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/pavgpool.h>
9 #include <xnnpack/math.h>
10
11
xnn_f32_pavgpool_minmax_ukernel_9x__wasm_c1(size_t output_pixels,size_t kernel_elements,size_t channels,const float ** input,size_t input_offset,const float * zero,const float * multiplier,float * output,size_t input_increment,size_t output_increment,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])12 void xnn_f32_pavgpool_minmax_ukernel_9x__wasm_c1(
13 size_t output_pixels,
14 size_t kernel_elements,
15 size_t channels,
16 const float** input,
17 size_t input_offset,
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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(output_pixels != 0);
26 assert(kernel_elements != 0);
27 assert(kernel_elements <= 9);
28 assert(channels != 0);
29
30 const float voutput_min = params->scalar.min;
31 const float voutput_max = params->scalar.max;
32
33 do {
34 const float* i0 = input[0];
35 assert(i0 != NULL);
36 const float* i1 = input[1];
37 const float* i2 = input[2];
38 const float* i3 = input[3];
39 const float* i4 = input[4];
40 const float* i5 = input[5];
41 const float* i6 = input[6];
42 const float* i7 = input[7];
43 const float* i8 = input[8];
44 input = (const float**) ((uintptr_t) input + input_increment);
45 if (kernel_elements < 2) {
46 i1 = zero;
47 }
48 assert(i1 != NULL);
49 if (kernel_elements <= 2) {
50 i2 = zero;
51 }
52 assert(i2 != NULL);
53 if (kernel_elements < 4) {
54 i3 = zero;
55 }
56 assert(i3 != NULL);
57 if (kernel_elements <= 4) {
58 i4 = zero;
59 }
60 assert(i4 != NULL);
61 if (kernel_elements < 6) {
62 i5 = zero;
63 }
64 assert(i5 != NULL);
65 if (kernel_elements <= 6) {
66 i6 = zero;
67 }
68 assert(i6 != NULL);
69 if (kernel_elements < 8) {
70 i7 = zero;
71 }
72 assert(i7 != NULL);
73 if (kernel_elements <= 8) {
74 i8 = zero;
75 }
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i0 != zero) {
78 i0 = (const float*) ((uintptr_t) i0 + input_offset);
79 }
80 if XNN_UNPREDICTABLE(i1 != zero) {
81 i1 = (const float*) ((uintptr_t) i1 + input_offset);
82 }
83 if XNN_UNPREDICTABLE(i2 != zero) {
84 i2 = (const float*) ((uintptr_t) i2 + input_offset);
85 }
86 if XNN_UNPREDICTABLE(i3 != zero) {
87 i3 = (const float*) ((uintptr_t) i3 + input_offset);
88 }
89 if XNN_UNPREDICTABLE(i4 != zero) {
90 i4 = (const float*) ((uintptr_t) i4 + input_offset);
91 }
92 if XNN_UNPREDICTABLE(i5 != zero) {
93 i5 = (const float*) ((uintptr_t) i5 + input_offset);
94 }
95 if XNN_UNPREDICTABLE(i6 != zero) {
96 i6 = (const float*) ((uintptr_t) i6 + input_offset);
97 }
98 if XNN_UNPREDICTABLE(i7 != zero) {
99 i7 = (const float*) ((uintptr_t) i7 + input_offset);
100 }
101 if XNN_UNPREDICTABLE(i8 != zero) {
102 i8 = (const float*) ((uintptr_t) i8 + input_offset);
103 }
104
105 const float vmultiplier = *multiplier++;
106
107 size_t c = channels;
108 do {
109 const float vi0 = *i0++;
110 const float vi1 = *i1++;
111 const float vi2 = *i2++;
112 const float vi3 = *i3++;
113 const float vi4 = *i4++;
114 const float vi5 = *i5++;
115 const float vi6 = *i6++;
116 const float vi7 = *i7++;
117 const float vi8 = *i8++;
118
119 const float vsum01 = vi0 + vi1;
120 const float vsum23 = vi2 + vi3;
121 const float vsum45 = vi4 + vi5;
122 const float vsum67 = vi6 + vi7;
123 const float vsum018 = vsum01 + vi8;
124 const float vsum2345 = vsum23 + vsum45;
125 const float vsum01678 = vsum018 + vsum67;
126 const float vsum = vsum2345 + vsum01678;
127
128 float vout = vsum * vmultiplier;
129 vout = __builtin_wasm_max_f32(vout, voutput_min);
130 vout = __builtin_wasm_min_f32(vout, voutput_max);
131
132 *output++ = vout;
133 } while (--c != 0);
134 output = (float*) ((uintptr_t) output + output_increment);
135 } while (--output_pixels != 0);
136 }
137