1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/neon-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 <arm_neon.h>
13
14 #include <xnnpack/vunary.h>
15 #include <xnnpack/common.h>
16
17
xnn_f32_velu_ukernel__neonfma_rr1_p6_x4(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_velu_ukernel__neonfma_rr1_p6_x4(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const float32x4_t vprescale = vld1q_dup_f32(¶ms->scalar.prescale);
30 const float32x4_t valpha = vld1q_dup_f32(¶ms->scalar.alpha);
31 const float32x4_t vbeta = vld1q_dup_f32(¶ms->scalar.beta);
32
33 const float32x4_t vsat_cutoff = vmovq_n_f32(-0x1.154246p+4f);
34 const float32x4_t vmagic_bias = vmovq_n_f32(0x1.8000FEp23f);
35 const float32x4_t vlog2e = vmovq_n_f32(0x1.715476p+0f);
36 const float32x4_t vminus_ln2 = vmovq_n_f32(-0x1.62E430p-1f);
37 const float32x4_t vc6 = vmovq_n_f32(0x1.6b7338p-10f);
38 const float32x4_t vc5 = vmovq_n_f32(0x1.12278Ep-7f);
39 const float32x4_t vc4 = vmovq_n_f32(0x1.555716p-5f);
40 const float32x4_t vc3 = vmovq_n_f32(0x1.5554B0p-3f);
41 const float32x4_t vc2 = vmovq_n_f32(0x1.FFFFFEp-2f);
42 const float32x4_t vone = vmovq_n_f32(1.0f);
43
44 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
45 float32x4_t vx = vld1q_f32(x); x += 4;
46
47 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
48
49 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vlog2e);
50 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
51 vn = vsubq_f32(vn, vmagic_bias);
52
53 float32x4_t vt = vfmaq_f32(vz, vn, vminus_ln2);
54
55 float32x4_t vp = vfmaq_f32(vc5, vc6, vt);
56 vp = vfmaq_f32(vc4, vp, vt);
57 vp = vfmaq_f32(vc3, vp, vt);
58 vp = vfmaq_f32(vc2, vp, vt);
59 vp = vmulq_f32(vp, vt);
60
61 vt = vmulq_f32(vt, vs);
62 vs = vsubq_f32(vs, vone);
63 vp = vfmaq_f32(vt, vp, vt);
64 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
65
66 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
67 vx = vmulq_f32(vx, vbeta);
68 const float32x4_t vy = vbslq_f32(vm, ve, vx);
69
70 vst1q_f32(y, vy); y += 4;
71 }
72 if XNN_UNLIKELY(n != 0) {
73 float32x4_t vx = vld1q_f32(x);
74
75 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
76
77 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vlog2e);
78 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
79 vn = vsubq_f32(vn, vmagic_bias);
80
81 float32x4_t vt = vfmaq_f32(vz, vn, vminus_ln2);
82
83 float32x4_t vp = vfmaq_f32(vc5, vc6, vt);
84 vp = vfmaq_f32(vc4, vp, vt);
85 vp = vfmaq_f32(vc3, vp, vt);
86 vp = vfmaq_f32(vc2, vp, vt);
87 vp = vmulq_f32(vp, vt);
88
89 vt = vmulq_f32(vt, vs);
90 vs = vsubq_f32(vs, vone);
91 vp = vfmaq_f32(vt, vp, vt);
92 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
93
94 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
95 vx = vmulq_f32(vx, vbeta);
96 const float32x4_t vy = vbslq_f32(vm, ve, vx);
97
98 float32x2_t vy_lo = vget_low_f32(vy);
99 if (n & (2 * sizeof(float))) {
100 vst1_f32(y, vy_lo); y += 2;
101 vy_lo = vget_high_f32(vy);
102 }
103 if (n & (1 * sizeof(float))) {
104 vst1_lane_f32(y, vy_lo, 0);
105 }
106 }
107 }
108