| /* |
| * Copyright (c) 2021-2022 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. |
| */ |
| #ifndef SRC_CORE_NEON_KERNELS_ELEMENTWISE_IMPL_H |
| #define SRC_CORE_NEON_KERNELS_ELEMENTWISE_IMPL_H |
| |
| #include "src/core/NEON/NEAsymm.h" |
| |
| namespace arm_compute |
| { |
| namespace cpu |
| { |
| template <ArithmeticOperation op, typename VectorType> |
| typename VectorType::type elementwise_arithm_op(const typename VectorType::type &a, const typename VectorType::type &b) |
| { |
| using vec_type = typename VectorType::type; |
| using scalar_type = typename VectorType::scalar_type; |
| using tag_type = typename VectorType::tag_type; |
| |
| vec_type res = wrapper::vdup_n(static_cast<scalar_type>(0), tag_type{}); |
| |
| switch(op) |
| { |
| case ArithmeticOperation::MAX: |
| res = wrapper::vmax(a, b); |
| break; |
| case ArithmeticOperation::MIN: |
| res = wrapper::vmin(a, b); |
| break; |
| case ArithmeticOperation::SQUARED_DIFF: |
| { |
| const vec_type tmp = wrapper::vsub(a, b); |
| res = wrapper::vmul(tmp, tmp); |
| break; |
| } |
| case ArithmeticOperation::PRELU: |
| { |
| const vec_type zero = wrapper::vdup_n(static_cast<scalar_type>(0), tag_type{}); |
| const vec_type tmp = wrapper::vmul(a, b); |
| const auto gt = wrapper::vcgt(a, zero); |
| |
| res = wrapper::vbsl(gt, a, tmp); |
| break; |
| } |
| |
| default: |
| ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); |
| } |
| |
| return res; |
| } |
| template <ArithmeticOperation op, typename ScalarType, typename VectorType> |
| typename VectorType::type elementwise_arithm_op_broadcast(const typename VectorType::type &a, const ScalarType &broadcast_value, const bool reorder) |
| { |
| using tag_type = typename VectorType::tag_type; |
| using vec_type = typename VectorType::type; |
| |
| vec_type broadcast_vector = wrapper::vdup_n(broadcast_value, tag_type{}); |
| return elementwise_arithm_op<op, VectorType>(reorder ? broadcast_vector : a, reorder ? a : broadcast_vector); |
| } |
| |
| template <typename InputScalarType, typename OutputScalarType, typename InputVectorType> |
| void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, |
| OutputScalarType (*scalar_func)(const InputScalarType &, const InputScalarType &), |
| int (*broadcast_func)(int, int, int, const InputScalarType *, const InputScalarType &, OutputScalarType *, const bool), |
| int (*neon_func)(int, int, int, const InputScalarType *, const InputScalarType *, OutputScalarType *)) |
| { |
| // Create input windows |
| Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(in2->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 = std::min(16 / static_cast<int>(sizeof(OutputScalarType)), 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 = in1->info()->tensor_shape().x() != in2->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 ? in2 : in1; |
| const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; |
| |
| // 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 output(out, win); |
| |
| execute_window_loop(win, [&](const Coordinates &) |
| { |
| auto output_ptr = reinterpret_cast<OutputScalarType *>(output.ptr()); |
| const auto non_broadcast_input_ptr = reinterpret_cast<const InputScalarType *>(non_broadcast_input.ptr()); |
| const InputScalarType broadcast_value = *reinterpret_cast<const InputScalarType *>(broadcast_input.ptr()); |
| |
| int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_value, output_ptr, !is_broadcast_input_2); |
| for(; x < window_end_x; ++x) |
| { |
| const auto a = *(non_broadcast_input_ptr + x); |
| *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? broadcast_value : a, !is_broadcast_input_2 ? a : broadcast_value); |
| } |
| }, |
| broadcast_input, non_broadcast_input, output); |
| } |
| 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(in1, input1_win); |
| Iterator input2(in2, input2_win); |
| Iterator output(out, win); |
| |
| execute_window_loop(win, [&](const Coordinates &) |
| { |
| auto output_ptr = reinterpret_cast<OutputScalarType *>(output.ptr()); |
| const auto input1_ptr = reinterpret_cast<const InputScalarType *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const InputScalarType *>(input2.ptr()); |
| |
| int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr); |
| for(; x < window_end_x; ++x) |
| { |
| const auto a = *(input1_ptr + x); |
| const auto b = *(input2_ptr + x); |
| *(output_ptr + x) = (*scalar_func)(a, b); |
| } |
| }, |
| input1, input2, output); |
| } |
| } |
| |
| template <ArithmeticOperation op, typename ScalarType> |
| inline ScalarType elementwise_arithm_op_scalar(const ScalarType &a, const ScalarType &b) |
| { |
| auto res = ScalarType(0); |
| |
| switch(op) |
| { |
| case ArithmeticOperation::MAX: |
| res = std::max(a, b); |
| break; |
| case ArithmeticOperation::MIN: |
| res = std::min(a, b); |
| break; |
| case ArithmeticOperation::SQUARED_DIFF: |
| { |
| res = (a - b) * (a - b); |
| break; |
| } |
| case ArithmeticOperation::PRELU: |
| { |
| res = (a > 0 ? a : a * b); |
| break; |
| } |
| case ArithmeticOperation::DIV: |
| { |
| res = a / b; |
| if(std::is_integral<ScalarType>::value) |
| { |
| res = (b == 0) ? 0 : res; |
| if(static_cast<int32_t>(a) % static_cast<int32_t>(b) != 0 && ((a < 0) != (b < 0))) |
| { |
| --res; |
| } |
| } |
| break; |
| } |
| case ArithmeticOperation::POWER: |
| { |
| res = std::pow(a, b); |
| break; |
| } |
| default: |
| ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); |
| } |
| return res; |
| } |
| |
| template <> |
| inline int32x4_t elementwise_arithm_op<ArithmeticOperation::DIV, typename wrapper::traits::neon_vector<int32_t, 4>>(const int32x4_t &a, const int32x4_t &b) |
| { |
| return vcvtq_s32_f32(vfloorq_f32(wrapper::vdiv(vcvtq_f32_s32(a), vcvtq_f32_s32(b)))); |
| } |
| |
| template <> |
| inline float32x4_t elementwise_arithm_op<ArithmeticOperation::DIV, typename wrapper::traits::neon_vector<float, 4>>(const float32x4_t &a, const float32x4_t &b) |
| { |
| return wrapper::vdiv(a, b); |
| } |
| |
| template <> |
| inline float32x4_t elementwise_arithm_op<ArithmeticOperation::POWER, typename wrapper::traits::neon_vector<float, 4>>(const float32x4_t &a, const float32x4_t &b) |
| { |
| return wrapper::vpow(a, b); |
| } |
| |
| #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| template <> |
| inline float16x8_t elementwise_arithm_op<ArithmeticOperation::DIV, typename wrapper::traits::neon_vector<float16_t, 8>>(const float16x8_t &a, const float16x8_t &b) |
| { |
| return wrapper::vdiv(a, b); |
| } |
| |
| template <> |
| inline float16x8_t elementwise_arithm_op<ArithmeticOperation::POWER, typename wrapper::traits::neon_vector<float16_t, 8>>(const float16x8_t &a, const float16x8_t &b) |
| { |
| return wrapper::vpow(a, b); |
| } |
| #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| |
| template <ArithmeticOperation op, typename ScalarType, typename VectorType> |
| inline int elementwise_arithm_op_loop(int window_start_x, int window_end_x, int window_step_x, |
| const ScalarType *input1_ptr, const ScalarType *input2_ptr, ScalarType *output_ptr) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const auto a = wrapper::vloadq(input1_ptr + x); |
| const auto b = wrapper::vloadq(input2_ptr + x); |
| wrapper::vstore(output_ptr + x, elementwise_arithm_op<op, VectorType>(a, b)); |
| } |
| return x; |
| } |
| |
| template <ArithmeticOperation op, typename ScalarType, typename VectorType> |
| inline int elementwise_arithm_op_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, |
| const ScalarType *non_broadcast_input_ptr, const ScalarType &broadcast_value, ScalarType *output_ptr, const bool reorder) |
| { |
| 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 + x)); |
| wrapper::vstore(output_ptr + x, elementwise_arithm_op_broadcast<op, ScalarType, VectorType>(a, broadcast_value, reorder)); |
| } |
| return x; |
| } |
| |
| template <ArithmeticOperation op, typename VectorType> |
| void elementwise_arithm_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| { |
| using scalar_type = typename VectorType::scalar_type; |
| |
| elementwise_op<scalar_type, scalar_type, VectorType>(in1, in2, out, window, |
| &elementwise_arithm_op_scalar<op, scalar_type>, |
| &elementwise_arithm_op_broadcast_loop<op, scalar_type, VectorType>, |
| &elementwise_arithm_op_loop<op, scalar_type, VectorType>); |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType> |
| inline uint8_t elementwise_comp_op_scalar(const InputScalarType &a, const InputScalarType &b) |
| { |
| bool res = false; |
| |
| switch(op) |
| { |
| case ComparisonOperation::Equal: |
| res = (a == b); |
| break; |
| case ComparisonOperation::NotEqual: |
| res = (a != b); |
| break; |
| case ComparisonOperation::Greater: |
| res = (a > b); |
| break; |
| case ComparisonOperation::GreaterEqual: |
| res = (a >= b); |
| break; |
| case ComparisonOperation::Less: |
| res = (a < b); |
| break; |
| case ComparisonOperation::LessEqual: |
| res = (a <= b); |
| break; |
| default: |
| ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); |
| } |
| return res ? ~static_cast<uint8_t>(0) : static_cast<uint8_t>(0); |
| } |
| |
| template <ComparisonOperation op, typename InputVectorType, typename OutputVectorType> |
| inline OutputVectorType elementwise_comp_op(const InputVectorType &a, const InputVectorType &b) |
| { |
| OutputVectorType res = { 0, 0, 0, 0 }; |
| |
| switch(op) |
| { |
| case ComparisonOperation::Equal: |
| res = wrapper::vceq(a, b); |
| break; |
| case ComparisonOperation::NotEqual: |
| res = wrapper::vnot(wrapper::vceq(a, b)); |
| break; |
| case ComparisonOperation::Greater: |
| res = wrapper::vcgt(a, b); |
| break; |
| case ComparisonOperation::GreaterEqual: |
| res = wrapper::vcge(a, b); |
| break; |
| case ComparisonOperation::Less: |
| res = wrapper::vcgt(b, a); |
| break; |
| case ComparisonOperation::LessEqual: |
| res = wrapper::vcge(b, a); |
| break; |
| default: |
| ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); |
| } |
| |
| return res; |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType, typename InputVectorType, typename OutputVectorType> |
| inline OutputVectorType elementwise_comp_op_broadcast(const InputVectorType &a, const InputScalarType &broadcast_value, const bool reorder) |
| { |
| InputVectorType broadcast_vector = wrapper::vdup_n(broadcast_value, wrapper::traits::vector_128_tag()); |
| return elementwise_comp_op<op, InputVectorType, OutputVectorType>(reorder ? broadcast_vector : a, reorder ? a : broadcast_vector); |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType, typename InputVectorType> |
| inline int elementwise_comp_op_broadcast_8_loop(int window_start_x, int window_end_x, int window_step_x, |
| const InputScalarType *non_broadcast_input_ptr, const InputScalarType &broadcast_value, uint8_t *output_ptr, const bool reorder) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const auto a = elementwise_comp_op_broadcast<op, InputScalarType, InputVectorType, uint8x16_t>(wrapper::vloadq((non_broadcast_input_ptr + x)), broadcast_value, reorder); |
| wrapper::vstore(output_ptr + x, a); |
| } |
| return x; |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType, typename InputVectorType> |
| inline int elementwise_comp_op_broadcast_16_loop(int window_start_x, int window_end_x, int window_step_x, |
| const InputScalarType *non_broadcast_input_ptr, const InputScalarType &broadcast_value, uint8_t *output_ptr, const bool reorder) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const auto a = elementwise_comp_op_broadcast<op, InputScalarType, InputVectorType, uint16x8_t>(wrapper::vloadq((non_broadcast_input_ptr + x)), broadcast_value, reorder); |
| wrapper::vstore(output_ptr + x, wrapper::vmovn(a)); |
| } |
| return x; |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType, typename InputVectorType> |
| inline int elementwise_comp_op_broadcast_32_loop(int window_start_x, int window_end_x, int window_step_x, |
| const InputScalarType *non_broadcast_input_ptr, const InputScalarType &broadcast_value, uint8_t *output_ptr, const bool reorder) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const auto a = elementwise_comp_op_broadcast<op, InputScalarType, InputVectorType, uint32x4_t>(wrapper::vloadq(non_broadcast_input_ptr + x), broadcast_value, reorder); |
| const auto b = elementwise_comp_op_broadcast<op, InputScalarType, InputVectorType, uint32x4_t>(wrapper::vloadq(non_broadcast_input_ptr + x + 4), broadcast_value, reorder); |
| wrapper::vstore(output_ptr + x, wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(a), wrapper::vmovn(b)))); |
| } |
| if(x <= window_end_x - 4) |
| { |
| const auto a = elementwise_comp_op_broadcast<op, InputScalarType, InputVectorType, uint32x4_t>(wrapper::vloadq((non_broadcast_input_ptr + x)), broadcast_value, reorder); |
| for(int i = 0; i < 4; i++) |
| { |
| *(output_ptr + x + i) = wrapper::vgetlane(a, i); |
| } |
| x = +4; |
| } |
| return x; |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType, typename InputVectorType> |
| inline int elementwise_comp_op_8_loop(int window_start_x, int window_end_x, int window_step_x, |
| const InputScalarType *input1_ptr, const InputScalarType *input2_ptr, uint8_t *output_ptr) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const auto a = wrapper::vloadq(input1_ptr + x); |
| const auto b = wrapper::vloadq(input2_ptr + x); |
| const auto res = elementwise_comp_op<op, InputVectorType, uint8x16_t>(a, b); |
| wrapper::vstore(output_ptr + x, res); |
| } |
| return x; |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType, typename InputVectorType> |
| inline int elementwise_comp_op_16_loop(int window_start_x, int window_end_x, int window_step_x, |
| const InputScalarType *input1_ptr, const InputScalarType *input2_ptr, uint8_t *output_ptr) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const auto a = wrapper::vloadq(input1_ptr + x); |
| const auto b = wrapper::vloadq(input2_ptr + x); |
| const auto res = elementwise_comp_op<op, InputVectorType, uint16x8_t>(a, b); |
| wrapper::vstore(output_ptr + x, wrapper::vmovn(res)); |
| } |
| return x; |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType, typename InputVectorType> |
| inline int elementwise_comp_op_32_loop(int window_start_x, int window_end_x, int window_step_x, |
| const InputScalarType *input1_ptr, const InputScalarType *input2_ptr, uint8_t *output_ptr) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| auto a = wrapper::vloadq(input1_ptr + x); |
| auto b = wrapper::vloadq(input2_ptr + x); |
| const auto res = elementwise_comp_op<op, InputVectorType, uint32x4_t>(a, b); |
| a = wrapper::vloadq(input1_ptr + x + 4); |
| b = wrapper::vloadq(input2_ptr + x + 4); |
| const auto res2 = elementwise_comp_op<op, InputVectorType, uint32x4_t>(a, b); |
| wrapper::vstore(output_ptr + x, wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(res), wrapper::vmovn(res2)))); |
| } |
| if(x <= window_end_x - 4) |
| { |
| const auto a = wrapper::vloadq(input1_ptr + x); |
| const auto b = wrapper::vloadq(input2_ptr + x); |
| const auto res = elementwise_comp_op<op, InputVectorType, uint32x4_t>(a, b); |
| for(int i = 0; i < 4; i++) |
| { |
| *(output_ptr + x + i) = wrapper::vgetlane(res, i); |
| } |
| x = +4; |
| } |
| return x; |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType, typename InputVectorType> |
| void elementwise_comp_op_8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| { |
| elementwise_op<InputScalarType, uint8_t, InputVectorType>(in1, in2, out, window, |
| &elementwise_comp_op_scalar<op, InputScalarType>, |
| &elementwise_comp_op_broadcast_8_loop<op, InputScalarType, InputVectorType>, |
| &elementwise_comp_op_8_loop<op, InputScalarType, InputVectorType>); |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType, typename InputVectorType> |
| void elementwise_comp_op_16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| { |
| elementwise_op<InputScalarType, uint8_t, InputVectorType>(in1, in2, out, window, |
| &elementwise_comp_op_scalar<op, InputScalarType>, |
| &elementwise_comp_op_broadcast_16_loop<op, InputScalarType, InputVectorType>, |
| &elementwise_comp_op_16_loop<op, InputScalarType, InputVectorType>); |
| } |
| |
| template <ComparisonOperation op, typename InputScalarType, typename InputVectorType> |
| void elementwise_comp_op_32(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| { |
| elementwise_op<InputScalarType, uint8_t, InputVectorType>(in1, in2, out, window, |
| &elementwise_comp_op_scalar<op, InputScalarType>, |
| &elementwise_comp_op_broadcast_32_loop<op, InputScalarType, InputVectorType>, |
| &elementwise_comp_op_32_loop<op, InputScalarType, InputVectorType>); |
| } |
| |
| inline float32x4x4_t load_quantized(const uint8_t *input1_ptr, const int32x4_t &offset, const float32x4_t &scale) |
| { |
| qasymm8x16_t x = vld1q_u8(input1_ptr); |
| const float32x4x4_t out = |
| { |
| { |
| vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(x))))), offset)), scale), |
| vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(x))))), offset)), scale), |
| vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(x))))), offset)), scale), |
| vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(x))))), offset)), scale), |
| } |
| }; |
| return out; |
| } |
| |
| inline float32x4x4_t load_quantized_signed(const int8_t *input1_ptr, const int32x4_t &offset, const float32x4_t &scale) |
| { |
| qasymm8x16_signed_t x = vld1q_s8(input1_ptr); |
| const float32x4x4_t out = |
| { |
| { |
| vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(x)))), offset)), scale), |
| vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(x)))), offset)), scale), |
| vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(x)))), offset)), scale), |
| vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(x)))), offset)), scale), |
| } |
| }; |
| return out; |
| } |
| |
| inline void store_quantized(uint8_t *output_ptr, const uint32x4x4_t &out) |
| { |
| const uint8x8_t pa = vqmovn_u16(vcombine_u16(vqmovn_u32(out.val[0]), vqmovn_u32(out.val[1]))); |
| const uint8x8_t pb = vqmovn_u16(vcombine_u16(vqmovn_u32(out.val[2]), vqmovn_u32(out.val[3]))); |
| vst1q_u8(output_ptr, vcombine_u8(pa, pb)); |
| } |
| |
| inline void store_quantized(uint8_t *output_ptr, const int32x4x4_t &out) |
| { |
| const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(out.val[0]), vqmovn_s32(out.val[1]))); |
| const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(out.val[2]), vqmovn_s32(out.val[3]))); |
| vst1q_u8(output_ptr, vcombine_u8(pa, pb)); |
| } |
| |
| inline void store_quantized(uint8_t *output_ptr, const float32x4x4_t &rf, const float32x4_t &offset, const float32x4_t &invscale) |
| { |
| int32x4x4_t out = |
| { |
| { |
| vcvtq_s32_f32(vmlaq_f32(offset, rf.val[0], invscale)), |
| vcvtq_s32_f32(vmlaq_f32(offset, rf.val[1], invscale)), |
| vcvtq_s32_f32(vmlaq_f32(offset, rf.val[2], invscale)), |
| vcvtq_s32_f32(vmlaq_f32(offset, rf.val[3], invscale)), |
| } |
| }; |
| store_quantized(output_ptr, out); |
| } |
| |
| inline void store_quantized_signed(int8_t *output_ptr, const int32x4x4_t &out) |
| { |
| const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(out.val[0]), vqmovn_s32(out.val[1]))); |
| const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(out.val[2]), vqmovn_s32(out.val[3]))); |
| vst1q_s8(output_ptr, vcombine_s8(pa, pb)); |
| } |
| |
| inline void store_quantized_signed(int8_t *output_ptr, const float32x4x4_t &rf, const float32x4_t &offset, const float32x4_t &invscale) |
| { |
| int32x4x4_t out = |
| { |
| { |
| vcvtq_s32_f32(vmlaq_f32(offset, rf.val[0], invscale)), |
| vcvtq_s32_f32(vmlaq_f32(offset, rf.val[1], invscale)), |
| vcvtq_s32_f32(vmlaq_f32(offset, rf.val[2], invscale)), |
| vcvtq_s32_f32(vmlaq_f32(offset, rf.val[3], invscale)), |
| } |
| }; |
| store_quantized_signed(output_ptr, out); |
| } |
| |
| template <ArithmeticOperation op> |
| inline uint8_t elementwise_arithm_op_quantized_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo) |
| { |
| return quantize_qasymm8(elementwise_arithm_op_scalar<op>(a, b), qinfo); |
| } |
| |
| template <ArithmeticOperation op> |
| inline int8_t elementwise_arithm_op_quantized_signed_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo) |
| { |
| return quantize_qasymm8_signed(elementwise_arithm_op_scalar<op>(a, b), qinfo); |
| } |
| |
| template <ArithmeticOperation op> |
| float32x4x4_t elementwise_arithm_op(const float32x4x4_t &a, const float32x4x4_t &b) |
| { |
| using neon_vector_float = wrapper::traits::neon_vector<float, 4>; |
| float32x4x4_t out = |
| { |
| { |
| elementwise_arithm_op<op, neon_vector_float>(a.val[0], b.val[0]), |
| elementwise_arithm_op<op, neon_vector_float>(a.val[1], b.val[1]), |
| elementwise_arithm_op<op, neon_vector_float>(a.val[2], b.val[2]), |
| elementwise_arithm_op<op, neon_vector_float>(a.val[3], b.val[3]), |
| } |
| }; |
| return out; |
| } |
| |
| template <ComparisonOperation op> |
| inline uint8_t elementwise_comp_op_quantized_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo) |
| { |
| ARM_COMPUTE_UNUSED(qinfo); |
| return elementwise_comp_op_scalar<op>(a, b); |
| } |
| |
| template <ComparisonOperation op> |
| inline uint32x4x4_t elementwise_comp_op(const float32x4x4_t &a, const float32x4x4_t &b) |
| { |
| uint32x4x4_t out = |
| { |
| { |
| elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[0], b.val[0]), |
| elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[1], b.val[1]), |
| elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[2], b.val[2]), |
| elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[3], b.val[3]) |
| } |
| }; |
| return out; |
| } |
| |
| template <ArithmeticOperation op> |
| inline int elementwise_arithm_op_quantized_loop(int window_start_x, int window_end_x, int window_step_x, |
| const uint8_t *input1_ptr, const uint8_t *input2_ptr, uint8_t *output_ptr, |
| int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2, |
| float32x4_t voffseto, float32x4_t invvscaleo) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| // Get inputs and compute output |
| const float32x4x4_t af = load_quantized(input1_ptr + x, voffset1, vscale1); |
| const float32x4x4_t bf = load_quantized(input2_ptr + x, voffset2, vscale2); |
| const float32x4x4_t rf = elementwise_arithm_op<op>(af, bf); |
| store_quantized(output_ptr + x, rf, voffseto, invvscaleo); |
| } |
| return x; |
| } |
| |
| template <ArithmeticOperation op> |
| inline int elementwise_arithm_op_quantized_singed_loop(int window_start_x, int window_end_x, int window_step_x, |
| const int8_t *input1_ptr, const int8_t *input2_ptr, int8_t *output_ptr, |
| int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2, |
| float32x4_t voffseto, float32x4_t invvscaleo) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| // Get inputs and compute output |
| const float32x4x4_t af = load_quantized_signed(input1_ptr + x, voffset1, vscale1); |
| const float32x4x4_t bf = load_quantized_signed(input2_ptr + x, voffset2, vscale2); |
| const float32x4x4_t rf = elementwise_arithm_op<op>(af, bf); |
| store_quantized_signed(output_ptr + x, rf, voffseto, invvscaleo); |
| } |
| return x; |
| } |
| |
| template <ArithmeticOperation op> |
| inline int elementwise_arithm_op_quantized_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, |
| const uint8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr, |
| int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast, |
| float32x4_t voffseto, float32x4_t invvscaleo, bool reorder) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const float32x4x4_t af = load_quantized(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast); |
| const float32x4x4_t rf = elementwise_arithm_op<op>(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector); |
| store_quantized(output_ptr + x, rf, voffseto, invvscaleo); |
| } |
| return x; |
| } |
| template <ArithmeticOperation op> |
| inline int elementwise_arithm_op_quantized_signed_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, |
| const int8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, int8_t *output_ptr, |
| int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast, |
| float32x4_t voffseto, float32x4_t invvscaleo, bool reorder) |
| { |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const float32x4x4_t af = load_quantized_signed(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast); |
| const float32x4x4_t rf = elementwise_arithm_op<op>(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector); |
| store_quantized_signed(output_ptr + x, rf, voffseto, invvscaleo); |
| } |
| return x; |
| } |
| |
| template <ComparisonOperation op> |
| inline int elementwise_comp_op_quantized_loop(int window_start_x, int window_end_x, int window_step_x, |
| const uint8_t *input1_ptr, const uint8_t *input2_ptr, uint8_t *output_ptr, |
| int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2, |
| float32x4_t voffseto, float32x4_t invvscaleo) |
| { |
| ARM_COMPUTE_UNUSED(voffseto, invvscaleo); |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const float32x4x4_t af = load_quantized(input1_ptr + x, voffset1, vscale1); |
| const float32x4x4_t bf = load_quantized(input2_ptr + x, voffset2, vscale2); |
| const uint32x4x4_t rf = elementwise_comp_op<op>(af, bf); |
| store_quantized(output_ptr + x, rf); |
| } |
| return x; |
| } |
| |
| template <ComparisonOperation op> |
| inline int elementwise_comp_op_quantized_signed_loop(int window_start_x, int window_end_x, int window_step_x, |
| const int8_t *input1_ptr, const int8_t *input2_ptr, uint8_t *output_ptr, |
| int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2, |
| float32x4_t voffseto, float32x4_t invvscaleo) |
| { |
| ARM_COMPUTE_UNUSED(voffseto, invvscaleo); |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const float32x4x4_t af = load_quantized_signed(input1_ptr + x, voffset1, vscale1); |
| const float32x4x4_t bf = load_quantized_signed(input2_ptr + x, voffset2, vscale2); |
| const uint32x4x4_t rf = elementwise_comp_op<op>(af, bf); |
| store_quantized(output_ptr + x, rf); |
| } |
| return x; |
| } |
| |
| template <ComparisonOperation op> |
| inline int elementwise_comp_op_quantized_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, |
| const uint8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr, |
| int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast, |
| float32x4_t voffseto, float32x4_t invvscaleo, bool reorder) |
| { |
| ARM_COMPUTE_UNUSED(voffseto, invvscaleo); |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const float32x4x4_t af = load_quantized(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast); |
| const uint32x4x4_t rf = elementwise_comp_op<op>(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector); |
| store_quantized(output_ptr + x, rf); |
| } |
| return x; |
| } |
| |
| template <ComparisonOperation op> |
| inline int elementwise_comp_op_quantized_signed_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, |
| const int8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr, |
| int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast, |
| float32x4_t voffseto, float32x4_t invvscaleo, bool reorder) |
| { |
| ARM_COMPUTE_UNUSED(voffseto, invvscaleo); |
| int x = window_start_x; |
| for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| { |
| const float32x4x4_t af = load_quantized_signed(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast); |
| const uint32x4x4_t rf = elementwise_comp_op<op>(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector); |
| store_quantized(output_ptr + x, rf); |
| } |
| return x; |
| } |
| |
| inline void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, |
| uint8_t (*scalar_func)(const float &, const float &, UniformQuantizationInfo), |
| int (*broadcast_func)(int, int, int, const uint8_t *, float32x4x4_t, uint8_t *, int32x4_t, float32x4_t, |
| float32x4_t, float32x4_t, const bool), |
| int (*neon_func)(int, int, int, const uint8_t *, const uint8_t *, uint8_t *, |
| int32x4_t, int32x4_t, float32x4_t, float32x4_t, |
| float32x4_t, float32x4_t)) |
| { |
| // Create input windows |
| Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(in2->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 = 16; |
| 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 = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x(); |
| |
| const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform(); |
| |
| // Output quantization info (add 0.5 to round toward the nearest integer - 0.5 rounds away from zero) |
| const float32x4_t voffseto = vdupq_n_f32(output_qinfo.offset + 0.5f); |
| const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.scale); |
| |
| if(is_broadcast_across_x) |
| { |
| // Select the broadcast input on the X axis |
| 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 ? in2 : in1; |
| const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; |
| |
| const UniformQuantizationInfo broadcast_qinfo = broadcast_tensor->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform(); |
| |
| const int32x4_t voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.offset); |
| const float32x4_t vscale_non_broadcast = vdupq_n_f32(non_broadcast_qinfo.scale); |
| |
| // 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 output(out, win); |
| |
| execute_window_loop(win, [&](const Coordinates &) |
| { |
| const auto non_broadcast_input_ptr = reinterpret_cast<const uint8_t *>(non_broadcast_input.ptr()); |
| const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr()); |
| |
| const uint8_t broadcast_value = *reinterpret_cast<const uint8_t *>(broadcast_input.ptr()); |
| const float32x4x4_t broadcast_vector = vdequantize(vdupq_n_u8(broadcast_value), broadcast_qinfo); |
| |
| int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr, |
| voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2); |
| for(; x < window_end_x; ++x) |
| { |
| const float afs = dequantize_qasymm8(*(non_broadcast_input_ptr + x), non_broadcast_qinfo); |
| const float bfs = dequantize_qasymm8(broadcast_value, broadcast_qinfo); |
| *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? bfs : afs, !is_broadcast_input_2 ? afs : bfs, output_qinfo); |
| } |
| }, |
| broadcast_input, non_broadcast_input, output); |
| } |
| else |
| { |
| const UniformQuantizationInfo input1_qinfo = in1->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo input2_qinfo = in2->info()->quantization_info().uniform(); |
| |
| // Input1 quantization info |
| const int32x4_t voffset1 = vdupq_n_s32(input1_qinfo.offset); |
| const float32x4_t vscale1 = vdupq_n_f32(input1_qinfo.scale); |
| |
| // Input2 quantization info |
| const int32x4_t voffset2 = vdupq_n_s32(input2_qinfo.offset); |
| const float32x4_t vscale2 = vdupq_n_f32(input2_qinfo.scale); |
| |
| // 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(in1, input1_win); |
| Iterator input2(in2, input2_win); |
| Iterator output(out, win); |
| |
| 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 *>(output.ptr()); |
| |
| int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2, |
| vscale1, vscale2, voffseto, invvscaleo); |
| for(; x < window_end_x; ++x) |
| { |
| const float afs = dequantize_qasymm8(*(input1_ptr + x), input1_qinfo); |
| const float bfs = dequantize_qasymm8(*(input2_ptr + x), input2_qinfo); |
| *(output_ptr + x) = (*scalar_func)(afs, bfs, output_qinfo); |
| } |
| }, |
| input1, input2, output); |
| } |
| } |
| |
| inline void elementwise_comp_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, |
| uint8_t (*scalar_func)(const float &, const float &, UniformQuantizationInfo), |
| int (*broadcast_func)(int, int, int, const int8_t *, float32x4x4_t, uint8_t *, int32x4_t, float32x4_t, |
| float32x4_t, float32x4_t, const bool), |
| int (*neon_func)(int, int, int, const int8_t *, const int8_t *, uint8_t *, |
| int32x4_t, int32x4_t, float32x4_t, float32x4_t, |
| float32x4_t, float32x4_t)) |
| { |
| // Create input windows |
| Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(in2->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 = 16; |
| 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 = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x(); |
| |
| const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform(); |
| |
| const float32x4_t voffseto = vdupq_n_f32(output_qinfo.offset); |
| const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.scale); |
| |
| if(is_broadcast_across_x) |
| { |
| // Select the broadcast input on the X axis |
| 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 ? in2 : in1; |
| const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; |
| |
| const UniformQuantizationInfo broadcast_qinfo = broadcast_tensor->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform(); |
| |
| const int32x4_t voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.offset); |
| const float32x4_t vscale_non_broadcast = vdupq_n_f32(non_broadcast_qinfo.scale); |
| |
| // 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 output(out, win); |
| |
| execute_window_loop(win, [&](const Coordinates &) |
| { |
| const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr()); |
| const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr()); |
| |
| const int8_t broadcast_value = *reinterpret_cast<const int8_t *>(broadcast_input.ptr()); |
| const float32x4x4_t broadcast_vector = vdequantize(vdupq_n_s8(broadcast_value), broadcast_qinfo); |
| |
| int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr, |
| voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2); |
| for(; x < window_end_x; ++x) |
| { |
| const float afs = dequantize_qasymm8_signed(*(non_broadcast_input_ptr + x), non_broadcast_qinfo); |
| const float bfs = dequantize_qasymm8_signed(broadcast_value, broadcast_qinfo); |
| *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? bfs : afs, !is_broadcast_input_2 ? afs : bfs, output_qinfo); |
| } |
| }, |
| broadcast_input, non_broadcast_input, output); |
| } |
| else |
| { |
| const UniformQuantizationInfo input1_qinfo = in1->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo input2_qinfo = in2->info()->quantization_info().uniform(); |
| |
| // Input1 quantization info |
| const int32x4_t voffset1 = vdupq_n_s32(input1_qinfo.offset); |
| const float32x4_t vscale1 = vdupq_n_f32(input1_qinfo.scale); |
| |
| // Input2 quantization info |
| const int32x4_t voffset2 = vdupq_n_s32(input2_qinfo.offset); |
| const float32x4_t vscale2 = vdupq_n_f32(input2_qinfo.scale); |
| |
| // 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(in1, input1_win); |
| Iterator input2(in2, input2_win); |
| Iterator output(out, win); |
| |
| execute_window_loop(win, [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr()); |
| |
| int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2, |
| vscale1, vscale2, voffseto, invvscaleo); |
| for(; x < window_end_x; ++x) |
| { |
| const float afs = dequantize_qasymm8_signed(*(input1_ptr + x), input1_qinfo); |
| const float bfs = dequantize_qasymm8_signed(*(input2_ptr + x), input2_qinfo); |
| *(output_ptr + x) = (*scalar_func)(afs, bfs, output_qinfo); |
| } |
| }, |
| input1, input2, output); |
| } |
| } |
| |
| inline void elementwise_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, |
| int8_t (*scalar_func)(const float &, const float &, UniformQuantizationInfo), |
| int (*broadcast_func)(int, int, int, const int8_t *, float32x4x4_t, int8_t *, int32x4_t, float32x4_t, |
| float32x4_t, float32x4_t, const bool), |
| int (*neon_func)(int, int, int, const int8_t *, const int8_t *, int8_t *, |
| int32x4_t, int32x4_t, float32x4_t, float32x4_t, |
| float32x4_t, float32x4_t)) |
| { |
| // Create input windows |
| Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()); |
| Window input2_win = window.broadcast_if_dimension_le_one(in2->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 = 16; |
| 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 = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x(); |
| |
| const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform(); |
| |
| const float32x4_t voffseto = vdupq_n_f32(output_qinfo.offset); |
| const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.scale); |
| |
| if(is_broadcast_across_x) |
| { |
| // Select the broadcast input on the X axis |
| 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 ? in2 : in1; |
| const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; |
| |
| const UniformQuantizationInfo broadcast_qinfo = broadcast_tensor->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform(); |
| |
| const int32x4_t voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.offset); |
| const float32x4_t vscale_non_broadcast = vdupq_n_f32(non_broadcast_qinfo.scale); |
| |
| // 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 output(out, win); |
| |
| execute_window_loop(win, [&](const Coordinates &) |
| { |
| const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr()); |
| const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr()); |
| |
| const int8_t broadcast_value = *reinterpret_cast<const int8_t *>(broadcast_input.ptr()); |
| const float32x4x4_t broadcast_vector = vdequantize(vdupq_n_s8(broadcast_value), broadcast_qinfo); |
| |
| int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr, |
| voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2); |
| for(; x < window_end_x; ++x) |
| { |
| const float afs = dequantize_qasymm8_signed(*(non_broadcast_input_ptr + x), non_broadcast_qinfo); |
| const float bfs = dequantize_qasymm8_signed(broadcast_value, broadcast_qinfo); |
| *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? bfs : afs, !is_broadcast_input_2 ? afs : bfs, output_qinfo); |
| } |
| }, |
| broadcast_input, non_broadcast_input, output); |
| } |
| else |
| { |
| const UniformQuantizationInfo input1_qinfo = in1->info()->quantization_info().uniform(); |
| const UniformQuantizationInfo input2_qinfo = in2->info()->quantization_info().uniform(); |
| |
| // Input1 quantization info |
| const int32x4_t voffset1 = vdupq_n_s32(input1_qinfo.offset); |
| const float32x4_t vscale1 = vdupq_n_f32(input1_qinfo.scale); |
| |
| // Input2 quantization info |
| const int32x4_t voffset2 = vdupq_n_s32(input2_qinfo.offset); |
| const float32x4_t vscale2 = vdupq_n_f32(input2_qinfo.scale); |
| |
| // 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(in1, input1_win); |
| Iterator input2(in2, input2_win); |
| Iterator output(out, win); |
| |
| execute_window_loop(win, [&](const Coordinates &) |
| { |
| const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr()); |
| const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr()); |
| const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr()); |
| |
| int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2, |
| vscale1, vscale2, voffseto, invvscaleo); |
| for(; x < window_end_x; ++x) |
| { |
| const float afs = dequantize_qasymm8_signed(*(input1_ptr + x), input1_qinfo); |
| const float bfs = dequantize_qasymm8_signed(*(input2_ptr + x), input2_qinfo); |
| *(output_ptr + x) = (*scalar_func)(afs, bfs, output_qinfo); |
| } |
| }, |
| input1, input2, output); |
| } |
| } |
| |
| template <ArithmeticOperation op> |
| void elementwise_arithm_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| { |
| elementwise_op_quantized(in1, in2, out, window, &elementwise_arithm_op_quantized_scalar<op>, |
| &elementwise_arithm_op_quantized_broadcast_loop<op>, |
| &elementwise_arithm_op_quantized_loop<op>); |
| } |
| |
| template <ArithmeticOperation op> |
| void elementwise_arithm_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| { |
| elementwise_op_quantized_signed(in1, in2, out, window, &elementwise_arithm_op_quantized_signed_scalar<op>, |
| &elementwise_arithm_op_quantized_signed_broadcast_loop<op>, |
| &elementwise_arithm_op_quantized_singed_loop<op>); |
| } |
| |
| template <ComparisonOperation op> |
| void elementwise_comp_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| { |
| elementwise_op_quantized(in1, in2, out, window, &elementwise_comp_op_quantized_scalar<op>, |
| &elementwise_comp_op_quantized_broadcast_loop<op>, |
| &elementwise_comp_op_quantized_loop<op>); |
| } |
| |
| template <ComparisonOperation op> |
| void elementwise_comp_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| { |
| elementwise_comp_quantized_signed(in1, in2, out, window, &elementwise_comp_op_quantized_scalar<op>, |
| &elementwise_comp_op_quantized_signed_broadcast_loop<op>, |
| &elementwise_comp_op_quantized_signed_loop<op>); |
| } |
| } // namespace cpu |
| } // namespace arm_compute |
| |
| #endif /* SRC_CORE_NEON_KERNELS_ELEMENTWISE_IMPL_H */ |
