| /* |
| * Copyright (c) 2016-2023 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 "src/cpu/kernels/CpuMulKernel.h" |
| |
| #include "arm_compute/core/ITensor.h" |
| #include "arm_compute/core/TensorInfo.h" |
| |
| #include "src/core/common/Registrars.h" |
| #include "src/core/CPP/Validate.h" |
| #include "src/core/helpers/AutoConfiguration.h" |
| #include "src/core/helpers/WindowHelpers.h" |
| #include "src/core/NEON/NEAsymm.h" |
| #include "src/core/NEON/NESymm.h" |
| #include "src/core/NEON/wrapper/wrapper.h" |
| #include "src/cpu/kernels/mul/generic/neon/list.h" |
| |
| #include <arm_neon.h> |
| |
| namespace |
| { |
| #if defined(ENABLE_FP32_KERNELS) |
| static constexpr size_t default_mws_N1_fp32_neon = 22447; |
| static constexpr size_t default_mws_V1_fp32_neon = 38982; |
| #endif /* ENABLE_FP32_KERNELS */ |
| static constexpr size_t default_mws_other_platforms_1d_tensor = 10240; |
| } // namespace |
| namespace arm_compute |
| { |
| namespace cpu |
| { |
| namespace kernels |
| { |
| namespace |
| { |
| const float scale255_constant = 1.f / 255.f; |
| const float32x4_t scale255_constant_f32q = vdupq_n_f32(scale255_constant); |
| const float32x4_t positive_round_f32q = vdupq_n_f32(0.5f); |
| |
| inline Status validate_arguments(const ITensorInfo *src1, |
| const ITensorInfo *src2, |
| const ITensorInfo *dst, |
| float scale, |
| ConvertPolicy overflow_policy, |
| RoundingPolicy rounding_policy) |
| { |
| ARM_COMPUTE_UNUSED(overflow_policy); |
| ARM_COMPUTE_UNUSED(rounding_policy); |
| |
| ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src1); |
| ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 1, DataType::U8, DataType::QASYMM8, |
| DataType::QASYMM8_SIGNED, DataType::S16, DataType::S32, |
| DataType::QSYMM16, DataType::F16, DataType::F32); |
| ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src2, 1, DataType::U8, DataType::QASYMM8, |
| DataType::QASYMM8_SIGNED, DataType::S16, DataType::S32, |
| DataType::QSYMM16, DataType::F16, DataType::F32); |
| ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::U8, DataType::QASYMM8, |
| DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, |
| DataType::S32, DataType::F16, DataType::F32); |
| if (is_data_type_quantized(src1->data_type()) || is_data_type_quantized(src2->data_type())) |
| { |
| ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src1, src2); |
| ARM_COMPUTE_RETURN_ERROR_ON_MSG(overflow_policy == ConvertPolicy::WRAP, |
| "ConvertPolicy cannot be WRAP if datatype is quantized"); |
| } |
| |
| if (dst->total_size() > 0) |
| { |
| const TensorShape &out_shape = TensorShape::broadcast_shape(src1->tensor_shape(), src2->tensor_shape()); |
| ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, dst->tensor_shape(), 0), |
| "Wrong shape for dst"); |
| ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); |
| // clang-format off |
| ARM_COMPUTE_RETURN_ERROR_ON_MSG( |
| !(src1->data_type() == src2->data_type() && src2->data_type() == dst->data_type()) && |
| !(src1->data_type() == DataType::U8 && src2->data_type() == DataType::U8 && dst->data_type() == DataType::S16) && |
| !(src1->data_type() == DataType::U8 && src2->data_type() == DataType::S16 && dst->data_type() == DataType::S16) && |
| !(src1->data_type() == DataType::S16 && src2->data_type() == DataType::U8 && dst->data_type() == DataType::S16) && |
| !(src1->data_type() == DataType::S16 && src2->data_type() == DataType::U8 && dst->data_type() == DataType::S16) && |
| !(src1->data_type() == DataType::QSYMM16 && src2->data_type() == DataType::QSYMM16 && dst->data_type() == DataType::S32) |
| , "Invalid data type combination"); |
| // clang-format on |
| ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->data_type() == DataType::S16 && dst->data_type() == DataType::S32 && |
| scale != 1.f, |
| "Unsupported scale for QSYMM16 inputs and S32 dst"); |
| } |
| |
| if (std::abs(scale - scale255_constant) < 0.00001f) |
| { |
| ARM_COMPUTE_RETURN_ERROR_ON(rounding_policy != RoundingPolicy::TO_NEAREST_UP && |
| rounding_policy != RoundingPolicy::TO_NEAREST_EVEN); |
| ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->data_type() == DataType::S32 && src2->data_type() == DataType::S32 && |
| dst->data_type() == DataType::S32, |
| "Scale == 1/255 is not supported if input and dst are of data type S32"); |
| } |
| else |
| { |
| ARM_COMPUTE_RETURN_ERROR_ON(rounding_policy != RoundingPolicy::TO_ZERO); |
| |
| int exponent = 0; |
| const float normalized_mantissa = std::frexp(scale, &exponent); |
| |
| // Use int scaling if factor is equal to 1/2^n for 0 <= n <= 15 |
| // frexp returns 0.5 as mantissa which means that the exponent will be in the range of -1 <= e <= 14 |
| // Moreover, it will be negative as we deal with 1/2^n |
| ARM_COMPUTE_RETURN_ERROR_ON_MSG(!((normalized_mantissa == 0.5f) && (-14 <= exponent) && (exponent <= 1)), |
| "Scale value not supported (Should be 1/(2^n) or 1/255"); |
| } |
| |
| return Status{}; |
| } |
| |
| /* Scales a given vector by 1/255. |
| * |
| * @note This does not work for all cases. e.g. for float of 0.49999999999999994 and large floats. |
| * |
| * @param in Input vector to scale. |
| * @return Scaled dst rounded to nearest (round half up). |
| */ |
| inline int32x4_t scale255_S32_S32(int32x4_t in) |
| { |
| // Scale |
| const float32x4_t tmp = vmulq_f32(vcvtq_f32_s32(in), scale255_constant_f32q); |
| // Round to nearest (round half up) |
| // Add +0.5 for all values |
| // Afterwards vcvt rounds toward zero |
| return vcvtq_s32_f32(vaddq_f32(tmp, positive_round_f32q)); |
| } |
| |
| inline uint16x8_t scale255_U16_U16(uint16x8_t in) |
| { |
| const int32x4_t tmp_s1 = scale255_S32_S32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(in)))); |
| const int32x4_t tmp_s2 = scale255_S32_S32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(in)))); |
| return vreinterpretq_u16_s16(vcombine_s16(vmovn_s32(tmp_s2), vmovn_s32(tmp_s1))); |
| } |
| |
| template <typename T> |
| inline typename std::enable_if<std::is_same<T, int8_t>::value, int8x16_t>::type |
| vquantize(float32x4x4_t val, const UniformQuantizationInfo &info) |
| { |
| return vquantize_signed(val, info); |
| } |
| |
| template <typename T> |
| inline typename std::enable_if<std::is_same<T, uint8_t>::value, uint8x16_t>::type |
| vquantize(float32x4x4_t val, const UniformQuantizationInfo &info) |
| { |
| return vquantize(val, info); |
| } |
| |
| template <typename T> |
| void mul_saturate_quantized_8(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, float scale) |
| { |
| // Create input windows |
| Window win = window; |
| Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); |
| |
| // Clear X Dimension on execution window as we handle manually |
| win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| const int window_step_x = 16 / sizeof(T); |
| const auto window_start_x = static_cast<int>(window.x().start()); |
| const auto window_end_x = static_cast<int>(window.x().end()); |
| const bool is_broadcast_across_x = src1->info()->tensor_shape().x() != src2->info()->tensor_shape().x(); |
| |
| const UniformQuantizationInfo output_qua_info = out->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo tmp_qua_info = {output_qua_info.scale / scale, output_qua_info.offset}; |
| |
| if (is_broadcast_across_x) |
| { |
| const bool is_broadcast_input_2 = input2_win.x().step() == 0; |
| Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; |
| Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; |
| const ITensor *broadcast_tensor = is_broadcast_input_2 ? src2 : src1; |
| const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src2 : src1; |
| const UniformQuantizationInfo broadcast_qinfo = broadcast_tensor->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform(); |
| |
| // Clear X Dimension on execution window as we handle manually |
| non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator broadcast_input(broadcast_tensor, broadcast_win); |
| Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); |
| Iterator dst(out, win); |
| |
| using ExactTagType = typename wrapper::traits::neon_vector<T, window_step_x>::tag_type; |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr()); |
| const auto output_ptr = reinterpret_cast<T *>(dst.ptr()); |
| |
| const auto broadcast_value = *reinterpret_cast<const T *>(broadcast_input.ptr()); |
| const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{}); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x); |
| |
| // Dequantize inputs |
| const float32x4x4_t in1_f32x4x4 = vdequantize(non_broadcast_v, non_broadcast_qinfo); |
| const float32x4x4_t in2_f32x4x4 = vdequantize(broadcast_value_vec, broadcast_qinfo); |
| |
| const float32x4x4_t out_f32x4x4 = { |
| vmulq_f32(in1_f32x4x4.val[0], in2_f32x4x4.val[0]), |
| vmulq_f32(in1_f32x4x4.val[1], in2_f32x4x4.val[1]), |
| vmulq_f32(in1_f32x4x4.val[2], in2_f32x4x4.val[2]), |
| vmulq_f32(in1_f32x4x4.val[3], in2_f32x4x4.val[3]), |
| }; |
| |
| // Quantize dst |
| const auto result = vquantize<T>(out_f32x4x4, tmp_qua_info); |
| wrapper::vstore(output_ptr + x, result); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| // Dequantize inputs |
| const T src1 = *(non_broadcast_input_ptr + x); |
| const float tmp_in1 = Qasymm8QuantizationHelper<T>::dequantize(src1, non_broadcast_qinfo); |
| const float tmp_in2 = Qasymm8QuantizationHelper<T>::dequantize(broadcast_value, broadcast_qinfo); |
| const float tmp_f = tmp_in1 * tmp_in2; |
| |
| // Quantize dst |
| const auto tmp_qua = Qasymm8QuantizationHelper<T>::quantize(tmp_f, tmp_qua_info); |
| *(output_ptr + x) = tmp_qua; |
| } |
| }, |
| broadcast_input, non_broadcast_input, dst); |
| } |
| else |
| { |
| const UniformQuantizationInfo input1_qua_info = src1->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo input2_qua_info = src2->info()->quantization_info().uniform(); |
| |
| // Clear X Dimension on execution window as we handle manually |
| input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator input1(src1, input1_win); |
| Iterator input2(src2, input2_win); |
| Iterator dst(out, win); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<T *>(dst.ptr()); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const auto input1_q = wrapper::vloadq(input1_ptr + x); |
| const auto input2_q = wrapper::vloadq(input2_ptr + x); |
| |
| // Dequantize inputs |
| const float32x4x4_t in1_f32x4x4 = vdequantize(input1_q, input1_qua_info); |
| const float32x4x4_t in2_f32x4x4 = vdequantize(input2_q, input2_qua_info); |
| |
| const float32x4x4_t out_f32x4x4 = { |
| vmulq_f32(in1_f32x4x4.val[0], in2_f32x4x4.val[0]), |
| vmulq_f32(in1_f32x4x4.val[1], in2_f32x4x4.val[1]), |
| vmulq_f32(in1_f32x4x4.val[2], in2_f32x4x4.val[2]), |
| vmulq_f32(in1_f32x4x4.val[3], in2_f32x4x4.val[3]), |
| }; |
| |
| // Quantize dst |
| const auto result = vquantize<T>(out_f32x4x4, tmp_qua_info); |
| wrapper::vstore(output_ptr + x, result); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| // Dequantize inputs |
| const T src1 = *(input1_ptr + x); |
| const T src2 = *(input2_ptr + x); |
| const float tmp_in1 = Qasymm8QuantizationHelper<T>::dequantize(src1, input1_qua_info); |
| const float tmp_in2 = Qasymm8QuantizationHelper<T>::dequantize(src2, input2_qua_info); |
| const float tmp_f = tmp_in1 * tmp_in2; |
| |
| // Quantize dst |
| const auto tmp_qua = Qasymm8QuantizationHelper<T>::quantize(tmp_f, tmp_qua_info); |
| *(output_ptr + x) = tmp_qua; |
| } |
| }, |
| input1, input2, dst); |
| } |
| } |
| |
| bool mul_q8_neon_fixedpoint_possible(const ITensorInfo *src0, |
| const ITensorInfo *src1, |
| const ITensorInfo *dst, |
| float scale) |
| { |
| const auto iq0 = src0->quantization_info().uniform(); |
| const auto iq1 = src1->quantization_info().uniform(); |
| const auto oq = dst->quantization_info().uniform(); |
| |
| const auto multiplier = ((iq0.scale * iq1.scale) / oq.scale) * scale; |
| |
| if (multiplier < -8191.f || multiplier > 8191.f) |
| { |
| //The multiplier cannot be stored as a 14.18 signed fixed-point number |
| return false; |
| } |
| |
| const auto offset_out = float(oq.offset); |
| |
| const auto max_result = multiplier * (256) * (256) + offset_out; |
| |
| if (max_result > 8191.f) |
| { |
| //It might not be possible to store the result as a 14.18 signed fixed-point number. |
| return false; |
| } |
| |
| return true; |
| } |
| |
| template <typename ScalarType> |
| void mul_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const Window &window, float scale) |
| { |
| const auto in0_info = src0->info(); |
| const auto in1_info = src1->info(); |
| |
| const auto &in0_shape = in0_info->tensor_shape(); |
| const auto &in1_shape = in1_info->tensor_shape(); |
| |
| // Create input windows. |
| Window in0_win = window.broadcast_if_dimension_le_one(in0_shape); |
| Window in1_win = window.broadcast_if_dimension_le_one(in1_shape); |
| |
| // Clear the x dimension on the execution window as we process the whole row each iteration. |
| Window win = window; |
| win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| constexpr int window_step_x = 16; |
| const auto window_start_x = window.x().start(); |
| const auto window_end_x = window.x().end(); |
| const auto is_broadcast_across_x = in0_shape.x() != in1_shape.x(); |
| |
| const auto iq0_info = in0_info->quantization_info().uniform(); |
| const auto iq1_info = in1_info->quantization_info().uniform(); |
| const auto oq_info = dst->info()->quantization_info().uniform(); |
| |
| const auto in0_offset = iq0_info.offset; |
| const auto in1_offset = iq1_info.offset; |
| const auto out_offset = oq_info.offset; |
| const auto multiplier = ((iq0_info.scale * iq1_info.scale) / oq_info.scale) * scale; |
| |
| constexpr int32_t two_pwr18i = 262144; |
| constexpr float two_pwr18f = 262144.f; |
| |
| const auto in0_offset_16p0 = static_cast<int16_t>(in0_offset); |
| const auto in1_offset_16p0 = static_cast<int16_t>(in1_offset); |
| const auto out_offset_14p18 = static_cast<int32_t>(out_offset * two_pwr18i); |
| const auto multiplier_14p18 = static_cast<int32_t>(multiplier * two_pwr18f); |
| |
| if (is_broadcast_across_x) |
| { |
| // Prefix: a = non-broadcast, b = broadcast. |
| |
| const auto is_broadcast_input_1 = in1_win.x().step() == 0; |
| auto a_win = is_broadcast_input_1 ? in0_win : in1_win; |
| auto b_win = is_broadcast_input_1 ? in1_win : in0_win; |
| const auto a_tensor = is_broadcast_input_1 ? src0 : src1; |
| const auto b_tensor = is_broadcast_input_1 ? src1 : src0; |
| |
| const auto a_offset_16p0 = is_broadcast_input_1 ? in0_offset_16p0 : in1_offset_16p0; |
| const auto b_offset_16p0 = is_broadcast_input_1 ? in1_offset : in0_offset; |
| #ifndef __aarch64__ |
| const auto a_offset = is_broadcast_input_1 ? in0_offset : in1_offset; |
| const auto b_offset = is_broadcast_input_1 ? in1_offset : in0_offset; |
| #endif //__aarch64__ |
| const auto a_voffset_16p0 = wrapper::vdup_n(a_offset_16p0, wrapper::traits::vector_64_tag()); |
| |
| // Clear the x dimension on the execution window as we process the whole row each iteration. |
| a_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator a_input_it(a_tensor, a_win); |
| Iterator b_input_it(b_tensor, b_win); |
| Iterator out_it(dst, win); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto a_ptr = reinterpret_cast<const ScalarType *>(a_input_it.ptr()); |
| const auto b_ptr = reinterpret_cast<const ScalarType *>(b_input_it.ptr()); |
| const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()); |
| |
| const auto b_val = *b_ptr; |
| const auto b_offseted_32p0 = static_cast<int32_t>(b_val - b_offset_16p0); |
| const auto b_voffseted_32p0 = wrapper::vdup_n(b_offseted_32p0, wrapper::traits::vector_128_tag()); |
| |
| const auto vmultiplier_14p18 = wrapper::vdup_n(multiplier_14p18, wrapper::traits::vector_128_tag()); |
| const auto voffsetout_14p18 = wrapper::vdup_n(out_offset_14p18, wrapper::traits::vector_128_tag()); |
| |
| int x = window_start_x; |
| |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| // Load the inputs. |
| const auto a_vin_8p0 = wrapper::vloadq(a_ptr + x); |
| |
| // Widen the non-broadcast elements to signed 16-bit regardless of the input signedness. |
| const auto a_vin_16p0_0 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(a_vin_8p0))); |
| const auto a_vin_16p0_1 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(a_vin_8p0))); |
| |
| const auto voffseted_32p0_00 = wrapper::vsubl(wrapper::vgetlow(a_vin_16p0_0), a_voffset_16p0); |
| const auto voffseted_32p0_01 = wrapper::vsubl(wrapper::vgethigh(a_vin_16p0_0), a_voffset_16p0); |
| const auto voffseted_32p0_10 = wrapper::vsubl(wrapper::vgetlow(a_vin_16p0_1), a_voffset_16p0); |
| const auto voffseted_32p0_11 = wrapper::vsubl(wrapper::vgethigh(a_vin_16p0_1), a_voffset_16p0); |
| |
| const auto vinnermul_32p0_00 = wrapper::vmul(voffseted_32p0_00, b_voffseted_32p0); |
| const auto vinnermul_32p0_01 = wrapper::vmul(voffseted_32p0_01, b_voffseted_32p0); |
| const auto vinnermul_32p0_10 = wrapper::vmul(voffseted_32p0_10, b_voffseted_32p0); |
| const auto vinnermul_32p0_11 = wrapper::vmul(voffseted_32p0_11, b_voffseted_32p0); |
| |
| const auto vout_14p18_00 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_00, vmultiplier_14p18); |
| const auto vout_14p18_01 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_01, vmultiplier_14p18); |
| const auto vout_14p18_10 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_10, vmultiplier_14p18); |
| const auto vout_14p18_11 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_11, vmultiplier_14p18); |
| |
| // These shift rights are to revert the multiplication by twopwr18. Hard limit of a maximum shift by 8 requires multiple shift instructions to achieve this. |
| const auto vout_15p1_00 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_00)); |
| const auto vout_15p1_01 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_01)); |
| const auto vout_15p1_10 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_10)); |
| const auto vout_15p1_11 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_11)); |
| |
| const auto vout_15p1_0 = wrapper::vcombine(vout_15p1_00, vout_15p1_01); |
| |
| const auto vout_15p1_1 = wrapper::vcombine(vout_15p1_10, vout_15p1_11); |
| const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()); |
| |
| const auto vout_8p0 = |
| wrapper::vcombine(wrapper::vqrshrn<2>(vout_15p1_0), wrapper::vqrshrn<2>(vout_15p1_1)); |
| wrapper::vstore(out_ptr + x, vout_8p0); |
| } |
| |
| //Process the left-over elements. |
| for (; x < window_end_x; ++x) |
| { |
| #ifdef __aarch64__ |
| out_ptr[x] = wrapper::vqrshrn<2>(wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>( |
| (multiplier_14p18 * (int32_t(a_ptr[x]) - a_offset_16p0) * (int32_t(b_val) - b_offset_16p0)) + |
| out_offset_14p18))); |
| #else //__aarch64__ |
| out_ptr[x] = utility::clamp<int32_t, ScalarType>(support::cpp11::lround( |
| multiplier * ((float(a_ptr[x]) - a_offset) * (float(b_val) - b_offset)) + float(out_offset))); |
| #endif //__aarch64__ |
| } |
| }, |
| a_input_it, b_input_it, out_it); |
| } |
| else |
| { |
| const auto voffset0_16p0 = wrapper::vdup_n(in0_offset_16p0, wrapper::traits::vector_64_tag()); |
| const auto voffset1_16p0 = wrapper::vdup_n(in1_offset_16p0, wrapper::traits::vector_64_tag()); |
| const auto voffsetout_14p18 = wrapper::vdup_n(out_offset_14p18, wrapper::traits::vector_128_tag()); |
| const auto vmultiplier_14p18 = wrapper::vdup_n(multiplier_14p18, wrapper::traits::vector_128_tag()); |
| |
| // Clear the x dimension on the execution window as we process the whole row each iteration. |
| in0_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| in1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator in0_it(src0, in0_win); |
| Iterator in1_it(src1, in1_win); |
| Iterator out_it(dst, win); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto in0_ptr = reinterpret_cast<const ScalarType *>(in0_it.ptr()); |
| const auto in1_ptr = reinterpret_cast<const ScalarType *>(in1_it.ptr()); |
| const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()); |
| |
| int x = window_start_x; |
| |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| // Load the inputs. |
| const auto vin0_8p0 = wrapper::vloadq(in0_ptr + x); |
| const auto vin1_8p0 = wrapper::vloadq(in1_ptr + x); |
| |
| // Widen the input elements to signed 16-bit regardless of the input signedness. |
| const auto vin0_16p0_0 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(vin0_8p0))); |
| const auto vin0_16p0_1 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(vin0_8p0))); |
| const auto vin1_16p0_0 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(vin1_8p0))); |
| const auto vin1_16p0_1 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(vin1_8p0))); |
| |
| const auto voffseted0_32p0_00 = wrapper::vsubl(wrapper::vgetlow(vin0_16p0_0), voffset0_16p0); |
| const auto voffseted0_32p0_01 = wrapper::vsubl(wrapper::vgethigh(vin0_16p0_0), voffset0_16p0); |
| const auto voffseted0_32p0_10 = wrapper::vsubl(wrapper::vgetlow(vin0_16p0_1), voffset0_16p0); |
| const auto voffseted0_32p0_11 = wrapper::vsubl(wrapper::vgethigh(vin0_16p0_1), voffset0_16p0); |
| |
| const auto voffseted1_32p0_00 = wrapper::vsubl(wrapper::vgetlow(vin1_16p0_0), voffset1_16p0); |
| const auto voffseted1_32p0_01 = wrapper::vsubl(wrapper::vgethigh(vin1_16p0_0), voffset1_16p0); |
| const auto voffseted1_32p0_10 = wrapper::vsubl(wrapper::vgetlow(vin1_16p0_1), voffset1_16p0); |
| const auto voffseted1_32p0_11 = wrapper::vsubl(wrapper::vgethigh(vin1_16p0_1), voffset1_16p0); |
| |
| const auto vinnermul_32p0_00 = wrapper::vmul(voffseted0_32p0_00, voffseted1_32p0_00); |
| const auto vinnermul_32p0_01 = wrapper::vmul(voffseted0_32p0_01, voffseted1_32p0_01); |
| const auto vinnermul_32p0_10 = wrapper::vmul(voffseted0_32p0_10, voffseted1_32p0_10); |
| const auto vinnermul_32p0_11 = wrapper::vmul(voffseted0_32p0_11, voffseted1_32p0_11); |
| |
| const auto vout_14p18_00 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_00, vmultiplier_14p18); |
| const auto vout_14p18_01 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_01, vmultiplier_14p18); |
| const auto vout_14p18_10 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_10, vmultiplier_14p18); |
| const auto vout_14p18_11 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_11, vmultiplier_14p18); |
| |
| // These shift rights are to revert the multiplication by twopwr18. Hard limit of a maximum shift by 8 requires multiple shift instructions to achieve this. |
| const auto vout_14p2_00 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_00)); |
| const auto vout_14p2_01 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_01)); |
| const auto vout_14p2_10 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_10)); |
| const auto vout_14p2_11 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_11)); |
| |
| const auto vout_14p2_0 = wrapper::vcombine(vout_14p2_00, vout_14p2_01); |
| |
| const auto vout_14p2_1 = wrapper::vcombine(vout_14p2_10, vout_14p2_11); |
| |
| const auto vout_8p0 = |
| wrapper::vcombine(wrapper::vqrshrn<2>(vout_14p2_0), wrapper::vqrshrn<2>(vout_14p2_1)); |
| wrapper::vstore(out_ptr + x, vout_8p0); |
| } |
| |
| //Process the left-over elements. |
| for (; x < window_end_x; ++x) |
| { |
| #ifdef __aarch64__ |
| out_ptr[x] = wrapper::vqrshrn<2>(wrapper::vqrshrn_ex<8, ScalarType>( |
| wrapper::vshrq_n<8>((multiplier_14p18 * (int32_t(in0_ptr[x]) - in0_offset_16p0) * |
| (int32_t(in1_ptr[x]) - in1_offset_16p0)) + |
| out_offset_14p18))); |
| #else //__aarch64__ |
| out_ptr[x] = utility::clamp<int32_t, ScalarType>(support::cpp11::lround( |
| multiplier * ((float(in0_ptr[x]) - in0_offset) * (float(in1_ptr[x]) - in1_offset)) + |
| float(out_offset))); |
| #endif //__aarch64__ |
| } |
| }, |
| in0_it, in1_it, out_it); |
| } |
| } |
| |
| void mul_saturate_QSYMM16_QSYMM16_QSYMM16( |
| const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, float scale) |
| { |
| const UniformQuantizationInfo input1_qua_info = src1->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo input2_qua_info = src2->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo output_qua_info = out->info()->quantization_info().uniform(); |
| |
| // Create input windows |
| Window win = window; |
| Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); |
| |
| // Clear X Dimension on execution window as we handle manually |
| win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator input1(src1, input1_win); |
| Iterator input2(src2, input2_win); |
| Iterator dst(out, win); |
| |
| const int window_step_x = 16; |
| const auto window_start_x = static_cast<int>(window.x().start()); |
| const auto window_end_x = static_cast<int>(window.x().end()); |
| |
| const UniformQuantizationInfo tmp_qua_info = {output_qua_info.scale / scale, output_qua_info.offset}; |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const qsymm16_t *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const qsymm16_t *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<qsymm16_t *>(dst.ptr()); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const qsymm16x8x2_t input1_q = {{ |
| vld1q_s16(input1_ptr + x), |
| vld1q_s16(input1_ptr + x + 8), |
| }}; |
| const qsymm16x8x2_t input2_q = {{ |
| vld1q_s16(input2_ptr + x), |
| vld1q_s16(input2_ptr + x + 8), |
| }}; |
| |
| // Dequantize inputs |
| const float32x4x4_t in1_f32x4x4 = vdequantize(input1_q, input1_qua_info); |
| const float32x4x4_t in2_f32x4x4 = vdequantize(input2_q, input2_qua_info); |
| |
| const float32x4x4_t out_f32x4x4 = { |
| vmulq_f32(in1_f32x4x4.val[0], in2_f32x4x4.val[0]), |
| vmulq_f32(in1_f32x4x4.val[1], in2_f32x4x4.val[1]), |
| vmulq_f32(in1_f32x4x4.val[2], in2_f32x4x4.val[2]), |
| vmulq_f32(in1_f32x4x4.val[3], in2_f32x4x4.val[3]), |
| }; |
| |
| const qsymm16x8x2_t result = vquantize_qsymm16(out_f32x4x4, tmp_qua_info); |
| vst1q_s16(output_ptr + x, result.val[0]); |
| vst1q_s16(output_ptr + x + 8, result.val[1]); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| // Dequantize inputs |
| float tmp_in1 = static_cast<float>(*(input1_ptr + x)) * input1_qua_info.scale; |
| float tmp_in2 = static_cast<float>(*(input2_ptr + x)) * input2_qua_info.scale; |
| float tmp_f = tmp_in1 * tmp_in2; |
| |
| // Quantize dst, lrintf() has same rounding mode as vcombine_s16 |
| int32_t tmp = lrintf(tmp_f / tmp_qua_info.scale); |
| qsymm16_t tmp_qua = |
| static_cast<qsymm16_t>(tmp > SHRT_MAX) ? SHRT_MAX : ((tmp < SHRT_MIN) ? SHRT_MIN : tmp); |
| *(output_ptr + x) = tmp_qua; |
| } |
| }, |
| input1, input2, dst); |
| } |
| |
| void mul_QSYMM16_QSYMM16_S32(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, int scale) |
| { |
| ARM_COMPUTE_UNUSED(scale); |
| |
| // Create input windows |
| Window win = window; |
| Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); |
| |
| // Clear X Dimension on execution window as we handle manually |
| win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator input1(src1, input1_win); |
| Iterator input2(src2, input2_win); |
| Iterator dst(out, win); |
| |
| const int window_step_x = 16; |
| const auto window_start_x = static_cast<int>(window.x().start()); |
| const auto window_end_x = static_cast<int>(window.x().end()); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const qsymm16_t *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const qsymm16_t *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<int32_t *>(dst.ptr()); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const qsymm16x8x2_t input1_q = {{ |
| vld1q_s16(input1_ptr + x), |
| vld1q_s16(input1_ptr + x + 8), |
| }}; |
| const qsymm16x8x2_t input2_q = {{ |
| vld1q_s16(input2_ptr + x), |
| vld1q_s16(input2_ptr + x + 8), |
| }}; |
| |
| const int32x4x4_t in1_s32 = {{ |
| vmovl_s16(vget_low_s16(input1_q.val[0])), |
| vmovl_s16(vget_high_s16(input1_q.val[0])), |
| vmovl_s16(vget_low_s16(input1_q.val[1])), |
| vmovl_s16(vget_high_s16(input1_q.val[1])), |
| }}; |
| const int32x4x4_t in2_s32 = {{ |
| vmovl_s16(vget_low_s16(input2_q.val[0])), |
| vmovl_s16(vget_high_s16(input2_q.val[0])), |
| vmovl_s16(vget_low_s16(input2_q.val[1])), |
| vmovl_s16(vget_high_s16(input2_q.val[1])), |
| }}; |
| |
| const int32x4x4_t result = {{ |
| vmulq_s32(in1_s32.val[0], in2_s32.val[0]), |
| vmulq_s32(in1_s32.val[1], in2_s32.val[1]), |
| vmulq_s32(in1_s32.val[2], in2_s32.val[2]), |
| vmulq_s32(in1_s32.val[3], in2_s32.val[3]), |
| }}; |
| |
| vst1q_s32(output_ptr + x, result.val[0]); |
| vst1q_s32(output_ptr + x + 4, result.val[1]); |
| vst1q_s32(output_ptr + x + 8, result.val[2]); |
| vst1q_s32(output_ptr + x + 12, result.val[3]); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| int32_t tmp = static_cast<int32_t>(*(input1_ptr + x)) * static_cast<int32_t>(*(input2_ptr + x)); |
| *(output_ptr + x) = tmp; |
| } |
| }, |
| input1, input2, dst); |
| } |
| |
| template <bool is_scale255, bool is_sat> |
| void mul_U8_U8_U8(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, int n) |
| { |
| // Create input windows |
| Window win = window; |
| Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); |
| |
| // Clear X Dimension on execution window as we handle manually |
| win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator input1(src1, input1_win); |
| Iterator input2(src2, input2_win); |
| Iterator dst(out, win); |
| |
| const int window_step_x = 16 / sizeof(uint8_t); |
| const auto window_start_x = static_cast<int>(window.x().start()); |
| const auto window_end_x = static_cast<int>(window.x().end()); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<uint8_t *>(dst.ptr()); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const uint8x16_t ta1 = wrapper::vloadq(input1_ptr + x); |
| const uint8x16_t ta2 = wrapper::vloadq(input2_ptr + x); |
| |
| uint16x8_t tmp1_high = vmovl_u8(vget_high_u8(ta1)); |
| const uint16x8_t tmp2_high = vmovl_u8(vget_high_u8(ta2)); |
| uint16x8_t tmp1_low = vmovl_u8(vget_low_u8(ta1)); |
| const uint16x8_t tmp2_low = vmovl_u8(vget_low_u8(ta2)); |
| |
| tmp1_high = vmulq_u16(tmp1_high, tmp2_high); |
| tmp1_low = vmulq_u16(tmp1_low, tmp2_low); |
| |
| if (is_scale255) |
| { |
| tmp1_high = scale255_U16_U16(tmp1_high); |
| tmp1_low = scale255_U16_U16(tmp1_low); |
| } |
| else |
| { |
| const int16x8_t vn = vdupq_n_s16(-n); |
| |
| if (is_sat) |
| { |
| tmp1_high = vqshlq_u16(tmp1_high, vn); |
| tmp1_low = vqshlq_u16(tmp1_low, vn); |
| } |
| else |
| { |
| tmp1_high = vshlq_u16(tmp1_high, vn); |
| tmp1_low = vshlq_u16(tmp1_low, vn); |
| } |
| } |
| if (is_sat) |
| { |
| vst1q_u8(output_ptr + x, vcombine_u8(vqmovn_u16(tmp1_low), vqmovn_u16(tmp1_high))); |
| } |
| else |
| { |
| vst1q_u8(output_ptr + x, vcombine_u8(vmovn_u16(tmp1_low), vmovn_u16(tmp1_high))); |
| } |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| uint16_t tmp = static_cast<uint16_t>(*(input1_ptr + x)) * static_cast<uint16_t>(*(input2_ptr + x)); |
| |
| if (is_scale255) |
| { |
| float tmp_f = static_cast<float>(tmp) * scale255_constant; |
| tmp = static_cast<uint16_t>(tmp_f + 0.5f); |
| } |
| else |
| { |
| tmp >>= n; |
| } |
| if (is_sat && tmp > 255) |
| { |
| tmp = 255; |
| } |
| *(output_ptr + x) = static_cast<uint8_t>(tmp); |
| } |
| }, |
| input1, input2, dst); |
| } |
| |
| template <bool is_scale255, bool is_sat> |
| inline int16x8_t mul_S16_S16_S16_n_loop(const int16x8_t &src1, const int16x8_t &src2, int n) |
| { |
| int32x4_t tmp1_high = vmovl_s16(vget_high_s16(src1)); |
| const int32x4_t tmp2_high = vmovl_s16(vget_high_s16(src2)); |
| int32x4_t tmp1_low = vmovl_s16(vget_low_s16(src1)); |
| const int32x4_t tmp2_low = vmovl_s16(vget_low_s16(src2)); |
| |
| tmp1_high = vmulq_s32(tmp1_high, tmp2_high); |
| tmp1_low = vmulq_s32(tmp1_low, tmp2_low); |
| |
| if (is_scale255) |
| { |
| tmp1_high = scale255_S32_S32(tmp1_high); |
| tmp1_low = scale255_S32_S32(tmp1_low); |
| } |
| else |
| { |
| // Right shift amount |
| const int32x4_t vn = vdupq_n_s32(-n); |
| // Left shift amount |
| const int32x4_t vnl = vdupq_n_s32(n); |
| // Calculate conversion bit |
| const uint32x4_t tmp1_high_u = vreinterpretq_u32_s32(tmp1_high); |
| const uint32x4_t tmp1_low_u = vreinterpretq_u32_s32(tmp1_low); |
| const uint32x4_t sign_high = vshrq_n_u32(tmp1_high_u, 31); |
| const uint32x4_t sign_low = vshrq_n_u32(tmp1_low_u, 31); |
| const int32x4_t sign_high_s = vreinterpretq_s32_u32(sign_high); |
| const int32x4_t sign_low_s = vreinterpretq_s32_u32(sign_low); |
| const int32x4_t convert_high = vsubq_s32(vshlq_s32(sign_high_s, vnl), sign_high_s); |
| const int32x4_t convert_low = vsubq_s32(vshlq_s32(sign_low_s, vnl), sign_low_s); |
| if (is_sat) |
| { |
| tmp1_high = vqshlq_s32(vaddq_s32(tmp1_high, convert_high), vn); |
| tmp1_low = vqshlq_s32(vaddq_s32(tmp1_low, convert_low), vn); |
| } |
| else |
| { |
| tmp1_high = vshlq_s32(vaddq_s32(tmp1_high, convert_high), vn); |
| tmp1_low = vshlq_s32(vaddq_s32(tmp1_low, convert_low), vn); |
| } |
| } |
| |
| if (is_sat) |
| { |
| return vcombine_s16(vqmovn_s32(tmp1_low), vqmovn_s32(tmp1_high)); |
| } |
| else |
| { |
| return vcombine_s16(vmovn_s32(tmp1_low), vmovn_s32(tmp1_high)); |
| } |
| } |
| |
| template <bool is_scale255, bool is_sat> |
| inline int16x8x2_t mul_S16_S16_S16_n_k(const int16x8x2_t &src1, const int16x8x2_t &src2, int n) |
| { |
| const int16x8x2_t result = {{// First 8 elements |
| mul_S16_S16_S16_n_loop<is_scale255, is_sat>(src1.val[0], src2.val[0], n), |
| // Second 8 elements |
| mul_S16_S16_S16_n_loop<is_scale255, is_sat>(src1.val[1], src2.val[1], n)}}; |
| |
| return result; |
| } |
| |
| template <bool is_scale255, bool is_sat> |
| void mul_S16_S16_S16(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, int n) |
| { |
| // Create input windows |
| Window win = window; |
| Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); |
| |
| // Clear X Dimension on execution window as we handle manually |
| win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator input1(src1, input1_win); |
| Iterator input2(src2, input2_win); |
| Iterator dst(out, win); |
| |
| const int window_step_x = 16; |
| const auto window_start_x = static_cast<int>(window.x().start()); |
| const auto window_end_x = static_cast<int>(window.x().end()); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const int16_t *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<int16_t *>(dst.ptr()); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const int16x8x2_t ta1 = {{ |
| vld1q_s16(input1_ptr + x), |
| vld1q_s16(input1_ptr + x + 8), |
| }}; |
| const int16x8x2_t ta2 = {{ |
| vld1q_s16(input2_ptr + x), |
| vld1q_s16(input2_ptr + x + 8), |
| }}; |
| const int16x8x2_t result = mul_S16_S16_S16_n_k<is_scale255, is_sat>(ta1, ta2, n); |
| |
| vst1q_s16(output_ptr + x, result.val[0]); |
| vst1q_s16(output_ptr + x + 8, result.val[1]); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| int32_t tmp = static_cast<int32_t>(*(input1_ptr + x)) * static_cast<int32_t>(*(input2_ptr + x)); |
| |
| if (is_scale255) |
| { |
| float tmp_f = static_cast<float>(tmp) * scale255_constant; |
| |
| tmp = static_cast<int32_t>(tmp_f + 0.5f); |
| } |
| else |
| { |
| if (tmp >= 0) |
| { |
| tmp >>= n; |
| } |
| else |
| { |
| uint32_t mask = (1u << n) - 1; |
| tmp = (tmp + static_cast<int32_t>(mask)) >> n; |
| } |
| } |
| if (is_sat) |
| { |
| tmp = (tmp > SHRT_MAX) ? SHRT_MAX : ((tmp < SHRT_MIN) ? SHRT_MIN : tmp); |
| } |
| *(output_ptr + x) = static_cast<int16_t>(tmp); |
| } |
| }, |
| input1, input2, dst); |
| } |
| |
| template <bool is_sat> |
| inline int32x4_t mul_S32_S32_S32_n_loop(const int32x4_t &src1, const int32x4_t &src2, int n) |
| { |
| const int32x2_t input1_1 = vget_low_s32(src1); |
| const int32x2_t input2_1 = vget_low_s32(src2); |
| const int32x2_t input1_2 = vget_high_s32(src1); |
| const int32x2_t input2_2 = vget_high_s32(src2); |
| |
| int64x2_t tmp_1 = vmull_s32(input1_1, input2_1); |
| int64x2_t tmp_2 = vmull_s32(input1_2, input2_2); |
| |
| // Apply scaling, conversion and rounding (round to zero) |
| // Right shift amount |
| const int64x2_t vn = vdupq_n_s64(-n); |
| // Left shift amount |
| const int64x2_t vnl = vdupq_n_s64(n); |
| // Calculate conversion bit |
| const uint64x2_t tmp_1_u = vreinterpretq_u64_s64(tmp_1); |
| const uint64x2_t sign_1 = vshrq_n_u64(tmp_1_u, 63); |
| const int64x2_t sign_1_s = vreinterpretq_s64_u64(sign_1); |
| const int64x2_t convert_1 = vsubq_s64(vshlq_s64(sign_1_s, vnl), sign_1_s); |
| |
| const uint64x2_t tmp_2_u = vreinterpretq_u64_s64(tmp_2); |
| const uint64x2_t sign_2 = vshrq_n_u64(tmp_2_u, 63); |
| const int64x2_t sign_2_s = vreinterpretq_s64_u64(sign_2); |
| const int64x2_t convert_2 = vsubq_s64(vshlq_s64(sign_2_s, vnl), sign_2_s); |
| if (is_sat) |
| { |
| tmp_1 = vqshlq_s64(vaddq_s64(tmp_1, convert_1), vn); |
| tmp_2 = vqshlq_s64(vaddq_s64(tmp_2, convert_2), vn); |
| return vcombine_s32(vqmovn_s64(tmp_1), vqmovn_s64(tmp_2)); |
| } |
| else |
| { |
| tmp_1 = vshlq_s64(vaddq_s64(tmp_1, convert_1), vn); |
| tmp_2 = vshlq_s64(vaddq_s64(tmp_2, convert_2), vn); |
| return vcombine_s32(vmovn_s64(tmp_1), vmovn_s64(tmp_2)); |
| } |
| } |
| |
| template <bool is_sat> |
| inline int32x4x2_t mul_S32_S32_S32_n_k(const int32x4x2_t &src1, const int32x4x2_t &src2, int n) |
| { |
| const int32x4x2_t result = {{// First 4 elements |
| mul_S32_S32_S32_n_loop<is_sat>(src1.val[0], src2.val[0], n), |
| // Second 4 elements |
| mul_S32_S32_S32_n_loop<is_sat>(src1.val[1], src2.val[1], n)}}; |
| |
| return result; |
| } |
| |
| template <bool is_sat> |
| void mul_S32_S32_S32(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, int n) |
| { |
| // Create input windows |
| Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); |
| |
| // Clear X Dimension on execution window as we handle manually |
| Window win = window; |
| win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| const int window_step_x = 8; |
| const auto window_start_x = static_cast<int>(window.x().start()); |
| const auto window_end_x = static_cast<int>(window.x().end()); |
| const bool is_broadcast_across_x = src1->info()->tensor_shape().x() != src2->info()->tensor_shape().x(); |
| |
| if (is_broadcast_across_x) |
| { |
| const bool is_broadcast_input_2 = input2_win.x().step() == 0; |
| Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; |
| Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; |
| const ITensor *broadcast_tensor = is_broadcast_input_2 ? src2 : src1; |
| const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src2 : src1; |
| |
| // Clear X Dimension on execution window as we handle manually |
| non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator broadcast_input(broadcast_tensor, broadcast_win); |
| Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); |
| Iterator dst(out, win); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto non_broadcast_input_ptr = reinterpret_cast<const int32_t *>(non_broadcast_input.ptr()); |
| const auto output_ptr = reinterpret_cast<int32_t *>(dst.ptr()); |
| |
| const int32_t broadcast_value = *reinterpret_cast<const int32_t *>(broadcast_input.ptr()); |
| const auto broadcast_value_vec = vdupq_n_s32(broadcast_value); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const int32x4x2_t broadcast_v = {{ |
| broadcast_value_vec, |
| broadcast_value_vec, |
| }}; |
| const int32x4x2_t non_broadcast_v = {{ |
| vld1q_s32(non_broadcast_input_ptr + x), |
| vld1q_s32(non_broadcast_input_ptr + x + 4), |
| }}; |
| const int32x4x2_t result = mul_S32_S32_S32_n_k<is_sat>(broadcast_v, non_broadcast_v, n); |
| |
| vst1q_s32(output_ptr + x, result.val[0]); |
| vst1q_s32(output_ptr + x + 4, result.val[1]); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| int64_t tmp = |
| static_cast<int64_t>(broadcast_value) * static_cast<int64_t>(*(non_broadcast_input_ptr + x)); |
| |
| if (tmp >= 0) |
| { |
| tmp >>= n; |
| } |
| else |
| { |
| uint64_t mask = ((uint64_t)1u << n) - 1; |
| tmp = (tmp + static_cast<int64_t>(mask)) >> n; |
| } |
| if (is_sat) |
| { |
| tmp = utility::clamp<int64_t, int32_t>(tmp); |
| } |
| *(output_ptr + x) = static_cast<int32_t>(tmp); |
| } |
| }, |
| broadcast_input, non_broadcast_input, dst); |
| } |
| else |
| { |
| // Clear X Dimension on execution window as we handle manually |
| input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator input1(src1, input1_win); |
| Iterator input2(src2, input2_win); |
| Iterator dst(out, win); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const int32_t *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const int32_t *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<int32_t *>(dst.ptr()); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const int32x4x2_t ta1 = {{ |
| vld1q_s32(input1_ptr + x), |
| vld1q_s32(input1_ptr + x + 4), |
| }}; |
| const int32x4x2_t ta2 = {{ |
| vld1q_s32(input2_ptr + x), |
| vld1q_s32(input2_ptr + x + 4), |
| }}; |
| const int32x4x2_t result = mul_S32_S32_S32_n_k<is_sat>(ta1, ta2, n); |
| |
| vst1q_s32(output_ptr + x, result.val[0]); |
| vst1q_s32(output_ptr + x + 4, result.val[1]); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| int64_t tmp = static_cast<int64_t>(*(input1_ptr + x)) * static_cast<int64_t>(*(input2_ptr + x)); |
| |
| if (tmp >= 0) |
| { |
| tmp >>= n; |
| } |
| else |
| { |
| uint64_t mask = ((uint64_t)1u << n) - 1; |
| tmp = (tmp + static_cast<int64_t>(mask)) >> n; |
| } |
| if (is_sat) |
| { |
| tmp = utility::clamp<int64_t, int32_t>(tmp); |
| } |
| *(output_ptr + x) = static_cast<int32_t>(tmp); |
| } |
| }, |
| input1, input2, dst); |
| } |
| } |
| |
| void c_mul_F32_F32_F32_n(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window) |
| { |
| // Create input windows |
| Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); |
| |
| // Clear X Dimension on execution window as we handle manually |
| Window win = window; |
| win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| constexpr int window_step_x = 8 / sizeof(float); |
| const auto window_start_x = static_cast<int>(window.x().start()); |
| const auto window_end_x = static_cast<int>(window.x().end()); |
| const bool is_broadcast_across_x = src1->info()->tensor_shape().x() != src2->info()->tensor_shape().x(); |
| |
| using ExactTagType = typename wrapper::traits::neon_vector<float, 2>::tag_type; |
| |
| if (is_broadcast_across_x) |
| { |
| const bool is_broadcast_input_2 = input2_win.x().step() == 0; |
| Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; |
| Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; |
| const ITensor *broadcast_tensor = is_broadcast_input_2 ? src2 : src1; |
| const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src2 : src1; |
| |
| // Clear X Dimension on execution window as we handle manually |
| non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator broadcast_input(broadcast_tensor, broadcast_win); |
| Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); |
| Iterator dst(out, win); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto non_broadcast_input_ptr = reinterpret_cast<const float *>(non_broadcast_input.ptr()); |
| const auto output_ptr = reinterpret_cast<float *>(dst.ptr()); |
| |
| const float broadcast_value = *reinterpret_cast<const float *>(broadcast_input.ptr()); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const auto a = wrapper::vloadq(non_broadcast_input_ptr + 2 * x); |
| float32x4_t b = vdupq_n_f32(broadcast_value); |
| |
| const float32x4_t mask = {-1.0f, 1.0f, -1.0f, 1.0f}; |
| const float32x2_t tmp00 = wrapper::vdup_n(wrapper::vgetlane(a, 0), ExactTagType{}); |
| const float32x2_t tmp01 = wrapper::vdup_n(wrapper::vgetlane(a, 1), ExactTagType{}); |
| const float32x2_t tmp10 = wrapper::vdup_n(wrapper::vgetlane(a, 2), ExactTagType{}); |
| const float32x2_t tmp11 = wrapper::vdup_n(wrapper::vgetlane(a, 3), ExactTagType{}); |
| |
| const float32x4_t tmp0 = wrapper::vcombine(tmp00, tmp10); |
| const float32x4_t tmp1 = wrapper::vcombine(tmp01, tmp11); |
| |
| float32x4_t res = wrapper::vmul(tmp0, b); |
| b = wrapper::vmul(b, mask); |
| |
| res = wrapper::vmla(res, tmp1, b); |
| wrapper::vstore(output_ptr + 2 * x, res); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| const auto non_broadcast_value0 = *(non_broadcast_input_ptr + 2 * x); |
| const auto non_broadcast_value1 = *(non_broadcast_input_ptr + 2 * x + 1); |
| auto res1 = broadcast_value * (non_broadcast_value0 - non_broadcast_value1); |
| auto res2 = broadcast_value * (non_broadcast_value1 + non_broadcast_value0); |
| *(output_ptr + 2 * x) = res1; |
| *(output_ptr + 2 * x + 1) = res2; |
| } |
| }, |
| broadcast_input, non_broadcast_input, dst); |
| } |
| else |
| { |
| // Clear X Dimension on execution window as we handle manually |
| input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator input1(src1, input1_win); |
| Iterator input2(src2, input2_win); |
| Iterator dst(out, win); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const float *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const float *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<float *>(dst.ptr()); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const float32x4_t a = wrapper::vloadq(input1_ptr + 2 * x); |
| float32x4_t b = wrapper::vloadq(input2_ptr + 2 * x); |
| |
| const float32x4_t mask = {-1.0f, 1.0f, -1.0f, 1.0f}; |
| const float32x2_t tmp00 = wrapper::vdup_n(wrapper::vgetlane(a, 0), ExactTagType{}); |
| const float32x2_t tmp01 = wrapper::vdup_n(wrapper::vgetlane(a, 1), ExactTagType{}); |
| const float32x2_t tmp10 = wrapper::vdup_n(wrapper::vgetlane(a, 2), ExactTagType{}); |
| const float32x2_t tmp11 = wrapper::vdup_n(wrapper::vgetlane(a, 3), ExactTagType{}); |
| |
| const float32x4_t tmp0 = wrapper::vcombine(tmp00, tmp10); |
| const float32x4_t tmp1 = wrapper::vcombine(tmp01, tmp11); |
| |
| float32x4_t res = wrapper::vmul(tmp0, b); |
| |
| b = wrapper::vrev64(b); |
| b = wrapper::vmul(b, mask); |
| |
| res = wrapper::vmla(res, tmp1, b); |
| wrapper::vstore(output_ptr + 2 * x, res); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| const auto a0 = *(input1_ptr + 2 * x); |
| const auto a1 = *(input1_ptr + 2 * x + 1); |
| const auto b0 = *(input2_ptr + 2 * x); |
| const auto b1 = *(input2_ptr + 2 * x + 1); |
| auto res1 = a0 * b0 - a1 * b1; |
| auto res2 = a0 * b1 + a1 * b0; |
| *(output_ptr + 2 * x) = res1; |
| *(output_ptr + 2 * x + 1) = res2; |
| } |
| }, |
| input1, input2, dst); |
| } |
| } |
| |
| template <bool is_scale255, bool is_sat> |
| void mul_U8_U8_S16(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, int n) |
| { |
| // Create input windows |
| Window win = window; |
| Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); |
| |
| // Clear X Dimension on execution window as we handle manually |
| win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator input1(src1, input1_win); |
| Iterator input2(src2, input2_win); |
| Iterator dst(out, win); |
| |
| const int window_step_x = 16 / sizeof(uint8_t); |
| const auto window_start_x = static_cast<int>(window.x().start()); |
| const auto window_end_x = static_cast<int>(window.x().end()); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<int16_t *>(dst.ptr()); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const uint8x16_t bv = wrapper::vloadq(input2_ptr + x); |
| const uint8x16_t av = wrapper::vloadq(input1_ptr + x); |
| |
| uint16x8_t tmp_low = vmovl_u8(vget_low_u8(av)); |
| uint16x8_t tmp_high = vmovl_u8(vget_high_u8(av)); |
| tmp_low = vmulq_u16(tmp_low, vmovl_u8(vget_low_u8(bv))); |
| tmp_high = vmulq_u16(tmp_high, vmovl_u8(vget_high_u8(bv))); |
| |
| if (is_scale255) |
| { |
| tmp_low = scale255_U16_U16(tmp_low); |
| tmp_high = scale255_U16_U16(tmp_high); |
| } |
| else |
| { |
| const int16x8_t vn = vdupq_n_s16(-n); |
| |
| if (is_sat) |
| { |
| tmp_low = vqshlq_u16(tmp_low, vn); |
| tmp_high = vqshlq_u16(tmp_high, vn); |
| } |
| else |
| { |
| tmp_low = vshlq_u16(tmp_low, vn); |
| tmp_high = vshlq_u16(tmp_high, vn); |
| } |
| } |
| |
| if (is_sat) |
| { |
| static const uint16x8_t max = vdupq_n_u16(SHRT_MAX); |
| |
| tmp_low = vminq_u16(tmp_low, max); |
| tmp_high = vminq_u16(tmp_high, max); |
| } |
| |
| vst1q_s16(output_ptr + x, vreinterpretq_s16_u16(tmp_low)); |
| vst1q_s16(output_ptr + x + 8, vreinterpretq_s16_u16(tmp_high)); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| int32_t tmp = static_cast<int32_t>(*(input1_ptr + x)) * static_cast<int32_t>(*(input2_ptr + x)); |
| |
| if (is_scale255) |
| { |
| float tmp_f = static_cast<float>(tmp) * scale255_constant; |
| tmp = static_cast<int32_t>(tmp_f + 0.5f); |
| } |
| else |
| { |
| tmp >>= n; |
| } |
| |
| if (is_sat) |
| { |
| tmp = (tmp > SHRT_MAX) ? SHRT_MAX : tmp; |
| } |
| |
| *(output_ptr + x) = static_cast<int16_t>(tmp); |
| } |
| }, |
| input1, input2, dst); |
| } |
| |
| template <bool is_scale255, bool is_sat> |
| void mul_S16_U8_S16(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, int n) |
| { |
| // Create input windows |
| Window win = window; |
| Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); |
| |
| // Clear X Dimension on execution window as we handle manually |
| win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| |
| Iterator input1(src1, input1_win); |
| Iterator input2(src2, input2_win); |
| Iterator dst(out, win); |
| |
| const int window_step_x = 16; |
| const auto window_start_x = static_cast<int>(window.x().start()); |
| const auto window_end_x = static_cast<int>(window.x().end()); |
| |
| execute_window_loop( |
| win, |
| [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<int16_t *>(dst.ptr()); |
| |
| // Compute window_step_x elements per iteration |
| int x = window_start_x; |
| for (; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const int16x8x2_t ta1 = {{ |
| vld1q_s16(input1_ptr + x), |
| vld1q_s16(input1_ptr + x + 8), |
| }}; |
| const uint8x8x2_t ta2u = {{ |
| vld1_u8(input2_ptr + x), |
| vld1_u8(input2_ptr + x + 8), |
| }}; |
| const int16x8x2_t ta2 = { |
| {vreinterpretq_s16_u16(vmovl_u8(ta2u.val[0])), vreinterpretq_s16_u16(vmovl_u8(ta2u.val[1]))}}; |
| |
| const int16x8x2_t result = mul_S16_S16_S16_n_k<is_scale255, is_sat>(ta1, ta2, n); |
| |
| vst1q_s16(output_ptr + x, result.val[0]); |
| vst1q_s16(output_ptr + x + 8, result.val[1]); |
| } |
| |
| // Compute left-over elements |
| for (; x < window_end_x; ++x) |
| { |
| int32_t tmp = static_cast<int32_t>(*(input1_ptr + x)) * static_cast<int32_t>(*(input2_ptr + x)); |
| |
| if (is_scale255) |
| { |
| float tmp_f = static_cast<float>(tmp) * scale255_constant; |
| |
| tmp = static_cast<int32_t>(tmp_f + 0.5f); |
| } |
| else |
| { |
| if (tmp >= 0) |
| { |
| tmp >>= n; |
| } |
| else |
| { |
| uint32_t mask = (1u << n) - 1; |
| tmp = (tmp + static_cast<int32_t>(mask)) >> n; |
| } |
| } |
| if (is_sat) |
| { |
| tmp = (tmp > SHRT_MAX) ? SHRT_MAX : ((tmp < SHRT_MIN) ? SHRT_MIN : tmp); |
| } |
| *(output_ptr + x) = static_cast<int16_t>(tmp); |
| } |
| }, |
| input1, input2, dst); |
| } |
| |
| template <bool is_scale255, bool is_sat> |
| void mul_U8_S16_S16(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, int n) |
| { |
| // Simply swap the two input buffers |
| mul_S16_U8_S16<is_scale255, is_sat>(src2, src1, out, window, n); |
| } |
| } // namespace |
| |
| void CpuMulKernel::configure(ITensorInfo *src1, |
| ITensorInfo *src2, |
| ITensorInfo *dst, |
| float scale, |
| ConvertPolicy overflow_policy, |
| RoundingPolicy rounding_policy) |
| { |
| ARM_COMPUTE_UNUSED(rounding_policy); |
| ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, dst); |
| |
| ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src1, src2, dst, scale, overflow_policy, rounding_policy)); |
| |
| const TensorShape &out_shape = TensorShape::broadcast_shape(src1->tensor_shape(), src2->tensor_shape()); |
| |
| // Auto initialize dst if not initialized |
| set_shape_if_empty(*dst, out_shape); |
| |
| _scale = scale; |
| _scale_exponent = 0; |
| _func_quantized = nullptr; |
| _func_int = nullptr; |
| _func_float = nullptr; |
| |
| bool is_scale_255 = false; |
| // Check and validate scaling factor |
| if (std::abs(scale - scale255_constant) < 0.00001f) |
| { |
| is_scale_255 = true; |
| } |
| else |
| { |
| int exponent = 0; |
| |
| std::frexp(scale, &exponent); |
| |
| // Store the positive exponent. We know that we compute 1/2^n |
| // Additionally we need to subtract 1 to compensate that frexp used a mantissa of 0.5 |
| _scale_exponent = std::abs(exponent - 1); |
| } |
| |
| const DataType dt_input1 = src1->data_type(); |
| const DataType dt_input2 = src2->data_type(); |
| const DataType dt_output = dst->data_type(); |
| const bool is_sat = (overflow_policy == ConvertPolicy::SATURATE); |
| |
| switch (dt_input1) |
| { |
| case DataType::QASYMM8: |
| if (dt_input2 == DataType::QASYMM8 && dt_output == DataType::QASYMM8) |
| { |
| if (mul_q8_neon_fixedpoint_possible(src1, src2, dst, scale)) |
| { |
| _func_quantized = &mul_q8_neon_fixedpoint<uint8_t>; |
| } |
| else |
| { |
| _func_quantized = &mul_saturate_quantized_8<uint8_t>; |
| } |
| } |
| break; |
| case DataType::QASYMM8_SIGNED: |
| if (dt_input2 == DataType::QASYMM8_SIGNED) |
| { |
| if (mul_q8_neon_fixedpoint_possible(src1, src2, dst, scale)) |
| { |
| _func_quantized = &mul_q8_neon_fixedpoint<int8_t>; |
| } |
| else |
| { |
| _func_quantized = &mul_saturate_quantized_8<int8_t>; |
| } |
| } |
| break; |
| case DataType::QSYMM16: |
| if (dt_input2 == DataType::QSYMM16 && dt_output == DataType::QSYMM16) |
| { |
| _func_quantized = &mul_saturate_QSYMM16_QSYMM16_QSYMM16; |
| } |
| else if (dt_input2 == DataType::QSYMM16 && dt_output == DataType::S32) |
| { |
| _func_int = &mul_QSYMM16_QSYMM16_S32; |
| } |
| break; |
| case DataType::S16: |
| if (DataType::U8 == dt_input2 && DataType::S16 == dt_output) |
| { |
| if (is_scale_255) |
| { |
| _func_int = is_sat ? &mul_S16_U8_S16<true, true> : &mul_S16_U8_S16<true, false>; |
| } |
| else |
| { |
| _func_int = is_sat ? &mul_S16_U8_S16<false, true> : &mul_S16_U8_S16<false, false>; |
| } |
| } |
| if (DataType::S16 == dt_input2 && DataType::S16 == dt_output) |
| { |
| if (is_scale_255) |
| { |
| _func_int = is_sat ? &mul_S16_S16_S16<true, true> : &mul_S16_S16_S16<true, false>; |
| } |
| else |
| { |
| _func_int = is_sat ? &mul_S16_S16_S16<false, true> : &mul_S16_S16_S16<false, false>; |
| } |
| } |
| break; |
| case DataType::S32: |
| if (DataType::S32 == dt_input2 && DataType::S32 == dt_output) |
| { |
| _func_int = is_sat ? &mul_S32_S32_S32<true> : &mul_S32_S32_S32<false>; |
| } |
| break; |
| case DataType::U8: |
| if (DataType::U8 == dt_input2 && DataType::U8 == dt_output) |
| { |
| if (is_scale_255) |
| { |
| _func_int = is_sat ? &mul_U8_U8_U8<true, true> : &mul_U8_U8_U8<true, false>; |
| } |
| else |
| { |
| _func_int = is_sat ? &mul_U8_U8_U8<false, true> : &mul_U8_U8_U8<false, false>; |
| } |
| } |
| else if (DataType::U8 == dt_input2 && DataType::S16 == dt_output) |
| { |
| if (is_scale_255) |
| { |
| _func_int = is_sat ? &mul_U8_U8_S16<true, true> : &mul_U8_U8_S16<true, false>; |
| } |
| else |
| { |
| _func_int = is_sat ? &mul_U8_U8_S16<false, true> : &mul_U8_U8_S16<false, false>; |
| } |
| } |
| else if (DataType::S16 == dt_input2 && DataType::S16 == dt_output) |
| { |
| if (is_scale_255) |
| { |
| _func_int = is_sat ? &mul_U8_S16_S16<true, true> : &mul_U8_S16_S16<true, false>; |
| } |
| else |
| { |
| _func_int = is_sat ? &mul_U8_S16_S16<false, true> : &mul_U8_S16_S16<false, false>; |
| } |
| } |
| break; |
| case DataType::F16: |
| _func_float = REGISTER_FP16_NEON(cpu::mul_F16_F16_F16); |
| break; |
| case DataType::F32: |
| _func_float = REGISTER_FP32_NEON(cpu::mul_F32_F32_F32); |
| break; |
| default: |
| ARM_COMPUTE_ERROR("You called with the wrong img formats"); |
| } |
| |
| // Configure kernel window |
| Window win; |
| std::tie(win, _split_dimension) = calculate_squashed_or_max_window(*src1, *src2); |
| |
| ICpuKernel::configure(win); |
| } |
| |
| size_t CpuMulKernel::get_mws(const CPUInfo &platform, size_t thread_count) const |
| { |
| ARM_COMPUTE_UNUSED(thread_count); |
| |
| #if defined(ENABLE_FP32_KERNELS) |
| if (this->_func_float == &mul_F32_F32_F32) |
| { |
| size_t mws = ICPPKernel::default_mws; |
| if (platform.get_cpu_model() == CPUModel::N1) |
| { |
| mws = default_mws_N1_fp32_neon; |
| } |
| else if (platform.get_cpu_model() == CPUModel::V1) |
| { |
| mws = default_mws_V1_fp32_neon; |
| } |
| else |
| { |
| if (_split_dimension == Window::DimX) |
| { |
| // Don't split the work load too small if the tensor has been reinterpreted as 1D. |
| // This number is loosely chosen as threading overhead in each platform varies wildly. |
| return default_mws_other_platforms_1d_tensor; |
| } |
| return default_mws; |
| } |
| |
| // tensor is 1D or was re-interpreted as 1D |
| if (this->window().shape().num_dimensions() == 1) |
| { |
| return mws; |
| } |
| else |
| { |
| // scale mws down by the number of elements along all the dimensions (x, z, w, etc) except the one |
| // that we parallelize along (the y dimension). This allows for parallelization when the Y_SIZE is small |
| // but the other sizes are large, which boosts performance. |
| mws = static_cast<size_t>(mws / (this->window().num_iterations_total() / this->window().num_iterations(1))); |
| return std::max(static_cast<size_t>(1), mws); |
| } |
| } |
| #else /* ENABLE_FP32_KERNELS */ |
| ARM_COMPUTE_UNUSED(platform); |
| #endif /* ENABLE_FP32_KERNELS */ |
| if (_split_dimension == Window::DimX) |
| { |
| // Don't split the work load too small if the tensor has been reinterpreted as 1D. |
| // This number is loosely chosen as threading overhead in each platform varies wildly. |
| return default_mws_other_platforms_1d_tensor; |
| } |
| return default_mws; |
| } |
| |
| Status CpuMulKernel::validate(const ITensorInfo *src1, |
| const ITensorInfo *src2, |
| const ITensorInfo *dst, |
| float scale, |
| ConvertPolicy overflow_policy, |
| RoundingPolicy rounding_policy) |
| { |
| ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, dst); |
| ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src1, src2, dst, scale, overflow_policy, rounding_policy)); |
| |
| return Status{}; |
| } |
| |
| void CpuMulKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) |
| { |
| ARM_COMPUTE_UNUSED(info); |
| ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); |
| |
| auto src1 = tensors.get_const_tensor(TensorType::ACL_SRC_0); |
| auto src2 = tensors.get_const_tensor(TensorType::ACL_SRC_1); |
| auto dst = tensors.get_tensor(TensorType::ACL_DST); |
| |
| if (_func_quantized != nullptr) |
| { |
| (*_func_quantized)(src1, src2, dst, window, _scale); |
| } |
| else if (_func_int != nullptr) |
| { |
| (*_func_int)(src1, src2, dst, window, _scale_exponent); |
| } |
| else |
| { |
| ARM_COMPUTE_ERROR_ON(_func_float == nullptr); |
| (*_func_float)(src1, src2, dst, window, _scale); |
| } |
| } |
| |
| const char *CpuMulKernel::name() const |
| { |
| return "CpuMulKernel"; |
| } |
| |
| namespace |
| { |
| Status validate_arguments_complex(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst) |
| { |
| ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 2, DataType::F32); |
| ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src2, 2, DataType::F32); |
| |
| const TensorShape &out_shape = TensorShape::broadcast_shape(src1->tensor_shape(), src2->tensor_shape()); |
| |
| ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); |
| |
| // Validate in case of configured dst |
| if (dst->total_size() > 0) |
| { |
| ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 2, DataType::F32); |
| ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, dst->tensor_shape(), 0), |
| "Wrong shape for dst"); |
| } |
| |
| return Status{}; |
| } |
| } // namespace |
| |
| void CpuComplexMulKernel::configure(ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst) |
| { |
| ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, dst); |
| ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_complex(src1, src2, dst)); |
| |
| const TensorShape &out_shape = TensorShape::broadcast_shape(src1->tensor_shape(), src2->tensor_shape()); |
| |
| // Auto initialize dst if not initialized |
| const TensorInfo out_info(out_shape, src1->num_channels(), src1->data_type()); |
| auto_init_if_empty(*dst, out_info); |
| |
| // Configure kernel window |
| Window win = calculate_max_window(out_shape); |
| |
| ICpuKernel::configure(win); |
| } |
| |
| Status CpuComplexMulKernel::validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst) |
| { |
| ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, dst); |
| ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_complex(src1, src2, dst)); |
| |
| return Status{}; |
| } |
| |
| void CpuComplexMulKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) |
| { |
| ARM_COMPUTE_UNUSED(info); |
| ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); |
| |
| auto src1 = tensors.get_const_tensor(TensorType::ACL_SRC_0); |
| auto src2 = tensors.get_const_tensor(TensorType::ACL_SRC_1); |
| auto dst = tensors.get_tensor(TensorType::ACL_DST); |
| |
| c_mul_F32_F32_F32_n(src1, src2, dst, window); |
| } |
| |
| const char *CpuComplexMulKernel::name() const |
| { |
| return "CpuComplexMulKernel"; |
| } |
| } // namespace kernels |
| } // namespace cpu |
| } // namespace arm_compute |