diff options
| author | Georgios Pinitas <georgios.pinitas@arm.com> | 2018-12-03 14:30:05 +0000 |
|---|---|---|
| committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2019-01-14 17:53:22 +0000 |
| commit | 5a5945387e70f62e6e1e95a177fae261d7570443 (patch) | |
| tree | ff8bd61c2e071b5a0b923f4a0d1bef72486435e9 /arm_compute/core/NEON/NEMath.inl | |
| parent | dea2d2d58fe3a742e6f66fe50befbe0044e15ad1 (diff) | |
| download | ComputeLibrary-5a5945387e70f62e6e1e95a177fae261d7570443.tar.gz | |
COMPMID-1809: Remove padding in NEGEMMConvolutionLayer 64-bit path.
Change-Id: I1806591a2c73a1f057f13d8c6107d7b9796a82c8
Reviewed-on: https://review.mlplatform.org/370
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Diffstat (limited to 'arm_compute/core/NEON/NEMath.inl')
| -rw-r--r-- | arm_compute/core/NEON/NEMath.inl | 34 |
1 files changed, 6 insertions, 28 deletions
diff --git a/arm_compute/core/NEON/NEMath.inl b/arm_compute/core/NEON/NEMath.inl index 4de80509f0..27b4fc2c1b 100644 --- a/arm_compute/core/NEON/NEMath.inl +++ b/arm_compute/core/NEON/NEMath.inl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2018 ARM Limited. + * Copyright (c) 2016-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -255,34 +255,12 @@ inline float16x8_t vexpq_f16(float16x8_t x) inline float16x8_t vlogq_f16(float16x8_t x) { - static const std::array<float16x8_t, 8> log_tab_f16 = - { - { - vdupq_n_f16(-2.29561495781f), - vdupq_n_f16(-2.47071170807f), - vdupq_n_f16(-5.68692588806f), - vdupq_n_f16(-0.165253549814f), - vdupq_n_f16(5.17591238022f), - vdupq_n_f16(0.844007015228f), - vdupq_n_f16(4.58445882797f), - vdupq_n_f16(0.0141278216615f), - } - }; - - static const int16x8_t CONST_127 = vdupq_n_s16(127); // 127 - static const float16x8_t CONST_LN2 = vdupq_n_f16(0.6931471805f); // ln(2) - - // Extract exponent - const int16x8_t m = vsubq_s16(vreinterpretq_s16_u16(vshrq_n_u16(vreinterpretq_u16_f16(x), 9)), CONST_127); - const float16x8_t val = vreinterpretq_f16_s16(vsubq_s16(vreinterpretq_s16_f16(x), vshlq_n_s16(m, 9))); - - // Polynomial Approximation - float16x8_t poly = vtaylor_polyq_f16(val, log_tab_f16); - - // Reconstruct - poly = vaddq_f16(poly, vmulq_f16(vcvtq_f16_s16(m), CONST_LN2)); + // TODO (COMPMID-1535) : Revisit FP16 approximations + const float32x4_t x_high = vcvt_f32_f16(vget_high_f16(x)); + const float32x4_t x_low = vcvt_f32_f16(vget_low_f16(x)); - return poly; + const float16x8_t res = vcvt_high_f16_f32(vcvt_f16_f32(vlogq_f32(x_low)), vlogq_f32(x_high)); + return res; } inline float16x8_t vpowq_f16(float16x8_t val, float16x8_t n) |