From 1c948d47f55ff8a39aa527f63ea7df93a13dd38e Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Tue, 20 Nov 2018 16:03:01 +0000 Subject: COMPMID-1800: (Nightly) Mismatches in SC9863 board for NEON FP16 Fixes for: - ReduceMean, reduction on the X axis for FP16 with 8 elements was performed only up to a certain point. The fix now takes into account the number of elements of the vector and does as many reductions as necessary. - YOLOLayer, activation for FP16 has to be performed on 32 bits until the FP16 approximations is fixed. Change-Id: I75373f4edd37de476e6fe1a56de3ef386b65c619 --- arm_compute/core/NEON/NEMath.inl | 35 +++++------------------------------ 1 file changed, 5 insertions(+), 30 deletions(-) (limited to 'arm_compute') diff --git a/arm_compute/core/NEON/NEMath.inl b/arm_compute/core/NEON/NEMath.inl index 2bc1ab7964..4de80509f0 100644 --- a/arm_compute/core/NEON/NEMath.inl +++ b/arm_compute/core/NEON/NEMath.inl @@ -245,37 +245,12 @@ inline float16x8_t vtaylor_polyq_f16(float16x8_t x, const std::array exp_tab_f16 = - { - { - vdupq_n_f16(1.f), - vdupq_n_f16(0.0416598916054f), - vdupq_n_f16(0.500000596046f), - vdupq_n_f16(0.0014122662833f), - vdupq_n_f16(1.00000011921f), - vdupq_n_f16(0.00833693705499f), - vdupq_n_f16(0.166665703058f), - vdupq_n_f16(0.000195780929062f), - } - }; - - static const float16x8_t CONST_LN2 = vdupq_n_f16(0.6931471805f); // ln(2) - static const float16x8_t CONST_INV_LN2 = vdupq_n_f16(1.4426950408f); // 1/ln(2) - static const float16x8_t CONST_0 = vdupq_n_f16(0.f); - static const int16x8_t CONST_NEGATIVE_126 = vdupq_n_s16(-126); - - // Perform range reduction [-log(2),log(2)] - const int16x8_t m = vcvtq_s16_f16(vmulq_f16(x, CONST_INV_LN2)); - const float16x8_t val = vsubq_f16(x, 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)); - // Polynomial Approximation - float16x8_t poly = vtaylor_polyq_f16(val, exp_tab_f16); - - // Reconstruct - poly = vreinterpretq_f16_s16(vqaddq_s16(vreinterpretq_s16_f16(poly), vqshlq_n_s16(m, 9))); - poly = vbslq_f16(vcltq_s16(m, CONST_NEGATIVE_126), CONST_0, poly); - - return poly; + const float16x8_t res = vcvt_high_f16_f32(vcvt_f16_f32(vexpq_f32(x_low)), vexpq_f32(x_high)); + return res; } inline float16x8_t vlogq_f16(float16x8_t x) -- cgit v1.2.1