diff options
Diffstat (limited to 'src/cpu/kernels/activation/generic/neon')
| -rw-r--r-- | src/cpu/kernels/activation/generic/neon/fp16.cpp | 43 | ||||
| -rw-r--r-- | src/cpu/kernels/activation/generic/neon/fp32.cpp | 39 | ||||
| -rw-r--r-- | src/cpu/kernels/activation/generic/neon/impl.h | 246 | ||||
| -rw-r--r-- | src/cpu/kernels/activation/generic/neon/lut.cpp | 57 | ||||
| -rw-r--r-- | src/cpu/kernels/activation/generic/neon/qasymm8.cpp | 310 | ||||
| -rw-r--r-- | src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp | 274 | ||||
| -rw-r--r-- | src/cpu/kernels/activation/generic/neon/qsymm16.cpp | 156 |
7 files changed, 1125 insertions, 0 deletions
diff --git a/src/cpu/kernels/activation/generic/neon/fp16.cpp b/src/cpu/kernels/activation/generic/neon/fp16.cpp new file mode 100644 index 0000000000..ddc6dc24cd --- /dev/null +++ b/src/cpu/kernels/activation/generic/neon/fp16.cpp @@ -0,0 +1,43 @@ +/* + * Copyright (c) 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. + */ +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) + +#include "src/cpu/kernels/activation/generic/neon/impl.h" + +namespace arm_compute +{ +namespace cpu +{ +namespace +{ +constexpr ActFpImplParams Fp16Params = {static_cast<float16_t>(1e-7), 8}; +} // namespace + +void neon_fp16_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + fp_neon_activation_impl<float16_t, Fp16Params>(src, dst, act_info, window); +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/cpu/kernels/activation/generic/neon/fp32.cpp b/src/cpu/kernels/activation/generic/neon/fp32.cpp new file mode 100644 index 0000000000..e558f8c73e --- /dev/null +++ b/src/cpu/kernels/activation/generic/neon/fp32.cpp @@ -0,0 +1,39 @@ +/* + * Copyright (c) 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 "src/cpu/kernels/activation/generic/neon/impl.h" +namespace arm_compute +{ +namespace cpu +{ +namespace +{ +constexpr ActFpImplParams Fp32Params = {static_cast<float>(1e-24), 4}; +} // namespace +void neon_fp32_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + fp_neon_activation_impl<float, Fp32Params>(src, dst, act_info, window); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/activation/generic/neon/impl.h b/src/cpu/kernels/activation/generic/neon/impl.h new file mode 100644 index 0000000000..afeb6f7f3d --- /dev/null +++ b/src/cpu/kernels/activation/generic/neon/impl.h @@ -0,0 +1,246 @@ +/* + * Copyright (c) 2020-2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/function_info/ActivationLayerInfo.h" + +#include "src/core/NEON/wrapper/wrapper.h" +namespace arm_compute +{ +namespace cpu +{ +/** Constant parameters needed by the activation implementation. + * These parameters differ for each floating type + * + * @note This are passed as a struct as C++ does not allow float as a template parameter until C++20 + **/ +struct ActFpImplParams +{ + float delta; /**< Minimum delta needed to avoid NaN on corner-cases of elementary functions */ + int step_x; /**< Window step at the x dimension */ +}; + +#ifndef __aarch64__ +inline float32x4_t mask_float_vector(const float32x4_t &in, const uint32x4_t &mask) +{ + auto int_in = vreinterpretq_u32_f32(in); + return vreinterpretq_f32_u32(wrapper::vand(int_in, mask)); +} +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +inline float16x8_t mask_float_vector(const float16x8_t &in, const uint16x8_t &mask) +{ + auto int_in = vreinterpretq_u16_f16(in); + return vreinterpretq_f16_u16(wrapper::vand(int_in, mask)); +} +#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +#endif /* __aarch64__ */ + +template <typename T, const ActFpImplParams &P> +void fp_neon_activation_impl(const ITensor *src, + ITensor *dst, + const ActivationLayerInfo &act_info, + const Window &window) +{ + /** SIMD vector tag type. */ + using ExactTagType = + typename arm_compute::wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>; + constexpr int window_step_x = P.step_x; + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + // In case of non-aarch64, a small delta value is added to the input + // to prevent NAN values caused by zeros in inputs to SQRT. + // In case of aarh64, we call vsqrt directly, so we don't use delta. +#ifndef __aarch64__ + const auto delta = wrapper::vdup_n(static_cast<T>(P.delta), ExactTagType{}); +#else /* #ifndef __aarch64__ */ + const auto const_inv_2 = wrapper::vdup_n(static_cast<T>(0.5f), ExactTagType{}); + const auto const_inv_sqrt_2 = wrapper::vdup_n(static_cast<T>(0.70710678118f), ExactTagType{}); +#endif /* __aarch64__ */ + const auto const_1 = wrapper::vdup_n(static_cast<T>(1.f), ExactTagType{}); + const auto const_0 = wrapper::vdup_n(static_cast<T>(0.f), ExactTagType{}); + const auto const_6 = wrapper::vdup_n(static_cast<T>(6.f), ExactTagType{}); + const auto const_3 = wrapper::vdup_n(static_cast<T>(3.f), ExactTagType{}); + const auto const_inv_6 = wrapper::vdup_n(static_cast<T>(0.166666667f), ExactTagType{}); + constexpr float soft_relu_thresh = 12.f; + const auto vsoft_relu_thresh = wrapper::vdup_n(static_cast<T>(soft_relu_thresh), ExactTagType{}); + const auto va = wrapper::vdup_n(static_cast<T>(act_info.a()), ExactTagType{}); + const auto vb = wrapper::vdup_n(static_cast<T>(act_info.b()), ExactTagType{}); + const auto a = static_cast<T>(act_info.a()); + const auto b = static_cast<T>(act_info.b()); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const T *>(input.ptr()); + const auto output_ptr = reinterpret_cast<T *>(output.ptr()); + wrapper::traits::neon_bitvector_t<T, wrapper::traits::BitWidth::W128> tmp; + // Compute S elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + switch (act) + { + case ActivationLayerInfo::ActivationFunction::ABS: + tmp = wrapper::vabs(vin); + break; + case ActivationLayerInfo::ActivationFunction::LINEAR: + tmp = wrapper::vmla(vb, va, vin); + break; + case ActivationLayerInfo::ActivationFunction::LOGISTIC: + tmp = wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin)))); + break; + case ActivationLayerInfo::ActivationFunction::RELU: + tmp = wrapper::vmax(const_0, vin); + break; + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: + tmp = wrapper::vmin(va, wrapper::vmax(vb, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: + tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin)); + break; + case ActivationLayerInfo::ActivationFunction::SOFT_RELU: + tmp = wrapper::vbsl(wrapper::vcgt(vin, vsoft_relu_thresh), vin, + wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin)))); + break; + case ActivationLayerInfo::ActivationFunction::ELU: + tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, + wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1))); + break; + case ActivationLayerInfo::ActivationFunction::SQRT: +#ifdef __aarch64__ + tmp = wrapper::vsqrt(vin); +#else /* __aarch64__ */ + { + const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0.f, ExactTagType{})); + tmp = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask)))); + tmp = mask_float_vector(tmp, wrapper::vnot(bitmask)); + } +#endif /* __aarch64__ */ + break; + case ActivationLayerInfo::ActivationFunction::SQUARE: + tmp = wrapper::vmul(vin, vin); + break; + case ActivationLayerInfo::ActivationFunction::TANH: + tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin))); + break; + case ActivationLayerInfo::ActivationFunction::IDENTITY: + tmp = vin; + break; + case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + tmp = wrapper::vmul( + vin, + wrapper::vmul(const_inv_6, + wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3))))); + break; + case ActivationLayerInfo::ActivationFunction::SWISH: + tmp = wrapper::vmul(vin, wrapper::vinv(wrapper::vadd( + const_1, wrapper::vexpq(wrapper::vneg(wrapper::vmul(va, vin)))))); + break; +#ifdef __aarch64__ + case ActivationLayerInfo::ActivationFunction::GELU: + tmp = wrapper::vmul( + vin, + wrapper::vmul(const_inv_2, + wrapper::vadd(const_1, wrapper::verf(wrapper::vmul(vin, const_inv_sqrt_2))))); + break; +#endif /* __aarch64__ */ + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); + } + // Compute left-over elements + for (; x < window_end_x; ++x) + { + const T in = *(reinterpret_cast<const T *>(input_ptr + x)); + T tmp; + switch (act) + { + case ActivationLayerInfo::ActivationFunction::ABS: + tmp = std::abs(in); + break; + case ActivationLayerInfo::ActivationFunction::LINEAR: + tmp = a * in + b; + break; + case ActivationLayerInfo::ActivationFunction::LOGISTIC: + tmp = static_cast<T>(1) / (static_cast<T>(1) + std::exp(-in)); + break; + case ActivationLayerInfo::ActivationFunction::RELU: + tmp = std::max<T>(static_cast<T>(0), in); + break; + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + tmp = std::min<T>(a, std::max(static_cast<T>(0), in)); + break; + case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: + tmp = std::min<T>(a, std::max<T>(b, in)); + break; + case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: + tmp = (in > 0) ? in : a * in; + break; + case ActivationLayerInfo::ActivationFunction::SOFT_RELU: + tmp = (in > soft_relu_thresh) ? in : std::log(static_cast<T>(1) + std::exp(in)); + break; + case ActivationLayerInfo::ActivationFunction::ELU: + tmp = (in >= 0) ? in : a * (std::exp(in) - 1); + break; + case ActivationLayerInfo::ActivationFunction::SQRT: + tmp = std::sqrt(in); + break; + case ActivationLayerInfo::ActivationFunction::SQUARE: + tmp = in * in; + break; + case ActivationLayerInfo::ActivationFunction::TANH: + tmp = a * std::tanh(b * in); + break; + case ActivationLayerInfo::ActivationFunction::IDENTITY: + tmp = in; + break; + case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + tmp = in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f); + break; + case ActivationLayerInfo::ActivationFunction::SWISH: + tmp = in / (static_cast<T>(1) + std::exp(-a * in)); + break; + case ActivationLayerInfo::ActivationFunction::GELU: + tmp = in * static_cast<T>(0.5f * (1.0f + erff(static_cast<float>(in) / 1.41421356237f))); + break; + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + *(output_ptr + x) = tmp; + } + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/activation/generic/neon/lut.cpp b/src/cpu/kernels/activation/generic/neon/lut.cpp new file mode 100644 index 0000000000..ddd186f9cb --- /dev/null +++ b/src/cpu/kernels/activation/generic/neon/lut.cpp @@ -0,0 +1,57 @@ +/* + * Copyright (c) 2022-2024 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/function_info/ActivationLayerInfo.h" + +#include "src/cpu/kernels/lut/list.h" + +namespace arm_compute +{ +namespace cpu +{ +#ifdef __aarch64__ +void neon_q8_activation_lut(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + ARM_COMPUTE_ERROR_ON( // LUT does not provide any performance benefit for ReLU as it's a single max() operation + (src->info()->data_type() != DataType::QASYMM8 && src->info()->data_type() != DataType::QASYMM8_SIGNED) || + act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU); + const auto window_end_x = window.x().end(); + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const uint8_t *>(input.ptr()); + auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr()); + lut_u8_neon(act_info.lut().data(), 1u, window_end_x, &input_ptr, &output_ptr); + }, + input, output); +} +#endif // __aarch64__ +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/activation/generic/neon/qasymm8.cpp b/src/cpu/kernels/activation/generic/neon/qasymm8.cpp new file mode 100644 index 0000000000..1451301ea2 --- /dev/null +++ b/src/cpu/kernels/activation/generic/neon/qasymm8.cpp @@ -0,0 +1,310 @@ +/* + * Copyright (c) 2020-2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/function_info/ActivationLayerInfo.h" + +#include "src/core/NEON/NEAsymm.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include <arm_neon.h> +#include <cmath> +#include <cstddef> +#include <cstdint> + +namespace arm_compute +{ +namespace cpu +{ +void neon_qasymm8_activation(const ITensor *src, + ITensor *dst, + const ActivationLayerInfo &act_info, + const Window &window) +{ + constexpr 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 ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); + const qasymm8x16_t va = vdupq_n_u8(quantize_qasymm8(act_info.a(), qi_in)); + const qasymm8x16_t vb = vdupq_n_u8(quantize_qasymm8(act_info.b(), qi_in)); + const qasymm8_t a = quantize_qasymm8(act_info.a(), qi_in); + const qasymm8_t b = quantize_qasymm8(act_info.b(), qi_in); + const qasymm8_t const_0 = quantize_qasymm8(0.f, qi_in); + const qasymm8x16_t vconst_0 = vdupq_n_u8(const_0); + const auto vconst_1 = vdupq_n_f32(1.f); + +#ifndef __aarch64__ + const auto vconst_0_f32 = vdupq_n_f32(0); +#else // #ifndef __aarch64__ + const auto const_inv_2 = vdupq_n_f32(0.5f); + const auto const_inv_sqrt_2 = vdupq_n_f32(0.70710678118f); +#endif // __aarch64__ + const float32x4_t va_f32 = vdupq_n_f32(act_info.a()); + const float32x4_t vb_f32 = vdupq_n_f32(act_info.b()); + const float a_f32 = act_info.a(); + const float b_f32 = act_info.b(); + +#ifndef __aarch64__ + const auto const_6_f32 = vdupq_n_f32(6.f); + const auto const_0_f32 = vdupq_n_f32(0.f); + const auto const_3_f32 = vdupq_n_f32(3.f); + const auto const_inv_6_f32 = vdupq_n_f32(0.166666667f); +#endif // __aarch64__ + + // Initialise scale/offset for re-quantization + float s = qi_in.scale / qi_out.scale; + float o = -qi_in.offset * s + qi_out.offset; + float32x4_t vs = vdupq_n_f32(s); + float32x4_t vo = vdupq_n_f32(o); + + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const qasymm8_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<qasymm8_t *>(output.ptr()); + + wrapper::traits::neon_bitvector_t<qasymm8_t, wrapper::traits::BitWidth::W128> tmp; + + // Compute S elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + if (act == ActivationLayerInfo::ActivationFunction::RELU) + { + // Perform activation + tmp = vmaxq_u8(vconst_0, vin); + // Re-quantize to new output space + tmp = vmlaq_qasymm8<RoundingPolicy::TO_NEAREST_UP>(tmp, vs, vo); + } + else if (act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + // Perform activation + tmp = vminq_u8(va, vmaxq_u8(vconst_0, vin)); + // Re-quantize to new output space + tmp = vmlaq_qasymm8<RoundingPolicy::TO_NEAREST_UP>(tmp, vs, vo); + } + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + // Perform activation + tmp = vminq_u8(va, vmaxq_u8(vb, vin)); + // Re-quantize to new output space + tmp = vmlaq_qasymm8<RoundingPolicy::TO_NEAREST_UP>(tmp, vs, vo); + } +#ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead. + else if (act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = {{ + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[2])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[3])))), + }}; + // Re-quantize to new output space + tmp = vquantize(tmp_dep, qi_out); + } +#endif // __aarch64__ + else if (act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = {{ + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[2], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[3], vb_f32))), + }}; + // Re-quantize to new output space + tmp = vquantize(tmp_dep, qi_out); + } +#ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead. + else if (act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = {{ + wrapper::vmul( + vin_deq.val[0], + wrapper::vmul( + const_inv_6_f32, + wrapper::vmin(const_6_f32, + wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[0], const_3_f32))))), + wrapper::vmul( + vin_deq.val[1], + wrapper::vmul( + const_inv_6_f32, + wrapper::vmin(const_6_f32, + wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[1], const_3_f32))))), + wrapper::vmul( + vin_deq.val[2], + wrapper::vmul( + const_inv_6_f32, + wrapper::vmin(const_6_f32, + wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[2], const_3_f32))))), + wrapper::vmul( + vin_deq.val[3], + wrapper::vmul( + const_inv_6_f32, + wrapper::vmin(const_6_f32, + wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[3], const_3_f32))))), + }}; + // Re-quantize to new output space + tmp = vquantize(tmp_dep, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + const auto vin_deq = vdequantize(vin, qi_in); + + const uint32x4x4_t pos_mask = {{ + wrapper::vcgt(vin_deq.val[0], vconst_0_f32), + wrapper::vcgt(vin_deq.val[1], vconst_0_f32), + wrapper::vcgt(vin_deq.val[2], vconst_0_f32), + wrapper::vcgt(vin_deq.val[3], vconst_0_f32), + }}; + + const float32x4x4_t tmp_dep = {{ + wrapper::vbsl(pos_mask.val[0], vin_deq.val[0], wrapper::vmul(va_f32, vin_deq.val[0])), + wrapper::vbsl(pos_mask.val[1], vin_deq.val[1], wrapper::vmul(va_f32, vin_deq.val[1])), + wrapper::vbsl(pos_mask.val[2], vin_deq.val[2], wrapper::vmul(va_f32, vin_deq.val[2])), + wrapper::vbsl(pos_mask.val[3], vin_deq.val[3], wrapper::vmul(va_f32, vin_deq.val[3])), + }}; + + tmp = vquantize(tmp_dep, qi_out); + } +#else // #ifndef __aarch64__ + else if (act == ActivationLayerInfo::ActivationFunction::GELU) + { + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = {{ + wrapper::vmul(vin_deq.val[0], + wrapper::vmul(const_inv_2, + wrapper::vadd(vconst_1, wrapper::verf(wrapper::vmul( + vin_deq.val[0], const_inv_sqrt_2))))), + wrapper::vmul(vin_deq.val[1], + wrapper::vmul(const_inv_2, + wrapper::vadd(vconst_1, wrapper::verf(wrapper::vmul( + vin_deq.val[1], const_inv_sqrt_2))))), + wrapper::vmul(vin_deq.val[2], + wrapper::vmul(const_inv_2, + wrapper::vadd(vconst_1, wrapper::verf(wrapper::vmul( + vin_deq.val[2], const_inv_sqrt_2))))), + wrapper::vmul(vin_deq.val[3], + wrapper::vmul(const_inv_2, + wrapper::vadd(vconst_1, wrapper::verf(wrapper::vmul( + vin_deq.val[3], const_inv_sqrt_2))))), + }}; + // Re-quantize to new output space + tmp = vquantize(tmp_dep, qi_out); + } +#endif // __aarch64__ + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + qasymm8_t in = *(reinterpret_cast<const qasymm8_t *>(input_ptr + x)); + qasymm8_t tmp = 0; + if (act == ActivationLayerInfo::ActivationFunction::RELU) + { + tmp = std::max(const_0, in); + tmp = utility::clamp<int32_t, qasymm8_t>(support::cpp11::lround(tmp * s + o)); + } + else if (act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + tmp = std::min(a, std::max(const_0, in)); + tmp = utility::clamp<int32_t, qasymm8_t>(support::cpp11::lround(tmp * s + o)); + } + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + tmp = std::min(a, std::max(b, in)); + tmp = utility::clamp<int32_t, qasymm8_t>(support::cpp11::lround(tmp * s + o)); + } +#ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead. + else if (act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + float tmp_f = dequantize_qasymm8(in, qi_in); + tmp_f = 1.f / (1.f + std::exp(-tmp_f)); + tmp = quantize_qasymm8(tmp_f, qi_out); + } +#endif // __aarch64__ + else if (act == ActivationLayerInfo::ActivationFunction::TANH) + { + float tmp_f = dequantize_qasymm8(in, qi_in); + tmp_f = a_f32 * std::tanh(b_f32 * tmp_f); + tmp = quantize_qasymm8(tmp_f, qi_out); + } +#ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead. + else if (act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) + { + float tmp_f = dequantize_qasymm8(in, qi_in); + tmp_f = tmp_f * ((std::min(std::max((tmp_f + 3), 0.0f), 6.0f)) * 0.166666667f); + tmp = quantize_qasymm8(tmp_f, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + float tmp_f = dequantize_qasymm8(in, qi_in); + tmp_f = tmp_f > 0 ? tmp_f : tmp_f * a_f32; + tmp = quantize_qasymm8(tmp_f, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::GELU) + { + float tmp_f = dequantize_qasymm8(in, qi_in); + tmp = tmp_f * 0.5f * (1.0f + std::erff(in / 1.41421356237f)); + tmp = quantize_qasymm8(tmp_f, qi_out); + } +#endif // __aarch64__ + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + *(output_ptr + x) = tmp; + } + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp b/src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp new file mode 100644 index 0000000000..a2f588245a --- /dev/null +++ b/src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp @@ -0,0 +1,274 @@ +/* + * Copyright (c) 2020-2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/function_info/ActivationLayerInfo.h" + +#include "src/core/NEON/NEAsymm.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include <arm_neon.h> +#include <cmath> +#include <cstddef> + +namespace arm_compute +{ +namespace cpu +{ +void neon_qasymm8_signed_activation(const ITensor *src, + ITensor *dst, + const ActivationLayerInfo &act_info, + const Window &window) +{ + constexpr 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 ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); + const qasymm8x16_signed_t va = vdupq_n_s8(quantize_qasymm8_signed(act_info.a(), qi_in)); + const qasymm8x16_signed_t vb = vdupq_n_s8(quantize_qasymm8_signed(act_info.b(), qi_in)); + const qasymm8_signed_t a = quantize_qasymm8_signed(act_info.a(), qi_in); + const qasymm8_signed_t b = quantize_qasymm8_signed(act_info.b(), qi_in); + const qasymm8_signed_t const_0 = quantize_qasymm8_signed(0.f, qi_in); + const qasymm8x16_signed_t vconst_0 = vdupq_n_s8(const_0); +#ifndef __aarch64__ + const auto vconst_1 = vdupq_n_f32(1.f); + const auto vconst_0_f32 = vdupq_n_f32(0.f); +#endif // __aarch64__ + const float32x4_t va_f32 = vdupq_n_f32(act_info.a()); + const float32x4_t vb_f32 = vdupq_n_f32(act_info.b()); + const float a_f32 = act_info.a(); + const float b_f32 = act_info.b(); + const auto const_6_f32 = vdupq_n_f32(6.f); + const auto const_0_f32 = vdupq_n_f32(0.f); + const auto const_3_f32 = vdupq_n_f32(3.f); + const auto const_inv_6_f32 = vdupq_n_f32(0.166666667f); + + // Initialise scale/offset for re-quantization + float s = qi_in.scale / qi_out.scale; + float o = -qi_in.offset * s + qi_out.offset; + float32x4_t vs = vdupq_n_f32(s); + float32x4_t vo = vdupq_n_f32(o); + + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const qasymm8_signed_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<qasymm8_signed_t *>(output.ptr()); + + wrapper::traits::neon_bitvector_t<qasymm8_signed_t, wrapper::traits::BitWidth::W128> tmp; + + // Compute S elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + if (act == ActivationLayerInfo::ActivationFunction::RELU) + { + // Perform activation + tmp = vmaxq_s8(vconst_0, vin); + // Re-quantize to new output space + tmp = vmlaq_qasymm8_signed<RoundingPolicy::TO_NEAREST_UP>(tmp, vs, vo); + } + else if (act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + // Perform activation + tmp = vminq_s8(va, vmaxq_s8(vconst_0, vin)); + // Re-quantize to new output space + tmp = vmlaq_qasymm8_signed<RoundingPolicy::TO_NEAREST_UP>(tmp, vs, vo); + } + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + // Perform activation + tmp = vminq_s8(va, vmaxq_s8(vb, vin)); + // Re-quantize to new output space + tmp = vmlaq_qasymm8_signed<RoundingPolicy::TO_NEAREST_UP>(tmp, vs, vo); + } +#ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead. + else if (act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = {{ + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[2])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[3])))), + }}; + // Re-quantize to new output space + tmp = vquantize_signed(tmp_dep, qi_out); + } +#endif // __aarch64__ + else if (act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = {{ + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[2], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[3], vb_f32))), + }}; + // Re-quantize to new output space + tmp = vquantize_signed(tmp_dep, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = {{ + wrapper::vmul( + vin_deq.val[0], + wrapper::vmul( + const_inv_6_f32, + wrapper::vmin(const_6_f32, + wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[0], const_3_f32))))), + wrapper::vmul( + vin_deq.val[1], + wrapper::vmul( + const_inv_6_f32, + wrapper::vmin(const_6_f32, + wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[1], const_3_f32))))), + wrapper::vmul( + vin_deq.val[2], + wrapper::vmul( + const_inv_6_f32, + wrapper::vmin(const_6_f32, + wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[2], const_3_f32))))), + wrapper::vmul( + vin_deq.val[3], + wrapper::vmul( + const_inv_6_f32, + wrapper::vmin(const_6_f32, + wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[3], const_3_f32))))), + }}; + // Re-quantize to new output space + tmp = vquantize_signed(tmp_dep, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + const auto vin_deq = vdequantize(vin, qi_in); + +#ifdef __aarch64__ + const uint32x4x4_t pos_mask = {{ + wrapper::vcgtz(vin_deq.val[0]), + wrapper::vcgtz(vin_deq.val[1]), + wrapper::vcgtz(vin_deq.val[2]), + wrapper::vcgtz(vin_deq.val[3]), + }}; +#else // __aarch64__ + const uint32x4x4_t pos_mask = {{ + wrapper::vcgt(vin_deq.val[0], vconst_0_f32), + wrapper::vcgt(vin_deq.val[1], vconst_0_f32), + wrapper::vcgt(vin_deq.val[2], vconst_0_f32), + wrapper::vcgt(vin_deq.val[3], vconst_0_f32), + }}; +#endif // __aarch64__ + + const float32x4x4_t tmp_dep = {{ + wrapper::vbsl(pos_mask.val[0], vin_deq.val[0], wrapper::vmul(va_f32, vin_deq.val[0])), + wrapper::vbsl(pos_mask.val[1], vin_deq.val[1], wrapper::vmul(va_f32, vin_deq.val[1])), + wrapper::vbsl(pos_mask.val[2], vin_deq.val[2], wrapper::vmul(va_f32, vin_deq.val[2])), + wrapper::vbsl(pos_mask.val[3], vin_deq.val[3], wrapper::vmul(va_f32, vin_deq.val[3])), + }}; + + tmp = vquantize_signed(tmp_dep, qi_out); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + qasymm8_signed_t in = *(reinterpret_cast<const qasymm8_signed_t *>(input_ptr + x)); + qasymm8_signed_t tmp = 0; + if (act == ActivationLayerInfo::ActivationFunction::RELU) + { + tmp = std::max(const_0, in); + tmp = utility::clamp<int32_t, qasymm8_signed_t>(support::cpp11::lround(tmp * s + o)); + } + else if (act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + tmp = std::min(a, std::max(const_0, in)); + tmp = utility::clamp<int32_t, qasymm8_signed_t>(support::cpp11::lround(tmp * s + o)); + } + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + tmp = std::min(a, std::max(b, in)); + tmp = utility::clamp<int32_t, qasymm8_signed_t>(support::cpp11::lround(tmp * s + o)); + } +#ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead. + else if (act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + float tmp_f = dequantize_qasymm8_signed(in, qi_in); + tmp_f = 1.f / (1.f + std::exp(-tmp_f)); + tmp = quantize_qasymm8_signed(tmp_f, qi_out); + } +#endif // __aarch64__ + else if (act == ActivationLayerInfo::ActivationFunction::TANH) + { + float tmp_f = dequantize_qasymm8_signed(in, qi_in); + tmp_f = a_f32 * std::tanh(b_f32 * tmp_f); + tmp = quantize_qasymm8_signed(tmp_f, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) + { + float tmp_f = dequantize_qasymm8_signed(in, qi_in); + tmp_f = tmp_f * ((std::min(std::max((tmp_f + 3), 0.0f), 6.0f)) * 0.166666667f); + tmp = quantize_qasymm8_signed(tmp_f, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + float tmp_f = dequantize_qasymm8_signed(in, qi_in); + tmp_f = tmp_f > 0 ? tmp_f : tmp_f * a_f32; + tmp = quantize_qasymm8_signed(tmp_f, qi_out); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + *(output_ptr + x) = tmp; + } + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/activation/generic/neon/qsymm16.cpp b/src/cpu/kernels/activation/generic/neon/qsymm16.cpp new file mode 100644 index 0000000000..891646ea00 --- /dev/null +++ b/src/cpu/kernels/activation/generic/neon/qsymm16.cpp @@ -0,0 +1,156 @@ +/* + * Copyright (c) 2020-2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/experimental/Types.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/function_info/ActivationLayerInfo.h" + +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/NESymm.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include <arm_neon.h> +#include <cmath> +#include <cstddef> + +namespace arm_compute +{ +namespace cpu +{ +void neon_qsymm16_activation(const ITensor *src, + ITensor *dst, + const ActivationLayerInfo &act_info, + const Window &window) +{ + constexpr 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 ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); + const auto vconst_1 = vdupq_n_f32(1.f); + const float32x4_t va_f32 = vdupq_n_f32(act_info.a()); + const float32x4_t vb_f32 = vdupq_n_f32(act_info.b()); + const float a_f32 = act_info.a(); + const float b_f32 = act_info.b(); + + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const qsymm16_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<qsymm16_t *>(output.ptr()); + + wrapper::traits::neon_bitvector_t<qsymm16_t, wrapper::traits::BitWidth::W128> tmp; + ARM_COMPUTE_UNUSED(tmp); + + // Compute S elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + if (act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = vdequantize_int16(vin, qi_in.scale); + // Perform activation + const float32x4x2_t tmp_dep = {{ + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))), + }}; + // Re-quantize to new output space + tmp = vquantize_int16(tmp_dep, qi_out.scale); + } + else if (act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + const auto vin_deq = vdequantize_int16(vin, qi_in.scale); + // Perform activation + const float32x4x2_t tmp_dep = {{ + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))), + }}; + // Re-quantize to new output space + tmp = vquantize_int16(tmp_dep, qi_out.scale); + } + + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + // De-quantize + const auto vin_deq = vdequantize_int16(vin, qi_in.scale); + // Perform activation + const float32x4x2_t tmp_dep = {{wrapper::vmin(va_f32, wrapper::vmax(vb_f32, vin_deq.val[0])), + wrapper::vmin(va_f32, wrapper::vmax(vb_f32, vin_deq.val[1]))}}; + // Re-quantize to new output space + tmp = vquantize_int16(tmp_dep, qi_out.scale); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + qsymm16_t in = *(reinterpret_cast<const qsymm16_t *>(input_ptr + x)); + qsymm16_t tmp = 0; + if (act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + float tmp_f = dequantize_qsymm16(in, qi_in.scale); + tmp_f = 1.f / (1.f + std::exp(-tmp_f)); + tmp = quantize_qsymm16(tmp_f, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::TANH) + { + float tmp_f = dequantize_qsymm16(in, qi_in.scale); + tmp_f = a_f32 * std::tanh(b_f32 * tmp_f); + tmp = quantize_qsymm16(tmp_f, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + float tmp_f = dequantize_qsymm16(in, qi_in.scale); + tmp_f = std::min<float>(a_f32, std::max<float>(b_f32, tmp_f)); + tmp = quantize_qsymm16(tmp_f, qi_out); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + *(output_ptr + x) = tmp; + } + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute |