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_OOB_READS
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->neon_rr2_p6.prescale);
30 const float32x4_t valpha = vld1q_dup_f32(¶ms->neon_rr2_p6.alpha);
31 const float32x4_t vbeta = vld1q_dup_f32(¶ms->neon_rr2_p6.beta);
32 const float32x4_t vsat_cutoff = vld1q_dup_f32(¶ms->neon_rr2_p6.sat_cutoff);
33 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->neon_rr2_p6.magic_bias);
34 const float32x4_t vlog2e = vld1q_dup_f32(¶ms->neon_rr2_p6.log2e);
35 const float32x4_t vminus_ln2_hi = vld1q_dup_f32(¶ms->neon_rr2_p6.minus_ln2_hi);
36 const float32x4_t vminus_ln2_lo = vld1q_dup_f32(¶ms->neon_rr2_p6.minus_ln2_lo);
37 const float32x4_t vc6 = vld1q_dup_f32(¶ms->neon_rr2_p6.c6);
38 const float32x4_t vc5 = vld1q_dup_f32(¶ms->neon_rr2_p6.c5);
39 const float32x4_t vc4 = vld1q_dup_f32(¶ms->neon_rr2_p6.c4);
40 const float32x4_t vc3 = vld1q_dup_f32(¶ms->neon_rr2_p6.c3);
41 const float32x4_t vc2 = vld1q_dup_f32(¶ms->neon_rr2_p6.c2);
42 const float32x4_t vone = vmovq_n_f32(1.0f);
43
44 for (; n >= 24 * sizeof(float); n -= 24 * 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 float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
50 float32x4_t vxKLMN = vld1q_f32(x); x += 4;
51
52 const float32x4_t vz0123 = vmaxq_f32(vmulq_f32(vx0123, vprescale), vsat_cutoff);
53 const float32x4_t vz4567 = vmaxq_f32(vmulq_f32(vx4567, vprescale), vsat_cutoff);
54 const float32x4_t vz89AB = vmaxq_f32(vmulq_f32(vx89AB, vprescale), vsat_cutoff);
55 const float32x4_t vzCDEF = vmaxq_f32(vmulq_f32(vxCDEF, vprescale), vsat_cutoff);
56 const float32x4_t vzGHIJ = vmaxq_f32(vmulq_f32(vxGHIJ, vprescale), vsat_cutoff);
57 const float32x4_t vzKLMN = vmaxq_f32(vmulq_f32(vxKLMN, vprescale), vsat_cutoff);
58
59 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vlog2e);
60 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vlog2e);
61 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vlog2e);
62 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vlog2e);
63 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vlog2e);
64 float32x4_t vnKLMN = vmlaq_f32(vmagic_bias, vzKLMN, vlog2e);
65
66 float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
67 vn0123 = vsubq_f32(vn0123, vmagic_bias);
68 float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
69 vn4567 = vsubq_f32(vn4567, vmagic_bias);
70 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
71 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
72 float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
73 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
74 float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23));
75 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
76 float32x4_t vsKLMN = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnKLMN), 23));
77 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias);
78
79 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vminus_ln2_hi);
80 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vminus_ln2_hi);
81 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vminus_ln2_hi);
82 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vminus_ln2_hi);
83 float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vminus_ln2_hi);
84 float32x4_t vtKLMN = vmlaq_f32(vzKLMN, vnKLMN, vminus_ln2_hi);
85
86 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);
87 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);
88 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo);
89 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo);
90 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo);
91 vtKLMN = vmlaq_f32(vtKLMN, vnKLMN, vminus_ln2_lo);
92
93 float32x4_t vp0123 = vmlaq_f32(vc5, vc6, vt0123);
94 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567);
95 float32x4_t vp89AB = vmlaq_f32(vc5, vc6, vt89AB);
96 float32x4_t vpCDEF = vmlaq_f32(vc5, vc6, vtCDEF);
97 float32x4_t vpGHIJ = vmlaq_f32(vc5, vc6, vtGHIJ);
98 float32x4_t vpKLMN = vmlaq_f32(vc5, vc6, vtKLMN);
99
100 vp0123 = vmlaq_f32(vc4, vp0123, vt0123);
101 vp4567 = vmlaq_f32(vc4, vp4567, vt4567);
102 vp89AB = vmlaq_f32(vc4, vp89AB, vt89AB);
103 vpCDEF = vmlaq_f32(vc4, vpCDEF, vtCDEF);
104 vpGHIJ = vmlaq_f32(vc4, vpGHIJ, vtGHIJ);
105 vpKLMN = vmlaq_f32(vc4, vpKLMN, vtKLMN);
106
107 vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
108 vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
109 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
110 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);
111 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ);
112 vpKLMN = vmlaq_f32(vc3, vpKLMN, vtKLMN);
113
114 vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
115 vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
116 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
117 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);
118 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ);
119 vpKLMN = vmlaq_f32(vc2, vpKLMN, vtKLMN);
120
121 vp0123 = vmulq_f32(vp0123, vt0123);
122 vp4567 = vmulq_f32(vp4567, vt4567);
123 vp89AB = vmulq_f32(vp89AB, vt89AB);
124 vpCDEF = vmulq_f32(vpCDEF, vtCDEF);
125 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ);
126 vpKLMN = vmulq_f32(vpKLMN, vtKLMN);
127
128 vt0123 = vmulq_f32(vt0123, vs0123);
129 vs0123 = vsubq_f32(vs0123, vone);
130 vt4567 = vmulq_f32(vt4567, vs4567);
131 vs4567 = vsubq_f32(vs4567, vone);
132 vt89AB = vmulq_f32(vt89AB, vs89AB);
133 vs89AB = vsubq_f32(vs89AB, vone);
134 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
135 vsCDEF = vsubq_f32(vsCDEF, vone);
136 vtGHIJ = vmulq_f32(vtGHIJ, vsGHIJ);
137 vsGHIJ = vsubq_f32(vsGHIJ, vone);
138 vtKLMN = vmulq_f32(vtKLMN, vsKLMN);
139 vsKLMN = vsubq_f32(vsKLMN, vone);
140
141 vp0123 = vmlaq_f32(vt0123, vp0123, vt0123);
142 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567);
143 vp89AB = vmlaq_f32(vt89AB, vp89AB, vt89AB);
144 vpCDEF = vmlaq_f32(vtCDEF, vpCDEF, vtCDEF);
145 vpGHIJ = vmlaq_f32(vtGHIJ, vpGHIJ, vtGHIJ);
146 vpKLMN = vmlaq_f32(vtKLMN, vpKLMN, vtKLMN);
147
148 const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
149 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
150 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha);
151 const float32x4_t veCDEF = vmulq_f32(vaddq_f32(vpCDEF, vsCDEF), valpha);
152 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha);
153 const float32x4_t veKLMN = vmulq_f32(vaddq_f32(vpKLMN, vsKLMN), valpha);
154
155 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
156 vx0123 = vmulq_f32(vx0123, vbeta);
157 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
158 vx4567 = vmulq_f32(vx4567, vbeta);
159 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
160 vx89AB = vmulq_f32(vx89AB, vbeta);
161 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
162 vxCDEF = vmulq_f32(vxCDEF, vbeta);
163 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
164 vxGHIJ = vmulq_f32(vxGHIJ, vbeta);
165 const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f));
166 vxKLMN = vmulq_f32(vxKLMN, vbeta);
167
168 const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
169 const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
170 const float32x4_t vy89AB = vbslq_f32(vm89AB, ve89AB, vx89AB);
171 const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF);
172 const float32x4_t vyGHIJ = vbslq_f32(vmGHIJ, veGHIJ, vxGHIJ);
173 const float32x4_t vyKLMN = vbslq_f32(vmKLMN, veKLMN, vxKLMN);
174
175 vst1q_f32(y, vy0123); y += 4;
176 vst1q_f32(y, vy4567); y += 4;
177 vst1q_f32(y, vy89AB); y += 4;
178 vst1q_f32(y, vyCDEF); y += 4;
179 vst1q_f32(y, vyGHIJ); y += 4;
180 vst1q_f32(y, vyKLMN); y += 4;
181 }
182 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
183 float32x4_t vx = vld1q_f32(x); x += 4;
184
185 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
186
187 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
188 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
189 vn = vsubq_f32(vn, vmagic_bias);
190
191 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
192 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
193
194 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
195 vp = vmlaq_f32(vc4, vp, vt);
196 vp = vmlaq_f32(vc3, vp, vt);
197 vp = vmlaq_f32(vc2, vp, vt);
198 vp = vmulq_f32(vp, vt);
199
200 vt = vmulq_f32(vt, vs);
201 vs = vsubq_f32(vs, vone);
202 vp = vmlaq_f32(vt, vp, vt);
203 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
204
205 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
206 vx = vmulq_f32(vx, vbeta);
207 const float32x4_t vy = vbslq_f32(vm, ve, vx);
208
209 vst1q_f32(y, vy); y += 4;
210 }
211 if XNN_UNLIKELY(n != 0) {
212 float32x4_t vx = vld1q_f32(x);
213
214 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
215
216 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
217 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
218 vn = vsubq_f32(vn, vmagic_bias);
219
220 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
221 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
222
223 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
224 vp = vmlaq_f32(vc4, vp, vt);
225 vp = vmlaq_f32(vc3, vp, vt);
226 vp = vmlaq_f32(vc2, vp, vt);
227 vp = vmulq_f32(vp, vt);
228
229 vt = vmulq_f32(vt, vs);
230 vs = vsubq_f32(vs, vone);
231 vp = vmlaq_f32(vt, vp, vt);
232 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
233
234 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
235 vx = vmulq_f32(vx, vbeta);
236 const float32x4_t vy = vbslq_f32(vm, ve, vx);
237
238 float32x2_t vy_lo = vget_low_f32(vy);
239 if (n & (2 * sizeof(float))) {
240 vst1_f32(y, vy_lo); y += 2;
241 vy_lo = vget_high_f32(vy);
242 }
243 if (n & (1 * sizeof(float))) {
244 vst1_lane_f32(y, vy_lo, 0);
245 }
246 }
247 }
248