1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-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 <smmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx_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__avx_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 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
164 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
165
166
167 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
168 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
169 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
170 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
171 i0 += 8;
172
173
174 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
175
176 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
177 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
178
179 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
180 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
181 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
182 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
183 i1 += 8;
184
185
186 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
187
188 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
189 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
190
191 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
192 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
193 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
194 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
195 i2 += 8;
196
197
198 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
199
200 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
201 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
202
203 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
204 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
205 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
206 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
207 i3 += 8;
208
209
210 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
211
212 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
213 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
214
215 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
216 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
217 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
218 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
219 i4 += 8;
220
221
222 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
223
224 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
225 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
226
227 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
228 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
229 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
230 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
231 i5 += 8;
232
233
234 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
235
236 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
237 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
238
239 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
240 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
241 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
242 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
243 i6 += 8;
244
245
246 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
247
248 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
249 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
250
251 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
252 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
253 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
254 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
255 i7 += 8;
256
257
258 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
259
260 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
261 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
262
263 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
264 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
265 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
266 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
267 i8 += 8;
268
269
270 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
271
272 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
273 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
274
275 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
276 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
277 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
278 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
279 i9 += 8;
280
281
282 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
283
284 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
285 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
286
287 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
288 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
289 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
290 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
291 i10 += 8;
292
293
294 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
295
296 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
297 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
298
299 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
300 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
301 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
302 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
303 i11 += 8;
304
305
306 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
307
308 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
309 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
310
311 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
312 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
313 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
314 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
315 i12 += 8;
316
317
318 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
319
320 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
321 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
322
323 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
324 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
325 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
326 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
327 i13 += 8;
328
329
330 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
331
332 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
333 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
334
335 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
336 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
337 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
338 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
339 i14 += 8;
340
341
342 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
343
344 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
345 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
346
347 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
348 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
349 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
350 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
351 i15 += 8;
352
353
354 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
355
356 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
357 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
358
359 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
360 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
361 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
362 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
363 i16 += 8;
364
365
366 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
367
368 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
369 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
370
371 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
372 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
373 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
374 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
375 i17 += 8;
376
377
378 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
379
380 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
381 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
382
383 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
384 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
385 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
386 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
387 i18 += 8;
388
389
390 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
391
392 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
393 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
394
395 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
396 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
397 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
398 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
399 i19 += 8;
400
401
402 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
403
404 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
405 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
406
407 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
408 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
409 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
410 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
411 i20 += 8;
412
413
414 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
415
416 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
417 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
418
419 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
420 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
421 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
422 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
423 i21 += 8;
424
425
426 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
427
428 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
429 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
430
431 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
432 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
433 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
434 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
435 i22 += 8;
436
437
438 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
439
440 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
441 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
442
443 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
444 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
445 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
446 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
447 i23 += 8;
448
449
450 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
451
452 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
453 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
454
455 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
456 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
457 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
458 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
459 i24 += 8;
460
461
462 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
463
464 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
465 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
466
467 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
468
469 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
470 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
471
472 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
473 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
474 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
475
476 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
477 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
478 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
479
480 vacc0123 = _mm_cvtps_epi32(vscaled0123);
481 vacc4567 = _mm_cvtps_epi32(vscaled4567);
482
483 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
484 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
485
486
487 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
488
489 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
490 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
491
492 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
493 output += 8;
494 }
495 if XNN_UNLIKELY(c != 0) {
496 {
497 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
498 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
499
500
501 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
502 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
503 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
504 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
505
506
507 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
508
509 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
510 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
511
512 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
513 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
514 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
515 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
516
517
518 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
519
520 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
521 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
522
523 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
524 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
525 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
526 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
527
528
529 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
530
531 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
532 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
533
534 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
535 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
536 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
537 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
538
539
540 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
541
542 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
543 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
544
545 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
546 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
547 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
548 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
549
550
551 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
552
553 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
554 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
555
556 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
557 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
558 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
559 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
560
561
562 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
563
564 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
565 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
566
567 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
568 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
569 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
570 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
571
572
573 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
574
575 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
576 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
577
578 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
579 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
580 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
581 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
582
583
584 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
585
586 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
587 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
588
589 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
590 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
591 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
592 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
593
594
595 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
596
597 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
598 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
599
600 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
601 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
602 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
603 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
604
605
606 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
607
608 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
609 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
610
611 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
612 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
613 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
614 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
615
616
617 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
618
619 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
620 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
621
622 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
623 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
624 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
625 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
626
627
628 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
629
630 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
631 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
632
633 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
634 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
635 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
636 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
637
638
639 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
640
641 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
642 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
643
644 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
645 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
646 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
647 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
648
649
650 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
651
652 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
653 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
654
655 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
656 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
657 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
658 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
659
660
661 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
662
663 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
664 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
665
666 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
667 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
668 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
669 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
670
671
672 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
673
674 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
675 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
676
677 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
678 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
679 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
680 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
681
682
683 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
684
685 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
686 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
687
688 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
689 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
690 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
691 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
692
693
694 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
695
696 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
697 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
698
699 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
700 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
701 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
702 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
703
704
705 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
706
707 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
708 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
709
710 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
711 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
712 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
713 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
714
715
716 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
717
718 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
719 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
720
721 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
722 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
723 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
724 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
725
726
727 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
728
729 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
730 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
731
732 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
733 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
734 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
735 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
736
737
738 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
739
740 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
741 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
742
743 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
744 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
745 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
746 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
747
748
749 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
750
751 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
752 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
753
754 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
755 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
756 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
757 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
758
759
760 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
761
762 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
763 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
764
765 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
766 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
767 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
768 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
769
770
771 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
772
773 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
774 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
775
776
777 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
778 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
779
780 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
781 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
782 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
783
784 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
785 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
786 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
787
788 vacc0123 = _mm_cvtps_epi32(vscaled0123);
789 vacc4567 = _mm_cvtps_epi32(vscaled4567);
790
791
792 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
793 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
794
795
796 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
797
798 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
799
800 if (c & 4) {
801 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
802 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
803 output += 4;
804 }
805 if (c & 2) {
806 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
807 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
808 output += 2;
809 }
810 if (c & 1) {
811 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
812 output += 1;
813 }
814 }
815 }
816
817 output = (int8_t*) ((uintptr_t) output + output_increment);
818 } while (--output_width != 0);
819 }
820