• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_x16(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_x16(
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(&params->scalar.prescale);
30   const float32x4_t valpha = vld1q_dup_f32(&params->scalar.alpha);
31   const float32x4_t vbeta = vld1q_dup_f32(&params->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 >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
45     float32x4_t vx0123 = vld1q_f32(x); x += 4;
46     float32x4_t vx4567 = vld1q_f32(x); x += 4;
47     float32x4_t vx89AB = vld1q_f32(x); x += 4;
48     float32x4_t vxCDEF = vld1q_f32(x); x += 4;
49 
50     const float32x4_t vz0123 = vmaxq_f32(vmulq_f32(vx0123, vprescale), vsat_cutoff);
51     const float32x4_t vz4567 = vmaxq_f32(vmulq_f32(vx4567, vprescale), vsat_cutoff);
52     const float32x4_t vz89AB = vmaxq_f32(vmulq_f32(vx89AB, vprescale), vsat_cutoff);
53     const float32x4_t vzCDEF = vmaxq_f32(vmulq_f32(vxCDEF, vprescale), vsat_cutoff);
54 
55     float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vlog2e);
56     float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vlog2e);
57     float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vlog2e);
58     float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vlog2e);
59 
60     float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
61     vn0123 = vsubq_f32(vn0123, vmagic_bias);
62     float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
63     vn4567 = vsubq_f32(vn4567, vmagic_bias);
64     float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
65     vn89AB = vsubq_f32(vn89AB, vmagic_bias);
66     float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
67     vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
68 
69     float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vminus_ln2);
70     float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vminus_ln2);
71     float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vminus_ln2);
72     float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vminus_ln2);
73 
74     float32x4_t vp0123 = vfmaq_f32(vc5, vc6, vt0123);
75     float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567);
76     float32x4_t vp89AB = vfmaq_f32(vc5, vc6, vt89AB);
77     float32x4_t vpCDEF = vfmaq_f32(vc5, vc6, vtCDEF);
78 
79     vp0123 = vfmaq_f32(vc4, vp0123, vt0123);
80     vp4567 = vfmaq_f32(vc4, vp4567, vt4567);
81     vp89AB = vfmaq_f32(vc4, vp89AB, vt89AB);
82     vpCDEF = vfmaq_f32(vc4, vpCDEF, vtCDEF);
83 
84     vp0123 = vfmaq_f32(vc3, vp0123, vt0123);
85     vp4567 = vfmaq_f32(vc3, vp4567, vt4567);
86     vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB);
87     vpCDEF = vfmaq_f32(vc3, vpCDEF, vtCDEF);
88 
89     vp0123 = vfmaq_f32(vc2, vp0123, vt0123);
90     vp4567 = vfmaq_f32(vc2, vp4567, vt4567);
91     vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB);
92     vpCDEF = vfmaq_f32(vc2, vpCDEF, vtCDEF);
93 
94     vp0123 = vmulq_f32(vp0123, vt0123);
95     vp4567 = vmulq_f32(vp4567, vt4567);
96     vp89AB = vmulq_f32(vp89AB, vt89AB);
97     vpCDEF = vmulq_f32(vpCDEF, vtCDEF);
98 
99     vt0123 = vmulq_f32(vt0123, vs0123);
100     vs0123 = vsubq_f32(vs0123, vone);
101     vt4567 = vmulq_f32(vt4567, vs4567);
102     vs4567 = vsubq_f32(vs4567, vone);
103     vt89AB = vmulq_f32(vt89AB, vs89AB);
104     vs89AB = vsubq_f32(vs89AB, vone);
105     vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
106     vsCDEF = vsubq_f32(vsCDEF, vone);
107 
108     vp0123 = vfmaq_f32(vt0123, vp0123, vt0123);
109     vp4567 = vfmaq_f32(vt4567, vp4567, vt4567);
110     vp89AB = vfmaq_f32(vt89AB, vp89AB, vt89AB);
111     vpCDEF = vfmaq_f32(vtCDEF, vpCDEF, vtCDEF);
112 
113     const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
114     const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
115     const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha);
116     const float32x4_t veCDEF = vmulq_f32(vaddq_f32(vpCDEF, vsCDEF), valpha);
117 
118     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
119     vx0123 = vmulq_f32(vx0123, vbeta);
120     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
121     vx4567 = vmulq_f32(vx4567, vbeta);
122     const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
123     vx89AB = vmulq_f32(vx89AB, vbeta);
124     const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
125     vxCDEF = vmulq_f32(vxCDEF, vbeta);
126 
127     const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
128     const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
129     const float32x4_t vy89AB = vbslq_f32(vm89AB, ve89AB, vx89AB);
130     const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF);
131 
132     vst1q_f32(y, vy0123); y += 4;
133     vst1q_f32(y, vy4567); y += 4;
134     vst1q_f32(y, vy89AB); y += 4;
135     vst1q_f32(y, vyCDEF); y += 4;
136   }
137   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
138     float32x4_t vx = vld1q_f32(x); x += 4;
139 
140     const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
141 
142     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vlog2e);
143     float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
144     vn = vsubq_f32(vn, vmagic_bias);
145 
146     float32x4_t vt = vfmaq_f32(vz, vn, vminus_ln2);
147 
148     float32x4_t vp = vfmaq_f32(vc5, vc6, vt);
149     vp = vfmaq_f32(vc4, vp, vt);
150     vp = vfmaq_f32(vc3, vp, vt);
151     vp = vfmaq_f32(vc2, vp, vt);
152     vp = vmulq_f32(vp, vt);
153 
154     vt = vmulq_f32(vt, vs);
155     vs = vsubq_f32(vs, vone);
156     vp = vfmaq_f32(vt, vp, vt);
157     const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
158 
159     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
160     vx = vmulq_f32(vx, vbeta);
161     const float32x4_t vy = vbslq_f32(vm, ve, vx);
162 
163     vst1q_f32(y, vy); y += 4;
164   }
165   if XNN_UNLIKELY(n != 0) {
166     float32x4_t vx = vld1q_f32(x);
167 
168     const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
169 
170     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vlog2e);
171     float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
172     vn = vsubq_f32(vn, vmagic_bias);
173 
174     float32x4_t vt = vfmaq_f32(vz, vn, vminus_ln2);
175 
176     float32x4_t vp = vfmaq_f32(vc5, vc6, vt);
177     vp = vfmaq_f32(vc4, vp, vt);
178     vp = vfmaq_f32(vc3, vp, vt);
179     vp = vfmaq_f32(vc2, vp, vt);
180     vp = vmulq_f32(vp, vt);
181 
182     vt = vmulq_f32(vt, vs);
183     vs = vsubq_f32(vs, vone);
184     vp = vfmaq_f32(vt, vp, vt);
185     const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
186 
187     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
188     vx = vmulq_f32(vx, vbeta);
189     const float32x4_t vy = vbslq_f32(vm, ve, vx);
190 
191     float32x2_t vy_lo = vget_low_f32(vy);
192     if (n & (2 * sizeof(float))) {
193       vst1_f32(y, vy_lo); y += 2;
194       vy_lo = vget_high_f32(vy);
195     }
196     if (n & (1 * sizeof(float))) {
197       vst1_lane_f32(y, vy_lo, 0);
198     }
199   }
200 }
201