diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2018-08-23 15:29:16 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:54:54 +0000 |
commit | f2cde9b29deee6423ea6fe9a1a9afc9ef61d2663 (patch) | |
tree | 8a770ff88dc93cee025c44de3c1b8d6f35b9e317 /arm_compute | |
parent | 13d96e061fe3be14f9693e6761f1795a2399b249 (diff) | |
download | ComputeLibrary-f2cde9b29deee6423ea6fe9a1a9afc9ef61d2663.tar.gz |
COMPMID-1534 - Fix NENormalizationLayer for FP16
Implemented vinvq_f16 with fp32 data type in order to avoid accuracy issue.
Change-Id: Ibfffd12e4a941c1388a982fc7bbe3e1832351feb
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/145416
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'arm_compute')
-rw-r--r-- | arm_compute/core/NEON/NEMath.inl | 11 |
1 files changed, 10 insertions, 1 deletions
diff --git a/arm_compute/core/NEON/NEMath.inl b/arm_compute/core/NEON/NEMath.inl index 61d25d115c..1ebc9c10af 100644 --- a/arm_compute/core/NEON/NEMath.inl +++ b/arm_compute/core/NEON/NEMath.inl @@ -303,7 +303,16 @@ inline float16x8_t vlogq_f16(float16x8_t x) inline float16x8_t vpowq_f16(float16x8_t val, float16x8_t n) { - return vexpq_f16(vmulq_f16(n, vlogq_f16(val))); + // TODO (giaiod01) - COMPMID-1535 + float32x4_t n0_f32 = vcvt_f32_f16(vget_low_f16(n)); + float32x4_t n1_f32 = vcvt_f32_f16(vget_high_f16(n)); + float32x4_t val0_f32 = vcvt_f32_f16(vget_low_f16(val)); + float32x4_t val1_f32 = vcvt_f32_f16(vget_high_f16(val)); + + float32x4_t res0_f32 = vexpq_f32(vmulq_f32(n0_f32, vlogq_f32(val0_f32))); + float32x4_t res1_f32 = vexpq_f32(vmulq_f32(n1_f32, vlogq_f32(val1_f32))); + + return vcombine_f16(vcvt_f16_f32(res0_f32), vcvt_f16_f32(res1_f32)); } #endif /* DOXYGEN_SKIP_THIS */ #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |