Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-25p8c-minmax-fp32-sse2-mul16-add16.c

raw
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 <emmintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/unaligned.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_25p8c__sse2_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 intptr_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
34 XNN_FORCE_REALIZATION(vscale);
35
36 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
37 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
38 const __m128i voutput_min = _mm_set1_epi16(params->fp32_scalar.output_min);
39 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
40 XNN_FORCE_REALIZATION(voutput_zero_point);
41 XNN_FORCE_REALIZATION(voutput_min);
42
43 do {
44 const int8_t* i0 = input[0];
45 assert(i0 != NULL);
46 if XNN_UNPREDICTABLE(i0 != zero) {
47 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
48 }
49 const int8_t* i1 = input[1];
50 assert(i1 != NULL);
51 if XNN_UNPREDICTABLE(i1 != zero) {
52 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
53 }
54 const int8_t* i2 = input[2];
55 assert(i2 != NULL);
56 if XNN_UNPREDICTABLE(i2 != zero) {
57 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
58 }
59 const int8_t* i3 = input[3];
60 assert(i3 != NULL);
61 if XNN_UNPREDICTABLE(i3 != zero) {
62 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
63 }
64 const int8_t* i4 = input[4];
65 assert(i4 != NULL);
66 if XNN_UNPREDICTABLE(i4 != zero) {
67 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
68 }
69 const int8_t* i5 = input[5];
70 assert(i5 != NULL);
71 if XNN_UNPREDICTABLE(i5 != zero) {
72 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
73 }
74 const int8_t* i6 = input[6];
75 assert(i6 != NULL);
76 if XNN_UNPREDICTABLE(i6 != zero) {
77 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
78 }
79 const int8_t* i7 = input[7];
80 assert(i7 != NULL);
81 if XNN_UNPREDICTABLE(i7 != zero) {
82 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
83 }
84 const int8_t* i8 = input[8];
85 assert(i8 != NULL);
86 if XNN_UNPREDICTABLE(i8 != zero) {
87 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
88 }
89 const int8_t* i9 = input[9];
90 assert(i9 != NULL);
91 if XNN_UNPREDICTABLE(i9 != zero) {
92 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
93 }
94 const int8_t* i10 = input[10];
95 assert(i10 != NULL);
96 if XNN_UNPREDICTABLE(i10 != zero) {
97 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
98 }
99 const int8_t* i11 = input[11];
100 assert(i11 != NULL);
101 if XNN_UNPREDICTABLE(i11 != zero) {
102 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
103 }
104 const int8_t* i12 = input[12];
105 assert(i12 != NULL);
106 if XNN_UNPREDICTABLE(i12 != zero) {
107 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
108 }
109 const int8_t* i13 = input[13];
110 assert(i13 != NULL);
111 if XNN_UNPREDICTABLE(i13 != zero) {
112 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
113 }
114 const int8_t* i14 = input[14];
115 assert(i14 != NULL);
116 if XNN_UNPREDICTABLE(i14 != zero) {
117 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
118 }
119 const int8_t* i15 = input[15];
120 assert(i15 != NULL);
121 if XNN_UNPREDICTABLE(i15 != zero) {
122 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
123 }
124 const int8_t* i16 = input[16];
125 assert(i16 != NULL);
126 if XNN_UNPREDICTABLE(i16 != zero) {
127 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
128 }
129 const int8_t* i17 = input[17];
130 assert(i17 != NULL);
131 if XNN_UNPREDICTABLE(i17 != zero) {
132 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
133 }
134 const int8_t* i18 = input[18];
135 assert(i18 != NULL);
136 if XNN_UNPREDICTABLE(i18 != zero) {
137 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
138 }
139 const int8_t* i19 = input[19];
140 assert(i19 != NULL);
141 if XNN_UNPREDICTABLE(i19 != zero) {
142 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
143 }
144 const int8_t* i20 = input[20];
145 assert(i20 != NULL);
146 if XNN_UNPREDICTABLE(i20 != zero) {
147 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
148 }
149 const int8_t* i21 = input[21];
150 assert(i21 != NULL);
151 if XNN_UNPREDICTABLE(i21 != zero) {
152 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
153 }
154 const int8_t* i22 = input[22];
155 assert(i22 != NULL);
156 if XNN_UNPREDICTABLE(i22 != zero) {
157 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
158 }
159 const int8_t* i23 = input[23];
160 assert(i23 != NULL);
161 if XNN_UNPREDICTABLE(i23 != zero) {
162 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
163 }
164 const int8_t* i24 = input[24];
165 assert(i24 != NULL);
166 if XNN_UNPREDICTABLE(i24 != zero) {
167 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
168 }
169 input = (const int8_t**) ((uintptr_t) input + input_stride);
170
171 size_t c = channels;
172 const void* w = weights;
173 for (; c >= 8; c -= 8) {
174 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
175 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
176
177
178 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
179 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
180 i0 += 8;
181
182 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
183 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
184
185 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
186
187
188 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
189 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
190 i1 += 8;
191
192 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
193 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
194
195 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
196
197 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
198 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
199 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
200
201 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
202 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
203 i2 += 8;
204
205 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
206 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
207
208 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
209
210
211 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
212 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
213 i3 += 8;
214
215 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
216 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
217
218 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
219
220 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
221 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
222 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
223
224 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
225 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
226 i4 += 8;
227
228 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
229 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
230
231 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
232
233
234 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
235 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
236 i5 += 8;
237
238 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
239 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
240
241 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
242
243 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
244 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
245 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
246
247 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
248 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
249 i6 += 8;
250
251 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
252 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
253
254 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
255
256
257 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
258 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
259 i7 += 8;
260
261 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
262 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
263
264 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
265
266 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
267 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
268 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
269
270 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
271 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
272 i8 += 8;
273
274 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
275 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
276
277 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
278
279
280 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
281 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
282 i9 += 8;
283
284 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
285 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
286
287 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
288
289 const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
290 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
291 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
292
293 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
294 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
295 i10 += 8;
296
297 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
298 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
299
300 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
301
302
303 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
304 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
305 i11 += 8;
306
307 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
308 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
309
310 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
311
312 const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
313 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
314 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
315
316 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
317 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
318 i12 += 8;
319
320 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
321 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
322
323 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
324
325
326 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
327 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
328 i13 += 8;
329
330 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
331 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
332
333 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
334
335 const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
336 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
337 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
338
339 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
340 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
341 i14 += 8;
342
343 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
344 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
345
346 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
347
348
349 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
350 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
351 i15 += 8;
352
353 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
354 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
355
356 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
357
358 const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
359 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
360 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
361
362 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
363 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
364 i16 += 8;
365
366 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
367 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
368
369 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
370
371
372 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
373 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
374 i17 += 8;
375
376 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
377 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
378
379 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
380
381 const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
382 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
383 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
384
385 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
386 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
387 i18 += 8;
388
389 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
390 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
391
392 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
393
394
395 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
396 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
397 i19 += 8;
398
399 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
400 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
401
402 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
403
404 const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
405 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
406 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
407
408 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
409 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
410 i20 += 8;
411
412 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
413 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
414
415 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
416
417
418 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
419 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
420 i21 += 8;
421
422 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
423 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
424
425 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
426
427 const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
428 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
429 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
430
431 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
432 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
433 i22 += 8;
434
435 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
436 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
437
438 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
439
440
441 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
442 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
443 i23 += 8;
444
445 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
446 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
447
448 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
449
450 const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
451 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
452 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
453
454 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
455 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
456 i24 += 8;
457
458 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
459 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
460
461 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
462
463 const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
464 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
465 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
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 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
473 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
474
475 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
476 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
477
478 vacc0123 = _mm_cvtps_epi32(vscaled0123);
479 vacc4567 = _mm_cvtps_epi32(vscaled4567);
480
481 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
482
483 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
484
485 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
486
487
488 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
489 output += 8;
490 }
491 if XNN_UNLIKELY(c != 0) {
492 {
493 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
494 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
495
496
497 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
498 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
499
500 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
501 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
502
503 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
504
505
506 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
507 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
508
509 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
510 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
511
512 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
513
514 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
515 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
516 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
517
518 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
519 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
520
521 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
522 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
523
524 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
525
526
527 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
528 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
529
530 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
531 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
532
533 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
534
535 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
536 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
537 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
538
539 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
540 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
541
542 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
543 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
544
545 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
546
547
548 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
549 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
550
551 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
552 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
553
554 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
555
556 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
557 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
558 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
559
560 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
561 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
562
563 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
564 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
565
566 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
567
568
569 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
570 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
571
572 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
573 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
574
575 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
576
577 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
578 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
579 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
580
581 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
582 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
583
584 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
585 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
586
587 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
588
589
590 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
591 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
592
593 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
594 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
595
596 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
597
598 const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
599 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
600 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
601
602 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
603 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
604
605 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
606 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
607
608 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
609
610
611 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
612 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
613
614 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
615 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
616
617 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
618
619 const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
620 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
621 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
622
623 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
624 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
625
626 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
627 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
628
629 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
630
631
632 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
633 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
634
635 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
636 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
637
638 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
639
640 const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
641 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
642 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
643
644 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
645 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
646
647 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
648 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
649
650 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
651
652
653 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
654 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
655
656 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
657 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
658
659 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
660
661 const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
662 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
663 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
664
665 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
666 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
667
668 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
669 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
670
671 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
672
673
674 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
675 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
676
677 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
678 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
679
680 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
681
682 const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
683 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
684 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
685
686 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
687 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
688
689 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
690 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
691
692 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
693
694
695 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
696 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
697
698 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
699 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
700
701 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
702
703 const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
704 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
705 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
706
707 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
708 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
709
710 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
711 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
712
713 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
714
715
716 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
717 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
718
719 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
720 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
721
722 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
723
724 const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
725 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
726 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
727
728 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
729 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
730
731 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
732 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
733
734 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
735
736
737 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
738 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
739
740 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
741 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
742
743 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
744
745 const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
746 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
747 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
748
749 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
750 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
751
752 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
753 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
754
755 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
756
757 const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
758 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
759 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
760
761
762 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
763 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
764
765 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
766 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
767
768 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
769 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
770
771 vacc0123 = _mm_cvtps_epi32(vscaled0123);
772 vacc4567 = _mm_cvtps_epi32(vscaled4567);
773
774
775 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
776
777 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
778
779 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
780
781
782 if (c & 4) {
783 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
784 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
785 output += 4;
786 }
787 if (c & 2) {
788 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
789 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
790 output += 2;
791 }
792 if (c & 1) {
793 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
794 output += 1;
795 }
796 }
797 }
798
799 output = (int8_t*) ((uintptr_t) output + output_increment);
800 } while (--output_width != 0);
801 }