1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <emmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse2_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__sse2_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 vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)));
174 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
175 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)));
176 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
177 const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)));
178 i0 += 24;
179
180 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
181 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
182 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
183 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
184 const __m128i vxi0xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi0xGHIJKLMN, vi0xGHIJKLMN), 8);
185 const __m128i vxk0xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk0xGHIJKLMN, vk0xGHIJKLMN), 8);
186
187 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
188 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
189 const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
190 const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
191 const __m128i vprod0xGHIJKLMNlo = _mm_mullo_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN);
192 const __m128i vprod0xGHIJKLMNhi = _mm_mulhi_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN);
193
194 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
195 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
196 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
197 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
198 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod0xGHIJKLMNlo, vprod0xGHIJKLMNhi));
199 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod0xGHIJKLMNlo, vprod0xGHIJKLMNhi));
200
201 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
202 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)));
203 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
204 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)));
205 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
206 const __m128i vk1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)));
207 i1 += 24;
208
209 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
210 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
211 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
212 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
213 const __m128i vxi1xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi1xGHIJKLMN, vi1xGHIJKLMN), 8);
214 const __m128i vxk1xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk1xGHIJKLMN, vk1xGHIJKLMN), 8);
215
216 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
217 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
218 const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
219 const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
220 const __m128i vprod1xGHIJKLMNlo = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN);
221 const __m128i vprod1xGHIJKLMNhi = _mm_mulhi_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN);
222
223 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
224 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
225 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
226 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
227 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod1xGHIJKLMNlo, vprod1xGHIJKLMNhi));
228 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod1xGHIJKLMNlo, vprod1xGHIJKLMNhi));
229
230 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
231 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)));
232 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
233 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)));
234 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
235 const __m128i vk2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)));
236 i2 += 24;
237
238 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
239 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
240 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
241 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
242 const __m128i vxi2xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi2xGHIJKLMN, vi2xGHIJKLMN), 8);
243 const __m128i vxk2xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk2xGHIJKLMN, vk2xGHIJKLMN), 8);
244
245 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
246 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
247 const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
248 const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
249 const __m128i vprod2xGHIJKLMNlo = _mm_mullo_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN);
250 const __m128i vprod2xGHIJKLMNhi = _mm_mulhi_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN);
251
252 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
253 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
254 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
255 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
256 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod2xGHIJKLMNlo, vprod2xGHIJKLMNhi));
257 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod2xGHIJKLMNlo, vprod2xGHIJKLMNhi));
258
259 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
260 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)));
261 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
262 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)));
263 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
264 const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)));
265 i3 += 24;
266
267 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
268 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
269 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
270 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
271 const __m128i vxi3xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi3xGHIJKLMN, vi3xGHIJKLMN), 8);
272 const __m128i vxk3xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk3xGHIJKLMN, vk3xGHIJKLMN), 8);
273
274 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
275 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
276 const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
277 const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
278 const __m128i vprod3xGHIJKLMNlo = _mm_mullo_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN);
279 const __m128i vprod3xGHIJKLMNhi = _mm_mulhi_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN);
280
281 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
282 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
283 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
284 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
285 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod3xGHIJKLMNlo, vprod3xGHIJKLMNhi));
286 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod3xGHIJKLMNlo, vprod3xGHIJKLMNhi));
287
288 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
289 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)));
290 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
291 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)));
292 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
293 const __m128i vk4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)));
294 i4 += 24;
295
296 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
297 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
298 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
299 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
300 const __m128i vxi4xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi4xGHIJKLMN, vi4xGHIJKLMN), 8);
301 const __m128i vxk4xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk4xGHIJKLMN, vk4xGHIJKLMN), 8);
302
303 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
304 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
305 const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
306 const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
307 const __m128i vprod4xGHIJKLMNlo = _mm_mullo_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN);
308 const __m128i vprod4xGHIJKLMNhi = _mm_mulhi_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN);
309
310 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
311 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
312 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
313 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
314 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod4xGHIJKLMNlo, vprod4xGHIJKLMNhi));
315 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod4xGHIJKLMNlo, vprod4xGHIJKLMNhi));
316
317 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
318 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)));
319 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
320 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)));
321 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
322 const __m128i vk5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)));
323 i5 += 24;
324
325 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
326 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
327 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
328 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
329 const __m128i vxi5xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi5xGHIJKLMN, vi5xGHIJKLMN), 8);
330 const __m128i vxk5xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk5xGHIJKLMN, vk5xGHIJKLMN), 8);
331
332 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
333 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
334 const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
335 const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
336 const __m128i vprod5xGHIJKLMNlo = _mm_mullo_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN);
337 const __m128i vprod5xGHIJKLMNhi = _mm_mulhi_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN);
338
339 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
340 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
341 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
342 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
343 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod5xGHIJKLMNlo, vprod5xGHIJKLMNhi));
344 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod5xGHIJKLMNlo, vprod5xGHIJKLMNhi));
345
346 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
347 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)));
348 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
349 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)));
350 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
351 const __m128i vk6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)));
352 i6 += 24;
353
354 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
355 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
356 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
357 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
358 const __m128i vxi6xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi6xGHIJKLMN, vi6xGHIJKLMN), 8);
359 const __m128i vxk6xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk6xGHIJKLMN, vk6xGHIJKLMN), 8);
360
361 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
362 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
363 const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
364 const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
365 const __m128i vprod6xGHIJKLMNlo = _mm_mullo_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN);
366 const __m128i vprod6xGHIJKLMNhi = _mm_mulhi_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN);
367
368 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
369 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
370 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
371 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
372 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod6xGHIJKLMNlo, vprod6xGHIJKLMNhi));
373 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod6xGHIJKLMNlo, vprod6xGHIJKLMNhi));
374
375 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
376 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)));
377 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
378 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)));
379 const __m128i vi7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i7 + 16));
380 const __m128i vk7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)));
381 i7 += 24;
382
383 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
384 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
385 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
386 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
387 const __m128i vxi7xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi7xGHIJKLMN, vi7xGHIJKLMN), 8);
388 const __m128i vxk7xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk7xGHIJKLMN, vk7xGHIJKLMN), 8);
389
390 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
391 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
392 const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
393 const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
394 const __m128i vprod7xGHIJKLMNlo = _mm_mullo_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN);
395 const __m128i vprod7xGHIJKLMNhi = _mm_mulhi_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN);
396
397 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
398 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
399 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
400 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
401 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod7xGHIJKLMNlo, vprod7xGHIJKLMNhi));
402 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod7xGHIJKLMNlo, vprod7xGHIJKLMNhi));
403
404 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
405 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)));
406 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
407 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)));
408 const __m128i vi8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i8 + 16));
409 const __m128i vk8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)));
410 i8 += 24;
411
412 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
413 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
414 const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
415 const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
416 const __m128i vxi8xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi8xGHIJKLMN, vi8xGHIJKLMN), 8);
417 const __m128i vxk8xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk8xGHIJKLMN, vk8xGHIJKLMN), 8);
418
419 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
420 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
421 const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
422 const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
423 const __m128i vprod8xGHIJKLMNlo = _mm_mullo_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN);
424 const __m128i vprod8xGHIJKLMNhi = _mm_mulhi_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN);
425
426 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
427 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
428 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
429 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
430 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod8xGHIJKLMNlo, vprod8xGHIJKLMNhi));
431 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod8xGHIJKLMNlo, vprod8xGHIJKLMNhi));
432
433 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
434 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)));
435 const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
436 const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 224 * sizeof(int8_t)));
437 const __m128i vi9xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i9 + 16));
438 const __m128i vk9xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 232 * sizeof(int8_t)));
439 i9 += 24;
440
441 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
442 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
443 const __m128i vxi9x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x89ABCDEF, vi9x89ABCDEF), 8);
444 const __m128i vxk9x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x89ABCDEF, vk9x89ABCDEF), 8);
445 const __m128i vxi9xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi9xGHIJKLMN, vi9xGHIJKLMN), 8);
446 const __m128i vxk9xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk9xGHIJKLMN, vk9xGHIJKLMN), 8);
447
448 const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
449 const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
450 const __m128i vprod9x89ABCDEFlo = _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
451 const __m128i vprod9x89ABCDEFhi = _mm_mulhi_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
452 const __m128i vprod9xGHIJKLMNlo = _mm_mullo_epi16(vxi9xGHIJKLMN, vxk9xGHIJKLMN);
453 const __m128i vprod9xGHIJKLMNhi = _mm_mulhi_epi16(vxi9xGHIJKLMN, vxk9xGHIJKLMN);
454
455 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
456 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
457 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod9x89ABCDEFlo, vprod9x89ABCDEFhi));
458 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod9x89ABCDEFlo, vprod9x89ABCDEFhi));
459 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod9xGHIJKLMNlo, vprod9xGHIJKLMNhi));
460 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod9xGHIJKLMNlo, vprod9xGHIJKLMNhi));
461
462 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
463 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 240 * sizeof(int8_t)));
464 const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
465 const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 248 * sizeof(int8_t)));
466 const __m128i vi10xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i10 + 16));
467 const __m128i vk10xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 256 * sizeof(int8_t)));
468 i10 += 24;
469
470 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
471 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
472 const __m128i vxi10x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x89ABCDEF, vi10x89ABCDEF), 8);
473 const __m128i vxk10x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x89ABCDEF, vk10x89ABCDEF), 8);
474 const __m128i vxi10xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi10xGHIJKLMN, vi10xGHIJKLMN), 8);
475 const __m128i vxk10xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk10xGHIJKLMN, vk10xGHIJKLMN), 8);
476
477 const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
478 const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
479 const __m128i vprod10x89ABCDEFlo = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
480 const __m128i vprod10x89ABCDEFhi = _mm_mulhi_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
481 const __m128i vprod10xGHIJKLMNlo = _mm_mullo_epi16(vxi10xGHIJKLMN, vxk10xGHIJKLMN);
482 const __m128i vprod10xGHIJKLMNhi = _mm_mulhi_epi16(vxi10xGHIJKLMN, vxk10xGHIJKLMN);
483
484 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
485 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
486 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod10x89ABCDEFlo, vprod10x89ABCDEFhi));
487 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod10x89ABCDEFlo, vprod10x89ABCDEFhi));
488 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod10xGHIJKLMNlo, vprod10xGHIJKLMNhi));
489 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod10xGHIJKLMNlo, vprod10xGHIJKLMNhi));
490
491 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
492 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 264 * sizeof(int8_t)));
493 const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
494 const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 272 * sizeof(int8_t)));
495 const __m128i vi11xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i11 + 16));
496 const __m128i vk11xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 280 * sizeof(int8_t)));
497 i11 += 24;
498
499 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
500 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
501 const __m128i vxi11x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x89ABCDEF, vi11x89ABCDEF), 8);
502 const __m128i vxk11x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x89ABCDEF, vk11x89ABCDEF), 8);
503 const __m128i vxi11xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi11xGHIJKLMN, vi11xGHIJKLMN), 8);
504 const __m128i vxk11xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk11xGHIJKLMN, vk11xGHIJKLMN), 8);
505
506 const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
507 const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
508 const __m128i vprod11x89ABCDEFlo = _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
509 const __m128i vprod11x89ABCDEFhi = _mm_mulhi_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
510 const __m128i vprod11xGHIJKLMNlo = _mm_mullo_epi16(vxi11xGHIJKLMN, vxk11xGHIJKLMN);
511 const __m128i vprod11xGHIJKLMNhi = _mm_mulhi_epi16(vxi11xGHIJKLMN, vxk11xGHIJKLMN);
512
513 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
514 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
515 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod11x89ABCDEFlo, vprod11x89ABCDEFhi));
516 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod11x89ABCDEFlo, vprod11x89ABCDEFhi));
517 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod11xGHIJKLMNlo, vprod11xGHIJKLMNhi));
518 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod11xGHIJKLMNlo, vprod11xGHIJKLMNhi));
519
520 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
521 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 288 * sizeof(int8_t)));
522 const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
523 const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 296 * sizeof(int8_t)));
524 const __m128i vi12xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i12 + 16));
525 const __m128i vk12xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 304 * sizeof(int8_t)));
526 i12 += 24;
527
528 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
529 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
530 const __m128i vxi12x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x89ABCDEF, vi12x89ABCDEF), 8);
531 const __m128i vxk12x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x89ABCDEF, vk12x89ABCDEF), 8);
532 const __m128i vxi12xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi12xGHIJKLMN, vi12xGHIJKLMN), 8);
533 const __m128i vxk12xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk12xGHIJKLMN, vk12xGHIJKLMN), 8);
534
535 const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
536 const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
537 const __m128i vprod12x89ABCDEFlo = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
538 const __m128i vprod12x89ABCDEFhi = _mm_mulhi_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
539 const __m128i vprod12xGHIJKLMNlo = _mm_mullo_epi16(vxi12xGHIJKLMN, vxk12xGHIJKLMN);
540 const __m128i vprod12xGHIJKLMNhi = _mm_mulhi_epi16(vxi12xGHIJKLMN, vxk12xGHIJKLMN);
541
542 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
543 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
544 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod12x89ABCDEFlo, vprod12x89ABCDEFhi));
545 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod12x89ABCDEFlo, vprod12x89ABCDEFhi));
546 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod12xGHIJKLMNlo, vprod12xGHIJKLMNhi));
547 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod12xGHIJKLMNlo, vprod12xGHIJKLMNhi));
548
549 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
550 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 312 * sizeof(int8_t)));
551 const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
552 const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 320 * sizeof(int8_t)));
553 const __m128i vi13xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i13 + 16));
554 const __m128i vk13xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 328 * sizeof(int8_t)));
555 i13 += 24;
556
557 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
558 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
559 const __m128i vxi13x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x89ABCDEF, vi13x89ABCDEF), 8);
560 const __m128i vxk13x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x89ABCDEF, vk13x89ABCDEF), 8);
561 const __m128i vxi13xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi13xGHIJKLMN, vi13xGHIJKLMN), 8);
562 const __m128i vxk13xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk13xGHIJKLMN, vk13xGHIJKLMN), 8);
563
564 const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
565 const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
566 const __m128i vprod13x89ABCDEFlo = _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
567 const __m128i vprod13x89ABCDEFhi = _mm_mulhi_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
568 const __m128i vprod13xGHIJKLMNlo = _mm_mullo_epi16(vxi13xGHIJKLMN, vxk13xGHIJKLMN);
569 const __m128i vprod13xGHIJKLMNhi = _mm_mulhi_epi16(vxi13xGHIJKLMN, vxk13xGHIJKLMN);
570
571 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
572 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
573 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod13x89ABCDEFlo, vprod13x89ABCDEFhi));
574 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod13x89ABCDEFlo, vprod13x89ABCDEFhi));
575 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod13xGHIJKLMNlo, vprod13xGHIJKLMNhi));
576 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod13xGHIJKLMNlo, vprod13xGHIJKLMNhi));
577
578 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
579 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 336 * sizeof(int8_t)));
580 const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
581 const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 344 * sizeof(int8_t)));
582 const __m128i vi14xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i14 + 16));
583 const __m128i vk14xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 352 * sizeof(int8_t)));
584 i14 += 24;
585
586 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
587 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
588 const __m128i vxi14x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x89ABCDEF, vi14x89ABCDEF), 8);
589 const __m128i vxk14x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x89ABCDEF, vk14x89ABCDEF), 8);
590 const __m128i vxi14xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi14xGHIJKLMN, vi14xGHIJKLMN), 8);
591 const __m128i vxk14xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk14xGHIJKLMN, vk14xGHIJKLMN), 8);
592
593 const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
594 const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
595 const __m128i vprod14x89ABCDEFlo = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
596 const __m128i vprod14x89ABCDEFhi = _mm_mulhi_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
597 const __m128i vprod14xGHIJKLMNlo = _mm_mullo_epi16(vxi14xGHIJKLMN, vxk14xGHIJKLMN);
598 const __m128i vprod14xGHIJKLMNhi = _mm_mulhi_epi16(vxi14xGHIJKLMN, vxk14xGHIJKLMN);
599
600 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
601 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
602 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod14x89ABCDEFlo, vprod14x89ABCDEFhi));
603 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod14x89ABCDEFlo, vprod14x89ABCDEFhi));
604 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod14xGHIJKLMNlo, vprod14xGHIJKLMNhi));
605 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod14xGHIJKLMNlo, vprod14xGHIJKLMNhi));
606
607 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
608 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 360 * sizeof(int8_t)));
609 const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
610 const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 368 * sizeof(int8_t)));
611 const __m128i vi15xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i15 + 16));
612 const __m128i vk15xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 376 * sizeof(int8_t)));
613 i15 += 24;
614
615 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
616 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
617 const __m128i vxi15x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x89ABCDEF, vi15x89ABCDEF), 8);
618 const __m128i vxk15x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x89ABCDEF, vk15x89ABCDEF), 8);
619 const __m128i vxi15xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi15xGHIJKLMN, vi15xGHIJKLMN), 8);
620 const __m128i vxk15xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk15xGHIJKLMN, vk15xGHIJKLMN), 8);
621
622 const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
623 const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
624 const __m128i vprod15x89ABCDEFlo = _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
625 const __m128i vprod15x89ABCDEFhi = _mm_mulhi_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
626 const __m128i vprod15xGHIJKLMNlo = _mm_mullo_epi16(vxi15xGHIJKLMN, vxk15xGHIJKLMN);
627 const __m128i vprod15xGHIJKLMNhi = _mm_mulhi_epi16(vxi15xGHIJKLMN, vxk15xGHIJKLMN);
628
629 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
630 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
631 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod15x89ABCDEFlo, vprod15x89ABCDEFhi));
632 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod15x89ABCDEFlo, vprod15x89ABCDEFhi));
633 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod15xGHIJKLMNlo, vprod15xGHIJKLMNhi));
634 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod15xGHIJKLMNlo, vprod15xGHIJKLMNhi));
635
636 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
637 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 384 * sizeof(int8_t)));
638 const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
639 const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 392 * sizeof(int8_t)));
640 const __m128i vi16xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i16 + 16));
641 const __m128i vk16xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 400 * sizeof(int8_t)));
642 i16 += 24;
643
644 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
645 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
646 const __m128i vxi16x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x89ABCDEF, vi16x89ABCDEF), 8);
647 const __m128i vxk16x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x89ABCDEF, vk16x89ABCDEF), 8);
648 const __m128i vxi16xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi16xGHIJKLMN, vi16xGHIJKLMN), 8);
649 const __m128i vxk16xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk16xGHIJKLMN, vk16xGHIJKLMN), 8);
650
651 const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
652 const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
653 const __m128i vprod16x89ABCDEFlo = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
654 const __m128i vprod16x89ABCDEFhi = _mm_mulhi_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
655 const __m128i vprod16xGHIJKLMNlo = _mm_mullo_epi16(vxi16xGHIJKLMN, vxk16xGHIJKLMN);
656 const __m128i vprod16xGHIJKLMNhi = _mm_mulhi_epi16(vxi16xGHIJKLMN, vxk16xGHIJKLMN);
657
658 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
659 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
660 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod16x89ABCDEFlo, vprod16x89ABCDEFhi));
661 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod16x89ABCDEFlo, vprod16x89ABCDEFhi));
662 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod16xGHIJKLMNlo, vprod16xGHIJKLMNhi));
663 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod16xGHIJKLMNlo, vprod16xGHIJKLMNhi));
664
665 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
666 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 408 * sizeof(int8_t)));
667 const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
668 const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 416 * sizeof(int8_t)));
669 const __m128i vi17xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i17 + 16));
670 const __m128i vk17xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 424 * sizeof(int8_t)));
671 i17 += 24;
672
673 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
674 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
675 const __m128i vxi17x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x89ABCDEF, vi17x89ABCDEF), 8);
676 const __m128i vxk17x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x89ABCDEF, vk17x89ABCDEF), 8);
677 const __m128i vxi17xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi17xGHIJKLMN, vi17xGHIJKLMN), 8);
678 const __m128i vxk17xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk17xGHIJKLMN, vk17xGHIJKLMN), 8);
679
680 const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
681 const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
682 const __m128i vprod17x89ABCDEFlo = _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
683 const __m128i vprod17x89ABCDEFhi = _mm_mulhi_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
684 const __m128i vprod17xGHIJKLMNlo = _mm_mullo_epi16(vxi17xGHIJKLMN, vxk17xGHIJKLMN);
685 const __m128i vprod17xGHIJKLMNhi = _mm_mulhi_epi16(vxi17xGHIJKLMN, vxk17xGHIJKLMN);
686
687 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
688 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
689 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod17x89ABCDEFlo, vprod17x89ABCDEFhi));
690 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod17x89ABCDEFlo, vprod17x89ABCDEFhi));
691 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod17xGHIJKLMNlo, vprod17xGHIJKLMNhi));
692 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod17xGHIJKLMNlo, vprod17xGHIJKLMNhi));
693
694 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
695 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 432 * sizeof(int8_t)));
696 const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
697 const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 440 * sizeof(int8_t)));
698 const __m128i vi18xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i18 + 16));
699 const __m128i vk18xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 448 * sizeof(int8_t)));
700 i18 += 24;
701
702 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
703 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
704 const __m128i vxi18x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x89ABCDEF, vi18x89ABCDEF), 8);
705 const __m128i vxk18x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x89ABCDEF, vk18x89ABCDEF), 8);
706 const __m128i vxi18xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi18xGHIJKLMN, vi18xGHIJKLMN), 8);
707 const __m128i vxk18xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk18xGHIJKLMN, vk18xGHIJKLMN), 8);
708
709 const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
710 const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
711 const __m128i vprod18x89ABCDEFlo = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
712 const __m128i vprod18x89ABCDEFhi = _mm_mulhi_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
713 const __m128i vprod18xGHIJKLMNlo = _mm_mullo_epi16(vxi18xGHIJKLMN, vxk18xGHIJKLMN);
714 const __m128i vprod18xGHIJKLMNhi = _mm_mulhi_epi16(vxi18xGHIJKLMN, vxk18xGHIJKLMN);
715
716 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
717 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
718 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod18x89ABCDEFlo, vprod18x89ABCDEFhi));
719 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod18x89ABCDEFlo, vprod18x89ABCDEFhi));
720 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod18xGHIJKLMNlo, vprod18xGHIJKLMNhi));
721 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod18xGHIJKLMNlo, vprod18xGHIJKLMNhi));
722
723 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
724 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 456 * sizeof(int8_t)));
725 const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
726 const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 464 * sizeof(int8_t)));
727 const __m128i vi19xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i19 + 16));
728 const __m128i vk19xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 472 * sizeof(int8_t)));
729 i19 += 24;
730
731 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
732 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
733 const __m128i vxi19x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x89ABCDEF, vi19x89ABCDEF), 8);
734 const __m128i vxk19x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x89ABCDEF, vk19x89ABCDEF), 8);
735 const __m128i vxi19xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi19xGHIJKLMN, vi19xGHIJKLMN), 8);
736 const __m128i vxk19xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk19xGHIJKLMN, vk19xGHIJKLMN), 8);
737
738 const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
739 const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
740 const __m128i vprod19x89ABCDEFlo = _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
741 const __m128i vprod19x89ABCDEFhi = _mm_mulhi_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
742 const __m128i vprod19xGHIJKLMNlo = _mm_mullo_epi16(vxi19xGHIJKLMN, vxk19xGHIJKLMN);
743 const __m128i vprod19xGHIJKLMNhi = _mm_mulhi_epi16(vxi19xGHIJKLMN, vxk19xGHIJKLMN);
744
745 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
746 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
747 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod19x89ABCDEFlo, vprod19x89ABCDEFhi));
748 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod19x89ABCDEFlo, vprod19x89ABCDEFhi));
749 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod19xGHIJKLMNlo, vprod19xGHIJKLMNhi));
750 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod19xGHIJKLMNlo, vprod19xGHIJKLMNhi));
751
752 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
753 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 480 * sizeof(int8_t)));
754 const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
755 const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 488 * sizeof(int8_t)));
756 const __m128i vi20xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i20 + 16));
757 const __m128i vk20xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 496 * sizeof(int8_t)));
758 i20 += 24;
759
760 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
761 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
762 const __m128i vxi20x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x89ABCDEF, vi20x89ABCDEF), 8);
763 const __m128i vxk20x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x89ABCDEF, vk20x89ABCDEF), 8);
764 const __m128i vxi20xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi20xGHIJKLMN, vi20xGHIJKLMN), 8);
765 const __m128i vxk20xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk20xGHIJKLMN, vk20xGHIJKLMN), 8);
766
767 const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
768 const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
769 const __m128i vprod20x89ABCDEFlo = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
770 const __m128i vprod20x89ABCDEFhi = _mm_mulhi_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
771 const __m128i vprod20xGHIJKLMNlo = _mm_mullo_epi16(vxi20xGHIJKLMN, vxk20xGHIJKLMN);
772 const __m128i vprod20xGHIJKLMNhi = _mm_mulhi_epi16(vxi20xGHIJKLMN, vxk20xGHIJKLMN);
773
774 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
775 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
776 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod20x89ABCDEFlo, vprod20x89ABCDEFhi));
777 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod20x89ABCDEFlo, vprod20x89ABCDEFhi));
778 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod20xGHIJKLMNlo, vprod20xGHIJKLMNhi));
779 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod20xGHIJKLMNlo, vprod20xGHIJKLMNhi));
780
781 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
782 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 504 * sizeof(int8_t)));
783 const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
784 const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 512 * sizeof(int8_t)));
785 const __m128i vi21xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i21 + 16));
786 const __m128i vk21xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 520 * sizeof(int8_t)));
787 i21 += 24;
788
789 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
790 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
791 const __m128i vxi21x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x89ABCDEF, vi21x89ABCDEF), 8);
792 const __m128i vxk21x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x89ABCDEF, vk21x89ABCDEF), 8);
793 const __m128i vxi21xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi21xGHIJKLMN, vi21xGHIJKLMN), 8);
794 const __m128i vxk21xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk21xGHIJKLMN, vk21xGHIJKLMN), 8);
795
796 const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
797 const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
798 const __m128i vprod21x89ABCDEFlo = _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
799 const __m128i vprod21x89ABCDEFhi = _mm_mulhi_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
800 const __m128i vprod21xGHIJKLMNlo = _mm_mullo_epi16(vxi21xGHIJKLMN, vxk21xGHIJKLMN);
801 const __m128i vprod21xGHIJKLMNhi = _mm_mulhi_epi16(vxi21xGHIJKLMN, vxk21xGHIJKLMN);
802
803 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
804 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
805 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod21x89ABCDEFlo, vprod21x89ABCDEFhi));
806 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod21x89ABCDEFlo, vprod21x89ABCDEFhi));
807 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod21xGHIJKLMNlo, vprod21xGHIJKLMNhi));
808 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod21xGHIJKLMNlo, vprod21xGHIJKLMNhi));
809
810 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
811 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 528 * sizeof(int8_t)));
812 const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
813 const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 536 * sizeof(int8_t)));
814 const __m128i vi22xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i22 + 16));
815 const __m128i vk22xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 544 * sizeof(int8_t)));
816 i22 += 24;
817
818 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
819 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
820 const __m128i vxi22x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x89ABCDEF, vi22x89ABCDEF), 8);
821 const __m128i vxk22x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x89ABCDEF, vk22x89ABCDEF), 8);
822 const __m128i vxi22xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi22xGHIJKLMN, vi22xGHIJKLMN), 8);
823 const __m128i vxk22xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk22xGHIJKLMN, vk22xGHIJKLMN), 8);
824
825 const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
826 const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
827 const __m128i vprod22x89ABCDEFlo = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
828 const __m128i vprod22x89ABCDEFhi = _mm_mulhi_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
829 const __m128i vprod22xGHIJKLMNlo = _mm_mullo_epi16(vxi22xGHIJKLMN, vxk22xGHIJKLMN);
830 const __m128i vprod22xGHIJKLMNhi = _mm_mulhi_epi16(vxi22xGHIJKLMN, vxk22xGHIJKLMN);
831
832 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
833 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
834 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod22x89ABCDEFlo, vprod22x89ABCDEFhi));
835 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod22x89ABCDEFlo, vprod22x89ABCDEFhi));
836 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod22xGHIJKLMNlo, vprod22xGHIJKLMNhi));
837 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod22xGHIJKLMNlo, vprod22xGHIJKLMNhi));
838
839 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
840 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 552 * sizeof(int8_t)));
841 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
842 const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 560 * sizeof(int8_t)));
843 const __m128i vi23xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i23 + 16));
844 const __m128i vk23xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 568 * sizeof(int8_t)));
845 i23 += 24;
846
847 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
848 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
849 const __m128i vxi23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x89ABCDEF, vi23x89ABCDEF), 8);
850 const __m128i vxk23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x89ABCDEF, vk23x89ABCDEF), 8);
851 const __m128i vxi23xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi23xGHIJKLMN, vi23xGHIJKLMN), 8);
852 const __m128i vxk23xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk23xGHIJKLMN, vk23xGHIJKLMN), 8);
853
854 const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
855 const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
856 const __m128i vprod23x89ABCDEFlo = _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
857 const __m128i vprod23x89ABCDEFhi = _mm_mulhi_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
858 const __m128i vprod23xGHIJKLMNlo = _mm_mullo_epi16(vxi23xGHIJKLMN, vxk23xGHIJKLMN);
859 const __m128i vprod23xGHIJKLMNhi = _mm_mulhi_epi16(vxi23xGHIJKLMN, vxk23xGHIJKLMN);
860
861 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
862 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
863 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod23x89ABCDEFlo, vprod23x89ABCDEFhi));
864 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod23x89ABCDEFlo, vprod23x89ABCDEFhi));
865 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod23xGHIJKLMNlo, vprod23xGHIJKLMNhi));
866 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod23xGHIJKLMNlo, vprod23xGHIJKLMNhi));
867
868 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
869 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 576 * sizeof(int8_t)));
870 const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
871 const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 584 * sizeof(int8_t)));
872 const __m128i vi24xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i24 + 16));
873 const __m128i vk24xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 592 * sizeof(int8_t)));
874 i24 += 24;
875
876 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
877 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
878 const __m128i vxi24x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x89ABCDEF, vi24x89ABCDEF), 8);
879 const __m128i vxk24x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x89ABCDEF, vk24x89ABCDEF), 8);
880 const __m128i vxi24xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi24xGHIJKLMN, vi24xGHIJKLMN), 8);
881 const __m128i vxk24xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk24xGHIJKLMN, vk24xGHIJKLMN), 8);
882
883 const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
884 const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
885 const __m128i vprod24x89ABCDEFlo = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
886 const __m128i vprod24x89ABCDEFhi = _mm_mulhi_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
887 const __m128i vprod24xGHIJKLMNlo = _mm_mullo_epi16(vxi24xGHIJKLMN, vxk24xGHIJKLMN);
888 const __m128i vprod24xGHIJKLMNhi = _mm_mulhi_epi16(vxi24xGHIJKLMN, vxk24xGHIJKLMN);
889
890 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
891 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
892 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod24x89ABCDEFlo, vprod24x89ABCDEFhi));
893 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod24x89ABCDEFlo, vprod24x89ABCDEFhi));
894 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod24xGHIJKLMNlo, vprod24xGHIJKLMNhi));
895 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod24xGHIJKLMNlo, vprod24xGHIJKLMNhi));
896
897 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t));
898
899 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
900 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
901 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
902 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
903 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
904 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
905
906 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
907 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
908 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
909 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
910 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
911 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscale);
912 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscale);
913
914 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
915 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
916 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
917 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
918 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
919 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
920 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
921
922 vacc0123 = _mm_cvtps_epi32(vscaled0123);
923 vacc4567 = _mm_cvtps_epi32(vscaled4567);
924 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
925 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
926 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
927 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
928
929 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
930 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
931 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
932 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
933
934 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
935 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
936 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
937 voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
938
939 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
940 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
941
942
943 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
944 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
945 output += 24;
946 }
947 if XNN_UNLIKELY(c != 0) {
948 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
949 do {
950 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
951 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
952
953
954 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
955 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
956 i0 += 8;
957
958 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
959 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
960
961 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
962 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
963
964 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
965 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
966
967 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
968 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 24));
969 i1 += 8;
970
971 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
972 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
973
974 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
975 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
976
977 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
978 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
979
980 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
981 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
982 i2 += 8;
983
984 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
985 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
986
987 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
988 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
989
990 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
991 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
992
993 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
994 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 72));
995 i3 += 8;
996
997 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
998 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
999
1000 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
1001 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
1002
1003 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
1004 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
1005
1006 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
1007 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
1008 i4 += 8;
1009
1010 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
1011 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
1012
1013 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
1014 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
1015
1016 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
1017 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
1018
1019 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
1020 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 120));
1021 i5 += 8;
1022
1023 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
1024 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
1025
1026 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
1027 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
1028
1029 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
1030 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
1031
1032 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
1033 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
1034 i6 += 8;
1035
1036 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
1037 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
1038
1039 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
1040 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
1041
1042 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
1043 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
1044
1045 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
1046 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 168));
1047 i7 += 8;
1048
1049 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
1050 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
1051
1052 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
1053 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
1054
1055 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
1056 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
1057
1058 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
1059 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
1060 i8 += 8;
1061
1062 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
1063 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
1064
1065 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
1066 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
1067
1068 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
1069 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
1070
1071 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
1072 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 216));
1073 i9 += 8;
1074
1075 const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
1076 const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
1077
1078 const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
1079 const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
1080
1081 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
1082 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
1083
1084 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
1085 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
1086 i10 += 8;
1087
1088 const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
1089 const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
1090
1091 const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
1092 const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
1093
1094 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
1095 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
1096
1097 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
1098 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 264));
1099 i11 += 8;
1100
1101 const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
1102 const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
1103
1104 const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
1105 const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
1106
1107 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
1108 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
1109
1110 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
1111 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
1112 i12 += 8;
1113
1114 const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
1115 const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
1116
1117 const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
1118 const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
1119
1120 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
1121 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
1122
1123 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
1124 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 312));
1125 i13 += 8;
1126
1127 const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
1128 const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
1129
1130 const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
1131 const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
1132
1133 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
1134 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
1135
1136 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
1137 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
1138 i14 += 8;
1139
1140 const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
1141 const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
1142
1143 const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
1144 const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
1145
1146 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
1147 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
1148
1149 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
1150 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 360));
1151 i15 += 8;
1152
1153 const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
1154 const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
1155
1156 const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
1157 const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
1158
1159 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
1160 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
1161
1162 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
1163 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
1164 i16 += 8;
1165
1166 const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
1167 const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
1168
1169 const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
1170 const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
1171
1172 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
1173 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
1174
1175 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
1176 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 408));
1177 i17 += 8;
1178
1179 const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
1180 const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
1181
1182 const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
1183 const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
1184
1185 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
1186 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
1187
1188 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
1189 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 432));
1190 i18 += 8;
1191
1192 const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
1193 const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
1194
1195 const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
1196 const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
1197
1198 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
1199 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
1200
1201 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
1202 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 456));
1203 i19 += 8;
1204
1205 const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
1206 const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
1207
1208 const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
1209 const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
1210
1211 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
1212 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
1213
1214 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
1215 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 480));
1216 i20 += 8;
1217
1218 const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
1219 const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
1220
1221 const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
1222 const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
1223
1224 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
1225 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
1226
1227 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
1228 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 504));
1229 i21 += 8;
1230
1231 const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
1232 const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
1233
1234 const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
1235 const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
1236
1237 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
1238 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
1239
1240 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
1241 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 528));
1242 i22 += 8;
1243
1244 const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
1245 const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
1246
1247 const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
1248 const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
1249
1250 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
1251 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
1252
1253 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
1254 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 552));
1255 i23 += 8;
1256
1257 const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
1258 const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
1259
1260 const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
1261 const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
1262
1263 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
1264 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
1265
1266 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
1267 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 576));
1268 i24 += 8;
1269
1270 const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
1271 const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
1272
1273 const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
1274 const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
1275
1276 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
1277 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
1278
1279 k += 8;
1280
1281 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
1282 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
1283
1284 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
1285 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
1286 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
1287
1288 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
1289 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
1290 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1291
1292 vacc0123 = _mm_cvtps_epi32(vscaled0123);
1293 vacc4567 = _mm_cvtps_epi32(vscaled4567);
1294
1295 w = (const void*) ((const int32_t*) w + 8);
1296
1297 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
1298 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1299
1300 vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
1301
1302 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
1303
1304
1305 if XNN_LIKELY(c >= 8) {
1306 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
1307 output += 8;
1308 c -= 8;
1309 } else {
1310 if (c & 4) {
1311 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
1312 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
1313 output += 4;
1314 }
1315 if (c & 2) {
1316 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
1317 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1318 output += 2;
1319 }
1320 if (c & 1) {
1321 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
1322 output += 1;
1323 }
1324 c = 0;
1325 }
1326 } while (c != 0);
1327 }
1328
1329 output = (int8_t*) ((uintptr_t) output + output_increment);
1330 } while (--output_width != 0);
1331 }
1332