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_up24x25__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_up24x25__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 >= 24; c -= 24) {
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 __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
169 __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
170
171
172 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
173 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
174 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)));
175 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
176 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
177 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
178 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)));
179 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
180 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
181 const __m128i vxi0xGHIJKLMN = _mm_cvtepi8_epi16(vi0xGHIJKLMN);
182 const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)));
183 const __m128i vxk0xGHIJKLMN = _mm_cvtepi8_epi16(vk0xGHIJKLMN);
184 i0 += 24;
185
186
187 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
188 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
189 __m128i vprodGHIJKLMN = _mm_mullo_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN);
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
193 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
194 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
195 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
196 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
197
198 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
199 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
200 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)));
201 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
202 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
203 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
204 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)));
205 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
206 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
207 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(vi1xGHIJKLMN);
208 const __m128i vk1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)));
209 const __m128i vxk1xGHIJKLMN = _mm_cvtepi8_epi16(vk1xGHIJKLMN);
210 i1 += 24;
211
212
213 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
214 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
215 vprodGHIJKLMN = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN);
216
217 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
218 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
219 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
220 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
221 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
222 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
223
224 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
225 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
226 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)));
227 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
228 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
229 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
230 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)));
231 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
232 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
233 const __m128i vxi2xGHIJKLMN = _mm_cvtepi8_epi16(vi2xGHIJKLMN);
234 const __m128i vk2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)));
235 const __m128i vxk2xGHIJKLMN = _mm_cvtepi8_epi16(vk2xGHIJKLMN);
236 i2 += 24;
237
238
239 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
240 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
241 vprodGHIJKLMN = _mm_mullo_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN);
242
243 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
244 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
245 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
246 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
247 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
248 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
249
250 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
251 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
252 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)));
253 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
254 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
255 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
256 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)));
257 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
258 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
259 const __m128i vxi3xGHIJKLMN = _mm_cvtepi8_epi16(vi3xGHIJKLMN);
260 const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)));
261 const __m128i vxk3xGHIJKLMN = _mm_cvtepi8_epi16(vk3xGHIJKLMN);
262 i3 += 24;
263
264
265 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
266 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
267 vprodGHIJKLMN = _mm_mullo_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN);
268
269 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
270 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
271 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
272 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
273 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
274 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
275
276 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
277 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
278 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)));
279 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
280 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
281 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
282 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)));
283 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
284 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
285 const __m128i vxi4xGHIJKLMN = _mm_cvtepi8_epi16(vi4xGHIJKLMN);
286 const __m128i vk4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)));
287 const __m128i vxk4xGHIJKLMN = _mm_cvtepi8_epi16(vk4xGHIJKLMN);
288 i4 += 24;
289
290
291 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
292 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
293 vprodGHIJKLMN = _mm_mullo_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN);
294
295 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
296 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
297 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
298 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
299 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
300 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
301
302 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
303 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
304 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)));
305 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
306 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
307 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
308 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)));
309 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
310 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
311 const __m128i vxi5xGHIJKLMN = _mm_cvtepi8_epi16(vi5xGHIJKLMN);
312 const __m128i vk5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)));
313 const __m128i vxk5xGHIJKLMN = _mm_cvtepi8_epi16(vk5xGHIJKLMN);
314 i5 += 24;
315
316
317 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
318 vprod89ABCDEF = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
319 vprodGHIJKLMN = _mm_mullo_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN);
320
321 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
322 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
323 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
324 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
325 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
326 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
327
328 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
329 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
330 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)));
331 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
332 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
333 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
334 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)));
335 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
336 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
337 const __m128i vxi6xGHIJKLMN = _mm_cvtepi8_epi16(vi6xGHIJKLMN);
338 const __m128i vk6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)));
339 const __m128i vxk6xGHIJKLMN = _mm_cvtepi8_epi16(vk6xGHIJKLMN);
340 i6 += 24;
341
342
343 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
344 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
345 vprodGHIJKLMN = _mm_mullo_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN);
346
347 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
348 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
349 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
350 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
351 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
352 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
353
354 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
355 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
356 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)));
357 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
358 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
359 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
360 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)));
361 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
362 const __m128i vi7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i7 + 16));
363 const __m128i vxi7xGHIJKLMN = _mm_cvtepi8_epi16(vi7xGHIJKLMN);
364 const __m128i vk7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)));
365 const __m128i vxk7xGHIJKLMN = _mm_cvtepi8_epi16(vk7xGHIJKLMN);
366 i7 += 24;
367
368
369 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
370 vprod89ABCDEF = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
371 vprodGHIJKLMN = _mm_mullo_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN);
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 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
378 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
379
380 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
381 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
382 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)));
383 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
384 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
385 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
386 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)));
387 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
388 const __m128i vi8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i8 + 16));
389 const __m128i vxi8xGHIJKLMN = _mm_cvtepi8_epi16(vi8xGHIJKLMN);
390 const __m128i vk8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)));
391 const __m128i vxk8xGHIJKLMN = _mm_cvtepi8_epi16(vk8xGHIJKLMN);
392 i8 += 24;
393
394
395 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
396 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
397 vprodGHIJKLMN = _mm_mullo_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN);
398
399 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
400 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
401 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
402 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
403 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
404 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
405
406 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
407 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
408 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)));
409 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
410 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
411 const __m128i vxi9x89ABCDEF = _mm_cvtepi8_epi16(vi9x89ABCDEF);
412 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 224 * sizeof(int8_t)));
413 const __m128i vxk9x89ABCDEF = _mm_cvtepi8_epi16(vk9x89ABCDEF);
414 const __m128i vi9xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i9 + 16));
415 const __m128i vxi9xGHIJKLMN = _mm_cvtepi8_epi16(vi9xGHIJKLMN);
416 const __m128i vk9xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 232 * sizeof(int8_t)));
417 const __m128i vxk9xGHIJKLMN = _mm_cvtepi8_epi16(vk9xGHIJKLMN);
418 i9 += 24;
419
420
421 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
422 vprod89ABCDEF = _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
423 vprodGHIJKLMN = _mm_mullo_epi16(vxi9xGHIJKLMN, vxk9xGHIJKLMN);
424
425 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
426 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
427 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
428 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
429 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
430 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
431
432 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
433 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
434 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 240 * sizeof(int8_t)));
435 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
436 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
437 const __m128i vxi10x89ABCDEF = _mm_cvtepi8_epi16(vi10x89ABCDEF);
438 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 248 * sizeof(int8_t)));
439 const __m128i vxk10x89ABCDEF = _mm_cvtepi8_epi16(vk10x89ABCDEF);
440 const __m128i vi10xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i10 + 16));
441 const __m128i vxi10xGHIJKLMN = _mm_cvtepi8_epi16(vi10xGHIJKLMN);
442 const __m128i vk10xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 256 * sizeof(int8_t)));
443 const __m128i vxk10xGHIJKLMN = _mm_cvtepi8_epi16(vk10xGHIJKLMN);
444 i10 += 24;
445
446
447 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
448 vprod89ABCDEF = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
449 vprodGHIJKLMN = _mm_mullo_epi16(vxi10xGHIJKLMN, vxk10xGHIJKLMN);
450
451 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
452 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
453 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
454 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
455 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
456 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
457
458 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
459 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
460 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 264 * sizeof(int8_t)));
461 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
462 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
463 const __m128i vxi11x89ABCDEF = _mm_cvtepi8_epi16(vi11x89ABCDEF);
464 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 272 * sizeof(int8_t)));
465 const __m128i vxk11x89ABCDEF = _mm_cvtepi8_epi16(vk11x89ABCDEF);
466 const __m128i vi11xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i11 + 16));
467 const __m128i vxi11xGHIJKLMN = _mm_cvtepi8_epi16(vi11xGHIJKLMN);
468 const __m128i vk11xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 280 * sizeof(int8_t)));
469 const __m128i vxk11xGHIJKLMN = _mm_cvtepi8_epi16(vk11xGHIJKLMN);
470 i11 += 24;
471
472
473 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
474 vprod89ABCDEF = _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
475 vprodGHIJKLMN = _mm_mullo_epi16(vxi11xGHIJKLMN, vxk11xGHIJKLMN);
476
477 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
478 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
479 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
480 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
481 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
482 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
483
484 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
485 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
486 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 288 * sizeof(int8_t)));
487 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
488 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
489 const __m128i vxi12x89ABCDEF = _mm_cvtepi8_epi16(vi12x89ABCDEF);
490 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 296 * sizeof(int8_t)));
491 const __m128i vxk12x89ABCDEF = _mm_cvtepi8_epi16(vk12x89ABCDEF);
492 const __m128i vi12xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i12 + 16));
493 const __m128i vxi12xGHIJKLMN = _mm_cvtepi8_epi16(vi12xGHIJKLMN);
494 const __m128i vk12xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 304 * sizeof(int8_t)));
495 const __m128i vxk12xGHIJKLMN = _mm_cvtepi8_epi16(vk12xGHIJKLMN);
496 i12 += 24;
497
498
499 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
500 vprod89ABCDEF = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
501 vprodGHIJKLMN = _mm_mullo_epi16(vxi12xGHIJKLMN, vxk12xGHIJKLMN);
502
503 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
504 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
505 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
506 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
507 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
508 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
509
510 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
511 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
512 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 312 * sizeof(int8_t)));
513 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
514 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
515 const __m128i vxi13x89ABCDEF = _mm_cvtepi8_epi16(vi13x89ABCDEF);
516 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 320 * sizeof(int8_t)));
517 const __m128i vxk13x89ABCDEF = _mm_cvtepi8_epi16(vk13x89ABCDEF);
518 const __m128i vi13xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i13 + 16));
519 const __m128i vxi13xGHIJKLMN = _mm_cvtepi8_epi16(vi13xGHIJKLMN);
520 const __m128i vk13xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 328 * sizeof(int8_t)));
521 const __m128i vxk13xGHIJKLMN = _mm_cvtepi8_epi16(vk13xGHIJKLMN);
522 i13 += 24;
523
524
525 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
526 vprod89ABCDEF = _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
527 vprodGHIJKLMN = _mm_mullo_epi16(vxi13xGHIJKLMN, vxk13xGHIJKLMN);
528
529 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
530 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
531 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
532 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
533 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
534 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
535
536 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
537 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
538 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 336 * sizeof(int8_t)));
539 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
540 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
541 const __m128i vxi14x89ABCDEF = _mm_cvtepi8_epi16(vi14x89ABCDEF);
542 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 344 * sizeof(int8_t)));
543 const __m128i vxk14x89ABCDEF = _mm_cvtepi8_epi16(vk14x89ABCDEF);
544 const __m128i vi14xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i14 + 16));
545 const __m128i vxi14xGHIJKLMN = _mm_cvtepi8_epi16(vi14xGHIJKLMN);
546 const __m128i vk14xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 352 * sizeof(int8_t)));
547 const __m128i vxk14xGHIJKLMN = _mm_cvtepi8_epi16(vk14xGHIJKLMN);
548 i14 += 24;
549
550
551 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
552 vprod89ABCDEF = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
553 vprodGHIJKLMN = _mm_mullo_epi16(vxi14xGHIJKLMN, vxk14xGHIJKLMN);
554
555 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
556 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
557 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
558 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
559 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
560 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
561
562 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
563 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
564 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 360 * sizeof(int8_t)));
565 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
566 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
567 const __m128i vxi15x89ABCDEF = _mm_cvtepi8_epi16(vi15x89ABCDEF);
568 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 368 * sizeof(int8_t)));
569 const __m128i vxk15x89ABCDEF = _mm_cvtepi8_epi16(vk15x89ABCDEF);
570 const __m128i vi15xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i15 + 16));
571 const __m128i vxi15xGHIJKLMN = _mm_cvtepi8_epi16(vi15xGHIJKLMN);
572 const __m128i vk15xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 376 * sizeof(int8_t)));
573 const __m128i vxk15xGHIJKLMN = _mm_cvtepi8_epi16(vk15xGHIJKLMN);
574 i15 += 24;
575
576
577 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
578 vprod89ABCDEF = _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
579 vprodGHIJKLMN = _mm_mullo_epi16(vxi15xGHIJKLMN, vxk15xGHIJKLMN);
580
581 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
582 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
583 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
584 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
585 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
586 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
587
588 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
589 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
590 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 384 * sizeof(int8_t)));
591 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
592 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
593 const __m128i vxi16x89ABCDEF = _mm_cvtepi8_epi16(vi16x89ABCDEF);
594 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 392 * sizeof(int8_t)));
595 const __m128i vxk16x89ABCDEF = _mm_cvtepi8_epi16(vk16x89ABCDEF);
596 const __m128i vi16xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i16 + 16));
597 const __m128i vxi16xGHIJKLMN = _mm_cvtepi8_epi16(vi16xGHIJKLMN);
598 const __m128i vk16xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 400 * sizeof(int8_t)));
599 const __m128i vxk16xGHIJKLMN = _mm_cvtepi8_epi16(vk16xGHIJKLMN);
600 i16 += 24;
601
602
603 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
604 vprod89ABCDEF = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
605 vprodGHIJKLMN = _mm_mullo_epi16(vxi16xGHIJKLMN, vxk16xGHIJKLMN);
606
607 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
608 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
609 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
610 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
611 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
612 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
613
614 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
615 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
616 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 408 * sizeof(int8_t)));
617 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
618 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
619 const __m128i vxi17x89ABCDEF = _mm_cvtepi8_epi16(vi17x89ABCDEF);
620 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 416 * sizeof(int8_t)));
621 const __m128i vxk17x89ABCDEF = _mm_cvtepi8_epi16(vk17x89ABCDEF);
622 const __m128i vi17xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i17 + 16));
623 const __m128i vxi17xGHIJKLMN = _mm_cvtepi8_epi16(vi17xGHIJKLMN);
624 const __m128i vk17xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 424 * sizeof(int8_t)));
625 const __m128i vxk17xGHIJKLMN = _mm_cvtepi8_epi16(vk17xGHIJKLMN);
626 i17 += 24;
627
628
629 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
630 vprod89ABCDEF = _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
631 vprodGHIJKLMN = _mm_mullo_epi16(vxi17xGHIJKLMN, vxk17xGHIJKLMN);
632
633 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
634 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
635 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
636 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
637 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
638 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
639
640 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
641 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
642 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 432 * sizeof(int8_t)));
643 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
644 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
645 const __m128i vxi18x89ABCDEF = _mm_cvtepi8_epi16(vi18x89ABCDEF);
646 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 440 * sizeof(int8_t)));
647 const __m128i vxk18x89ABCDEF = _mm_cvtepi8_epi16(vk18x89ABCDEF);
648 const __m128i vi18xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i18 + 16));
649 const __m128i vxi18xGHIJKLMN = _mm_cvtepi8_epi16(vi18xGHIJKLMN);
650 const __m128i vk18xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 448 * sizeof(int8_t)));
651 const __m128i vxk18xGHIJKLMN = _mm_cvtepi8_epi16(vk18xGHIJKLMN);
652 i18 += 24;
653
654
655 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
656 vprod89ABCDEF = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
657 vprodGHIJKLMN = _mm_mullo_epi16(vxi18xGHIJKLMN, vxk18xGHIJKLMN);
658
659 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
660 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
661 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
662 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
663 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
664 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
665
666 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
667 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
668 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 456 * sizeof(int8_t)));
669 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
670 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
671 const __m128i vxi19x89ABCDEF = _mm_cvtepi8_epi16(vi19x89ABCDEF);
672 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 464 * sizeof(int8_t)));
673 const __m128i vxk19x89ABCDEF = _mm_cvtepi8_epi16(vk19x89ABCDEF);
674 const __m128i vi19xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i19 + 16));
675 const __m128i vxi19xGHIJKLMN = _mm_cvtepi8_epi16(vi19xGHIJKLMN);
676 const __m128i vk19xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 472 * sizeof(int8_t)));
677 const __m128i vxk19xGHIJKLMN = _mm_cvtepi8_epi16(vk19xGHIJKLMN);
678 i19 += 24;
679
680
681 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
682 vprod89ABCDEF = _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
683 vprodGHIJKLMN = _mm_mullo_epi16(vxi19xGHIJKLMN, vxk19xGHIJKLMN);
684
685 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
686 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
687 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
688 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
689 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
690 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
691
692 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
693 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
694 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 480 * sizeof(int8_t)));
695 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
696 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
697 const __m128i vxi20x89ABCDEF = _mm_cvtepi8_epi16(vi20x89ABCDEF);
698 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 488 * sizeof(int8_t)));
699 const __m128i vxk20x89ABCDEF = _mm_cvtepi8_epi16(vk20x89ABCDEF);
700 const __m128i vi20xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i20 + 16));
701 const __m128i vxi20xGHIJKLMN = _mm_cvtepi8_epi16(vi20xGHIJKLMN);
702 const __m128i vk20xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 496 * sizeof(int8_t)));
703 const __m128i vxk20xGHIJKLMN = _mm_cvtepi8_epi16(vk20xGHIJKLMN);
704 i20 += 24;
705
706
707 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
708 vprod89ABCDEF = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
709 vprodGHIJKLMN = _mm_mullo_epi16(vxi20xGHIJKLMN, vxk20xGHIJKLMN);
710
711 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
712 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
713 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
714 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
715 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
716 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
717
718 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
719 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
720 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 504 * sizeof(int8_t)));
721 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
722 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
723 const __m128i vxi21x89ABCDEF = _mm_cvtepi8_epi16(vi21x89ABCDEF);
724 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 512 * sizeof(int8_t)));
725 const __m128i vxk21x89ABCDEF = _mm_cvtepi8_epi16(vk21x89ABCDEF);
726 const __m128i vi21xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i21 + 16));
727 const __m128i vxi21xGHIJKLMN = _mm_cvtepi8_epi16(vi21xGHIJKLMN);
728 const __m128i vk21xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 520 * sizeof(int8_t)));
729 const __m128i vxk21xGHIJKLMN = _mm_cvtepi8_epi16(vk21xGHIJKLMN);
730 i21 += 24;
731
732
733 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
734 vprod89ABCDEF = _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
735 vprodGHIJKLMN = _mm_mullo_epi16(vxi21xGHIJKLMN, vxk21xGHIJKLMN);
736
737 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
738 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
739 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
740 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
741 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
742 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
743
744 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
745 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
746 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 528 * sizeof(int8_t)));
747 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
748 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
749 const __m128i vxi22x89ABCDEF = _mm_cvtepi8_epi16(vi22x89ABCDEF);
750 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 536 * sizeof(int8_t)));
751 const __m128i vxk22x89ABCDEF = _mm_cvtepi8_epi16(vk22x89ABCDEF);
752 const __m128i vi22xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i22 + 16));
753 const __m128i vxi22xGHIJKLMN = _mm_cvtepi8_epi16(vi22xGHIJKLMN);
754 const __m128i vk22xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 544 * sizeof(int8_t)));
755 const __m128i vxk22xGHIJKLMN = _mm_cvtepi8_epi16(vk22xGHIJKLMN);
756 i22 += 24;
757
758
759 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
760 vprod89ABCDEF = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
761 vprodGHIJKLMN = _mm_mullo_epi16(vxi22xGHIJKLMN, vxk22xGHIJKLMN);
762
763 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
764 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
765 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
766 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
767 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
768 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
769
770 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
771 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
772 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 552 * sizeof(int8_t)));
773 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
774 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
775 const __m128i vxi23x89ABCDEF = _mm_cvtepi8_epi16(vi23x89ABCDEF);
776 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 560 * sizeof(int8_t)));
777 const __m128i vxk23x89ABCDEF = _mm_cvtepi8_epi16(vk23x89ABCDEF);
778 const __m128i vi23xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i23 + 16));
779 const __m128i vxi23xGHIJKLMN = _mm_cvtepi8_epi16(vi23xGHIJKLMN);
780 const __m128i vk23xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 568 * sizeof(int8_t)));
781 const __m128i vxk23xGHIJKLMN = _mm_cvtepi8_epi16(vk23xGHIJKLMN);
782 i23 += 24;
783
784
785 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
786 vprod89ABCDEF = _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
787 vprodGHIJKLMN = _mm_mullo_epi16(vxi23xGHIJKLMN, vxk23xGHIJKLMN);
788
789 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
790 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
791 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
792 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
793 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
794 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
795
796 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
797 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
798 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 576 * sizeof(int8_t)));
799 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
800 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
801 const __m128i vxi24x89ABCDEF = _mm_cvtepi8_epi16(vi24x89ABCDEF);
802 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 584 * sizeof(int8_t)));
803 const __m128i vxk24x89ABCDEF = _mm_cvtepi8_epi16(vk24x89ABCDEF);
804 const __m128i vi24xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i24 + 16));
805 const __m128i vxi24xGHIJKLMN = _mm_cvtepi8_epi16(vi24xGHIJKLMN);
806 const __m128i vk24xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 592 * sizeof(int8_t)));
807 const __m128i vxk24xGHIJKLMN = _mm_cvtepi8_epi16(vk24xGHIJKLMN);
808 i24 += 24;
809
810
811 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
812 vprod89ABCDEF = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
813 vprodGHIJKLMN = _mm_mullo_epi16(vxi24xGHIJKLMN, vxk24xGHIJKLMN);
814
815 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
816 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
817 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
818 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
819 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
820 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
821
822 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t));
823
824 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
825 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
826 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
827 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
828 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
829 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
830
831 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
832 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
833 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
834 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
835 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
836 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscale);
837 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscale);
838
839 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
840 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
841 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
842 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
843 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
844 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
845 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
846
847 vacc0123 = _mm_cvtps_epi32(vscaled0123);
848 vacc4567 = _mm_cvtps_epi32(vscaled4567);
849 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
850 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
851 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
852 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
853
854 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
855 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
856 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
857 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
858
859
860 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
861 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
862
863 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
864 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
865 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
866
867 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
868 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
869 output += 24;
870 }
871 if XNN_UNLIKELY(c != 0) {
872 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
873 do {
874 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
875 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
876
877
878 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
879 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
880 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
881 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
882 i0 += 8;
883
884
885 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
886
887 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
888 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
889
890 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
891 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
892 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 24));
893 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
894 i1 += 8;
895
896
897 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
898
899 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
900 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
901
902 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
903 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
904 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
905 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
906 i2 += 8;
907
908
909 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
910
911 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
912 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
913
914 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
915 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
916 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 72));
917 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
918 i3 += 8;
919
920
921 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
922
923 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
924 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
925
926 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
927 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
928 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
929 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
930 i4 += 8;
931
932
933 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
934
935 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
936 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
937
938 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
939 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
940 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 120));
941 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
942 i5 += 8;
943
944
945 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
946
947 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
948 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
949
950 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
951 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
952 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
953 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
954 i6 += 8;
955
956
957 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
958
959 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
960 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
961
962 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
963 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
964 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 168));
965 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
966 i7 += 8;
967
968
969 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
970
971 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
972 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
973
974 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
975 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
976 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
977 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
978 i8 += 8;
979
980
981 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
982
983 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
984 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
985
986 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
987 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
988 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 216));
989 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
990 i9 += 8;
991
992
993 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
994
995 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
996 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
997
998 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
999 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
1000 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
1001 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
1002 i10 += 8;
1003
1004
1005 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
1006
1007 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1008 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1009
1010 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
1011 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
1012 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 264));
1013 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
1014 i11 += 8;
1015
1016
1017 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
1018
1019 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1020 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1021
1022 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
1023 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
1024 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
1025 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
1026 i12 += 8;
1027
1028
1029 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
1030
1031 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1032 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1033
1034 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
1035 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
1036 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 312));
1037 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
1038 i13 += 8;
1039
1040
1041 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
1042
1043 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1044 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1045
1046 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
1047 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
1048 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
1049 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
1050 i14 += 8;
1051
1052
1053 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
1054
1055 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1056 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1057
1058 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
1059 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
1060 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 360));
1061 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
1062 i15 += 8;
1063
1064
1065 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
1066
1067 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1068 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1069
1070 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
1071 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
1072 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
1073 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
1074 i16 += 8;
1075
1076
1077 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
1078
1079 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1080 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1081
1082 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
1083 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
1084 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 408));
1085 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
1086 i17 += 8;
1087
1088
1089 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
1090
1091 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1092 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1093
1094 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
1095 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
1096 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 432));
1097 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
1098 i18 += 8;
1099
1100
1101 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
1102
1103 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1104 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1105
1106 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
1107 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
1108 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 456));
1109 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
1110 i19 += 8;
1111
1112
1113 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
1114
1115 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1116 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1117
1118 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
1119 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
1120 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 480));
1121 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
1122 i20 += 8;
1123
1124
1125 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
1126
1127 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1128 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1129
1130 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
1131 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
1132 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 504));
1133 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
1134 i21 += 8;
1135
1136
1137 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
1138
1139 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1140 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1141
1142 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
1143 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
1144 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 528));
1145 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
1146 i22 += 8;
1147
1148
1149 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
1150
1151 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1152 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1153
1154 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
1155 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
1156 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 552));
1157 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
1158 i23 += 8;
1159
1160
1161 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
1162
1163 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1164 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1165
1166 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
1167 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
1168 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 576));
1169 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
1170 i24 += 8;
1171
1172
1173 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
1174
1175 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1176 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1177
1178 k += 8;
1179
1180 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
1181 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
1182
1183 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
1184 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
1185 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
1186
1187 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
1188 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
1189 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1190
1191 vacc0123 = _mm_cvtps_epi32(vscaled0123);
1192 vacc4567 = _mm_cvtps_epi32(vscaled4567);
1193
1194 w = (const void*) ((const int32_t*) w + 8);
1195
1196 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
1197 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1198
1199
1200 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
1201
1202 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
1203
1204 if XNN_LIKELY(c >= 8) {
1205 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
1206 output += 8;
1207 c -= 8;
1208 } else {
1209 if (c & 4) {
1210 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
1211 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
1212 output += 4;
1213 }
1214 if (c & 2) {
1215 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
1216 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1217 output += 2;
1218 }
1219 if (c & 1) {
1220 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
1221 output += 1;
1222 }
1223 c = 0;
1224 }
1225 } while (c != 0);
1226 }
1227
1228 output = (int8_t*) ((uintptr_t) output + output_increment);
1229 } while (--output_width != 0);
1230 }
1231