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
16
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse2_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse2_mul16_add16(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 const int8_t* i9 = input[9];
79 assert(i9 != NULL);
80 if XNN_UNPREDICTABLE(i9 != zero) {
81 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
82 }
83 const int8_t* i10 = input[10];
84 assert(i10 != NULL);
85 if XNN_UNPREDICTABLE(i10 != zero) {
86 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
87 }
88 const int8_t* i11 = input[11];
89 assert(i11 != NULL);
90 if XNN_UNPREDICTABLE(i11 != zero) {
91 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
92 }
93 const int8_t* i12 = input[12];
94 assert(i12 != NULL);
95 if XNN_UNPREDICTABLE(i12 != zero) {
96 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
97 }
98 const int8_t* i13 = input[13];
99 assert(i13 != NULL);
100 if XNN_UNPREDICTABLE(i13 != zero) {
101 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
102 }
103 const int8_t* i14 = input[14];
104 assert(i14 != NULL);
105 if XNN_UNPREDICTABLE(i14 != zero) {
106 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
107 }
108 const int8_t* i15 = input[15];
109 assert(i15 != NULL);
110 if XNN_UNPREDICTABLE(i15 != zero) {
111 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
112 }
113 const int8_t* i16 = input[16];
114 assert(i16 != NULL);
115 if XNN_UNPREDICTABLE(i16 != zero) {
116 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
117 }
118 const int8_t* i17 = input[17];
119 assert(i17 != NULL);
120 if XNN_UNPREDICTABLE(i17 != zero) {
121 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
122 }
123 const int8_t* i18 = input[18];
124 assert(i18 != NULL);
125 if XNN_UNPREDICTABLE(i18 != zero) {
126 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
127 }
128 const int8_t* i19 = input[19];
129 assert(i19 != NULL);
130 if XNN_UNPREDICTABLE(i19 != zero) {
131 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
132 }
133 const int8_t* i20 = input[20];
134 assert(i20 != NULL);
135 if XNN_UNPREDICTABLE(i20 != zero) {
136 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
137 }
138 const int8_t* i21 = input[21];
139 assert(i21 != NULL);
140 if XNN_UNPREDICTABLE(i21 != zero) {
141 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
142 }
143 const int8_t* i22 = input[22];
144 assert(i22 != NULL);
145 if XNN_UNPREDICTABLE(i22 != zero) {
146 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
147 }
148 const int8_t* i23 = input[23];
149 assert(i23 != NULL);
150 if XNN_UNPREDICTABLE(i23 != zero) {
151 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
152 }
153 const int8_t* i24 = input[24];
154 assert(i24 != NULL);
155 if XNN_UNPREDICTABLE(i24 != zero) {
156 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
157 }
158 input = (const int8_t**) ((uintptr_t) input + input_stride);
159
160 size_t c = channels;
161 const void* w = weights;
162 for (; c >= 16; c -= 16) {
163 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
164 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
165 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
166 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
167
168
169 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
170 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
171 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
172 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
173 i0 += 16;
174
175 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
176 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
177 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
178 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
179
180 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
181 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
182
183
184 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
185 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
186 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
187 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
188 i1 += 16;
189
190 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
191 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
192 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
193 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
194
195 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
196 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
197
198 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
199 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
200 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
201 const __m128i vsignprod1x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
202 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod1x89ABCDEF));
203 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod1x89ABCDEF));
204
205 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
206 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
207 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
208 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
209 i2 += 16;
210
211 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
212 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
213 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
214 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
215
216 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
217 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
218
219
220 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
221 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
222 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
223 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
224 i3 += 16;
225
226 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
227 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
228 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
229 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
230
231 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
232 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
233
234 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
235 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
236 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
237 const __m128i vsignprod3x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
238 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod3x89ABCDEF));
239 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod3x89ABCDEF));
240
241 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
242 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
243 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
244 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
245 i4 += 16;
246
247 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
248 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
249 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
250 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
251
252 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
253 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
254
255
256 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
257 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
258 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
259 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
260 i5 += 16;
261
262 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
263 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
264 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
265 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
266
267 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
268 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
269
270 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
271 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
272 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
273 const __m128i vsignprod5x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
274 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod5x89ABCDEF));
275 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod5x89ABCDEF));
276
277 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
278 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
279 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
280 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
281 i6 += 16;
282
283 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
284 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
285 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
286 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
287
288 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
289 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
290
291
292 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
293 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
294 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
295 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
296 i7 += 16;
297
298 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
299 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
300 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
301 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
302
303 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
304 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
305
306 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
307 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
308 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
309 const __m128i vsignprod7x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
310 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod7x89ABCDEF));
311 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod7x89ABCDEF));
312
313 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
314 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
315 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
316 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
317 i8 += 16;
318
319 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
320 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
321 const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
322 const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
323
324 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
325 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
326
327
328 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
329 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
330 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
331 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
332 i9 += 16;
333
334 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
335 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
336 const __m128i vxi9x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x89ABCDEF, vi9x89ABCDEF), 8);
337 const __m128i vxk9x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x89ABCDEF, vk9x89ABCDEF), 8);
338
339 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
340 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF));
341
342 const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
343 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
344 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
345 const __m128i vsignprod9x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
346 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod9x89ABCDEF));
347 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod9x89ABCDEF));
348
349 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
350 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
351 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
352 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
353 i10 += 16;
354
355 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
356 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
357 const __m128i vxi10x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x89ABCDEF, vi10x89ABCDEF), 8);
358 const __m128i vxk10x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x89ABCDEF, vk10x89ABCDEF), 8);
359
360 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
361 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
362
363
364 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
365 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
366 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
367 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
368 i11 += 16;
369
370 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
371 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
372 const __m128i vxi11x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x89ABCDEF, vi11x89ABCDEF), 8);
373 const __m128i vxk11x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x89ABCDEF, vk11x89ABCDEF), 8);
374
375 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
376 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF));
377
378 const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
379 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
380 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
381 const __m128i vsignprod11x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
382 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod11x89ABCDEF));
383 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod11x89ABCDEF));
384
385 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
386 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
387 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
388 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
389 i12 += 16;
390
391 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
392 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
393 const __m128i vxi12x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x89ABCDEF, vi12x89ABCDEF), 8);
394 const __m128i vxk12x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x89ABCDEF, vk12x89ABCDEF), 8);
395
396 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
397 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
398
399
400 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
401 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
402 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
403 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
404 i13 += 16;
405
406 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
407 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
408 const __m128i vxi13x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x89ABCDEF, vi13x89ABCDEF), 8);
409 const __m128i vxk13x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x89ABCDEF, vk13x89ABCDEF), 8);
410
411 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
412 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF));
413
414 const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
415 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
416 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
417 const __m128i vsignprod13x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
418 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod13x89ABCDEF));
419 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod13x89ABCDEF));
420
421 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
422 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
423 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
424 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
425 i14 += 16;
426
427 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
428 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
429 const __m128i vxi14x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x89ABCDEF, vi14x89ABCDEF), 8);
430 const __m128i vxk14x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x89ABCDEF, vk14x89ABCDEF), 8);
431
432 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
433 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
434
435
436 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
437 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
438 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
439 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
440 i15 += 16;
441
442 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
443 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
444 const __m128i vxi15x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x89ABCDEF, vi15x89ABCDEF), 8);
445 const __m128i vxk15x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x89ABCDEF, vk15x89ABCDEF), 8);
446
447 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
448 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF));
449
450 const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
451 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
452 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
453 const __m128i vsignprod15x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
454 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod15x89ABCDEF));
455 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod15x89ABCDEF));
456
457 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
458 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
459 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
460 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
461 i16 += 16;
462
463 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
464 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
465 const __m128i vxi16x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x89ABCDEF, vi16x89ABCDEF), 8);
466 const __m128i vxk16x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x89ABCDEF, vk16x89ABCDEF), 8);
467
468 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
469 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
470
471
472 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
473 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
474 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
475 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
476 i17 += 16;
477
478 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
479 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
480 const __m128i vxi17x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x89ABCDEF, vi17x89ABCDEF), 8);
481 const __m128i vxk17x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x89ABCDEF, vk17x89ABCDEF), 8);
482
483 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
484 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF));
485
486 const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
487 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
488 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
489 const __m128i vsignprod17x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
490 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod17x89ABCDEF));
491 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod17x89ABCDEF));
492
493 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
494 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
495 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
496 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
497 i18 += 16;
498
499 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
500 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
501 const __m128i vxi18x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x89ABCDEF, vi18x89ABCDEF), 8);
502 const __m128i vxk18x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x89ABCDEF, vk18x89ABCDEF), 8);
503
504 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
505 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
506
507
508 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
509 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
510 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
511 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
512 i19 += 16;
513
514 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
515 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
516 const __m128i vxi19x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x89ABCDEF, vi19x89ABCDEF), 8);
517 const __m128i vxk19x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x89ABCDEF, vk19x89ABCDEF), 8);
518
519 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
520 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF));
521
522 const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
523 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
524 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
525 const __m128i vsignprod19x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
526 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod19x89ABCDEF));
527 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod19x89ABCDEF));
528
529 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
530 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
531 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
532 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
533 i20 += 16;
534
535 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
536 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
537 const __m128i vxi20x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x89ABCDEF, vi20x89ABCDEF), 8);
538 const __m128i vxk20x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x89ABCDEF, vk20x89ABCDEF), 8);
539
540 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
541 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
542
543
544 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
545 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
546 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
547 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
548 i21 += 16;
549
550 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
551 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
552 const __m128i vxi21x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x89ABCDEF, vi21x89ABCDEF), 8);
553 const __m128i vxk21x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x89ABCDEF, vk21x89ABCDEF), 8);
554
555 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
556 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF));
557
558 const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
559 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
560 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
561 const __m128i vsignprod21x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
562 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod21x89ABCDEF));
563 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod21x89ABCDEF));
564
565 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
566 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
567 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
568 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
569 i22 += 16;
570
571 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
572 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
573 const __m128i vxi22x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x89ABCDEF, vi22x89ABCDEF), 8);
574 const __m128i vxk22x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x89ABCDEF, vk22x89ABCDEF), 8);
575
576 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
577 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
578
579
580 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
581 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
582 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
583 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
584 i23 += 16;
585
586 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
587 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
588 const __m128i vxi23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x89ABCDEF, vi23x89ABCDEF), 8);
589 const __m128i vxk23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x89ABCDEF, vk23x89ABCDEF), 8);
590
591 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
592 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF));
593
594 const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
595 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
596 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
597 const __m128i vsignprod23x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
598 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod23x89ABCDEF));
599 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod23x89ABCDEF));
600
601 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
602 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
603 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
604 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
605 i24 += 16;
606
607 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
608 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
609 const __m128i vxi24x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x89ABCDEF, vi24x89ABCDEF), 8);
610 const __m128i vxk24x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x89ABCDEF, vk24x89ABCDEF), 8);
611
612 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
613 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
614
615 const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
616 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
617 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
618 const __m128i vsignprod24x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
619 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod24x89ABCDEF));
620 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod24x89ABCDEF));
621
622 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
623
624 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
625 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
626 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
627 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
628
629 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
630 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
631 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
632 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
633 w = (const void*) ((const float*) w + 16);
634 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
635 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
636 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
637 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
638
639 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point);
640 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
641 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
642 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
643 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
644
645 vacc0123 = _mm_cvtps_epi32(vscaled0123);
646 vacc4567 = _mm_cvtps_epi32(vscaled4567);
647 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
648 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
649
650 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
651 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
652 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
653
654 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
655 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
656 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
657
658 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
659
660
661 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
662 output += 16;
663 }
664 if XNN_UNLIKELY(c != 0) {
665 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
666 do {
667 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
668 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
669
670
671 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
672 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
673 i0 += 8;
674
675 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
676 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
677
678 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
679
680
681 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
682 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
683 i1 += 8;
684
685 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
686 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
687
688 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
689
690 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
691 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
692 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
693
694 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
695 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
696 i2 += 8;
697
698 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
699 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
700
701 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
702
703
704 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
705 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
706 i3 += 8;
707
708 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
709 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
710
711 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
712
713 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
714 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
715 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
716
717 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
718 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
719 i4 += 8;
720
721 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
722 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
723
724 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
725
726
727 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
728 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
729 i5 += 8;
730
731 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
732 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
733
734 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
735
736 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
737 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
738 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
739
740 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
741 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
742 i6 += 8;
743
744 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
745 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
746
747 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
748
749
750 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
751 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
752 i7 += 8;
753
754 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
755 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
756
757 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
758
759 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
760 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
761 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
762
763 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
764 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
765 i8 += 8;
766
767 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
768 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
769
770 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
771
772
773 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
774 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
775 i9 += 8;
776
777 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
778 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
779
780 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
781
782 const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
783 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
784 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
785
786 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
787 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
788 i10 += 8;
789
790 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
791 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
792
793 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
794
795
796 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
797 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
798 i11 += 8;
799
800 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
801 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
802
803 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
804
805 const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
806 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
807 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
808
809 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
810 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
811 i12 += 8;
812
813 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
814 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
815
816 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
817
818
819 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
820 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
821 i13 += 8;
822
823 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
824 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
825
826 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
827
828 const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
829 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
830 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
831
832 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
833 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
834 i14 += 8;
835
836 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
837 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
838
839 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
840
841
842 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
843 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
844 i15 += 8;
845
846 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
847 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
848
849 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
850
851 const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
852 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
853 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
854
855 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
856 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
857 i16 += 8;
858
859 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
860 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
861
862 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
863
864
865 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
866 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
867 i17 += 8;
868
869 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
870 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
871
872 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
873
874 const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
875 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
876 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
877
878 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
879 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
880 i18 += 8;
881
882 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
883 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
884
885 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
886
887
888 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
889 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
890 i19 += 8;
891
892 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
893 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
894
895 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
896
897 const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
898 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
899 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
900
901 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
902 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
903 i20 += 8;
904
905 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
906 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
907
908 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
909
910
911 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
912 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
913 i21 += 8;
914
915 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
916 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
917
918 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
919
920 const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
921 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
922 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
923
924 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
925 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
926 i22 += 8;
927
928 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
929 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
930
931 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
932
933
934 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
935 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
936 i23 += 8;
937
938 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
939 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
940
941 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
942
943 const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
944 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
945 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
946
947 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
948 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
949 i24 += 8;
950
951 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
952 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
953
954 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
955
956 const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
957 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
958 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
959
960 k += 8;
961
962 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
963 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
964
965 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t)));
966 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t) + 4 * sizeof(float)));
967 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
968 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
969
970 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point);
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 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
980 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
981
982 vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->sse2.output_min));
983
984 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
985
986
987 if XNN_LIKELY(c >= 8) {
988 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
989 output += 8;
990 c -= 8;
991 } else {
992 if (c & 4) {
993 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
994 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
995 output += 4;
996 }
997 if (c & 2) {
998 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
999 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1000 output += 2;
1001 }
1002 if (c & 1) {
1003 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
1004 output += 1;
1005 }
1006 c = 0;
1007 }
1008 } while (c != 0);
1009 }
1010
1011 output = (int8_t*) ((uintptr_t) output + output_increment);
1012 } while (--output_width != 0);
1013 }
1014