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 #include <xnnpack/unaligned.h>
16
17
xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul16_add16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 do {
34 const int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const int8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const int8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const int8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const int8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const int8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const int8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 const int8_t* i9 = input[9];
80 assert(i9 != NULL);
81 if XNN_UNPREDICTABLE(i9 != zero) {
82 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
83 }
84 const int8_t* i10 = input[10];
85 assert(i10 != NULL);
86 if XNN_UNPREDICTABLE(i10 != zero) {
87 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
88 }
89 const int8_t* i11 = input[11];
90 assert(i11 != NULL);
91 if XNN_UNPREDICTABLE(i11 != zero) {
92 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
93 }
94 const int8_t* i12 = input[12];
95 assert(i12 != NULL);
96 if XNN_UNPREDICTABLE(i12 != zero) {
97 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
98 }
99 const int8_t* i13 = input[13];
100 assert(i13 != NULL);
101 if XNN_UNPREDICTABLE(i13 != zero) {
102 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
103 }
104 const int8_t* i14 = input[14];
105 assert(i14 != NULL);
106 if XNN_UNPREDICTABLE(i14 != zero) {
107 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
108 }
109 const int8_t* i15 = input[15];
110 assert(i15 != NULL);
111 if XNN_UNPREDICTABLE(i15 != zero) {
112 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
113 }
114 const int8_t* i16 = input[16];
115 assert(i16 != NULL);
116 if XNN_UNPREDICTABLE(i16 != zero) {
117 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
118 }
119 const int8_t* i17 = input[17];
120 assert(i17 != NULL);
121 if XNN_UNPREDICTABLE(i17 != zero) {
122 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
123 }
124 const int8_t* i18 = input[18];
125 assert(i18 != NULL);
126 if XNN_UNPREDICTABLE(i18 != zero) {
127 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
128 }
129 const int8_t* i19 = input[19];
130 assert(i19 != NULL);
131 if XNN_UNPREDICTABLE(i19 != zero) {
132 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
133 }
134 const int8_t* i20 = input[20];
135 assert(i20 != NULL);
136 if XNN_UNPREDICTABLE(i20 != zero) {
137 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
138 }
139 const int8_t* i21 = input[21];
140 assert(i21 != NULL);
141 if XNN_UNPREDICTABLE(i21 != zero) {
142 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
143 }
144 const int8_t* i22 = input[22];
145 assert(i22 != NULL);
146 if XNN_UNPREDICTABLE(i22 != zero) {
147 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
148 }
149 const int8_t* i23 = input[23];
150 assert(i23 != NULL);
151 if XNN_UNPREDICTABLE(i23 != zero) {
152 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
153 }
154 const int8_t* i24 = input[24];
155 assert(i24 != NULL);
156 if XNN_UNPREDICTABLE(i24 != zero) {
157 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
158 }
159 input = (const int8_t**) ((uintptr_t) input + input_stride);
160
161 size_t c = channels;
162 const void* w = weights;
163 for (; c >= 8; c -= 8) {
164 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
165 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
166
167
168 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
169 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
170 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
171 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
172 i0 += 8;
173
174
175 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
176
177
178 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
179 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
180 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
181 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
182 i1 += 8;
183
184
185 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
186
187 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
188 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
189
190 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
191 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
192 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
193 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
194 i2 += 8;
195
196
197 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
198
199
200 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
201 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
202 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
203 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
204 i3 += 8;
205
206
207 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
208
209 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
210 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
211
212 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
213 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
214 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
215 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
216 i4 += 8;
217
218
219 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
220
221
222 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
223 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
224 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
225 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
226 i5 += 8;
227
228
229 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
230
231 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
232 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
233
234 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
235 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
236 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
237 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
238 i6 += 8;
239
240
241 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
242
243
244 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
245 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
246 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
247 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
248 i7 += 8;
249
250
251 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
252
253 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
254 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
255
256 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
257 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
258 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
259 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
260 i8 += 8;
261
262
263 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
264
265
266 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
267 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
268 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
269 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
270 i9 += 8;
271
272
273 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
274
275 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
276 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
277
278 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
279 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
280 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
281 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
282 i10 += 8;
283
284
285 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
286
287
288 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
289 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
290 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
291 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
292 i11 += 8;
293
294
295 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
296
297 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
298 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
299
300 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
301 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
302 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
303 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
304 i12 += 8;
305
306
307 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
308
309
310 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
311 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
312 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
313 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
314 i13 += 8;
315
316
317 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
318
319 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
320 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
321
322 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
323 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
324 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
325 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
326 i14 += 8;
327
328
329 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
330
331
332 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
333 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
334 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
335 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
336 i15 += 8;
337
338
339 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
340
341 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
342 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
343
344 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
345 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
346 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
347 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
348 i16 += 8;
349
350
351 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
352
353
354 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
355 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
356 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
357 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
358 i17 += 8;
359
360
361 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
362
363 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
364 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
365
366 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
367 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
368 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
369 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
370 i18 += 8;
371
372
373 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
374
375
376 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
377 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
378 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
379 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
380 i19 += 8;
381
382
383 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
384
385 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
386 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
387
388 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
389 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
390 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
391 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
392 i20 += 8;
393
394
395 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
396
397
398 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
399 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
400 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
401 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
402 i21 += 8;
403
404
405 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
406
407 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
408 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
409
410 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
411 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
412 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
413 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
414 i22 += 8;
415
416
417 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
418
419
420 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
421 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
422 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
423 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
424 i23 += 8;
425
426
427 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
428
429 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
430 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
431
432 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
433 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
434 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
435 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
436 i24 += 8;
437
438
439 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
440
441 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
442 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
443
444 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
445
446 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
447 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
448
449 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
450 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
451 w = (const void*) ((const float*) w + 8);
452 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
453 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
454
455 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
456 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
457 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
458
459 vacc0123 = _mm_cvtps_epi32(vscaled0123);
460 vacc4567 = _mm_cvtps_epi32(vscaled4567);
461
462 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
463 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
464
465
466 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
467
468 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
469 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
470
471 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
472 output += 8;
473 }
474 if XNN_UNLIKELY(c != 0) {
475 {
476 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
477 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
478
479
480 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
481 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
482 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
483 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
484
485
486 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
487
488
489 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
490 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
491 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
492 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
493
494
495 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
496
497 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
498 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
499
500 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
501 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
502 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
503 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
504
505
506 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
507
508
509 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
510 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
511 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
512 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
513
514
515 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
516
517 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
518 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
519
520 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
521 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
522 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
523 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
524
525
526 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
527
528
529 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
530 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
531 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
532 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
533
534
535 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
536
537 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
538 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
539
540 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
541 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
542 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
543 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
544
545
546 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
547
548
549 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
550 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
551 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
552 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
553
554
555 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
556
557 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
558 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
559
560 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
561 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
562 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
563 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
564
565
566 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
567
568
569 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
570 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
571 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
572 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
573
574
575 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
576
577 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
578 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
579
580 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
581 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
582 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
583 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
584
585
586 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
587
588
589 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
590 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
591 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
592 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
593
594
595 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
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 vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
601 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
602 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
603 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
604
605
606 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
607
608
609 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
610 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
611 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
612 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
613
614
615 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
616
617 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
618 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
619
620 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
621 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
622 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
623 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
624
625
626 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
627
628
629 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
630 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
631 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
632 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
633
634
635 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
636
637 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
638 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
639
640 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
641 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
642 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
643 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
644
645
646 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
647
648
649 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
650 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
651 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
652 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
653
654
655 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
656
657 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
658 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
659
660 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
661 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
662 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
663 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
664
665
666 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
667
668
669 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
670 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
671 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
672 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
673
674
675 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
676
677 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
678 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
679
680 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
681 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
682 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
683 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
684
685
686 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
687
688
689 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
690 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
691 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
692 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
693
694
695 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
696
697 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
698 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
699
700 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
701 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
702 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
703 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
704
705
706 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
707
708
709 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
710 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
711 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
712 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
713
714
715 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
716
717 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
718 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
719
720 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
721 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
722 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
723 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
724
725
726 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
727
728 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
729 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
730
731
732 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
733 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
734
735 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t)));
736 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t) + 4 * sizeof(float)));
737 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
738 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
739
740 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
741 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
742 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
743
744 vacc0123 = _mm_cvtps_epi32(vscaled0123);
745 vacc4567 = _mm_cvtps_epi32(vscaled4567);
746
747
748 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
749 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
750
751
752 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
753
754 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
755
756 if (c & 4) {
757 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
758 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
759 output += 4;
760 }
761 if (c & 2) {
762 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
763 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
764 output += 2;
765 }
766 if (c & 1) {
767 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
768 output += 1;
769 }
770 }
771 }
772
773 output = (int8_t*) ((uintptr_t) output + output_increment);
774 } while (--output_width != 0);
775 }
776