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_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_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)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul16(
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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 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 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
178 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
179
180 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
181 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
182 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
183 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
184 i1 += 8;
185
186
187 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
188
189 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
190 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
191
192 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
193 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
194 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
195 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
196 i2 += 8;
197
198
199 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
200
201 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
202 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
203
204 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
205 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
206 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
207 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
208 i3 += 8;
209
210
211 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
212
213 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
214 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
215
216 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
217 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
218 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
219 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
220 i4 += 8;
221
222
223 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
224
225 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
226 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
227
228 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
229 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
230 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
231 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
232 i5 += 8;
233
234
235 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
236
237 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
238 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
239
240 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
241 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
242 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
243 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
244 i6 += 8;
245
246
247 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
248
249 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
250 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
251
252 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
253 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
254 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
255 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
256 i7 += 8;
257
258
259 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
260
261 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
262 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
263
264 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
265 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
266 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
267 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
268 i8 += 8;
269
270
271 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
272
273 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
274 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
275
276 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
277 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
278 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
279 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
280 i9 += 8;
281
282
283 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
284
285 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
286 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
287
288 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
289 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
290 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
291 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
292 i10 += 8;
293
294
295 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
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 vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
301 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
302 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
303 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
304 i11 += 8;
305
306
307 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
308
309 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
310 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
311
312 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
313 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
314 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
315 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
316 i12 += 8;
317
318
319 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
320
321 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
322 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
323
324 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
325 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
326 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
327 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
328 i13 += 8;
329
330
331 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
332
333 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
334 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
335
336 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
337 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
338 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
339 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
340 i14 += 8;
341
342
343 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
344
345 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
346 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
347
348 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
349 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
350 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
351 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
352 i15 += 8;
353
354
355 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
356
357 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
358 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
359
360 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
361 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
362 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
363 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
364 i16 += 8;
365
366
367 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
368
369 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
370 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
371
372 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
373 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
374 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
375 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
376 i17 += 8;
377
378
379 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
380
381 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
382 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
383
384 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
385 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
386 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
387 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
388 i18 += 8;
389
390
391 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
392
393 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
394 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
395
396 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
397 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
398 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
399 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
400 i19 += 8;
401
402
403 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
404
405 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
406 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
407
408 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
409 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
410 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
411 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
412 i20 += 8;
413
414
415 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
416
417 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
418 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
419
420 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
421 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
422 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
423 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
424 i21 += 8;
425
426
427 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
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 vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
433 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
434 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
435 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
436 i22 += 8;
437
438
439 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
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 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
445 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
446 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
447 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
448 i23 += 8;
449
450
451 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
452
453 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
454 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
455
456 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
457 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
458 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
459 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
460 i24 += 8;
461
462
463 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
464
465 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
466 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
467
468 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
469
470 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
471 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
472
473 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
474 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
475 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
476
477 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
478 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
479 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
480
481 vacc0123 = _mm_cvtps_epi32(vscaled0123);
482 vacc4567 = _mm_cvtps_epi32(vscaled4567);
483
484 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
485 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
486
487
488 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
489
490 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
491 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
492
493 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
494 output += 8;
495 }
496 if XNN_UNLIKELY(c != 0) {
497 {
498 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
499 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
500
501
502 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
503 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
504 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
505 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
506
507
508 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
509
510 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
511 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
512
513 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
514 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
515 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
516 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
517
518
519 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
520
521 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
522 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
523
524 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
525 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
526 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
527 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
528
529
530 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
531
532 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
533 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
534
535 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
536 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
537 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
538 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
539
540
541 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
542
543 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
544 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
545
546 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
547 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
548 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
549 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
550
551
552 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
553
554 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
555 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
556
557 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
558 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
559 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
560 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
561
562
563 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
564
565 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
566 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
567
568 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
569 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
570 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
571 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
572
573
574 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
575
576 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
577 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
578
579 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
580 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
581 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
582 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
583
584
585 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
586
587 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
588 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
589
590 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
591 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
592 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
593 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
594
595
596 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
597
598 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
599 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
600
601 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
602 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
603 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
604 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
605
606
607 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
608
609 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
610 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
611
612 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
613 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
614 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
615 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
616
617
618 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
619
620 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
621 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
622
623 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
624 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
625 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
626 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
627
628
629 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
630
631 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
632 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
633
634 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
635 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
636 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
637 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
638
639
640 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
641
642 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
643 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
644
645 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
646 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
647 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
648 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
649
650
651 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
652
653 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
654 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
655
656 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
657 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
658 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
659 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
660
661
662 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
663
664 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
665 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
666
667 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
668 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
669 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
670 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
671
672
673 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
674
675 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
676 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
677
678 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
679 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
680 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
681 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
682
683
684 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
685
686 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
687 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
688
689 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
690 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
691 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
692 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
693
694
695 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
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 vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
701 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
702 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
703 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
704
705
706 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
707
708 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
709 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
710
711 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
712 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
713 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
714 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
715
716
717 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
718
719 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
720 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
721
722 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
723 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
724 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
725 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
726
727
728 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
729
730 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
731 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
732
733 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
734 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
735 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
736 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
737
738
739 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
740
741 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
742 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
743
744 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
745 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
746 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
747 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
748
749
750 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
751
752 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
753 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
754
755 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
756 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
757 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
758 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
759
760
761 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
762
763 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
764 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
765
766 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
767 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
768 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
769 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
770
771
772 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
773
774 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
775 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
776
777
778 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
779 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
780
781 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
782 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
783 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
784
785 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
786 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
787 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
788
789 vacc0123 = _mm_cvtps_epi32(vscaled0123);
790 vacc4567 = _mm_cvtps_epi32(vscaled4567);
791
792
793 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
794 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
795
796
797 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
798
799 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
800
801 if (c & 4) {
802 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
803 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
804 output += 4;
805 }
806 if (c & 2) {
807 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
808 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
809 output += 2;
810 }
811 if (c & 1) {
812 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
813 output += 1;
814 }
815 }
816 }
817
818 output = (int8_t*) ((uintptr_t) output + output_increment);
819 } while (--output_width != 0);
820 }
821