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 <wasm_simd128.h>
10
11 #include <xnnpack/common.h>
12 #include <xnnpack/math-stubs.h>
13
14
xnn_math_f32_sigmoid__wasmsimd_rr2_p5_div(size_t n,const float * input,float * output)15 void xnn_math_f32_sigmoid__wasmsimd_rr2_p5_div(
16 size_t n,
17 const float* input,
18 float* output)
19 {
20 assert(n % (4 * sizeof(float)) == 0);
21
22 // Large number such that ulp(magic bias) == 1 and magic bias === 127 mod 2**22.
23 const v128_t vmagic_bias = wasm_f32x4_splat(0x1.8000FEp23f);
24 const v128_t vminus_log2e = wasm_f32x4_splat(-0x1.715476p+0f);
25 // Last 7 bits are zeroes
26 const v128_t vln2_hi = wasm_f32x4_splat(0x1.62E400p-1f);
27 const v128_t vln2_lo = wasm_f32x4_splat(0x1.7F7D1Cp-20f);
28 // Coefficient of polynomial approximation of
29 // exp(-t) ~ 1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) on [-log(2)/2, log(2)/2]
30 const v128_t vc5 = wasm_f32x4_splat(-0x1.0F9F9Cp-7f);
31 const v128_t vc4 = wasm_f32x4_splat( 0x1.573A1Ap-5f);
32 const v128_t vc3 = wasm_f32x4_splat(-0x1.555A80p-3f);
33 const v128_t vc2 = wasm_f32x4_splat( 0x1.FFFDC6p-2f);
34 const v128_t vc1 = wasm_f32x4_splat(-0x1.FFFFF6p-1f);
35 const v128_t vone = wasm_f32x4_splat(1.0f);
36 // The largest z for which sigmoidf(-z) is normalized.
37 // This number is also the largest z for which expf(-z) is normalized.
38 const v128_t vdenorm_cutoff = wasm_f32x4_splat(0x1.5D589Ep+6f);
39
40 for (; n != 0; n -= 4 * sizeof(float)) {
41 const v128_t vx = wasm_v128_load(input);
42 input += 4;
43
44 // General structure of the algorithm:
45 //
46 // / exp(x) / (1 + exp(x)) if x <= 0
47 // f[x] :=
48 // \ 1 - f[-x] if x >= 0
49 //
50 // First we compute f[-z] := exp(-z) / (1 + exp(-z)) where z = abs(x),
51 // then replace result with 1 - f[-z] if x >= 0.
52 const v128_t vz = wasm_f32x4_abs(vx);
53
54 // Compute reduced argument n := round(-z / log(2)).
55 // We do it by adding a large number (magic bias), which cause rounding of the result to integer, then subtracing
56 // the large number back. The trick with adding large number is valid only within certain bounds
57 // (|-z / log(2)| <= 2**22, i.e. |z| <= 0x1.62E43p+21 = 2907270.0), but that is acceptable, because inputs x
58 // outside of [-87.336544, 17.328678] (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup
59 // the result for such inputs at the very end of the algorithm.
60 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
61
62 // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
63 // -87.336544 <= -z <= 0.0, and -126 <= n <= 0 accordingly.
64 const v128_t vs = wasm_i32x4_shl(vn, 23);
65
66 // Subtract the large number back to get the final n := round(-z / log(2)) as a floating-point number.
67 vn = wasm_f32x4_sub(vn, vmagic_bias);
68
69 // Compute reduced argument t := z + n * log(2). Note that -t = -z - n * log(2).
70 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
71 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
72 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
73
74 // Compute degree-5 polynomial approximation for exp(-t) on [-log(2)/2, log(2)/2]:
75 // P(t) = 1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) = 1 + t * p
76 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt, vc5));
77 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt, vp));
78 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt, vp));
79 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt, vp));
80
81 // Reconstruct the exp(-z) value:
82 // e = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
83 // = s * (1 + t * p)
84 // = s + (t * s) * p
85 vt = wasm_f32x4_mul(vt, vs);
86 const v128_t ve = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
87
88 // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
89 v128_t vf = wasm_f32x4_div(ve, wasm_f32x4_add(ve, vone));
90
91 // For inputs below denormal cutoff, replace output with +0.0f.
92 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
93 vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
94
95 // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
96 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
97
98 wasm_v128_store(output, vf);
99 output += 4;
100 }
101 }
102