• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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__neonfma_rr1_lut16_p3_x16(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__neonfma_rr1_lut16_p3_x16(
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_OOB_READS
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->neonfma_rr1_lut16_p3.prescale);
32   const float32x4_t valpha = vld1q_dup_f32(&params->neonfma_rr1_lut16_p3.alpha);
33   const float32x4_t vbeta = vld1q_dup_f32(&params->neonfma_rr1_lut16_p3.beta);
34   const float32x4_t vsat_cutoff = vld1q_dup_f32(&params->neonfma_rr1_lut16_p3.sat_cutoff);
35   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->neonfma_rr1_lut16_p3.magic_bias);
36   const float32x4_t vlog2e = vld1q_dup_f32(&params->neonfma_rr1_lut16_p3.log2e);
37   const int32x4_t vindex_mask = vmovq_n_s32(0xF);
38   const float32x4_t vminus_ln2 = vld1q_dup_f32(&params->neonfma_rr1_lut16_p3.minus_ln2);
39   const float32x4_t vc3 = vld1q_dup_f32(&params->neonfma_rr1_lut16_p3.c3);
40   const float32x4_t vc2 = vld1q_dup_f32(&params->neonfma_rr1_lut16_p3.c2);
41   const float32x4_t vone = vmovq_n_f32(1.0f);
42 
43   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
44     float32x4_t vx0123 = vld1q_f32(x); x += 4;
45     float32x4_t vx4567 = vld1q_f32(x); x += 4;
46     float32x4_t vx89AB = vld1q_f32(x); x += 4;
47     float32x4_t vxCDEF = vld1q_f32(x); x += 4;
48 
49     const float32x4_t vz0123 = vmaxq_f32(vmulq_f32(vx0123, vprescale), vsat_cutoff);
50     const float32x4_t vz4567 = vmaxq_f32(vmulq_f32(vx4567, vprescale), vsat_cutoff);
51     const float32x4_t vz89AB = vmaxq_f32(vmulq_f32(vx89AB, vprescale), vsat_cutoff);
52     const float32x4_t vzCDEF = vmaxq_f32(vmulq_f32(vxCDEF, vprescale), vsat_cutoff);
53 
54     float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vlog2e);
55     float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vlog2e);
56     float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vlog2e);
57     float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vlog2e);
58 
59     const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask), 2));
60     const int32x4_t ven0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 19);
61     const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask), 2));
62     const int32x4_t ven4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 19);
63     const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask), 2));
64     const int32x4_t ven89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 19);
65     const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_mask), 2));
66     const int32x4_t venCDEF = vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 19);
67 
68     const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
69     const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
70     int32x2_t vl01 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01));
71     int32x2_t vl23 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23));
72     vl01 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32)), vl01, 1);
73     vl23 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32)), vl23, 1);
74     const int32x4_t vl0123 = vcombine_s32(vl01, vl23);
75     const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
76     const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
77     int32x2_t vl45 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45));
78     int32x2_t vl67 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67));
79     vl45 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32)), vl45, 1);
80     vl67 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32)), vl67, 1);
81     const int32x4_t vl4567 = vcombine_s32(vl45, vl67);
82     const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
83     const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
84     int32x2_t vl89 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx89));
85     int32x2_t vlAB = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxAB));
86     vl89 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx89 >> 32)), vl89, 1);
87     vlAB = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxAB >> 32)), vlAB, 1);
88     const int32x4_t vl89AB = vcombine_s32(vl89, vlAB);
89     const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0);
90     const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1);
91     int32x2_t vlCD = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxCD));
92     int32x2_t vlEF = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxEF));
93     vlCD = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxCD >> 32)), vlCD, 1);
94     vlEF = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxEF >> 32)), vlEF, 1);
95     const int32x4_t vlCDEF = vcombine_s32(vlCD, vlEF);
96 
97     vn0123 = vsubq_f32(vn0123, vmagic_bias);
98     float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vl0123, ven0123));
99     vn4567 = vsubq_f32(vn4567, vmagic_bias);
100     float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vl4567, ven4567));
101     vn89AB = vsubq_f32(vn89AB, vmagic_bias);
102     float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB));
103     vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
104     float32x4_t vsCDEF = vreinterpretq_f32_s32(vaddq_s32(vlCDEF, venCDEF));
105 
106     float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vminus_ln2);
107     float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vminus_ln2);
108     float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vminus_ln2);
109     float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vminus_ln2);
110 
111     float32x4_t vp0123 = vfmaq_f32(vc2, vc3, vt0123);
112     float32x4_t vp4567 = vfmaq_f32(vc2, vc3, vt4567);
113     float32x4_t vp89AB = vfmaq_f32(vc2, vc3, vt89AB);
114     float32x4_t vpCDEF = vfmaq_f32(vc2, vc3, vtCDEF);
115 
116     vp0123 = vmulq_f32(vp0123, vt0123);
117     vp4567 = vmulq_f32(vp4567, vt4567);
118     vp89AB = vmulq_f32(vp89AB, vt89AB);
119     vpCDEF = vmulq_f32(vpCDEF, vtCDEF);
120 
121     vt0123 = vmulq_f32(vt0123, vs0123);
122     vs0123 = vsubq_f32(vs0123, vone);
123     vt4567 = vmulq_f32(vt4567, vs4567);
124     vs4567 = vsubq_f32(vs4567, vone);
125     vt89AB = vmulq_f32(vt89AB, vs89AB);
126     vs89AB = vsubq_f32(vs89AB, vone);
127     vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
128     vsCDEF = vsubq_f32(vsCDEF, vone);
129 
130     vp0123 = vfmaq_f32(vt0123, vp0123, vt0123);
131     vp4567 = vfmaq_f32(vt4567, vp4567, vt4567);
132     vp89AB = vfmaq_f32(vt89AB, vp89AB, vt89AB);
133     vpCDEF = vfmaq_f32(vtCDEF, vpCDEF, vtCDEF);
134 
135     const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
136     const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
137     const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha);
138     const float32x4_t veCDEF = vmulq_f32(vaddq_f32(vpCDEF, vsCDEF), valpha);
139 
140     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
141     vx0123 = vmulq_f32(vx0123, vbeta);
142     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
143     vx4567 = vmulq_f32(vx4567, vbeta);
144     const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
145     vx89AB = vmulq_f32(vx89AB, vbeta);
146     const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
147     vxCDEF = vmulq_f32(vxCDEF, vbeta);
148 
149     const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
150     const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
151     const float32x4_t vy89AB = vbslq_f32(vm89AB, ve89AB, vx89AB);
152     const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF);
153 
154     vst1q_f32(y, vy0123); y += 4;
155     vst1q_f32(y, vy4567); y += 4;
156     vst1q_f32(y, vy89AB); y += 4;
157     vst1q_f32(y, vyCDEF); y += 4;
158   }
159   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
160     float32x4_t vx = vld1q_f32(x); x += 4;
161 
162     const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
163 
164     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vlog2e);
165     const uint64x2_t vidx = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask), 2));
166     const int32x4_t ven = vshlq_n_s32(vreinterpretq_s32_f32(vn), 19);
167 
168     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
169     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
170     int32x2_t vl_lo = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
171     int32x2_t vl_hi = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
172     vl_lo = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)), vl_lo, 1);
173     vl_hi = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)), vl_hi, 1);
174 
175     vn = vsubq_f32(vn, vmagic_bias);
176     const int32x4_t vl = vcombine_s32(vl_lo, vl_hi);
177 
178     float32x4_t vt = vfmaq_f32(vz, vn, vminus_ln2);
179     float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vl, ven));
180 
181     float32x4_t vp = vfmaq_f32(vc2, vc3, vt);
182     vp = vmulq_f32(vp, vt);
183 
184     vt = vmulq_f32(vt, vs);
185     vs = vsubq_f32(vs, vone);
186     vp = vfmaq_f32(vt, vp, vt);
187     const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
188 
189     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
190     vx = vmulq_f32(vx, vbeta);
191     const float32x4_t vy = vbslq_f32(vm, ve, vx);
192 
193     vst1q_f32(y, vy); y += 4;
194   }
195   if XNN_UNLIKELY(n != 0) {
196     float32x4_t vx = vld1q_f32(x);
197 
198     const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
199 
200     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vlog2e);
201     const uint64x2_t vidx = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask), 2));
202     const int32x4_t ven = vshlq_n_s32(vreinterpretq_s32_f32(vn), 19);
203 
204     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
205     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
206     int32x2_t vl_lo = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
207     int32x2_t vl_hi = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
208     vl_lo = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)), vl_lo, 1);
209     vl_hi = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)), vl_hi, 1);
210 
211     vn = vsubq_f32(vn, vmagic_bias);
212     const int32x4_t vl = vcombine_s32(vl_lo, vl_hi);
213 
214     float32x4_t vt = vfmaq_f32(vz, vn, vminus_ln2);
215     float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vl, ven));
216 
217     float32x4_t vp = vfmaq_f32(vc2, vc3, vt);
218     vp = vmulq_f32(vp, vt);
219 
220     vt = vmulq_f32(vt, vs);
221     vs = vsubq_f32(vs, vone);
222     vp = vfmaq_f32(vt, vp, vt);
223     const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
224 
225     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
226     vx = vmulq_f32(vx, vbeta);
227     const float32x4_t vy = vbslq_f32(vm, ve, vx);
228 
229     float32x2_t vy_lo = vget_low_f32(vy);
230     if (n & (2 * sizeof(float))) {
231       vst1_f32(y, vy_lo); y += 2;
232       vy_lo = vget_high_f32(vy);
233     }
234     if (n & (1 * sizeof(float))) {
235       vst1_lane_f32(y, vy_lo, 0);
236     }
237   }
238 }
239