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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
45 float32x4_t vx0123 = vld1q_f32(x); x += 4;
46 float32x4_t vx4567 = vld1q_f32(x); x += 4;
47
48 const float32x4_t vz0123 = vmaxq_f32(vmulq_f32(vx0123, vprescale), vsat_cutoff);
49 const float32x4_t vz4567 = vmaxq_f32(vmulq_f32(vx4567, vprescale), vsat_cutoff);
50
51 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vlog2e);
52 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vlog2e);
53
54 float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
55 vn0123 = vsubq_f32(vn0123, vmagic_bias);
56 float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
57 vn4567 = vsubq_f32(vn4567, vmagic_bias);
58
59 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vminus_ln2);
60 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vminus_ln2);
61
62 float32x4_t vp0123 = vfmaq_f32(vc5, vc6, vt0123);
63 float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567);
64
65 vp0123 = vfmaq_f32(vc4, vp0123, vt0123);
66 vp4567 = vfmaq_f32(vc4, vp4567, vt4567);
67
68 vp0123 = vfmaq_f32(vc3, vp0123, vt0123);
69 vp4567 = vfmaq_f32(vc3, vp4567, vt4567);
70
71 vp0123 = vfmaq_f32(vc2, vp0123, vt0123);
72 vp4567 = vfmaq_f32(vc2, vp4567, vt4567);
73
74 vp0123 = vmulq_f32(vp0123, vt0123);
75 vp4567 = vmulq_f32(vp4567, vt4567);
76
77 vt0123 = vmulq_f32(vt0123, vs0123);
78 vs0123 = vsubq_f32(vs0123, vone);
79 vt4567 = vmulq_f32(vt4567, vs4567);
80 vs4567 = vsubq_f32(vs4567, vone);
81
82 vp0123 = vfmaq_f32(vt0123, vp0123, vt0123);
83 vp4567 = vfmaq_f32(vt4567, vp4567, vt4567);
84
85 const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
86 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
87
88 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
89 vx0123 = vmulq_f32(vx0123, vbeta);
90 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
91 vx4567 = vmulq_f32(vx4567, vbeta);
92
93 const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
94 const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
95
96 vst1q_f32(y, vy0123); y += 4;
97 vst1q_f32(y, vy4567); y += 4;
98 }
99 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
100 float32x4_t vx = vld1q_f32(x); x += 4;
101
102 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
103
104 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vlog2e);
105 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
106 vn = vsubq_f32(vn, vmagic_bias);
107
108 float32x4_t vt = vfmaq_f32(vz, vn, vminus_ln2);
109
110 float32x4_t vp = vfmaq_f32(vc5, vc6, vt);
111 vp = vfmaq_f32(vc4, vp, vt);
112 vp = vfmaq_f32(vc3, vp, vt);
113 vp = vfmaq_f32(vc2, vp, vt);
114 vp = vmulq_f32(vp, vt);
115
116 vt = vmulq_f32(vt, vs);
117 vs = vsubq_f32(vs, vone);
118 vp = vfmaq_f32(vt, vp, vt);
119 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
120
121 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
122 vx = vmulq_f32(vx, vbeta);
123 const float32x4_t vy = vbslq_f32(vm, ve, vx);
124
125 vst1q_f32(y, vy); y += 4;
126 }
127 if XNN_UNLIKELY(n != 0) {
128 float32x4_t vx = vld1q_f32(x);
129
130 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
131
132 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vlog2e);
133 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
134 vn = vsubq_f32(vn, vmagic_bias);
135
136 float32x4_t vt = vfmaq_f32(vz, vn, vminus_ln2);
137
138 float32x4_t vp = vfmaq_f32(vc5, vc6, vt);
139 vp = vfmaq_f32(vc4, vp, vt);
140 vp = vfmaq_f32(vc3, vp, vt);
141 vp = vfmaq_f32(vc2, vp, vt);
142 vp = vmulq_f32(vp, vt);
143
144 vt = vmulq_f32(vt, vs);
145 vs = vsubq_f32(vs, vone);
146 vp = vfmaq_f32(vt, vp, vt);
147 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
148
149 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
150 vx = vmulq_f32(vx, vbeta);
151 const float32x4_t vy = vbslq_f32(vm, ve, vx);
152
153 float32x2_t vy_lo = vget_low_f32(vy);
154 if (n & (2 * sizeof(float))) {
155 vst1_f32(y, vy_lo); y += 2;
156 vy_lo = vget_high_f32(vy);
157 }
158 if (n & (1 * sizeof(float))) {
159 vst1_lane_f32(y, vy_lo, 0);
160 }
161 }
162 }
163