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 <xnnpack/gavgpool.h>
9 #include <xnnpack/math.h>
10
11
xnn_f32_gavgpool_ukernel_mp7p7q__scalar(size_t m,size_t n,const float * input,size_t input_stride,const float * zero,float * buffer,float * output,const union xnn_f32_avgpool_params params[restrict static1])12 void xnn_f32_gavgpool_ukernel_mp7p7q__scalar(
13 size_t m,
14 size_t n,
15 const float* input,
16 size_t input_stride,
17 const float* zero,
18 float* buffer,
19 float* output,
20 const union xnn_f32_avgpool_params params[restrict static 1])
21 {
22 assert(m > 7);
23 assert(n != 0);
24
25 const float* i0 = input;
26 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
27 const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
28 const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
29 const float* i4 = (const float*) ((uintptr_t) i3 + input_stride);
30 const float* i5 = (const float*) ((uintptr_t) i4 + input_stride);
31 const float* i6 = (const float*) ((uintptr_t) i5 + input_stride);
32 const size_t input_increment = 7 * input_stride - n * sizeof(float);
33
34 float* b = buffer;
35 size_t k = n;
36 do {
37 const float vi0 = *i0++;
38 const float vi1 = *i1++;
39 const float vi2 = *i2++;
40 const float vi3 = *i3++;
41 const float vi4 = *i4++;
42 const float vi5 = *i5++;
43 const float vi6 = *i6++;
44
45 const float vsum01 = vi0 + vi1;
46 const float vsum23 = vi2 + vi3;
47 const float vsum45 = vi4 + vi5;
48
49 const float vsum016 = vsum01 + vi6;
50 const float vsum2345 = vsum23 + vsum45;
51
52 const float vsum = vsum016 + vsum2345;
53
54 *b++ = vsum;
55 } while (--k != 0);
56 for (m -= 7; m > 7; m -= 7) {
57 b = buffer;
58
59 i0 = (const float*) ((uintptr_t) i0 + input_increment);
60 i1 = (const float*) ((uintptr_t) i1 + input_increment);
61 i2 = (const float*) ((uintptr_t) i2 + input_increment);
62 i3 = (const float*) ((uintptr_t) i3 + input_increment);
63 i4 = (const float*) ((uintptr_t) i4 + input_increment);
64 i5 = (const float*) ((uintptr_t) i5 + input_increment);
65 i6 = (const float*) ((uintptr_t) i6 + input_increment);
66
67 size_t k = n;
68 do {
69 const float vi0 = *i0++;
70 const float vi1 = *i1++;
71 const float vi2 = *i2++;
72 const float vi3 = *i3++;
73 const float vi4 = *i4++;
74 const float vi5 = *i5++;
75 const float vi6 = *i6++;
76 const float vacc = *b;
77
78 const float vsum01 = vi0 + vi1;
79 const float vsum23 = vi2 + vi3;
80 const float vsum45 = vi4 + vi5;
81 const float vsum6a = vi6 + vacc;
82
83 const float vsum0123 = vsum01 + vsum23;
84 const float vsum456a = vsum45 + vsum6a;
85
86 const float vsum = vsum0123 + vsum456a;
87
88 *b++ = vsum;
89 } while (--k != 0);
90 }
91
92 i0 = (const float*) ((uintptr_t) i0 + input_increment);
93 i1 = (const float*) ((uintptr_t) i1 + input_increment);
94 if (m < 2) {
95 i1 = zero;
96 }
97 i2 = (const float*) ((uintptr_t) i2 + input_increment);
98 if (m <= 2) {
99 i2 = zero;
100 }
101 i3 = (const float*) ((uintptr_t) i3 + input_increment);
102 if (m < 4) {
103 i3 = zero;
104 }
105 i4 = (const float*) ((uintptr_t) i4 + input_increment);
106 if (m <= 4) {
107 i4 = zero;
108 }
109 i5 = (const float*) ((uintptr_t) i5 + input_increment);
110 if (m < 6) {
111 i5 = zero;
112 }
113 i6 = (const float*) ((uintptr_t) i6 + input_increment);
114 if (m <= 6) {
115 i6 = zero;
116 }
117 const float vmultiplier = params->scalar.multiplier;
118 const float voutput_min = params->scalar.output_min;
119 const float voutput_max = params->scalar.output_max;
120
121 b = buffer;
122 do {
123 const float vi0 = *i0++;
124 const float vi1 = *i1++;
125 const float vi2 = *i2++;
126 const float vi3 = *i3++;
127 const float vi4 = *i4++;
128 const float vi5 = *i5++;
129 const float vi6 = *i6++;
130 const float vacc = *b++;
131
132 const float vsum01 = vi0 + vi1;
133 const float vsum23 = vi2 + vi3;
134 const float vsum45 = vi4 + vi5;
135 const float vsum6a = vi6 + vacc;
136
137 const float vsum0123 = vsum01 + vsum23;
138 const float vsum456a = vsum45 + vsum6a;
139
140 const float vsum = vsum0123 + vsum456a;
141
142 float vout = vsum * vmultiplier;
143 vout = math_max_f32(vout, voutput_min);
144 vout = math_min_f32(vout, voutput_max);
145
146 *output++ = vout;
147 } while (--n != 0);
148 }
149