Marat Dukhan | ffd6840 | 2019-11-15 15:19:11 -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 <emmintrin.h> |
| 11 | |
| 12 | #include <xnnpack/math-stubs.h> |
| 13 | |
| 14 | |
| 15 | void xnn_math_f32_exp__sse2_p5( |
| 16 | size_t n, |
| 17 | const float* input, |
| 18 | float* output) |
| 19 | { |
| 20 | assert(n % (8 * sizeof(float)) == 0); |
| 21 | |
| 22 | const __m128 vmagic_bias = _mm_set1_ps(0x1.800000p+23f); |
| 23 | // The smallest x for which expf(x) is non-zero. |
| 24 | const __m128 vzero_cutoff = _mm_set1_ps(-0x1.9FE368p+6f); |
| 25 | // The largest x for which expf(x) is finite. |
| 26 | const __m128 vinf_cutoff = _mm_set1_ps(0x1.62E42Ep+6f); |
| 27 | const __m128 vlog2e = _mm_set1_ps(0x1.715476p+0f); |
| 28 | // Last 8 bits are zeroes |
| 29 | const __m128 vminus_ln2_hi = _mm_set1_ps(-0x1.62E400p-1f); |
| 30 | const __m128 vminus_ln2_lo = _mm_set1_ps(-0x1.7F7D1Cp-20f); |
| 31 | const __m128 vplus_inf = _mm_set1_ps(INFINITY); |
| 32 | |
| 33 | const __m128 vc1 = _mm_set1_ps(0x1.FFFFF6p-1f); |
| 34 | const __m128 vc2 = _mm_set1_ps(0x1.FFFDC6p-2f); |
| 35 | const __m128 vc3 = _mm_set1_ps(0x1.555A80p-3f); |
| 36 | const __m128 vc4 = _mm_set1_ps(0x1.573A1Ap-5f); |
| 37 | const __m128 vc5 = _mm_set1_ps(0x1.0F9F9Cp-7f); |
| 38 | |
| 39 | const __m128i vmin_exponent = _mm_set1_epi32(0xC1000000); |
| 40 | const __m128i vmax_exponent = _mm_set1_epi32(0x3F800000); |
| 41 | const __m128i vdefault_exponent = vmax_exponent; |
| 42 | |
| 43 | for (; n != 0; n -= 4 * sizeof(float)) { |
| 44 | const __m128 vx = _mm_loadu_ps(input); |
| 45 | |
| 46 | // Compute reduced argument n := round(x / log(2)). |
| 47 | // We do it by adding a large number (magic bias) to the product x * (1/log(2)), which cause rounding of the result |
| 48 | // to an integer, then subtracing the large number back. The trick with adding large number is valid only within |
| 49 | // certain bounds (|x| <= 2**22), but thats ok, because inputs outside of [-103.97207, 88.72283] underflow or |
| 50 | // overflow expf(x) anyway. We fixup the result for such inputs at the very end of the algorithm. |
| 51 | __m128 vn = _mm_add_ps(_mm_mul_ps(vx, vlog2e), vmagic_bias); |
| 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 | __m128i veo = _mm_slli_epi32(_mm_castps_si128(vn), 23); |
| 61 | __m128i ven = _mm_max_epi16(veo, vmin_exponent); |
| 62 | ven = _mm_min_epi16(ven, vmax_exponent); |
| 63 | veo = _mm_sub_epi32(veo, ven); |
| 64 | const __m128 vsn = _mm_castsi128_ps(_mm_add_epi32(ven, vdefault_exponent)); |
| 65 | const __m128 vso = _mm_castsi128_ps(_mm_add_epi32(veo, vdefault_exponent)); |
| 66 | |
| 67 | // Subtract the large number back to get final n := round(x / log(2)). |
| 68 | vn = _mm_sub_ps(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 | __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vx); |
| 73 | vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt); |
| 74 | |
| 75 | // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2]. |
| 76 | __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4); |
| 77 | vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3); |
| 78 | vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2); |
| 79 | vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1); |
| 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 = _mm_mul_ps(vt, vso); |
| 86 | __m128 vf = _mm_mul_ps(vsn, _mm_add_ps(_mm_mul_ps(vt, vp), vso)); |
| 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 = _mm_andnot_ps(_mm_cmplt_ps(vx, vzero_cutoff), vf); |
| 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 | const __m128 vm = _mm_cmpgt_ps(vx, vinf_cutoff); |
| 94 | vf = _mm_or_ps(_mm_and_ps(vplus_inf, vm), _mm_andnot_ps(vm, vf)); |
| 95 | _mm_storeu_ps(output, vf); |
| 96 | |
| 97 | input += 4; |
| 98 | output += 4; |
| 99 | } |
| 100 | } |