1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <smmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__avx_mul16(
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 >= 24; c -= 24) {
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 __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
168 __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
169
170
171 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
172 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
173 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)));
174 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
175 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
176 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
177 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)));
178 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
179 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
180 const __m128i vxi0xGHIJKLMN = _mm_cvtepi8_epi16(vi0xGHIJKLMN);
181 const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)));
182 const __m128i vxk0xGHIJKLMN = _mm_cvtepi8_epi16(vk0xGHIJKLMN);
183 i0 += 24;
184
185
186 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
187 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
188 __m128i vprodGHIJKLMN = _mm_mullo_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN);
189
190 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
191 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
192 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
193 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
194 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
195 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
196
197 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
198 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
199 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)));
200 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
201 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
202 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
203 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)));
204 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
205 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
206 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(vi1xGHIJKLMN);
207 const __m128i vk1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)));
208 const __m128i vxk1xGHIJKLMN = _mm_cvtepi8_epi16(vk1xGHIJKLMN);
209 i1 += 24;
210
211
212 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
213 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
214 vprodGHIJKLMN = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN);
215
216 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
217 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
218 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
219 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
220 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
221 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
222
223 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
224 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
225 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)));
226 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
227 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
228 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
229 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)));
230 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
231 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
232 const __m128i vxi2xGHIJKLMN = _mm_cvtepi8_epi16(vi2xGHIJKLMN);
233 const __m128i vk2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)));
234 const __m128i vxk2xGHIJKLMN = _mm_cvtepi8_epi16(vk2xGHIJKLMN);
235 i2 += 24;
236
237
238 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
239 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
240 vprodGHIJKLMN = _mm_mullo_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN);
241
242 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
243 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
244 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
245 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
246 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
247 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
248
249 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
250 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
251 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)));
252 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
253 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
254 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
255 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)));
256 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
257 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
258 const __m128i vxi3xGHIJKLMN = _mm_cvtepi8_epi16(vi3xGHIJKLMN);
259 const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)));
260 const __m128i vxk3xGHIJKLMN = _mm_cvtepi8_epi16(vk3xGHIJKLMN);
261 i3 += 24;
262
263
264 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
265 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
266 vprodGHIJKLMN = _mm_mullo_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN);
267
268 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
269 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
270 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
271 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
272 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
273 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
274
275 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
276 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
277 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)));
278 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
279 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
280 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
281 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)));
282 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
283 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
284 const __m128i vxi4xGHIJKLMN = _mm_cvtepi8_epi16(vi4xGHIJKLMN);
285 const __m128i vk4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)));
286 const __m128i vxk4xGHIJKLMN = _mm_cvtepi8_epi16(vk4xGHIJKLMN);
287 i4 += 24;
288
289
290 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
291 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
292 vprodGHIJKLMN = _mm_mullo_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN);
293
294 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
295 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
296 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
297 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
298 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
299 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
300
301 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
302 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
303 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)));
304 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
305 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
306 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
307 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)));
308 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
309 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
310 const __m128i vxi5xGHIJKLMN = _mm_cvtepi8_epi16(vi5xGHIJKLMN);
311 const __m128i vk5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)));
312 const __m128i vxk5xGHIJKLMN = _mm_cvtepi8_epi16(vk5xGHIJKLMN);
313 i5 += 24;
314
315
316 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
317 vprod89ABCDEF = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
318 vprodGHIJKLMN = _mm_mullo_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN);
319
320 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
321 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
322 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
323 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
324 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
325 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
326
327 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
328 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
329 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)));
330 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
331 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
332 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
333 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)));
334 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
335 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
336 const __m128i vxi6xGHIJKLMN = _mm_cvtepi8_epi16(vi6xGHIJKLMN);
337 const __m128i vk6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)));
338 const __m128i vxk6xGHIJKLMN = _mm_cvtepi8_epi16(vk6xGHIJKLMN);
339 i6 += 24;
340
341
342 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
343 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
344 vprodGHIJKLMN = _mm_mullo_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN);
345
346 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
347 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
348 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
349 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
350 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
351 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
352
353 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
354 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
355 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)));
356 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
357 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
358 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
359 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)));
360 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
361 const __m128i vi7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i7 + 16));
362 const __m128i vxi7xGHIJKLMN = _mm_cvtepi8_epi16(vi7xGHIJKLMN);
363 const __m128i vk7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)));
364 const __m128i vxk7xGHIJKLMN = _mm_cvtepi8_epi16(vk7xGHIJKLMN);
365 i7 += 24;
366
367
368 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
369 vprod89ABCDEF = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
370 vprodGHIJKLMN = _mm_mullo_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN);
371
372 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
373 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
374 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
375 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
376 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
377 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
378
379 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
380 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
381 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)));
382 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
383 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
384 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
385 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)));
386 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
387 const __m128i vi8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i8 + 16));
388 const __m128i vxi8xGHIJKLMN = _mm_cvtepi8_epi16(vi8xGHIJKLMN);
389 const __m128i vk8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)));
390 const __m128i vxk8xGHIJKLMN = _mm_cvtepi8_epi16(vk8xGHIJKLMN);
391 i8 += 24;
392
393
394 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
395 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
396 vprodGHIJKLMN = _mm_mullo_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN);
397
398 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
399 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
400 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
401 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
402 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
403 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
404
405 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
406 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
407 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)));
408 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
409 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
410 const __m128i vxi9x89ABCDEF = _mm_cvtepi8_epi16(vi9x89ABCDEF);
411 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 224 * sizeof(int8_t)));
412 const __m128i vxk9x89ABCDEF = _mm_cvtepi8_epi16(vk9x89ABCDEF);
413 const __m128i vi9xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i9 + 16));
414 const __m128i vxi9xGHIJKLMN = _mm_cvtepi8_epi16(vi9xGHIJKLMN);
415 const __m128i vk9xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 232 * sizeof(int8_t)));
416 const __m128i vxk9xGHIJKLMN = _mm_cvtepi8_epi16(vk9xGHIJKLMN);
417 i9 += 24;
418
419
420 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
421 vprod89ABCDEF = _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
422 vprodGHIJKLMN = _mm_mullo_epi16(vxi9xGHIJKLMN, vxk9xGHIJKLMN);
423
424 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
425 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
426 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
427 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
428 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
429 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
430
431 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
432 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
433 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 240 * sizeof(int8_t)));
434 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
435 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
436 const __m128i vxi10x89ABCDEF = _mm_cvtepi8_epi16(vi10x89ABCDEF);
437 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 248 * sizeof(int8_t)));
438 const __m128i vxk10x89ABCDEF = _mm_cvtepi8_epi16(vk10x89ABCDEF);
439 const __m128i vi10xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i10 + 16));
440 const __m128i vxi10xGHIJKLMN = _mm_cvtepi8_epi16(vi10xGHIJKLMN);
441 const __m128i vk10xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 256 * sizeof(int8_t)));
442 const __m128i vxk10xGHIJKLMN = _mm_cvtepi8_epi16(vk10xGHIJKLMN);
443 i10 += 24;
444
445
446 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
447 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
448 vprodGHIJKLMN = _mm_mullo_epi16(vxi10xGHIJKLMN, vxk10xGHIJKLMN);
449
450 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
451 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
452 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
453 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
454 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
455 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
456
457 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
458 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
459 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 264 * sizeof(int8_t)));
460 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
461 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
462 const __m128i vxi11x89ABCDEF = _mm_cvtepi8_epi16(vi11x89ABCDEF);
463 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 272 * sizeof(int8_t)));
464 const __m128i vxk11x89ABCDEF = _mm_cvtepi8_epi16(vk11x89ABCDEF);
465 const __m128i vi11xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i11 + 16));
466 const __m128i vxi11xGHIJKLMN = _mm_cvtepi8_epi16(vi11xGHIJKLMN);
467 const __m128i vk11xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 280 * sizeof(int8_t)));
468 const __m128i vxk11xGHIJKLMN = _mm_cvtepi8_epi16(vk11xGHIJKLMN);
469 i11 += 24;
470
471
472 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
473 vprod89ABCDEF = _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
474 vprodGHIJKLMN = _mm_mullo_epi16(vxi11xGHIJKLMN, vxk11xGHIJKLMN);
475
476 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
477 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
478 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
479 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
480 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
481 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
482
483 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
484 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
485 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 288 * sizeof(int8_t)));
486 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
487 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
488 const __m128i vxi12x89ABCDEF = _mm_cvtepi8_epi16(vi12x89ABCDEF);
489 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 296 * sizeof(int8_t)));
490 const __m128i vxk12x89ABCDEF = _mm_cvtepi8_epi16(vk12x89ABCDEF);
491 const __m128i vi12xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i12 + 16));
492 const __m128i vxi12xGHIJKLMN = _mm_cvtepi8_epi16(vi12xGHIJKLMN);
493 const __m128i vk12xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 304 * sizeof(int8_t)));
494 const __m128i vxk12xGHIJKLMN = _mm_cvtepi8_epi16(vk12xGHIJKLMN);
495 i12 += 24;
496
497
498 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
499 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
500 vprodGHIJKLMN = _mm_mullo_epi16(vxi12xGHIJKLMN, vxk12xGHIJKLMN);
501
502 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
503 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
504 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
505 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
506 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
507 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
508
509 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
510 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
511 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 312 * sizeof(int8_t)));
512 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
513 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
514 const __m128i vxi13x89ABCDEF = _mm_cvtepi8_epi16(vi13x89ABCDEF);
515 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 320 * sizeof(int8_t)));
516 const __m128i vxk13x89ABCDEF = _mm_cvtepi8_epi16(vk13x89ABCDEF);
517 const __m128i vi13xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i13 + 16));
518 const __m128i vxi13xGHIJKLMN = _mm_cvtepi8_epi16(vi13xGHIJKLMN);
519 const __m128i vk13xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 328 * sizeof(int8_t)));
520 const __m128i vxk13xGHIJKLMN = _mm_cvtepi8_epi16(vk13xGHIJKLMN);
521 i13 += 24;
522
523
524 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
525 vprod89ABCDEF = _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
526 vprodGHIJKLMN = _mm_mullo_epi16(vxi13xGHIJKLMN, vxk13xGHIJKLMN);
527
528 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
529 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
530 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
531 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
532 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
533 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
534
535 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
536 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
537 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 336 * sizeof(int8_t)));
538 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
539 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
540 const __m128i vxi14x89ABCDEF = _mm_cvtepi8_epi16(vi14x89ABCDEF);
541 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 344 * sizeof(int8_t)));
542 const __m128i vxk14x89ABCDEF = _mm_cvtepi8_epi16(vk14x89ABCDEF);
543 const __m128i vi14xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i14 + 16));
544 const __m128i vxi14xGHIJKLMN = _mm_cvtepi8_epi16(vi14xGHIJKLMN);
545 const __m128i vk14xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 352 * sizeof(int8_t)));
546 const __m128i vxk14xGHIJKLMN = _mm_cvtepi8_epi16(vk14xGHIJKLMN);
547 i14 += 24;
548
549
550 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
551 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
552 vprodGHIJKLMN = _mm_mullo_epi16(vxi14xGHIJKLMN, vxk14xGHIJKLMN);
553
554 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
555 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
556 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
557 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
558 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
559 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
560
561 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
562 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
563 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 360 * sizeof(int8_t)));
564 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
565 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
566 const __m128i vxi15x89ABCDEF = _mm_cvtepi8_epi16(vi15x89ABCDEF);
567 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 368 * sizeof(int8_t)));
568 const __m128i vxk15x89ABCDEF = _mm_cvtepi8_epi16(vk15x89ABCDEF);
569 const __m128i vi15xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i15 + 16));
570 const __m128i vxi15xGHIJKLMN = _mm_cvtepi8_epi16(vi15xGHIJKLMN);
571 const __m128i vk15xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 376 * sizeof(int8_t)));
572 const __m128i vxk15xGHIJKLMN = _mm_cvtepi8_epi16(vk15xGHIJKLMN);
573 i15 += 24;
574
575
576 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
577 vprod89ABCDEF = _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
578 vprodGHIJKLMN = _mm_mullo_epi16(vxi15xGHIJKLMN, vxk15xGHIJKLMN);
579
580 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
581 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
582 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
583 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
584 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
585 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
586
587 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
588 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
589 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 384 * sizeof(int8_t)));
590 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
591 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
592 const __m128i vxi16x89ABCDEF = _mm_cvtepi8_epi16(vi16x89ABCDEF);
593 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 392 * sizeof(int8_t)));
594 const __m128i vxk16x89ABCDEF = _mm_cvtepi8_epi16(vk16x89ABCDEF);
595 const __m128i vi16xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i16 + 16));
596 const __m128i vxi16xGHIJKLMN = _mm_cvtepi8_epi16(vi16xGHIJKLMN);
597 const __m128i vk16xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 400 * sizeof(int8_t)));
598 const __m128i vxk16xGHIJKLMN = _mm_cvtepi8_epi16(vk16xGHIJKLMN);
599 i16 += 24;
600
601
602 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
603 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
604 vprodGHIJKLMN = _mm_mullo_epi16(vxi16xGHIJKLMN, vxk16xGHIJKLMN);
605
606 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
607 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
608 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
609 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
610 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
611 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
612
613 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
614 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
615 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 408 * sizeof(int8_t)));
616 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
617 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
618 const __m128i vxi17x89ABCDEF = _mm_cvtepi8_epi16(vi17x89ABCDEF);
619 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 416 * sizeof(int8_t)));
620 const __m128i vxk17x89ABCDEF = _mm_cvtepi8_epi16(vk17x89ABCDEF);
621 const __m128i vi17xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i17 + 16));
622 const __m128i vxi17xGHIJKLMN = _mm_cvtepi8_epi16(vi17xGHIJKLMN);
623 const __m128i vk17xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 424 * sizeof(int8_t)));
624 const __m128i vxk17xGHIJKLMN = _mm_cvtepi8_epi16(vk17xGHIJKLMN);
625 i17 += 24;
626
627
628 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
629 vprod89ABCDEF = _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
630 vprodGHIJKLMN = _mm_mullo_epi16(vxi17xGHIJKLMN, vxk17xGHIJKLMN);
631
632 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
633 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
634 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
635 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
636 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
637 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
638
639 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
640 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
641 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 432 * sizeof(int8_t)));
642 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
643 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
644 const __m128i vxi18x89ABCDEF = _mm_cvtepi8_epi16(vi18x89ABCDEF);
645 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 440 * sizeof(int8_t)));
646 const __m128i vxk18x89ABCDEF = _mm_cvtepi8_epi16(vk18x89ABCDEF);
647 const __m128i vi18xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i18 + 16));
648 const __m128i vxi18xGHIJKLMN = _mm_cvtepi8_epi16(vi18xGHIJKLMN);
649 const __m128i vk18xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 448 * sizeof(int8_t)));
650 const __m128i vxk18xGHIJKLMN = _mm_cvtepi8_epi16(vk18xGHIJKLMN);
651 i18 += 24;
652
653
654 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
655 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
656 vprodGHIJKLMN = _mm_mullo_epi16(vxi18xGHIJKLMN, vxk18xGHIJKLMN);
657
658 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
659 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
660 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
661 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
662 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
663 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
664
665 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
666 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
667 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 456 * sizeof(int8_t)));
668 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
669 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
670 const __m128i vxi19x89ABCDEF = _mm_cvtepi8_epi16(vi19x89ABCDEF);
671 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 464 * sizeof(int8_t)));
672 const __m128i vxk19x89ABCDEF = _mm_cvtepi8_epi16(vk19x89ABCDEF);
673 const __m128i vi19xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i19 + 16));
674 const __m128i vxi19xGHIJKLMN = _mm_cvtepi8_epi16(vi19xGHIJKLMN);
675 const __m128i vk19xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 472 * sizeof(int8_t)));
676 const __m128i vxk19xGHIJKLMN = _mm_cvtepi8_epi16(vk19xGHIJKLMN);
677 i19 += 24;
678
679
680 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
681 vprod89ABCDEF = _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
682 vprodGHIJKLMN = _mm_mullo_epi16(vxi19xGHIJKLMN, vxk19xGHIJKLMN);
683
684 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
685 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
686 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
687 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
688 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
689 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
690
691 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
692 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
693 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 480 * sizeof(int8_t)));
694 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
695 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
696 const __m128i vxi20x89ABCDEF = _mm_cvtepi8_epi16(vi20x89ABCDEF);
697 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 488 * sizeof(int8_t)));
698 const __m128i vxk20x89ABCDEF = _mm_cvtepi8_epi16(vk20x89ABCDEF);
699 const __m128i vi20xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i20 + 16));
700 const __m128i vxi20xGHIJKLMN = _mm_cvtepi8_epi16(vi20xGHIJKLMN);
701 const __m128i vk20xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 496 * sizeof(int8_t)));
702 const __m128i vxk20xGHIJKLMN = _mm_cvtepi8_epi16(vk20xGHIJKLMN);
703 i20 += 24;
704
705
706 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
707 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
708 vprodGHIJKLMN = _mm_mullo_epi16(vxi20xGHIJKLMN, vxk20xGHIJKLMN);
709
710 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
711 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
712 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
713 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
714 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
715 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
716
717 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
718 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
719 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 504 * sizeof(int8_t)));
720 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
721 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
722 const __m128i vxi21x89ABCDEF = _mm_cvtepi8_epi16(vi21x89ABCDEF);
723 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 512 * sizeof(int8_t)));
724 const __m128i vxk21x89ABCDEF = _mm_cvtepi8_epi16(vk21x89ABCDEF);
725 const __m128i vi21xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i21 + 16));
726 const __m128i vxi21xGHIJKLMN = _mm_cvtepi8_epi16(vi21xGHIJKLMN);
727 const __m128i vk21xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 520 * sizeof(int8_t)));
728 const __m128i vxk21xGHIJKLMN = _mm_cvtepi8_epi16(vk21xGHIJKLMN);
729 i21 += 24;
730
731
732 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
733 vprod89ABCDEF = _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
734 vprodGHIJKLMN = _mm_mullo_epi16(vxi21xGHIJKLMN, vxk21xGHIJKLMN);
735
736 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
737 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
738 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
739 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
740 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
741 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
742
743 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
744 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
745 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 528 * sizeof(int8_t)));
746 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
747 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
748 const __m128i vxi22x89ABCDEF = _mm_cvtepi8_epi16(vi22x89ABCDEF);
749 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 536 * sizeof(int8_t)));
750 const __m128i vxk22x89ABCDEF = _mm_cvtepi8_epi16(vk22x89ABCDEF);
751 const __m128i vi22xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i22 + 16));
752 const __m128i vxi22xGHIJKLMN = _mm_cvtepi8_epi16(vi22xGHIJKLMN);
753 const __m128i vk22xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 544 * sizeof(int8_t)));
754 const __m128i vxk22xGHIJKLMN = _mm_cvtepi8_epi16(vk22xGHIJKLMN);
755 i22 += 24;
756
757
758 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
759 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
760 vprodGHIJKLMN = _mm_mullo_epi16(vxi22xGHIJKLMN, vxk22xGHIJKLMN);
761
762 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
763 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
764 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
765 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
766 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
767 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
768
769 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
770 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
771 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 552 * sizeof(int8_t)));
772 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
773 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
774 const __m128i vxi23x89ABCDEF = _mm_cvtepi8_epi16(vi23x89ABCDEF);
775 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 560 * sizeof(int8_t)));
776 const __m128i vxk23x89ABCDEF = _mm_cvtepi8_epi16(vk23x89ABCDEF);
777 const __m128i vi23xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i23 + 16));
778 const __m128i vxi23xGHIJKLMN = _mm_cvtepi8_epi16(vi23xGHIJKLMN);
779 const __m128i vk23xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 568 * sizeof(int8_t)));
780 const __m128i vxk23xGHIJKLMN = _mm_cvtepi8_epi16(vk23xGHIJKLMN);
781 i23 += 24;
782
783
784 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
785 vprod89ABCDEF = _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
786 vprodGHIJKLMN = _mm_mullo_epi16(vxi23xGHIJKLMN, vxk23xGHIJKLMN);
787
788 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
789 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
790 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
791 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
792 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
793 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
794
795 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
796 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
797 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 576 * sizeof(int8_t)));
798 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
799 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
800 const __m128i vxi24x89ABCDEF = _mm_cvtepi8_epi16(vi24x89ABCDEF);
801 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 584 * sizeof(int8_t)));
802 const __m128i vxk24x89ABCDEF = _mm_cvtepi8_epi16(vk24x89ABCDEF);
803 const __m128i vi24xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i24 + 16));
804 const __m128i vxi24xGHIJKLMN = _mm_cvtepi8_epi16(vi24xGHIJKLMN);
805 const __m128i vk24xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 592 * sizeof(int8_t)));
806 const __m128i vxk24xGHIJKLMN = _mm_cvtepi8_epi16(vk24xGHIJKLMN);
807 i24 += 24;
808
809
810 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
811 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
812 vprodGHIJKLMN = _mm_mullo_epi16(vxi24xGHIJKLMN, vxk24xGHIJKLMN);
813
814 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
815 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
816 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
817 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
818 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
819 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
820
821 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t));
822
823 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
824 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
825 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
826 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
827 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
828 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
829
830 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
831 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
832 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
833 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
834 const __m128 vscaleGHIJ = _mm_loadu_ps((const float*) w + 16);
835 const __m128 vscaleKLMN = _mm_loadu_ps((const float*) w + 20);
836 w = (const void*) ((const float*) w + 24);
837 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
838 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
839 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
840 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
841 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscaleGHIJ);
842 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscaleKLMN);
843
844 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
845 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
846 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
847 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
848 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
849 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
850 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
851
852 vacc0123 = _mm_cvtps_epi32(vscaled0123);
853 vacc4567 = _mm_cvtps_epi32(vscaled4567);
854 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
855 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
856 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
857 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
858
859 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
860 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
861 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
862 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
863
864
865 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
866 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
867
868 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
869 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
870 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
871
872 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
873 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
874 output += 24;
875 }
876 if XNN_UNLIKELY(c != 0) {
877 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
878 do {
879 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
880 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
881
882
883 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
884 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
885 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
886 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
887 i0 += 8;
888
889
890 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
891
892 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
893 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
894
895 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
896 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
897 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 24));
898 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
899 i1 += 8;
900
901
902 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
903
904 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
905 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
906
907 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
908 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
909 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
910 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
911 i2 += 8;
912
913
914 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
915
916 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
917 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
918
919 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
920 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
921 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 72));
922 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
923 i3 += 8;
924
925
926 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
927
928 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
929 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
930
931 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
932 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
933 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
934 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
935 i4 += 8;
936
937
938 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
939
940 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
941 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
942
943 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
944 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
945 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 120));
946 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
947 i5 += 8;
948
949
950 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
951
952 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
953 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
954
955 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
956 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
957 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
958 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
959 i6 += 8;
960
961
962 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
963
964 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
965 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
966
967 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
968 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
969 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 168));
970 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
971 i7 += 8;
972
973
974 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
975
976 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
977 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
978
979 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
980 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
981 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
982 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
983 i8 += 8;
984
985
986 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
987
988 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
989 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
990
991 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
992 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
993 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 216));
994 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
995 i9 += 8;
996
997
998 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
999
1000 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1001 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1002
1003 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
1004 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
1005 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
1006 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
1007 i10 += 8;
1008
1009
1010 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
1011
1012 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1013 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1014
1015 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
1016 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
1017 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 264));
1018 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
1019 i11 += 8;
1020
1021
1022 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
1023
1024 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1025 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1026
1027 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
1028 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
1029 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
1030 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
1031 i12 += 8;
1032
1033
1034 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
1035
1036 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1037 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1038
1039 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
1040 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
1041 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 312));
1042 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
1043 i13 += 8;
1044
1045
1046 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
1047
1048 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1049 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1050
1051 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
1052 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
1053 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
1054 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
1055 i14 += 8;
1056
1057
1058 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
1059
1060 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1061 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1062
1063 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
1064 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
1065 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 360));
1066 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
1067 i15 += 8;
1068
1069
1070 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
1071
1072 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1073 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1074
1075 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
1076 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
1077 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
1078 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
1079 i16 += 8;
1080
1081
1082 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
1083
1084 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1085 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1086
1087 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
1088 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
1089 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 408));
1090 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
1091 i17 += 8;
1092
1093
1094 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
1095
1096 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1097 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1098
1099 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
1100 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
1101 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 432));
1102 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
1103 i18 += 8;
1104
1105
1106 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
1107
1108 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1109 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1110
1111 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
1112 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
1113 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 456));
1114 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
1115 i19 += 8;
1116
1117
1118 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
1119
1120 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1121 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1122
1123 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
1124 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
1125 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 480));
1126 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
1127 i20 += 8;
1128
1129
1130 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
1131
1132 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1133 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1134
1135 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
1136 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
1137 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 504));
1138 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
1139 i21 += 8;
1140
1141
1142 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
1143
1144 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1145 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1146
1147 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
1148 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
1149 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 528));
1150 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
1151 i22 += 8;
1152
1153
1154 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
1155
1156 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1157 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1158
1159 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
1160 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
1161 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 552));
1162 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
1163 i23 += 8;
1164
1165
1166 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
1167
1168 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1169 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1170
1171 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
1172 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
1173 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 576));
1174 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
1175 i24 += 8;
1176
1177
1178 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
1179
1180 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1181 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1182
1183 k += 8;
1184
1185 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
1186 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
1187
1188 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t)));
1189 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t) + 4 * sizeof(float)));
1190 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
1191 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
1192
1193 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
1194 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
1195 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1196
1197 vacc0123 = _mm_cvtps_epi32(vscaled0123);
1198 vacc4567 = _mm_cvtps_epi32(vscaled4567);
1199
1200 w = (const void*) ((const int32_t*) w + 8);
1201
1202 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
1203 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1204
1205
1206 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
1207
1208 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->sse4.output_min));
1209
1210 if XNN_LIKELY(c >= 8) {
1211 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
1212 output += 8;
1213 c -= 8;
1214 } else {
1215 if (c & 4) {
1216 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
1217 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
1218 output += 4;
1219 }
1220 if (c & 2) {
1221 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
1222 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1223 output += 2;
1224 }
1225 if (c & 1) {
1226 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
1227 output += 1;
1228 }
1229 c = 0;
1230 }
1231 } while (c != 0);
1232 }
1233
1234 output = (int8_t*) ((uintptr_t) output + output_increment);
1235 } while (--output_width != 0);
1236 }
1237