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 #include <xnnpack/unaligned.h>
16
17
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)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 do {
34 const int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const int8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const int8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const int8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const int8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const int8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const int8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 const int8_t* i9 = input[9];
80 assert(i9 != NULL);
81 if XNN_UNPREDICTABLE(i9 != zero) {
82 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
83 }
84 const int8_t* i10 = input[10];
85 assert(i10 != NULL);
86 if XNN_UNPREDICTABLE(i10 != zero) {
87 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
88 }
89 const int8_t* i11 = input[11];
90 assert(i11 != NULL);
91 if XNN_UNPREDICTABLE(i11 != zero) {
92 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
93 }
94 const int8_t* i12 = input[12];
95 assert(i12 != NULL);
96 if XNN_UNPREDICTABLE(i12 != zero) {
97 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
98 }
99 const int8_t* i13 = input[13];
100 assert(i13 != NULL);
101 if XNN_UNPREDICTABLE(i13 != zero) {
102 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
103 }
104 const int8_t* i14 = input[14];
105 assert(i14 != NULL);
106 if XNN_UNPREDICTABLE(i14 != zero) {
107 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
108 }
109 const int8_t* i15 = input[15];
110 assert(i15 != NULL);
111 if XNN_UNPREDICTABLE(i15 != zero) {
112 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
113 }
114 const int8_t* i16 = input[16];
115 assert(i16 != NULL);
116 if XNN_UNPREDICTABLE(i16 != zero) {
117 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
118 }
119 const int8_t* i17 = input[17];
120 assert(i17 != NULL);
121 if XNN_UNPREDICTABLE(i17 != zero) {
122 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
123 }
124 const int8_t* i18 = input[18];
125 assert(i18 != NULL);
126 if XNN_UNPREDICTABLE(i18 != zero) {
127 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
128 }
129 const int8_t* i19 = input[19];
130 assert(i19 != NULL);
131 if XNN_UNPREDICTABLE(i19 != zero) {
132 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
133 }
134 const int8_t* i20 = input[20];
135 assert(i20 != NULL);
136 if XNN_UNPREDICTABLE(i20 != zero) {
137 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
138 }
139 const int8_t* i21 = input[21];
140 assert(i21 != NULL);
141 if XNN_UNPREDICTABLE(i21 != zero) {
142 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
143 }
144 const int8_t* i22 = input[22];
145 assert(i22 != NULL);
146 if XNN_UNPREDICTABLE(i22 != zero) {
147 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
148 }
149 const int8_t* i23 = input[23];
150 assert(i23 != NULL);
151 if XNN_UNPREDICTABLE(i23 != zero) {
152 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
153 }
154 const int8_t* i24 = input[24];
155 assert(i24 != NULL);
156 if XNN_UNPREDICTABLE(i24 != zero) {
157 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
158 }
159 input = (const int8_t**) ((uintptr_t) input + input_stride);
160
161 size_t c = channels;
162 const void* w = weights;
163 for (; c >= 16; c -= 16) {
164 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
165 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
166 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
167 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
168
169
170 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
171 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
172 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
173 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
174 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
175 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
176 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
177 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
178 i0 += 16;
179
180
181 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
182 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
183
184 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
185 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
186 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
187 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
188
189 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
190 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
191 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
192 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
193 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
194 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
195 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
196 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
197 i1 += 16;
198
199
200 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
201 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
202
203 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
204 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
205 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
206 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
207
208 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
209 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
210 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
211 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
212 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
213 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
214 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
215 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
216 i2 += 16;
217
218
219 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
220 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
221
222 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
223 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
224 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
225 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
226
227 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
228 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
229 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
230 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
231 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
232 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
233 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
234 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
235 i3 += 16;
236
237
238 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
239 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
240
241 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
242 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
243 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
244 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
245
246 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
247 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
248 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
249 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
250 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
251 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
252 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
253 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
254 i4 += 16;
255
256
257 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
258 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
259
260 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
261 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
262 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
263 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
264
265 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
266 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
267 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
268 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
269 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
270 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
271 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
272 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
273 i5 += 16;
274
275
276 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
277 vprod89ABCDEF = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
278
279 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
280 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
281 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
282 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
283
284 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
285 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
286 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
287 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
288 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
289 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
290 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
291 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
292 i6 += 16;
293
294
295 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
296 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
297
298 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
299 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
300 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
301 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
302
303 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
304 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
305 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
306 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
307 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
308 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
309 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
310 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
311 i7 += 16;
312
313
314 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
315 vprod89ABCDEF = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
316
317 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
318 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
319 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
320 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
321
322 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
323 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
324 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
325 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
326 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
327 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
328 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
329 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
330 i8 += 16;
331
332
333 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
334 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
335
336 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
337 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
338 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
339 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
340
341 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
342 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
343 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
344 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
345 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
346 const __m128i vxi9x89ABCDEF = _mm_cvtepi8_epi16(vi9x89ABCDEF);
347 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
348 const __m128i vxk9x89ABCDEF = _mm_cvtepi8_epi16(vk9x89ABCDEF);
349 i9 += 16;
350
351
352 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
353 vprod89ABCDEF = _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
354
355 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
356 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
357 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
358 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
359
360 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
361 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
362 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
363 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
364 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
365 const __m128i vxi10x89ABCDEF = _mm_cvtepi8_epi16(vi10x89ABCDEF);
366 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
367 const __m128i vxk10x89ABCDEF = _mm_cvtepi8_epi16(vk10x89ABCDEF);
368 i10 += 16;
369
370
371 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
372 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
373
374 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
375 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
376 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
377 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
378
379 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
380 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
381 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
382 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
383 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
384 const __m128i vxi11x89ABCDEF = _mm_cvtepi8_epi16(vi11x89ABCDEF);
385 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
386 const __m128i vxk11x89ABCDEF = _mm_cvtepi8_epi16(vk11x89ABCDEF);
387 i11 += 16;
388
389
390 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
391 vprod89ABCDEF = _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
392
393 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
394 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
395 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
396 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
397
398 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
399 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
400 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
401 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
402 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
403 const __m128i vxi12x89ABCDEF = _mm_cvtepi8_epi16(vi12x89ABCDEF);
404 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
405 const __m128i vxk12x89ABCDEF = _mm_cvtepi8_epi16(vk12x89ABCDEF);
406 i12 += 16;
407
408
409 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
410 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
411
412 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
413 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
414 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
415 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
416
417 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
418 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
419 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
420 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
421 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
422 const __m128i vxi13x89ABCDEF = _mm_cvtepi8_epi16(vi13x89ABCDEF);
423 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
424 const __m128i vxk13x89ABCDEF = _mm_cvtepi8_epi16(vk13x89ABCDEF);
425 i13 += 16;
426
427
428 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
429 vprod89ABCDEF = _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
430
431 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
432 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
433 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
434 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
435
436 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
437 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
438 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
439 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
440 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
441 const __m128i vxi14x89ABCDEF = _mm_cvtepi8_epi16(vi14x89ABCDEF);
442 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
443 const __m128i vxk14x89ABCDEF = _mm_cvtepi8_epi16(vk14x89ABCDEF);
444 i14 += 16;
445
446
447 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
448 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
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
455 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
456 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
457 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
458 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
459 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
460 const __m128i vxi15x89ABCDEF = _mm_cvtepi8_epi16(vi15x89ABCDEF);
461 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
462 const __m128i vxk15x89ABCDEF = _mm_cvtepi8_epi16(vk15x89ABCDEF);
463 i15 += 16;
464
465
466 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
467 vprod89ABCDEF = _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
468
469 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
470 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
471 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
472 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
473
474 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
475 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
476 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
477 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
478 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
479 const __m128i vxi16x89ABCDEF = _mm_cvtepi8_epi16(vi16x89ABCDEF);
480 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
481 const __m128i vxk16x89ABCDEF = _mm_cvtepi8_epi16(vk16x89ABCDEF);
482 i16 += 16;
483
484
485 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
486 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
487
488 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
489 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
490 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
491 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
492
493 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
494 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
495 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
496 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
497 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
498 const __m128i vxi17x89ABCDEF = _mm_cvtepi8_epi16(vi17x89ABCDEF);
499 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
500 const __m128i vxk17x89ABCDEF = _mm_cvtepi8_epi16(vk17x89ABCDEF);
501 i17 += 16;
502
503
504 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
505 vprod89ABCDEF = _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
506
507 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
508 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
509 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
510 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
511
512 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
513 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
514 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
515 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
516 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
517 const __m128i vxi18x89ABCDEF = _mm_cvtepi8_epi16(vi18x89ABCDEF);
518 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
519 const __m128i vxk18x89ABCDEF = _mm_cvtepi8_epi16(vk18x89ABCDEF);
520 i18 += 16;
521
522
523 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
524 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
525
526 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
527 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
528 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
529 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
530
531 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
532 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
533 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
534 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
535 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
536 const __m128i vxi19x89ABCDEF = _mm_cvtepi8_epi16(vi19x89ABCDEF);
537 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
538 const __m128i vxk19x89ABCDEF = _mm_cvtepi8_epi16(vk19x89ABCDEF);
539 i19 += 16;
540
541
542 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
543 vprod89ABCDEF = _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
544
545 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
546 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
547 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
548 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
549
550 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
551 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
552 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
553 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
554 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
555 const __m128i vxi20x89ABCDEF = _mm_cvtepi8_epi16(vi20x89ABCDEF);
556 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
557 const __m128i vxk20x89ABCDEF = _mm_cvtepi8_epi16(vk20x89ABCDEF);
558 i20 += 16;
559
560
561 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
562 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
563
564 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
565 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
566 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
567 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
568
569 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
570 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
571 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
572 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
573 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
574 const __m128i vxi21x89ABCDEF = _mm_cvtepi8_epi16(vi21x89ABCDEF);
575 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
576 const __m128i vxk21x89ABCDEF = _mm_cvtepi8_epi16(vk21x89ABCDEF);
577 i21 += 16;
578
579
580 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
581 vprod89ABCDEF = _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
582
583 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
584 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
585 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
586 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
587
588 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
589 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
590 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
591 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
592 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
593 const __m128i vxi22x89ABCDEF = _mm_cvtepi8_epi16(vi22x89ABCDEF);
594 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
595 const __m128i vxk22x89ABCDEF = _mm_cvtepi8_epi16(vk22x89ABCDEF);
596 i22 += 16;
597
598
599 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
600 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
601
602 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
603 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
604 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
605 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
606
607 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
608 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
609 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
610 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
611 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
612 const __m128i vxi23x89ABCDEF = _mm_cvtepi8_epi16(vi23x89ABCDEF);
613 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
614 const __m128i vxk23x89ABCDEF = _mm_cvtepi8_epi16(vk23x89ABCDEF);
615 i23 += 16;
616
617
618 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
619 vprod89ABCDEF = _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
620
621 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
622 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
623 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
624 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
625
626 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
627 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
628 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
629 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
630 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
631 const __m128i vxi24x89ABCDEF = _mm_cvtepi8_epi16(vi24x89ABCDEF);
632 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
633 const __m128i vxk24x89ABCDEF = _mm_cvtepi8_epi16(vk24x89ABCDEF);
634 i24 += 16;
635
636
637 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
638 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
639
640 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
641 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
642 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
643 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
644
645 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
646
647 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
648 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
649 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
650 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
651
652 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
653 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
654 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
655 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
656 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
657
658 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
659 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
660 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
661 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
662 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
663
664 vacc0123 = _mm_cvtps_epi32(vscaled0123);
665 vacc4567 = _mm_cvtps_epi32(vscaled4567);
666 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
667 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
668
669 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
670 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
671 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
672
673
674 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
675
676 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
677 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
678
679 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
680 output += 16;
681 }
682 if XNN_UNLIKELY(c != 0) {
683 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
684 do {
685 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
686 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
687
688
689 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
690 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
691 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
692 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
693 i0 += 8;
694
695
696 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
697
698 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
699 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
700
701 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
702 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
703 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
704 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
705 i1 += 8;
706
707
708 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
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
713 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
714 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
715 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
716 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
717 i2 += 8;
718
719
720 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
721
722 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
723 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
724
725 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
726 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
727 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
728 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
729 i3 += 8;
730
731
732 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
733
734 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
735 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
736
737 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
738 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
739 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
740 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
741 i4 += 8;
742
743
744 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
745
746 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
747 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
748
749 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
750 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
751 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
752 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
753 i5 += 8;
754
755
756 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
757
758 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
759 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
760
761 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
762 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
763 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
764 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
765 i6 += 8;
766
767
768 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
769
770 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
771 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
772
773 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
774 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
775 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
776 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
777 i7 += 8;
778
779
780 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
781
782 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
783 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
784
785 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
786 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
787 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
788 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
789 i8 += 8;
790
791
792 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
793
794 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
795 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
796
797 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
798 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
799 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
800 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
801 i9 += 8;
802
803
804 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
805
806 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
807 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
808
809 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
810 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
811 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
812 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
813 i10 += 8;
814
815
816 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
817
818 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
819 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
820
821 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
822 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
823 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
824 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
825 i11 += 8;
826
827
828 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
829
830 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
831 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
832
833 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
834 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
835 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
836 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
837 i12 += 8;
838
839
840 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
841
842 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
843 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
844
845 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
846 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
847 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
848 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
849 i13 += 8;
850
851
852 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
853
854 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
855 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
856
857 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
858 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
859 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
860 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
861 i14 += 8;
862
863
864 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
865
866 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
867 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
868
869 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
870 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
871 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
872 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
873 i15 += 8;
874
875
876 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
877
878 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
879 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
880
881 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
882 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
883 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
884 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
885 i16 += 8;
886
887
888 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
889
890 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
891 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
892
893 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
894 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
895 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
896 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
897 i17 += 8;
898
899
900 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
901
902 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
903 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
904
905 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
906 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
907 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
908 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
909 i18 += 8;
910
911
912 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
913
914 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
915 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
916
917 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
918 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
919 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
920 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
921 i19 += 8;
922
923
924 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
925
926 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
927 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
928
929 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
930 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
931 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
932 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
933 i20 += 8;
934
935
936 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
937
938 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
939 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
940
941 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
942 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
943 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
944 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
945 i21 += 8;
946
947
948 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
949
950 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
951 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
952
953 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
954 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
955 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
956 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
957 i22 += 8;
958
959
960 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
961
962 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
963 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
964
965 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
966 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
967 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
968 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
969 i23 += 8;
970
971
972 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
973
974 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
975 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
976
977 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
978 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
979 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
980 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
981 i24 += 8;
982
983
984 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
985
986 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
987 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
988
989 k += 8;
990
991 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
992 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
993
994 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
995 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
996 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
997
998 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
999 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
1000 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1001
1002 vacc0123 = _mm_cvtps_epi32(vscaled0123);
1003 vacc4567 = _mm_cvtps_epi32(vscaled4567);
1004
1005 w = (const void*) ((const int32_t*) w + 8);
1006
1007 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
1008 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1009
1010
1011 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
1012
1013 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
1014
1015 if XNN_LIKELY(c >= 8) {
1016 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
1017 output += 8;
1018 c -= 8;
1019 } else {
1020 if (c & 4) {
1021 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
1022 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
1023 output += 4;
1024 }
1025 if (c & 2) {
1026 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
1027 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1028 output += 2;
1029 }
1030 if (c & 1) {
1031 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
1032 output += 1;
1033 }
1034 c = 0;
1035 }
1036 } while (c != 0);
1037 }
1038
1039 output = (int8_t*) ((uintptr_t) output + output_increment);
1040 } while (--output_width != 0);
1041 }
1042