diff options
| author | Georgios Pinitas <georgios.pinitas@arm.com> | 2017-06-26 15:54:06 +0100 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-09-17 14:14:20 +0100 |
| commit | ee12254d4aae887ba654aa26bebc2e1330fc0a88 (patch) | |
| tree | b6821c0b718c128a044e27561c9afad55d2e3ee5 /arm_compute/core/NEON/NEMath.inl | |
| parent | 659abc022b42ad8d11f58b3f3b7b4175f3cf1038 (diff) | |
| download | ComputeLibrary-ee12254d4aae887ba654aa26bebc2e1330fc0a88.tar.gz | |
COMPMID-417: Fix negative overflowing in NESoftmaxLayer.
Change-Id: Ic8fa05ac44855f591ae8f5de63204973f5694c83
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/78880
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Diffstat (limited to 'arm_compute/core/NEON/NEMath.inl')
| -rw-r--r-- | arm_compute/core/NEON/NEMath.inl | 9 |
1 files changed, 6 insertions, 3 deletions
diff --git a/arm_compute/core/NEON/NEMath.inl b/arm_compute/core/NEON/NEMath.inl index a31a4c0dc5..9a49493cf6 100644 --- a/arm_compute/core/NEON/NEMath.inl +++ b/arm_compute/core/NEON/NEMath.inl @@ -85,8 +85,10 @@ inline float32x4_t vtaylor_polyq_f32(float32x4_t x, const std::array<float32x4_t inline float32x4_t vexpq_f32(float32x4_t x) { - static const float32x4_t CONST_LN2 = vdupq_n_f32(0.6931471805f); // ln(2) - static const float32x4_t CONST_INV_LN2 = vdupq_n_f32(1.4426950408f); // 1/ln(2) + static const float32x4_t CONST_LN2 = vdupq_n_f32(0.6931471805f); // ln(2) + static const float32x4_t CONST_INV_LN2 = vdupq_n_f32(1.4426950408f); // 1/ln(2) + static const float32x4_t CONST_0 = vdupq_n_f32(0.f); + static const int32x4_t CONST_NEGATIVE_126 = vdupq_n_s32(-126); // Perform range reduction [-log(2),log(2)] int32x4_t m = vcvtq_s32_f32(vmulq_f32(x, CONST_INV_LN2)); @@ -96,7 +98,8 @@ inline float32x4_t vexpq_f32(float32x4_t x) float32x4_t poly = vtaylor_polyq_f32(val, exp_tab); // Reconstruct - poly = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(poly), vshlq_n_s32(m, 23))); + poly = vreinterpretq_f32_s32(vqaddq_s32(vreinterpretq_s32_f32(poly), vqshlq_n_s32(m, 23))); + poly = vbslq_f32(vcltq_s32(m, CONST_NEGATIVE_126), CONST_0, poly); return poly; } |