1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-psimd.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <psimd.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_f32_dwconv_ukernel_up8x4__psimd(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_dwconv_ukernel_up8x4__psimd(
18 size_t channels,
19 size_t output_width,
20 const float** input,
21 const float* weights,
22 float* output,
23 size_t input_stride,
24 size_t output_increment,
25 const union xnn_f32_output_params params[restrict static 1])
26 {
27 assert(channels != 0);
28 assert(output_width != 0);
29
30 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
31 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
32 do {
33 const float* i0 = input[0];
34 assert(i0 != NULL);
35 const float* i1 = input[1];
36 assert(i1 != NULL);
37 const float* i2 = input[2];
38 assert(i2 != NULL);
39 const float* i3 = input[3];
40 assert(i3 != NULL);
41 input = (const float**) ((uintptr_t) input + input_stride);
42
43 size_t c = channels;
44 const float* w = weights;
45 for (; c >= 8; c -= 8) {
46 psimd_f32 vacc0123p0 = psimd_load_f32(w);
47 psimd_f32 vacc4567p0 = psimd_load_f32(w + 4);
48
49
50 const psimd_f32 vi0x0123 = psimd_load_f32(i0);
51 const psimd_f32 vi0x4567 = psimd_load_f32(i0 + 4);
52 i0 += 8;
53
54 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8);
55 const psimd_f32 vk0x4567 = psimd_load_f32(w + 12);
56 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123);
57 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi0x4567, vk0x4567);
58
59 const psimd_f32 vi1x0123 = psimd_load_f32(i1);
60 const psimd_f32 vi1x4567 = psimd_load_f32(i1 + 4);
61 i1 += 8;
62
63 const psimd_f32 vk1x0123 = psimd_load_f32(w + 16);
64 const psimd_f32 vk1x4567 = psimd_load_f32(w + 20);
65 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi1x0123, vk1x0123);
66 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi1x4567, vk1x4567);
67
68 const psimd_f32 vi2x0123 = psimd_load_f32(i2);
69 const psimd_f32 vi2x4567 = psimd_load_f32(i2 + 4);
70 i2 += 8;
71
72 const psimd_f32 vk2x0123 = psimd_load_f32(w + 24);
73 const psimd_f32 vk2x4567 = psimd_load_f32(w + 28);
74 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi2x0123, vk2x0123);
75 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi2x4567, vk2x4567);
76
77 const psimd_f32 vi3x0123 = psimd_load_f32(i3);
78 const psimd_f32 vi3x4567 = psimd_load_f32(i3 + 4);
79 i3 += 8;
80
81 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32);
82 const psimd_f32 vk3x4567 = psimd_load_f32(w + 36);
83 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi3x0123, vk3x0123);
84 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi3x4567, vk3x4567);
85
86 w += 40;
87
88
89 psimd_f32 vacc0123 = psimd_max_f32(vacc0123p0, vmin);
90 psimd_f32 vacc4567 = psimd_max_f32(vacc4567p0, vmin);
91 vacc0123 = psimd_min_f32(vacc0123, vmax);
92 vacc4567 = psimd_min_f32(vacc4567, vmax);
93
94 psimd_store_f32(output, vacc0123);
95 psimd_store_f32(output + 4, vacc4567);
96 output += 8;
97 }
98 for (; c >= 4; c -= 4) {
99 psimd_f32 vacc0123p0 = psimd_load_f32(w);
100
101 const psimd_f32 vi0x0123 = psimd_load_f32(i0);
102 i0 += 4;
103
104 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8);
105 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123);
106
107 const psimd_f32 vi1x0123 = psimd_load_f32(i1);
108 i1 += 4;
109
110 const psimd_f32 vk1x0123 = psimd_load_f32(w + 16);
111 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi1x0123, vk1x0123);
112
113 const psimd_f32 vi2x0123 = psimd_load_f32(i2);
114 i2 += 4;
115
116 const psimd_f32 vk2x0123 = psimd_load_f32(w + 24);
117 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi2x0123, vk2x0123);
118
119 const psimd_f32 vi3x0123 = psimd_load_f32(i3);
120 i3 += 4;
121
122 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32);
123 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi3x0123, vk3x0123);
124
125 w += 4;
126
127
128 psimd_f32 vacc0123 = psimd_max_f32(vacc0123p0, vmin);
129 vacc0123 = psimd_min_f32(vacc0123, vmax);
130
131 psimd_store_f32(output, vacc0123);
132 output += 4;
133 }
134 if XNN_UNLIKELY(c != 0) {
135 psimd_f32 vacc0123p0 = psimd_load_f32(w);
136
137 const psimd_f32 vi0x0123 = psimd_load_f32(i0);
138 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8);
139 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123);
140
141 const psimd_f32 vi1x0123 = psimd_load_f32(i1);
142 const psimd_f32 vk1x0123 = psimd_load_f32(w + 16);
143 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi1x0123, vk1x0123);
144
145 const psimd_f32 vi2x0123 = psimd_load_f32(i2);
146 const psimd_f32 vk2x0123 = psimd_load_f32(w + 24);
147 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi2x0123, vk2x0123);
148
149 const psimd_f32 vi3x0123 = psimd_load_f32(i3);
150 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32);
151 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi3x0123, vk3x0123);
152
153
154 psimd_f32 vacc0123 = psimd_max_f32(vacc0123p0, vmin);
155 vacc0123 = psimd_min_f32(vacc0123, vmax);
156
157 if (c & 2) {
158 psimd_store2_f32(output, vacc0123);
159 vacc0123 = psimd_concat_hi_f32(vacc0123, vacc0123);
160 output += 2;
161 }
162 if (c & 1) {
163 psimd_store1_f32(output, vacc0123);
164 output += 1;
165 }
166 }
167
168 output = (float*) ((uintptr_t) output + output_increment);
169 } while (--output_width != 0);
170 }
171