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_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_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_conv_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 vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
171 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
172 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
173 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
174 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
175 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
176 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
177 i0 += 16;
178
179
180 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
181 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
182
183 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
184 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
185 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
186 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
187
188 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
189 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
190 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
191 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
192 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
193 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
194 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
195 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
196 i1 += 16;
197
198
199 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
200 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
201
202 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
203 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
204 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
205 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
206
207 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
208 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
209 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
210 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
211 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
212 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
213 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
214 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
215 i2 += 16;
216
217
218 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
219 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
220
221 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
222 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
223 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
224 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
225
226 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
227 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
228 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
229 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
230 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
231 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
232 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
233 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
234 i3 += 16;
235
236
237 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
238 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
239
240 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
241 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
242 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
243 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
244
245 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
246 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
247 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
248 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
249 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
250 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
251 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
252 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
253 i4 += 16;
254
255
256 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
257 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
258
259 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
260 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
261 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
262 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
263
264 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
265 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
266 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
267 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
268 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
269 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
270 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
271 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
272 i5 += 16;
273
274
275 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
276 vprod89ABCDEF = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
277
278 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
279 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
280 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
281 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
282
283 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
284 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
285 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
286 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
287 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
288 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
289 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
290 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
291 i6 += 16;
292
293
294 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
295 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
296
297 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
298 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
299 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
300 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
301
302 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
303 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
304 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
305 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
306 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
307 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
308 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
309 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
310 i7 += 16;
311
312
313 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
314 vprod89ABCDEF = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
315
316 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
317 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
318 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
319 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
320
321 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
322 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
323 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
324 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
325 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
326 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
327 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
328 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
329 i8 += 16;
330
331
332 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
333 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
334
335 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
336 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
337 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
338 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
339
340 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
341 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
342 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
343 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
344 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
345 const __m128i vxi9x89ABCDEF = _mm_cvtepi8_epi16(vi9x89ABCDEF);
346 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
347 const __m128i vxk9x89ABCDEF = _mm_cvtepi8_epi16(vk9x89ABCDEF);
348 i9 += 16;
349
350
351 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
352 vprod89ABCDEF = _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
353
354 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
355 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
356 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
357 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
358
359 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
360 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
361 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
362 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
363 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
364 const __m128i vxi10x89ABCDEF = _mm_cvtepi8_epi16(vi10x89ABCDEF);
365 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
366 const __m128i vxk10x89ABCDEF = _mm_cvtepi8_epi16(vk10x89ABCDEF);
367 i10 += 16;
368
369
370 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
371 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
372
373 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
374 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
375 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
376 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
377
378 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
379 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
380 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
381 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
382 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
383 const __m128i vxi11x89ABCDEF = _mm_cvtepi8_epi16(vi11x89ABCDEF);
384 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
385 const __m128i vxk11x89ABCDEF = _mm_cvtepi8_epi16(vk11x89ABCDEF);
386 i11 += 16;
387
388
389 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
390 vprod89ABCDEF = _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
391
392 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
393 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
394 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
395 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
396
397 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
398 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
399 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
400 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
401 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
402 const __m128i vxi12x89ABCDEF = _mm_cvtepi8_epi16(vi12x89ABCDEF);
403 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
404 const __m128i vxk12x89ABCDEF = _mm_cvtepi8_epi16(vk12x89ABCDEF);
405 i12 += 16;
406
407
408 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
409 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
410
411 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
412 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
413 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
414 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
415
416 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
417 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
418 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
419 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
420 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
421 const __m128i vxi13x89ABCDEF = _mm_cvtepi8_epi16(vi13x89ABCDEF);
422 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
423 const __m128i vxk13x89ABCDEF = _mm_cvtepi8_epi16(vk13x89ABCDEF);
424 i13 += 16;
425
426
427 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
428 vprod89ABCDEF = _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
429
430 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
431 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
432 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
433 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
434
435 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
436 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
437 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
438 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
439 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
440 const __m128i vxi14x89ABCDEF = _mm_cvtepi8_epi16(vi14x89ABCDEF);
441 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
442 const __m128i vxk14x89ABCDEF = _mm_cvtepi8_epi16(vk14x89ABCDEF);
443 i14 += 16;
444
445
446 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
447 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
448
449 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
450 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
451 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
452 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
453
454 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
455 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
456 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
457 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
458 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
459 const __m128i vxi15x89ABCDEF = _mm_cvtepi8_epi16(vi15x89ABCDEF);
460 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
461 const __m128i vxk15x89ABCDEF = _mm_cvtepi8_epi16(vk15x89ABCDEF);
462 i15 += 16;
463
464
465 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
466 vprod89ABCDEF = _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
467
468 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
469 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
470 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
471 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
472
473 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
474 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
475 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
476 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
477 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
478 const __m128i vxi16x89ABCDEF = _mm_cvtepi8_epi16(vi16x89ABCDEF);
479 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
480 const __m128i vxk16x89ABCDEF = _mm_cvtepi8_epi16(vk16x89ABCDEF);
481 i16 += 16;
482
483
484 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
485 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
486
487 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
488 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
489 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
490 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
491
492 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
493 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
494 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
495 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
496 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
497 const __m128i vxi17x89ABCDEF = _mm_cvtepi8_epi16(vi17x89ABCDEF);
498 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
499 const __m128i vxk17x89ABCDEF = _mm_cvtepi8_epi16(vk17x89ABCDEF);
500 i17 += 16;
501
502
503 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
504 vprod89ABCDEF = _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
505
506 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
507 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
508 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
509 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
510
511 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
512 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
513 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
514 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
515 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
516 const __m128i vxi18x89ABCDEF = _mm_cvtepi8_epi16(vi18x89ABCDEF);
517 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
518 const __m128i vxk18x89ABCDEF = _mm_cvtepi8_epi16(vk18x89ABCDEF);
519 i18 += 16;
520
521
522 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
523 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
524
525 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
526 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
527 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
528 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
529
530 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
531 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
532 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
533 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
534 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
535 const __m128i vxi19x89ABCDEF = _mm_cvtepi8_epi16(vi19x89ABCDEF);
536 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
537 const __m128i vxk19x89ABCDEF = _mm_cvtepi8_epi16(vk19x89ABCDEF);
538 i19 += 16;
539
540
541 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
542 vprod89ABCDEF = _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
543
544 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
545 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
546 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
547 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
548
549 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
550 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
551 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
552 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
553 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
554 const __m128i vxi20x89ABCDEF = _mm_cvtepi8_epi16(vi20x89ABCDEF);
555 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
556 const __m128i vxk20x89ABCDEF = _mm_cvtepi8_epi16(vk20x89ABCDEF);
557 i20 += 16;
558
559
560 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
561 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
562
563 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
564 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
565 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
566 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
567
568 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
569 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
570 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
571 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
572 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
573 const __m128i vxi21x89ABCDEF = _mm_cvtepi8_epi16(vi21x89ABCDEF);
574 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
575 const __m128i vxk21x89ABCDEF = _mm_cvtepi8_epi16(vk21x89ABCDEF);
576 i21 += 16;
577
578
579 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
580 vprod89ABCDEF = _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
581
582 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
583 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
584 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
585 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
586
587 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
588 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
589 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
590 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
591 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
592 const __m128i vxi22x89ABCDEF = _mm_cvtepi8_epi16(vi22x89ABCDEF);
593 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
594 const __m128i vxk22x89ABCDEF = _mm_cvtepi8_epi16(vk22x89ABCDEF);
595 i22 += 16;
596
597
598 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
599 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
600
601 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
602 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
603 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
604 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
605
606 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
607 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
608 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
609 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
610 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
611 const __m128i vxi23x89ABCDEF = _mm_cvtepi8_epi16(vi23x89ABCDEF);
612 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
613 const __m128i vxk23x89ABCDEF = _mm_cvtepi8_epi16(vk23x89ABCDEF);
614 i23 += 16;
615
616
617 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
618 vprod89ABCDEF = _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
619
620 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
621 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
622 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
623 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
624
625 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
626 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
627 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
628 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
629 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
630 const __m128i vxi24x89ABCDEF = _mm_cvtepi8_epi16(vi24x89ABCDEF);
631 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
632 const __m128i vxk24x89ABCDEF = _mm_cvtepi8_epi16(vk24x89ABCDEF);
633 i24 += 16;
634
635
636 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
637 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
638
639 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
640 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
641 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
642 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
643
644 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
645
646 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
647 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
648 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
649 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
650
651 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
652 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
653 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
654 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
655 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
656
657 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
658 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
659 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
660 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
661 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
662
663 vacc0123 = _mm_cvtps_epi32(vscaled0123);
664 vacc4567 = _mm_cvtps_epi32(vscaled4567);
665 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
666 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
667
668 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
669 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
670 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
671
672
673 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
674
675 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
676 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
677
678 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
679 output += 16;
680 }
681 if XNN_UNLIKELY(c != 0) {
682 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
683 do {
684 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
685 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
686
687
688 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
689 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
690 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
691 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
692 i0 += 8;
693
694
695 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
696
697 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
698 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
699
700 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
701 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
702 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
703 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
704 i1 += 8;
705
706
707 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
708
709 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
710 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
711
712 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
713 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
714 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
715 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
716 i2 += 8;
717
718
719 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
720
721 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
722 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
723
724 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
725 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
726 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
727 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
728 i3 += 8;
729
730
731 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
732
733 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
734 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
735
736 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
737 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
738 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
739 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
740 i4 += 8;
741
742
743 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
744
745 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
746 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
747
748 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
749 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
750 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
751 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
752 i5 += 8;
753
754
755 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
756
757 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
758 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
759
760 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
761 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
762 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
763 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
764 i6 += 8;
765
766
767 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
768
769 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
770 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
771
772 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
773 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
774 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
775 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
776 i7 += 8;
777
778
779 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
780
781 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
782 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
783
784 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
785 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
786 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
787 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
788 i8 += 8;
789
790
791 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
792
793 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
794 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
795
796 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
797 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
798 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
799 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
800 i9 += 8;
801
802
803 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
804
805 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
806 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
807
808 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
809 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
810 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
811 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
812 i10 += 8;
813
814
815 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
816
817 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
818 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
819
820 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
821 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
822 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
823 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
824 i11 += 8;
825
826
827 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
828
829 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
830 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
831
832 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
833 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
834 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
835 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
836 i12 += 8;
837
838
839 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
840
841 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
842 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
843
844 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
845 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
846 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
847 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
848 i13 += 8;
849
850
851 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
852
853 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
854 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
855
856 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
857 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
858 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
859 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
860 i14 += 8;
861
862
863 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
864
865 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
866 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
867
868 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
869 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
870 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
871 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
872 i15 += 8;
873
874
875 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
876
877 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
878 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
879
880 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
881 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
882 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
883 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
884 i16 += 8;
885
886
887 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
888
889 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
890 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
891
892 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
893 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
894 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
895 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
896 i17 += 8;
897
898
899 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
900
901 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
902 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
903
904 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
905 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
906 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
907 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
908 i18 += 8;
909
910
911 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
912
913 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
914 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
915
916 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
917 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
918 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
919 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
920 i19 += 8;
921
922
923 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
924
925 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
926 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
927
928 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
929 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
930 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
931 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
932 i20 += 8;
933
934
935 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
936
937 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
938 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
939
940 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
941 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
942 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
943 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
944 i21 += 8;
945
946
947 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
948
949 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
950 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
951
952 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
953 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
954 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
955 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
956 i22 += 8;
957
958
959 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
960
961 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
962 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
963
964 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
965 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
966 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
967 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
968 i23 += 8;
969
970
971 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
972
973 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
974 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
975
976 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
977 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
978 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
979 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
980 i24 += 8;
981
982
983 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
984
985 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
986 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
987
988 k += 8;
989
990 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
991 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
992
993 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
994 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
995 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
996
997 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
998 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
999 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1000
1001 vacc0123 = _mm_cvtps_epi32(vscaled0123);
1002 vacc4567 = _mm_cvtps_epi32(vscaled4567);
1003
1004 w = (const void*) ((const int32_t*) w + 8);
1005
1006 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
1007 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1008
1009
1010 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
1011
1012 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
1013
1014 if XNN_LIKELY(c >= 8) {
1015 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
1016 output += 8;
1017 c -= 8;
1018 } else {
1019 if (c & 4) {
1020 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
1021 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
1022 output += 4;
1023 }
1024 if (c & 2) {
1025 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
1026 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1027 output += 2;
1028 }
1029 if (c & 1) {
1030 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
1031 output += 1;
1032 }
1033 c = 0;
1034 }
1035 } while (c != 0);
1036 }
1037
1038 output = (int8_t*) ((uintptr_t) output + output_increment);
1039 } while (--output_width != 0);
1040 }
1041