perf: remove switch in Arm(R) Neon(TM) activations

Move activation-function selection to new template dispatch_activation_function(), which facilitates inlining of code to increase performance.
This commit addresses the Neon kernels.
The patch improves the performance for ReLU by 4-7%, Logistic by 11% and Hyperbolic tangent by 2-5%. Performance improvements can vary by platform and problem size. Generally this patch improves the performance the most for larger problem sizes.

Partially Resolves: COMPMID-8359
Change-Id: I661e5d51b9993cda9c0fef2f198b1a7438f8ce8c
Signed-off-by: Dennis Wildmark <[email protected]>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/14722
Tested-by: Arm Jenkins <[email protected]>
Reviewed-by: Pablo Marquez Tello <[email protected]>
Benchmark: Arm Jenkins <[email protected]>
Comments-Addressed: Arm Jenkins <[email protected]>
Reviewed-by: Dongsung Kim <[email protected]>

Author: dwildmark

src/cpu/kernels/activation/generic/neon/fp16.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021, 2025 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,7 +23,7 @@
  */
 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
 
-#include "src/cpu/kernels/activation/generic/neon/impl.h"
+#include "src/cpu/kernels/activation/generic/neon/fp_impl.h"
 
 namespace arm_compute
 {

src/cpu/kernels/activation/generic/neon/fp32.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021, 2025 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,7 +22,7 @@
  * SOFTWARE.
  */
 
-#include "src/cpu/kernels/activation/generic/neon/impl.h"
+#include "src/cpu/kernels/activation/generic/neon/fp_impl.h"
 namespace arm_compute
 {
 namespace cpu

src/cpu/kernels/activation/generic/neon/fp_impl.h

@@ -0,0 +1,249 @@
+/*
+ * Copyright (c) 2020-2023, 2025 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef ACL_SRC_CPU_KERNELS_ACTIVATION_GENERIC_NEON_FP_IMPL_H
+#define ACL_SRC_CPU_KERNELS_ACTIVATION_GENERIC_NEON_FP_IMPL_H
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/function_info/ActivationLayerInfo.h"
+
+#include "src/core/NEON/wrapper/wrapper.h"
+namespace arm_compute
+{
+namespace cpu
+{
+/** Constant parameters needed by the activation implementation.
+ *  These parameters differ for each floating type
+ *
+ * @note This are passed as a struct as C++ does not allow float as a template parameter until C++20
+ **/
+struct ActFpImplParams
+{
+    float delta;  /**< Minimum delta needed to avoid NaN on corner-cases of elementary functions */
+    int   step_x; /**< Window step at the x dimension */
+};
+
+#ifndef __aarch64__
+inline float32x4_t mask_float_vector(const float32x4_t &in, const uint32x4_t &mask)
+{
+    auto int_in = vreinterpretq_u32_f32(in);
+    return vreinterpretq_f32_u32(wrapper::vand(int_in, mask));
+}
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+inline float16x8_t mask_float_vector(const float16x8_t &in, const uint16x8_t &mask)
+{
+    auto int_in = vreinterpretq_u16_f16(in);
+    return vreinterpretq_f16_u16(wrapper::vand(int_in, mask));
+}
+#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+#endif /* __aarch64__ */
+
+template <typename T, const ActFpImplParams &P, typename F>
+void dispatch_fp_neon_activation_function(ActivationLayerInfo::ActivationFunction act,
+                                          const ActivationLayerInfo              &act_info,
+                                          F                                     &&fn)
+{
+    using ExactTagType =
+        typename arm_compute::wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
+#ifndef __aarch64__
+    const auto delta = wrapper::vdup_n(static_cast<T>(P.delta), ExactTagType{});
+#else  /* #ifndef __aarch64__ */
+    const auto const_inv_2      = wrapper::vdup_n(static_cast<T>(0.5f), ExactTagType{});
+    const auto const_inv_sqrt_2 = wrapper::vdup_n(static_cast<T>(0.70710678118f), ExactTagType{});
+#endif /* __aarch64__ */
+    const auto      const_1           = wrapper::vdup_n(static_cast<T>(1.f), ExactTagType{});
+    const auto      const_0           = wrapper::vdup_n(static_cast<T>(0.f), ExactTagType{});
+    const auto      const_6           = wrapper::vdup_n(static_cast<T>(6.f), ExactTagType{});
+    const auto      const_3           = wrapper::vdup_n(static_cast<T>(3.f), ExactTagType{});
+    const auto      const_inv_6       = wrapper::vdup_n(static_cast<T>(0.166666667f), ExactTagType{});
+    constexpr float soft_relu_thresh  = 12.f;
+    const auto      vsoft_relu_thresh = wrapper::vdup_n(static_cast<T>(soft_relu_thresh), ExactTagType{});
+    const auto      va                = wrapper::vdup_n(static_cast<T>(act_info.a()), ExactTagType{});
+    const auto      vb                = wrapper::vdup_n(static_cast<T>(act_info.b()), ExactTagType{});
+    const auto      a                 = static_cast<T>(act_info.a());
+    const auto      b                 = static_cast<T>(act_info.b());
+
+    switch (act)
+    {
+        case ActivationLayerInfo::ActivationFunction::ABS:
+            fn([](auto vin) { return wrapper::vabs(vin); }, [](auto in) { return std::abs(in); });
+            break;
+        case ActivationLayerInfo::ActivationFunction::LINEAR:
+            fn([&](auto vin) { return wrapper::vmla(vb, va, vin); }, [&](auto in) { return a * in + b; });
+            break;
+        case ActivationLayerInfo::ActivationFunction::LOGISTIC:
+            fn([&](auto vin) { return wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin)))); },
+               [](auto in) { return static_cast<T>(1) / (static_cast<T>(1) + std::exp(-in)); });
+            break;
+        case ActivationLayerInfo::ActivationFunction::RELU:
+            fn([&](auto vin) { return wrapper::vmax(const_0, vin); },
+               [](auto in) { return std::max<T>(static_cast<T>(0), in); });
+            break;
+        case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+            fn([&](auto vin) { return wrapper::vmin(va, wrapper::vmax(const_0, vin)); },
+               [&](auto in) { return std::min<T>(a, std::max(static_cast<T>(0), in)); });
+            break;
+        case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
+            fn([&](auto vin) { return wrapper::vmin(va, wrapper::vmax(vb, vin)); },
+               [&](auto in) { return std::min<T>(a, std::max<T>(b, in)); });
+            break;
+        case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
+            fn([&](auto vin) { return wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin)); },
+               [&](auto in) { return (in > 0) ? in : a * in; });
+            break;
+        case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
+            fn(
+                [&](auto vin)
+                {
+                    return wrapper::vbsl(wrapper::vcgt(vin, vsoft_relu_thresh), vin,
+                                         wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin))));
+                },
+                [](auto in) { return (in > soft_relu_thresh) ? in : std::log(static_cast<T>(1) + std::exp(in)); });
+            break;
+        case ActivationLayerInfo::ActivationFunction::ELU:
+            fn(
+                [&](auto vin)
+                {
+                    return wrapper::vbsl(wrapper::vcge(vin, const_0), vin,
+                                         wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1)));
+                },
+                [&](auto in) { return (in >= 0) ? in : a * (std::exp(in) - 1); });
+            break;
+        case ActivationLayerInfo::ActivationFunction::SQRT:
+            fn(
+#ifdef __aarch64__
+                [](auto vin) { return wrapper::vsqrt(vin); },
+#else  /* __aarch64__ */
+                [&](auto vin)
+                {
+                    const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0.f, ExactTagType{}));
+                    auto tmp = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask))));
+                    return mask_float_vector(tmp, wrapper::vnot(bitmask));
+                },
+#endif /* __aarch64__ */
+                [](auto in) { return std::sqrt(in); });
+            break;
+        case ActivationLayerInfo::ActivationFunction::SQUARE:
+            fn([](auto vin) { return wrapper::vmul(vin, vin); }, [](auto in) { return in * in; });
+            break;
+        case ActivationLayerInfo::ActivationFunction::TANH:
+            fn([&](auto vin) { return wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin))); },
+               [&](auto in) { return a * std::tanh(b * in); });
+            break;
+        case ActivationLayerInfo::ActivationFunction::IDENTITY:
+            fn([](auto vin) { return vin; }, [](auto in) { return in; });
+            break;
+        case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
+            fn(
+                [&](auto vin)
+                {
+                    return wrapper::vmul(
+                        vin,
+                        wrapper::vmul(const_inv_6,
+                                      wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3)))));
+                },
+                [](auto in) { return in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f); });
+            break;
+        case ActivationLayerInfo::ActivationFunction::SWISH:
+            fn(
+                [&](auto vin)
+                {
+                    return wrapper::vmul(vin, wrapper::vinv(wrapper::vadd(
+                                                  const_1, wrapper::vexpq(wrapper::vneg(wrapper::vmul(va, vin))))));
+                },
+                [&](auto in) { return in / (static_cast<T>(1) + std::exp(-a * in)); });
+            break;
+#ifdef __aarch64__
+        case ActivationLayerInfo::ActivationFunction::GELU:
+            fn(
+                [&](auto vin)
+                {
+                    return wrapper::vmul(
+                        vin,
+                        wrapper::vmul(const_inv_2,
+                                      wrapper::vadd(const_1, wrapper::verf(wrapper::vmul(vin, const_inv_sqrt_2)))));
+                },
+                [](auto in)
+                { return in * static_cast<T>(0.5f * (1.0f + erff(static_cast<float>(in) / 1.41421356237f))); });
+            break;
+#endif /* __aarch64__ */
+        default:
+            ARM_COMPUTE_ERROR("Unsupported activation function");
+    }
+}
+
+template <typename T, const ActFpImplParams &P>
+void fp_neon_activation_impl(const ITensor             *src,
+                             ITensor                   *dst,
+                             const ActivationLayerInfo &act_info,
+                             const Window              &window)
+{
+    /** SIMD vector tag type. */
+    // using ExactTagType =
+    //     typename arm_compute::wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
+    constexpr int                                 window_step_x  = P.step_x;
+    const auto                                    window_start_x = static_cast<int>(window.x().start());
+    const auto                                    window_end_x   = static_cast<int>(window.x().end());
+    const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
+    Window                                        win_collapsed  = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+    // In case of non-aarch64, a small delta value is added to the input
+    // to prevent NAN values caused by zeros in inputs to SQRT.
+    // In case of aarh64, we call vsqrt directly, so we don't use delta.
+    dispatch_fp_neon_activation_function<T, P>(
+        act, act_info,
+        [&](auto activation_op_vec, auto activation_op_tail)
+        {
+            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());
+                    // 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);
+                        wrapper::traits::neon_bitvector_t<T, wrapper::traits::BitWidth::W128> tmp =
+                            activation_op_vec(vin);
+                        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));
+                        *(output_ptr + x) = activation_op_tail(in);
+                    }
+                },
+                input, output);
+        });
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // ACL_SRC_CPU_KERNELS_ACTIVATION_GENERIC_NEON_FP_IMPL_H