xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/avx2-rr1-p5.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <immintrin.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16 
17 
xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x16(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x16(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24   assert(n % sizeof(float) == 0);
25 
26   const __m256 vsign_mask = _mm256_load_ps(params->avx2_rr1_p5.sign_mask);
27   const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p5.magic_bias);
28   const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p5.log2e);
29   const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p5.minus_ln2);
30   const __m256 vc5 = _mm256_load_ps(params->avx2_rr1_p5.c5);
31   const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_p5.c4);
32   const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p5.c3);
33   const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p5.c2);
34   const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p5.c1);
35   const __m256 vone = _mm256_load_ps(params->avx2_rr1_p5.one);
36   const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p5.denorm_cutoff);
37 
38   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
39     const __m256 vx0 = _mm256_loadu_ps(x);
40     const __m256 vx1 = _mm256_loadu_ps(x + 8);
41     x += 16;
42 
43     const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
44     const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
45 
46     __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
47     __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
48 
49     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
50     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
51 
52     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
53     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
54 
55     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
56     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
57 
58     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
59     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
60 
61     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
62     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
63 
64     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
65     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
66 
67     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
68     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
69 
70     vt0 = _mm256_mul_ps(vt0, vs0);
71     vt1 = _mm256_mul_ps(vt1, vs1);
72 
73     const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
74     const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
75 
76     const __m256 vd0 = _mm256_add_ps(ve0, vone);
77     const __m256 vd1 = _mm256_add_ps(ve1, vone);
78 
79     __m256 vr0 = _mm256_rcp_ps(vd0);
80     __m256 vr1 = _mm256_rcp_ps(vd1);
81 
82     vr0 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
83     vr1 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
84 
85     vr0 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
86     vr1 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
87 
88     __m256 vf0 = _mm256_mul_ps(ve0, vr0);
89     __m256 vf1 = _mm256_mul_ps(ve1, vr1);
90 
91     vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
92     vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
93 
94     vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
95     vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
96 
97     _mm256_storeu_ps(y, vf0);
98     _mm256_storeu_ps(y + 8, vf1);
99     y += 16;
100   }
101   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
102     const __m256 vx = _mm256_loadu_ps(x);
103     x += 8;
104 
105     const __m256 vz = _mm256_or_ps(vx, vsign_mask);
106 
107     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
108     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
109     vn = _mm256_sub_ps(vn, vmagic_bias);
110 
111     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
112 
113     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
114     vp = _mm256_fmadd_ps(vp, vt, vc3);
115     vp = _mm256_fmadd_ps(vp, vt, vc2);
116     vp = _mm256_fmadd_ps(vp, vt, vc1);
117 
118     vt = _mm256_mul_ps(vt, vs);
119     const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
120 
121     const __m256 vd = _mm256_add_ps(ve, vone);
122     __m256 vr = _mm256_rcp_ps(vd);
123     vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
124     vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
125     __m256 vf = _mm256_mul_ps(ve, vr);
126 
127     vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
128     vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
129 
130     _mm256_storeu_ps(y, vf);
131     y += 8;
132   }
133   if XNN_UNLIKELY(n != 0) {
134     assert(n >= 1 * sizeof(float));
135     assert(n <= 7 * sizeof(float));
136     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx2_rr1_p5.mask_table[7] - n));
137 
138     const __m256 vx = _mm256_maskload_ps(x, vmask);
139 
140     const __m256 vz = _mm256_or_ps(vx, vsign_mask);
141 
142     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
143     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
144     vn = _mm256_sub_ps(vn, vmagic_bias);
145 
146     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
147 
148     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
149     vp = _mm256_fmadd_ps(vp, vt, vc3);
150     vp = _mm256_fmadd_ps(vp, vt, vc2);
151     vp = _mm256_fmadd_ps(vp, vt, vc1);
152 
153     vt = _mm256_mul_ps(vt, vs);
154     const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
155 
156     const __m256 vd = _mm256_add_ps(ve, vone);
157     __m256 vr = _mm256_rcp_ps(vd);
158     vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
159     vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
160     __m256 vf = _mm256_mul_ps(ve, vr);
161 
162     vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
163     vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
164 
165     __m128 vf_lo = _mm256_castps256_ps128(vf);
166     if (n & (4 * sizeof(float))) {
167       _mm_storeu_ps(y, vf_lo);
168       vf_lo = _mm256_extractf128_ps(vf, 1);
169       y += 4;
170     }
171     if (n & (2 * sizeof(float))) {
172       _mm_storel_pi((__m64*) y, vf_lo);
173       vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
174       y += 2;
175     }
176     if (n & (1 * sizeof(float))) {
177       _mm_store_ss(y, vf_lo);
178     }
179   }
180 }
181