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_up16x25__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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__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_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 >= 16; c -= 16) {
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 v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
166 v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
167
168
169 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
170 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
171 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
172 const v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
173 i0 += 16;
174
175 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
176 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
177
178
179 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
180 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
181 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
182 const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
183 i1 += 16;
184
185 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
186 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF));
187
188 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
189 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
190 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
191 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
192
193 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
194 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
195 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
196 const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
197 i2 += 16;
198
199 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
200 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
201
202
203 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
204 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
205 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
206 const v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
207 i3 += 16;
208
209 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
210 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF));
211
212 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
213 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
214 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
215 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
216
217 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
218 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
219 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
220 const v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
221 i4 += 16;
222
223 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
224 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
225
226
227 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
228 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
229 const v128_t vi5x89ABCDEF = wasm_i16x8_load8x8(i5 + 8);
230 const v128_t vk5x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
231 i5 += 16;
232
233 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
234 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF));
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 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
239 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
240
241 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
242 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
243 const v128_t vi6x89ABCDEF = wasm_i16x8_load8x8(i6 + 8);
244 const v128_t vk6x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
245 i6 += 16;
246
247 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
248 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
249
250
251 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
252 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
253 const v128_t vi7x89ABCDEF = wasm_i16x8_load8x8(i7 + 8);
254 const v128_t vk7x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
255 i7 += 16;
256
257 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
258 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF));
259
260 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
261 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
262 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
263 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
264
265 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
266 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
267 const v128_t vi8x89ABCDEF = wasm_i16x8_load8x8(i8 + 8);
268 const v128_t vk8x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
269 i8 += 16;
270
271 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
272 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
273
274
275 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
276 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
277 const v128_t vi9x89ABCDEF = wasm_i16x8_load8x8(i9 + 8);
278 const v128_t vk9x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
279 i9 += 16;
280
281 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi9x01234567, vk9x01234567));
282 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi9x89ABCDEF, vk9x89ABCDEF));
283
284 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
285 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
286 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
287 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
288
289 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
290 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
291 const v128_t vi10x89ABCDEF = wasm_i16x8_load8x8(i10 + 8);
292 const v128_t vk10x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
293 i10 += 16;
294
295 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
296 vprod89ABCDEF = wasm_i16x8_mul(vi10x89ABCDEF, vk10x89ABCDEF);
297
298
299 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
300 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
301 const v128_t vi11x89ABCDEF = wasm_i16x8_load8x8(i11 + 8);
302 const v128_t vk11x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
303 i11 += 16;
304
305 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi11x01234567, vk11x01234567));
306 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi11x89ABCDEF, vk11x89ABCDEF));
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 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
311 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
312
313 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
314 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
315 const v128_t vi12x89ABCDEF = wasm_i16x8_load8x8(i12 + 8);
316 const v128_t vk12x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
317 i12 += 16;
318
319 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
320 vprod89ABCDEF = wasm_i16x8_mul(vi12x89ABCDEF, vk12x89ABCDEF);
321
322
323 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
324 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
325 const v128_t vi13x89ABCDEF = wasm_i16x8_load8x8(i13 + 8);
326 const v128_t vk13x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
327 i13 += 16;
328
329 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi13x01234567, vk13x01234567));
330 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi13x89ABCDEF, vk13x89ABCDEF));
331
332 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
333 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
334 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
335 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
336
337 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
338 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
339 const v128_t vi14x89ABCDEF = wasm_i16x8_load8x8(i14 + 8);
340 const v128_t vk14x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
341 i14 += 16;
342
343 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
344 vprod89ABCDEF = wasm_i16x8_mul(vi14x89ABCDEF, vk14x89ABCDEF);
345
346
347 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
348 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
349 const v128_t vi15x89ABCDEF = wasm_i16x8_load8x8(i15 + 8);
350 const v128_t vk15x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
351 i15 += 16;
352
353 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi15x01234567, vk15x01234567));
354 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi15x89ABCDEF, vk15x89ABCDEF));
355
356 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
357 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
358 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
359 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
360
361 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
362 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
363 const v128_t vi16x89ABCDEF = wasm_i16x8_load8x8(i16 + 8);
364 const v128_t vk16x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
365 i16 += 16;
366
367 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
368 vprod89ABCDEF = wasm_i16x8_mul(vi16x89ABCDEF, vk16x89ABCDEF);
369
370
371 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
372 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
373 const v128_t vi17x89ABCDEF = wasm_i16x8_load8x8(i17 + 8);
374 const v128_t vk17x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
375 i17 += 16;
376
377 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi17x01234567, vk17x01234567));
378 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi17x89ABCDEF, vk17x89ABCDEF));
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 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
383 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
384
385 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
386 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
387 const v128_t vi18x89ABCDEF = wasm_i16x8_load8x8(i18 + 8);
388 const v128_t vk18x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
389 i18 += 16;
390
391 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
392 vprod89ABCDEF = wasm_i16x8_mul(vi18x89ABCDEF, vk18x89ABCDEF);
393
394
395 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
396 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
397 const v128_t vi19x89ABCDEF = wasm_i16x8_load8x8(i19 + 8);
398 const v128_t vk19x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
399 i19 += 16;
400
401 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi19x01234567, vk19x01234567));
402 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi19x89ABCDEF, vk19x89ABCDEF));
403
404 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
405 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
406 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
407 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
408
409 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
410 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
411 const v128_t vi20x89ABCDEF = wasm_i16x8_load8x8(i20 + 8);
412 const v128_t vk20x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
413 i20 += 16;
414
415 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
416 vprod89ABCDEF = wasm_i16x8_mul(vi20x89ABCDEF, vk20x89ABCDEF);
417
418
419 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
420 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
421 const v128_t vi21x89ABCDEF = wasm_i16x8_load8x8(i21 + 8);
422 const v128_t vk21x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
423 i21 += 16;
424
425 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi21x01234567, vk21x01234567));
426 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi21x89ABCDEF, vk21x89ABCDEF));
427
428 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
429 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
430 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
431 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
432
433 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
434 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
435 const v128_t vi22x89ABCDEF = wasm_i16x8_load8x8(i22 + 8);
436 const v128_t vk22x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
437 i22 += 16;
438
439 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
440 vprod89ABCDEF = wasm_i16x8_mul(vi22x89ABCDEF, vk22x89ABCDEF);
441
442
443 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
444 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
445 const v128_t vi23x89ABCDEF = wasm_i16x8_load8x8(i23 + 8);
446 const v128_t vk23x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
447 i23 += 16;
448
449 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi23x01234567, vk23x01234567));
450 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi23x89ABCDEF, vk23x89ABCDEF));
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 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
455 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
456
457 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
458 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
459 const v128_t vi24x89ABCDEF = wasm_i16x8_load8x8(i24 + 8);
460 const v128_t vk24x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
461 i24 += 16;
462
463 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
464 vprod89ABCDEF = wasm_i16x8_mul(vi24x89ABCDEF, vk24x89ABCDEF);
465
466 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
467 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
468 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
469 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
470
471
472 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
473
474 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
475 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
476 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
477 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
478
479 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
480 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
481 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
482 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
483 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
484
485 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
486 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
487 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
488 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
489 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
490
491 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
492 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
493 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
494 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
495 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
496
497 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
498 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
499 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
500 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
501 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
502
503 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
504 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
505
506 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
507
508 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
509 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
510
511 wasm_v128_store(output, vout0123456789ABCDEF);
512 output += 16;
513 }
514 if XNN_UNLIKELY(c != 0) {
515 const int8_t* k = (const int8_t*) ((uintptr_t) w + 16 * sizeof(int32_t));
516 do {
517 v128_t vacc0123 = wasm_v128_load(w);
518 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
519
520
521 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
522 const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
523 i0 += 8;
524
525 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
526
527
528 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
529 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 16));
530 i1 += 8;
531
532 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
533
534 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
535 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
536
537 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
538 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 32));
539 i2 += 8;
540
541 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
542
543
544 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
545 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) (k + 48));
546 i3 += 8;
547
548 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
549
550 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
551 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
552
553 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
554 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) (k + 64));
555 i4 += 8;
556
557 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
558
559
560 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
561 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) (k + 80));
562 i5 += 8;
563
564 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
565
566 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
567 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
568
569 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
570 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) (k + 96));
571 i6 += 8;
572
573 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
574
575
576 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
577 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) (k + 112));
578 i7 += 8;
579
580 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
581
582 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
583 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
584
585 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
586 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) (k + 128));
587 i8 += 8;
588
589 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
590
591
592 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
593 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) (k + 144));
594 i9 += 8;
595
596 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi9x01234567, vk9x01234567));
597
598 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
599 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
600
601 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
602 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) (k + 160));
603 i10 += 8;
604
605 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
606
607
608 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
609 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) (k + 176));
610 i11 += 8;
611
612 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi11x01234567, vk11x01234567));
613
614 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
615 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
616
617 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
618 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) (k + 192));
619 i12 += 8;
620
621 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
622
623
624 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
625 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) (k + 208));
626 i13 += 8;
627
628 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi13x01234567, vk13x01234567));
629
630 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
631 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
632
633 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
634 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) (k + 224));
635 i14 += 8;
636
637 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
638
639
640 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
641 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) (k + 240));
642 i15 += 8;
643
644 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi15x01234567, vk15x01234567));
645
646 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
647 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
648
649 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
650 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) (k + 256));
651 i16 += 8;
652
653 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
654
655
656 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
657 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) (k + 272));
658 i17 += 8;
659
660 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi17x01234567, vk17x01234567));
661
662 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
663 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
664
665 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
666 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) (k + 288));
667 i18 += 8;
668
669 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
670
671
672 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
673 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) (k + 304));
674 i19 += 8;
675
676 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi19x01234567, vk19x01234567));
677
678 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
679 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
680
681 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
682 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) (k + 320));
683 i20 += 8;
684
685 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
686
687
688 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
689 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) (k + 336));
690 i21 += 8;
691
692 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi21x01234567, vk21x01234567));
693
694 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
695 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
696
697 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
698 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) (k + 352));
699 i22 += 8;
700
701 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
702
703
704 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
705 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) (k + 368));
706 i23 += 8;
707
708 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi23x01234567, vk23x01234567));
709
710 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
711 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
712
713 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
714 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) (k + 384));
715 i24 += 8;
716
717 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
718
719 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
720 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
721
722 k += 8;
723
724
725 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
726 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
727
728 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
729 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
730 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
731
732 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
733 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
734 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
735
736 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
737 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
738 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
739
740 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
741 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
742 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
743
744 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
745 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
746
747 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
748 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
749
750 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
751
752 if XNN_LIKELY(c >= 8) {
753 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
754 output += 8;
755 c -= 8;
756 } else {
757 if (c & 4) {
758 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
759 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
760 output += 4;
761 }
762 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
763 if (c & 2) {
764 *((uint16_t*) output) = (uint16_t) vout0123;
765 vout0123 >>= 16;
766 output += 2;
767 }
768 if (c & 1) {
769 *output = (int8_t) vout0123;
770 output += 1;
771 }
772 c = 0;
773 }
774 } while (c != 0);
775 }
776
777 output = (int8_t*) ((uintptr_t) output + output_increment);
778 } while (--output_width != 0);
779 }
780