1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-wasmsimd-mul16.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_qu8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16(
18 size_t channels,
19 size_t output_width,
20 const uint8_t** input,
21 const void* weights,
22 uint8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const uint8_t* zero,
27 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const v128_t vkernel_zero_point = wasm_u32x4_load16x4(params->fp32_wasmsimd.kernel_zero_point);
33 do {
34 const uint8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const uint8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const uint8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const uint8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const uint8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const uint8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const uint8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const uint8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const uint8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 const uint8_t* i9 = input[9];
80 assert(i9 != NULL);
81 if XNN_UNPREDICTABLE(i9 != zero) {
82 i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
83 }
84 const uint8_t* i10 = input[10];
85 assert(i10 != NULL);
86 if XNN_UNPREDICTABLE(i10 != zero) {
87 i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
88 }
89 const uint8_t* i11 = input[11];
90 assert(i11 != NULL);
91 if XNN_UNPREDICTABLE(i11 != zero) {
92 i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
93 }
94 const uint8_t* i12 = input[12];
95 assert(i12 != NULL);
96 if XNN_UNPREDICTABLE(i12 != zero) {
97 i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
98 }
99 const uint8_t* i13 = input[13];
100 assert(i13 != NULL);
101 if XNN_UNPREDICTABLE(i13 != zero) {
102 i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
103 }
104 const uint8_t* i14 = input[14];
105 assert(i14 != NULL);
106 if XNN_UNPREDICTABLE(i14 != zero) {
107 i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
108 }
109 const uint8_t* i15 = input[15];
110 assert(i15 != NULL);
111 if XNN_UNPREDICTABLE(i15 != zero) {
112 i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
113 }
114 const uint8_t* i16 = input[16];
115 assert(i16 != NULL);
116 if XNN_UNPREDICTABLE(i16 != zero) {
117 i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
118 }
119 const uint8_t* i17 = input[17];
120 assert(i17 != NULL);
121 if XNN_UNPREDICTABLE(i17 != zero) {
122 i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
123 }
124 const uint8_t* i18 = input[18];
125 assert(i18 != NULL);
126 if XNN_UNPREDICTABLE(i18 != zero) {
127 i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
128 }
129 const uint8_t* i19 = input[19];
130 assert(i19 != NULL);
131 if XNN_UNPREDICTABLE(i19 != zero) {
132 i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
133 }
134 const uint8_t* i20 = input[20];
135 assert(i20 != NULL);
136 if XNN_UNPREDICTABLE(i20 != zero) {
137 i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
138 }
139 const uint8_t* i21 = input[21];
140 assert(i21 != NULL);
141 if XNN_UNPREDICTABLE(i21 != zero) {
142 i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
143 }
144 const uint8_t* i22 = input[22];
145 assert(i22 != NULL);
146 if XNN_UNPREDICTABLE(i22 != zero) {
147 i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
148 }
149 const uint8_t* i23 = input[23];
150 assert(i23 != NULL);
151 if XNN_UNPREDICTABLE(i23 != zero) {
152 i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
153 }
154 const uint8_t* i24 = input[24];
155 assert(i24 != NULL);
156 if XNN_UNPREDICTABLE(i24 != zero) {
157 i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
158 }
159 input = (const uint8_t**) ((uintptr_t) input + input_stride);
160
161 size_t c = channels;
162 const void* w = weights;
163 for (; c >= 8; c -= 8) {
164 v128_t vacc0123 = wasm_v128_load(w);
165 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
166
167
168 const v128_t vi0x01234567 = wasm_u16x8_load8x8(i0);
169 const v128_t vk0x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
170 i0 += 8;
171
172 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
173
174 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
175 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
176
177 const v128_t vi1x01234567 = wasm_u16x8_load8x8(i1);
178 const v128_t vk1x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
179 v128_t vsumx01234567 = wasm_i16x8_add(vi0x01234567, vi1x01234567);
180 i1 += 8;
181
182 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
183
184 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
185 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
186
187 const v128_t vi2x01234567 = wasm_u16x8_load8x8(i2);
188 const v128_t vk2x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
189 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi2x01234567);
190 i2 += 8;
191
192 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
193
194 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
195 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
196
197 const v128_t vi3x01234567 = wasm_u16x8_load8x8(i3);
198 const v128_t vk3x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
199 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi3x01234567);
200 i3 += 8;
201
202 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
203
204 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
205 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
206
207 const v128_t vi4x01234567 = wasm_u16x8_load8x8(i4);
208 const v128_t vk4x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
209 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi4x01234567);
210 i4 += 8;
211
212 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
213
214 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
215 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
216
217 const v128_t vi5x01234567 = wasm_u16x8_load8x8(i5);
218 const v128_t vk5x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
219 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi5x01234567);
220 i5 += 8;
221
222 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
223
224 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
225 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
226
227 const v128_t vi6x01234567 = wasm_u16x8_load8x8(i6);
228 const v128_t vk6x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
229 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi6x01234567);
230 i6 += 8;
231
232 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
233
234 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
235 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
236
237 const v128_t vi7x01234567 = wasm_u16x8_load8x8(i7);
238 const v128_t vk7x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
239 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi7x01234567);
240 i7 += 8;
241
242 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
243
244 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
245 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
246
247 const v128_t vi8x01234567 = wasm_u16x8_load8x8(i8);
248 const v128_t vk8x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
249 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi8x01234567);
250 i8 += 8;
251
252 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
253
254 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
255 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
256
257 const v128_t vi9x01234567 = wasm_u16x8_load8x8(i9);
258 const v128_t vk9x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
259 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi9x01234567);
260 i9 += 8;
261
262 vprod01234567 = wasm_i16x8_mul(vi9x01234567, vk9x01234567);
263
264 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
265 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
266
267 const v128_t vi10x01234567 = wasm_u16x8_load8x8(i10);
268 const v128_t vk10x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
269 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi10x01234567);
270 i10 += 8;
271
272 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
273
274 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
275 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
276
277 const v128_t vi11x01234567 = wasm_u16x8_load8x8(i11);
278 const v128_t vk11x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
279 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi11x01234567);
280 i11 += 8;
281
282 vprod01234567 = wasm_i16x8_mul(vi11x01234567, vk11x01234567);
283
284 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
285 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
286
287 const v128_t vi12x01234567 = wasm_u16x8_load8x8(i12);
288 const v128_t vk12x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
289 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi12x01234567);
290 i12 += 8;
291
292 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
293
294 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
295 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
296
297 const v128_t vi13x01234567 = wasm_u16x8_load8x8(i13);
298 const v128_t vk13x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
299 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi13x01234567);
300 i13 += 8;
301
302 vprod01234567 = wasm_i16x8_mul(vi13x01234567, vk13x01234567);
303
304 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
305 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
306
307 const v128_t vi14x01234567 = wasm_u16x8_load8x8(i14);
308 const v128_t vk14x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
309 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi14x01234567);
310 i14 += 8;
311
312 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
313
314 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
315 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
316
317 const v128_t vi15x01234567 = wasm_u16x8_load8x8(i15);
318 const v128_t vk15x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
319 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi15x01234567);
320 i15 += 8;
321
322 vprod01234567 = wasm_i16x8_mul(vi15x01234567, vk15x01234567);
323
324 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
325 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
326
327 const v128_t vi16x01234567 = wasm_u16x8_load8x8(i16);
328 const v128_t vk16x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
329 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi16x01234567);
330 i16 += 8;
331
332 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
333
334 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
335 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
336
337 const v128_t vi17x01234567 = wasm_u16x8_load8x8(i17);
338 const v128_t vk17x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
339 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi17x01234567);
340 i17 += 8;
341
342 vprod01234567 = wasm_i16x8_mul(vi17x01234567, vk17x01234567);
343
344 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
345 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
346
347 const v128_t vi18x01234567 = wasm_u16x8_load8x8(i18);
348 const v128_t vk18x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
349 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi18x01234567);
350 i18 += 8;
351
352 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
353
354 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
355 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
356
357 const v128_t vi19x01234567 = wasm_u16x8_load8x8(i19);
358 const v128_t vk19x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
359 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi19x01234567);
360 i19 += 8;
361
362 vprod01234567 = wasm_i16x8_mul(vi19x01234567, vk19x01234567);
363
364 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
365 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
366
367 const v128_t vi20x01234567 = wasm_u16x8_load8x8(i20);
368 const v128_t vk20x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
369 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi20x01234567);
370 i20 += 8;
371
372 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
373
374 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
375 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
376
377 const v128_t vi21x01234567 = wasm_u16x8_load8x8(i21);
378 const v128_t vk21x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
379 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi21x01234567);
380 i21 += 8;
381
382 vprod01234567 = wasm_i16x8_mul(vi21x01234567, vk21x01234567);
383
384 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
385 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
386
387 const v128_t vi22x01234567 = wasm_u16x8_load8x8(i22);
388 const v128_t vk22x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
389 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi22x01234567);
390 i22 += 8;
391
392 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
393
394 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
395 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
396
397 const v128_t vi23x01234567 = wasm_u16x8_load8x8(i23);
398 const v128_t vk23x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
399 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi23x01234567);
400 i23 += 8;
401
402 vprod01234567 = wasm_i16x8_mul(vi23x01234567, vk23x01234567);
403
404 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
405 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
406
407 const v128_t vi24x01234567 = wasm_u16x8_load8x8(i24);
408 const v128_t vk24x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
409 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi24x01234567);
410 i24 += 8;
411
412 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
413
414 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
415 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
416
417 vacc0123 = wasm_i32x4_sub(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx01234567), vkernel_zero_point));
418 vacc4567 = wasm_i32x4_sub(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx01234567), vkernel_zero_point));
419
420 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(uint8_t));
421
422 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
423 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
424
425 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
426 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
427 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
428
429 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
430 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
431 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
432
433 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
434 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
435 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
436
437 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
438 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
439 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
440
441 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
442
443 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
444
445 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
446 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
447
448 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
449 output += 8;
450 }
451 if XNN_UNLIKELY(c != 0) {
452 {
453 v128_t vacc0123 = wasm_v128_load(w);
454 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
455
456
457 const v128_t vi0x01234567 = wasm_u16x8_load8x8(i0);
458 const v128_t vk0x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
459
460 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
461
462 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
463 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
464
465 const v128_t vi1x01234567 = wasm_u16x8_load8x8(i1);
466 const v128_t vk1x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
467 v128_t vsumx01234567 = wasm_i16x8_add(vi0x01234567, vi1x01234567);
468
469 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
470
471 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
472 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
473
474 const v128_t vi2x01234567 = wasm_u16x8_load8x8(i2);
475 const v128_t vk2x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
476 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi2x01234567);
477
478 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
479
480 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
481 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
482
483 const v128_t vi3x01234567 = wasm_u16x8_load8x8(i3);
484 const v128_t vk3x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
485 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi3x01234567);
486
487 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
488
489 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
490 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
491
492 const v128_t vi4x01234567 = wasm_u16x8_load8x8(i4);
493 const v128_t vk4x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
494 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi4x01234567);
495
496 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
497
498 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
499 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
500
501 const v128_t vi5x01234567 = wasm_u16x8_load8x8(i5);
502 const v128_t vk5x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
503 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi5x01234567);
504
505 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
506
507 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
508 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
509
510 const v128_t vi6x01234567 = wasm_u16x8_load8x8(i6);
511 const v128_t vk6x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
512 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi6x01234567);
513
514 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
515
516 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
517 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
518
519 const v128_t vi7x01234567 = wasm_u16x8_load8x8(i7);
520 const v128_t vk7x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
521 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi7x01234567);
522
523 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
524
525 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
526 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
527
528 const v128_t vi8x01234567 = wasm_u16x8_load8x8(i8);
529 const v128_t vk8x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
530 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi8x01234567);
531
532 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
533
534 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
535 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
536
537 const v128_t vi9x01234567 = wasm_u16x8_load8x8(i9);
538 const v128_t vk9x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
539 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi9x01234567);
540
541 vprod01234567 = wasm_i16x8_mul(vi9x01234567, vk9x01234567);
542
543 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
544 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
545
546 const v128_t vi10x01234567 = wasm_u16x8_load8x8(i10);
547 const v128_t vk10x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
548 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi10x01234567);
549
550 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
551
552 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
553 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
554
555 const v128_t vi11x01234567 = wasm_u16x8_load8x8(i11);
556 const v128_t vk11x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
557 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi11x01234567);
558
559 vprod01234567 = wasm_i16x8_mul(vi11x01234567, vk11x01234567);
560
561 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
562 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
563
564 const v128_t vi12x01234567 = wasm_u16x8_load8x8(i12);
565 const v128_t vk12x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
566 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi12x01234567);
567
568 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
569
570 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
571 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
572
573 const v128_t vi13x01234567 = wasm_u16x8_load8x8(i13);
574 const v128_t vk13x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
575 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi13x01234567);
576
577 vprod01234567 = wasm_i16x8_mul(vi13x01234567, vk13x01234567);
578
579 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
580 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
581
582 const v128_t vi14x01234567 = wasm_u16x8_load8x8(i14);
583 const v128_t vk14x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
584 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi14x01234567);
585
586 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
587
588 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
589 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
590
591 const v128_t vi15x01234567 = wasm_u16x8_load8x8(i15);
592 const v128_t vk15x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
593 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi15x01234567);
594
595 vprod01234567 = wasm_i16x8_mul(vi15x01234567, vk15x01234567);
596
597 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
598 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
599
600 const v128_t vi16x01234567 = wasm_u16x8_load8x8(i16);
601 const v128_t vk16x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
602 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi16x01234567);
603
604 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
605
606 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
607 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
608
609 const v128_t vi17x01234567 = wasm_u16x8_load8x8(i17);
610 const v128_t vk17x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
611 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi17x01234567);
612
613 vprod01234567 = wasm_i16x8_mul(vi17x01234567, vk17x01234567);
614
615 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
616 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
617
618 const v128_t vi18x01234567 = wasm_u16x8_load8x8(i18);
619 const v128_t vk18x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
620 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi18x01234567);
621
622 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
623
624 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
625 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
626
627 const v128_t vi19x01234567 = wasm_u16x8_load8x8(i19);
628 const v128_t vk19x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
629 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi19x01234567);
630
631 vprod01234567 = wasm_i16x8_mul(vi19x01234567, vk19x01234567);
632
633 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
634 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
635
636 const v128_t vi20x01234567 = wasm_u16x8_load8x8(i20);
637 const v128_t vk20x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
638 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi20x01234567);
639
640 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
641
642 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
643 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
644
645 const v128_t vi21x01234567 = wasm_u16x8_load8x8(i21);
646 const v128_t vk21x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
647 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi21x01234567);
648
649 vprod01234567 = wasm_i16x8_mul(vi21x01234567, vk21x01234567);
650
651 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
652 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
653
654 const v128_t vi22x01234567 = wasm_u16x8_load8x8(i22);
655 const v128_t vk22x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
656 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi22x01234567);
657
658 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
659
660 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
661 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
662
663 const v128_t vi23x01234567 = wasm_u16x8_load8x8(i23);
664 const v128_t vk23x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
665 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi23x01234567);
666
667 vprod01234567 = wasm_i16x8_mul(vi23x01234567, vk23x01234567);
668
669 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
670 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
671
672 const v128_t vi24x01234567 = wasm_u16x8_load8x8(i24);
673 const v128_t vk24x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
674 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi24x01234567);
675
676 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
677
678 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
679 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
680
681
682 vacc0123 = wasm_i32x4_sub(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx01234567), vkernel_zero_point));
683 vacc4567 = wasm_i32x4_sub(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx01234567), vkernel_zero_point));
684
685 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
686 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
687
688 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
689 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
690 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
691
692 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
693 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
694 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
695
696 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
697 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
698 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
699
700 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
701 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
702 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
703
704 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
705 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
706
707 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
708 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
709
710
711 if (c & 4) {
712 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
713 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
714 output += 4;
715 }
716 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
717 if (c & 2) {
718 *((uint16_t*) output) = (uint16_t) vout0123;
719 vout0123 >>= 16;
720 output += 2;
721 }
722 if (c & 1) {
723 *output = (uint8_t) vout0123;
724 output += 1;
725 }
726 }
727 }
728
729 output = (uint8_t*) ((uintptr_t) output + output_increment);
730 } while (--output_width != 0);
731 }
732