xref: /aosp_15_r20/external/XNNPACK/src/f16-vsigmoid/gen/vsigmoid-avx2-rr1-p2-rcp-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-vsigmoid/avx2.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2022 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/intrinsics-polyfill.h>
16 #include <xnnpack/vunary.h>
17 
18 
xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x16(size_t batch,const void * input,void * output,const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x16(
20     size_t batch,
21     const void* input,
22     void* output,
23     const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25   assert(batch % sizeof(uint16_t) == 0);
26 
27   const __m256 vsign_mask = _mm256_load_ps(params->avx2_rr1_p2.sign_mask);
28   const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p2.magic_bias);
29   const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p2.log2e);
30   const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p2.minus_ln2);
31   const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p2.c2);
32   const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p2.c1);
33   const __m256 vone = _mm256_load_ps(params->avx2_rr1_p2.one);
34   const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p2.denorm_cutoff);
35 
36   const uint16_t* i = (const uint16_t*) input;
37   uint16_t* o = (uint16_t*) output;
38   for (; batch >= 16 * sizeof(uint16_t); batch -= 16 * sizeof(uint16_t)) {
39     const __m256 vx0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
40     const __m256 vx1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 8)));
41     i += 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     const __m256 vp0 = _mm256_fmadd_ps(vc2, vt0, vc1);
59     const __m256 vp1 = _mm256_fmadd_ps(vc2, vt1, vc1);
60 
61     vt0 = _mm256_mul_ps(vt0, vs0);
62     vt1 = _mm256_mul_ps(vt1, vs1);
63 
64     const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
65     const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
66 
67     const __m256 vd0 = _mm256_add_ps(ve0, vone);
68     const __m256 vd1 = _mm256_add_ps(ve1, vone);
69 
70     const __m256 vr0 = _mm256_rcp_ps(vd0);
71     const __m256 vr1 = _mm256_rcp_ps(vd1);
72 
73     __m256 vf0 = _mm256_mul_ps(ve0, vr0);
74     __m256 vf1 = _mm256_mul_ps(ve1, vr1);
75 
76     vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
77     vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
78 
79     vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
80     vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
81 
82     _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf0, _MM_FROUND_NO_EXC));
83     _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vf1, _MM_FROUND_NO_EXC));
84     o += 16;
85   }
86   for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
87     const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
88     i += 8;
89 
90     const __m256 vz = _mm256_or_ps(vx, vsign_mask);
91 
92     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
93     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
94     vn = _mm256_sub_ps(vn, vmagic_bias);
95 
96     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
97 
98     const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
99     vt = _mm256_mul_ps(vt, vs);
100     const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
101 
102     const __m256 vd = _mm256_add_ps(ve, vone);
103     const __m256 vr = _mm256_rcp_ps(vd);
104     __m256 vf = _mm256_mul_ps(ve, vr);
105 
106     vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
107     vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
108 
109     _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC));
110     o += 8;
111   }
112   if XNN_UNLIKELY(batch != 0) {
113     assert(batch >= 1 * sizeof(uint16_t));
114     assert(batch <= 7 * sizeof(uint16_t));
115     const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
116 
117     const __m256 vz = _mm256_or_ps(vx, vsign_mask);
118 
119     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
120     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
121     vn = _mm256_sub_ps(vn, vmagic_bias);
122 
123     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
124 
125     const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
126     vt = _mm256_mul_ps(vt, vs);
127     const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
128 
129     const __m256 vd = _mm256_add_ps(ve, vone);
130     const __m256 vr = _mm256_rcp_ps(vd);
131     __m256 vf = _mm256_mul_ps(ve, vr);
132 
133     vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
134     vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
135 
136     __m128i vh = _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC);
137     if (batch & (4 * sizeof(uint16_t))) {
138       _mm_storel_epi64((__m128i*) o, vh);
139       vh = _mm_unpackhi_epi64(vh, vh);
140       o += 4;
141     }
142     if (batch & (2 * sizeof(uint16_t))) {
143       _mm_storeu_si32(o, vh);
144       vh = _mm_srli_epi64(vh, 32);
145       o += 2;
146     }
147     if (batch & (1 * sizeof(uint16_t))) {
148       *o = (uint16_t) _mm_extract_epi16(vh, 0);
149     }
150   }
151 }
152