xref: /aosp_15_r20/external/XNNPACK/src/math/sigmoid-f32-avx-rr2-lut64-p2-div.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Copyright 2020 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 <stddef.h>
8 
9 #include <immintrin.h>
10 
11 #include <xnnpack/common.h>
12 #include <xnnpack/math-stubs.h>
13 
14 
15 // Table of exp2(k / 64) values decremented (as integer) by (k << 17), k = 0..63
16 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64];
17 
xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div(size_t n,const float * input,float * output)18 void xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div(
19     size_t n,
20     const float* input,
21     float* output)
22 {
23   assert(n % (8 * sizeof(float)) == 0);
24 
25   // Floating-point mask with only the sign bit set
26   const __m256 vsign_mask = _mm256_set1_ps(-0.0f);
27   // Large number such that ulp(magic bias) == exp2(-6)
28   const __m256 vmagic_bias = _mm256_set1_ps(0x1.800000p17f);
29   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p0f);
30   // Mask for the lowest 6 bits
31   const __m256 vindex_mask = _mm256_castsi256_ps(_mm256_set1_epi32(INT32_C(0x3F)));
32   // Last 13 bits are zeroes
33   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.630000p-1f);
34   const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.BD0106p-13f);
35   // Coefficient of polynomial approximation of exp(t) ~ 1 + t * (1 + t * c2) on [-log(2)/128, log(2)/128]
36   const __m256 vc2 = _mm256_set1_ps(0x1.FFFF0Ap-2f);
37   const __m256 vone = _mm256_set1_ps(1.0f);
38   // The smallest x for which sigmoidf(x) is normalized.
39   // This number is also the smallest x for which expf(x) is normalized.
40   const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);
41 
42   for (; n != 0; n -= 8 * sizeof(float)) {
43     const __m256 vx = _mm256_loadu_ps(input);
44 
45     // General structure of the algorithm:
46     //
47     //           / exp(x) / (1 + exp(x)) if x <= 0
48     //   f[x] :=
49     //           \ 1 - f[-x] if x >= 0
50     //
51     // First we compute f[z] := exp(z) / (1 + exp(z)) where z = -abs(x), then replace result with 1 - f[z] if x >= 0.
52     const __m256 vz = _mm256_or_ps(vx, vsign_mask);
53 
54     // Compute reduced argument n := round(z / log(2), 6).
55     // We do it by adding a large number (magic bias), which cause rounding of the result to 6 fractional bits, then
56     // subtracing the large number back. The trick with adding large number is valid only within certain bounds
57     // (|z / log(2)| <= 2**16, i.e. |z| <= 0x1.62E43p+15 = 45426.09375), but that is acceptable, because inputs x
58     // outside of [-87.336544, 17.328678] (i.e. z outsize [87.336544, 0]) underflow or saturate sigmoidf(x). We fixup
59     // the result  for such inputs at the very end of the algorithm.
60     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
61 
62     // Create a floating-point number s (scale) such that s := 2**n for such inputs that sigmoidf(z) is normalized,
63     // i.e. -87.33642 <= z <= 0. As n has 6 fractional bits, we split s == 2**n = 2**int(n) * 2**frac(n). We create s
64     // in two steps:
65     // 1. Fetch 2**frac(n) from the table using the 6 low bits of n, as integer. Note that the fetched values are in
66     //    the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
67     // 2. Adjust fecthed value by addition of int(n) to its floating-point exponent. The result is always a normalized
68     //    number, because for -87.33642 <= z <= 0 (inputs for which sigmoidf(z) is normalized) we have
69     //    -126 <= int(n) <= 0, and thus the adjusted exponent is not lower than -126.
70     //
71     // Shift bits 6:14 into 23:31 (position of floating-point exponent).
72     __m128i ve_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 17);
73     __m128i ve_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 17);
74 
75     // Use bits 0:6 of n, as integer, as an index for table lookup of l := 2**frac(n).
76     const __m256 vidx = _mm256_and_ps(vn, vindex_mask);
77     const __m128i vidx_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx)), 2);
78     const __m128i vidx_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vidx, 1)), 2);
79 #if XNN_ARCH_X86_64
80     const uint64_t vidx_ll = (uint64_t) _mm_cvtsi128_si64(vidx_lo);
81     const uint64_t vidx_lh = (uint64_t) _mm_extract_epi64(vidx_lo, 1);
82     const uint64_t vidx_hl = (uint64_t) _mm_cvtsi128_si64(vidx_hi);
83     const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1);
84     __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_ll)));
85     __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lh)));
86     __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hl)));
87     __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hh)));
88     vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_ll >> 32))), 1);
89     vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lh >> 32))), 1);
90     vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hl >> 32))), 1);
91     vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hh >> 32))), 1);
92 #else
93     __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx_lo))));
94     __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_lo, 2))));
95     __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx_hi))));
96     __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_hi, 2))));
97     vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))), 1);
98     vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))), 1);
99     vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_hi, 1))), 1);
100     vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_hi, 3))), 1);
101 #endif
102     const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh);
103     const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh);
104     // Adjust exponent of the value l fetched from the table to get the final s value.
105     const __m128 vs_lo = _mm_castsi128_ps(_mm_add_epi32(vl_lo, ve_lo));
106     const __m128 vs_hi = _mm_castsi128_ps(_mm_add_epi32(vl_hi, ve_hi));
107     const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
108 
109     // Subtract the large number back to get the final n := round(z / log(2), 6) as a floating-point number.
110     vn = _mm256_sub_ps(vn, vmagic_bias);
111 
112     // Compute reduced argument t := z - n * log(2).
113     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
114     __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
115     vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
116 
117     // Compute degree-2 polynomial approximation for exp(t) on [-log(2)/128, log(2)/128].
118     //   P(t) = 1 + t * (1 + t * c2) = 1 + (t + t * (t * c2)) = 1 + p
119     __m256 vp = _mm256_mul_ps(vt, vc2);
120     vp = _mm256_add_ps(vt, _mm256_mul_ps(vp, vt));
121 
122     // Reconstruct the exp(z) value:
123     //   e = s * (1 + t * (1 + t * c2))
124     //     = s * (1 + p)
125     //     = s + s * p
126     const __m256 ve = _mm256_add_ps(vs, _mm256_mul_ps(vs, vp));
127 
128     // Denominator of the sigmoid fraction: 1.0 + exp(z)
129     const __m256 vd = _mm256_add_ps(ve, vone);
130 
131     // Reconstruct sigmoid(z) = exp(z) / (1.0 + exp(z))
132     __m256 vf = _mm256_div_ps(ve, vd);
133 
134     // For inputs below denormal cutoff, replace output with +0.0f.
135     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
136     vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
137 
138     // Reconstruct sigmoid(x) = x < 0 ? sigmoid(z) : 1.0 - sigmoid(z)
139     vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
140 
141     _mm256_storeu_ps(output, vf);
142 
143     input += 8;
144     output += 8;
145   }
146 }
147