diff options
| author | morgolock <pablo.tello@arm.com> | 2020-05-11 16:00:04 +0100 |
|---|---|---|
| committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2020-05-15 11:24:27 +0000 |
| commit | 3155f77d23eb332974a6b9cb5d37a1329b8493a6 (patch) | |
| tree | a581a882a677e8d39177c84f1f321c9ff27707ad | |
| parent | 090502887d87f52d28e98e90c0e17c582b9e63d6 (diff) | |
| download | armcl-3155f77d23eb332974a6b9cb5d37a1329b8493a6.tar.gz armcl-3155f77d23eb332974a6b9cb5d37a1329b8493a6.tar.bz2 armcl-3155f77d23eb332974a6b9cb5d37a1329b8493a6.zip | |
COMPMID-3457: vexp failure in QASYMM8_SIGNED overflowing
Change-Id: Ied11a4a3e9d04615a1a1f0bfa552f3dd8293a170
Signed-off-by: morgolock <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3178
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
| -rw-r--r-- | arm_compute/core/NEON/NEMath.h | 1 | ||||
| -rw-r--r-- | arm_compute/core/NEON/NEMath.inl | 6 |
2 files changed, 6 insertions, 1 deletions
diff --git a/arm_compute/core/NEON/NEMath.h b/arm_compute/core/NEON/NEMath.h index 3905f67e2..8827bbf45 100644 --- a/arm_compute/core/NEON/NEMath.h +++ b/arm_compute/core/NEON/NEMath.h @@ -25,6 +25,7 @@ #define ARM_COMPUTE_NEMATH_H #include <arm_neon.h> +#include <array> namespace arm_compute { diff --git a/arm_compute/core/NEON/NEMath.inl b/arm_compute/core/NEON/NEMath.inl index 49870d06a..032bfde23 100644 --- a/arm_compute/core/NEON/NEMath.inl +++ b/arm_compute/core/NEON/NEMath.inl @@ -22,6 +22,7 @@ * SOFTWARE. */ #include <cmath> +#include <limits> #ifndef M_PI #define M_PI (3.14159265358979323846) @@ -147,6 +148,8 @@ 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_INF = vdupq_n_f32(std::numeric_limits<float>::infinity()); + static const float32x4_t CONST_MAX_INPUT = vdupq_n_f32(88.7f); static const float32x4_t CONST_0 = vdupq_n_f32(0.f); static const int32x4_t CONST_NEGATIVE_126 = vdupq_n_s32(-126); @@ -159,7 +162,8 @@ inline float32x4_t vexpq_f32(float32x4_t x) // Reconstruct 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); + poly = vbslq_f32(vcltq_s32(m, CONST_NEGATIVE_126), CONST_0, poly); // Handle underflow + poly = vbslq_f32(vcgtq_f32(x, CONST_MAX_INPUT), CONST_INF, poly); // Handle overflow return poly; } |