Marat Dukhan | 797a8fe | 2019-11-14 20:21:57 -0800 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <assert.h> |
| 7 | #include <math.h> |
| 8 | #include <stddef.h> |
| 9 | |
| 10 | #include <arm_neon.h> |
| 11 | |
| 12 | #include <xnnpack/math-stubs.h> |
| 13 | |
| 14 | |
| 15 | void xnn_math_f32_exp__neonfma_p5( |
| 16 | size_t n, |
| 17 | const float* input, |
| 18 | float* output) |
| 19 | { |
| 20 | assert(n % (4 * sizeof(float)) == 0); |
| 21 | |
| 22 | const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p+23f); |
| 23 | // The smallest x for which expf(x) is non-zero. |
| 24 | const float32x4_t vzero_cutoff = vmovq_n_f32(-0x1.9FE368p+6f); |
| 25 | // The largest x for which expf(x) is finite. |
| 26 | const float32x4_t vinf_cutoff = vmovq_n_f32(0x1.62E42Ep+6f); |
| 27 | const float32x4_t vlog2e = vmovq_n_f32(0x1.715476p+0f); |
| 28 | const float32x4_t vminus_ln2_hi = vmovq_n_f32(-0x1.62E43p-1f); |
| 29 | const float32x4_t vminus_ln2_lo = vmovq_n_f32(0x1.05C61p-29f); |
| 30 | const float32x4_t vplus_inf = vmovq_n_f32(INFINITY); |
| 31 | |
| 32 | const float32x4_t vc1 = vmovq_n_f32(0x1.FFFFF6p-1f); |
| 33 | const float32x4_t vc2 = vmovq_n_f32(0x1.FFFDC6p-2f); |
| 34 | const float32x4_t vc3 = vmovq_n_f32(0x1.555A80p-3f); |
| 35 | const float32x4_t vc4 = vmovq_n_f32(0x1.573A1Ap-5f); |
| 36 | const float32x4_t vc5 = vmovq_n_f32(0x1.0F9F9Cp-7f); |
| 37 | |
| 38 | const int32x4_t vmin_exponent = vmovq_n_s32(INT32_C(0xC1000000)); |
| 39 | const int32x4_t vmax_exponent = vmovq_n_s32(INT32_C(0x3F800000)); |
| 40 | const int32x4_t vdefault_exponent = vmax_exponent; |
| 41 | |
| 42 | for (; n != 0; n -= 4 * sizeof(float)) { |
| 43 | const float32x4_t vx = vld1q_f32(input); input += 4; |
| 44 | |
| 45 | // Compute reduced argument n := round(x / log(2)). |
| 46 | // We do it by adding a large number (magic bias), which cause rounding of result to an integer, then subtracing the |
| 47 | // large number back. The first addition is combined with multiplication by log2e into a single FMA instruction. |
| 48 | // The trick with adding large number is valid only within certain bounds (|x| <= 2**22), but thats ok, because |
| 49 | // inputs outside of [-103.97207, 88.72283] underflow or overflow expf(x) anyway. We fixup the result for such |
| 50 | // inputs at the very end of the algorithm. |
| 51 | float32x4_t vn = vfmaq_f32(vmagic_bias, vx, vlog2e); |
| 52 | |
| 53 | // Create two floating-point numbers, sn (scale, normal) and so (scale, overflow) such that sn * so == 2**n |
| 54 | // for inputs which don't cause overflow, i.e. -103.97207 <= x <= 88.72283, and -150 <= n <= 128 accordingly. |
| 55 | // We need to use two numbers rather than one because a normalized single-precision exponent must be in [-127, 126] |
| 56 | // range, which is insufficient to cover [-150, 128] range of n. |
| 57 | // - When n is within [-127, 126], sn == 2**n and so == 1.0. |
| 58 | // - When n < -127, sn == 2**(-127) and so == 2**(n + 127). |
| 59 | // - When n > 126, sn == 2**126 and so == 2**(n - 126). |
| 60 | int32x4_t veo = vshlq_n_s32(vreinterpretq_s32_f32(vn), 23); |
| 61 | int32x4_t ven = vmaxq_s32(veo, vmin_exponent); |
| 62 | ven = vminq_s32(ven, vmax_exponent); |
| 63 | veo = vsubq_s32(veo, ven); |
| 64 | const float32x4_t vsn = vreinterpretq_f32_s32(vaddq_s32(ven, vdefault_exponent)); |
| 65 | const float32x4_t vso = vreinterpretq_f32_s32(vaddq_s32(veo, vdefault_exponent)); |
| 66 | |
| 67 | // Subtract the large number back to get final n := round(x / log(2)). |
| 68 | vn = vsubq_f32(vn, vmagic_bias); |
| 69 | |
| 70 | // Compute reduced argument t := x - n * log(2). |
| 71 | // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 72 | float32x4_t vt = vfmaq_f32(vx, vn, vminus_ln2_hi); |
| 73 | vt = vfmaq_f32(vt, vn, vminus_ln2_lo); |
| 74 | |
| 75 | // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2]. |
| 76 | float32x4_t vp = vfmaq_f32(vc4, vc5, vt); |
| 77 | vp = vfmaq_f32(vc3, vp, vt); |
| 78 | vp = vfmaq_f32(vc2, vp, vt); |
| 79 | vp = vfmaq_f32(vc1, vp, vt); |
| 80 | |
| 81 | // Reconstruct the final f value: |
| 82 | // f = so * sn * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 83 | // = sn * (so + (t * so) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))) |
| 84 | // = sn * (so + (t * so) * p) |
| 85 | vt = vmulq_f32(vt, vso); |
| 86 | float32x4_t vf = vmulq_f32(vsn, vfmaq_f32(vso, vt, vp)); |
| 87 | |
| 88 | // For inputs below zero cutoff, replace output with +0.0f. |
| 89 | // Note that for NaN inputs, comparison result is false, and outputs are left unchanged. |
| 90 | vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vzero_cutoff))); |
| 91 | // For inputs above inf cutoff, replace output with +inf. |
| 92 | // Note that for NaN inputs, comparison result is false, and outputs are left unchanged. |
| 93 | vf = vbslq_f32(vcgtq_f32(vx, vinf_cutoff), vplus_inf, vf); |
| 94 | vst1q_f32(output, vf); output += 4; |
| 95 | } |
| 96 | } |