xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up24x25-minmax-fp32-xop-mul32.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-sse-mul32.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 #if defined(__GNUC__) || defined(__clang__)
13   #include <x86intrin.h>
14 #else
15   #include <immintrin.h>
16   #include <ammintrin.h>
17 #endif
18 
19 #include <xnnpack/dwconv.h>
20 #include <xnnpack/intrinsics-polyfill.h>
21 #include <xnnpack/unaligned.h>
22 
23 
xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__xop_mul32(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)])24 void xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__xop_mul32(
25     size_t channels,
26     size_t output_width,
27     const int8_t** input,
28     const void* weights,
29     int8_t* output,
30     size_t input_stride,
31     size_t output_increment,
32     size_t input_offset,
33     const int8_t* zero,
34     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36   assert(channels != 0);
37   assert(output_width != 0);
38 
39   do {
40     const int8_t* i0 = input[0];
41     assert(i0 != NULL);
42     if XNN_UNPREDICTABLE(i0 != zero) {
43       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
44     }
45     const int8_t* i1 = input[1];
46     assert(i1 != NULL);
47     if XNN_UNPREDICTABLE(i1 != zero) {
48       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
49     }
50     const int8_t* i2 = input[2];
51     assert(i2 != NULL);
52     if XNN_UNPREDICTABLE(i2 != zero) {
53       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
54     }
55     const int8_t* i3 = input[3];
56     assert(i3 != NULL);
57     if XNN_UNPREDICTABLE(i3 != zero) {
58       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
59     }
60     const int8_t* i4 = input[4];
61     assert(i4 != NULL);
62     if XNN_UNPREDICTABLE(i4 != zero) {
63       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
64     }
65     const int8_t* i5 = input[5];
66     assert(i5 != NULL);
67     if XNN_UNPREDICTABLE(i5 != zero) {
68       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
69     }
70     const int8_t* i6 = input[6];
71     assert(i6 != NULL);
72     if XNN_UNPREDICTABLE(i6 != zero) {
73       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
74     }
75     const int8_t* i7 = input[7];
76     assert(i7 != NULL);
77     if XNN_UNPREDICTABLE(i7 != zero) {
78       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
79     }
80     const int8_t* i8 = input[8];
81     assert(i8 != NULL);
82     if XNN_UNPREDICTABLE(i8 != zero) {
83       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
84     }
85     const int8_t* i9 = input[9];
86     assert(i9 != NULL);
87     if XNN_UNPREDICTABLE(i9 != zero) {
88       i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
89     }
90     const int8_t* i10 = input[10];
91     assert(i10 != NULL);
92     if XNN_UNPREDICTABLE(i10 != zero) {
93       i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
94     }
95     const int8_t* i11 = input[11];
96     assert(i11 != NULL);
97     if XNN_UNPREDICTABLE(i11 != zero) {
98       i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
99     }
100     const int8_t* i12 = input[12];
101     assert(i12 != NULL);
102     if XNN_UNPREDICTABLE(i12 != zero) {
103       i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
104     }
105     const int8_t* i13 = input[13];
106     assert(i13 != NULL);
107     if XNN_UNPREDICTABLE(i13 != zero) {
108       i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
109     }
110     const int8_t* i14 = input[14];
111     assert(i14 != NULL);
112     if XNN_UNPREDICTABLE(i14 != zero) {
113       i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
114     }
115     const int8_t* i15 = input[15];
116     assert(i15 != NULL);
117     if XNN_UNPREDICTABLE(i15 != zero) {
118       i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
119     }
120     const int8_t* i16 = input[16];
121     assert(i16 != NULL);
122     if XNN_UNPREDICTABLE(i16 != zero) {
123       i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
124     }
125     const int8_t* i17 = input[17];
126     assert(i17 != NULL);
127     if XNN_UNPREDICTABLE(i17 != zero) {
128       i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
129     }
130     const int8_t* i18 = input[18];
131     assert(i18 != NULL);
132     if XNN_UNPREDICTABLE(i18 != zero) {
133       i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
134     }
135     const int8_t* i19 = input[19];
136     assert(i19 != NULL);
137     if XNN_UNPREDICTABLE(i19 != zero) {
138       i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
139     }
140     const int8_t* i20 = input[20];
141     assert(i20 != NULL);
142     if XNN_UNPREDICTABLE(i20 != zero) {
143       i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
144     }
145     const int8_t* i21 = input[21];
146     assert(i21 != NULL);
147     if XNN_UNPREDICTABLE(i21 != zero) {
148       i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
149     }
150     const int8_t* i22 = input[22];
151     assert(i22 != NULL);
152     if XNN_UNPREDICTABLE(i22 != zero) {
153       i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
154     }
155     const int8_t* i23 = input[23];
156     assert(i23 != NULL);
157     if XNN_UNPREDICTABLE(i23 != zero) {
158       i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
159     }
160     const int8_t* i24 = input[24];
161     assert(i24 != NULL);
162     if XNN_UNPREDICTABLE(i24 != zero) {
163       i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
164     }
165     input = (const int8_t**) ((uintptr_t) input + input_stride);
166 
167     size_t c = channels;
168     const void* w = weights;
169     for (; c >= 24; c -= 24) {
170       __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
171       __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
172       __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
173       __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
174       __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
175       __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
176 
177 
178       const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
179       const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
180       const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
181       const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
182       const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
183       const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
184       const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
185       const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
186       const __m128i vi0xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 16)));
187       const __m128i vk0xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
188       const __m128i vi0xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 20)));
189       const __m128i vk0xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
190       i0 += 24;
191 
192       vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
193       vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
194       vacc89AB = _mm_macc_epi32(vi0x89AB, vk0x89AB, vacc89AB);
195       vaccCDEF = _mm_macc_epi32(vi0xCDEF, vk0xCDEF, vaccCDEF);
196       vaccGHIJ = _mm_macc_epi32(vi0xGHIJ, vk0xGHIJ, vaccGHIJ);
197       vaccKLMN = _mm_macc_epi32(vi0xKLMN, vk0xKLMN, vaccKLMN);
198 
199       const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
200       const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
201       const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
202       const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
203       const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
204       const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
205       const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
206       const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
207       const __m128i vi1xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 16)));
208       const __m128i vk1xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
209       const __m128i vi1xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 20)));
210       const __m128i vk1xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
211       i1 += 24;
212 
213       vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
214       vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
215       vacc89AB = _mm_macc_epi32(vi1x89AB, vk1x89AB, vacc89AB);
216       vaccCDEF = _mm_macc_epi32(vi1xCDEF, vk1xCDEF, vaccCDEF);
217       vaccGHIJ = _mm_macc_epi32(vi1xGHIJ, vk1xGHIJ, vaccGHIJ);
218       vaccKLMN = _mm_macc_epi32(vi1xKLMN, vk1xKLMN, vaccKLMN);
219 
220       const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
221       const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)))));
222       const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
223       const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 52 * sizeof(int8_t)))));
224       const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
225       const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)))));
226       const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
227       const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 60 * sizeof(int8_t)))));
228       const __m128i vi2xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 16)));
229       const __m128i vk2xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)))));
230       const __m128i vi2xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 20)));
231       const __m128i vk2xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 68 * sizeof(int8_t)))));
232       i2 += 24;
233 
234       vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
235       vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
236       vacc89AB = _mm_macc_epi32(vi2x89AB, vk2x89AB, vacc89AB);
237       vaccCDEF = _mm_macc_epi32(vi2xCDEF, vk2xCDEF, vaccCDEF);
238       vaccGHIJ = _mm_macc_epi32(vi2xGHIJ, vk2xGHIJ, vaccGHIJ);
239       vaccKLMN = _mm_macc_epi32(vi2xKLMN, vk2xKLMN, vaccKLMN);
240 
241       const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
242       const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)))));
243       const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
244       const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 76 * sizeof(int8_t)))));
245       const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
246       const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)))));
247       const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
248       const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 84 * sizeof(int8_t)))));
249       const __m128i vi3xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 16)));
250       const __m128i vk3xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)))));
251       const __m128i vi3xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 20)));
252       const __m128i vk3xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 92 * sizeof(int8_t)))));
253       i3 += 24;
254 
255       vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
256       vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
257       vacc89AB = _mm_macc_epi32(vi3x89AB, vk3x89AB, vacc89AB);
258       vaccCDEF = _mm_macc_epi32(vi3xCDEF, vk3xCDEF, vaccCDEF);
259       vaccGHIJ = _mm_macc_epi32(vi3xGHIJ, vk3xGHIJ, vaccGHIJ);
260       vaccKLMN = _mm_macc_epi32(vi3xKLMN, vk3xKLMN, vaccKLMN);
261 
262       const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
263       const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)))));
264       const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
265       const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 100 * sizeof(int8_t)))));
266       const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
267       const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)))));
268       const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
269       const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 108 * sizeof(int8_t)))));
270       const __m128i vi4xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 16)));
271       const __m128i vk4xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)))));
272       const __m128i vi4xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 20)));
273       const __m128i vk4xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 116 * sizeof(int8_t)))));
274       i4 += 24;
275 
276       vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
277       vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
278       vacc89AB = _mm_macc_epi32(vi4x89AB, vk4x89AB, vacc89AB);
279       vaccCDEF = _mm_macc_epi32(vi4xCDEF, vk4xCDEF, vaccCDEF);
280       vaccGHIJ = _mm_macc_epi32(vi4xGHIJ, vk4xGHIJ, vaccGHIJ);
281       vaccKLMN = _mm_macc_epi32(vi4xKLMN, vk4xKLMN, vaccKLMN);
282 
283       const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
284       const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)))));
285       const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
286       const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 124 * sizeof(int8_t)))));
287       const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
288       const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)))));
289       const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
290       const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 132 * sizeof(int8_t)))));
291       const __m128i vi5xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 16)));
292       const __m128i vk5xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)))));
293       const __m128i vi5xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 20)));
294       const __m128i vk5xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 140 * sizeof(int8_t)))));
295       i5 += 24;
296 
297       vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
298       vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
299       vacc89AB = _mm_macc_epi32(vi5x89AB, vk5x89AB, vacc89AB);
300       vaccCDEF = _mm_macc_epi32(vi5xCDEF, vk5xCDEF, vaccCDEF);
301       vaccGHIJ = _mm_macc_epi32(vi5xGHIJ, vk5xGHIJ, vaccGHIJ);
302       vaccKLMN = _mm_macc_epi32(vi5xKLMN, vk5xKLMN, vaccKLMN);
303 
304       const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
305       const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)))));
306       const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
307       const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 148 * sizeof(int8_t)))));
308       const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
309       const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)))));
310       const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
311       const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 156 * sizeof(int8_t)))));
312       const __m128i vi6xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 16)));
313       const __m128i vk6xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)))));
314       const __m128i vi6xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 20)));
315       const __m128i vk6xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 164 * sizeof(int8_t)))));
316       i6 += 24;
317 
318       vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
319       vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
320       vacc89AB = _mm_macc_epi32(vi6x89AB, vk6x89AB, vacc89AB);
321       vaccCDEF = _mm_macc_epi32(vi6xCDEF, vk6xCDEF, vaccCDEF);
322       vaccGHIJ = _mm_macc_epi32(vi6xGHIJ, vk6xGHIJ, vaccGHIJ);
323       vaccKLMN = _mm_macc_epi32(vi6xKLMN, vk6xKLMN, vaccKLMN);
324 
325       const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
326       const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)))));
327       const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
328       const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 172 * sizeof(int8_t)))));
329       const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 8)));
330       const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)))));
331       const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 12)));
332       const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 180 * sizeof(int8_t)))));
333       const __m128i vi7xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 16)));
334       const __m128i vk7xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)))));
335       const __m128i vi7xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 20)));
336       const __m128i vk7xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 188 * sizeof(int8_t)))));
337       i7 += 24;
338 
339       vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
340       vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
341       vacc89AB = _mm_macc_epi32(vi7x89AB, vk7x89AB, vacc89AB);
342       vaccCDEF = _mm_macc_epi32(vi7xCDEF, vk7xCDEF, vaccCDEF);
343       vaccGHIJ = _mm_macc_epi32(vi7xGHIJ, vk7xGHIJ, vaccGHIJ);
344       vaccKLMN = _mm_macc_epi32(vi7xKLMN, vk7xKLMN, vaccKLMN);
345 
346       const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
347       const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)))));
348       const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
349       const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 196 * sizeof(int8_t)))));
350       const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 8)));
351       const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)))));
352       const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 12)));
353       const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 204 * sizeof(int8_t)))));
354       const __m128i vi8xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 16)));
355       const __m128i vk8xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)))));
356       const __m128i vi8xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 20)));
357       const __m128i vk8xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 212 * sizeof(int8_t)))));
358       i8 += 24;
359 
360       vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
361       vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
362       vacc89AB = _mm_macc_epi32(vi8x89AB, vk8x89AB, vacc89AB);
363       vaccCDEF = _mm_macc_epi32(vi8xCDEF, vk8xCDEF, vaccCDEF);
364       vaccGHIJ = _mm_macc_epi32(vi8xGHIJ, vk8xGHIJ, vaccGHIJ);
365       vaccKLMN = _mm_macc_epi32(vi8xKLMN, vk8xKLMN, vaccKLMN);
366 
367       const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
368       const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)))));
369       const __m128i vi9x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 4)));
370       const __m128i vk9x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 220 * sizeof(int8_t)))));
371       const __m128i vi9x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 8)));
372       const __m128i vk9x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 224 * sizeof(int8_t)))));
373       const __m128i vi9xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 12)));
374       const __m128i vk9xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 228 * sizeof(int8_t)))));
375       const __m128i vi9xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 16)));
376       const __m128i vk9xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 232 * sizeof(int8_t)))));
377       const __m128i vi9xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 20)));
378       const __m128i vk9xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 236 * sizeof(int8_t)))));
379       i9 += 24;
380 
381       vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
382       vacc4567 = _mm_macc_epi32(vi9x4567, vk9x4567, vacc4567);
383       vacc89AB = _mm_macc_epi32(vi9x89AB, vk9x89AB, vacc89AB);
384       vaccCDEF = _mm_macc_epi32(vi9xCDEF, vk9xCDEF, vaccCDEF);
385       vaccGHIJ = _mm_macc_epi32(vi9xGHIJ, vk9xGHIJ, vaccGHIJ);
386       vaccKLMN = _mm_macc_epi32(vi9xKLMN, vk9xKLMN, vaccKLMN);
387 
388       const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
389       const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 240 * sizeof(int8_t)))));
390       const __m128i vi10x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 4)));
391       const __m128i vk10x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 244 * sizeof(int8_t)))));
392       const __m128i vi10x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 8)));
393       const __m128i vk10x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 248 * sizeof(int8_t)))));
394       const __m128i vi10xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 12)));
395       const __m128i vk10xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 252 * sizeof(int8_t)))));
396       const __m128i vi10xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 16)));
397       const __m128i vk10xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 256 * sizeof(int8_t)))));
398       const __m128i vi10xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 20)));
399       const __m128i vk10xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 260 * sizeof(int8_t)))));
400       i10 += 24;
401 
402       vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
403       vacc4567 = _mm_macc_epi32(vi10x4567, vk10x4567, vacc4567);
404       vacc89AB = _mm_macc_epi32(vi10x89AB, vk10x89AB, vacc89AB);
405       vaccCDEF = _mm_macc_epi32(vi10xCDEF, vk10xCDEF, vaccCDEF);
406       vaccGHIJ = _mm_macc_epi32(vi10xGHIJ, vk10xGHIJ, vaccGHIJ);
407       vaccKLMN = _mm_macc_epi32(vi10xKLMN, vk10xKLMN, vaccKLMN);
408 
409       const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
410       const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 264 * sizeof(int8_t)))));
411       const __m128i vi11x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 4)));
412       const __m128i vk11x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 268 * sizeof(int8_t)))));
413       const __m128i vi11x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 8)));
414       const __m128i vk11x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 272 * sizeof(int8_t)))));
415       const __m128i vi11xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 12)));
416       const __m128i vk11xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 276 * sizeof(int8_t)))));
417       const __m128i vi11xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 16)));
418       const __m128i vk11xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 280 * sizeof(int8_t)))));
419       const __m128i vi11xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 20)));
420       const __m128i vk11xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 284 * sizeof(int8_t)))));
421       i11 += 24;
422 
423       vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
424       vacc4567 = _mm_macc_epi32(vi11x4567, vk11x4567, vacc4567);
425       vacc89AB = _mm_macc_epi32(vi11x89AB, vk11x89AB, vacc89AB);
426       vaccCDEF = _mm_macc_epi32(vi11xCDEF, vk11xCDEF, vaccCDEF);
427       vaccGHIJ = _mm_macc_epi32(vi11xGHIJ, vk11xGHIJ, vaccGHIJ);
428       vaccKLMN = _mm_macc_epi32(vi11xKLMN, vk11xKLMN, vaccKLMN);
429 
430       const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
431       const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 288 * sizeof(int8_t)))));
432       const __m128i vi12x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 4)));
433       const __m128i vk12x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 292 * sizeof(int8_t)))));
434       const __m128i vi12x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 8)));
435       const __m128i vk12x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 296 * sizeof(int8_t)))));
436       const __m128i vi12xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 12)));
437       const __m128i vk12xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 300 * sizeof(int8_t)))));
438       const __m128i vi12xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 16)));
439       const __m128i vk12xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 304 * sizeof(int8_t)))));
440       const __m128i vi12xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 20)));
441       const __m128i vk12xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 308 * sizeof(int8_t)))));
442       i12 += 24;
443 
444       vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
445       vacc4567 = _mm_macc_epi32(vi12x4567, vk12x4567, vacc4567);
446       vacc89AB = _mm_macc_epi32(vi12x89AB, vk12x89AB, vacc89AB);
447       vaccCDEF = _mm_macc_epi32(vi12xCDEF, vk12xCDEF, vaccCDEF);
448       vaccGHIJ = _mm_macc_epi32(vi12xGHIJ, vk12xGHIJ, vaccGHIJ);
449       vaccKLMN = _mm_macc_epi32(vi12xKLMN, vk12xKLMN, vaccKLMN);
450 
451       const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
452       const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 312 * sizeof(int8_t)))));
453       const __m128i vi13x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 4)));
454       const __m128i vk13x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 316 * sizeof(int8_t)))));
455       const __m128i vi13x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 8)));
456       const __m128i vk13x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 320 * sizeof(int8_t)))));
457       const __m128i vi13xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 12)));
458       const __m128i vk13xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 324 * sizeof(int8_t)))));
459       const __m128i vi13xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 16)));
460       const __m128i vk13xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 328 * sizeof(int8_t)))));
461       const __m128i vi13xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 20)));
462       const __m128i vk13xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 332 * sizeof(int8_t)))));
463       i13 += 24;
464 
465       vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
466       vacc4567 = _mm_macc_epi32(vi13x4567, vk13x4567, vacc4567);
467       vacc89AB = _mm_macc_epi32(vi13x89AB, vk13x89AB, vacc89AB);
468       vaccCDEF = _mm_macc_epi32(vi13xCDEF, vk13xCDEF, vaccCDEF);
469       vaccGHIJ = _mm_macc_epi32(vi13xGHIJ, vk13xGHIJ, vaccGHIJ);
470       vaccKLMN = _mm_macc_epi32(vi13xKLMN, vk13xKLMN, vaccKLMN);
471 
472       const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
473       const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 336 * sizeof(int8_t)))));
474       const __m128i vi14x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 4)));
475       const __m128i vk14x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 340 * sizeof(int8_t)))));
476       const __m128i vi14x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 8)));
477       const __m128i vk14x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 344 * sizeof(int8_t)))));
478       const __m128i vi14xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 12)));
479       const __m128i vk14xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 348 * sizeof(int8_t)))));
480       const __m128i vi14xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 16)));
481       const __m128i vk14xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 352 * sizeof(int8_t)))));
482       const __m128i vi14xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 20)));
483       const __m128i vk14xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 356 * sizeof(int8_t)))));
484       i14 += 24;
485 
486       vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
487       vacc4567 = _mm_macc_epi32(vi14x4567, vk14x4567, vacc4567);
488       vacc89AB = _mm_macc_epi32(vi14x89AB, vk14x89AB, vacc89AB);
489       vaccCDEF = _mm_macc_epi32(vi14xCDEF, vk14xCDEF, vaccCDEF);
490       vaccGHIJ = _mm_macc_epi32(vi14xGHIJ, vk14xGHIJ, vaccGHIJ);
491       vaccKLMN = _mm_macc_epi32(vi14xKLMN, vk14xKLMN, vaccKLMN);
492 
493       const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
494       const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 360 * sizeof(int8_t)))));
495       const __m128i vi15x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 4)));
496       const __m128i vk15x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 364 * sizeof(int8_t)))));
497       const __m128i vi15x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 8)));
498       const __m128i vk15x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 368 * sizeof(int8_t)))));
499       const __m128i vi15xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 12)));
500       const __m128i vk15xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 372 * sizeof(int8_t)))));
501       const __m128i vi15xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 16)));
502       const __m128i vk15xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 376 * sizeof(int8_t)))));
503       const __m128i vi15xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 20)));
504       const __m128i vk15xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 380 * sizeof(int8_t)))));
505       i15 += 24;
506 
507       vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
508       vacc4567 = _mm_macc_epi32(vi15x4567, vk15x4567, vacc4567);
509       vacc89AB = _mm_macc_epi32(vi15x89AB, vk15x89AB, vacc89AB);
510       vaccCDEF = _mm_macc_epi32(vi15xCDEF, vk15xCDEF, vaccCDEF);
511       vaccGHIJ = _mm_macc_epi32(vi15xGHIJ, vk15xGHIJ, vaccGHIJ);
512       vaccKLMN = _mm_macc_epi32(vi15xKLMN, vk15xKLMN, vaccKLMN);
513 
514       const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
515       const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 384 * sizeof(int8_t)))));
516       const __m128i vi16x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 4)));
517       const __m128i vk16x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 388 * sizeof(int8_t)))));
518       const __m128i vi16x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 8)));
519       const __m128i vk16x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 392 * sizeof(int8_t)))));
520       const __m128i vi16xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 12)));
521       const __m128i vk16xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 396 * sizeof(int8_t)))));
522       const __m128i vi16xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 16)));
523       const __m128i vk16xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 400 * sizeof(int8_t)))));
524       const __m128i vi16xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 20)));
525       const __m128i vk16xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 404 * sizeof(int8_t)))));
526       i16 += 24;
527 
528       vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
529       vacc4567 = _mm_macc_epi32(vi16x4567, vk16x4567, vacc4567);
530       vacc89AB = _mm_macc_epi32(vi16x89AB, vk16x89AB, vacc89AB);
531       vaccCDEF = _mm_macc_epi32(vi16xCDEF, vk16xCDEF, vaccCDEF);
532       vaccGHIJ = _mm_macc_epi32(vi16xGHIJ, vk16xGHIJ, vaccGHIJ);
533       vaccKLMN = _mm_macc_epi32(vi16xKLMN, vk16xKLMN, vaccKLMN);
534 
535       const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
536       const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 408 * sizeof(int8_t)))));
537       const __m128i vi17x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 4)));
538       const __m128i vk17x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 412 * sizeof(int8_t)))));
539       const __m128i vi17x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 8)));
540       const __m128i vk17x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 416 * sizeof(int8_t)))));
541       const __m128i vi17xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 12)));
542       const __m128i vk17xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 420 * sizeof(int8_t)))));
543       const __m128i vi17xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 16)));
544       const __m128i vk17xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 424 * sizeof(int8_t)))));
545       const __m128i vi17xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 20)));
546       const __m128i vk17xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 428 * sizeof(int8_t)))));
547       i17 += 24;
548 
549       vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
550       vacc4567 = _mm_macc_epi32(vi17x4567, vk17x4567, vacc4567);
551       vacc89AB = _mm_macc_epi32(vi17x89AB, vk17x89AB, vacc89AB);
552       vaccCDEF = _mm_macc_epi32(vi17xCDEF, vk17xCDEF, vaccCDEF);
553       vaccGHIJ = _mm_macc_epi32(vi17xGHIJ, vk17xGHIJ, vaccGHIJ);
554       vaccKLMN = _mm_macc_epi32(vi17xKLMN, vk17xKLMN, vaccKLMN);
555 
556       const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
557       const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 432 * sizeof(int8_t)))));
558       const __m128i vi18x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 4)));
559       const __m128i vk18x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 436 * sizeof(int8_t)))));
560       const __m128i vi18x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 8)));
561       const __m128i vk18x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 440 * sizeof(int8_t)))));
562       const __m128i vi18xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 12)));
563       const __m128i vk18xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 444 * sizeof(int8_t)))));
564       const __m128i vi18xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 16)));
565       const __m128i vk18xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 448 * sizeof(int8_t)))));
566       const __m128i vi18xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 20)));
567       const __m128i vk18xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 452 * sizeof(int8_t)))));
568       i18 += 24;
569 
570       vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
571       vacc4567 = _mm_macc_epi32(vi18x4567, vk18x4567, vacc4567);
572       vacc89AB = _mm_macc_epi32(vi18x89AB, vk18x89AB, vacc89AB);
573       vaccCDEF = _mm_macc_epi32(vi18xCDEF, vk18xCDEF, vaccCDEF);
574       vaccGHIJ = _mm_macc_epi32(vi18xGHIJ, vk18xGHIJ, vaccGHIJ);
575       vaccKLMN = _mm_macc_epi32(vi18xKLMN, vk18xKLMN, vaccKLMN);
576 
577       const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
578       const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 456 * sizeof(int8_t)))));
579       const __m128i vi19x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 4)));
580       const __m128i vk19x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 460 * sizeof(int8_t)))));
581       const __m128i vi19x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 8)));
582       const __m128i vk19x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 464 * sizeof(int8_t)))));
583       const __m128i vi19xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 12)));
584       const __m128i vk19xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 468 * sizeof(int8_t)))));
585       const __m128i vi19xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 16)));
586       const __m128i vk19xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 472 * sizeof(int8_t)))));
587       const __m128i vi19xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 20)));
588       const __m128i vk19xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 476 * sizeof(int8_t)))));
589       i19 += 24;
590 
591       vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
592       vacc4567 = _mm_macc_epi32(vi19x4567, vk19x4567, vacc4567);
593       vacc89AB = _mm_macc_epi32(vi19x89AB, vk19x89AB, vacc89AB);
594       vaccCDEF = _mm_macc_epi32(vi19xCDEF, vk19xCDEF, vaccCDEF);
595       vaccGHIJ = _mm_macc_epi32(vi19xGHIJ, vk19xGHIJ, vaccGHIJ);
596       vaccKLMN = _mm_macc_epi32(vi19xKLMN, vk19xKLMN, vaccKLMN);
597 
598       const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
599       const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 480 * sizeof(int8_t)))));
600       const __m128i vi20x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 4)));
601       const __m128i vk20x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 484 * sizeof(int8_t)))));
602       const __m128i vi20x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 8)));
603       const __m128i vk20x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 488 * sizeof(int8_t)))));
604       const __m128i vi20xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 12)));
605       const __m128i vk20xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 492 * sizeof(int8_t)))));
606       const __m128i vi20xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 16)));
607       const __m128i vk20xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 496 * sizeof(int8_t)))));
608       const __m128i vi20xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 20)));
609       const __m128i vk20xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 500 * sizeof(int8_t)))));
610       i20 += 24;
611 
612       vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
613       vacc4567 = _mm_macc_epi32(vi20x4567, vk20x4567, vacc4567);
614       vacc89AB = _mm_macc_epi32(vi20x89AB, vk20x89AB, vacc89AB);
615       vaccCDEF = _mm_macc_epi32(vi20xCDEF, vk20xCDEF, vaccCDEF);
616       vaccGHIJ = _mm_macc_epi32(vi20xGHIJ, vk20xGHIJ, vaccGHIJ);
617       vaccKLMN = _mm_macc_epi32(vi20xKLMN, vk20xKLMN, vaccKLMN);
618 
619       const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
620       const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 504 * sizeof(int8_t)))));
621       const __m128i vi21x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 4)));
622       const __m128i vk21x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 508 * sizeof(int8_t)))));
623       const __m128i vi21x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 8)));
624       const __m128i vk21x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 512 * sizeof(int8_t)))));
625       const __m128i vi21xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 12)));
626       const __m128i vk21xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 516 * sizeof(int8_t)))));
627       const __m128i vi21xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 16)));
628       const __m128i vk21xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 520 * sizeof(int8_t)))));
629       const __m128i vi21xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 20)));
630       const __m128i vk21xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 524 * sizeof(int8_t)))));
631       i21 += 24;
632 
633       vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
634       vacc4567 = _mm_macc_epi32(vi21x4567, vk21x4567, vacc4567);
635       vacc89AB = _mm_macc_epi32(vi21x89AB, vk21x89AB, vacc89AB);
636       vaccCDEF = _mm_macc_epi32(vi21xCDEF, vk21xCDEF, vaccCDEF);
637       vaccGHIJ = _mm_macc_epi32(vi21xGHIJ, vk21xGHIJ, vaccGHIJ);
638       vaccKLMN = _mm_macc_epi32(vi21xKLMN, vk21xKLMN, vaccKLMN);
639 
640       const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
641       const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 528 * sizeof(int8_t)))));
642       const __m128i vi22x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 4)));
643       const __m128i vk22x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 532 * sizeof(int8_t)))));
644       const __m128i vi22x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 8)));
645       const __m128i vk22x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 536 * sizeof(int8_t)))));
646       const __m128i vi22xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 12)));
647       const __m128i vk22xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 540 * sizeof(int8_t)))));
648       const __m128i vi22xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 16)));
649       const __m128i vk22xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 544 * sizeof(int8_t)))));
650       const __m128i vi22xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 20)));
651       const __m128i vk22xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 548 * sizeof(int8_t)))));
652       i22 += 24;
653 
654       vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
655       vacc4567 = _mm_macc_epi32(vi22x4567, vk22x4567, vacc4567);
656       vacc89AB = _mm_macc_epi32(vi22x89AB, vk22x89AB, vacc89AB);
657       vaccCDEF = _mm_macc_epi32(vi22xCDEF, vk22xCDEF, vaccCDEF);
658       vaccGHIJ = _mm_macc_epi32(vi22xGHIJ, vk22xGHIJ, vaccGHIJ);
659       vaccKLMN = _mm_macc_epi32(vi22xKLMN, vk22xKLMN, vaccKLMN);
660 
661       const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
662       const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 552 * sizeof(int8_t)))));
663       const __m128i vi23x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 4)));
664       const __m128i vk23x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 556 * sizeof(int8_t)))));
665       const __m128i vi23x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 8)));
666       const __m128i vk23x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 560 * sizeof(int8_t)))));
667       const __m128i vi23xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 12)));
668       const __m128i vk23xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 564 * sizeof(int8_t)))));
669       const __m128i vi23xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 16)));
670       const __m128i vk23xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 568 * sizeof(int8_t)))));
671       const __m128i vi23xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 20)));
672       const __m128i vk23xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 572 * sizeof(int8_t)))));
673       i23 += 24;
674 
675       vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
676       vacc4567 = _mm_macc_epi32(vi23x4567, vk23x4567, vacc4567);
677       vacc89AB = _mm_macc_epi32(vi23x89AB, vk23x89AB, vacc89AB);
678       vaccCDEF = _mm_macc_epi32(vi23xCDEF, vk23xCDEF, vaccCDEF);
679       vaccGHIJ = _mm_macc_epi32(vi23xGHIJ, vk23xGHIJ, vaccGHIJ);
680       vaccKLMN = _mm_macc_epi32(vi23xKLMN, vk23xKLMN, vaccKLMN);
681 
682       const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
683       const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 576 * sizeof(int8_t)))));
684       const __m128i vi24x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 4)));
685       const __m128i vk24x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 580 * sizeof(int8_t)))));
686       const __m128i vi24x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 8)));
687       const __m128i vk24x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 584 * sizeof(int8_t)))));
688       const __m128i vi24xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 12)));
689       const __m128i vk24xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 588 * sizeof(int8_t)))));
690       const __m128i vi24xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 16)));
691       const __m128i vk24xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 592 * sizeof(int8_t)))));
692       const __m128i vi24xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 20)));
693       const __m128i vk24xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 596 * sizeof(int8_t)))));
694       i24 += 24;
695 
696       vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
697       vacc4567 = _mm_macc_epi32(vi24x4567, vk24x4567, vacc4567);
698       vacc89AB = _mm_macc_epi32(vi24x89AB, vk24x89AB, vacc89AB);
699       vaccCDEF = _mm_macc_epi32(vi24xCDEF, vk24xCDEF, vaccCDEF);
700       vaccGHIJ = _mm_macc_epi32(vi24xGHIJ, vk24xGHIJ, vaccGHIJ);
701       vaccKLMN = _mm_macc_epi32(vi24xKLMN, vk24xKLMN, vaccKLMN);
702 
703       w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t));
704 
705       __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
706       __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
707       __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
708       __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
709       __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
710       __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
711 
712       const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
713       vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
714       vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
715       vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
716       vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
717       vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscale);
718       vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscale);
719 
720       const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
721       vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
722       vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
723       vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
724       vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
725       vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
726       vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
727 
728       vacc0123 = _mm_cvtps_epi32(vscaled0123);
729       vacc4567 = _mm_cvtps_epi32(vscaled4567);
730       vacc89AB = _mm_cvtps_epi32(vscaled89AB);
731       vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
732       vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
733       vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
734 
735       const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
736       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
737       __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
738       __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
739 
740       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
741       __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
742       vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
743       __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
744       voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
745 
746       _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
747       _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
748       output += 24;
749     }
750     if XNN_UNLIKELY(c != 0) {
751       const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
752       do {
753         __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
754 
755         const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
756         const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) k)));
757         i0 += 4;
758 
759         vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
760         const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
761         const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 24))));
762         i1 += 4;
763 
764         vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
765         const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
766         const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48))));
767         i2 += 4;
768 
769         vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
770         const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
771         const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 72))));
772         i3 += 4;
773 
774         vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
775         const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
776         const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96))));
777         i4 += 4;
778 
779         vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
780         const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
781         const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 120))));
782         i5 += 4;
783 
784         vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
785         const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
786         const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 144))));
787         i6 += 4;
788 
789         vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
790         const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
791         const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 168))));
792         i7 += 4;
793 
794         vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
795         const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
796         const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 192))));
797         i8 += 4;
798 
799         vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
800         const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
801         const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 216))));
802         i9 += 4;
803 
804         vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
805         const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
806         const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 240))));
807         i10 += 4;
808 
809         vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
810         const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
811         const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 264))));
812         i11 += 4;
813 
814         vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
815         const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
816         const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 288))));
817         i12 += 4;
818 
819         vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
820         const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
821         const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 312))));
822         i13 += 4;
823 
824         vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
825         const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
826         const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 336))));
827         i14 += 4;
828 
829         vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
830         const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
831         const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 360))));
832         i15 += 4;
833 
834         vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
835         const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
836         const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 384))));
837         i16 += 4;
838 
839         vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
840         const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
841         const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 408))));
842         i17 += 4;
843 
844         vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
845         const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
846         const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 432))));
847         i18 += 4;
848 
849         vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
850         const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
851         const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 456))));
852         i19 += 4;
853 
854         vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
855         const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
856         const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 480))));
857         i20 += 4;
858 
859         vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
860         const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
861         const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 504))));
862         i21 += 4;
863 
864         vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
865         const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
866         const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 528))));
867         i22 += 4;
868 
869         vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
870         const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
871         const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 552))));
872         i23 += 4;
873 
874         vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
875         const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
876         const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 576))));
877         i24 += 4;
878 
879         vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
880 
881         k += 4;
882 
883         __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
884         vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
885         vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
886         vacc0123 = _mm_cvtps_epi32(vscaled0123);
887 
888         w = (const void*) ((const int32_t*) w + 4);
889 
890         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
891         __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
892 
893         vout0123 = _mm_packs_epi16(vout0123, vout0123);
894         vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
895 
896         if XNN_LIKELY(c >= 4) {
897           _mm_storeu_si32(output, vout0123);
898           output += 4;
899           c -= 4;
900         } else {
901           if (c & 2) {
902             unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
903             vout0123 = _mm_srli_epi32(vout0123, 16);
904             output += 2;
905           }
906           if (c & 1) {
907             *output = (int8_t) _mm_extract_epi8(vout0123, 0);
908             output += 1;
909           }
910           c = 0;
911         }
912       } while (c != 0);
913     }
914 
915     output = (int8_t*) ((uintptr_t) output + output_increment);
916   } while (--output_width != 0);
917 }
918