/* * Copyright (c) 2016-2020 ARM Limited. * * SPDX-License-Identifier: MIT * * Permission is hereby granted, free of charge, to any person obtaining a copy * of this software and associated documentation files (the "Software"), to * deal in the Software without restriction, including without limitation the * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or * sell copies of the Software, and to permit persons to whom the Software is * furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included in all * copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #include "arm_compute/core/NEON/kernels/NEDepthConvertLayerKernel.h" #include "arm_compute/core/CPP/Validate.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/NEON/NEFixedPoint.h" #include "arm_compute/core/NEON/NEMath.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/SaturateCast.h" #include "arm_compute/core/NEON/wrapper/wrapper.h" using namespace arm_compute; namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift) { ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(output); ARM_COMPUTE_RETURN_ERROR_ON_CPU_BF16_UNSUPPORTED(input); ARM_COMPUTE_RETURN_ERROR_ON_CPU_BF16_UNSUPPORTED(output); ARM_COMPUTE_UNUSED(policy); ARM_COMPUTE_RETURN_ERROR_ON(input == output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8_SIGNED, DataType::QASYMM8, DataType::U8, DataType::S16, DataType::U16, DataType::BFLOAT16, DataType::F16, DataType::F32, DataType::S32); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8_SIGNED, DataType::QASYMM8, DataType::U8, DataType::S16, DataType::U16, DataType::BFLOAT16, DataType::F16, DataType::U32, DataType::S32, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(shift >= 8); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QASYMM8_SIGNED && (output->data_type() != DataType::S16 && output->data_type() != DataType::S32 && output->data_type() != DataType::F16 && output->data_type() != DataType::F32), "Only data_types supported [in] QASYMM8 -> [out] U16, S16, S32, F16, F32"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QASYMM8 && (output->data_type() != DataType::S16 && output->data_type() != DataType::U16 && output->data_type() != DataType::S32 && output->data_type() != DataType::F16 && output->data_type() != DataType::F32), "Only data_types supported [in] QASYMM8 -> [out] U16, S16, S32, F16, F32"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::U8 && (output->data_type() != DataType::S16 && output->data_type() != DataType::U16 && output->data_type() != DataType::S32 && output->data_type() != DataType::F16 && output->data_type() != DataType::F32), "Only data_types supported [in] U8 -> [out] U16, S16, S32, F16, F32"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::U16 && (output->data_type() != DataType::U8 && output->data_type() != DataType::U32), "Only data_types supported [in] U16 -> [out] U8, U32"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::S16 && (output->data_type() != DataType::QASYMM8_SIGNED && output->data_type() != DataType::U8 && output->data_type() != DataType::S32), "Only data_types supported [in] S16 -> [out] U8, S32"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::BFLOAT16 && output->data_type() != DataType::F32, "Only data_types supported [in] BFLOAT16 -> [out] F32"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::F16 && (output->data_type() != DataType::QASYMM8_SIGNED && output->data_type() != DataType::QASYMM8 && output->data_type() != DataType::U8 && output->data_type() != DataType::F32 && output->data_type() != DataType::S32), "Only data_types supported [in] F16 -> [out] QASYMM8, F32, S32, U8"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::F32 && (output->data_type() != DataType::QASYMM8_SIGNED && output->data_type() != DataType::QASYMM8 && output->data_type() != DataType::F16 && output->data_type() != DataType::BFLOAT16 && output->data_type() != DataType::S32 && output->data_type() != DataType::U8), "Only data_types supported [in] F32 -> [out] QASYMM8, BFLOAT16, F16, S32, U8"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::S32 && (output->data_type() != DataType::QASYMM8_SIGNED && output->data_type() != DataType::QASYMM8 && output->data_type() != DataType::F16 && output->data_type() != DataType::F32 && output->data_type() != DataType::U8), "Only data_types supported [in] S32 -> [out] QASYMM8, F16, F32, U8"); // Validate in case of configured output if(output->total_size() > 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); } return Status{}; } } // namespace NEDepthConvertLayerKernel::NEDepthConvertLayerKernel() : _input(nullptr), _output(nullptr), _policy(), _shift(0) { } void NEDepthConvertLayerKernel::configure(const ITensor *input, ITensor *output, ConvertPolicy policy, uint32_t shift) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); // 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()); _input = input; _output = output; _policy = policy; _shift = shift; ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), policy, shift)); // Configure kernel window Window win = calculate_max_window(*input->info(), Steps()); Coordinates coord; coord.set_num_dimensions(output->info()->num_dimensions()); output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape())); ICPPKernel::configure(win); } Status NEDepthConvertLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, policy, shift)); return Status{}; } void NEDepthConvertLayerKernel::run(const Window &window, const ThreadInfo &info) { ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); ARM_COMPUTE_ERROR_ON_NULLPTR(_input, _output); ARM_COMPUTE_ERROR_ON(_input == _output); const auto window_start_x = static_cast(window.x().start()); const auto window_end_x = static_cast(window.x().end()); const int window_step_x = 16; Window win{ window }; win.set(Window::DimX, Window::Dimension(0, 1, 1)); Iterator input(_input, win); Iterator output(_output, win); switch(_input->info()->data_type()) { case DataType::QASYMM8_SIGNED: { const int16x8_t b = vdupq_n_s16(_shift); switch(_output->info()->data_type()) { case DataType::S16: { /* Up-conversion QASYMM8_SIGNED -> S16 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int8x16_t texels_s8 = vld1q_s8(input_ptr + x); const int16x8x2_t texels = { { vshlq_s16(vmovl_s8(vget_low_s8(texels_s8)), b), vshlq_s16(vmovl_s8(vget_high_s8(texels_s8)), b) } }; vst1q_s16(output_ptr + x, texels.val[0]); vst1q_s16(output_ptr + x + 8, texels.val[1]); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) << _shift); } }, input, output); break; } case DataType::S32: { /* Up-conversion QASYMM8_SIGNED -> S32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int8x16_t texels_s8 = vld1q_s8(input_ptr + x); const int16x8x2_t texels = { { vshlq_s16(vmovl_s8(vget_low_s8(texels_s8)), b), vshlq_s16(vmovl_s8(vget_high_s8(texels_s8)), b) } }; vst1q_s32(output_ptr + x, vmovl_s16(vget_low_s16(texels.val[0]))); vst1q_s32(output_ptr + x + 4, vmovl_s16(vget_high_s16(texels.val[0]))); vst1q_s32(output_ptr + x + 8, vmovl_s16(vget_low_s16(texels.val[1]))); vst1q_s32(output_ptr + x + 12, vmovl_s16(vget_high_s16(texels.val[1]))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) << _shift); } }, input, output); break; } case DataType::F32: { /* Up-conversion QASYMM8_SIGNED -> F32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int8x16_t texels_s8 = vld1q_s8(reinterpret_cast(input.ptr())); const int16x8x2_t texels = { { vshlq_s16(vmovl_s8(vget_low_s8(texels_s8)), b), vshlq_s16(vmovl_s8(vget_high_s8(texels_s8)), b) } }; vst1q_f32(output_ptr + x, vcvtq_f32_s32(vmovl_s16(vget_low_s16(texels.val[0])))); vst1q_f32(output_ptr + x + 4, vcvtq_f32_s32(vmovl_s16(vget_high_s16(texels.val[0])))); vst1q_f32(output_ptr + x + 8, vcvtq_f32_s32(vmovl_s16(vget_low_s16(texels.val[1])))); vst1q_f32(output_ptr + x + 12, vcvtq_f32_s32(vmovl_s16(vget_high_s16(texels.val[1])))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) << _shift); } }, input, output); break; } #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: { /* Up-conversion QASYMM8_SIGNED -> F16 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int8x16_t texels_s8 = vld1q_s8(input_ptr + x); const int16x8x2_t texels = { { vshlq_s16(vmovl_s8(vget_low_s8(texels_s8)), b), vshlq_s16(vmovl_s8(vget_high_s8(texels_s8)), b) } }; vst1q_f16(output_ptr + x, vcvtq_f16_s16(texels.val[0])); vst1q_f16(output_ptr + x + 8, vcvtq_f16_s16(texels.val[1])); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) << _shift); } }, input, output); break; } #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC default: ARM_COMPUTE_ERROR("Output data type not supported"); } break; } case DataType::QASYMM8: case DataType::U8: { const int16x8_t b = vdupq_n_s16(_shift); switch(_output->info()->data_type()) { case DataType::S16: { /* Up-conversion U8 -> S16 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const uint8x16_t texels_u8 = vld1q_u8(input_ptr + x); const int16x8x2_t texels = { { vshlq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(texels_u8))), b), vshlq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(texels_u8))), b) } }; vst1q_s16(output_ptr + x, texels.val[0]); vst1q_s16(output_ptr + x + 8, texels.val[1]); } // Compute left-over elements for(; x < window_end_x; ++x) { auto in = static_cast(*(input_ptr + x)); *(output_ptr + x) = in << _shift; } }, input, output); break; } case DataType::S32: { /* Up-conversion U8 -> S32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const uint8x16_t texels_u8 = vld1q_u8(input_ptr + x); const int16x8x2_t texels = { { vshlq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(texels_u8))), b), vshlq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(texels_u8))), b) } }; vst1q_s32(output_ptr + x, vmovl_s16(vget_low_s16(texels.val[0]))); vst1q_s32(output_ptr + x + 4, vmovl_s16(vget_high_s16(texels.val[0]))); vst1q_s32(output_ptr + x + 8, vmovl_s16(vget_low_s16(texels.val[1]))); vst1q_s32(output_ptr + x + 12, vmovl_s16(vget_high_s16(texels.val[1]))); } // Compute left-over elements for(; x < window_end_x; ++x) { auto in = static_cast(*(input_ptr + x)); *(output_ptr + x) = in << _shift; } }, input, output); break; } case DataType::F32: { /* Up-conversion U8 -> F32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const uint8x16_t texels_u8 = vld1q_u8(input_ptr + x); const int16x8x2_t texels = { { vshlq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(texels_u8))), b), vshlq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(texels_u8))), b) } }; vst1q_f32(output_ptr + x, vcvtq_f32_s32(vmovl_s16(vget_low_s16(texels.val[0])))); vst1q_f32(output_ptr + x + 4, vcvtq_f32_s32(vmovl_s16(vget_high_s16(texels.val[0])))); vst1q_f32(output_ptr + x + 8, vcvtq_f32_s32(vmovl_s16(vget_low_s16(texels.val[1])))); vst1q_f32(output_ptr + x + 12, vcvtq_f32_s32(vmovl_s16(vget_high_s16(texels.val[1])))); } // Compute left-over elements for(; x < window_end_x; ++x) { auto in = static_cast(*(input_ptr + x)); *(output_ptr + x) = static_cast(in << _shift); } }, input, output); break; } #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: { /* Up-conversion U8 -> F16 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const uint8x16_t texels_u8 = vld1q_u8(input_ptr + x); const int16x8x2_t texels = { { vshlq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(texels_u8))), b), vshlq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(texels_u8))), b) } }; vst1q_f16(output_ptr + x, vcvtq_f16_s16(texels.val[0])); vst1q_f16(output_ptr + x + 8, vcvtq_f16_s16(texels.val[1])); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) << _shift); } }, input, output); break; } #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::U16: { /* Up-conversion U8 -> U16 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const uint8x16_t texels_u8 = vld1q_u8(input_ptr + x); const uint16x8x2_t texels = { { vshlq_u16(vmovl_u8(vget_low_u8(texels_u8)), b), vshlq_u16(vmovl_u8(vget_high_u8(texels_u8)), b) } }; vst1q_u16(output_ptr + x, texels.val[0]); vst1q_u16(output_ptr + x + 8, texels.val[1]); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x)) << _shift; } }, input, output); break; } default: ARM_COMPUTE_ERROR("Output data type not supported"); } break; } case DataType::S16: { switch(_output->info()->data_type()) { case DataType::QASYMM8_SIGNED: { const int16x8_t b = vdupq_n_s16(-static_cast(_shift)); /* Down-conversion S16 -> QASYMM8_SIGNED */ if(ConvertPolicy::SATURATE == _policy) { execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int16x8x2_t texels = { { vqshlq_s16(vld1q_s16(input_ptr + x), b), vqshlq_s16(vld1q_s16(input_ptr + x + 8), b) } }; vst1q_s8(output_ptr + x, vcombine_s8(vqmovn_s16(texels.val[0]), vqmovn_s16(texels.val[1]))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = utils::cast::saturate_cast(*(input_ptr + x) >> _shift); } }, input, output); } else { execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int16x8x2_t texels = { { vshlq_s16(vld1q_s16(input_ptr + x), b), vshlq_s16(vld1q_s16(input_ptr + x + 8), b) } }; vst1q_s8(output_ptr + x, vcombine_s8(vmovn_s16(texels.val[0]), vmovn_s16(texels.val[1]))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) >> _shift); } }, input, output); } break; } case DataType::U8: { const int16x8_t b = vdupq_n_s16(-static_cast(_shift)); /* Down-conversion S16 -> U8 */ if(ConvertPolicy::SATURATE == _policy) { execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int16x8x2_t texels = { { vqshlq_s16(vld1q_s16(input_ptr + x), b), vqshlq_s16(vld1q_s16(input_ptr + x + 8), b) } }; vst1q_u8(output_ptr + x, vcombine_u8(vqmovun_s16(texels.val[0]), vqmovun_s16(texels.val[1]))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = utils::cast::saturate_cast(*(input_ptr + x) >> _shift); } }, input, output); } else { execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int16x8x2_t texels = { { vshlq_s16(vld1q_s16(input_ptr + x), b), vshlq_s16(vld1q_s16(input_ptr + x + 8), b) } }; vst1q_u8(output_ptr + x, vcombine_u8(vmovn_u16(vreinterpretq_u16_s16(texels.val[0])), vmovn_u16(vreinterpretq_u16_s16(texels.val[1])))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) >> _shift); } }, input, output); } break; } case DataType::S32: { const int32x4_t b = vdupq_n_s32(_shift); /* Up-conversion S16 -> S32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int16x8x2_t texels = { { vld1q_s16(input_ptr + x), vld1q_s16(input_ptr + x + 8) } }; const int32x4x4_t texels_s32 = { { vshlq_s32(vmovl_s16(vget_low_s16(texels.val[0])), b), vshlq_s32(vmovl_s16(vget_high_s16(texels.val[0])), b), vshlq_s32(vmovl_s16(vget_low_s16(texels.val[1])), b), vshlq_s32(vmovl_s16(vget_high_s16(texels.val[1])), b) } }; vst1q_s32(output_ptr + x, texels_s32.val[0]); vst1q_s32(output_ptr + x + 4, texels_s32.val[1]); vst1q_s32(output_ptr + x + 8, texels_s32.val[2]); vst1q_s32(output_ptr + x + 12, texels_s32.val[3]); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) << _shift); } }, input, output); break; } default: ARM_COMPUTE_ERROR("Output data type not supported"); } break; } case DataType::U16: { switch(_output->info()->data_type()) { case DataType::U8: { const int16x8_t b = vdupq_n_s16(-static_cast(_shift)); /* Down-conversion U16 -> U8 */ if(ConvertPolicy::SATURATE == _policy) { execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const uint16x8x2_t texels = { { vqshlq_u16(vld1q_u16(input_ptr + x), b), vqshlq_u16(vld1q_u16(input_ptr + x + 8), b) } }; vst1q_u8(output_ptr + x, vcombine_u8(vqmovn_u16(texels.val[0]), vqmovn_u16(texels.val[1]))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = utils::cast::saturate_cast(*(input_ptr + x) >> _shift); } }, input, output); } else { execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const uint16x8x2_t texels = { { vshlq_u16(vld1q_u16(input_ptr + x), b), vshlq_u16(vld1q_u16(input_ptr + x + 8), b) } }; vst1q_u8(output_ptr + x, vcombine_u8(vmovn_u16(texels.val[0]), vmovn_u16(texels.val[1]))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) >> _shift); } }, input, output); } break; } case DataType::U32: { const int32x4_t b = vdupq_n_s32(_shift); /* Up-conversion U16 -> U32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const uint16x8x2_t texels = { { vld1q_u16(input_ptr + x), vld1q_u16(input_ptr + x + 8) } }; vst1q_u32(output_ptr + x, vshlq_u32(vmovl_u16(vget_low_u16(texels.val[0])), b)); vst1q_u32(output_ptr + x + 4, vshlq_u32(vmovl_u16(vget_high_u16(texels.val[0])), b)); vst1q_u32(output_ptr + x + 8, vshlq_u32(vmovl_u16(vget_low_u16(texels.val[1])), b)); vst1q_u32(output_ptr + x + 12, vshlq_u32(vmovl_u16(vget_high_u16(texels.val[1])), b)); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) << _shift); } }, input, output); break; } default: ARM_COMPUTE_ERROR("Output data type not supported"); } break; } #if defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) case DataType::BFLOAT16: switch(_output->info()->data_type()) { case DataType::F32: { /* Up-conversion BFLOAT16 -> F32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const uint16x8x2_t texels = { { vld1q_u16(reinterpret_cast(input.ptr())), vld1q_u16(reinterpret_cast(input.ptr()) + 8) } }; vst1q_f32(reinterpret_cast(output.ptr()), vreinterpretq_f32_u32(vshlq_n_u32(vmovl_u16(vget_low_u16(texels.val[0])), 16))); vst1q_f32(reinterpret_cast(output.ptr()) + 4, vreinterpretq_f32_u32(vshlq_n_u32(vmovl_u16(vget_high_u16(texels.val[0])), 16))); vst1q_f32(reinterpret_cast(output.ptr()) + 8, vreinterpretq_f32_u32(vshlq_n_u32(vmovl_u16(vget_low_u16(texels.val[1])), 16))); vst1q_f32(reinterpret_cast(output.ptr()) + 12, vreinterpretq_f32_u32(vshlq_n_u32(vmovl_u16(vget_high_u16(texels.val[1])), 16))); } for(; x < window_end_x; ++x) { *(output_ptr + x) = float(*(input_ptr + x)); } }, input, output); break; } default: ARM_COMPUTE_ERROR("Output data type unsupported"); } break; #endif /* defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) */ #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: switch(_output->info()->data_type()) { case DataType::QASYMM8_SIGNED: { const float16_t scale_s = 1 << _shift; const float16x8_t scale = vdupq_n_f16(scale_s); /* Up-conversion F16 -> QASYMM8_SIGNED */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const float16x8x2_t texels = { { vmulq_f16(vld1q_f16(input_ptr + x), scale), vmulq_f16(vld1q_f16(input_ptr + x + 8), scale), } }; vst1q_s8(output_ptr + x, vcombine_s8(vqmovn_s16(vcvtq_s16_f16(texels.val[0])), vqmovn_s16(vcvtq_s16_f16(texels.val[1])))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) * scale_s); } }, input, output); break; } case DataType::QASYMM8: case DataType::U8: { const float16_t scale_s = 1 << _shift; const float16x8_t scale = vdupq_n_f16(scale_s); /* Up-conversion F16 -> U8 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const float16x8x2_t texels = { { vmulq_f16(vld1q_f16(input_ptr + x), scale), vmulq_f16(vld1q_f16(input_ptr + x + 8), scale), } }; vst1q_u8(output_ptr + x, vcombine_u8(vqmovun_s16(vcvtq_s16_f16(texels.val[0])), vqmovun_s16(vcvtq_s16_f16(texels.val[1])))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) * scale_s); } }, input, output); break; } case DataType::F32: { const float scale_s = 1 << _shift; const float32x4_t scale = vdupq_n_f32(scale_s); /* Up-conversion F16 -> F32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const float16x8x2_t texels = { { vld1q_f16(input_ptr + x), vld1q_f16(input_ptr + x + 8) } }; vst1q_f32(output_ptr + x, vmulq_f32(vcvt_f32_f16(vget_low_f16(texels.val[0])), scale)); vst1q_f32(output_ptr + x + 4, vmulq_f32(vcvt_f32_f16(vget_high_f16(texels.val[0])), scale)); vst1q_f32(output_ptr + x + 8, vmulq_f32(vcvt_f32_f16(vget_low_f16(texels.val[1])), scale)); vst1q_f32(output_ptr + x + 12, vmulq_f32(vcvt_f32_f16(vget_high_f16(texels.val[1])), scale)); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) * scale_s); } }, input, output); break; } case DataType::S32: { const float scale_s = 1 << _shift; const float32x4_t scale = vdupq_n_f32(scale_s); /* Up-conversion F16 -> S32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const float16x8x2_t texels = { { vld1q_f16(input_ptr + x), vld1q_f16(input_ptr + x + 8) } }; vst1q_s32(output_ptr + x, vcvtq_s32_f32(vmulq_f32(vcvt_f32_f16(vget_low_f16(texels.val[0])), scale))); vst1q_s32(output_ptr + x + 4, vcvtq_s32_f32(vmulq_f32(vcvt_f32_f16(vget_high_f16(texels.val[0])), scale))); vst1q_s32(output_ptr + x + 8, vcvtq_s32_f32(vmulq_f32(vcvt_f32_f16(vget_low_f16(texels.val[1])), scale))); vst1q_s32(output_ptr + x + 12, vcvtq_s32_f32(vmulq_f32(vcvt_f32_f16(vget_high_f16(texels.val[1])), scale))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) * scale_s); } }, input, output); break; } default: ARM_COMPUTE_ERROR("Output data type not supported"); } break; #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ case DataType::F32: switch(_output->info()->data_type()) { #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: { const float scale_s = 1.f / (1 << _shift); const float32x4_t scale = vdupq_n_f32(scale_s); /* Down-conversion F32 -> F16 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const float32x4x4_t texels = { { vmulq_f32(vld1q_f32(input_ptr + x), scale), vmulq_f32(vld1q_f32(input_ptr + x + 4), scale), vmulq_f32(vld1q_f32(input_ptr + x + 8), scale), vmulq_f32(vld1q_f32(input_ptr + x + 12), scale) } }; vst1q_f16(output_ptr + x, vcombine_f16(vcvt_f16_f32(texels.val[0]), vcvt_f16_f32(texels.val[1]))); vst1q_f16(output_ptr + x + 8, vcombine_f16(vcvt_f16_f32(texels.val[2]), vcvt_f16_f32(texels.val[3]))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) * scale_s); } }, input, output); break; } #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ #if defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) case DataType::BFLOAT16: { /* Down-conversion F32 -> BFLOAT16 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { wrapper::vcvt_bf16_f32(reinterpret_cast(input.ptr()), reinterpret_cast(output.ptr())); wrapper::vcvt_bf16_f32(reinterpret_cast(input.ptr()) + 8, reinterpret_cast(output.ptr()) + 8); } for(; x < window_end_x; ++x) { *(output_ptr + x) = *(input_ptr + x); } }, input, output); break; } #endif /* defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) */ case DataType::S32: { const float scale_s = 1.f / (1 << _shift); const float32x4_t scale = vdupq_n_f32(scale_s); /* Conversion F32 -> S32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const float32x4x4_t texels = { { vmulq_f32(vld1q_f32(input_ptr + x), scale), vmulq_f32(vld1q_f32(input_ptr + x + 4), scale), vmulq_f32(vld1q_f32(input_ptr + x + 8), scale), vmulq_f32(vld1q_f32(input_ptr + x + 12), scale), } }; vst1q_s32(output_ptr + x, vcvtq_s32_f32(texels.val[0])); vst1q_s32(output_ptr + x + 4, vcvtq_s32_f32(texels.val[1])); vst1q_s32(output_ptr + x + 8, vcvtq_s32_f32(texels.val[2])); vst1q_s32(output_ptr + x + 12, vcvtq_s32_f32(texels.val[3])); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) * scale_s); } }, input, output); break; } case DataType::QASYMM8: case DataType::U8: { const float scale_s = 1.f / (1 << _shift); const float32x4_t scale = vdupq_n_f32(scale_s); /* Down-conversion F32 -> U8 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const float32x4x4_t texels = { { vmulq_f32(vld1q_f32(input_ptr + x), scale), vmulq_f32(vld1q_f32(input_ptr + x + 4), scale), vmulq_f32(vld1q_f32(input_ptr + x + 8), scale), vmulq_f32(vld1q_f32(input_ptr + x + 12), scale), } }; vst1_u8(output_ptr + x, vqmovn_u16(vcombine_u16(vqmovun_s32(vcvtq_s32_f32(texels.val[0])), vqmovun_s32(vcvtq_s32_f32(texels.val[1]))))); vst1_u8(output_ptr + x + 8, vqmovn_u16(vcombine_u16(vqmovun_s32(vcvtq_s32_f32(texels.val[2])), vqmovun_s32(vcvtq_s32_f32(texels.val[3]))))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = utils::cast::saturate_cast(*(input_ptr + x) * scale_s); } }, input, output); break; } case DataType::QASYMM8_SIGNED: { const float scale_s = 1.f / (1 << _shift); const float32x4_t scale = vdupq_n_f32(scale_s); /* Down-conversion F32 -> QASYMM8_SIGNED */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const float32x4x4_t texels = { { vmulq_f32(vld1q_f32(input_ptr + x), scale), vmulq_f32(vld1q_f32(input_ptr + x + 4), scale), vmulq_f32(vld1q_f32(input_ptr + x + 8), scale), vmulq_f32(vld1q_f32(input_ptr + x + 12), scale), } }; vst1_s8(output_ptr + x, vqmovn_s16(vcombine_s16(vqmovn_s32(vcvtq_s32_f32(texels.val[0])), vqmovn_s32(vcvtq_s32_f32(texels.val[1]))))); vst1_s8(output_ptr + x + 8, vqmovn_s16(vcombine_s16(vqmovn_s32(vcvtq_s32_f32(texels.val[2])), vqmovn_s32(vcvtq_s32_f32(texels.val[3]))))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = utils::cast::saturate_cast(*(input_ptr + x) * scale_s); } }, input, output); break; } default: ARM_COMPUTE_ERROR("Output data type not supported"); } break; case DataType::S32: switch(_output->info()->data_type()) { #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: { const float scale_s = 1.f / (1 << _shift); const float32x4_t scale = vdupq_n_f32(scale_s); /* Down-conversion S32 -> F16 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const float32x4x4_t texels = { { vmulq_f32(vcvtq_f32_s32(vld1q_s32(input_ptr + x)), scale), vmulq_f32(vcvtq_f32_s32(vld1q_s32(input_ptr + x + 4)), scale), vmulq_f32(vcvtq_f32_s32(vld1q_s32(input_ptr + x + 8)), scale), vmulq_f32(vcvtq_f32_s32(vld1q_s32(input_ptr + x + 12)), scale) } }; vst1q_f16(output_ptr + x, vcombine_f16(vcvt_f16_f32(texels.val[0]), vcvt_f16_f32(texels.val[1]))); vst1q_f16(output_ptr + x + 8, vcombine_f16(vcvt_f16_f32(texels.val[2]), vcvt_f16_f32(texels.val[3]))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) * scale_s); } }, input, output); break; } #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ case DataType::F32: { const int scale_s = 1.f / (1 << _shift); const int32x4_t scale = vdupq_n_s32(scale_s); /* Conversion S32 -> F32 */ execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int32x4x4_t texels = { { vmulq_s32(vld1q_s32(input_ptr + x), scale), vmulq_s32(vld1q_s32(input_ptr + x + 4), scale), vmulq_s32(vld1q_s32(input_ptr + x + 8), scale), vmulq_s32(vld1q_s32(input_ptr + x + 12), scale), } }; vst1q_f32(output_ptr + x, vcvtq_f32_s32(texels.val[0])); vst1q_f32(output_ptr + x + 4, vcvtq_f32_s32(texels.val[1])); vst1q_f32(output_ptr + x + 8, vcvtq_f32_s32(texels.val[2])); vst1q_f32(output_ptr + x + 12, vcvtq_f32_s32(texels.val[3])); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) * scale_s); } }, input, output); break; } case DataType::QASYMM8_SIGNED: { const int32x4_t b = vdupq_n_s32(-static_cast(_shift)); /* Down-conversion S32 -> QASYMM8_SIGNED */ if(ConvertPolicy::SATURATE == _policy) { execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int32x4x4_t texels = { { vqshlq_s32(vld1q_s32(input_ptr + x), b), vqshlq_s32(vld1q_s32(input_ptr + x + 4), b), vqshlq_s32(vld1q_s32(input_ptr + x + 8), b), vqshlq_s32(vld1q_s32(input_ptr + x + 12), b) } }; vst1_s8(output_ptr + x, vqmovn_s16(vcombine_s16(vqmovn_s32(texels.val[0]), vqmovn_s32(texels.val[1])))); vst1_s8(output_ptr + x + 8, vqmovn_s16(vcombine_s16(vqmovn_s32(texels.val[2]), vqmovn_s32(texels.val[3])))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = utils::cast::saturate_cast(*(input_ptr + x) >> _shift); } }, input, output); } else { execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int32x4x4_t texels = { { vshlq_s32(vld1q_s32(input_ptr + x), b), vshlq_s32(vld1q_s32(input_ptr + x + 4), b), vshlq_s32(vld1q_s32(input_ptr + x + 8), b), vshlq_s32(vld1q_s32(input_ptr + x + 12), b) } }; vst1_s8(output_ptr + x, vmovn_s16(vcombine_s16(vmovn_s32(texels.val[0]), vmovn_s32(texels.val[1])))); vst1_s8(output_ptr + x + 8, vmovn_s16(vcombine_s16(vmovn_s32(texels.val[2]), vmovn_s32(texels.val[3])))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) >> _shift); } }, input, output); } break; } case DataType::QASYMM8: case DataType::U8: { const int32x4_t b = vdupq_n_s32(-static_cast(_shift)); /* Down-conversion S32 -> U8 */ if(ConvertPolicy::SATURATE == _policy) { execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int32x4x4_t texels = { { vqshlq_s32(vld1q_s32(input_ptr + x), b), vqshlq_s32(vld1q_s32(input_ptr + x + 4), b), vqshlq_s32(vld1q_s32(input_ptr + x + 8), b), vqshlq_s32(vld1q_s32(input_ptr + x + 12), b) } }; vst1_u8(output_ptr + x, vqmovn_u16(vcombine_u16(vqmovun_s32(texels.val[0]), vqmovun_s32(texels.val[1])))); vst1_u8(output_ptr + x + 8, vqmovn_u16(vcombine_u16(vqmovun_s32(texels.val[2]), vqmovun_s32(texels.val[3])))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = utils::cast::saturate_cast(*(input_ptr + x) >> _shift); } }, input, output); } else { execute_window_loop(win, [&](const Coordinates &) { const auto input_ptr = reinterpret_cast(input.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { const int32x4x4_t texels = { { vshlq_s32(vld1q_s32(input_ptr + x), b), vshlq_s32(vld1q_s32(input_ptr + x + 4), b), vshlq_s32(vld1q_s32(input_ptr + x + 8), b), vshlq_s32(vld1q_s32(input_ptr + x + 12), b) } }; vst1_u8(output_ptr + x, vmovn_u16(vcombine_u16(vmovn_u32(vreinterpretq_u32_s32(texels.val[0])), vmovn_u32(vreinterpretq_u32_s32(texels.val[1]))))); vst1_u8(output_ptr + x + 8, vmovn_u16(vcombine_u16(vmovn_u32(vreinterpretq_u32_s32(texels.val[2])), vmovn_u32(vreinterpretq_u32_s32(texels.val[3]))))); } // Compute left-over elements for(; x < window_end_x; ++x) { *(output_ptr + x) = static_cast(*(input_ptr + x) >> _shift); } }, input, output); } break; } default: ARM_COMPUTE_ERROR("Output data type not supported"); } break; default: ARM_COMPUTE_ERROR("Not supported"); } }