xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up16x25-minmax-fp32-sse2-mul16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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_up16x25__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_up16x25__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 >= 16; c -= 16) {
164       __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
165       __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
166       __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
167       __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
168 
169 
170       const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
171       const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
172       const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
173       const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
174       i0 += 16;
175 
176       const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
177       const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
178       const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
179       const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
180 
181       const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
182       const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
183       const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
184       const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
185 
186       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
187       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
188       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
189       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
190 
191       const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
192       const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
193       const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
194       const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
195       i1 += 16;
196 
197       const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
198       const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
199       const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
200       const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
201 
202       const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
203       const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
204       const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
205       const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
206 
207       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
208       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
209       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
210       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
211 
212       const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
213       const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
214       const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
215       const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
216       i2 += 16;
217 
218       const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
219       const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
220       const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
221       const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
222 
223       const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
224       const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
225       const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
226       const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
227 
228       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
229       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
230       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
231       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
232 
233       const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
234       const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
235       const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
236       const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
237       i3 += 16;
238 
239       const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
240       const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
241       const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
242       const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
243 
244       const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
245       const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
246       const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
247       const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
248 
249       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
250       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
251       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
252       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
253 
254       const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
255       const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
256       const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
257       const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
258       i4 += 16;
259 
260       const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
261       const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
262       const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
263       const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
264 
265       const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
266       const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
267       const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
268       const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
269 
270       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
271       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
272       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
273       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
274 
275       const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
276       const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
277       const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
278       const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
279       i5 += 16;
280 
281       const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
282       const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
283       const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
284       const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
285 
286       const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
287       const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
288       const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
289       const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
290 
291       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
292       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
293       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
294       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
295 
296       const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
297       const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
298       const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
299       const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
300       i6 += 16;
301 
302       const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
303       const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
304       const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
305       const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
306 
307       const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
308       const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
309       const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
310       const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
311 
312       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
313       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
314       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
315       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
316 
317       const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
318       const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
319       const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
320       const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
321       i7 += 16;
322 
323       const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
324       const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
325       const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
326       const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
327 
328       const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
329       const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
330       const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
331       const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
332 
333       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
334       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
335       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
336       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
337 
338       const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
339       const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
340       const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
341       const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
342       i8 += 16;
343 
344       const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
345       const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
346       const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
347       const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
348 
349       const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
350       const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
351       const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
352       const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
353 
354       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
355       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
356       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
357       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
358 
359       const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
360       const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
361       const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
362       const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
363       i9 += 16;
364 
365       const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
366       const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
367       const __m128i vxi9x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x89ABCDEF, vi9x89ABCDEF), 8);
368       const __m128i vxk9x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x89ABCDEF, vk9x89ABCDEF), 8);
369 
370       const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
371       const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
372       const __m128i vprod9x89ABCDEFlo = _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
373       const __m128i vprod9x89ABCDEFhi = _mm_mulhi_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
374 
375       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
376       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
377       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod9x89ABCDEFlo, vprod9x89ABCDEFhi));
378       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod9x89ABCDEFlo, vprod9x89ABCDEFhi));
379 
380       const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
381       const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
382       const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
383       const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
384       i10 += 16;
385 
386       const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
387       const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
388       const __m128i vxi10x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x89ABCDEF, vi10x89ABCDEF), 8);
389       const __m128i vxk10x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x89ABCDEF, vk10x89ABCDEF), 8);
390 
391       const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
392       const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
393       const __m128i vprod10x89ABCDEFlo = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
394       const __m128i vprod10x89ABCDEFhi = _mm_mulhi_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
395 
396       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
397       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
398       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod10x89ABCDEFlo, vprod10x89ABCDEFhi));
399       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod10x89ABCDEFlo, vprod10x89ABCDEFhi));
400 
401       const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
402       const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
403       const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
404       const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
405       i11 += 16;
406 
407       const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
408       const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
409       const __m128i vxi11x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x89ABCDEF, vi11x89ABCDEF), 8);
410       const __m128i vxk11x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x89ABCDEF, vk11x89ABCDEF), 8);
411 
412       const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
413       const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
414       const __m128i vprod11x89ABCDEFlo = _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
415       const __m128i vprod11x89ABCDEFhi = _mm_mulhi_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
416 
417       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
418       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
419       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod11x89ABCDEFlo, vprod11x89ABCDEFhi));
420       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod11x89ABCDEFlo, vprod11x89ABCDEFhi));
421 
422       const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
423       const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
424       const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
425       const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
426       i12 += 16;
427 
428       const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
429       const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
430       const __m128i vxi12x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x89ABCDEF, vi12x89ABCDEF), 8);
431       const __m128i vxk12x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x89ABCDEF, vk12x89ABCDEF), 8);
432 
433       const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
434       const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
435       const __m128i vprod12x89ABCDEFlo = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
436       const __m128i vprod12x89ABCDEFhi = _mm_mulhi_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
437 
438       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
439       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
440       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod12x89ABCDEFlo, vprod12x89ABCDEFhi));
441       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod12x89ABCDEFlo, vprod12x89ABCDEFhi));
442 
443       const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
444       const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
445       const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
446       const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
447       i13 += 16;
448 
449       const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
450       const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
451       const __m128i vxi13x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x89ABCDEF, vi13x89ABCDEF), 8);
452       const __m128i vxk13x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x89ABCDEF, vk13x89ABCDEF), 8);
453 
454       const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
455       const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
456       const __m128i vprod13x89ABCDEFlo = _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
457       const __m128i vprod13x89ABCDEFhi = _mm_mulhi_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
458 
459       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
460       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
461       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod13x89ABCDEFlo, vprod13x89ABCDEFhi));
462       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod13x89ABCDEFlo, vprod13x89ABCDEFhi));
463 
464       const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
465       const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
466       const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
467       const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
468       i14 += 16;
469 
470       const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
471       const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
472       const __m128i vxi14x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x89ABCDEF, vi14x89ABCDEF), 8);
473       const __m128i vxk14x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x89ABCDEF, vk14x89ABCDEF), 8);
474 
475       const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
476       const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
477       const __m128i vprod14x89ABCDEFlo = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
478       const __m128i vprod14x89ABCDEFhi = _mm_mulhi_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
479 
480       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
481       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
482       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod14x89ABCDEFlo, vprod14x89ABCDEFhi));
483       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod14x89ABCDEFlo, vprod14x89ABCDEFhi));
484 
485       const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
486       const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
487       const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
488       const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
489       i15 += 16;
490 
491       const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
492       const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
493       const __m128i vxi15x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x89ABCDEF, vi15x89ABCDEF), 8);
494       const __m128i vxk15x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x89ABCDEF, vk15x89ABCDEF), 8);
495 
496       const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
497       const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
498       const __m128i vprod15x89ABCDEFlo = _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
499       const __m128i vprod15x89ABCDEFhi = _mm_mulhi_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
500 
501       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
502       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
503       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod15x89ABCDEFlo, vprod15x89ABCDEFhi));
504       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod15x89ABCDEFlo, vprod15x89ABCDEFhi));
505 
506       const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
507       const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
508       const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
509       const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
510       i16 += 16;
511 
512       const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
513       const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
514       const __m128i vxi16x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x89ABCDEF, vi16x89ABCDEF), 8);
515       const __m128i vxk16x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x89ABCDEF, vk16x89ABCDEF), 8);
516 
517       const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
518       const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
519       const __m128i vprod16x89ABCDEFlo = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
520       const __m128i vprod16x89ABCDEFhi = _mm_mulhi_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
521 
522       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
523       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
524       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod16x89ABCDEFlo, vprod16x89ABCDEFhi));
525       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod16x89ABCDEFlo, vprod16x89ABCDEFhi));
526 
527       const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
528       const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
529       const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
530       const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
531       i17 += 16;
532 
533       const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
534       const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
535       const __m128i vxi17x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x89ABCDEF, vi17x89ABCDEF), 8);
536       const __m128i vxk17x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x89ABCDEF, vk17x89ABCDEF), 8);
537 
538       const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
539       const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
540       const __m128i vprod17x89ABCDEFlo = _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
541       const __m128i vprod17x89ABCDEFhi = _mm_mulhi_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
542 
543       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
544       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
545       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod17x89ABCDEFlo, vprod17x89ABCDEFhi));
546       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod17x89ABCDEFlo, vprod17x89ABCDEFhi));
547 
548       const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
549       const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
550       const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
551       const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
552       i18 += 16;
553 
554       const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
555       const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
556       const __m128i vxi18x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x89ABCDEF, vi18x89ABCDEF), 8);
557       const __m128i vxk18x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x89ABCDEF, vk18x89ABCDEF), 8);
558 
559       const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
560       const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
561       const __m128i vprod18x89ABCDEFlo = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
562       const __m128i vprod18x89ABCDEFhi = _mm_mulhi_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
563 
564       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
565       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
566       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod18x89ABCDEFlo, vprod18x89ABCDEFhi));
567       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod18x89ABCDEFlo, vprod18x89ABCDEFhi));
568 
569       const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
570       const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
571       const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
572       const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
573       i19 += 16;
574 
575       const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
576       const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
577       const __m128i vxi19x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x89ABCDEF, vi19x89ABCDEF), 8);
578       const __m128i vxk19x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x89ABCDEF, vk19x89ABCDEF), 8);
579 
580       const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
581       const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
582       const __m128i vprod19x89ABCDEFlo = _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
583       const __m128i vprod19x89ABCDEFhi = _mm_mulhi_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
584 
585       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
586       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
587       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod19x89ABCDEFlo, vprod19x89ABCDEFhi));
588       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod19x89ABCDEFlo, vprod19x89ABCDEFhi));
589 
590       const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
591       const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
592       const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
593       const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
594       i20 += 16;
595 
596       const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
597       const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
598       const __m128i vxi20x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x89ABCDEF, vi20x89ABCDEF), 8);
599       const __m128i vxk20x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x89ABCDEF, vk20x89ABCDEF), 8);
600 
601       const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
602       const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
603       const __m128i vprod20x89ABCDEFlo = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
604       const __m128i vprod20x89ABCDEFhi = _mm_mulhi_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
605 
606       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
607       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
608       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod20x89ABCDEFlo, vprod20x89ABCDEFhi));
609       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod20x89ABCDEFlo, vprod20x89ABCDEFhi));
610 
611       const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
612       const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
613       const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
614       const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
615       i21 += 16;
616 
617       const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
618       const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
619       const __m128i vxi21x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x89ABCDEF, vi21x89ABCDEF), 8);
620       const __m128i vxk21x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x89ABCDEF, vk21x89ABCDEF), 8);
621 
622       const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
623       const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
624       const __m128i vprod21x89ABCDEFlo = _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
625       const __m128i vprod21x89ABCDEFhi = _mm_mulhi_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
626 
627       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
628       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
629       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod21x89ABCDEFlo, vprod21x89ABCDEFhi));
630       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod21x89ABCDEFlo, vprod21x89ABCDEFhi));
631 
632       const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
633       const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
634       const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
635       const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
636       i22 += 16;
637 
638       const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
639       const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
640       const __m128i vxi22x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x89ABCDEF, vi22x89ABCDEF), 8);
641       const __m128i vxk22x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x89ABCDEF, vk22x89ABCDEF), 8);
642 
643       const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
644       const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
645       const __m128i vprod22x89ABCDEFlo = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
646       const __m128i vprod22x89ABCDEFhi = _mm_mulhi_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
647 
648       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
649       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
650       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod22x89ABCDEFlo, vprod22x89ABCDEFhi));
651       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod22x89ABCDEFlo, vprod22x89ABCDEFhi));
652 
653       const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
654       const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
655       const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
656       const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
657       i23 += 16;
658 
659       const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
660       const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
661       const __m128i vxi23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x89ABCDEF, vi23x89ABCDEF), 8);
662       const __m128i vxk23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x89ABCDEF, vk23x89ABCDEF), 8);
663 
664       const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
665       const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
666       const __m128i vprod23x89ABCDEFlo = _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
667       const __m128i vprod23x89ABCDEFhi = _mm_mulhi_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
668 
669       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
670       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
671       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod23x89ABCDEFlo, vprod23x89ABCDEFhi));
672       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod23x89ABCDEFlo, vprod23x89ABCDEFhi));
673 
674       const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
675       const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
676       const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
677       const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
678       i24 += 16;
679 
680       const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
681       const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
682       const __m128i vxi24x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x89ABCDEF, vi24x89ABCDEF), 8);
683       const __m128i vxk24x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x89ABCDEF, vk24x89ABCDEF), 8);
684 
685       const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
686       const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
687       const __m128i vprod24x89ABCDEFlo = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
688       const __m128i vprod24x89ABCDEFhi = _mm_mulhi_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
689 
690       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
691       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
692       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod24x89ABCDEFlo, vprod24x89ABCDEFhi));
693       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod24x89ABCDEFlo, vprod24x89ABCDEFhi));
694 
695       w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
696 
697       __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
698       __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
699       __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
700       __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
701 
702       const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
703       vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
704       vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
705       vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
706       vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
707 
708       const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
709       vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
710       vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
711       vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
712       vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
713 
714       vacc0123 = _mm_cvtps_epi32(vscaled0123);
715       vacc4567 = _mm_cvtps_epi32(vscaled4567);
716       vacc89AB = _mm_cvtps_epi32(vscaled89AB);
717       vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
718 
719       const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
720       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
721       __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
722 
723       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
724       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
725       vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
726 
727       __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
728 
729 
730       _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
731       output += 16;
732     }
733     if XNN_UNLIKELY(c != 0) {
734       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
735       do {
736         __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
737         __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
738 
739 
740         const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
741         const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
742         i0 += 8;
743 
744         const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
745         const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
746 
747         const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
748         const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
749 
750         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
751         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
752 
753         const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
754         const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
755         i1 += 8;
756 
757         const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
758         const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
759 
760         const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
761         const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
762 
763         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
764         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
765 
766         const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
767         const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
768         i2 += 8;
769 
770         const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
771         const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
772 
773         const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
774         const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
775 
776         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
777         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
778 
779         const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
780         const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
781         i3 += 8;
782 
783         const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
784         const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
785 
786         const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
787         const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
788 
789         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
790         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
791 
792         const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
793         const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
794         i4 += 8;
795 
796         const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
797         const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
798 
799         const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
800         const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
801 
802         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
803         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
804 
805         const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
806         const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
807         i5 += 8;
808 
809         const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
810         const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
811 
812         const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
813         const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
814 
815         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
816         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
817 
818         const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
819         const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
820         i6 += 8;
821 
822         const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
823         const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
824 
825         const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
826         const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
827 
828         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
829         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
830 
831         const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
832         const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
833         i7 += 8;
834 
835         const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
836         const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
837 
838         const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
839         const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
840 
841         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
842         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
843 
844         const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
845         const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
846         i8 += 8;
847 
848         const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
849         const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
850 
851         const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
852         const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
853 
854         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
855         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
856 
857         const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
858         const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
859         i9 += 8;
860 
861         const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
862         const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
863 
864         const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
865         const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
866 
867         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
868         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
869 
870         const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
871         const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
872         i10 += 8;
873 
874         const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
875         const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
876 
877         const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
878         const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
879 
880         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
881         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
882 
883         const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
884         const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
885         i11 += 8;
886 
887         const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
888         const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
889 
890         const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
891         const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
892 
893         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
894         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
895 
896         const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
897         const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
898         i12 += 8;
899 
900         const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
901         const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
902 
903         const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
904         const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
905 
906         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
907         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
908 
909         const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
910         const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
911         i13 += 8;
912 
913         const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
914         const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
915 
916         const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
917         const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
918 
919         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
920         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
921 
922         const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
923         const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
924         i14 += 8;
925 
926         const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
927         const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
928 
929         const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
930         const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
931 
932         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
933         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
934 
935         const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
936         const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
937         i15 += 8;
938 
939         const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
940         const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
941 
942         const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
943         const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
944 
945         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
946         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
947 
948         const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
949         const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
950         i16 += 8;
951 
952         const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
953         const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
954 
955         const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
956         const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
957 
958         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
959         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
960 
961         const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
962         const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
963         i17 += 8;
964 
965         const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
966         const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
967 
968         const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
969         const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
970 
971         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
972         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
973 
974         const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
975         const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
976         i18 += 8;
977 
978         const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
979         const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
980 
981         const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
982         const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
983 
984         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
985         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
986 
987         const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
988         const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
989         i19 += 8;
990 
991         const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
992         const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
993 
994         const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
995         const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
996 
997         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
998         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
999 
1000         const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
1001         const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
1002         i20 += 8;
1003 
1004         const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
1005         const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
1006 
1007         const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
1008         const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
1009 
1010         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
1011         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
1012 
1013         const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
1014         const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
1015         i21 += 8;
1016 
1017         const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
1018         const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
1019 
1020         const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
1021         const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
1022 
1023         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
1024         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
1025 
1026         const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
1027         const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
1028         i22 += 8;
1029 
1030         const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
1031         const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
1032 
1033         const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
1034         const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
1035 
1036         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
1037         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
1038 
1039         const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
1040         const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
1041         i23 += 8;
1042 
1043         const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
1044         const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
1045 
1046         const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
1047         const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
1048 
1049         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
1050         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
1051 
1052         const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
1053         const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
1054         i24 += 8;
1055 
1056         const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
1057         const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
1058 
1059         const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
1060         const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
1061 
1062         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
1063         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
1064 
1065         k += 8;
1066 
1067         __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
1068         __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
1069 
1070         const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
1071         vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
1072         vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
1073 
1074         const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
1075         vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
1076         vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1077 
1078         vacc0123 = _mm_cvtps_epi32(vscaled0123);
1079         vacc4567 = _mm_cvtps_epi32(vscaled4567);
1080 
1081         w = (const void*) ((const int32_t*) w + 8);
1082 
1083         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
1084         __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1085 
1086         vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
1087 
1088         __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
1089 
1090 
1091         if XNN_LIKELY(c >= 8) {
1092           _mm_storel_epi64((__m128i*) output, vout0123456701234567);
1093           output += 8;
1094           c -= 8;
1095         } else {
1096           if (c & 4) {
1097             unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
1098             vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
1099             output += 4;
1100           }
1101           if (c & 2) {
1102             unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
1103             vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1104             output += 2;
1105           }
1106           if (c & 1) {
1107             *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
1108             output += 1;
1109           }
1110           c = 0;
1111         }
1112       } while (c != 0);
1113     }
1114 
1115     output = (int8_t*) ((uintptr_t) output + output_increment);
1116   } while (--output_width != 0);
1117 }
1118