Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-25p8c-minmax-fp32-wasmsimd-mul16.c

raw
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 #include "xnnpack/math.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_25p8c__wasmsimd_mul16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 intptr_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33
34 const v128_t vscale = wasm_v128_load32_splat(&params->fp32_scalar.scale);
35 XNN_FORCE_REALIZATION(vscale);
36 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
37 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
38 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
39 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
40 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
41 XNN_FORCE_REALIZATION(vmagic_bias);
42 XNN_FORCE_REALIZATION(vmagic_min);
43 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
44 XNN_FORCE_REALIZATION(voutput_max);
45
46 do {
47 const int8_t* i0 = input[0];
48 assert(i0 != NULL);
49 if XNN_UNPREDICTABLE(i0 != zero) {
50 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
51 }
52 const int8_t* i1 = input[1];
53 assert(i1 != NULL);
54 if XNN_UNPREDICTABLE(i1 != zero) {
55 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
56 }
57 const int8_t* i2 = input[2];
58 assert(i2 != NULL);
59 if XNN_UNPREDICTABLE(i2 != zero) {
60 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
61 }
62 const int8_t* i3 = input[3];
63 assert(i3 != NULL);
64 if XNN_UNPREDICTABLE(i3 != zero) {
65 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
66 }
67 const int8_t* i4 = input[4];
68 assert(i4 != NULL);
69 if XNN_UNPREDICTABLE(i4 != zero) {
70 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
71 }
72 const int8_t* i5 = input[5];
73 assert(i5 != NULL);
74 if XNN_UNPREDICTABLE(i5 != zero) {
75 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
76 }
77 const int8_t* i6 = input[6];
78 assert(i6 != NULL);
79 if XNN_UNPREDICTABLE(i6 != zero) {
80 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
81 }
82 const int8_t* i7 = input[7];
83 assert(i7 != NULL);
84 if XNN_UNPREDICTABLE(i7 != zero) {
85 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
86 }
87 const int8_t* i8 = input[8];
88 assert(i8 != NULL);
89 if XNN_UNPREDICTABLE(i8 != zero) {
90 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
91 }
92 const int8_t* i9 = input[9];
93 assert(i9 != NULL);
94 if XNN_UNPREDICTABLE(i9 != zero) {
95 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
96 }
97 const int8_t* i10 = input[10];
98 assert(i10 != NULL);
99 if XNN_UNPREDICTABLE(i10 != zero) {
100 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
101 }
102 const int8_t* i11 = input[11];
103 assert(i11 != NULL);
104 if XNN_UNPREDICTABLE(i11 != zero) {
105 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
106 }
107 const int8_t* i12 = input[12];
108 assert(i12 != NULL);
109 if XNN_UNPREDICTABLE(i12 != zero) {
110 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
111 }
112 const int8_t* i13 = input[13];
113 assert(i13 != NULL);
114 if XNN_UNPREDICTABLE(i13 != zero) {
115 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
116 }
117 const int8_t* i14 = input[14];
118 assert(i14 != NULL);
119 if XNN_UNPREDICTABLE(i14 != zero) {
120 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
121 }
122 const int8_t* i15 = input[15];
123 assert(i15 != NULL);
124 if XNN_UNPREDICTABLE(i15 != zero) {
125 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
126 }
127 const int8_t* i16 = input[16];
128 assert(i16 != NULL);
129 if XNN_UNPREDICTABLE(i16 != zero) {
130 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
131 }
132 const int8_t* i17 = input[17];
133 assert(i17 != NULL);
134 if XNN_UNPREDICTABLE(i17 != zero) {
135 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
136 }
137 const int8_t* i18 = input[18];
138 assert(i18 != NULL);
139 if XNN_UNPREDICTABLE(i18 != zero) {
140 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
141 }
142 const int8_t* i19 = input[19];
143 assert(i19 != NULL);
144 if XNN_UNPREDICTABLE(i19 != zero) {
145 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
146 }
147 const int8_t* i20 = input[20];
148 assert(i20 != NULL);
149 if XNN_UNPREDICTABLE(i20 != zero) {
150 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
151 }
152 const int8_t* i21 = input[21];
153 assert(i21 != NULL);
154 if XNN_UNPREDICTABLE(i21 != zero) {
155 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
156 }
157 const int8_t* i22 = input[22];
158 assert(i22 != NULL);
159 if XNN_UNPREDICTABLE(i22 != zero) {
160 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
161 }
162 const int8_t* i23 = input[23];
163 assert(i23 != NULL);
164 if XNN_UNPREDICTABLE(i23 != zero) {
165 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
166 }
167 const int8_t* i24 = input[24];
168 assert(i24 != NULL);
169 if XNN_UNPREDICTABLE(i24 != zero) {
170 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
171 }
172 input = (const int8_t**) ((uintptr_t) input + input_stride);
173
174 size_t c = channels;
175 const void* w = weights;
176 for (; c >= 8; c -= 8) {
177 v128_t vacc0123 = wasm_v128_load(w);
178 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
179
180
181 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
182 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
183 i0 += 8;
184
185 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
186
187 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
188 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
189
190 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
191 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
192 i1 += 8;
193
194 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
195
196 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
197 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
198
199 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
200 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
201 i2 += 8;
202
203 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
204
205 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
206 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
207
208 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
209 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
210 i3 += 8;
211
212 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
213
214 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
215 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
216
217 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
218 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
219 i4 += 8;
220
221 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
222
223 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
224 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
225
226 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
227 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
228 i5 += 8;
229
230 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
231
232 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
233 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
234
235 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
236 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
237 i6 += 8;
238
239 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
240
241 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
242 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
243
244 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
245 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
246 i7 += 8;
247
248 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
249
250 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
251 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
252
253 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
254 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
255 i8 += 8;
256
257 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
258
259 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
260 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
261
262 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
263 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
264 i9 += 8;
265
266 vprod01234567 = wasm_i16x8_mul(vi9x01234567, vk9x01234567);
267
268 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
269 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
270
271 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
272 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
273 i10 += 8;
274
275 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
276
277 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
278 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
279
280 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
281 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
282 i11 += 8;
283
284 vprod01234567 = wasm_i16x8_mul(vi11x01234567, vk11x01234567);
285
286 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
287 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
288
289 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
290 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
291 i12 += 8;
292
293 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
294
295 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
296 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
297
298 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
299 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
300 i13 += 8;
301
302 vprod01234567 = wasm_i16x8_mul(vi13x01234567, vk13x01234567);
303
304 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
305 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
306
307 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
308 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
309 i14 += 8;
310
311 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
312
313 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
314 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
315
316 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
317 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
318 i15 += 8;
319
320 vprod01234567 = wasm_i16x8_mul(vi15x01234567, vk15x01234567);
321
322 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
323 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
324
325 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
326 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
327 i16 += 8;
328
329 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
330
331 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
332 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
333
334 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
335 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
336 i17 += 8;
337
338 vprod01234567 = wasm_i16x8_mul(vi17x01234567, vk17x01234567);
339
340 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
341 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
342
343 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
344 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
345 i18 += 8;
346
347 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
348
349 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
350 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
351
352 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
353 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
354 i19 += 8;
355
356 vprod01234567 = wasm_i16x8_mul(vi19x01234567, vk19x01234567);
357
358 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
359 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
360
361 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
362 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
363 i20 += 8;
364
365 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
366
367 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
368 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
369
370 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
371 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
372 i21 += 8;
373
374 vprod01234567 = wasm_i16x8_mul(vi21x01234567, vk21x01234567);
375
376 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
377 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
378
379 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
380 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
381 i22 += 8;
382
383 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
384
385 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
386 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
387
388 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
389 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
390 i23 += 8;
391
392 vprod01234567 = wasm_i16x8_mul(vi23x01234567, vk23x01234567);
393
394 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
395 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
396
397 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
398 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
399 i24 += 8;
400
401 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
402
403 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
404 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
405
406
407 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
408
409 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
410 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
411
412 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
413 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
414
415 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
416 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
417
418 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
419 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
420
421 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
422 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
423
424 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
425
426 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
427
428 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
429
430 wasm_v128_store64_lane(output, vout0123456701234567, 0);
431 output += 8;
432 }
433 if XNN_UNLIKELY(c != 0) {
434 {
435 v128_t vacc0123 = wasm_v128_load(w);
436 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
437
438
439 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
440 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
441
442 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
443
444 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
445 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
446
447 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
448 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
449
450 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
451
452 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
453 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
454
455 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
456 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
457
458 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
459
460 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
461 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
462
463 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
464 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
465
466 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
467
468 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
469 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
470
471 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
472 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
473
474 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
475
476 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
477 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
478
479 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
480 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
481
482 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
483
484 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
485 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
486
487 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
488 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
489
490 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
491
492 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
493 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
494
495 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
496 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
497
498 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
499
500 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
501 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
502
503 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
504 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
505
506 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
507
508 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
509 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
510
511 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
512 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
513
514 vprod01234567 = wasm_i16x8_mul(vi9x01234567, vk9x01234567);
515
516 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
517 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
518
519 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
520 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
521
522 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
523
524 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
525 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
526
527 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
528 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
529
530 vprod01234567 = wasm_i16x8_mul(vi11x01234567, vk11x01234567);
531
532 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
533 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
534
535 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
536 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
537
538 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
539
540 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
541 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
542
543 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
544 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
545
546 vprod01234567 = wasm_i16x8_mul(vi13x01234567, vk13x01234567);
547
548 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
549 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
550
551 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
552 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
553
554 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
555
556 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
557 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
558
559 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
560 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
561
562 vprod01234567 = wasm_i16x8_mul(vi15x01234567, vk15x01234567);
563
564 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
565 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
566
567 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
568 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
569
570 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
571
572 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
573 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
574
575 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
576 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
577
578 vprod01234567 = wasm_i16x8_mul(vi17x01234567, vk17x01234567);
579
580 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
581 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
582
583 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
584 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
585
586 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
587
588 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
589 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
590
591 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
592 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
593
594 vprod01234567 = wasm_i16x8_mul(vi19x01234567, vk19x01234567);
595
596 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
597 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
598
599 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
600 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
601
602 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
603
604 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
605 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
606
607 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
608 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
609
610 vprod01234567 = wasm_i16x8_mul(vi21x01234567, vk21x01234567);
611
612 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
613 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
614
615 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
616 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
617
618 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
619
620 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
621 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
622
623 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
624 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
625
626 vprod01234567 = wasm_i16x8_mul(vi23x01234567, vk23x01234567);
627
628 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
629 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
630
631 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
632 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
633
634 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
635
636 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
637 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
638
639
640
641 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
642 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
643
644 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
645 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
646
647 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
648 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
649
650 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
651 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
652
653 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
654 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
655
656 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
657 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
658
659 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
660
661
662 if (c & 4) {
663 wasm_v128_store32_lane(output, vout0123456701234567, 0);
664 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
665 output += 4;
666 }
667 if (c & 2) {
668 wasm_v128_store16_lane(output, vout0123456701234567, 0);
669 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
670 output += 2;
671 }
672 if (c & 1) {
673 wasm_v128_store8_lane(output, vout0123456701234567, 0);
674 output += 1;
675 }
676 }
677 }
678
679 output = (int8_t*) ((uintptr_t) output + output_increment);
680 } while (--output_width != 0);
681 }