• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_x8(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_x8(
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(&params->neon_rr2_p5.log2e);
30   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->neon_rr2_p5.magic_bias);
31   const float32x4_t vminus_ln2_hi = vld1q_dup_f32(&params->neon_rr2_p5.minus_ln2_hi);
32   const float32x4_t vminus_ln2_lo = vld1q_dup_f32(&params->neon_rr2_p5.minus_ln2_lo);
33   const float32x4_t vc5 = vld1q_dup_f32(&params->neon_rr2_p5.c5);
34   const float32x4_t vc4 = vld1q_dup_f32(&params->neon_rr2_p5.c4);
35   const float32x4_t vc3 = vld1q_dup_f32(&params->neon_rr2_p5.c3);
36   const float32x4_t vc2 = vld1q_dup_f32(&params->neon_rr2_p5.c2);
37   const float32x4_t vc1 = vld1q_dup_f32(&params->neon_rr2_p5.c1);
38   const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->neon_rr2_p5.denorm_cutoff);
39 
40   float32x4_t vacc0 = vmovq_n_f32(0.0f);
41   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
42     const float32x4_t vi0123 = vld1q_f32(input); input += 4;
43     const float32x4_t vi4567 = vld1q_f32(input); input += 4;
44 
45     const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max);
46     const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max);
47 
48     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vx0123, vlog2e);
49     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e);
50 
51     const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
52     const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
53 
54     vn0123 = vsubq_f32(vn0123, vmagic_bias);
55     vn4567 = vsubq_f32(vn4567, vmagic_bias);
56 
57     float32x4_t vt0123 = vmlaq_f32(vx0123, vn0123, vminus_ln2_hi);
58     float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi);
59 
60     vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);
61     vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);
62 
63     float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);
64     float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);
65 
66     vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
67     vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
68 
69     vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
70     vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
71 
72     vp0123 = vmlaq_f32(vc1, vp0123, vt0123);
73     vp4567 = vmlaq_f32(vc1, vp4567, vt4567);
74 
75     vt0123 = vmulq_f32(vt0123, vs0123);
76     vt4567 = vmulq_f32(vt4567, vs4567);
77 
78     float32x4_t vf0123 = vmlaq_f32(vs0123, vp0123, vt0123);
79     float32x4_t vf4567 = vmlaq_f32(vs4567, vp4567, vt4567);
80 
81     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_cutoff)));
82     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_cutoff)));
83 
84     vst1q_f32(output, vf0123); output += 4;
85     vst1q_f32(output, vf4567); output += 4;
86 
87     vacc0 = vaddq_f32(vacc0, vf0123);
88     vacc0 = vaddq_f32(vacc0, vf4567);
89   }
90 
91   float32x4_t vacc = vacc0;
92   for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
93     const float32x4_t vi = vld1q_f32(input); input += 4;
94 
95     const float32x4_t vx = vsubq_f32(vi, vi_max);
96 
97     float32x4_t vn = vmlaq_f32(vmagic_bias, vx, vlog2e);
98 
99     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
100 
101     vn = vsubq_f32(vn, vmagic_bias);
102 
103     float32x4_t vt = vmlaq_f32(vx, vn, vminus_ln2_hi);
104     vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
105 
106     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
107     vp = vmlaq_f32(vc3, vp, vt);
108     vp = vmlaq_f32(vc2, vp, vt);
109     vp = vmlaq_f32(vc1, vp, vt);
110 
111     vt = vmulq_f32(vt, vs);
112     float32x4_t vf = vmlaq_f32(vs, vp, vt);
113 
114     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));
115 
116     vst1q_f32(output, vf); output += 4;
117 
118     vacc = vaddq_f32(vacc, vf);
119   }
120 #if XNN_ARCH_ARM64
121   float vacc_lo = vaddvq_f32(vacc);
122 #else
123   float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc));
124 #endif
125   if (elements != 0) {
126     assert(elements >= 1 * sizeof(float));
127     assert(elements <= 3 * sizeof(float));
128     const float32x4_t vi = vld1q_f32(input); input += 4;
129 
130     const float32x4_t vx = vsubq_f32(vi, vi_max);
131 
132     float32x4_t vn = vmlaq_f32(vmagic_bias, vx, vlog2e);
133 
134     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
135 
136     vn = vsubq_f32(vn, vmagic_bias);
137 
138     float32x4_t vt = vmlaq_f32(vx, vn, vminus_ln2_hi);
139     vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
140 
141     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
142     vp = vmlaq_f32(vc3, vp, vt);
143     vp = vmlaq_f32(vc2, vp, vt);
144     vp = vmlaq_f32(vc1, vp, vt);
145 
146     vt = vmulq_f32(vt, vs);
147     float32x4_t vf = vmlaq_f32(vs, vp, vt);
148 
149     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));
150 
151     float32x2_t vf_lo = vget_low_f32(vf);
152     if (elements & (2 * sizeof(float))) {
153       vst1_f32(output, vf_lo); output += 2;
154 
155       #if XNN_ARCH_ARM64
156         vacc_lo += vaddv_f32(vf_lo);
157       #else
158         vacc_lo = vadd_f32(vacc_lo, vf_lo);
159       #endif
160 
161       vf_lo = vget_high_f32(vf);
162     }
163     if (elements & (1 * sizeof(float))) {
164       vst1_lane_f32(output, vf_lo, 0);
165 
166       #if XNN_ARCH_ARM64
167         vacc_lo += vget_lane_f32(vf_lo, 0);
168       #else
169         vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32)));
170       #endif
171     }
172   }
173 #if XNN_ARCH_ARM64
174   *sum = vacc_lo;
175 #else
176   vst1_lane_f32(sum, vpadd_f32(vacc_lo, vacc_lo), 0);
177 #endif
178 }
179