1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx512f-rr1-p6.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_f32_velu_ukernel__avx512f_rr1_p6_x16(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_velu_ukernel__avx512f_rr1_p6_x16(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27
28 const __m512 vprescale = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.prescale));
29 const __m512 valpha = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.alpha));
30 const __m512 vbeta = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.beta));
31
32 const __m512 vsat_cutoff = _mm512_set1_ps(-0x1.154246p+4f);
33 const __m512 vmagic_bias = _mm512_set1_ps(0x1.8000FEp23f);
34 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
35 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f);
36 const __m512 vc6 = _mm512_set1_ps(0x1.6b7338p-10f);
37 const __m512 vc5 = _mm512_set1_ps(0x1.12278Ep-7f);
38 const __m512 vc4 = _mm512_set1_ps(0x1.555716p-5f);
39 const __m512 vc3 = _mm512_set1_ps(0x1.5554B0p-3f);
40 const __m512 vc2 = _mm512_set1_ps(0x1.FFFFFEp-2f);
41
42 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
43 __m512 vx = _mm512_loadu_ps(x);
44 x += 16;
45
46 const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
47 const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
48
49 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
50 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23));
51 vn = _mm512_sub_ps(vn, vmagic_bias);
52
53 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
54
55 __m512 vp = _mm512_fmadd_ps(vc6, vt, vc5);
56 vp = _mm512_fmadd_ps(vp, vt, vc4);
57 vp = _mm512_fmadd_ps(vp, vt, vc3);
58 vp = _mm512_fmadd_ps(vp, vt, vc2);
59 vp = _mm512_mul_ps(vp, vt);
60
61 vt = _mm512_mul_ps(vt, vs);
62 vs = _mm512_fmsub_ps(vs, valpha, valpha);
63 vp = _mm512_fmadd_ps(vp, vt, vt);
64 __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
65
66 vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
67
68 _mm512_storeu_ps(y, vy);
69 y += 16;
70 }
71 if XNN_UNLIKELY(n != 0) {
72 assert(n >= 1 * sizeof(float));
73 assert(n <= 15 * sizeof(float));
74 // Prepare mask for valid 32-bit elements (depends on n).
75 n >>= 2 /* log2(sizeof(float)) */;
76 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
77
78 __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
79
80 const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
81 const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
82
83 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
84 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23));
85 vn = _mm512_sub_ps(vn, vmagic_bias);
86
87 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
88
89 __m512 vp = _mm512_fmadd_ps(vc6, vt, vc5);
90 vp = _mm512_fmadd_ps(vp, vt, vc4);
91 vp = _mm512_fmadd_ps(vp, vt, vc3);
92 vp = _mm512_fmadd_ps(vp, vt, vc2);
93 vp = _mm512_mul_ps(vp, vt);
94
95 vt = _mm512_mul_ps(vt, vs);
96 vs = _mm512_fmsub_ps(vs, valpha, valpha);
97 vp = _mm512_fmadd_ps(vp, vt, vt);
98 __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
99
100 vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
101
102 _mm512_mask_storeu_ps(y, vmask, vy);
103 }
104 }
105