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_qc8_dwconv_minmax_fp32_ukernel_up16x9__wasmsimd_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__wasmsimd_mul16_add16(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qc8_conv_minmax_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 int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 input = (const int8_t**) ((uintptr_t) input + input_stride);
79
80 size_t c = channels;
81 const void* w = weights;
82 for (; c >= 16; c -= 16) {
83 v128_t vacc0123 = wasm_v128_load(w);
84 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
85 v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
86 v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
87
88
89 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
90 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
91 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
92 const v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
93 i0 += 16;
94
95 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
96 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
97
98
99 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
100 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
101 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
102 const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
103 i1 += 16;
104
105 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
106 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF));
107
108 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
109 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
110 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
111 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
112
113 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
114 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
115 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
116 const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
117 i2 += 16;
118
119 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
120 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
121
122
123 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
124 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
125 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
126 const v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
127 i3 += 16;
128
129 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
130 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF));
131
132 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
133 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
134 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
135 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
136
137 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
138 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
139 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
140 const v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
141 i4 += 16;
142
143 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
144 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
145
146
147 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
148 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
149 const v128_t vi5x89ABCDEF = wasm_i16x8_load8x8(i5 + 8);
150 const v128_t vk5x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
151 i5 += 16;
152
153 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
154 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF));
155
156 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
157 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
158 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
159 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
160
161 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
162 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
163 const v128_t vi6x89ABCDEF = wasm_i16x8_load8x8(i6 + 8);
164 const v128_t vk6x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
165 i6 += 16;
166
167 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
168 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
169
170
171 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
172 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
173 const v128_t vi7x89ABCDEF = wasm_i16x8_load8x8(i7 + 8);
174 const v128_t vk7x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
175 i7 += 16;
176
177 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
178 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF));
179
180 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
181 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
182 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
183 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
184
185 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
186 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
187 const v128_t vi8x89ABCDEF = wasm_i16x8_load8x8(i8 + 8);
188 const v128_t vk8x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
189 i8 += 16;
190
191 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
192 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
193
194 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
195 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
196 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
197 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
198
199
200 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
201
202 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
203 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
204 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
205 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
206
207 const v128_t vscale0123 = wasm_v128_load(w);
208 const v128_t vscale4567 = wasm_v128_load((const float*) w + 4);
209 const v128_t vscale89AB = wasm_v128_load((const float*) w + 8);
210 const v128_t vscaleCDEF = wasm_v128_load((const float*) w + 12);
211 w = (const void*) ((const float*) w + 16);
212
213 vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
214 vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
215 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale89AB);
216 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscaleCDEF);
217
218 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
219 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
220 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
221 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
222 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
223
224 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
225 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
226 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
227 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
228 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
229
230 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
231 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
232 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
233 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
234 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
235
236 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
237 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
238
239 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
240
241 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
242 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
243
244 wasm_v128_store(output, vout0123456789ABCDEF);
245 output += 16;
246 }
247 if XNN_UNLIKELY(c != 0) {
248 const int8_t* k = (const int8_t*) ((uintptr_t) w + 16 * sizeof(int32_t));
249 do {
250 v128_t vacc0123 = wasm_v128_load(w);
251 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
252
253
254 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
255 const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
256 i0 += 8;
257
258 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
259
260
261 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
262 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 16));
263 i1 += 8;
264
265 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
266
267 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
268 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
269
270 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
271 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 32));
272 i2 += 8;
273
274 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
275
276
277 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
278 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) (k + 48));
279 i3 += 8;
280
281 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
282
283 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
284 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
285
286 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
287 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) (k + 64));
288 i4 += 8;
289
290 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
291
292
293 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
294 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) (k + 80));
295 i5 += 8;
296
297 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
298
299 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
300 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
301
302 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
303 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) (k + 96));
304 i6 += 8;
305
306 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
307
308
309 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
310 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) (k + 112));
311 i7 += 8;
312
313 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
314
315 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
316 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
317
318 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
319 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) (k + 128));
320 i8 += 8;
321
322 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
323
324 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
325 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
326
327 k += 8;
328
329
330 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
331 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
332
333 const v128_t vscale0123 = wasm_v128_load((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
334 const v128_t vscale4567 = wasm_v128_load((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t) + 4 * sizeof(float)));
335
336 vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
337 vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
338
339 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
340 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
341 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
342
343 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
344 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
345 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
346
347 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
348 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
349 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
350
351 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
352 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
353
354 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
355 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
356
357 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
358
359 if XNN_LIKELY(c >= 8) {
360 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
361 output += 8;
362 c -= 8;
363 } else {
364 if (c & 4) {
365 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
366 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
367 output += 4;
368 }
369 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
370 if (c & 2) {
371 *((uint16_t*) output) = (uint16_t) vout0123;
372 vout0123 >>= 16;
373 output += 2;
374 }
375 if (c & 1) {
376 *output = (int8_t) vout0123;
377 output += 1;
378 }
379 c = 0;
380 }
381 } while (c != 0);
382 }
383
384 output = (int8_t*) ((uintptr_t) output + output_increment);
385 } while (--output_width != 0);
386 }
387