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_up8x4__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_up8x4__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 const float* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const float*) ((uintptr_t) i3 + input_offset);
52 }
53 input = (const float**) ((uintptr_t) input + input_stride);
54
55 size_t c = channels;
56 const float* w = weights;
57 for (; c >= 8; c -= 8) {
58 v128_t vacc0123p0 = wasm_v128_load(w);
59 v128_t vacc4567p0 = wasm_v128_load(w + 4);
60
61
62 const v128_t vi0x0123 = wasm_v128_load(i0);
63 const v128_t vi0x4567 = wasm_v128_load(i0 + 4);
64 i0 += 8;
65
66 const v128_t vk0x0123 = wasm_v128_load(w + 8);
67 const v128_t vk0x4567 = wasm_v128_load(w + 12);
68 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
69 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi0x4567, vk0x4567));
70
71 const v128_t vi1x0123 = wasm_v128_load(i1);
72 const v128_t vi1x4567 = wasm_v128_load(i1 + 4);
73 i1 += 8;
74
75 const v128_t vk1x0123 = wasm_v128_load(w + 16);
76 const v128_t vk1x4567 = wasm_v128_load(w + 20);
77 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
78 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi1x4567, vk1x4567));
79
80 const v128_t vi2x0123 = wasm_v128_load(i2);
81 const v128_t vi2x4567 = wasm_v128_load(i2 + 4);
82 i2 += 8;
83
84 const v128_t vk2x0123 = wasm_v128_load(w + 24);
85 const v128_t vk2x4567 = wasm_v128_load(w + 28);
86 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
87 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567));
88
89 const v128_t vi3x0123 = wasm_v128_load(i3);
90 const v128_t vi3x4567 = wasm_v128_load(i3 + 4);
91 i3 += 8;
92
93 const v128_t vk3x0123 = wasm_v128_load(w + 32);
94 const v128_t vk3x4567 = wasm_v128_load(w + 36);
95 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
96 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi3x4567, vk3x4567));
97
98 w += 40;
99
100
101 const v128_t vacc0123 = vacc0123p0;
102 const v128_t vacc4567 = vacc4567p0;
103
104 wasm_v128_store(output, vacc0123);
105 wasm_v128_store(output + 4, vacc4567);
106 output += 8;
107 }
108 for (; c >= 4; c -= 4) {
109 v128_t vacc0123p0 = wasm_v128_load(w);
110
111 const v128_t vi0x0123 = wasm_v128_load(i0);
112 i0 += 4;
113
114 const v128_t vk0x0123 = wasm_v128_load(w + 8);
115 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
116
117 const v128_t vi1x0123 = wasm_v128_load(i1);
118 i1 += 4;
119
120 const v128_t vk1x0123 = wasm_v128_load(w + 16);
121 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
122
123 const v128_t vi2x0123 = wasm_v128_load(i2);
124 i2 += 4;
125
126 const v128_t vk2x0123 = wasm_v128_load(w + 24);
127 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
128
129 const v128_t vi3x0123 = wasm_v128_load(i3);
130 i3 += 4;
131
132 const v128_t vk3x0123 = wasm_v128_load(w + 32);
133 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
134
135 w += 4;
136
137
138 const v128_t vacc0123 = vacc0123p0;
139
140 wasm_v128_store(output, vacc0123);
141 output += 4;
142 }
143 if XNN_UNLIKELY(c != 0) {
144 v128_t vacc0123p0 = wasm_v128_load(w);
145
146 const v128_t vi0x0123 = wasm_v128_load(i0);
147 const v128_t vk0x0123 = wasm_v128_load(w + 8);
148 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
149
150 const v128_t vi1x0123 = wasm_v128_load(i1);
151 const v128_t vk1x0123 = wasm_v128_load(w + 16);
152 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
153
154 const v128_t vi2x0123 = wasm_v128_load(i2);
155 const v128_t vk2x0123 = wasm_v128_load(w + 24);
156 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
157
158 const v128_t vi3x0123 = wasm_v128_load(i3);
159 const v128_t vk3x0123 = wasm_v128_load(w + 32);
160 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
161
162
163 v128_t vacc0123 = vacc0123p0;
164
165 if (c & 2) {
166 *((double*) output) = wasm_f64x2_extract_lane(vacc0123, 0);
167 vacc0123 = wasm_v32x4_shuffle(vacc0123, vacc0123, 2, 3, 2, 3);
168 output += 2;
169 }
170 if (c & 1) {
171 *output = wasm_f32x4_extract_lane(vacc0123, 0);
172 output += 1;
173 }
174 }
175
176 output = (float*) ((uintptr_t) output + output_increment);
177 } while (--output_width != 0);
178 }
179