diff options
author | Michalis Spyrou <michalis.spyrou@arm.com> | 2020-12-08 21:02:16 +0000 |
---|---|---|
committer | Michalis Spyrou <michalis.spyrou@arm.com> | 2021-01-05 14:30:17 +0000 |
commit | a3c9a3b3d56f0369b199512fef832e6db958a601 (patch) | |
tree | 357bf1ea0c3ccf2ac314b0777036642a11b5f7cd /src/core/NEON/kernels/arithmetic_addition/impl | |
parent | b309fc249e4383b4d40ae03e377c3cbad3f9f5f7 (diff) | |
download | ComputeLibrary-a3c9a3b3d56f0369b199512fef832e6db958a601.tar.gz |
COMPMID-3874: Create ArithmeticAddition SVE/SVE2
Change-Id: I4ec7561a7f6a42a22b8187968ae302dbe75023bc
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4753
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arithmetic_addition/impl')
10 files changed, 1776 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/NEON/integer.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/integer.cpp new file mode 100644 index 0000000000..8dd58cec6d --- /dev/null +++ b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/integer.cpp @@ -0,0 +1,171 @@ +/* + * Copyright (c) 2020-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/wrapper.h" +#include "src/core/common/StdTypes.h" +#include "src/core/helpers/WindowHelpers.h" + +namespace arm_compute +{ +namespace cpu +{ +void arithmetic_addition_U8_U8_S16_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + // Create input windows + Window win = window; + 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 + 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(in1, input1_win); + Iterator input2(in2, input2_win); + Iterator output(out, win); + + 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()); + + 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 *>(output.ptr()); + + if(policy == ConvertPolicy::WRAP) + { + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x))); + const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x))); + wrapper::vstore(output_ptr + x, wrapper::vadd(vin1, vin2)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + *(output_ptr + x) = static_cast<int16_t>(*(input1_ptr + x)) + static_cast<int16_t>(*(input2_ptr + x)); + } + } + else + { + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x))); + const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x))); + wrapper::vstore(output_ptr + x, wrapper::vqadd(vin1, vin2)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + *(output_ptr + x) = wrapper::add_sat(static_cast<int16_t>(*(input1_ptr + x)), + static_cast<int16_t>(*(input2_ptr + x))); + } + } + }, + input1, input2, output); +} + +void arithmetic_addition_S16_U8_S16_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + // Create input windows + Window win = window; + 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 + 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(in1, input1_win); + Iterator input2(in2, input2_win); + Iterator output(out, win); + + 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()); + + 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 *>(output.ptr()); + + if(policy == ConvertPolicy::WRAP) + { + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin1 = wrapper::vloadq(input1_ptr + x); + const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x))); + wrapper::vstore(output_ptr + x, wrapper::vadd(vin1, vin2)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + *(output_ptr + x) = *(input1_ptr + x) + static_cast<int16_t>(*(input2_ptr + x)); + } + } + else + { + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin1 = wrapper::vloadq(input1_ptr + x); + const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x))); + wrapper::vstore(output_ptr + x, wrapper::vqadd(vin1, vin2)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + *(output_ptr + x) = wrapper::add_sat(*(input1_ptr + x), static_cast<int16_t>(*(input2_ptr + x))); + } + } + }, + input1, input2, output); +} + +void arithmetic_addition_U8_S16_S16_neon(const ITensor *input1, const ITensor *input2, ITensor *output, const ConvertPolicy &policy, const Window &window) +{ + // Simply swap the two input buffers: + arithmetic_addition_S16_U8_S16_neon(input2, input1, output, policy, window); +} +} // namespace cpu +} // namespace arm_compute
\ No newline at end of file diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/NEON/list.h b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/list.h new file mode 100644 index 0000000000..a8ab0910fd --- /dev/null +++ b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/list.h @@ -0,0 +1,146 @@ +/* + * Copyright (c) 2020-2021 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_ARITHMETIC_ADDITION_LIST_H +#define SRC_CORE_NEON_KERNELS_ARITHMETIC_ADDITION_LIST_H + +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/wrapper.h" + +namespace arm_compute +{ +namespace cpu +{ +#define DECLARE_ARITHMETIC_ADDITION_KERNEL(func_name) \ + void func_name(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) + +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qasymm8_neon); +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qasymm8_signed_neon); +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qsymm16_neon); +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_S16_U8_S16_neon); +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_U8_S16_S16_neon); +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_U8_U8_S16_neon); + +#undef DECLARE_ARITHMETIC_ADDITION_KERNEL + +template <typename ScalarType> +void arithmetic_addition_same_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + /** NEON vector tag type. */ + using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<ScalarType, wrapper::traits::BitWidth::W128>; + + // 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)); + + constexpr int window_step_x = 16 / sizeof(ScalarType); + 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 &) + { + const auto non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr()); + const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr()); + + const ScalarType broadcast_value = *reinterpret_cast<const ScalarType *>(broadcast_input.ptr()); + const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{}); + + // Compute S 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); + const auto res = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(broadcast_value_vec, non_broadcast_v) : wrapper::vadd(broadcast_value_vec, non_broadcast_v); + wrapper::vstore(output_ptr + x, res); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const auto non_broadcast_v = *(non_broadcast_input_ptr + x); + *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(broadcast_value, non_broadcast_v) : broadcast_value + non_broadcast_v; + } + }, + 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 &) + { + const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr()); + const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr()); + const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr()); + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto val1 = wrapper::vloadq(input1_ptr + x); + const auto val2 = wrapper::vloadq(input2_ptr + x); + const auto res = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(val1, val2) : wrapper::vadd(val1, val2); + wrapper::vstore(output_ptr + x, res); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const auto val1 = *(input1_ptr + x); + const auto val2 = *(input2_ptr + x); + *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(val1, val2) : val1 + val2; + } + }, + input1, input2, output); + } +} +} // namespace cpu +} // namespace arm_compute +#endif // SRC_CORE_NEON_KERNELS_ARITHMETIC_ADDITION_LIST_H
\ No newline at end of file diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8.cpp new file mode 100644 index 0000000000..b93dad20f1 --- /dev/null +++ b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8.cpp @@ -0,0 +1,210 @@ +/* + * Copyright (c) 2020-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/core/common/StdTypes.h" +#include "src/core/helpers/WindowHelpers.h" + +namespace arm_compute +{ +namespace cpu +{ +void arithmetic_addition_qasymm8_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + ARM_COMPUTE_UNUSED(policy); + + // 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 iq1_info = in1->info()->quantization_info().uniform(); + const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = out->info()->quantization_info().uniform(); + + const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale); + const float32x4_t voffseto = vdupq_n_f32(oq_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 ? 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 float32x4_t vscale1 = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale); + const float32x4_t vscale2 = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale); + const int32x4_t voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset); + const int32x4_t voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset); + + // 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 uint8x16_t broadcast_value_vec = vdupq_n_u8(broadcast_value); + + const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset2)), vscale2); + const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset2)), vscale2); + const auto bf_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset2)), vscale2); + const auto bf_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset2)), vscale2); + + const float bfs = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale; + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const uint8x16_t a = vld1q_u8(non_broadcast_input_ptr + x); + const auto af_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1); + const auto af_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1); + const auto af_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1); + const auto af_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1); + + int32x4_t rf_0{}; + int32x4_t rf_1{}; + int32x4_t rf_2{}; + int32x4_t rf_3{}; + +#ifdef __aarch64__ + rf_0 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo)); + rf_2 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo)); + rf_3 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo)); +#else //__aarch64__ + rf_0 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo)); + rf_2 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo)); + rf_3 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo)); +#endif //__aarch64__ + + const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1))); + const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3))); + vst1q_u8(output_ptr + x, vcombine_u8(pa, pb)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const float afs = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale; + *(output_ptr + x) = quantize_qasymm8((afs + bfs), oq_info); + } + }, + 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); + + const float32x4_t vscale1 = vdupq_n_f32(iq1_info.scale); + const float32x4_t vscale2 = vdupq_n_f32(iq2_info.scale); + const int32x4_t voffset1 = vdupq_n_s32(iq1_info.offset); + const int32x4_t voffset2 = vdupq_n_s32(iq2_info.offset); + + 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()); + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const uint8x16_t a = vld1q_u8(input1_ptr + x); + const uint8x16_t b = vld1q_u8(input2_ptr + x); + + const auto af_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1); + const auto af_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1); + const auto af_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1); + const auto af_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1); + + const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(b))))), voffset2)), vscale2); + const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(b))))), voffset2)), vscale2); + const auto bf_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(b))))), voffset2)), vscale2); + const auto bf_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(b))))), voffset2)), vscale2); + + int32x4_t rf_0{}; + int32x4_t rf_1{}; + int32x4_t rf_2{}; + int32x4_t rf_3{}; + +#ifdef __aarch64__ + rf_0 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo)); + rf_2 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo)); + rf_3 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo)); +#else //__aarch64__ + rf_0 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo)); + rf_2 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo)); + rf_3 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo)); +#endif //__aarch64__ + + const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1))); + const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3))); + vst1q_u8(output_ptr + x, vcombine_u8(pa, pb)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const float afs = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale; + const float bfs = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale; + *(output_ptr + x) = quantize_qasymm8((afs + bfs), out->info()->quantization_info()); + } + }, + input1, input2, output); + } +} +} // namespace cpu +} // namespace arm_compute
\ No newline at end of file diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8_signed.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8_signed.cpp new file mode 100644 index 0000000000..ba81cfcc03 --- /dev/null +++ b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8_signed.cpp @@ -0,0 +1,209 @@ +/* + * Copyright (c) 2020-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/core/common/StdTypes.h" +#include "src/core/helpers/WindowHelpers.h" + +namespace arm_compute +{ +namespace cpu +{ +void arithmetic_addition_qasymm8_signed_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + ARM_COMPUTE_UNUSED(policy); + + // 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 iq1_info = in1->info()->quantization_info().uniform(); + const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = out->info()->quantization_info().uniform(); + + const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale); + const float32x4_t voffseto = vdupq_n_f32(oq_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 ? 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 float32x4_t vscale1 = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale); + const float32x4_t vscale2 = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale); + const int32x4_t voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset); + const int32x4_t voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset); + + // 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 int8x16_t broadcast_value_vec = vdupq_n_s8(broadcast_value); + + const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(broadcast_value_vec)))), voffset2)), vscale2); + const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(broadcast_value_vec)))), voffset2)), vscale2); + const auto bf_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(broadcast_value_vec)))), voffset2)), vscale2); + const auto bf_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(broadcast_value_vec)))), voffset2)), vscale2); + const float bfs = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale; + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const int8x16_t a = vld1q_s8(non_broadcast_input_ptr + x); + + const auto af_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1); + const auto af_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1); + const auto af_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1); + const auto af_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1); + + int32x4_t rf_0{}; + int32x4_t rf_1{}; + int32x4_t rf_2{}; + int32x4_t rf_3{}; + +#ifdef __aarch64__ + rf_0 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo)); + rf_2 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo)); + rf_3 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo)); +#else //__aarch64__ + rf_0 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo)); + rf_2 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo)); + rf_3 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo)); +#endif //__aarch64__ + + const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1))); + const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3))); + vst1q_s8(output_ptr + x, vcombine_s8(pa, pb)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const float afs = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale; + *(output_ptr + x) = quantize_qasymm8_signed((afs + bfs), oq_info); + } + }, + 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); + + const float32x4_t vscale1 = vdupq_n_f32(iq1_info.scale); + const float32x4_t vscale2 = vdupq_n_f32(iq2_info.scale); + const int32x4_t voffset1 = vdupq_n_s32(iq1_info.offset); + const int32x4_t voffset2 = vdupq_n_s32(iq2_info.offset); + 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()); + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const int8x16_t a = vld1q_s8(input1_ptr + x); + const int8x16_t b = vld1q_s8(input2_ptr + x); + + const auto af_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1); + const auto af_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1); + const auto af_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1); + const auto af_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1); + + const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(b)))), voffset2)), vscale2); + const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(b)))), voffset2)), vscale2); + const auto bf_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(b)))), voffset2)), vscale2); + const auto bf_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(b)))), voffset2)), vscale2); + + int32x4_t rf_0{}; + int32x4_t rf_1{}; + int32x4_t rf_2{}; + int32x4_t rf_3{}; + +#ifdef __aarch64__ + rf_0 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo)); + rf_2 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo)); + rf_3 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo)); +#else //__aarch64__ + rf_0 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo)); + rf_2 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo)); + rf_3 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo)); +#endif //__aarch64__ + + const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1))); + const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3))); + vst1q_s8(output_ptr + x, vcombine_s8(pa, pb)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const float afs = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale; + const float bfs = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale; + *(output_ptr + x) = quantize_qasymm8_signed((afs + bfs), out->info()->quantization_info()); + } + }, + input1, input2, output); + } +} +} // namespace cpu +} // namespace arm_compute
\ No newline at end of file diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qsymm16.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qsymm16.cpp new file mode 100644 index 0000000000..538c600187 --- /dev/null +++ b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qsymm16.cpp @@ -0,0 +1,175 @@ +/* + * Copyright (c) 2020-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/core/common/StdTypes.h" +#include "src/core/helpers/WindowHelpers.h" + +namespace arm_compute +{ +namespace cpu +{ +void arithmetic_addition_qsymm16_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + ARM_COMPUTE_UNUSED(policy); + + // 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 = 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(); + + const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform(); + const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = out->info()->quantization_info().uniform(); + + const float32x4_t vscale1 = vdupq_n_f32(iq1_info.scale); + const float32x4_t vscale2 = vdupq_n_f32(iq2_info.scale); + const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale); + + 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; + 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 output(out, win); + + execute_window_loop(win, [&](const Coordinates &) + { + const auto non_broadcast_input_ptr = reinterpret_cast<const int16_t *>(non_broadcast_input.ptr()); + const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr()); + + const int16_t broadcast_value = *reinterpret_cast<const int16_t *>(broadcast_input.ptr()); + const int16x8_t broadcast_value_vec = vdupq_n_s16(broadcast_value); + + const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(broadcast_value_vec))), vscale2); + const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(broadcast_value_vec))), vscale2); + const float bfs = static_cast<int32_t>(broadcast_value) * broadcast_qinfo.scale; + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const int16x8_t a = vld1q_s16(non_broadcast_input_ptr + x); + const auto af_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1); + const auto af_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1); + + int32x4_t rf_0{}; + int32x4_t rf_1{}; +#ifdef __aarch64__ + rf_0 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); +#else //__aarch64__ + rf_0 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); +#endif //__aarch64__ + + const int16x8_t pa = vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)); + vst1q_s16(output_ptr + x, pa); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const float afs = static_cast<int32_t>(*(non_broadcast_input_ptr + x)) * non_broadcast_qinfo.scale; + *(output_ptr + x) = quantize_qsymm16((afs + bfs), oq_info); + } + }, + 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 &) + { + 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 *>(output.ptr()); + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const int16x8_t a = vld1q_s16(input1_ptr + x); + const int16x8_t b = vld1q_s16(input2_ptr + x); + + const auto af_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1); + const auto af_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1); + const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(b))), vscale2); + const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(b))), vscale2); + + int32x4_t rf_0{}; + int32x4_t rf_1{}; +#ifdef __aarch64__ + rf_0 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); +#else //__aarch64__ + rf_0 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); +#endif //__aarch64__ + + const int16x8_t pa = vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)); + vst1q_s16(output_ptr + x, pa); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const float afs = static_cast<int32_t>((*(input1_ptr + x))) * iq1_info.scale; + const float bfs = static_cast<int32_t>((*(input2_ptr + x))) * iq2_info.scale; + *(output_ptr + x) = quantize_qsymm16((afs + bfs), out->info()->quantization_info()); + } + }, + input1, input2, output); + } +} +} // namespace cpu +} // namespace arm_compute
\ No newline at end of file diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/SVE/integer.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/integer.cpp new file mode 100644 index 0000000000..c502a0235e --- /dev/null +++ b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/integer.cpp @@ -0,0 +1,201 @@ +/* + * Copyright (c) 2020-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#if defined(__ARM_FEATURE_SVE) +#include "src/core/NEON/SVEMath.h" +#include <arm_sve.h> + +namespace arm_compute +{ +namespace cpu +{ +void arithmetic_addition_U8_U8_S16_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + // Create input windows + Window win = window; + 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 + 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(in1, input1_win); + Iterator input2(in2, input2_win); + Iterator output(out, win); + + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + const auto all_true_pg = svptrue_b8(); + + 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 *>(output.ptr()); + + if(policy == ConvertPolicy::WRAP) + { + int x = window_start_x; + svbool_t pg_u = svwhilelt_b8(x, window_end_x); + svbool_t pg_0 = svwhilelt_b16(x, window_end_x); + svbool_t pg_1 = svwhilelt_b16(x, static_cast<int>(window_end_x + svcnth())); + do + { + const auto vin1 = svld1(pg_u, input1_ptr + x); + const auto vin2 = svld1(pg_u, input2_ptr + x); + + const auto vin1_lo = svreinterpret_s16_u16(svunpklo(vin1)); + const auto vin1_hi = svreinterpret_s16_u16(svunpkhi(vin1)); + const auto vin2_lo = svreinterpret_s16_u16(svunpklo(vin2)); + const auto vin2_hi = svreinterpret_s16_u16(svunpkhi(vin2)); + svst1(pg_0, output_ptr + x, svqadd(vin1_lo, vin2_lo)); + svst1(pg_1, output_ptr + x + svcnth(), svqadd(vin1_hi, vin2_hi)); + + x += svcntb(); + pg_u = svwhilelt_b8(x, window_end_x); + pg_0 = svwhilelt_b16(x, window_end_x); + pg_1 = svwhilelt_b16(x, static_cast<int>(window_end_x + svcnth())); + } + while(svptest_any(all_true_pg, pg_u)); + } + else + { + int x = window_start_x; + svbool_t pg_u = svwhilelt_b8(x, window_end_x); + svbool_t pg_0 = svwhilelt_b16(x, window_end_x); + svbool_t pg_1 = svwhilelt_b16(x, static_cast<int>(window_end_x + svcnth())); + do + { + const auto vin1 = svld1(pg_u, input1_ptr + x); + const auto vin2 = svld1(pg_u, input2_ptr + x); + + const auto vin1_lo = svreinterpret_s16_u16(svunpklo(vin1)); + const auto vin1_hi = svreinterpret_s16_u16(svunpkhi(vin1)); + const auto vin2_lo = svreinterpret_s16_u16(svunpklo(vin2)); + const auto vin2_hi = svreinterpret_s16_u16(svunpkhi(vin2)); + svst1(pg_0, output_ptr + x, svqadd(vin1_lo, vin2_lo)); + svst1(pg_1, output_ptr + x + svcnth(), svqadd(vin1_hi, vin2_hi)); + + x += svcntb(); + pg_u = svwhilelt_b8(x, window_end_x); + pg_0 = svwhilelt_b16(x, window_end_x); + pg_1 = svwhilelt_b16(x, static_cast<int>(window_end_x + svcnth())); + } + while(svptest_any(all_true_pg, pg_u)); + } + }, + input1, input2, output); +} + +void arithmetic_addition_S16_U8_S16_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + // Create input windows + Window win = window; + 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 + 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(in1, input1_win); + Iterator input2(in2, input2_win); + Iterator output(out, win); + + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + const auto all_true_pg = svptrue_b8(); + + 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 *>(output.ptr()); + + if(policy == ConvertPolicy::WRAP) + { + int x = window_start_x; + svbool_t pg_u = svwhilelt_b8(x, window_end_x); + svbool_t pg_0 = svwhilelt_b16(x, window_end_x); + svbool_t pg_1 = svwhilelt_b16(x + static_cast<int>(svcnth()), window_end_x); + do + { + const auto vin1_0 = svld1_s16(pg_0, input1_ptr + x); + const auto vin1_1 = svld1_s16(pg_1, input1_ptr + x + svcnth()); + const auto vin2_u8 = svld1_u8(pg_u, input2_ptr + x); + const auto vin2_0 = svreinterpret_s16_u16(svunpklo(vin2_u8)); + const auto vin2_1 = svreinterpret_s16_u16(svunpkhi(vin2_u8)); + svst1_s16(pg_0, output_ptr + x, svadd_s16_z(pg_0, vin1_0, vin2_0)); + svst1_s16(pg_1, output_ptr + x, svadd_s16_z(pg_1, vin1_1, vin2_1)); + + x += svcnth(); + pg_u = svwhilelt_b8(x, window_end_x); + pg_0 = svwhilelt_b16(x, window_end_x); + pg_1 = svwhilelt_b16(x + static_cast<int>(svcnth()), window_end_x); + } + while(svptest_any(all_true_pg, pg_u)); + } + else + { + int x = window_start_x; + svbool_t pg_u = svwhilelt_b8(x, window_end_x); + svbool_t pg_0 = svwhilelt_b16(x, window_end_x); + svbool_t pg_1 = svwhilelt_b16(x + static_cast<int>(svcnth()), window_end_x); + do + { + const auto vin1_0 = svld1_s16(pg_0, input1_ptr + x); + const auto vin1_1 = svld1_s16(pg_1, input1_ptr + x); + const auto vin2_u8 = svld1_u8(pg_u, input2_ptr + x); + const auto vin2_0 = svreinterpret_s16_u16(svunpklo(vin2_u8)); + const auto vin2_1 = svreinterpret_s16_u16(svunpkhi(vin2_u8)); + + svst1_s16(pg_0, output_ptr + x, svqadd(vin1_0, vin2_0)); + svst1_s16(pg_1, output_ptr + x, svqadd(vin1_1, vin2_1)); + + x += svcnth(); + pg_u = svwhilelt_b8(x, window_end_x); + pg_0 = svwhilelt_b16(x, window_end_x); + pg_1 = svwhilelt_b16(x + static_cast<int>(svcnth()), window_end_x); + } + while(svptest_any(all_true_pg, pg_u)); + } + }, + input1, input2, output); +} + +void arithmetic_addition_U8_S16_S16_sve(const ITensor *input1, const ITensor *input2, ITensor *output, const ConvertPolicy &policy, const Window &window) +{ + // Simply swap the two input buffers: + arithmetic_addition_S16_U8_S16_sve(input2, input1, output, policy, window); +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_SVE) */
\ No newline at end of file diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/SVE/list.h b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/list.h new file mode 100644 index 0000000000..3e238c40d0 --- /dev/null +++ b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/list.h @@ -0,0 +1,145 @@ +/* + * Copyright (c) 2020-2021 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_SVE_KERNELS_ARITHMETIC_ADDITION_LIST_H +#define SRC_CORE_SVE_KERNELS_ARITHMETIC_ADDITION_LIST_H + +#if defined(__ARM_FEATURE_SVE) +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/SVEMath.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include <arm_sve.h> + +namespace arm_compute +{ +namespace cpu +{ +#define DECLARE_ARITHMETIC_ADDITION_KERNEL(func_name) \ + void func_name(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) + +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qasymm8_sve); +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qasymm8_signed_sve); +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qsymm16_sve); +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_S16_U8_S16_sve); +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_U8_S16_S16_sve); +DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_U8_U8_S16_sve); + +#undef DECLARE_ARITHMETIC_ADDITION_KERNEL + +template <typename ScalarType> +void arithmetic_addition_same_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + const auto all_true_pg = wrapper::svptrue<ScalarType>(); + 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 bool is_sat = (policy == ConvertPolicy::SATURATE); + + // Clear X Dimension on execution window as we handle manually + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + // 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()); + + Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape())); + Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape())); + Iterator output(out, window); + + 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 &) + { + const auto non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr()); + const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr()); + + const ScalarType broadcast_value = *reinterpret_cast<const ScalarType *>(broadcast_input.ptr()); + const auto broadcast_value_vec = wrapper::svdup_n(broadcast_value); + + int x = window_start_x; + svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x); + do + { + const auto non_broadcast_v = svld1(pg, non_broadcast_input_ptr + x); + auto res = is_sat ? wrapper::svqadd(broadcast_value_vec, non_broadcast_v) : svadd_z(pg, broadcast_value_vec, non_broadcast_v); + svst1(pg, output_ptr + x, res); + + x += wrapper::svcnt<ScalarType>(); + pg = wrapper::svwhilelt<ScalarType>(x, window_end_x); + } + while(svptest_any(all_true_pg, pg)); + }, + 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 &) + { + const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr()); + const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr()); + const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr()); + + int x = window_start_x; + svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x); + do + { + const auto val1 = svld1(pg, input1_ptr + x); + const auto val2 = svld1(pg, input2_ptr + x); + const auto res = is_sat ? wrapper::svqadd(val1, val2) : svadd_z(pg, val1, val2); + svst1(pg, output_ptr + x, res); + + x += wrapper::svcnt<ScalarType>(); + pg = wrapper::svwhilelt<ScalarType>(x, window_end_x); + } + while(svptest_any(all_true_pg, pg)); + }, + input1, input2, output); + } +} +} // namespace cpu +} // namespace arm_compute +#endif // defined(__ARM_FEATURE_SVE) +#endif // SRC_CORE_SVE_KERNELS_ARITHMETIC_ADDITION_LIST_H
\ No newline at end of file diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8.cpp new file mode 100644 index 0000000000..871ee23ded --- /dev/null +++ b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8.cpp @@ -0,0 +1,182 @@ +/* + * Copyright (c) 2020-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#if defined(__ARM_FEATURE_SVE2) +#include "src/core/NEON/SVEMath.h" +#include <arm_sve.h> + +namespace arm_compute +{ +namespace cpu +{ +void arithmetic_addition_qasymm8_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + ARM_COMPUTE_UNUSED(policy); + + // 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 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 auto all_true_pg = svptrue_b8(); + + const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform(); + const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = out->info()->quantization_info().uniform(); + + const auto invvscaleo = svdup_n_f32(1.f / oq_info.scale); + const auto voffseto = svdup_n_f32(oq_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 ? in2 : in1; + const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; + + const svfloat32_t vscale1 = is_broadcast_input_2 ? svdup_n_f32(iq1_info.scale) : svdup_n_f32(iq2_info.scale); + const svfloat32_t vscale2 = is_broadcast_input_2 ? svdup_n_f32(iq2_info.scale) : svdup_n_f32(iq1_info.scale); + const svint32_t voffset1 = is_broadcast_input_2 ? svdup_n_s32(iq1_info.offset) : svdup_n_s32(iq2_info.offset); + const svint32_t voffset2 = is_broadcast_input_2 ? svdup_n_s32(iq2_info.offset) : svdup_n_s32(iq1_info.offset); + + // 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 svuint8_t broadcast_value_vec = svdup_n_u8(broadcast_value); + + int x = window_start_x; + svbool_t pg = svwhilelt_b8(x, window_end_x); + + const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(broadcast_value_vec))), voffset2)), vscale2); + const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(broadcast_value_vec))), voffset2)), vscale2); + const auto bf_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(broadcast_value_vec))), voffset2)), vscale2); + const auto bf_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(broadcast_value_vec))), voffset2)), vscale2); + + do + { + const svuint8_t a = svld1_u8(pg, non_broadcast_input_ptr + x); + + const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(a))), voffset1)), vscale1); + const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(a))), voffset1)), vscale1); + const auto af_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(a))), voffset1)), vscale1); + const auto af_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(a))), voffset1)), vscale1); + + const auto rf_0 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_0, bf_0), invvscaleo)); + const auto rf_1 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_1, bf_1), invvscaleo)); + const auto rf_2 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_2, bf_2), invvscaleo)); + const auto rf_3 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_3, bf_3), invvscaleo)); + + const auto pa = svqxtnt_u32(svqxtnb_u32(rf_0), rf_1); + const auto pb = svqxtnt_u32(svqxtnb_u32(rf_2), rf_3); + + const auto res = svqxtnt_u16(svqxtnb_u16(pa), pb); + svst1_u8(pg, output_ptr + x, res); + + x += svcntb(); + pg = svwhilelt_b8(x, window_end_x); + } + while(svptest_any(all_true_pg, pg)); + }, + 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); + + const auto vscale1 = svdup_n_f32(iq1_info.scale); + const auto vscale2 = svdup_n_f32(iq2_info.scale); + const auto voffset1 = svdup_n_s32(iq1_info.offset); + const auto voffset2 = svdup_n_s32(iq2_info.offset); + + 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 = window_start_x; + svbool_t pg = svwhilelt_b8(x, window_end_x); + do + { + const auto a = svld1_u8(pg, input1_ptr + x); + const auto b = svld1_u8(pg, input2_ptr + x); + const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(a))), voffset1)), vscale1); + const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(a))), voffset1)), vscale1); + const auto af_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(a))), voffset1)), vscale1); + const auto af_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(a))), voffset1)), vscale1); + + const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(b))), voffset2)), vscale2); + const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(b))), voffset2)), vscale2); + const auto bf_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(b))), voffset2)), vscale2); + const auto bf_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(b))), voffset2)), vscale2); + + const auto rf_0 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_0, bf_0), invvscaleo)); + const auto rf_1 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_1, bf_1), invvscaleo)); + const auto rf_2 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_2, bf_2), invvscaleo)); + const auto rf_3 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_3, bf_3), invvscaleo)); + + const auto pa = svqxtnt_u32(svqxtnb_u32(rf_0), rf_1); + const auto pb = svqxtnt_u32(svqxtnb_u32(rf_2), rf_3); + const auto res = svqxtnt_u16(svqxtnb_u16(pa), pb); + + svst1_u8(pg, output_ptr + x, res); + + x += svcntb(); + pg = svwhilelt_b8(x, window_end_x); + } + while(svptest_any(all_true_pg, pg)); + }, + input1, input2, output); + } +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_SVE2) */
\ No newline at end of file diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8_signed.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8_signed.cpp new file mode 100644 index 0000000000..2ba5d39400 --- /dev/null +++ b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8_signed.cpp @@ -0,0 +1,181 @@ +/* + * Copyright (c) 2020-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#if defined(__ARM_FEATURE_SVE2) +#include "src/core/NEON/SVEMath.h" +#include <arm_sve.h> + +namespace arm_compute +{ +namespace cpu +{ +void arithmetic_addition_qasymm8_signed_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + ARM_COMPUTE_UNUSED(policy); + + // 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 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 iq1_info = in1->info()->quantization_info().uniform(); + const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = out->info()->quantization_info().uniform(); + + const auto invvscaleo = svdup_n_f32(1.f / oq_info.scale); + const auto voffseto = svdup_n_f32(oq_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 ? in2 : in1; + const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; + const auto all_true_pg = svptrue_b8(); + + const auto vscale1 = is_broadcast_input_2 ? svdup_n_f32(iq1_info.scale) : svdup_n_f32(iq2_info.scale); + const auto vscale2 = is_broadcast_input_2 ? svdup_n_f32(iq2_info.scale) : svdup_n_f32(iq1_info.scale); + const auto voffset1 = is_broadcast_input_2 ? svdup_n_s32(iq1_info.offset) : svdup_n_s32(iq2_info.offset); + const auto voffset2 = is_broadcast_input_2 ? svdup_n_s32(iq2_info.offset) : svdup_n_s32(iq1_info.offset); + + // 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 auto broadcast_value_vec = svdup_n_s8(broadcast_value); + + int x = window_start_x; + svbool_t pg = svwhilelt_b8(x, window_end_x); + const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(broadcast_value_vec)), voffset2)), vscale2); + const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(broadcast_value_vec)), voffset2)), vscale2); + const auto bf_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(broadcast_value_vec)), voffset2)), vscale2); + const auto bf_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(broadcast_value_vec)), voffset2)), vscale2); + + do + { + const auto a = svld1_s8(pg, non_broadcast_input_ptr + x); + const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(a)), voffset1)), vscale1); + const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(a)), voffset1)), vscale1); + const auto af_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(a)), voffset1)), vscale1); + const auto af_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(a)), voffset1)), vscale1); + + const auto rf_0 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_0, bf_0), invvscaleo)); + const auto rf_1 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_1, bf_1), invvscaleo)); + const auto rf_2 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_2, bf_2), invvscaleo)); + const auto rf_3 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_3, bf_3), invvscaleo)); + + const auto pa = svqxtnt_s32(svqxtnb_s32(rf_0), rf_1); + const auto pb = svqxtnt_s32(svqxtnb_s32(rf_2), rf_3); + const auto res = svqxtnt_s16(svqxtnb_s16(pa), pb); + + svst1_s8(pg, output_ptr + x, res); + + x += svcntb(); + pg = svwhilelt_b8(x, window_end_x); + } + while(svptest_any(all_true_pg, pg)); + }, + 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); + + const auto vscale1 = svdup_n_f32(iq1_info.scale); + const auto vscale2 = svdup_n_f32(iq2_info.scale); + const auto voffset1 = svdup_n_s32(iq1_info.offset); + const auto voffset2 = svdup_n_s32(iq2_info.offset); + + 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 = window_start_x; + svbool_t pg = svwhilelt_b8(x, window_end_x); + do + { + const auto a = svld1_s8(pg, input1_ptr + x); + const auto b = svld1_s8(pg, input2_ptr + x); + + const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(a)), voffset1)), vscale1); + const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(a)), voffset1)), vscale1); + const auto af_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(a)), voffset1)), vscale1); + const auto af_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(a)), voffset1)), vscale1); + + const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(b)), voffset2)), vscale2); + const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(b)), voffset2)), vscale2); + const auto bf_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(b)), voffset2)), vscale2); + const auto bf_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(b)), voffset2)), vscale2); + + const auto rf_0 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_0, bf_0), invvscaleo)); + const auto rf_1 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_1, bf_1), invvscaleo)); + const auto rf_2 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_2, bf_2), invvscaleo)); + const auto rf_3 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_3, bf_3), invvscaleo)); + + const auto pa = svqxtnt_s32(svqxtnb_s32(rf_0), rf_1); + const auto pb = svqxtnt_s32(svqxtnb_s32(rf_2), rf_3); + const auto res = svqxtnt_s16(svqxtnb_s16(pa), pb); + + svst1_s8(pg, output_ptr + x, res); + + x += svcntb(); + pg = svwhilelt_b8(x, window_end_x); + } + while(svptest_any(svptrue_b8(), pg)); + }, + input1, input2, output); + } +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_SVE2) */
\ No newline at end of file diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qsymm16.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qsymm16.cpp new file mode 100644 index 0000000000..c072cdb249 --- /dev/null +++ b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qsymm16.cpp @@ -0,0 +1,156 @@ +/* + * Copyright (c) 2020-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#if defined(__ARM_FEATURE_SVE2) +#include "src/core/NEON/SVEMath.h" +#include <arm_sve.h> + +namespace arm_compute +{ +namespace cpu +{ +void arithmetic_addition_qsymm16_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window) +{ + ARM_COMPUTE_UNUSED(policy); + + // 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 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 iq1_info = in1->info()->quantization_info().uniform(); + const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = out->info()->quantization_info().uniform(); + + const auto vscale1 = svdup_n_f32(iq1_info.scale); + const auto vscale2 = svdup_n_f32(iq2_info.scale); + const auto invvscaleo = svdup_n_f32(1.f / oq_info.scale); + const auto all_true_pg = svptrue_b16(); + + 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 &) + { + const auto non_broadcast_input_ptr = reinterpret_cast<const int16_t *>(non_broadcast_input.ptr()); + const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr()); + + const int16_t broadcast_value = *reinterpret_cast<const int16_t *>(broadcast_input.ptr()); + const auto broadcast_value_vec = svdup_n_s16(broadcast_value); + + int x = window_start_x; + svbool_t pg = svwhilelt_b16(x, window_end_x); + + const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(broadcast_value_vec)), vscale2); + const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(broadcast_value_vec)), vscale2); + + do + { + const auto a = svld1_s16(pg, non_broadcast_input_ptr + x); + const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(a)), vscale1); + const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(a)), vscale1); + + const auto rf_0 = svcvt_s32_f32_z(pg, svmul_f32_z(pg, svadd_f32_z(pg, af_0, bf_0), invvscaleo)); + const auto rf_1 = svcvt_s32_f32_z(pg, svmul_f32_z(pg, svadd_f32_z(pg, af_1, bf_1), invvscaleo)); + + const auto res = svqxtnt_s32(svqxtnb_s32(rf_0), rf_1); + + svst1_s16(pg, output_ptr + x, res); + + x += svcnth(); + pg = svwhilelt_b16(x, window_end_x); + } + while(svptest_any(all_true_pg, pg)); + }, + 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 &) + { + 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 *>(output.ptr()); + + int x = window_start_x; + svbool_t pg = svwhilelt_b16(x, window_end_x); + do + { + auto a = svld1_s16(pg, input1_ptr + x); + auto b = svld1_s16(pg, input2_ptr + x); + + const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(a)), vscale1); + const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(a)), vscale1); + + const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(b)), vscale2); + const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(b)), vscale2); + + const auto rf_0 = svcvt_s32_f32_z(pg, svmul_f32_z(pg, svadd_f32_z(pg, af_0, bf_0), invvscaleo)); + const auto rf_1 = svcvt_s32_f32_z(pg, svmul_f32_z(pg, svadd_f32_z(pg, af_1, bf_1), invvscaleo)); + + const auto res = svqxtnt_s32(svqxtnb_s32(rf_0), rf_1); + svst1_s16(pg, output_ptr + x, res); + + x += svcnth(); + pg = svwhilelt_b16(x, window_end_x); + } + while(svptest_any(all_true_pg, pg)); + }, + input1, input2, output); + } +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_SVE2) */
\ No newline at end of file |