xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up8x25-minmax-fp32-wasmsimd-mul16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_qs8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16(
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_qs8_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     const int8_t* i9 = input[9];
79     assert(i9 != NULL);
80     if XNN_UNPREDICTABLE(i9 != zero) {
81       i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
82     }
83     const int8_t* i10 = input[10];
84     assert(i10 != NULL);
85     if XNN_UNPREDICTABLE(i10 != zero) {
86       i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
87     }
88     const int8_t* i11 = input[11];
89     assert(i11 != NULL);
90     if XNN_UNPREDICTABLE(i11 != zero) {
91       i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
92     }
93     const int8_t* i12 = input[12];
94     assert(i12 != NULL);
95     if XNN_UNPREDICTABLE(i12 != zero) {
96       i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
97     }
98     const int8_t* i13 = input[13];
99     assert(i13 != NULL);
100     if XNN_UNPREDICTABLE(i13 != zero) {
101       i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
102     }
103     const int8_t* i14 = input[14];
104     assert(i14 != NULL);
105     if XNN_UNPREDICTABLE(i14 != zero) {
106       i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
107     }
108     const int8_t* i15 = input[15];
109     assert(i15 != NULL);
110     if XNN_UNPREDICTABLE(i15 != zero) {
111       i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
112     }
113     const int8_t* i16 = input[16];
114     assert(i16 != NULL);
115     if XNN_UNPREDICTABLE(i16 != zero) {
116       i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
117     }
118     const int8_t* i17 = input[17];
119     assert(i17 != NULL);
120     if XNN_UNPREDICTABLE(i17 != zero) {
121       i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
122     }
123     const int8_t* i18 = input[18];
124     assert(i18 != NULL);
125     if XNN_UNPREDICTABLE(i18 != zero) {
126       i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
127     }
128     const int8_t* i19 = input[19];
129     assert(i19 != NULL);
130     if XNN_UNPREDICTABLE(i19 != zero) {
131       i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
132     }
133     const int8_t* i20 = input[20];
134     assert(i20 != NULL);
135     if XNN_UNPREDICTABLE(i20 != zero) {
136       i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
137     }
138     const int8_t* i21 = input[21];
139     assert(i21 != NULL);
140     if XNN_UNPREDICTABLE(i21 != zero) {
141       i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
142     }
143     const int8_t* i22 = input[22];
144     assert(i22 != NULL);
145     if XNN_UNPREDICTABLE(i22 != zero) {
146       i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
147     }
148     const int8_t* i23 = input[23];
149     assert(i23 != NULL);
150     if XNN_UNPREDICTABLE(i23 != zero) {
151       i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
152     }
153     const int8_t* i24 = input[24];
154     assert(i24 != NULL);
155     if XNN_UNPREDICTABLE(i24 != zero) {
156       i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
157     }
158     input = (const int8_t**) ((uintptr_t) input + input_stride);
159 
160     size_t c = channels;
161     const void* w = weights;
162     for (; c >= 8; c -= 8) {
163       v128_t vacc0123 = wasm_v128_load(w);
164       v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
165 
166 
167       const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
168       const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
169       i0 += 8;
170 
171       v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
172 
173       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
174       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
175 
176       const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
177       const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
178       i1 += 8;
179 
180       vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
181 
182       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
183       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
184 
185       const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
186       const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
187       i2 += 8;
188 
189       vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
190 
191       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
192       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
193 
194       const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
195       const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
196       i3 += 8;
197 
198       vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
199 
200       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
201       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
202 
203       const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
204       const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
205       i4 += 8;
206 
207       vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
208 
209       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
210       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
211 
212       const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
213       const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
214       i5 += 8;
215 
216       vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
217 
218       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
219       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
220 
221       const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
222       const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
223       i6 += 8;
224 
225       vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
226 
227       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
228       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
229 
230       const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
231       const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
232       i7 += 8;
233 
234       vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
235 
236       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
237       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
238 
239       const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
240       const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
241       i8 += 8;
242 
243       vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
244 
245       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
246       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
247 
248       const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
249       const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
250       i9 += 8;
251 
252       vprod01234567 = wasm_i16x8_mul(vi9x01234567, vk9x01234567);
253 
254       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
255       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
256 
257       const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
258       const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
259       i10 += 8;
260 
261       vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
262 
263       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
264       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
265 
266       const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
267       const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
268       i11 += 8;
269 
270       vprod01234567 = wasm_i16x8_mul(vi11x01234567, vk11x01234567);
271 
272       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
273       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
274 
275       const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
276       const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
277       i12 += 8;
278 
279       vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
280 
281       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
282       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
283 
284       const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
285       const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
286       i13 += 8;
287 
288       vprod01234567 = wasm_i16x8_mul(vi13x01234567, vk13x01234567);
289 
290       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
291       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
292 
293       const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
294       const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
295       i14 += 8;
296 
297       vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
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 vi15x01234567 = wasm_i16x8_load8x8(i15);
303       const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
304       i15 += 8;
305 
306       vprod01234567 = wasm_i16x8_mul(vi15x01234567, vk15x01234567);
307 
308       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
309       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
310 
311       const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
312       const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
313       i16 += 8;
314 
315       vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
316 
317       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
318       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
319 
320       const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
321       const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
322       i17 += 8;
323 
324       vprod01234567 = wasm_i16x8_mul(vi17x01234567, vk17x01234567);
325 
326       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
327       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
328 
329       const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
330       const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
331       i18 += 8;
332 
333       vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
334 
335       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
336       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
337 
338       const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
339       const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
340       i19 += 8;
341 
342       vprod01234567 = wasm_i16x8_mul(vi19x01234567, vk19x01234567);
343 
344       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
345       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
346 
347       const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
348       const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
349       i20 += 8;
350 
351       vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
352 
353       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
354       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
355 
356       const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
357       const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
358       i21 += 8;
359 
360       vprod01234567 = wasm_i16x8_mul(vi21x01234567, vk21x01234567);
361 
362       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
363       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
364 
365       const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
366       const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
367       i22 += 8;
368 
369       vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
370 
371       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
372       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
373 
374       const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
375       const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
376       i23 += 8;
377 
378       vprod01234567 = wasm_i16x8_mul(vi23x01234567, vk23x01234567);
379 
380       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
381       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
382 
383       const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
384       const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
385       i24 += 8;
386 
387       vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
388 
389       vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
390       vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
391 
392 
393       w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
394 
395       vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
396       vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
397 
398       const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
399       vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
400       vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
401 
402       const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
403       vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
404       vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
405 
406       const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
407       vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
408       vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
409 
410       const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
411       vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
412       vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
413 
414       v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
415 
416       v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
417 
418       const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
419       vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
420 
421       *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
422       output += 8;
423     }
424     if XNN_UNLIKELY(c != 0) {
425       {
426         v128_t vacc0123 = wasm_v128_load(w);
427         v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
428 
429 
430         const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
431         const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
432 
433         v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
434 
435         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
436         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
437 
438         const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
439         const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
440 
441         vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
442 
443         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
444         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
445 
446         const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
447         const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
448 
449         vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
450 
451         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
452         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
453 
454         const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
455         const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
456 
457         vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
458 
459         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
460         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
461 
462         const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
463         const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
464 
465         vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
466 
467         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
468         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
469 
470         const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
471         const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
472 
473         vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
474 
475         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
476         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
477 
478         const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
479         const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
480 
481         vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
482 
483         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
484         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
485 
486         const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
487         const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
488 
489         vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
490 
491         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
492         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
493 
494         const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
495         const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
496 
497         vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
498 
499         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
500         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
501 
502         const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
503         const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
504 
505         vprod01234567 = wasm_i16x8_mul(vi9x01234567, vk9x01234567);
506 
507         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
508         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
509 
510         const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
511         const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
512 
513         vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
514 
515         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
516         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
517 
518         const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
519         const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
520 
521         vprod01234567 = wasm_i16x8_mul(vi11x01234567, vk11x01234567);
522 
523         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
524         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
525 
526         const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
527         const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
528 
529         vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
530 
531         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
532         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
533 
534         const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
535         const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
536 
537         vprod01234567 = wasm_i16x8_mul(vi13x01234567, vk13x01234567);
538 
539         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
540         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
541 
542         const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
543         const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
544 
545         vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
546 
547         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
548         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
549 
550         const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
551         const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
552 
553         vprod01234567 = wasm_i16x8_mul(vi15x01234567, vk15x01234567);
554 
555         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
556         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
557 
558         const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
559         const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
560 
561         vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
562 
563         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
564         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
565 
566         const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
567         const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
568 
569         vprod01234567 = wasm_i16x8_mul(vi17x01234567, vk17x01234567);
570 
571         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
572         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
573 
574         const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
575         const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
576 
577         vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
578 
579         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
580         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
581 
582         const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
583         const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
584 
585         vprod01234567 = wasm_i16x8_mul(vi19x01234567, vk19x01234567);
586 
587         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
588         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
589 
590         const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
591         const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
592 
593         vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
594 
595         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
596         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
597 
598         const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
599         const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
600 
601         vprod01234567 = wasm_i16x8_mul(vi21x01234567, vk21x01234567);
602 
603         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
604         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
605 
606         const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
607         const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
608 
609         vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
610 
611         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
612         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
613 
614         const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
615         const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
616 
617         vprod01234567 = wasm_i16x8_mul(vi23x01234567, vk23x01234567);
618 
619         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
620         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
621 
622         const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
623         const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
624 
625         vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
626 
627         vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
628         vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
629 
630 
631 
632       vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
633       vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
634 
635       const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
636       vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
637       vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
638 
639       const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
640       vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
641       vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
642 
643       const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
644       vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
645       vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
646 
647       const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
648       vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
649       vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
650 
651       v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
652       v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
653 
654       const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
655       vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
656 
657 
658       if (c & 4) {
659         *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
660         vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
661         output += 4;
662       }
663       uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
664       if (c & 2) {
665         *((uint16_t*) output) = (uint16_t) vout0123;
666         vout0123 >>= 16;
667         output += 2;
668       }
669       if (c & 1) {
670         *output = (int8_t) vout0123;
671         output += 1;
672       }
673       }
674     }
675 
676     output = (int8_t*) ((uintptr_t) output + output_increment);
677   } while (--output_width != 0);
678 }
679