1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-velu/neon-lut16-p3.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 
18 extern XNN_INTERNAL const int32_t xnn_table_exp2minus_k_over_16[16];
19 
xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x20(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x20(
21     size_t n,
22     const float* x,
23     float* y,
24     const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
25 {
26   assert(n != 0);
27   assert(n % sizeof(float) == 0);
28   assert(x != NULL);
29   assert(y != NULL);
30 
31   const float32x4_t vprescale = vld1q_dup_f32(&params->scalar.prescale);
32   const float32x4_t valpha = vld1q_dup_f32(&params->scalar.alpha);
33   const float32x4_t vbeta = vld1q_dup_f32(&params->scalar.beta);
34 
35   const float32x4_t vsat_cutoff = vmovq_n_f32(-0x1.154246p+4f);
36   const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p19f);
37   const float32x4_t vlog2e = vmovq_n_f32(0x1.715476p+0f);
38   const int32x4_t vindex_mask = vmovq_n_s32(0xF);
39   const float32x4_t vminus_ln2_hi = vmovq_n_f32(-0x1.62E400p-1f);
40   const float32x4_t vminus_ln2_lo = vmovq_n_f32(-0x1.7F7D1Cp-20f);
41   const float32x4_t vc3 = vmovq_n_f32(0x1.55561Cp-3f);
42   const float32x4_t vc2 = vmovq_n_f32(0x1.0001ECp-1f);
43   const float32x4_t vone = vmovq_n_f32(1.0f);
44 
45   for (; n >= 20 * sizeof(float); n -= 20 * 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 
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 
58     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vlog2e);
59     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vlog2e);
60     float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vlog2e);
61     float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vlog2e);
62     float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vlog2e);
63 
64     const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask), 2));
65     const int32x4_t ven0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 19);
66     const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask), 2));
67     const int32x4_t ven4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 19);
68     const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask), 2));
69     const int32x4_t ven89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 19);
70     const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_mask), 2));
71     const int32x4_t venCDEF = vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 19);
72     const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_mask), 2));
73     const int32x4_t venGHIJ = vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 19);
74 
75     const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
76     const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
77     int32x2_t vl01 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01));
78     int32x2_t vl23 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23));
79     vl01 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32)), vl01, 1);
80     vl23 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32)), vl23, 1);
81     const int32x4_t vl0123 = vcombine_s32(vl01, vl23);
82     const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
83     const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
84     int32x2_t vl45 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45));
85     int32x2_t vl67 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67));
86     vl45 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32)), vl45, 1);
87     vl67 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32)), vl67, 1);
88     const int32x4_t vl4567 = vcombine_s32(vl45, vl67);
89     const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
90     const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
91     int32x2_t vl89 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx89));
92     int32x2_t vlAB = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxAB));
93     vl89 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx89 >> 32)), vl89, 1);
94     vlAB = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxAB >> 32)), vlAB, 1);
95     const int32x4_t vl89AB = vcombine_s32(vl89, vlAB);
96     const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0);
97     const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1);
98     int32x2_t vlCD = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxCD));
99     int32x2_t vlEF = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxEF));
100     vlCD = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxCD >> 32)), vlCD, 1);
101     vlEF = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxEF >> 32)), vlEF, 1);
102     const int32x4_t vlCDEF = vcombine_s32(vlCD, vlEF);
103     const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0);
104     const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1);
105     int32x2_t vlGH = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxGH));
106     int32x2_t vlIJ = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxIJ));
107     vlGH = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxGH >> 32)), vlGH, 1);
108     vlIJ = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxIJ >> 32)), vlIJ, 1);
109     const int32x4_t vlGHIJ = vcombine_s32(vlGH, vlIJ);
110 
111     vn0123 = vsubq_f32(vn0123, vmagic_bias);
112     float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vl0123, ven0123));
113     vn4567 = vsubq_f32(vn4567, vmagic_bias);
114     float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vl4567, ven4567));
115     vn89AB = vsubq_f32(vn89AB, vmagic_bias);
116     float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB));
117     vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
118     float32x4_t vsCDEF = vreinterpretq_f32_s32(vaddq_s32(vlCDEF, venCDEF));
119     vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
120     float32x4_t vsGHIJ = vreinterpretq_f32_s32(vaddq_s32(vlGHIJ, venGHIJ));
121 
122     float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vminus_ln2_hi);
123     float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vminus_ln2_hi);
124     float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vminus_ln2_hi);
125     float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vminus_ln2_hi);
126     float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vminus_ln2_hi);
127 
128     vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);
129     vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);
130     vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo);
131     vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo);
132     vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo);
133 
134     float32x4_t vp0123 = vmlaq_f32(vc2, vc3, vt0123);
135     float32x4_t vp4567 = vmlaq_f32(vc2, vc3, vt4567);
136     float32x4_t vp89AB = vmlaq_f32(vc2, vc3, vt89AB);
137     float32x4_t vpCDEF = vmlaq_f32(vc2, vc3, vtCDEF);
138     float32x4_t vpGHIJ = vmlaq_f32(vc2, vc3, vtGHIJ);
139 
140     vp0123 = vmulq_f32(vp0123, vt0123);
141     vp4567 = vmulq_f32(vp4567, vt4567);
142     vp89AB = vmulq_f32(vp89AB, vt89AB);
143     vpCDEF = vmulq_f32(vpCDEF, vtCDEF);
144     vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ);
145 
146     vt0123 = vmulq_f32(vt0123, vs0123);
147     vs0123 = vsubq_f32(vs0123, vone);
148     vt4567 = vmulq_f32(vt4567, vs4567);
149     vs4567 = vsubq_f32(vs4567, vone);
150     vt89AB = vmulq_f32(vt89AB, vs89AB);
151     vs89AB = vsubq_f32(vs89AB, vone);
152     vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
153     vsCDEF = vsubq_f32(vsCDEF, vone);
154     vtGHIJ = vmulq_f32(vtGHIJ, vsGHIJ);
155     vsGHIJ = vsubq_f32(vsGHIJ, vone);
156 
157     vp0123 = vmlaq_f32(vt0123, vp0123, vt0123);
158     vp4567 = vmlaq_f32(vt4567, vp4567, vt4567);
159     vp89AB = vmlaq_f32(vt89AB, vp89AB, vt89AB);
160     vpCDEF = vmlaq_f32(vtCDEF, vpCDEF, vtCDEF);
161     vpGHIJ = vmlaq_f32(vtGHIJ, vpGHIJ, vtGHIJ);
162 
163     const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
164     const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
165     const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha);
166     const float32x4_t veCDEF = vmulq_f32(vaddq_f32(vpCDEF, vsCDEF), valpha);
167     const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha);
168 
169     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
170     vx0123 = vmulq_f32(vx0123, vbeta);
171     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
172     vx4567 = vmulq_f32(vx4567, vbeta);
173     const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
174     vx89AB = vmulq_f32(vx89AB, vbeta);
175     const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
176     vxCDEF = vmulq_f32(vxCDEF, vbeta);
177     const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
178     vxGHIJ = vmulq_f32(vxGHIJ, vbeta);
179 
180     const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
181     const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
182     const float32x4_t vy89AB = vbslq_f32(vm89AB, ve89AB, vx89AB);
183     const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF);
184     const float32x4_t vyGHIJ = vbslq_f32(vmGHIJ, veGHIJ, vxGHIJ);
185 
186     vst1q_f32(y, vy0123); y += 4;
187     vst1q_f32(y, vy4567); y += 4;
188     vst1q_f32(y, vy89AB); y += 4;
189     vst1q_f32(y, vyCDEF); y += 4;
190     vst1q_f32(y, vyGHIJ); y += 4;
191   }
192   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
193     float32x4_t vx = vld1q_f32(x); x += 4;
194 
195     const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
196 
197     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
198     const uint64x2_t vidx = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask), 2));
199     const int32x4_t ven = vshlq_n_s32(vreinterpretq_s32_f32(vn), 19);
200 
201     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
202     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
203     int32x2_t vl_lo = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
204     int32x2_t vl_hi = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
205     vl_lo = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)), vl_lo, 1);
206     vl_hi = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)), vl_hi, 1);
207 
208     vn = vsubq_f32(vn, vmagic_bias);
209     const int32x4_t vl = vcombine_s32(vl_lo, vl_hi);
210 
211     float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
212     vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
213     float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vl, ven));
214 
215     float32x4_t vp = vmlaq_f32(vc2, vc3, vt);
216     vp = vmulq_f32(vp, vt);
217 
218     vt = vmulq_f32(vt, vs);
219     vs = vsubq_f32(vs, vone);
220     vp = vmlaq_f32(vt, vp, vt);
221     const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
222 
223     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
224     vx = vmulq_f32(vx, vbeta);
225     const float32x4_t vy = vbslq_f32(vm, ve, vx);
226 
227     vst1q_f32(y, vy); y += 4;
228   }
229   if XNN_UNLIKELY(n != 0) {
230     float32x4_t vx = vld1q_f32(x);
231 
232     const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
233 
234     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
235     const uint64x2_t vidx = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask), 2));
236     const int32x4_t ven = vshlq_n_s32(vreinterpretq_s32_f32(vn), 19);
237 
238     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
239     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
240     int32x2_t vl_lo = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
241     int32x2_t vl_hi = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
242     vl_lo = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)), vl_lo, 1);
243     vl_hi = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)), vl_hi, 1);
244 
245     vn = vsubq_f32(vn, vmagic_bias);
246     const int32x4_t vl = vcombine_s32(vl_lo, vl_hi);
247 
248     float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
249     vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
250     float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vl, ven));
251 
252     float32x4_t vp = vmlaq_f32(vc2, vc3, vt);
253     vp = vmulq_f32(vp, vt);
254 
255     vt = vmulq_f32(vt, vs);
256     vs = vsubq_f32(vs, vone);
257     vp = vmlaq_f32(vt, vp, vt);
258     const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
259 
260     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
261     vx = vmulq_f32(vx, vbeta);
262     const float32x4_t vy = vbslq_f32(vm, ve, vx);
263 
264     float32x2_t vy_lo = vget_low_f32(vy);
265     if (n & (2 * sizeof(float))) {
266       vst1_f32(y, vy_lo); y += 2;
267       vy_lo = vget_high_f32(vy);
268     }
269     if (n & (1 * sizeof(float))) {
270       vst1_lane_f32(y, vy_lo, 0);
271     }
272   }
273 }
274