1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <arm_neon.h>
9
10 #include <xnnpack/pavgpool.h>
11
12
xnn_f32_pavgpool_ukernel_up9__neon(size_t n,size_t ks,size_t kc,const float ** input,const float * zero,const float * multiplier,float * output,size_t input_increment,size_t output_increment,const union xnn_f32_output_params params[restrict static1])13 void xnn_f32_pavgpool_ukernel_up9__neon(
14 size_t n,
15 size_t ks,
16 size_t kc,
17 const float** input,
18 const float* zero,
19 const float* multiplier,
20 float* output,
21 size_t input_increment,
22 size_t output_increment,
23 const union xnn_f32_output_params params[restrict static 1])
24 {
25 assert(n != 0);
26 assert(ks != 0);
27 assert(ks <= 9);
28 assert(kc != 0);
29
30 const float32x4_t voutput_min = vld1q_dup_f32(¶ms->scalar.min);
31 const float32x4_t voutput_max = vld1q_dup_f32(¶ms->scalar.max);
32
33 do {
34 const float* i0 = input[0];
35 const float* i1 = input[1];
36 const float* i2 = input[2];
37 const float* i3 = input[3];
38 const float* i4 = input[4];
39 const float* i5 = input[5];
40 const float* i6 = input[6];
41 const float* i7 = input[7];
42 const float* i8 = input[8];
43 input = (const float**) ((uintptr_t) input + input_increment);
44 if (ks < 2) {
45 i1 = zero;
46 }
47 if (ks <= 2) {
48 i2 = zero;
49 }
50 if (ks < 4) {
51 i3 = zero;
52 }
53 if (ks <= 4) {
54 i4 = zero;
55 }
56 if (ks < 6) {
57 i5 = zero;
58 }
59 if (ks <= 6) {
60 i6 = zero;
61 }
62 if (ks < 8) {
63 i7 = zero;
64 }
65 if (ks <= 8) {
66 i8 = zero;
67 }
68
69 const float32x4_t vmultiplier = vld1q_dup_f32(multiplier); multiplier += 1;
70
71 size_t k = kc;
72 while (k >= 4) {
73 const float32x4_t vi0 = vld1q_f32(i0); i0 += 4;
74 const float32x4_t vi1 = vld1q_f32(i1); i1 += 4;
75 const float32x4_t vi2 = vld1q_f32(i2); i2 += 4;
76 const float32x4_t vi3 = vld1q_f32(i3); i3 += 4;
77 const float32x4_t vi4 = vld1q_f32(i4); i4 += 4;
78 const float32x4_t vi5 = vld1q_f32(i5); i5 += 4;
79 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4;
80 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4;
81 const float32x4_t vi8 = vld1q_f32(i8); i8 += 4;
82
83 const float32x4_t vsum01 = vaddq_f32(vi0, vi1);
84 const float32x4_t vsum23 = vaddq_f32(vi2, vi3);
85 const float32x4_t vsum45 = vaddq_f32(vi4, vi5);
86 const float32x4_t vsum67 = vaddq_f32(vi6, vi7);
87 const float32x4_t vsum018 = vaddq_f32(vsum01, vi8);
88 const float32x4_t vsum2345 = vaddq_f32(vsum23, vsum45);
89 const float32x4_t vsum01678 = vaddq_f32(vsum018, vsum67);
90 const float32x4_t vsum = vaddq_f32(vsum2345, vsum01678);
91
92 float32x4_t vout = vmulq_f32(vsum, vmultiplier);
93 vout = vmaxq_f32(vout, voutput_min);
94 vout = vminq_f32(vout, voutput_max);
95
96 vst1q_f32(output, vout); output += 4;
97
98 k -= 4;
99 }
100 if (k != 0) {
101 const float32x4_t vi0 = vld1q_f32(i0);
102 const float32x4_t vi1 = vld1q_f32(i1);
103 const float32x4_t vi2 = vld1q_f32(i2);
104 const float32x4_t vi3 = vld1q_f32(i3);
105 const float32x4_t vi4 = vld1q_f32(i4);
106 const float32x4_t vi5 = vld1q_f32(i5);
107 const float32x4_t vi6 = vld1q_f32(i6);
108 const float32x4_t vi7 = vld1q_f32(i7);
109 const float32x4_t vi8 = vld1q_f32(i8);
110
111 const float32x4_t vsum01 = vaddq_f32(vi0, vi1);
112 const float32x4_t vsum23 = vaddq_f32(vi2, vi3);
113 const float32x4_t vsum45 = vaddq_f32(vi4, vi5);
114 const float32x4_t vsum67 = vaddq_f32(vi6, vi7);
115 const float32x4_t vsum018 = vaddq_f32(vsum01, vi8);
116 const float32x4_t vsum2345 = vaddq_f32(vsum23, vsum45);
117 const float32x4_t vsum01678 = vaddq_f32(vsum018, vsum67);
118 const float32x4_t vsum = vaddq_f32(vsum2345, vsum01678);
119
120 float32x4_t vout = vmulq_f32(vsum, vmultiplier);
121 vout = vmaxq_f32(vout, voutput_min);
122 vout = vminq_f32(vout, voutput_max);
123
124 float32x2_t vout_lo = vget_low_f32(vout);
125 if (k & 2) {
126 vst1_f32(output, vout_lo); output += 2;
127 vout_lo = vget_high_f32(vout);
128 }
129 if (k & 1) {
130 vst1_lane_f32(output, vout_lo, 0); output += 1;
131 }
132 }
133 output = (float*) ((uintptr_t) output + output_increment);
134 } while (--n != 0);
135 }
136