1 // Auto-generated file. Do not edit!
2 // Template: src/f16-velu/avx2-rr1-p3.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_velu_ukernel__avx2_rr1_p3_x16(size_t n,const void * input,void * output,const union xnn_f16_elu_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f16_velu_ukernel__avx2_rr1_p3_x16(
20 size_t n,
21 const void* input,
22 void* output,
23 const union xnn_f16_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n % sizeof(uint16_t) == 0);
26
27 const __m256 vprescale = _mm256_load_ps(params->avx2_rr1_p3.prescale);
28 const __m256 vsat_cutoff = _mm256_load_ps(params->avx2_rr1_p3.sat_cutoff);
29 const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p3.magic_bias);
30 const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p3.log2e);
31 const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p3.minus_ln2);
32 const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p3.c3);
33 const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p3.c2);
34 const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p3.c1);
35 const __m256 valpha = _mm256_load_ps(params->avx2_rr1_p3.alpha);
36 const __m256 vbeta = _mm256_load_ps(params->avx2_rr1_p3.beta);
37
38 const uint16_t* i = (const uint16_t*) input;
39 uint16_t* o = (uint16_t*) output;
40 for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
41 __m256 vx0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
42 __m256 vx1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 8)));
43 i += 16;
44
45 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
46 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
47
48 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
49 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
50
51 __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
52 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
53 __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
54 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
55
56 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
57 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
58
59 __m256 vp0 = _mm256_fmadd_ps(vc3, vt0, vc2);
60 __m256 vp1 = _mm256_fmadd_ps(vc3, vt1, vc2);
61
62 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
63 vt0 = _mm256_mul_ps(vt0, valpha);
64 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
65 vt1 = _mm256_mul_ps(vt1, valpha);
66
67 vt0 = _mm256_mul_ps(vt0, vs0);
68 vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
69 vt1 = _mm256_mul_ps(vt1, vs1);
70 vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
71
72 const __m256 ve0 = _mm256_fmadd_ps(vp0, vt0, vs0);
73 vx0 = _mm256_mul_ps(vx0, vbeta);
74 const __m256 ve1 = _mm256_fmadd_ps(vp1, vt1, vs1);
75 vx1 = _mm256_mul_ps(vx1, vbeta);
76
77 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
78 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
79
80 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vy0, _MM_FROUND_NO_EXC));
81 _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vy1, _MM_FROUND_NO_EXC));
82 o += 16;
83 }
84 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
85 __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
86 i += 8;
87
88 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
89
90 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
91 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
92 vn = _mm256_sub_ps(vn, vmagic_bias);
93 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
94
95 __m256 vp = _mm256_fmadd_ps(vc3, vt, vc2);
96 vp = _mm256_fmadd_ps(vp, vt, vc1);
97 vt = _mm256_mul_ps(vt, valpha);
98 vt = _mm256_mul_ps(vt, vs);
99 vs = _mm256_fmsub_ps(vs, valpha, valpha);
100 const __m256 ve = _mm256_fmadd_ps(vp, vt, vs);
101 vx = _mm256_mul_ps(vx, vbeta);
102 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
103
104 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC));
105 o += 8;
106 }
107 if XNN_UNLIKELY(n != 0) {
108 assert(n >= 1 * sizeof(uint16_t));
109 assert(n <= 7 * sizeof(uint16_t));
110 __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
111
112 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
113
114 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
115 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
116 vn = _mm256_sub_ps(vn, vmagic_bias);
117 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
118
119 __m256 vp = _mm256_fmadd_ps(vc3, vt, vc2);
120 vp = _mm256_fmadd_ps(vp, vt, vc1);
121 vt = _mm256_mul_ps(vt, valpha);
122 vt = _mm256_mul_ps(vt, vs);
123 vs = _mm256_fmsub_ps(vs, valpha, valpha);
124 const __m256 ve = _mm256_fmadd_ps(vp, vt, vs);
125 vx = _mm256_mul_ps(vx, vbeta);
126 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
127
128 __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC);
129 if (n & (4 * sizeof(uint16_t))) {
130 _mm_storel_epi64((__m128i*) o, vh);
131 vh = _mm_unpackhi_epi64(vh, vh);
132 o += 4;
133 }
134 if (n & (2 * sizeof(uint16_t))) {
135 _mm_storeu_si32(o, vh);
136 vh = _mm_srli_epi64(vh, 32);
137 o += 2;
138 }
139 if (n & (1 * sizeof(uint16_t))) {
140 *o = (uint16_t) _mm_extract_epi16(vh, 0);
141 }
142 }
143 }
144