1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-wasmsimd.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <wasm_simd128.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_f32_dwconv_ukernel_up8x3__wasmsimd(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_ukernel_up8x3__wasmsimd(
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 size_t input_offset,
26 const float* zero,
27 const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const float* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const float*) ((uintptr_t) i0 + input_offset);
37 }
38 const float* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const float*) ((uintptr_t) i1 + input_offset);
42 }
43 const float* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const float*) ((uintptr_t) i2 + input_offset);
47 }
48 input = (const float**) ((uintptr_t) input + input_stride);
49
50 size_t c = channels;
51 const float* w = weights;
52 for (; c >= 8; c -= 8) {
53 v128_t vacc0123p0 = wasm_v128_load(w);
54 v128_t vacc4567p0 = wasm_v128_load(w + 4);
55
56
57 const v128_t vi0x0123 = wasm_v128_load(i0);
58 const v128_t vi0x4567 = wasm_v128_load(i0 + 4);
59 i0 += 8;
60
61 const v128_t vk0x0123 = wasm_v128_load(w + 8);
62 const v128_t vk0x4567 = wasm_v128_load(w + 12);
63 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
64 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi0x4567, vk0x4567));
65
66 const v128_t vi1x0123 = wasm_v128_load(i1);
67 const v128_t vi1x4567 = wasm_v128_load(i1 + 4);
68 i1 += 8;
69
70 const v128_t vk1x0123 = wasm_v128_load(w + 16);
71 const v128_t vk1x4567 = wasm_v128_load(w + 20);
72 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
73 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi1x4567, vk1x4567));
74
75 const v128_t vi2x0123 = wasm_v128_load(i2);
76 const v128_t vi2x4567 = wasm_v128_load(i2 + 4);
77 i2 += 8;
78
79 const v128_t vk2x0123 = wasm_v128_load(w + 24);
80 const v128_t vk2x4567 = wasm_v128_load(w + 28);
81 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
82 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567));
83
84 w += 32;
85
86
87 const v128_t vacc0123 = vacc0123p0;
88 const v128_t vacc4567 = vacc4567p0;
89
90 wasm_v128_store(output, vacc0123);
91 wasm_v128_store(output + 4, vacc4567);
92 output += 8;
93 }
94 for (; c >= 4; c -= 4) {
95 v128_t vacc0123p0 = wasm_v128_load(w);
96
97 const v128_t vi0x0123 = wasm_v128_load(i0);
98 i0 += 4;
99
100 const v128_t vk0x0123 = wasm_v128_load(w + 8);
101 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
102
103 const v128_t vi1x0123 = wasm_v128_load(i1);
104 i1 += 4;
105
106 const v128_t vk1x0123 = wasm_v128_load(w + 16);
107 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
108
109 const v128_t vi2x0123 = wasm_v128_load(i2);
110 i2 += 4;
111
112 const v128_t vk2x0123 = wasm_v128_load(w + 24);
113 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
114
115 w += 4;
116
117
118 const v128_t vacc0123 = vacc0123p0;
119
120 wasm_v128_store(output, vacc0123);
121 output += 4;
122 }
123 if XNN_UNLIKELY(c != 0) {
124 v128_t vacc0123p0 = wasm_v128_load(w);
125
126 const v128_t vi0x0123 = wasm_v128_load(i0);
127 const v128_t vk0x0123 = wasm_v128_load(w + 8);
128 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
129
130 const v128_t vi1x0123 = wasm_v128_load(i1);
131 const v128_t vk1x0123 = wasm_v128_load(w + 16);
132 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
133
134 const v128_t vi2x0123 = wasm_v128_load(i2);
135 const v128_t vk2x0123 = wasm_v128_load(w + 24);
136 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
137
138
139 v128_t vacc0123 = vacc0123p0;
140
141 if (c & 2) {
142 *((double*) output) = wasm_f64x2_extract_lane(vacc0123, 0);
143 vacc0123 = wasm_v32x4_shuffle(vacc0123, vacc0123, 2, 3, 2, 3);
144 output += 2;
145 }
146 if (c & 1) {
147 *output = wasm_f32x4_extract_lane(vacc0123, 0);
148 output += 1;
149 }
150 }
151
152 output = (float*) ((uintptr_t) output + output_increment);
153 } while (--output_width != 0);
154 }
155