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_x8(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_x8(
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 >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
39 const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
40 i += 8;
41
42 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
43
44 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
45 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
46 vn = _mm256_sub_ps(vn, vmagic_bias);
47
48 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
49
50 const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
51 vt = _mm256_mul_ps(vt, vs);
52 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
53
54 const __m256 vd = _mm256_add_ps(ve, vone);
55 const __m256 vr = _mm256_rcp_ps(vd);
56 __m256 vf = _mm256_mul_ps(ve, vr);
57
58 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
59 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
60
61 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC));
62 o += 8;
63 }
64 if XNN_UNLIKELY(batch != 0) {
65 assert(batch >= 1 * sizeof(uint16_t));
66 assert(batch <= 7 * sizeof(uint16_t));
67 const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
68
69 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
70
71 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
72 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
73 vn = _mm256_sub_ps(vn, vmagic_bias);
74
75 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
76
77 const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
78 vt = _mm256_mul_ps(vt, vs);
79 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
80
81 const __m256 vd = _mm256_add_ps(ve, vone);
82 const __m256 vr = _mm256_rcp_ps(vd);
83 __m256 vf = _mm256_mul_ps(ve, vr);
84
85 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
86 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
87
88 __m128i vh = _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC);
89 if (batch & (4 * sizeof(uint16_t))) {
90 _mm_storel_epi64((__m128i*) o, vh);
91 vh = _mm_unpackhi_epi64(vh, vh);
92 o += 4;
93 }
94 if (batch & (2 * sizeof(uint16_t))) {
95 _mm_storeu_si32(o, vh);
96 vh = _mm_srli_epi64(vh, 32);
97 o += 2;
98 }
99 if (batch & (1 * sizeof(uint16_t))) {
100 *o = (uint16_t) _mm_extract_epi16(vh, 0);
101 }
102 }
103 }
104