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/gavgpool.h>
11
12
xnn_f32_gavgpool_ukernel_up7__neon(size_t m,size_t n,const float * input,size_t input_stride,const float * zero,float * output,const union xnn_f32_avgpool_params params[restrict static1])13 void xnn_f32_gavgpool_ukernel_up7__neon(
14 size_t m,
15 size_t n,
16 const float* input,
17 size_t input_stride,
18 const float* zero,
19 float* output,
20 const union xnn_f32_avgpool_params params[restrict static 1])
21 {
22 assert(m != 0);
23 assert(m <= 7);
24 assert(n != 0);
25
26 const float* i0 = input;
27 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
28 if (m < 2) {
29 i1 = zero;
30 }
31 const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
32 if (m <= 2) {
33 i2 = zero;
34 }
35 const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
36 if (m < 4) {
37 i3 = zero;
38 }
39 const float* i4 = (const float*) ((uintptr_t) i3 + input_stride);
40 if (m <= 4) {
41 i4 = zero;
42 }
43 const float* i5 = (const float*) ((uintptr_t) i4 + input_stride);
44 if (m < 6) {
45 i5 = zero;
46 }
47 const float* i6 = (const float*) ((uintptr_t) i5 + input_stride);
48 if (m <= 6) {
49 i6 = zero;
50 }
51 const float32x4_t vmultiplier = vld1q_dup_f32(¶ms->scalar.multiplier);
52 const float32x4_t voutput_min = vld1q_dup_f32(¶ms->scalar.output_min);
53 const float32x4_t voutput_max = vld1q_dup_f32(¶ms->scalar.output_max);
54
55 while (n >= 4) {
56 const float32x4_t vi0 = vld1q_f32(i0); i0 += 4;
57 const float32x4_t vi1 = vld1q_f32(i1); i1 += 4;
58 const float32x4_t vi2 = vld1q_f32(i2); i2 += 4;
59 const float32x4_t vi3 = vld1q_f32(i3); i3 += 4;
60 const float32x4_t vi4 = vld1q_f32(i4); i4 += 4;
61 const float32x4_t vi5 = vld1q_f32(i5); i5 += 4;
62 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4;
63
64 const float32x4_t vsum01 = vaddq_f32(vi0, vi1);
65 const float32x4_t vsum23 = vaddq_f32(vi2, vi3);
66 const float32x4_t vsum45 = vaddq_f32(vi4, vi5);
67
68 const float32x4_t vsum016 = vaddq_f32(vsum01, vi6);
69 const float32x4_t vsum2345 = vaddq_f32(vsum23, vsum45);
70
71 const float32x4_t vsum = vaddq_f32(vsum016, vsum2345);
72
73 float32x4_t vout = vmulq_f32(vsum, vmultiplier);
74 vout = vmaxq_f32(vout, voutput_min);
75 vout = vminq_f32(vout, voutput_max);
76
77 vst1q_f32(output, vout); output += 4;
78
79 n -= 4;
80 }
81 if (n != 0) {
82 const float32x4_t vi0 = vld1q_f32(i0);
83 const float32x4_t vi1 = vld1q_f32(i1);
84 const float32x4_t vi2 = vld1q_f32(i2);
85 const float32x4_t vi3 = vld1q_f32(i3);
86 const float32x4_t vi4 = vld1q_f32(i4);
87 const float32x4_t vi5 = vld1q_f32(i5);
88 const float32x4_t vi6 = vld1q_f32(i6);
89
90 const float32x4_t vsum01 = vaddq_f32(vi0, vi1);
91 const float32x4_t vsum23 = vaddq_f32(vi2, vi3);
92 const float32x4_t vsum45 = vaddq_f32(vi4, vi5);
93
94 const float32x4_t vsum016 = vaddq_f32(vsum01, vi6);
95 const float32x4_t vsum2345 = vaddq_f32(vsum23, vsum45);
96
97 const float32x4_t vsum = vaddq_f32(vsum016, vsum2345);
98
99 float32x4_t vout = vmulq_f32(vsum, vmultiplier);
100 vout = vmaxq_f32(vout, voutput_min);
101 vout = vminq_f32(vout, voutput_max);
102
103 float32x2_t vout_lo = vget_low_f32(vout);
104 if (n & 2) {
105 vst1_f32(output, vout_lo); output += 2;
106 vout_lo = vget_high_f32(vout);
107 }
108 if (n & 1) {
109 vst1_lane_f32(output, vout_lo, 0);
110 }
111 }
112 }
113