Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-25p16c-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_25p16c__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 >= 16; c -= 16) {
174 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
175 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
176 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
177 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
178
179
180 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
181 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
182 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
183 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
184 i0 += 16;
185
186 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
187 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
188 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
189 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
190
191 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
192 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
193
194
195 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
196 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
197 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
198 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
199 i1 += 16;
200
201 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
202 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
203 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
204 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
205
206 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
207 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
208
209 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
210 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
211 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
212 const __m128i vsignprod1x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
213 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod1x89ABCDEF));
214 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod1x89ABCDEF));
215
216 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
217 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
218 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
219 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
220 i2 += 16;
221
222 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
223 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
224 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
225 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
226
227 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
228 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
229
230
231 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
232 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
233 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
234 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
235 i3 += 16;
236
237 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
238 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
239 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
240 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
241
242 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
243 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
244
245 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
246 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
247 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
248 const __m128i vsignprod3x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
249 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod3x89ABCDEF));
250 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod3x89ABCDEF));
251
252 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
253 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
254 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
255 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
256 i4 += 16;
257
258 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
259 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
260 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
261 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
262
263 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
264 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
265
266
267 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
268 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
269 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
270 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
271 i5 += 16;
272
273 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
274 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
275 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
276 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
277
278 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
279 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
280
281 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
282 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
283 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
284 const __m128i vsignprod5x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
285 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod5x89ABCDEF));
286 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod5x89ABCDEF));
287
288 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
289 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
290 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
291 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
292 i6 += 16;
293
294 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
295 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
296 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
297 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
298
299 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
300 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
301
302
303 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
304 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
305 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
306 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
307 i7 += 16;
308
309 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
310 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
311 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
312 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
313
314 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
315 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
316
317 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
318 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
319 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
320 const __m128i vsignprod7x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
321 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod7x89ABCDEF));
322 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod7x89ABCDEF));
323
324 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
325 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
326 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
327 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
328 i8 += 16;
329
330 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
331 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
332 const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
333 const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
334
335 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
336 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
337
338
339 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
340 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
341 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
342 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
343 i9 += 16;
344
345 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
346 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
347 const __m128i vxi9x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x89ABCDEF, vi9x89ABCDEF), 8);
348 const __m128i vxk9x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x89ABCDEF, vk9x89ABCDEF), 8);
349
350 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
351 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF));
352
353 const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
354 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
355 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
356 const __m128i vsignprod9x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
357 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod9x89ABCDEF));
358 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod9x89ABCDEF));
359
360 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
361 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
362 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
363 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
364 i10 += 16;
365
366 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
367 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
368 const __m128i vxi10x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x89ABCDEF, vi10x89ABCDEF), 8);
369 const __m128i vxk10x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x89ABCDEF, vk10x89ABCDEF), 8);
370
371 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
372 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
373
374
375 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
376 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
377 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
378 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
379 i11 += 16;
380
381 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
382 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
383 const __m128i vxi11x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x89ABCDEF, vi11x89ABCDEF), 8);
384 const __m128i vxk11x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x89ABCDEF, vk11x89ABCDEF), 8);
385
386 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
387 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF));
388
389 const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
390 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
391 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
392 const __m128i vsignprod11x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
393 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod11x89ABCDEF));
394 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod11x89ABCDEF));
395
396 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
397 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
398 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
399 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
400 i12 += 16;
401
402 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
403 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
404 const __m128i vxi12x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x89ABCDEF, vi12x89ABCDEF), 8);
405 const __m128i vxk12x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x89ABCDEF, vk12x89ABCDEF), 8);
406
407 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
408 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
409
410
411 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
412 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
413 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
414 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
415 i13 += 16;
416
417 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
418 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
419 const __m128i vxi13x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x89ABCDEF, vi13x89ABCDEF), 8);
420 const __m128i vxk13x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x89ABCDEF, vk13x89ABCDEF), 8);
421
422 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
423 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF));
424
425 const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
426 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
427 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
428 const __m128i vsignprod13x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
429 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod13x89ABCDEF));
430 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod13x89ABCDEF));
431
432 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
433 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
434 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
435 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
436 i14 += 16;
437
438 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
439 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
440 const __m128i vxi14x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x89ABCDEF, vi14x89ABCDEF), 8);
441 const __m128i vxk14x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x89ABCDEF, vk14x89ABCDEF), 8);
442
443 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
444 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
445
446
447 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
448 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
449 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
450 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
451 i15 += 16;
452
453 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
454 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
455 const __m128i vxi15x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x89ABCDEF, vi15x89ABCDEF), 8);
456 const __m128i vxk15x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x89ABCDEF, vk15x89ABCDEF), 8);
457
458 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
459 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF));
460
461 const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
462 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
463 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
464 const __m128i vsignprod15x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
465 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod15x89ABCDEF));
466 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod15x89ABCDEF));
467
468 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
469 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
470 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
471 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
472 i16 += 16;
473
474 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
475 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
476 const __m128i vxi16x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x89ABCDEF, vi16x89ABCDEF), 8);
477 const __m128i vxk16x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x89ABCDEF, vk16x89ABCDEF), 8);
478
479 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
480 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
481
482
483 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
484 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
485 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
486 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
487 i17 += 16;
488
489 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
490 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
491 const __m128i vxi17x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x89ABCDEF, vi17x89ABCDEF), 8);
492 const __m128i vxk17x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x89ABCDEF, vk17x89ABCDEF), 8);
493
494 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
495 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF));
496
497 const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
498 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
499 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
500 const __m128i vsignprod17x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
501 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod17x89ABCDEF));
502 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod17x89ABCDEF));
503
504 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
505 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
506 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
507 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
508 i18 += 16;
509
510 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
511 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
512 const __m128i vxi18x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x89ABCDEF, vi18x89ABCDEF), 8);
513 const __m128i vxk18x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x89ABCDEF, vk18x89ABCDEF), 8);
514
515 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
516 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
517
518
519 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
520 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
521 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
522 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
523 i19 += 16;
524
525 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
526 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
527 const __m128i vxi19x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x89ABCDEF, vi19x89ABCDEF), 8);
528 const __m128i vxk19x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x89ABCDEF, vk19x89ABCDEF), 8);
529
530 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
531 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF));
532
533 const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
534 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
535 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
536 const __m128i vsignprod19x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
537 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod19x89ABCDEF));
538 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod19x89ABCDEF));
539
540 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
541 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
542 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
543 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
544 i20 += 16;
545
546 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
547 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
548 const __m128i vxi20x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x89ABCDEF, vi20x89ABCDEF), 8);
549 const __m128i vxk20x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x89ABCDEF, vk20x89ABCDEF), 8);
550
551 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
552 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
553
554
555 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
556 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
557 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
558 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
559 i21 += 16;
560
561 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
562 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
563 const __m128i vxi21x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x89ABCDEF, vi21x89ABCDEF), 8);
564 const __m128i vxk21x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x89ABCDEF, vk21x89ABCDEF), 8);
565
566 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
567 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF));
568
569 const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
570 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
571 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
572 const __m128i vsignprod21x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
573 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod21x89ABCDEF));
574 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod21x89ABCDEF));
575
576 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
577 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
578 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
579 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
580 i22 += 16;
581
582 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
583 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
584 const __m128i vxi22x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x89ABCDEF, vi22x89ABCDEF), 8);
585 const __m128i vxk22x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x89ABCDEF, vk22x89ABCDEF), 8);
586
587 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
588 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
589
590
591 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
592 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
593 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
594 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
595 i23 += 16;
596
597 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
598 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
599 const __m128i vxi23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x89ABCDEF, vi23x89ABCDEF), 8);
600 const __m128i vxk23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x89ABCDEF, vk23x89ABCDEF), 8);
601
602 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
603 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF));
604
605 const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
606 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
607 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
608 const __m128i vsignprod23x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
609 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod23x89ABCDEF));
610 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod23x89ABCDEF));
611
612 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
613 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
614 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
615 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
616 i24 += 16;
617
618 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
619 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
620 const __m128i vxi24x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x89ABCDEF, vi24x89ABCDEF), 8);
621 const __m128i vxk24x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x89ABCDEF, vk24x89ABCDEF), 8);
622
623 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
624 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
625
626 const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
627 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
628 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
629 const __m128i vsignprod24x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
630 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod24x89ABCDEF));
631 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod24x89ABCDEF));
632
633 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
634
635 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
636 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
637 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
638 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
639
640 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
641 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
642 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
643 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
644
645 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
646 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
647 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
648 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
649
650 vacc0123 = _mm_cvtps_epi32(vscaled0123);
651 vacc4567 = _mm_cvtps_epi32(vscaled4567);
652 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
653 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
654
655 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
656 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
657
658 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
659 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
660
661 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
662
663
664 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
665 output += 16;
666 }
667 if XNN_UNLIKELY(c != 0) {
668 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
669 do {
670 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
671 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
672
673
674 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
675 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
676 i0 += 8;
677
678 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
679 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
680
681 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
682
683
684 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
685 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
686 i1 += 8;
687
688 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
689 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
690
691 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
692
693 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
694 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
695 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
696
697 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
698 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
699 i2 += 8;
700
701 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
702 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
703
704 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
705
706
707 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
708 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
709 i3 += 8;
710
711 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
712 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
713
714 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
715
716 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
717 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
718 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
719
720 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
721 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
722 i4 += 8;
723
724 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
725 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
726
727 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
728
729
730 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
731 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
732 i5 += 8;
733
734 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
735 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
736
737 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
738
739 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
740 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
741 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
742
743 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
744 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
745 i6 += 8;
746
747 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
748 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
749
750 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
751
752
753 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
754 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
755 i7 += 8;
756
757 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
758 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
759
760 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
761
762 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
763 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
764 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
765
766 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
767 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
768 i8 += 8;
769
770 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
771 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
772
773 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
774
775
776 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
777 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
778 i9 += 8;
779
780 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
781 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
782
783 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
784
785 const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
786 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
787 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
788
789 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
790 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
791 i10 += 8;
792
793 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
794 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
795
796 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
797
798
799 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
800 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
801 i11 += 8;
802
803 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
804 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
805
806 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
807
808 const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
809 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
810 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
811
812 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
813 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
814 i12 += 8;
815
816 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
817 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
818
819 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
820
821
822 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
823 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
824 i13 += 8;
825
826 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
827 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
828
829 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
830
831 const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
832 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
833 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
834
835 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
836 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
837 i14 += 8;
838
839 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
840 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
841
842 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
843
844
845 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
846 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
847 i15 += 8;
848
849 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
850 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
851
852 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
853
854 const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
855 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
856 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
857
858 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
859 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
860 i16 += 8;
861
862 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
863 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
864
865 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
866
867
868 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
869 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
870 i17 += 8;
871
872 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
873 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
874
875 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
876
877 const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
878 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
879 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
880
881 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
882 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
883 i18 += 8;
884
885 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
886 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
887
888 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
889
890
891 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
892 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
893 i19 += 8;
894
895 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
896 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
897
898 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
899
900 const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
901 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
902 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
903
904 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
905 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
906 i20 += 8;
907
908 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
909 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
910
911 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
912
913
914 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
915 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
916 i21 += 8;
917
918 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
919 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
920
921 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
922
923 const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
924 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
925 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
926
927 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
928 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
929 i22 += 8;
930
931 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
932 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
933
934 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
935
936
937 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
938 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
939 i23 += 8;
940
941 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
942 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
943
944 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
945
946 const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
947 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
948 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
949
950 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
951 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
952 i24 += 8;
953
954 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
955 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
956
957 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
958
959 const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
960 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
961 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
962
963 k += 8;
964
965 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
966 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
967
968 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
969 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
970
971 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
972 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
973
974 vacc0123 = _mm_cvtps_epi32(vscaled0123);
975 vacc4567 = _mm_cvtps_epi32(vscaled4567);
976
977 w = (const void*) ((const int32_t*) w + 8);
978
979 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
980
981 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
982
983 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
984
985
986 if XNN_LIKELY(c >= 8) {
987 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
988 output += 8;
989 c -= 8;
990 } else {
991 if (c & 4) {
992 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
993 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
994 output += 4;
995 }
996 if (c & 2) {
997 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
998 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
999 output += 2;
1000 }
1001 if (c & 1) {
1002 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
1003 output += 1;
1004 }
1005 c = 0;
1006 }
1007 } while (c != 0);
1008 }
1009
1010 output = (int8_t*) ((uintptr_t) output + output_increment);
1011 } while (--output_width != 0);
1012 }