x2_t texels_qs16 = - { - { - vld1q_qs16(reinterpret_cast(input.ptr())), - vld1q_qs16(reinterpret_cast(input.ptr()) + 8) - } - }; - vst1q_qs16(reinterpret_cast(output.ptr()), vqrshlq_s16(texels_qs16.val[0], relative_shift_vec)); - vst1q_qs16(reinterpret_cast(output.ptr()) + 8, vqrshlq_s16(texels_qs16.val[1], relative_shift_vec)); - }, - input, output); - } - break; - } - case DataType::F32: - { - /* Up-conversion QS16 -> F32 */ - execute_window_loop(window, [&](const Coordinates & id) - { - const int16x8x2_t texels_qs16 = - { - { - vld1q_s16(reinterpret_cast(input.ptr())), - vld1q_s16(reinterpret_cast(input.ptr()) + 8) - } - }; - - vst1q_f32(reinterpret_cast(output.ptr()), vcvt_f32_qs16(vget_low_s16(texels_qs16.val[0]), _fixed_point_position_input)); - vst1q_f32(reinterpret_cast(output.ptr()) + 4, vcvt_f32_qs16(vget_high_s16(texels_qs16.val[0]), _fixed_point_position_input)); - vst1q_f32(reinterpret_cast(output.ptr()) + 8, vcvt_f32_qs16(vget_low_s16(texels_qs16.val[1]), _fixed_point_position_input)); - vst1q_f32(reinterpret_cast(output.ptr()) + 12, vcvt_f32_qs16(vget_high_s16(texels_qs16.val[1]), _fixed_point_position_input)); - }, - input, output); - break; - } - default: - ARM_COMPUTE_ERROR("Output data type not supported"); - } - break; - } - case DataType::F32: - { - switch(_output->info()->data_type()) - { - case DataType::QS8: - { - /* Down-conversion F32 -> QS8 */ - execute_window_loop(window, [&](const Coordinates & id) - { - const float32x4x4_t texels_f32 = - { - { - vld1q_f32(reinterpret_cast(input.ptr())), - vld1q_f32(reinterpret_cast(input.ptr()) + 4), - vld1q_f32(reinterpret_cast(input.ptr()) + 8), - vld1q_f32(reinterpret_cast(input.ptr()) + 12) - } - }; - - const qint8x16_t texels_s8 = vqcvtq_qs8_f32(texels_f32, _fixed_point_position_output); - - vst1q_s8(reinterpret_cast(output.ptr()), texels_s8); - }, - input, output); - break; - } - case DataType::QS16: - { - /* Down-conversion F32 -> QS16 */ - execute_window_loop(window, [&](const Coordinates & id) - { - const float32x4x2_t texels_f32_1 = - { - { - vld1q_f32(reinterpret_cast(input.ptr())), - vld1q_f32(reinterpret_cast(input.ptr()) + 4), - } - }; - const float32x4x2_t texels_f32_2 = - { - { - vld1q_f32(reinterpret_cast(input.ptr()) + 8), - vld1q_f32(reinterpret_cast(input.ptr()) + 12) - } - }; - - vst1q_s16(reinterpret_cast(output.ptr()), vqcvtq_qs16_f32(texels_f32_1, _fixed_point_position_output)); - vst1q_s16(reinterpret_cast(output.ptr()) + 8, vqcvtq_qs16_f32(texels_f32_2, _fixed_point_position_output)); - }, - input, output); - break; - } - default: - ARM_COMPUTE_ERROR("Output data type not supported"); - } - break; - } default: ARM_COMPUTE_ERROR("Not supported"); } -- cgit v1.2.1