xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/up16x25-minmax-fp32-sse41-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 <smmintrin.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16 
17 
xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul16(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__sse41_mul16(
19     size_t channels,
20     size_t output_width,
21     const uint8_t** input,
22     const void* weights,
23     uint8_t* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const uint8_t* zero,
28     const union xnn_qu8_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 uint8_t* i0 = input[0];
35     assert(i0 != NULL);
36     if XNN_UNPREDICTABLE(i0 != zero) {
37       i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
38     }
39     const uint8_t* i1 = input[1];
40     assert(i1 != NULL);
41     if XNN_UNPREDICTABLE(i1 != zero) {
42       i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
43     }
44     const uint8_t* i2 = input[2];
45     assert(i2 != NULL);
46     if XNN_UNPREDICTABLE(i2 != zero) {
47       i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
48     }
49     const uint8_t* i3 = input[3];
50     assert(i3 != NULL);
51     if XNN_UNPREDICTABLE(i3 != zero) {
52       i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
53     }
54     const uint8_t* i4 = input[4];
55     assert(i4 != NULL);
56     if XNN_UNPREDICTABLE(i4 != zero) {
57       i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
58     }
59     const uint8_t* i5 = input[5];
60     assert(i5 != NULL);
61     if XNN_UNPREDICTABLE(i5 != zero) {
62       i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
63     }
64     const uint8_t* i6 = input[6];
65     assert(i6 != NULL);
66     if XNN_UNPREDICTABLE(i6 != zero) {
67       i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
68     }
69     const uint8_t* i7 = input[7];
70     assert(i7 != NULL);
71     if XNN_UNPREDICTABLE(i7 != zero) {
72       i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
73     }
74     const uint8_t* i8 = input[8];
75     assert(i8 != NULL);
76     if XNN_UNPREDICTABLE(i8 != zero) {
77       i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
78     }
79     const uint8_t* i9 = input[9];
80     assert(i9 != NULL);
81     if XNN_UNPREDICTABLE(i9 != zero) {
82       i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
83     }
84     const uint8_t* i10 = input[10];
85     assert(i10 != NULL);
86     if XNN_UNPREDICTABLE(i10 != zero) {
87       i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
88     }
89     const uint8_t* i11 = input[11];
90     assert(i11 != NULL);
91     if XNN_UNPREDICTABLE(i11 != zero) {
92       i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
93     }
94     const uint8_t* i12 = input[12];
95     assert(i12 != NULL);
96     if XNN_UNPREDICTABLE(i12 != zero) {
97       i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
98     }
99     const uint8_t* i13 = input[13];
100     assert(i13 != NULL);
101     if XNN_UNPREDICTABLE(i13 != zero) {
102       i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
103     }
104     const uint8_t* i14 = input[14];
105     assert(i14 != NULL);
106     if XNN_UNPREDICTABLE(i14 != zero) {
107       i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
108     }
109     const uint8_t* i15 = input[15];
110     assert(i15 != NULL);
111     if XNN_UNPREDICTABLE(i15 != zero) {
112       i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
113     }
114     const uint8_t* i16 = input[16];
115     assert(i16 != NULL);
116     if XNN_UNPREDICTABLE(i16 != zero) {
117       i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
118     }
119     const uint8_t* i17 = input[17];
120     assert(i17 != NULL);
121     if XNN_UNPREDICTABLE(i17 != zero) {
122       i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
123     }
124     const uint8_t* i18 = input[18];
125     assert(i18 != NULL);
126     if XNN_UNPREDICTABLE(i18 != zero) {
127       i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
128     }
129     const uint8_t* i19 = input[19];
130     assert(i19 != NULL);
131     if XNN_UNPREDICTABLE(i19 != zero) {
132       i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
133     }
134     const uint8_t* i20 = input[20];
135     assert(i20 != NULL);
136     if XNN_UNPREDICTABLE(i20 != zero) {
137       i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
138     }
139     const uint8_t* i21 = input[21];
140     assert(i21 != NULL);
141     if XNN_UNPREDICTABLE(i21 != zero) {
142       i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
143     }
144     const uint8_t* i22 = input[22];
145     assert(i22 != NULL);
146     if XNN_UNPREDICTABLE(i22 != zero) {
147       i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
148     }
149     const uint8_t* i23 = input[23];
150     assert(i23 != NULL);
151     if XNN_UNPREDICTABLE(i23 != zero) {
152       i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
153     }
154     const uint8_t* i24 = input[24];
155     assert(i24 != NULL);
156     if XNN_UNPREDICTABLE(i24 != zero) {
157       i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
158     }
159     input = (const uint8_t**) ((uintptr_t) input + input_stride);
160 
161     size_t c = channels;
162     const void* w = weights;
163     const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
164     for (; c >= 16; c -= 16) {
165       __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
166       __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
167       __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
168       __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
169 
170 
171       const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
172       const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
173       const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
174       const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
175       const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
176       const __m128i vxi0x89ABCDEF = _mm_cvtepu8_epi16(vi0x89ABCDEF);
177       const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
178       const __m128i vxk0x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x89ABCDEF), vk_zero_point);
179       i0 += 16;
180 
181 
182       const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
183       const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
184       const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
185       const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
186 
187       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
188       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
189       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
190       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
191 
192       const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
193       const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
194       const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
195       const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
196       const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
197       const __m128i vxi1x89ABCDEF = _mm_cvtepu8_epi16(vi1x89ABCDEF);
198       const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
199       const __m128i vxk1x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x89ABCDEF), vk_zero_point);
200       i1 += 16;
201 
202 
203       const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
204       const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
205       const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
206       const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
207 
208       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
209       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
210       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
211       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
212 
213       const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
214       const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
215       const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
216       const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
217       const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
218       const __m128i vxi2x89ABCDEF = _mm_cvtepu8_epi16(vi2x89ABCDEF);
219       const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
220       const __m128i vxk2x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x89ABCDEF), vk_zero_point);
221       i2 += 16;
222 
223 
224       const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
225       const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
226       const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
227       const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
228 
229       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
230       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
231       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
232       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
233 
234       const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
235       const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
236       const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
237       const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
238       const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
239       const __m128i vxi3x89ABCDEF = _mm_cvtepu8_epi16(vi3x89ABCDEF);
240       const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
241       const __m128i vxk3x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x89ABCDEF), vk_zero_point);
242       i3 += 16;
243 
244 
245       const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
246       const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
247       const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
248       const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
249 
250       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
251       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
252       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
253       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
254 
255       const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
256       const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
257       const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
258       const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
259       const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
260       const __m128i vxi4x89ABCDEF = _mm_cvtepu8_epi16(vi4x89ABCDEF);
261       const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
262       const __m128i vxk4x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x89ABCDEF), vk_zero_point);
263       i4 += 16;
264 
265 
266       const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
267       const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
268       const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
269       const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
270 
271       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
272       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
273       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
274       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
275 
276       const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
277       const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
278       const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
279       const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
280       const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
281       const __m128i vxi5x89ABCDEF = _mm_cvtepu8_epi16(vi5x89ABCDEF);
282       const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
283       const __m128i vxk5x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x89ABCDEF), vk_zero_point);
284       i5 += 16;
285 
286 
287       const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
288       const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
289       const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
290       const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
291 
292       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
293       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
294       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
295       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
296 
297       const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
298       const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
299       const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
300       const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
301       const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
302       const __m128i vxi6x89ABCDEF = _mm_cvtepu8_epi16(vi6x89ABCDEF);
303       const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
304       const __m128i vxk6x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x89ABCDEF), vk_zero_point);
305       i6 += 16;
306 
307 
308       const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
309       const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
310       const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
311       const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
312 
313       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
314       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
315       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
316       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
317 
318       const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
319       const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
320       const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
321       const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
322       const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
323       const __m128i vxi7x89ABCDEF = _mm_cvtepu8_epi16(vi7x89ABCDEF);
324       const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
325       const __m128i vxk7x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x89ABCDEF), vk_zero_point);
326       i7 += 16;
327 
328 
329       const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
330       const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
331       const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
332       const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
333 
334       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
335       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
336       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
337       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
338 
339       const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
340       const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
341       const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
342       const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
343       const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
344       const __m128i vxi8x89ABCDEF = _mm_cvtepu8_epi16(vi8x89ABCDEF);
345       const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
346       const __m128i vxk8x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x89ABCDEF), vk_zero_point);
347       i8 += 16;
348 
349 
350       const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
351       const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
352       const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
353       const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
354 
355       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
356       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
357       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
358       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
359 
360       const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
361       const __m128i vxi9x01234567 = _mm_cvtepu8_epi16(vi9x01234567);
362       const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
363       const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk9x01234567), vk_zero_point);
364       const __m128i vi9x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i9 + 8));
365       const __m128i vxi9x89ABCDEF = _mm_cvtepu8_epi16(vi9x89ABCDEF);
366       const __m128i vk9x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
367       const __m128i vxk9x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk9x89ABCDEF), vk_zero_point);
368       i9 += 16;
369 
370 
371       const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
372       const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
373       const __m128i vprod9x89ABCDEFlo = _mm_mullo_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
374       const __m128i vprod9x89ABCDEFhi = _mm_mulhi_epi16(vxi9x89ABCDEF, vxk9x89ABCDEF);
375 
376       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
377       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
378       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod9x89ABCDEFlo, vprod9x89ABCDEFhi));
379       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod9x89ABCDEFlo, vprod9x89ABCDEFhi));
380 
381       const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
382       const __m128i vxi10x01234567 = _mm_cvtepu8_epi16(vi10x01234567);
383       const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
384       const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk10x01234567), vk_zero_point);
385       const __m128i vi10x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i10 + 8));
386       const __m128i vxi10x89ABCDEF = _mm_cvtepu8_epi16(vi10x89ABCDEF);
387       const __m128i vk10x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
388       const __m128i vxk10x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk10x89ABCDEF), vk_zero_point);
389       i10 += 16;
390 
391 
392       const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
393       const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
394       const __m128i vprod10x89ABCDEFlo = _mm_mullo_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
395       const __m128i vprod10x89ABCDEFhi = _mm_mulhi_epi16(vxi10x89ABCDEF, vxk10x89ABCDEF);
396 
397       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
398       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
399       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod10x89ABCDEFlo, vprod10x89ABCDEFhi));
400       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod10x89ABCDEFlo, vprod10x89ABCDEFhi));
401 
402       const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
403       const __m128i vxi11x01234567 = _mm_cvtepu8_epi16(vi11x01234567);
404       const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
405       const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk11x01234567), vk_zero_point);
406       const __m128i vi11x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i11 + 8));
407       const __m128i vxi11x89ABCDEF = _mm_cvtepu8_epi16(vi11x89ABCDEF);
408       const __m128i vk11x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
409       const __m128i vxk11x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk11x89ABCDEF), vk_zero_point);
410       i11 += 16;
411 
412 
413       const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
414       const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
415       const __m128i vprod11x89ABCDEFlo = _mm_mullo_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
416       const __m128i vprod11x89ABCDEFhi = _mm_mulhi_epi16(vxi11x89ABCDEF, vxk11x89ABCDEF);
417 
418       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
419       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
420       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod11x89ABCDEFlo, vprod11x89ABCDEFhi));
421       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod11x89ABCDEFlo, vprod11x89ABCDEFhi));
422 
423       const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
424       const __m128i vxi12x01234567 = _mm_cvtepu8_epi16(vi12x01234567);
425       const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
426       const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk12x01234567), vk_zero_point);
427       const __m128i vi12x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i12 + 8));
428       const __m128i vxi12x89ABCDEF = _mm_cvtepu8_epi16(vi12x89ABCDEF);
429       const __m128i vk12x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(uint8_t)));
430       const __m128i vxk12x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk12x89ABCDEF), vk_zero_point);
431       i12 += 16;
432 
433 
434       const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
435       const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
436       const __m128i vprod12x89ABCDEFlo = _mm_mullo_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
437       const __m128i vprod12x89ABCDEFhi = _mm_mulhi_epi16(vxi12x89ABCDEF, vxk12x89ABCDEF);
438 
439       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
440       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
441       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod12x89ABCDEFlo, vprod12x89ABCDEFhi));
442       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod12x89ABCDEFlo, vprod12x89ABCDEFhi));
443 
444       const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
445       const __m128i vxi13x01234567 = _mm_cvtepu8_epi16(vi13x01234567);
446       const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(uint8_t)));
447       const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk13x01234567), vk_zero_point);
448       const __m128i vi13x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i13 + 8));
449       const __m128i vxi13x89ABCDEF = _mm_cvtepu8_epi16(vi13x89ABCDEF);
450       const __m128i vk13x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(uint8_t)));
451       const __m128i vxk13x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk13x89ABCDEF), vk_zero_point);
452       i13 += 16;
453 
454 
455       const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
456       const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
457       const __m128i vprod13x89ABCDEFlo = _mm_mullo_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
458       const __m128i vprod13x89ABCDEFhi = _mm_mulhi_epi16(vxi13x89ABCDEF, vxk13x89ABCDEF);
459 
460       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
461       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
462       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod13x89ABCDEFlo, vprod13x89ABCDEFhi));
463       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod13x89ABCDEFlo, vprod13x89ABCDEFhi));
464 
465       const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
466       const __m128i vxi14x01234567 = _mm_cvtepu8_epi16(vi14x01234567);
467       const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(uint8_t)));
468       const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk14x01234567), vk_zero_point);
469       const __m128i vi14x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i14 + 8));
470       const __m128i vxi14x89ABCDEF = _mm_cvtepu8_epi16(vi14x89ABCDEF);
471       const __m128i vk14x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(uint8_t)));
472       const __m128i vxk14x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk14x89ABCDEF), vk_zero_point);
473       i14 += 16;
474 
475 
476       const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
477       const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
478       const __m128i vprod14x89ABCDEFlo = _mm_mullo_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
479       const __m128i vprod14x89ABCDEFhi = _mm_mulhi_epi16(vxi14x89ABCDEF, vxk14x89ABCDEF);
480 
481       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
482       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
483       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod14x89ABCDEFlo, vprod14x89ABCDEFhi));
484       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod14x89ABCDEFlo, vprod14x89ABCDEFhi));
485 
486       const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
487       const __m128i vxi15x01234567 = _mm_cvtepu8_epi16(vi15x01234567);
488       const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(uint8_t)));
489       const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk15x01234567), vk_zero_point);
490       const __m128i vi15x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i15 + 8));
491       const __m128i vxi15x89ABCDEF = _mm_cvtepu8_epi16(vi15x89ABCDEF);
492       const __m128i vk15x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(uint8_t)));
493       const __m128i vxk15x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk15x89ABCDEF), vk_zero_point);
494       i15 += 16;
495 
496 
497       const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
498       const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
499       const __m128i vprod15x89ABCDEFlo = _mm_mullo_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
500       const __m128i vprod15x89ABCDEFhi = _mm_mulhi_epi16(vxi15x89ABCDEF, vxk15x89ABCDEF);
501 
502       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
503       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
504       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod15x89ABCDEFlo, vprod15x89ABCDEFhi));
505       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod15x89ABCDEFlo, vprod15x89ABCDEFhi));
506 
507       const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
508       const __m128i vxi16x01234567 = _mm_cvtepu8_epi16(vi16x01234567);
509       const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(uint8_t)));
510       const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk16x01234567), vk_zero_point);
511       const __m128i vi16x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i16 + 8));
512       const __m128i vxi16x89ABCDEF = _mm_cvtepu8_epi16(vi16x89ABCDEF);
513       const __m128i vk16x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(uint8_t)));
514       const __m128i vxk16x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk16x89ABCDEF), vk_zero_point);
515       i16 += 16;
516 
517 
518       const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
519       const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
520       const __m128i vprod16x89ABCDEFlo = _mm_mullo_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
521       const __m128i vprod16x89ABCDEFhi = _mm_mulhi_epi16(vxi16x89ABCDEF, vxk16x89ABCDEF);
522 
523       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
524       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
525       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod16x89ABCDEFlo, vprod16x89ABCDEFhi));
526       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod16x89ABCDEFlo, vprod16x89ABCDEFhi));
527 
528       const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
529       const __m128i vxi17x01234567 = _mm_cvtepu8_epi16(vi17x01234567);
530       const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(uint8_t)));
531       const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk17x01234567), vk_zero_point);
532       const __m128i vi17x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i17 + 8));
533       const __m128i vxi17x89ABCDEF = _mm_cvtepu8_epi16(vi17x89ABCDEF);
534       const __m128i vk17x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(uint8_t)));
535       const __m128i vxk17x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk17x89ABCDEF), vk_zero_point);
536       i17 += 16;
537 
538 
539       const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
540       const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
541       const __m128i vprod17x89ABCDEFlo = _mm_mullo_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
542       const __m128i vprod17x89ABCDEFhi = _mm_mulhi_epi16(vxi17x89ABCDEF, vxk17x89ABCDEF);
543 
544       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
545       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
546       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod17x89ABCDEFlo, vprod17x89ABCDEFhi));
547       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod17x89ABCDEFlo, vprod17x89ABCDEFhi));
548 
549       const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
550       const __m128i vxi18x01234567 = _mm_cvtepu8_epi16(vi18x01234567);
551       const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(uint8_t)));
552       const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk18x01234567), vk_zero_point);
553       const __m128i vi18x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i18 + 8));
554       const __m128i vxi18x89ABCDEF = _mm_cvtepu8_epi16(vi18x89ABCDEF);
555       const __m128i vk18x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(uint8_t)));
556       const __m128i vxk18x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk18x89ABCDEF), vk_zero_point);
557       i18 += 16;
558 
559 
560       const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
561       const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
562       const __m128i vprod18x89ABCDEFlo = _mm_mullo_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
563       const __m128i vprod18x89ABCDEFhi = _mm_mulhi_epi16(vxi18x89ABCDEF, vxk18x89ABCDEF);
564 
565       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
566       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
567       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod18x89ABCDEFlo, vprod18x89ABCDEFhi));
568       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod18x89ABCDEFlo, vprod18x89ABCDEFhi));
569 
570       const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
571       const __m128i vxi19x01234567 = _mm_cvtepu8_epi16(vi19x01234567);
572       const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(uint8_t)));
573       const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk19x01234567), vk_zero_point);
574       const __m128i vi19x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i19 + 8));
575       const __m128i vxi19x89ABCDEF = _mm_cvtepu8_epi16(vi19x89ABCDEF);
576       const __m128i vk19x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(uint8_t)));
577       const __m128i vxk19x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk19x89ABCDEF), vk_zero_point);
578       i19 += 16;
579 
580 
581       const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
582       const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
583       const __m128i vprod19x89ABCDEFlo = _mm_mullo_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
584       const __m128i vprod19x89ABCDEFhi = _mm_mulhi_epi16(vxi19x89ABCDEF, vxk19x89ABCDEF);
585 
586       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
587       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
588       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod19x89ABCDEFlo, vprod19x89ABCDEFhi));
589       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod19x89ABCDEFlo, vprod19x89ABCDEFhi));
590 
591       const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
592       const __m128i vxi20x01234567 = _mm_cvtepu8_epi16(vi20x01234567);
593       const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(uint8_t)));
594       const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk20x01234567), vk_zero_point);
595       const __m128i vi20x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i20 + 8));
596       const __m128i vxi20x89ABCDEF = _mm_cvtepu8_epi16(vi20x89ABCDEF);
597       const __m128i vk20x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(uint8_t)));
598       const __m128i vxk20x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk20x89ABCDEF), vk_zero_point);
599       i20 += 16;
600 
601 
602       const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
603       const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
604       const __m128i vprod20x89ABCDEFlo = _mm_mullo_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
605       const __m128i vprod20x89ABCDEFhi = _mm_mulhi_epi16(vxi20x89ABCDEF, vxk20x89ABCDEF);
606 
607       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
608       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
609       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod20x89ABCDEFlo, vprod20x89ABCDEFhi));
610       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod20x89ABCDEFlo, vprod20x89ABCDEFhi));
611 
612       const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
613       const __m128i vxi21x01234567 = _mm_cvtepu8_epi16(vi21x01234567);
614       const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(uint8_t)));
615       const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk21x01234567), vk_zero_point);
616       const __m128i vi21x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i21 + 8));
617       const __m128i vxi21x89ABCDEF = _mm_cvtepu8_epi16(vi21x89ABCDEF);
618       const __m128i vk21x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(uint8_t)));
619       const __m128i vxk21x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk21x89ABCDEF), vk_zero_point);
620       i21 += 16;
621 
622 
623       const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
624       const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
625       const __m128i vprod21x89ABCDEFlo = _mm_mullo_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
626       const __m128i vprod21x89ABCDEFhi = _mm_mulhi_epi16(vxi21x89ABCDEF, vxk21x89ABCDEF);
627 
628       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
629       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
630       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod21x89ABCDEFlo, vprod21x89ABCDEFhi));
631       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod21x89ABCDEFlo, vprod21x89ABCDEFhi));
632 
633       const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
634       const __m128i vxi22x01234567 = _mm_cvtepu8_epi16(vi22x01234567);
635       const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(uint8_t)));
636       const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk22x01234567), vk_zero_point);
637       const __m128i vi22x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i22 + 8));
638       const __m128i vxi22x89ABCDEF = _mm_cvtepu8_epi16(vi22x89ABCDEF);
639       const __m128i vk22x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(uint8_t)));
640       const __m128i vxk22x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk22x89ABCDEF), vk_zero_point);
641       i22 += 16;
642 
643 
644       const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
645       const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
646       const __m128i vprod22x89ABCDEFlo = _mm_mullo_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
647       const __m128i vprod22x89ABCDEFhi = _mm_mulhi_epi16(vxi22x89ABCDEF, vxk22x89ABCDEF);
648 
649       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
650       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
651       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod22x89ABCDEFlo, vprod22x89ABCDEFhi));
652       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod22x89ABCDEFlo, vprod22x89ABCDEFhi));
653 
654       const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
655       const __m128i vxi23x01234567 = _mm_cvtepu8_epi16(vi23x01234567);
656       const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(uint8_t)));
657       const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk23x01234567), vk_zero_point);
658       const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8));
659       const __m128i vxi23x89ABCDEF = _mm_cvtepu8_epi16(vi23x89ABCDEF);
660       const __m128i vk23x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(uint8_t)));
661       const __m128i vxk23x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk23x89ABCDEF), vk_zero_point);
662       i23 += 16;
663 
664 
665       const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
666       const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
667       const __m128i vprod23x89ABCDEFlo = _mm_mullo_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
668       const __m128i vprod23x89ABCDEFhi = _mm_mulhi_epi16(vxi23x89ABCDEF, vxk23x89ABCDEF);
669 
670       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
671       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
672       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod23x89ABCDEFlo, vprod23x89ABCDEFhi));
673       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod23x89ABCDEFlo, vprod23x89ABCDEFhi));
674 
675       const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
676       const __m128i vxi24x01234567 = _mm_cvtepu8_epi16(vi24x01234567);
677       const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(uint8_t)));
678       const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk24x01234567), vk_zero_point);
679       const __m128i vi24x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i24 + 8));
680       const __m128i vxi24x89ABCDEF = _mm_cvtepu8_epi16(vi24x89ABCDEF);
681       const __m128i vk24x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(uint8_t)));
682       const __m128i vxk24x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk24x89ABCDEF), vk_zero_point);
683       i24 += 16;
684 
685 
686       const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
687       const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
688       const __m128i vprod24x89ABCDEFlo = _mm_mullo_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
689       const __m128i vprod24x89ABCDEFhi = _mm_mulhi_epi16(vxi24x89ABCDEF, vxk24x89ABCDEF);
690 
691       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
692       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
693       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod24x89ABCDEFlo, vprod24x89ABCDEFhi));
694       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod24x89ABCDEFlo, vprod24x89ABCDEFhi));
695 
696       w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(uint8_t));
697 
698       __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
699       __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
700       __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
701       __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
702 
703       const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
704       vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
705       vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
706       vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
707       vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
708 
709       const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
710       vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
711       vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
712       vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
713       vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
714 
715       vacc0123 = _mm_cvtps_epi32(vscaled0123);
716       vacc4567 = _mm_cvtps_epi32(vscaled4567);
717       vacc89AB = _mm_cvtps_epi32(vscaled89AB);
718       vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
719 
720       const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
721       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
722       __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
723 
724       __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
725 
726       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
727       vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
728 
729       _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
730       output += 16;
731     }
732     if XNN_UNLIKELY(c != 0) {
733       const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
734       do {
735         __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
736         __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
737 
738 
739         const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
740         const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
741         const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
742         const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
743         i0 += 8;
744 
745 
746         const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
747         const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
748 
749         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
750         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
751 
752         const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
753         const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
754         const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
755         const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
756         i1 += 8;
757 
758 
759         const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
760         const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
761 
762         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
763         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
764 
765         const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
766         const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
767         const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
768         const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
769         i2 += 8;
770 
771 
772         const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
773         const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
774 
775         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
776         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
777 
778         const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
779         const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
780         const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
781         const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
782         i3 += 8;
783 
784 
785         const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
786         const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
787 
788         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
789         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
790 
791         const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
792         const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
793         const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
794         const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
795         i4 += 8;
796 
797 
798         const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
799         const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
800 
801         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
802         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
803 
804         const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
805         const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
806         const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
807         const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
808         i5 += 8;
809 
810 
811         const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
812         const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
813 
814         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
815         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
816 
817         const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
818         const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
819         const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
820         const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
821         i6 += 8;
822 
823 
824         const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
825         const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
826 
827         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
828         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
829 
830         const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
831         const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
832         const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
833         const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
834         i7 += 8;
835 
836 
837         const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
838         const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
839 
840         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
841         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
842 
843         const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
844         const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
845         const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
846         const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
847         i8 += 8;
848 
849 
850         const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
851         const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
852 
853         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
854         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
855 
856         const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
857         const __m128i vxi9x01234567 = _mm_cvtepu8_epi16(vi9x01234567);
858         const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
859         const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk9x01234567), vk_zero_point);
860         i9 += 8;
861 
862 
863         const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
864         const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567);
865 
866         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi));
867         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi));
868 
869         const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
870         const __m128i vxi10x01234567 = _mm_cvtepu8_epi16(vi10x01234567);
871         const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) (k + 160));
872         const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk10x01234567), vk_zero_point);
873         i10 += 8;
874 
875 
876         const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
877         const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567);
878 
879         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi));
880         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi));
881 
882         const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
883         const __m128i vxi11x01234567 = _mm_cvtepu8_epi16(vi11x01234567);
884         const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) (k + 176));
885         const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk11x01234567), vk_zero_point);
886         i11 += 8;
887 
888 
889         const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
890         const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567);
891 
892         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi));
893         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi));
894 
895         const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
896         const __m128i vxi12x01234567 = _mm_cvtepu8_epi16(vi12x01234567);
897         const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
898         const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk12x01234567), vk_zero_point);
899         i12 += 8;
900 
901 
902         const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
903         const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567);
904 
905         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi));
906         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi));
907 
908         const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
909         const __m128i vxi13x01234567 = _mm_cvtepu8_epi16(vi13x01234567);
910         const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) (k + 208));
911         const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk13x01234567), vk_zero_point);
912         i13 += 8;
913 
914 
915         const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
916         const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567);
917 
918         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi));
919         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi));
920 
921         const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
922         const __m128i vxi14x01234567 = _mm_cvtepu8_epi16(vi14x01234567);
923         const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) (k + 224));
924         const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk14x01234567), vk_zero_point);
925         i14 += 8;
926 
927 
928         const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
929         const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567);
930 
931         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi));
932         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi));
933 
934         const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
935         const __m128i vxi15x01234567 = _mm_cvtepu8_epi16(vi15x01234567);
936         const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) (k + 240));
937         const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk15x01234567), vk_zero_point);
938         i15 += 8;
939 
940 
941         const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
942         const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567);
943 
944         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi));
945         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi));
946 
947         const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
948         const __m128i vxi16x01234567 = _mm_cvtepu8_epi16(vi16x01234567);
949         const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) (k + 256));
950         const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk16x01234567), vk_zero_point);
951         i16 += 8;
952 
953 
954         const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
955         const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567);
956 
957         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi));
958         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi));
959 
960         const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
961         const __m128i vxi17x01234567 = _mm_cvtepu8_epi16(vi17x01234567);
962         const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) (k + 272));
963         const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk17x01234567), vk_zero_point);
964         i17 += 8;
965 
966 
967         const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
968         const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567);
969 
970         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi));
971         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi));
972 
973         const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
974         const __m128i vxi18x01234567 = _mm_cvtepu8_epi16(vi18x01234567);
975         const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) (k + 288));
976         const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk18x01234567), vk_zero_point);
977         i18 += 8;
978 
979 
980         const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
981         const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567);
982 
983         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi));
984         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi));
985 
986         const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
987         const __m128i vxi19x01234567 = _mm_cvtepu8_epi16(vi19x01234567);
988         const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) (k + 304));
989         const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk19x01234567), vk_zero_point);
990         i19 += 8;
991 
992 
993         const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
994         const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567);
995 
996         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi));
997         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi));
998 
999         const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
1000         const __m128i vxi20x01234567 = _mm_cvtepu8_epi16(vi20x01234567);
1001         const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) (k + 320));
1002         const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk20x01234567), vk_zero_point);
1003         i20 += 8;
1004 
1005 
1006         const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
1007         const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567);
1008 
1009         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi));
1010         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi));
1011 
1012         const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
1013         const __m128i vxi21x01234567 = _mm_cvtepu8_epi16(vi21x01234567);
1014         const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) (k + 336));
1015         const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk21x01234567), vk_zero_point);
1016         i21 += 8;
1017 
1018 
1019         const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
1020         const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567);
1021 
1022         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi));
1023         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi));
1024 
1025         const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
1026         const __m128i vxi22x01234567 = _mm_cvtepu8_epi16(vi22x01234567);
1027         const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) (k + 352));
1028         const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk22x01234567), vk_zero_point);
1029         i22 += 8;
1030 
1031 
1032         const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
1033         const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567);
1034 
1035         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi));
1036         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi));
1037 
1038         const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
1039         const __m128i vxi23x01234567 = _mm_cvtepu8_epi16(vi23x01234567);
1040         const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) (k + 368));
1041         const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk23x01234567), vk_zero_point);
1042         i23 += 8;
1043 
1044 
1045         const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
1046         const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567);
1047 
1048         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi));
1049         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi));
1050 
1051         const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
1052         const __m128i vxi24x01234567 = _mm_cvtepu8_epi16(vi24x01234567);
1053         const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) (k + 384));
1054         const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk24x01234567), vk_zero_point);
1055         i24 += 8;
1056 
1057 
1058         const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
1059         const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567);
1060 
1061         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi));
1062         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi));
1063 
1064         k += 8;
1065 
1066         __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
1067         __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
1068 
1069         const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
1070         vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
1071         vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
1072 
1073         const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
1074         vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
1075         vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1076 
1077         vacc0123 = _mm_cvtps_epi32(vscaled0123);
1078         vacc4567 = _mm_cvtps_epi32(vscaled4567);
1079 
1080         w = (const void*) ((const int32_t*) w + 8);
1081 
1082         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
1083         __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1084 
1085         __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
1086 
1087         vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
1088 
1089         if XNN_LIKELY(c >= 8) {
1090           _mm_storel_epi64((__m128i*) output, vout0123456701234567);
1091           output += 8;
1092           c -= 8;
1093         } else {
1094           if (c & 4) {
1095             unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
1096             vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
1097             output += 4;
1098           }
1099           if (c & 2) {
1100             unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
1101             vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1102             output += 2;
1103           }
1104           if (c & 1) {
1105             *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
1106             output += 1;
1107           }
1108           c = 0;
1109         }
1110       } while (c != 0);
1111     }
1112 
1113     output = (uint8_t*) ((uintptr_t) output + output_increment);
1114   } while (--output_width != 0);
1115 }
1116