• 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__neon_rr2_p6_x24(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__neon_rr2_p6_x24(
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_hi = vmovq_n_f32(-0x1.62E440p-1f);
37   const float32x4_t vminus_ln2_lo = vmovq_n_f32(0x1.0105C6p-21f);
38   const float32x4_t vc6 = vmovq_n_f32(0x1.6b7338p-10f);
39   const float32x4_t vc5 = vmovq_n_f32(0x1.12278Ep-7f);
40   const float32x4_t vc4 = vmovq_n_f32(0x1.555716p-5f);
41   const float32x4_t vc3 = vmovq_n_f32(0x1.5554B0p-3f);
42   const float32x4_t vc2 = vmovq_n_f32(0x1.FFFFFEp-2f);
43   const float32x4_t vone = vmovq_n_f32(1.0f);
44 
45   for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
46     float32x4_t vx0123 = vld1q_f32(x); x += 4;
47     float32x4_t vx4567 = vld1q_f32(x); x += 4;
48     float32x4_t vx89AB = vld1q_f32(x); x += 4;
49     float32x4_t vxCDEF = vld1q_f32(x); x += 4;
50     float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
51     float32x4_t vxKLMN = vld1q_f32(x); x += 4;
52 
53     const float32x4_t vz0123 = vmaxq_f32(vmulq_f32(vx0123, vprescale), vsat_cutoff);
54     const float32x4_t vz4567 = vmaxq_f32(vmulq_f32(vx4567, vprescale), vsat_cutoff);
55     const float32x4_t vz89AB = vmaxq_f32(vmulq_f32(vx89AB, vprescale), vsat_cutoff);
56     const float32x4_t vzCDEF = vmaxq_f32(vmulq_f32(vxCDEF, vprescale), vsat_cutoff);
57     const float32x4_t vzGHIJ = vmaxq_f32(vmulq_f32(vxGHIJ, vprescale), vsat_cutoff);
58     const float32x4_t vzKLMN = vmaxq_f32(vmulq_f32(vxKLMN, vprescale), vsat_cutoff);
59 
60     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vlog2e);
61     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vlog2e);
62     float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vlog2e);
63     float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vlog2e);
64     float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vlog2e);
65     float32x4_t vnKLMN = vmlaq_f32(vmagic_bias, vzKLMN, vlog2e);
66 
67     float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
68     vn0123 = vsubq_f32(vn0123, vmagic_bias);
69     float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
70     vn4567 = vsubq_f32(vn4567, vmagic_bias);
71     float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
72     vn89AB = vsubq_f32(vn89AB, vmagic_bias);
73     float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
74     vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
75     float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23));
76     vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
77     float32x4_t vsKLMN = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnKLMN), 23));
78     vnKLMN = vsubq_f32(vnKLMN, vmagic_bias);
79 
80     float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vminus_ln2_hi);
81     float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vminus_ln2_hi);
82     float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vminus_ln2_hi);
83     float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vminus_ln2_hi);
84     float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vminus_ln2_hi);
85     float32x4_t vtKLMN = vmlaq_f32(vzKLMN, vnKLMN, vminus_ln2_hi);
86 
87     vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);
88     vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);
89     vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo);
90     vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo);
91     vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo);
92     vtKLMN = vmlaq_f32(vtKLMN, vnKLMN, vminus_ln2_lo);
93 
94     float32x4_t vp0123 = vmlaq_f32(vc5, vc6, vt0123);
95     float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567);
96     float32x4_t vp89AB = vmlaq_f32(vc5, vc6, vt89AB);
97     float32x4_t vpCDEF = vmlaq_f32(vc5, vc6, vtCDEF);
98     float32x4_t vpGHIJ = vmlaq_f32(vc5, vc6, vtGHIJ);
99     float32x4_t vpKLMN = vmlaq_f32(vc5, vc6, vtKLMN);
100 
101     vp0123 = vmlaq_f32(vc4, vp0123, vt0123);
102     vp4567 = vmlaq_f32(vc4, vp4567, vt4567);
103     vp89AB = vmlaq_f32(vc4, vp89AB, vt89AB);
104     vpCDEF = vmlaq_f32(vc4, vpCDEF, vtCDEF);
105     vpGHIJ = vmlaq_f32(vc4, vpGHIJ, vtGHIJ);
106     vpKLMN = vmlaq_f32(vc4, vpKLMN, vtKLMN);
107 
108     vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
109     vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
110     vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
111     vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);
112     vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ);
113     vpKLMN = vmlaq_f32(vc3, vpKLMN, vtKLMN);
114 
115     vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
116     vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
117     vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
118     vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);
119     vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ);
120     vpKLMN = vmlaq_f32(vc2, vpKLMN, vtKLMN);
121 
122     vp0123 = vmulq_f32(vp0123, vt0123);
123     vp4567 = vmulq_f32(vp4567, vt4567);
124     vp89AB = vmulq_f32(vp89AB, vt89AB);
125     vpCDEF = vmulq_f32(vpCDEF, vtCDEF);
126     vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ);
127     vpKLMN = vmulq_f32(vpKLMN, vtKLMN);
128 
129     vt0123 = vmulq_f32(vt0123, vs0123);
130     vs0123 = vsubq_f32(vs0123, vone);
131     vt4567 = vmulq_f32(vt4567, vs4567);
132     vs4567 = vsubq_f32(vs4567, vone);
133     vt89AB = vmulq_f32(vt89AB, vs89AB);
134     vs89AB = vsubq_f32(vs89AB, vone);
135     vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
136     vsCDEF = vsubq_f32(vsCDEF, vone);
137     vtGHIJ = vmulq_f32(vtGHIJ, vsGHIJ);
138     vsGHIJ = vsubq_f32(vsGHIJ, vone);
139     vtKLMN = vmulq_f32(vtKLMN, vsKLMN);
140     vsKLMN = vsubq_f32(vsKLMN, vone);
141 
142     vp0123 = vmlaq_f32(vt0123, vp0123, vt0123);
143     vp4567 = vmlaq_f32(vt4567, vp4567, vt4567);
144     vp89AB = vmlaq_f32(vt89AB, vp89AB, vt89AB);
145     vpCDEF = vmlaq_f32(vtCDEF, vpCDEF, vtCDEF);
146     vpGHIJ = vmlaq_f32(vtGHIJ, vpGHIJ, vtGHIJ);
147     vpKLMN = vmlaq_f32(vtKLMN, vpKLMN, vtKLMN);
148 
149     const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
150     const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
151     const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha);
152     const float32x4_t veCDEF = vmulq_f32(vaddq_f32(vpCDEF, vsCDEF), valpha);
153     const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha);
154     const float32x4_t veKLMN = vmulq_f32(vaddq_f32(vpKLMN, vsKLMN), valpha);
155 
156     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
157     vx0123 = vmulq_f32(vx0123, vbeta);
158     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
159     vx4567 = vmulq_f32(vx4567, vbeta);
160     const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
161     vx89AB = vmulq_f32(vx89AB, vbeta);
162     const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
163     vxCDEF = vmulq_f32(vxCDEF, vbeta);
164     const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
165     vxGHIJ = vmulq_f32(vxGHIJ, vbeta);
166     const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f));
167     vxKLMN = vmulq_f32(vxKLMN, vbeta);
168 
169     const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
170     const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
171     const float32x4_t vy89AB = vbslq_f32(vm89AB, ve89AB, vx89AB);
172     const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF);
173     const float32x4_t vyGHIJ = vbslq_f32(vmGHIJ, veGHIJ, vxGHIJ);
174     const float32x4_t vyKLMN = vbslq_f32(vmKLMN, veKLMN, vxKLMN);
175 
176     vst1q_f32(y, vy0123); y += 4;
177     vst1q_f32(y, vy4567); y += 4;
178     vst1q_f32(y, vy89AB); y += 4;
179     vst1q_f32(y, vyCDEF); y += 4;
180     vst1q_f32(y, vyGHIJ); y += 4;
181     vst1q_f32(y, vyKLMN); y += 4;
182   }
183   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
184     float32x4_t vx = vld1q_f32(x); x += 4;
185 
186     const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
187 
188     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
189     float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
190     vn = vsubq_f32(vn, vmagic_bias);
191 
192     float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
193     vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
194 
195     float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
196     vp = vmlaq_f32(vc4, vp, vt);
197     vp = vmlaq_f32(vc3, vp, vt);
198     vp = vmlaq_f32(vc2, vp, vt);
199     vp = vmulq_f32(vp, vt);
200 
201     vt = vmulq_f32(vt, vs);
202     vs = vsubq_f32(vs, vone);
203     vp = vmlaq_f32(vt, vp, vt);
204     const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
205 
206     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
207     vx = vmulq_f32(vx, vbeta);
208     const float32x4_t vy = vbslq_f32(vm, ve, vx);
209 
210     vst1q_f32(y, vy); y += 4;
211   }
212   if XNN_UNLIKELY(n != 0) {
213     float32x4_t vx = vld1q_f32(x);
214 
215     const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
216 
217     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
218     float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
219     vn = vsubq_f32(vn, vmagic_bias);
220 
221     float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
222     vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
223 
224     float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
225     vp = vmlaq_f32(vc4, vp, vt);
226     vp = vmlaq_f32(vc3, vp, vt);
227     vp = vmlaq_f32(vc2, vp, vt);
228     vp = vmulq_f32(vp, vt);
229 
230     vt = vmulq_f32(vt, vs);
231     vs = vsubq_f32(vs, vone);
232     vp = vmlaq_f32(vt, vp, vt);
233     const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
234 
235     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
236     vx = vmulq_f32(vx, vbeta);
237     const float32x4_t vy = vbslq_f32(vm, ve, vx);
238 
239     float32x2_t vy_lo = vget_low_f32(vy);
240     if (n & (2 * sizeof(float))) {
241       vst1_f32(y, vy_lo); y += 2;
242       vy_lo = vget_high_f32(vy);
243     }
244     if (n & (1 * sizeof(float))) {
245       vst1_lane_f32(y, vy_lo, 0);
246     }
247   }
248 }
249