• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <psimd.h>
9 
10 #include <xnnpack/avgpool.h>
11 
12 
xnn_f32_avgpool_ukernel_up9__psimd(size_t n,size_t ks,size_t kc,const float ** input,const float * zero,float * output,size_t input_increment,size_t output_increment,const union xnn_f32_avgpool_params params[restrict static1])13 void xnn_f32_avgpool_ukernel_up9__psimd(
14     size_t n,
15     size_t ks,
16     size_t kc,
17     const float** input,
18     const float* zero,
19     float* output,
20     size_t input_increment,
21     size_t output_increment,
22     const union xnn_f32_avgpool_params params[restrict static 1])
23 {
24   assert(n != 0);
25   assert(ks != 0);
26   assert(ks <= 9);
27   assert(kc != 0);
28 
29   const psimd_f32 vmultiplier = psimd_load_splat_f32(&params->scalar.multiplier);
30   const psimd_f32 voutput_min = psimd_load_splat_f32(&params->scalar.output_min);
31   const psimd_f32 voutput_max = psimd_load_splat_f32(&params->scalar.output_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     size_t k = kc;
70     while (k >= 4) {
71       const psimd_f32 vi0 = psimd_load_f32(i0);
72       i0 += 4;
73       const psimd_f32 vi1 = psimd_load_f32(i1);
74       i1 += 4;
75       const psimd_f32 vi2 = psimd_load_f32(i2);
76       i2 += 4;
77       const psimd_f32 vi3 = psimd_load_f32(i3);
78       i3 += 4;
79       const psimd_f32 vi4 = psimd_load_f32(i4);
80       i4 += 4;
81       const psimd_f32 vi5 = psimd_load_f32(i5);
82       i5 += 4;
83       const psimd_f32 vi6 = psimd_load_f32(i6);
84       i6 += 4;
85       const psimd_f32 vi7 = psimd_load_f32(i7);
86       i7 += 4;
87       const psimd_f32 vi8 = psimd_load_f32(i8);
88       i8 += 4;
89 
90       const psimd_f32 vsum018 = psimd_add_f32(psimd_add_f32(vi0, vi1), vi8);
91       const psimd_f32 vsum23 = psimd_add_f32(vi2, vi3);
92       const psimd_f32 vsum45 = psimd_add_f32(vi4, vi5);
93       const psimd_f32 vsum67 = psimd_add_f32(vi6, vi7);
94 
95       const psimd_f32 vsum2345 = psimd_add_f32(vsum23, vsum45);
96       const psimd_f32 vsum01678 = psimd_add_f32(vsum018, vsum67);
97       const psimd_f32 vsum = psimd_add_f32(vsum2345, vsum01678);
98 
99       psimd_f32 vout = psimd_mul_f32(vsum, vmultiplier);
100       vout = psimd_max_f32(vout, voutput_min);
101       vout = psimd_min_f32(vout, voutput_max);
102 
103       psimd_store_f32(output, vout);
104       output += 4;
105 
106       k -= 4;
107     }
108     if (k != 0) {
109       const psimd_f32 vi0 = psimd_load_f32(i0);
110       const psimd_f32 vi1 = psimd_load_f32(i1);
111       const psimd_f32 vi2 = psimd_load_f32(i2);
112       const psimd_f32 vi3 = psimd_load_f32(i3);
113       const psimd_f32 vi4 = psimd_load_f32(i4);
114       const psimd_f32 vi5 = psimd_load_f32(i5);
115       const psimd_f32 vi6 = psimd_load_f32(i6);
116       const psimd_f32 vi7 = psimd_load_f32(i7);
117       const psimd_f32 vi8 = psimd_load_f32(i8);
118 
119       const psimd_f32 vsum01 = psimd_add_f32(vi0, vi1);
120       const psimd_f32 vsum23 = psimd_add_f32(vi2, vi3);
121       const psimd_f32 vsum45 = psimd_add_f32(vi4, vi5);
122       const psimd_f32 vsum67 = psimd_add_f32(vi6, vi7);
123       const psimd_f32 vsum018 = psimd_add_f32(vsum01, vi8);
124       const psimd_f32 vsum2345 = psimd_add_f32(vsum23, vsum45);
125       const psimd_f32 vsum01678 = psimd_add_f32(vsum018, vsum67);
126       const psimd_f32 vsum = psimd_add_f32(vsum2345, vsum01678);
127 
128       psimd_f32 vout = psimd_mul_f32(vsum, vmultiplier);
129       vout = psimd_max_f32(vout, voutput_min);
130       vout = psimd_min_f32(vout, voutput_max);
131 
132       if (k & 2) {
133         psimd_store2_f32(output, vout);
134         output += 2;
135         vout = psimd_concat_hi_f32(vout, vout);
136       }
137       if (k & 1) {
138         psimd_store1_f32(output, vout);
139         output += 1;
140       }
141     }
142     output = (float*) ((uintptr_t) output + output_increment);
143   } while (--n != 0);
144 }
145