1 // Auto-generated file. Do not edit!
2 // Template: src/f32-raddstoreexpminusmax/neon-p5.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/common.h>
15 #include <xnnpack/raddstoreexpminusmax.h>
16
17
xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16(size_t elements,const float * input,const float * max,float * output,float * sum,const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x16(
19 size_t elements,
20 const float* input,
21 const float* max,
22 float* output,
23 float* sum,
24 const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(elements % sizeof(float) == 0);
27
28 const float32x4_t vi_max = vld1q_dup_f32(max);
29 const float32x4_t vlog2e = vld1q_dup_f32(¶ms->neon_rr2_p5.log2e);
30 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->neon_rr2_p5.magic_bias);
31 const float32x4_t vminus_ln2_hi = vld1q_dup_f32(¶ms->neon_rr2_p5.minus_ln2_hi);
32 const float32x4_t vminus_ln2_lo = vld1q_dup_f32(¶ms->neon_rr2_p5.minus_ln2_lo);
33 const float32x4_t vc5 = vld1q_dup_f32(¶ms->neon_rr2_p5.c5);
34 const float32x4_t vc4 = vld1q_dup_f32(¶ms->neon_rr2_p5.c4);
35 const float32x4_t vc3 = vld1q_dup_f32(¶ms->neon_rr2_p5.c3);
36 const float32x4_t vc2 = vld1q_dup_f32(¶ms->neon_rr2_p5.c2);
37 const float32x4_t vc1 = vld1q_dup_f32(¶ms->neon_rr2_p5.c1);
38 const float32x4_t vdenorm_cutoff = vld1q_dup_f32(¶ms->neon_rr2_p5.denorm_cutoff);
39
40 float32x4_t vacc0 = vmovq_n_f32(0.0f);
41 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
42 const float32x4_t vi0123 = vld1q_f32(input); input += 4;
43 const float32x4_t vi4567 = vld1q_f32(input); input += 4;
44 const float32x4_t vi89AB = vld1q_f32(input); input += 4;
45 const float32x4_t viCDEF = vld1q_f32(input); input += 4;
46
47 const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max);
48 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max);
49 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max);
50 const float32x4_t vxCDEF = vsubq_f32(viCDEF, vi_max);
51
52 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vx0123, vlog2e);
53 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e);
54 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vx89AB, vlog2e);
55 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vxCDEF, vlog2e);
56
57 const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
58 const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
59 const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
60 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
61
62 vn0123 = vsubq_f32(vn0123, vmagic_bias);
63 vn4567 = vsubq_f32(vn4567, vmagic_bias);
64 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
65 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
66
67 float32x4_t vt0123 = vmlaq_f32(vx0123, vn0123, vminus_ln2_hi);
68 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi);
69 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi);
70 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi);
71
72 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);
73 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);
74 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo);
75 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo);
76
77 float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);
78 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);
79 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB);
80 float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF);
81
82 vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
83 vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
84 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
85 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);
86
87 vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
88 vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
89 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
90 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);
91
92 vp0123 = vmlaq_f32(vc1, vp0123, vt0123);
93 vp4567 = vmlaq_f32(vc1, vp4567, vt4567);
94 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB);
95 vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF);
96
97 vt0123 = vmulq_f32(vt0123, vs0123);
98 vt4567 = vmulq_f32(vt4567, vs4567);
99 vt89AB = vmulq_f32(vt89AB, vs89AB);
100 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
101
102 float32x4_t vf0123 = vmlaq_f32(vs0123, vp0123, vt0123);
103 float32x4_t vf4567 = vmlaq_f32(vs4567, vp4567, vt4567);
104 float32x4_t vf89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB);
105 float32x4_t vfCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF);
106
107 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_cutoff)));
108 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_cutoff)));
109 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_cutoff)));
110 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_cutoff)));
111
112 vst1q_f32(output, vf0123); output += 4;
113 vst1q_f32(output, vf4567); output += 4;
114 vst1q_f32(output, vf89AB); output += 4;
115 vst1q_f32(output, vfCDEF); output += 4;
116
117 vacc0 = vaddq_f32(vacc0, vf0123);
118 vacc0 = vaddq_f32(vacc0, vf4567);
119 vacc0 = vaddq_f32(vacc0, vf89AB);
120 vacc0 = vaddq_f32(vacc0, vfCDEF);
121 }
122
123 float32x4_t vacc = vacc0;
124 for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
125 const float32x4_t vi = vld1q_f32(input); input += 4;
126
127 const float32x4_t vx = vsubq_f32(vi, vi_max);
128
129 float32x4_t vn = vmlaq_f32(vmagic_bias, vx, vlog2e);
130
131 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
132
133 vn = vsubq_f32(vn, vmagic_bias);
134
135 float32x4_t vt = vmlaq_f32(vx, vn, vminus_ln2_hi);
136 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
137
138 float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
139 vp = vmlaq_f32(vc3, vp, vt);
140 vp = vmlaq_f32(vc2, vp, vt);
141 vp = vmlaq_f32(vc1, vp, vt);
142
143 vt = vmulq_f32(vt, vs);
144 float32x4_t vf = vmlaq_f32(vs, vp, vt);
145
146 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));
147
148 vst1q_f32(output, vf); output += 4;
149
150 vacc = vaddq_f32(vacc, vf);
151 }
152 #if XNN_ARCH_ARM64
153 float vacc_lo = vaddvq_f32(vacc);
154 #else
155 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc));
156 #endif
157 if (elements != 0) {
158 assert(elements >= 1 * sizeof(float));
159 assert(elements <= 3 * sizeof(float));
160 const float32x4_t vi = vld1q_f32(input); input += 4;
161
162 const float32x4_t vx = vsubq_f32(vi, vi_max);
163
164 float32x4_t vn = vmlaq_f32(vmagic_bias, vx, vlog2e);
165
166 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
167
168 vn = vsubq_f32(vn, vmagic_bias);
169
170 float32x4_t vt = vmlaq_f32(vx, vn, vminus_ln2_hi);
171 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
172
173 float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
174 vp = vmlaq_f32(vc3, vp, vt);
175 vp = vmlaq_f32(vc2, vp, vt);
176 vp = vmlaq_f32(vc1, vp, vt);
177
178 vt = vmulq_f32(vt, vs);
179 float32x4_t vf = vmlaq_f32(vs, vp, vt);
180
181 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));
182
183 float32x2_t vf_lo = vget_low_f32(vf);
184 if (elements & (2 * sizeof(float))) {
185 vst1_f32(output, vf_lo); output += 2;
186
187 #if XNN_ARCH_ARM64
188 vacc_lo += vaddv_f32(vf_lo);
189 #else
190 vacc_lo = vadd_f32(vacc_lo, vf_lo);
191 #endif
192
193 vf_lo = vget_high_f32(vf);
194 }
195 if (elements & (1 * sizeof(float))) {
196 vst1_lane_f32(output, vf_lo, 0);
197
198 #if XNN_ARCH_ARM64
199 vacc_lo += vget_lane_f32(vf_lo, 0);
200 #else
201 vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32)));
202 #endif
203 }
204 }
205 #if XNN_ARCH_ARM64
206 *sum = vacc_lo;
207 #else
208 vst1_lane_f32(sum, vpadd_f32(vacc_lo, vacc_lo), 0);
209 #endif
210 }
211