diff options
| author | Murray Kornelsen <murray.kornelsen@mail.mcgill.ca> | 2022-07-13 21:22:39 -0400 |
|---|---|---|
| committer | Pablo Marquez Tello <pablo.tello@arm.com> | 2022-09-14 09:15:03 +0000 |
| commit | 926f502ca731fa49bcdf949408ce25728616e5f2 (patch) | |
| tree | 7e221103a9c0c5c0e4c054abc07cbdf11c7c7b4e /src/core/NEON/NEMath.h | |
| parent | 6e09e1404c635d948cf20eb6b4b5747dfb6656f2 (diff) | |
| download | ComputeLibrary-926f502ca731fa49bcdf949408ce25728616e5f2.tar.gz | |
Adding GELU activation
OpenCL implementation uses built in erf.
NEON implementation requires new vectorized erf.
Uses the following approximation:
erf(x) = 1 - 1 / (1 + a1x + a2x^2 + a3x^3 + a4x^4)^4
a1 = 0.278393, a2 = 0.230389, a3 = 0.000972, a4 = 0.078108
From https://en.wikipedia.org/wiki/Error_function#Numerical_approximations
Signed-off-by: Murray Kornelsen <murray.kornelsen@mail.mcgill.ca>
Change-Id: I2d3964b2c26a4334166b17135f9104bc6324fad2
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7921
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Pablo Marquez Tello <pablo.tello@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/NEMath.h')
| -rw-r--r-- | src/core/NEON/NEMath.h | 18 |
1 files changed, 17 insertions, 1 deletions
diff --git a/src/core/NEON/NEMath.h b/src/core/NEON/NEMath.h index 8118c4701f..9e81c38ad8 100644 --- a/src/core/NEON/NEMath.h +++ b/src/core/NEON/NEMath.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2021 Arm Limited. + * Copyright (c) 2016-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -94,6 +94,14 @@ float32x4_t vtaylor_polyq_f32(float32x4_t x, const std::array<float32x4_t, 8> &c */ float32x4_t vexpq_f32(float32x4_t x); +/** Calculate error function + * + * @param[in] x Input vector in F32 format. + * + * @return The calculated erf. + */ +float32x4_t verfq_f32(float32x4_t x); + /** Calculate logarithm * * @param[in] x Input vector value in F32 format. @@ -308,6 +316,14 @@ float16x8_t vinvsqrtq_f16(float16x8_t x); */ float16x8_t vexpq_f16(float16x8_t x); +/** Calculate error function + * + * @param[in] x Input vector in F16 format. + * + * @return The calculated erf. + */ +float16x8_t verfq_f16(float16x8_t x); + /** Calculate n power of a number. * * pow(x,n) = e^(n*log(x)) |