diff options
Diffstat (limited to 'src/cpu/kernels/activation/generic')
-rw-r--r-- | src/cpu/kernels/activation/generic/neon/fp16.cpp | 4 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/neon/fp32.cpp | 2 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/neon/impl.h | 274 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/neon/lut.cpp | 20 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/neon/qasymm8.cpp | 344 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp | 300 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/neon/qsymm16.cpp | 163 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/sve/fp16.cpp | 148 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/sve/fp32.cpp | 149 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/sve2/lut.cpp | 20 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/sve2/qasymm8.cpp | 264 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp | 306 | ||||
-rw-r--r-- | src/cpu/kernels/activation/generic/sve2/qsymm16.cpp | 121 |
13 files changed, 1130 insertions, 985 deletions
diff --git a/src/cpu/kernels/activation/generic/neon/fp16.cpp b/src/cpu/kernels/activation/generic/neon/fp16.cpp index e51b5b3423..ddc6dc24cd 100644 --- a/src/cpu/kernels/activation/generic/neon/fp16.cpp +++ b/src/cpu/kernels/activation/generic/neon/fp16.cpp @@ -31,7 +31,7 @@ namespace cpu { namespace { -constexpr ActFpImplParams Fp16Params = { static_cast<float16_t>(1e-7), 8 }; +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) @@ -40,4 +40,4 @@ void neon_fp16_activation(const ITensor *src, ITensor *dst, const ActivationLaye } } // namespace cpu } // namespace arm_compute -#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
\ No newline at end of file +#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 index 2a3b8a0bfd..e558f8c73e 100644 --- a/src/cpu/kernels/activation/generic/neon/fp32.cpp +++ b/src/cpu/kernels/activation/generic/neon/fp32.cpp @@ -29,7 +29,7 @@ namespace cpu { namespace { -constexpr ActFpImplParams Fp32Params = { static_cast<float>(1e-24), 4 }; +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) { diff --git a/src/cpu/kernels/activation/generic/neon/impl.h b/src/cpu/kernels/activation/generic/neon/impl.h index 05885d8476..afeb6f7f3d 100644 --- a/src/cpu/kernels/activation/generic/neon/impl.h +++ b/src/cpu/kernels/activation/generic/neon/impl.h @@ -24,6 +24,7 @@ #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 { @@ -56,10 +57,14 @@ inline float16x8_t mask_float_vector(const float16x8_t &in, const uint16x8_t &ma #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) +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>; + 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()); @@ -72,12 +77,12 @@ void fp_neon_activation_impl(const ITensor *src, ITensor *dst, const ActivationL // 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 {}); + 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_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_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{}); @@ -88,143 +93,154 @@ void fp_neon_activation_impl(const ITensor *src, ITensor *dst, const ActivationL 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) + execute_window_loop( + win_collapsed, + [&](const Coordinates &) { - const auto vin = wrapper::vloadq(input_ptr + x); - switch(act) + 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) { - 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: + 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); + 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)); + 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; + 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; + 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"); + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); } - 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) + // Compute left-over elements + for (; x < window_end_x; ++x) { - 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"); + 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; } - *(output_ptr + x) = tmp; - } - }, - input, output); + }, + 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 index c973e964e4..f289c80d4b 100644 --- a/src/cpu/kernels/activation/generic/neon/lut.cpp +++ b/src/cpu/kernels/activation/generic/neon/lut.cpp @@ -24,6 +24,7 @@ #include "arm_compute/core/Helpers.h" #include "arm_compute/function_info/ActivationLayerInfo.h" + #include "src/cpu/kernels/lut/list.h" namespace arm_compute @@ -33,19 +34,22 @@ 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(src->info()->data_type() != DataType::QASYMM8 && src->info()->data_type() != DataType::QASYMM8_SIGNED); + ARM_COMPUTE_ERROR_ON(src->info()->data_type() != DataType::QASYMM8 && + src->info()->data_type() != DataType::QASYMM8_SIGNED); 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); + 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 diff --git a/src/cpu/kernels/activation/generic/neon/qasymm8.cpp b/src/cpu/kernels/activation/generic/neon/qasymm8.cpp index e7c146e46f..1451301ea2 100644 --- a/src/cpu/kernels/activation/generic/neon/qasymm8.cpp +++ b/src/cpu/kernels/activation/generic/neon/qasymm8.cpp @@ -25,6 +25,7 @@ #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" @@ -38,7 +39,10 @@ namespace arm_compute { namespace cpu { -void neon_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +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()); @@ -85,206 +89,222 @@ void neon_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationL 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()); + 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; + 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) + // Compute S elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - // 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); - } + 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 = + 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); - } + }}; + // 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 = + 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); - } + }}; + // 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 = + else if (act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) { - { - 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 = + // 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 = - { - { + 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); - } + 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 = + else if (act == ActivationLayerInfo::ActivationFunction::GELU) { - { - 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); - } + 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"); + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); } - 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) + // Compute left-over elements + for (; x < window_end_x; ++x) { - 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)); - } + 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); - } + 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); - } + 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); - } + 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"); + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + *(output_ptr + x) = tmp; } - *(output_ptr + x) = tmp; - } - }, - input, output); + }, + 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 index 52c396459b..a2f588245a 100644 --- a/src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp +++ b/src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp @@ -24,6 +24,7 @@ #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" @@ -36,7 +37,10 @@ namespace arm_compute { namespace cpu { -void neon_qasymm8_signed_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +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()); @@ -76,191 +80,195 @@ void neon_qasymm8_signed_activation(const ITensor *src, ITensor *dst, const Acti 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()); + 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; + 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) + // Compute S elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - // 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); - } + 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 = + 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); - } + }}; + // 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 = + 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 = + }}; + // Re-quantize to new output space + tmp = vquantize_signed(tmp_dep, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) { - { - 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); + // 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 = - { - { + 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 = - { - { + 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 = - { - { + 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"); + tmp = vquantize_signed(tmp_dep, qi_out); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); } - 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) + // Compute left-over elements + for (; x < window_end_x; ++x) { - tmp = std::min(a, std::max(b, in)); - tmp = utility::clamp<int32_t, qasymm8_signed_t>(support::cpp11::lround(tmp * s + o)); - } + 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); - } + 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"); + 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; } - *(output_ptr + x) = tmp; - } - }, - input, output); + }, + 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 index 2aea6cba3c..891646ea00 100644 --- a/src/cpu/kernels/activation/generic/neon/qsymm16.cpp +++ b/src/cpu/kernels/activation/generic/neon/qsymm16.cpp @@ -21,11 +21,12 @@ * 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/core/experimental/Types.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" @@ -38,7 +39,10 @@ namespace arm_compute { namespace cpu { -void neon_qsymm16_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +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()); @@ -59,103 +63,94 @@ void neon_qsymm16_activation(const ITensor *src, ITensor *dst, const ActivationL 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()); + 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); + 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) + // Compute S elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - // De-quantize - const auto vin_deq = vdequantize_int16(vin, qi_in.scale); - // Perform activation - const float32x4x2_t tmp_dep = + 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 = + }}; + // 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); - } + }}; + // 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 = + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) { - { - 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"); + // 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); } - 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) + // Compute left-over elements + for (; x < window_end_x; ++x) { - 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"); + 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; } - *(output_ptr + x) = tmp; - } - }, - input, output); + }, + input, output); } } // namespace cpu } // namespace arm_compute diff --git a/src/cpu/kernels/activation/generic/sve/fp16.cpp b/src/cpu/kernels/activation/generic/sve/fp16.cpp index 4757c60d8f..97399e01e0 100644 --- a/src/cpu/kernels/activation/generic/sve/fp16.cpp +++ b/src/cpu/kernels/activation/generic/sve/fp16.cpp @@ -29,11 +29,11 @@ #include "arm_compute/core/Window.h" #include "arm_compute/function_info/ActivationLayerInfo.h" -#include <cmath> -#include <cstddef> - #include "src/core/NEON/SVEMath.h" + #include <arm_sve.h> +#include <cmath> +#include <cstddef> namespace arm_compute { @@ -59,77 +59,87 @@ void sve_fp16_activation(const ITensor *src, ITensor *dst, const ActivationLayer const auto va = svdup_n_f16(act_info.a()); const auto vb = svdup_n_f16(act_info.b()); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast<const float16_t *>(input.ptr()); - const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr()); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const float16_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr()); - svfloat16_t tmp; + svfloat16_t tmp; - int x = window_start_x; - svbool_t pg = svwhilelt_b16(x, window_end_x); - do - { - const auto vin = svld1_f16(pg, input_ptr + x); - switch(act) + int x = window_start_x; + svbool_t pg = svwhilelt_b16(x, window_end_x); + do { - case ActivationLayerInfo::ActivationFunction::ABS: - tmp = svabs_f16_z(pg, vin); - break; - case ActivationLayerInfo::ActivationFunction::LINEAR: - tmp = svmla_f16_z(pg, vb, va, vin); - break; - case ActivationLayerInfo::ActivationFunction::LOGISTIC: - tmp = svinv_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, svneg_f16_z(pg, vin)))); - break; - case ActivationLayerInfo::ActivationFunction::RELU: - tmp = svmax_f16_z(pg, const_0, vin); - break; - case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: - tmp = svmin_f16_z(pg, va, svmax_f16_z(pg, const_0, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: - tmp = svmin_f16_z(pg, va, svmax_f16_z(pg, vb, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: - tmp = svadd_f16_z(pg, svmul_f16_z(pg, svmin_f16_z(pg, vin, const_0), va), svmax_f16_z(pg, vin, const_0)); - break; - case ActivationLayerInfo::ActivationFunction::SOFT_RELU: - tmp = svlog_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, vin))); - break; - case ActivationLayerInfo::ActivationFunction::ELU: - tmp = svsel_f16(svcmpgt_f16(pg, vin, const_0), vin, svmul_f16_z(pg, va, svsub_f16_z(pg, svexp_f16_z(pg, vin), const_1))); - break; - case ActivationLayerInfo::ActivationFunction::SQRT: - tmp = svsqrt_f16_z(pg, vin); - break; - case ActivationLayerInfo::ActivationFunction::SQUARE: - tmp = svmul_f16_z(pg, vin, vin); - break; - case ActivationLayerInfo::ActivationFunction::TANH: - tmp = svmul_f16_z(pg, va, svtanh_f16_z(pg, svmul_f16_z(pg, vb, vin))); - break; - case ActivationLayerInfo::ActivationFunction::IDENTITY: - tmp = vin; - break; - case ActivationLayerInfo::ActivationFunction::HARD_SWISH: - tmp = svmul_f16_z(pg, vin, svmul_f16_z(pg, const_inv_6, svmin_f16_z(pg, const_6, svmax_f16_z(pg, const_0, svadd_f16_z(pg, vin, const_3))))); - break; - case ActivationLayerInfo::ActivationFunction::SWISH: - tmp = svmul_f16_z(pg, vin, svinv_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, svneg_f16_z(pg, svmul_f16_z(pg, va, vin)))))); - break; - default: - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - svst1_f16(pg, output_ptr + x, tmp); + const auto vin = svld1_f16(pg, input_ptr + x); + switch (act) + { + case ActivationLayerInfo::ActivationFunction::ABS: + tmp = svabs_f16_z(pg, vin); + break; + case ActivationLayerInfo::ActivationFunction::LINEAR: + tmp = svmla_f16_z(pg, vb, va, vin); + break; + case ActivationLayerInfo::ActivationFunction::LOGISTIC: + tmp = svinv_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, svneg_f16_z(pg, vin)))); + break; + case ActivationLayerInfo::ActivationFunction::RELU: + tmp = svmax_f16_z(pg, const_0, vin); + break; + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + tmp = svmin_f16_z(pg, va, svmax_f16_z(pg, const_0, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: + tmp = svmin_f16_z(pg, va, svmax_f16_z(pg, vb, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: + tmp = svadd_f16_z(pg, svmul_f16_z(pg, svmin_f16_z(pg, vin, const_0), va), + svmax_f16_z(pg, vin, const_0)); + break; + case ActivationLayerInfo::ActivationFunction::SOFT_RELU: + tmp = svlog_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, vin))); + break; + case ActivationLayerInfo::ActivationFunction::ELU: + tmp = svsel_f16(svcmpgt_f16(pg, vin, const_0), vin, + svmul_f16_z(pg, va, svsub_f16_z(pg, svexp_f16_z(pg, vin), const_1))); + break; + case ActivationLayerInfo::ActivationFunction::SQRT: + tmp = svsqrt_f16_z(pg, vin); + break; + case ActivationLayerInfo::ActivationFunction::SQUARE: + tmp = svmul_f16_z(pg, vin, vin); + break; + case ActivationLayerInfo::ActivationFunction::TANH: + tmp = svmul_f16_z(pg, va, svtanh_f16_z(pg, svmul_f16_z(pg, vb, vin))); + break; + case ActivationLayerInfo::ActivationFunction::IDENTITY: + tmp = vin; + break; + case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + tmp = svmul_f16_z( + pg, vin, + svmul_f16_z( + pg, const_inv_6, + svmin_f16_z(pg, const_6, svmax_f16_z(pg, const_0, svadd_f16_z(pg, vin, const_3))))); + break; + case ActivationLayerInfo::ActivationFunction::SWISH: + tmp = svmul_f16_z( + pg, vin, + svinv_f16_z(pg, svadd_f16_z(pg, const_1, + svexp_f16_z(pg, svneg_f16_z(pg, svmul_f16_z(pg, va, vin)))))); + break; + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + svst1_f16(pg, output_ptr + x, tmp); - x += svcnth(); - pg = svwhilelt_b16(x, window_end_x); + x += svcnth(); + pg = svwhilelt_b16(x, window_end_x); - } - while(svptest_any(svptrue_b16(), pg)); - }, - input, output); + } while (svptest_any(svptrue_b16(), pg)); + }, + input, output); } } // namespace cpu } // namespace arm_compute diff --git a/src/cpu/kernels/activation/generic/sve/fp32.cpp b/src/cpu/kernels/activation/generic/sve/fp32.cpp index 87f04c255a..d1b075d52c 100644 --- a/src/cpu/kernels/activation/generic/sve/fp32.cpp +++ b/src/cpu/kernels/activation/generic/sve/fp32.cpp @@ -26,13 +26,13 @@ #include "arm_compute/core/ITensorPack.h" #include "arm_compute/core/Window.h" #include "arm_compute/function_info/ActivationLayerInfo.h" + #include "src/core/NEON/SVEMath.h" +#include <arm_sve.h> #include <cmath> #include <cstddef> -#include <arm_sve.h> - namespace arm_compute { namespace cpu @@ -58,78 +58,89 @@ void sve_fp32_activation(const ITensor *src, ITensor *dst, const ActivationLayer const auto va = svdup_n_f32(act_info.a()); const auto vb = svdup_n_f32(act_info.b()); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast<const float *>(input.ptr()); - const auto output_ptr = reinterpret_cast<float *>(output.ptr()); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const float *>(input.ptr()); + const auto output_ptr = reinterpret_cast<float *>(output.ptr()); - svfloat32_t tmp; + svfloat32_t tmp; - // Compute S elements per iteration - int x = window_start_x; - svbool_t pg = svwhilelt_b32(x, window_end_x); - do - { - const auto vin = svld1_f32(pg, input_ptr + x); - switch(act) + // Compute S elements per iteration + int x = window_start_x; + svbool_t pg = svwhilelt_b32(x, window_end_x); + do { - case ActivationLayerInfo::ActivationFunction::ABS: - tmp = svabs_f32_z(pg, vin); - break; - case ActivationLayerInfo::ActivationFunction::LINEAR: - tmp = svmla_f32_z(pg, vb, va, vin); - break; - case ActivationLayerInfo::ActivationFunction::LOGISTIC: - tmp = svinv_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, svneg_f32_z(pg, vin)))); - break; - case ActivationLayerInfo::ActivationFunction::RELU: - tmp = svmax_f32_z(pg, const_0, vin); - break; - case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: - tmp = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: - tmp = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: - tmp = svadd_f32_z(pg, svmul_f32_z(pg, svmin_f32_z(pg, vin, const_0), va), svmax_f32_z(pg, vin, const_0)); - break; - case ActivationLayerInfo::ActivationFunction::SOFT_RELU: - tmp = svsel_f32(svcmpgt_f32(pg, vin, soft_relu_thresh), vin, svlog_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, vin)))); - break; - case ActivationLayerInfo::ActivationFunction::ELU: - tmp = svsel_f32(svcmpgt_f32(pg, vin, const_0), vin, svmul_f32_z(pg, va, svsub_f32_z(pg, svexp_f32_z(pg, vin), const_1))); - break; - case ActivationLayerInfo::ActivationFunction::SQRT: - tmp = svsqrt_f32_z(pg, vin); - break; - case ActivationLayerInfo::ActivationFunction::SQUARE: - tmp = svmul_f32_z(pg, vin, vin); - break; - case ActivationLayerInfo::ActivationFunction::TANH: - tmp = svmul_f32_z(pg, va, svtanh_f32_z(pg, svmul_f32_z(pg, vb, vin))); - break; - case ActivationLayerInfo::ActivationFunction::IDENTITY: - tmp = vin; - break; - case ActivationLayerInfo::ActivationFunction::HARD_SWISH: - tmp = svmul_f32_z(pg, vin, svmul_f32_z(pg, const_inv_6, svmin_f32_z(pg, const_6, svmax_f32_z(pg, const_0, svadd_f32_z(pg, vin, const_3))))); - break; - case ActivationLayerInfo::ActivationFunction::SWISH: - tmp = svmul_f32_z(pg, vin, svinv_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, svneg_f32_z(pg, svmul_f32_z(pg, va, vin)))))); - break; - default: - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - svst1_f32(pg, output_ptr + x, tmp); + const auto vin = svld1_f32(pg, input_ptr + x); + switch (act) + { + case ActivationLayerInfo::ActivationFunction::ABS: + tmp = svabs_f32_z(pg, vin); + break; + case ActivationLayerInfo::ActivationFunction::LINEAR: + tmp = svmla_f32_z(pg, vb, va, vin); + break; + case ActivationLayerInfo::ActivationFunction::LOGISTIC: + tmp = svinv_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, svneg_f32_z(pg, vin)))); + break; + case ActivationLayerInfo::ActivationFunction::RELU: + tmp = svmax_f32_z(pg, const_0, vin); + break; + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + tmp = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: + tmp = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: + tmp = svadd_f32_z(pg, svmul_f32_z(pg, svmin_f32_z(pg, vin, const_0), va), + svmax_f32_z(pg, vin, const_0)); + break; + case ActivationLayerInfo::ActivationFunction::SOFT_RELU: + tmp = svsel_f32(svcmpgt_f32(pg, vin, soft_relu_thresh), vin, + svlog_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, vin)))); + break; + case ActivationLayerInfo::ActivationFunction::ELU: + tmp = svsel_f32(svcmpgt_f32(pg, vin, const_0), vin, + svmul_f32_z(pg, va, svsub_f32_z(pg, svexp_f32_z(pg, vin), const_1))); + break; + case ActivationLayerInfo::ActivationFunction::SQRT: + tmp = svsqrt_f32_z(pg, vin); + break; + case ActivationLayerInfo::ActivationFunction::SQUARE: + tmp = svmul_f32_z(pg, vin, vin); + break; + case ActivationLayerInfo::ActivationFunction::TANH: + tmp = svmul_f32_z(pg, va, svtanh_f32_z(pg, svmul_f32_z(pg, vb, vin))); + break; + case ActivationLayerInfo::ActivationFunction::IDENTITY: + tmp = vin; + break; + case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + tmp = svmul_f32_z( + pg, vin, + svmul_f32_z( + pg, const_inv_6, + svmin_f32_z(pg, const_6, svmax_f32_z(pg, const_0, svadd_f32_z(pg, vin, const_3))))); + break; + case ActivationLayerInfo::ActivationFunction::SWISH: + tmp = svmul_f32_z( + pg, vin, + svinv_f32_z(pg, svadd_f32_z(pg, const_1, + svexp_f32_z(pg, svneg_f32_z(pg, svmul_f32_z(pg, va, vin)))))); + break; + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + svst1_f32(pg, output_ptr + x, tmp); - x += svcntw(); - pg = svwhilelt_b32(x, window_end_x); + x += svcntw(); + pg = svwhilelt_b32(x, window_end_x); - } - while(svptest_any(svptrue_b32(), pg)); - }, - input, output); + } while (svptest_any(svptrue_b32(), pg)); + }, + input, output); } } // namespace cpu } // namespace arm_compute diff --git a/src/cpu/kernels/activation/generic/sve2/lut.cpp b/src/cpu/kernels/activation/generic/sve2/lut.cpp index d65de8d649..2ed667debf 100644 --- a/src/cpu/kernels/activation/generic/sve2/lut.cpp +++ b/src/cpu/kernels/activation/generic/sve2/lut.cpp @@ -24,6 +24,7 @@ #include "arm_compute/core/Helpers.h" #include "arm_compute/function_info/ActivationLayerInfo.h" + #include "src/cpu/kernels/lut/list.h" namespace arm_compute @@ -33,19 +34,22 @@ namespace cpu #ifdef __aarch64__ void sve2_q8_activation_lut(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) { - ARM_COMPUTE_ERROR_ON(src->info()->data_type() != DataType::QASYMM8 && src->info()->data_type() != DataType::QASYMM8_SIGNED); + ARM_COMPUTE_ERROR_ON(src->info()->data_type() != DataType::QASYMM8 && + src->info()->data_type() != DataType::QASYMM8_SIGNED); 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 = input.ptr(); - auto output_ptr = output.ptr(); - lut_u8_sve2(act_info.lut().data(), 1u, window_end_x, &input_ptr, &output_ptr); - }, - input, output); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = input.ptr(); + auto output_ptr = output.ptr(); + lut_u8_sve2(act_info.lut().data(), 1u, window_end_x, &input_ptr, &output_ptr); + }, + input, output); } #endif // __aarch64__ } // namespace cpu diff --git a/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp b/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp index bc9bc7aa3c..7efa9e4b72 100644 --- a/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp +++ b/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp @@ -26,18 +26,21 @@ #include "arm_compute/core/Window.h" #include "arm_compute/function_info/ActivationLayerInfo.h" -#include <cmath> -#include <cstddef> - #include "src/core/NEON/SVEAsymm.h" #include "src/core/NEON/SVEMath.h" + #include <arm_sve.h> +#include <cmath> +#include <cstddef> namespace arm_compute { namespace cpu { -void sve2_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +void sve2_qasymm8_activation(const ITensor *src, + ITensor *dst, + const ActivationLayerInfo &act_info, + const Window &window) { const auto window_start_x = static_cast<int>(window.x().start()); const auto window_end_x = static_cast<int>(window.x().end()); @@ -61,7 +64,7 @@ void sve2_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationL // Initialise scale/offset for re-quantization bool requant = true; - if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset) + if (qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset) { requant = false; } @@ -78,139 +81,160 @@ void sve2_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationL const auto vo_s32 = svdup_n_s32(o_s32); // Initialise scale/offset for re-quantization for leaky relu - int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8), - arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8), + arm_compute::RoundingPolicy::TO_NEAREST_EVEN); const auto vs_leaky_s32 = svdup_n_s32(s_leaky_s32); const auto vo_leaky_s32 = svdup_n_s32(o_leaky_s32); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast<const uint8_t *>(input.ptr()); - const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr()); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const uint8_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr()); - svuint8_t tmp; + svuint8_t tmp; - int x = window_start_x; - svbool_t pg = svwhilelt_b8(x, window_end_x); - do - { - const auto vin = svld1_u8(pg, input_ptr + x); - if(act == ActivationLayerInfo::ActivationFunction::RELU) - { - // Perform activation - tmp = svmax_u8_z(pg, vconst_0, vin); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - // Perform activation - tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vconst_0, vin)); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) - { - // Perform activation - tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vb, vin)); - // Re-quantize to new output space - tmp = svmla_qasymm8_z(pg, tmp, vs, vo); - } - else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = svcreate4_f32(svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3)))))); - - // Re-quantize to new output space - tmp = svquantize_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) + int x = window_start_x; + svbool_t pg = svwhilelt_b8(x, window_end_x); + do { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = svcreate4_f32(svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32)))); - - // Re-quantize to new output space - tmp = svquantize_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) - { - svbool_t p0, p1, p2, p3; - svint32x4_t tmp_dep; - - // Expand to int32 - const svint32x4_t vin_s32 = svcreate4_s32( - svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(vin))), - svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(vin))), - svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(vin))), - svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(vin)))); - - // Compare elements to input offset - if(qi_in.scale >= 0) + const auto vin = svld1_u8(pg, input_ptr + x); + if (act == ActivationLayerInfo::ActivationFunction::RELU) { - p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); - p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); - p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); - p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + // Perform activation + tmp = svmax_u8_z(pg, vconst_0, vin); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; } - else + else if (act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) { - p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); - p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); - p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); - p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + // Perform activation + tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vconst_0, vin)); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; } - - // Multiply negative elements and requantize if necessary - if(requant) + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + // Perform activation + tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vb, vin)); + // Re-quantize to new output space + tmp = svmla_qasymm8_z(pg, tmp, vs, vo); + } + else if (act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = svcreate4_f32( + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))), + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))), + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))), + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3)))))); + + // Re-quantize to new output space + tmp = svquantize_z(pg, tmp_dep, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::TANH) { - tmp_dep = svcreate4_s32( - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), svsel(p0, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), svsel(p1, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), svsel(p2, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), svsel(p3, vs_leaky_s32, vs_s32)), 8)); + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = svcreate4_f32( + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32)))); + + // Re-quantize to new output space + tmp = svquantize_z(pg, tmp_dep, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + svbool_t p0, p1, p2, p3; + svint32x4_t tmp_dep; + + // Expand to int32 + const svint32x4_t vin_s32 = svcreate4_s32(svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(vin))), + svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(vin))), + svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(vin))), + svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(vin)))); + + // Compare elements to input offset + if (qi_in.scale >= 0) + { + p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); + p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); + p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); + p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + } + else + { + p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); + p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); + p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); + p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + } + + // Multiply negative elements and requantize if necessary + if (requant) + { + tmp_dep = svcreate4_s32( + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), + svsel(p0, vs_leaky_s32, vs_s32)), + 8), + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), + svsel(p1, vs_leaky_s32, vs_s32)), + 8), + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), + svsel(p2, vs_leaky_s32, vs_s32)), + 8), + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), + svsel(p3, vs_leaky_s32, vs_s32)), + 8)); + } + else + { + tmp_dep = svcreate4_s32( + svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8)); + } + + // Convert uint32 vectors to uint16 vectors (with saturation) + const auto v_low_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1)); + const auto v_high_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3)); + + // convert uint16 vectors to uint8 vectors (with saturation) + tmp = svqxtnt_u16(svqxtnb_u16(v_low_u16), v_high_u16); } else { - tmp_dep = svcreate4_s32( - svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8)); + ARM_COMPUTE_ERROR("Unsupported activation function"); } - // Convert uint32 vectors to uint16 vectors (with saturation) - const auto v_low_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1)); - const auto v_high_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3)); - - // convert uint16 vectors to uint8 vectors (with saturation) - tmp = svqxtnt_u16(svqxtnb_u16(v_low_u16), v_high_u16); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - - svst1_u8(pg, output_ptr + x, tmp); - - x += svcntb(); - pg = svwhilelt_b8(x, window_end_x); + svst1_u8(pg, output_ptr + x, tmp); - } - while(svptest_any(svptrue_b8(), pg)); + x += svcntb(); + pg = svwhilelt_b8(x, window_end_x); - }, - input, output); + } while (svptest_any(svptrue_b8(), pg)); + }, + input, output); } } // namespace cpu } // namespace arm_compute diff --git a/src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp b/src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp index d20684f54d..e4667522dd 100644 --- a/src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp +++ b/src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp @@ -24,20 +24,23 @@ #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" - -#include <cmath> -#include <cstddef> #include "src/core/NEON/SVEAsymm.h" #include "src/core/NEON/SVEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" + #include <arm_sve.h> +#include <cmath> +#include <cstddef> namespace arm_compute { namespace cpu { -void sve2_qasymm8_signed_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +void sve2_qasymm8_signed_activation(const ITensor *src, + ITensor *dst, + const ActivationLayerInfo &act_info, + const Window &window) { const auto window_start_x = static_cast<int>(window.x().start()); const auto window_end_x = static_cast<int>(window.x().end()); @@ -65,7 +68,7 @@ void sve2_qasymm8_signed_activation(const ITensor *src, ITensor *dst, const Acti // Initialise scale/offset for re-quantization bool requant = true; - if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset) + if (qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset) { requant = false; } @@ -82,151 +85,190 @@ void sve2_qasymm8_signed_activation(const ITensor *src, ITensor *dst, const Acti const auto vo_s32 = svdup_n_s32(o_s32); // Initialise scale/offset for re-quantization for leaky relu - int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8), - arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8), + arm_compute::RoundingPolicy::TO_NEAREST_EVEN); const auto vs_leaky_s32 = svdup_n_s32(s_leaky_s32); const auto vo_leaky_s32 = svdup_n_s32(o_leaky_s32); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast<const int8_t *>(input.ptr()); - const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr()); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const int8_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr()); - svint8_t tmp; + svint8_t tmp; - int x = window_start_x; - svbool_t pg = svwhilelt_b8(x, window_end_x); - do - { - const auto vin = svld1_s8(pg, input_ptr + x); - if(act == ActivationLayerInfo::ActivationFunction::RELU) - { - // Perform activation - tmp = svmax_s8_z(pg, vconst_0, vin); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - // Perform activation - tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vconst_0, vin)); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) - { - // Perform activation - tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vb, vin)); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = svcreate4_f32( - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3)))))); - // Re-quantize to new output space - tmp = svquantize_signed_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = svcreate4_f32( - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32)))); - // Re-quantize to new output space - tmp = svquantize_signed_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = svcreate4_f32( - svmul_f32_z(pg, svget4_f32(vin_deq, 0), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 0), const_3_f32))))), - svmul_f32_z(pg, svget4_f32(vin_deq, 1), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 1), const_3_f32))))), - svmul_f32_z(pg, svget4_f32(vin_deq, 2), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 2), const_3_f32))))), - svmul_f32_z(pg, svget4_f32(vin_deq, 3), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 3), const_3_f32)))))); - // Re-quantize to new output space - tmp = svquantize_signed_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + int x = window_start_x; + svbool_t pg = svwhilelt_b8(x, window_end_x); + do { - svbool_t p0, p1, p2, p3; - svint32x4_t tmp_dep; - - // Expand to int32 - const svint32x4_t vin_s32 = svcreate4_s32( - svmovlb_s32(svmovlb_s16(vin)), - svmovlt_s32(svmovlb_s16(vin)), - svmovlb_s32(svmovlt_s16(vin)), - svmovlt_s32(svmovlt_s16(vin))); - - // Compare elements to input offset - if(qi_in.scale >= 0) + const auto vin = svld1_s8(pg, input_ptr + x); + if (act == ActivationLayerInfo::ActivationFunction::RELU) { - p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); - p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); - p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); - p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + // Perform activation + tmp = svmax_s8_z(pg, vconst_0, vin); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; } - else + else if (act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) { - p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); - p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); - p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); - p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + // Perform activation + tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vconst_0, vin)); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; } - - // Multiply negative elements and requantize if necessary - if(requant) + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) { - tmp_dep = svcreate4_s32( - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), svsel(p0, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), svsel(p1, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), svsel(p2, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), svsel(p3, vs_leaky_s32, vs_s32)), 8)); + // Perform activation + tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vb, vin)); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; } - else + else if (act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = svcreate4_f32( + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))), + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))), + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))), + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3)))))); + // Re-quantize to new output space + tmp = svquantize_signed_z(pg, tmp_dep, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = svcreate4_f32( + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32)))); + // Re-quantize to new output space + tmp = svquantize_signed_z(pg, tmp_dep, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) { - tmp_dep = svcreate4_s32( - svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8)); + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = svcreate4_f32( + svmul_f32_z(pg, svget4_f32(vin_deq, 0), + svmul_f32_z(pg, const_inv_6_f32, + svmin_f32_z(pg, const_6_f32, + svmax_f32_z(pg, const_0_f32, + svadd_f32_z(pg, svget4_f32(vin_deq, 0), + const_3_f32))))), + svmul_f32_z(pg, svget4_f32(vin_deq, 1), + svmul_f32_z(pg, const_inv_6_f32, + svmin_f32_z(pg, const_6_f32, + svmax_f32_z(pg, const_0_f32, + svadd_f32_z(pg, svget4_f32(vin_deq, 1), + const_3_f32))))), + svmul_f32_z(pg, svget4_f32(vin_deq, 2), + svmul_f32_z(pg, const_inv_6_f32, + svmin_f32_z(pg, const_6_f32, + svmax_f32_z(pg, const_0_f32, + svadd_f32_z(pg, svget4_f32(vin_deq, 2), + const_3_f32))))), + svmul_f32_z(pg, svget4_f32(vin_deq, 3), + svmul_f32_z(pg, const_inv_6_f32, + svmin_f32_z(pg, const_6_f32, + svmax_f32_z(pg, const_0_f32, + svadd_f32_z(pg, svget4_f32(vin_deq, 3), + const_3_f32)))))); + // Re-quantize to new output space + tmp = svquantize_signed_z(pg, tmp_dep, qi_out); } + else if (act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + svbool_t p0, p1, p2, p3; + svint32x4_t tmp_dep; - // Convert uint32 vectors to uint16 vectors (with saturation) - const auto v_low_s16 = svqxtnt_s32(svqxtnb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1)); - const auto v_high_s16 = svqxtnt_s32(svqxtnb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3)); + // Expand to int32 + const svint32x4_t vin_s32 = + svcreate4_s32(svmovlb_s32(svmovlb_s16(vin)), svmovlt_s32(svmovlb_s16(vin)), + svmovlb_s32(svmovlt_s16(vin)), svmovlt_s32(svmovlt_s16(vin))); - // convert uint16 vectors to uint8 vectors (with saturation) - tmp = svqxtnt_s16(svqxtnb_s16(v_low_s16), v_high_s16); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } + // Compare elements to input offset + if (qi_in.scale >= 0) + { + p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); + p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); + p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); + p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + } + else + { + p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); + p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); + p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); + p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + } + + // Multiply negative elements and requantize if necessary + if (requant) + { + tmp_dep = svcreate4_s32( + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), + svsel(p0, vs_leaky_s32, vs_s32)), + 8), + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), + svsel(p1, vs_leaky_s32, vs_s32)), + 8), + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), + svsel(p2, vs_leaky_s32, vs_s32)), + 8), + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), + svsel(p3, vs_leaky_s32, vs_s32)), + 8)); + } + else + { + tmp_dep = svcreate4_s32( + svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8)); + } + + // Convert uint32 vectors to uint16 vectors (with saturation) + const auto v_low_s16 = svqxtnt_s32(svqxtnb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1)); + const auto v_high_s16 = svqxtnt_s32(svqxtnb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3)); + + // convert uint16 vectors to uint8 vectors (with saturation) + tmp = svqxtnt_s16(svqxtnb_s16(v_low_s16), v_high_s16); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } - svst1_s8(pg, output_ptr + x, tmp); + svst1_s8(pg, output_ptr + x, tmp); - x += svcntb(); - pg = svwhilelt_b8(x, window_end_x); + x += svcntb(); + pg = svwhilelt_b8(x, window_end_x); - } - while(svptest_any(svptrue_b8(), pg)); - }, - input, output); + } while (svptest_any(svptrue_b8(), pg)); + }, + input, output); } } // namespace cpu } // namespace arm_compute diff --git a/src/cpu/kernels/activation/generic/sve2/qsymm16.cpp b/src/cpu/kernels/activation/generic/sve2/qsymm16.cpp index 5154fac8a7..f955893307 100644 --- a/src/cpu/kernels/activation/generic/sve2/qsymm16.cpp +++ b/src/cpu/kernels/activation/generic/sve2/qsymm16.cpp @@ -21,24 +21,27 @@ * 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/core/experimental/Types.h" #include "arm_compute/function_info/ActivationLayerInfo.h" -#include <cmath> -#include <cstddef> - #include "src/core/NEON/SVEMath.h" #include "src/core/NEON/SVESymm.h" + #include <arm_sve.h> +#include <cmath> +#include <cstddef> namespace arm_compute { namespace cpu { -void sve2_qsymm16_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +void sve2_qsymm16_activation(const ITensor *src, + ITensor *dst, + const ActivationLayerInfo &act_info, + const Window &window) { const auto window_start_x = static_cast<int>(window.x().start()); const auto window_end_x = static_cast<int>(window.x().end()); @@ -56,62 +59,70 @@ void sve2_qsymm16_activation(const ITensor *src, ITensor *dst, const ActivationL const auto va_f32 = svdup_n_f32(act_info.a()); const auto vb_f32 = svdup_n_f32(act_info.b()); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast<const int16_t *>(input.ptr()); - const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr()); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const int16_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr()); - svint16_t tmp; + svint16_t tmp; - int x = window_start_x; - svbool_t pg = svwhilelt_b16(x, window_end_x); - do - { - const auto vin = svld1_s16(pg, input_ptr + x); - if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - // De-quantize - auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale); - // Perform activation - const svfloat32x2_t tmp_dep = svcreate2_f32(svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 0))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 1)))))); - // Re-quantize to new output space - tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - // De-quantize - auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale); - // Perform activation - const svfloat32x2_t tmp_dep = svcreate2_f32(svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 0), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 1), vb_f32)))); - // Re-quantize to new output space - tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale); - } - else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) - { - // De-quantize - auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale); - // Perform activation - const svfloat32x2_t tmp_dep = svcreate2_f32(svmin_f32_z(pg, va_f32, svmax_f32_z(pg, vb_f32, svget2_f32(vin_deq, 0))), - svmin_f32_z(pg, va_f32, svmax_f32_z(pg, vb_f32, svget2_f32(vin_deq, 1)))); - // Re-quantize to new output space - tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale); - } - else + int x = window_start_x; + svbool_t pg = svwhilelt_b16(x, window_end_x); + do { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } + const auto vin = svld1_s16(pg, input_ptr + x); + if (act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale); + // Perform activation + const svfloat32x2_t tmp_dep = svcreate2_f32( + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 0))))), + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 1)))))); + // Re-quantize to new output space + tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale); + } + else if (act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale); + // Perform activation + const svfloat32x2_t tmp_dep = svcreate2_f32( + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 0), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 1), vb_f32)))); + // Re-quantize to new output space + tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale); + } + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + // De-quantize + auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale); + // Perform activation + const svfloat32x2_t tmp_dep = + svcreate2_f32(svmin_f32_z(pg, va_f32, svmax_f32_z(pg, vb_f32, svget2_f32(vin_deq, 0))), + svmin_f32_z(pg, va_f32, svmax_f32_z(pg, vb_f32, svget2_f32(vin_deq, 1)))); + // Re-quantize to new output space + tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } - svst1_s16(pg, output_ptr + x, tmp); + svst1_s16(pg, output_ptr + x, tmp); - x += svcnth(); - pg = svwhilelt_b16(x, window_end_x); + x += svcnth(); + pg = svwhilelt_b16(x, window_end_x); - } - while(svptest_any(svptrue_b16(), pg)); - }, - input, output); + } while (svptest_any(svptrue_b16(), pg)); + }, + input, output); } } // namespace cpu } // namespace arm_compute |