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_up8x9__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_up8x9__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 const float* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const float*) ((uintptr_t) i4 + input_offset);
57 }
58 const float* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const float*) ((uintptr_t) i5 + input_offset);
62 }
63 const float* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const float*) ((uintptr_t) i6 + input_offset);
67 }
68 const float* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const float*) ((uintptr_t) i7 + input_offset);
72 }
73 const float* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const float*) ((uintptr_t) i8 + input_offset);
77 }
78 input = (const float**) ((uintptr_t) input + input_stride);
79
80 size_t c = channels;
81 const float* w = weights;
82 for (; c >= 8; c -= 8) {
83 v128_t vacc0123p0 = wasm_v128_load(w);
84 v128_t vacc4567p0 = wasm_v128_load(w + 4);
85
86
87 const v128_t vi0x0123 = wasm_v128_load(i0);
88 const v128_t vi0x4567 = wasm_v128_load(i0 + 4);
89 i0 += 8;
90
91 const v128_t vk0x0123 = wasm_v128_load(w + 8);
92 const v128_t vk0x4567 = wasm_v128_load(w + 12);
93 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
94 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi0x4567, vk0x4567));
95
96 const v128_t vi1x0123 = wasm_v128_load(i1);
97 const v128_t vi1x4567 = wasm_v128_load(i1 + 4);
98 i1 += 8;
99
100 const v128_t vk1x0123 = wasm_v128_load(w + 16);
101 const v128_t vk1x4567 = wasm_v128_load(w + 20);
102 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
103 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi1x4567, vk1x4567));
104
105 const v128_t vi2x0123 = wasm_v128_load(i2);
106 const v128_t vi2x4567 = wasm_v128_load(i2 + 4);
107 i2 += 8;
108
109 const v128_t vk2x0123 = wasm_v128_load(w + 24);
110 const v128_t vk2x4567 = wasm_v128_load(w + 28);
111 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
112 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567));
113
114 const v128_t vi3x0123 = wasm_v128_load(i3);
115 const v128_t vi3x4567 = wasm_v128_load(i3 + 4);
116 i3 += 8;
117
118 const v128_t vk3x0123 = wasm_v128_load(w + 32);
119 const v128_t vk3x4567 = wasm_v128_load(w + 36);
120 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
121 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi3x4567, vk3x4567));
122
123 const v128_t vi4x0123 = wasm_v128_load(i4);
124 const v128_t vi4x4567 = wasm_v128_load(i4 + 4);
125 i4 += 8;
126
127 const v128_t vk4x0123 = wasm_v128_load(w + 40);
128 const v128_t vk4x4567 = wasm_v128_load(w + 44);
129 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));
130 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567));
131
132 const v128_t vi5x0123 = wasm_v128_load(i5);
133 const v128_t vi5x4567 = wasm_v128_load(i5 + 4);
134 i5 += 8;
135
136 const v128_t vk5x0123 = wasm_v128_load(w + 48);
137 const v128_t vk5x4567 = wasm_v128_load(w + 52);
138 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123));
139 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi5x4567, vk5x4567));
140
141 const v128_t vi6x0123 = wasm_v128_load(i6);
142 const v128_t vi6x4567 = wasm_v128_load(i6 + 4);
143 i6 += 8;
144
145 const v128_t vk6x0123 = wasm_v128_load(w + 56);
146 const v128_t vk6x4567 = wasm_v128_load(w + 60);
147 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));
148 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi6x4567, vk6x4567));
149
150 const v128_t vi7x0123 = wasm_v128_load(i7);
151 const v128_t vi7x4567 = wasm_v128_load(i7 + 4);
152 i7 += 8;
153
154 const v128_t vk7x0123 = wasm_v128_load(w + 64);
155 const v128_t vk7x4567 = wasm_v128_load(w + 68);
156 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123));
157 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi7x4567, vk7x4567));
158
159 const v128_t vi8x0123 = wasm_v128_load(i8);
160 const v128_t vi8x4567 = wasm_v128_load(i8 + 4);
161 i8 += 8;
162
163 const v128_t vk8x0123 = wasm_v128_load(w + 72);
164 const v128_t vk8x4567 = wasm_v128_load(w + 76);
165 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));
166 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi8x4567, vk8x4567));
167
168 w += 80;
169
170
171 const v128_t vacc0123 = vacc0123p0;
172 const v128_t vacc4567 = vacc4567p0;
173
174 wasm_v128_store(output, vacc0123);
175 wasm_v128_store(output + 4, vacc4567);
176 output += 8;
177 }
178 for (; c >= 4; c -= 4) {
179 v128_t vacc0123p0 = wasm_v128_load(w);
180
181 const v128_t vi0x0123 = wasm_v128_load(i0);
182 i0 += 4;
183
184 const v128_t vk0x0123 = wasm_v128_load(w + 8);
185 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
186
187 const v128_t vi1x0123 = wasm_v128_load(i1);
188 i1 += 4;
189
190 const v128_t vk1x0123 = wasm_v128_load(w + 16);
191 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
192
193 const v128_t vi2x0123 = wasm_v128_load(i2);
194 i2 += 4;
195
196 const v128_t vk2x0123 = wasm_v128_load(w + 24);
197 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
198
199 const v128_t vi3x0123 = wasm_v128_load(i3);
200 i3 += 4;
201
202 const v128_t vk3x0123 = wasm_v128_load(w + 32);
203 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
204
205 const v128_t vi4x0123 = wasm_v128_load(i4);
206 i4 += 4;
207
208 const v128_t vk4x0123 = wasm_v128_load(w + 40);
209 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));
210
211 const v128_t vi5x0123 = wasm_v128_load(i5);
212 i5 += 4;
213
214 const v128_t vk5x0123 = wasm_v128_load(w + 48);
215 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123));
216
217 const v128_t vi6x0123 = wasm_v128_load(i6);
218 i6 += 4;
219
220 const v128_t vk6x0123 = wasm_v128_load(w + 56);
221 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));
222
223 const v128_t vi7x0123 = wasm_v128_load(i7);
224 i7 += 4;
225
226 const v128_t vk7x0123 = wasm_v128_load(w + 64);
227 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123));
228
229 const v128_t vi8x0123 = wasm_v128_load(i8);
230 i8 += 4;
231
232 const v128_t vk8x0123 = wasm_v128_load(w + 72);
233 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));
234
235 w += 4;
236
237
238 const v128_t vacc0123 = vacc0123p0;
239
240 wasm_v128_store(output, vacc0123);
241 output += 4;
242 }
243 if XNN_UNLIKELY(c != 0) {
244 v128_t vacc0123p0 = wasm_v128_load(w);
245
246 const v128_t vi0x0123 = wasm_v128_load(i0);
247 const v128_t vk0x0123 = wasm_v128_load(w + 8);
248 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
249
250 const v128_t vi1x0123 = wasm_v128_load(i1);
251 const v128_t vk1x0123 = wasm_v128_load(w + 16);
252 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
253
254 const v128_t vi2x0123 = wasm_v128_load(i2);
255 const v128_t vk2x0123 = wasm_v128_load(w + 24);
256 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
257
258 const v128_t vi3x0123 = wasm_v128_load(i3);
259 const v128_t vk3x0123 = wasm_v128_load(w + 32);
260 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
261
262 const v128_t vi4x0123 = wasm_v128_load(i4);
263 const v128_t vk4x0123 = wasm_v128_load(w + 40);
264 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123));
265
266 const v128_t vi5x0123 = wasm_v128_load(i5);
267 const v128_t vk5x0123 = wasm_v128_load(w + 48);
268 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123));
269
270 const v128_t vi6x0123 = wasm_v128_load(i6);
271 const v128_t vk6x0123 = wasm_v128_load(w + 56);
272 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123));
273
274 const v128_t vi7x0123 = wasm_v128_load(i7);
275 const v128_t vk7x0123 = wasm_v128_load(w + 64);
276 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123));
277
278 const v128_t vi8x0123 = wasm_v128_load(i8);
279 const v128_t vk8x0123 = wasm_v128_load(w + 72);
280 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123));
281
282
283 v128_t vacc0123 = vacc0123p0;
284
285 if (c & 2) {
286 *((double*) output) = wasm_f64x2_extract_lane(vacc0123, 0);
287 vacc0123 = wasm_v32x4_shuffle(vacc0123, vacc0123, 2, 3, 2, 3);
288 output += 2;
289 }
290 if (c & 1) {
291 *output = wasm_f32x4_extract_lane(vacc0123, 0);
292 output += 1;
293 }
294 }
295
296 output = (float*) ((uintptr_t) output + output_increment);
297 } while (--output_width != 0);
298 }
299