Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-25p16c-minmax-fp32-sse41-mul32.c

raw
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 #include <immintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_dwconv_minmax_fp32_ukernel_25p16c__sse41_mul32(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 intptr_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(channels != 0);
32 assert(output_width != 0);
33
34
35 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
36 XNN_FORCE_REALIZATION(vscale);
37 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
38 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
39 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
40 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
41 XNN_FORCE_REALIZATION(voutput_zero_point);
42 XNN_FORCE_REALIZATION(voutput_min);
43
44 do {
45 const int8_t* i0 = input[0];
46 assert(i0 != NULL);
47 if XNN_UNPREDICTABLE(i0 != zero) {
48 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
49 }
50 const int8_t* i1 = input[1];
51 assert(i1 != NULL);
52 if XNN_UNPREDICTABLE(i1 != zero) {
53 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
54 }
55 const int8_t* i2 = input[2];
56 assert(i2 != NULL);
57 if XNN_UNPREDICTABLE(i2 != zero) {
58 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
59 }
60 const int8_t* i3 = input[3];
61 assert(i3 != NULL);
62 if XNN_UNPREDICTABLE(i3 != zero) {
63 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
64 }
65 const int8_t* i4 = input[4];
66 assert(i4 != NULL);
67 if XNN_UNPREDICTABLE(i4 != zero) {
68 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
69 }
70 const int8_t* i5 = input[5];
71 assert(i5 != NULL);
72 if XNN_UNPREDICTABLE(i5 != zero) {
73 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
74 }
75 const int8_t* i6 = input[6];
76 assert(i6 != NULL);
77 if XNN_UNPREDICTABLE(i6 != zero) {
78 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
79 }
80 const int8_t* i7 = input[7];
81 assert(i7 != NULL);
82 if XNN_UNPREDICTABLE(i7 != zero) {
83 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
84 }
85 const int8_t* i8 = input[8];
86 assert(i8 != NULL);
87 if XNN_UNPREDICTABLE(i8 != zero) {
88 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
89 }
90 const int8_t* i9 = input[9];
91 assert(i9 != NULL);
92 if XNN_UNPREDICTABLE(i9 != zero) {
93 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
94 }
95 const int8_t* i10 = input[10];
96 assert(i10 != NULL);
97 if XNN_UNPREDICTABLE(i10 != zero) {
98 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
99 }
100 const int8_t* i11 = input[11];
101 assert(i11 != NULL);
102 if XNN_UNPREDICTABLE(i11 != zero) {
103 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
104 }
105 const int8_t* i12 = input[12];
106 assert(i12 != NULL);
107 if XNN_UNPREDICTABLE(i12 != zero) {
108 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
109 }
110 const int8_t* i13 = input[13];
111 assert(i13 != NULL);
112 if XNN_UNPREDICTABLE(i13 != zero) {
113 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
114 }
115 const int8_t* i14 = input[14];
116 assert(i14 != NULL);
117 if XNN_UNPREDICTABLE(i14 != zero) {
118 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
119 }
120 const int8_t* i15 = input[15];
121 assert(i15 != NULL);
122 if XNN_UNPREDICTABLE(i15 != zero) {
123 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
124 }
125 const int8_t* i16 = input[16];
126 assert(i16 != NULL);
127 if XNN_UNPREDICTABLE(i16 != zero) {
128 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
129 }
130 const int8_t* i17 = input[17];
131 assert(i17 != NULL);
132 if XNN_UNPREDICTABLE(i17 != zero) {
133 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
134 }
135 const int8_t* i18 = input[18];
136 assert(i18 != NULL);
137 if XNN_UNPREDICTABLE(i18 != zero) {
138 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
139 }
140 const int8_t* i19 = input[19];
141 assert(i19 != NULL);
142 if XNN_UNPREDICTABLE(i19 != zero) {
143 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
144 }
145 const int8_t* i20 = input[20];
146 assert(i20 != NULL);
147 if XNN_UNPREDICTABLE(i20 != zero) {
148 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
149 }
150 const int8_t* i21 = input[21];
151 assert(i21 != NULL);
152 if XNN_UNPREDICTABLE(i21 != zero) {
153 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
154 }
155 const int8_t* i22 = input[22];
156 assert(i22 != NULL);
157 if XNN_UNPREDICTABLE(i22 != zero) {
158 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
159 }
160 const int8_t* i23 = input[23];
161 assert(i23 != NULL);
162 if XNN_UNPREDICTABLE(i23 != zero) {
163 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
164 }
165 const int8_t* i24 = input[24];
166 assert(i24 != NULL);
167 if XNN_UNPREDICTABLE(i24 != zero) {
168 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
169 }
170 input = (const int8_t**) ((uintptr_t) input + input_stride);
171
172 size_t c = channels;
173 const void* w = weights;
174 for (; c >= 16; c -= 16) {
175 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
176 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
177 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
178 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
179
180
181 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
182 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
183 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
184 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
185 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
186 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
187 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
188 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
189 i0 += 16;
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
193 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
194 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
195
196 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
197 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
198 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
199 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
200 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
201 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
202 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
203 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
204 i1 += 16;
205
206 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
207 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
208 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
209 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
210
211 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
212 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
213 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
214 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
215 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
216 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
217 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
218 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
219 i2 += 16;
220
221 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
222 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
223 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
224 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
225
226 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
227 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)))));
228 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
229 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(int8_t)))));
230 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
231 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)))));
232 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
233 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(int8_t)))));
234 i3 += 16;
235
236 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
237 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
238 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
239 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
240
241 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
242 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)))));
243 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
244 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(int8_t)))));
245 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
246 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)))));
247 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
248 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(int8_t)))));
249 i4 += 16;
250
251 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
252 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
253 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
254 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
255
256 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
257 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)))));
258 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
259 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(int8_t)))));
260 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
261 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)))));
262 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
263 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(int8_t)))));
264 i5 += 16;
265
266 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
267 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
268 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
269 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
270
271 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
272 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)))));
273 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
274 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 100 * sizeof(int8_t)))));
275 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
276 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)))));
277 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
278 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 108 * sizeof(int8_t)))));
279 i6 += 16;
280
281 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
282 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
283 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi6x89AB, vk6x89AB));
284 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi6xCDEF, vk6xCDEF));
285
286 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
287 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)))));
288 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
289 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 116 * sizeof(int8_t)))));
290 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 8)));
291 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)))));
292 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 12)));
293 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 124 * sizeof(int8_t)))));
294 i7 += 16;
295
296 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
297 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
298 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi7x89AB, vk7x89AB));
299 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi7xCDEF, vk7xCDEF));
300
301 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
302 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)))));
303 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
304 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 132 * sizeof(int8_t)))));
305 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 8)));
306 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)))));
307 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 12)));
308 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 140 * sizeof(int8_t)))));
309 i8 += 16;
310
311 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
312 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
313 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi8x89AB, vk8x89AB));
314 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi8xCDEF, vk8xCDEF));
315
316 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
317 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)))));
318 const __m128i vi9x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 4)));
319 const __m128i vk9x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 148 * sizeof(int8_t)))));
320 const __m128i vi9x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 8)));
321 const __m128i vk9x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)))));
322 const __m128i vi9xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 12)));
323 const __m128i vk9xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 156 * sizeof(int8_t)))));
324 i9 += 16;
325
326 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
327 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi9x4567, vk9x4567));
328 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi9x89AB, vk9x89AB));
329 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi9xCDEF, vk9xCDEF));
330
331 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
332 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)))));
333 const __m128i vi10x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 4)));
334 const __m128i vk10x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 164 * sizeof(int8_t)))));
335 const __m128i vi10x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 8)));
336 const __m128i vk10x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)))));
337 const __m128i vi10xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 12)));
338 const __m128i vk10xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 172 * sizeof(int8_t)))));
339 i10 += 16;
340
341 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
342 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi10x4567, vk10x4567));
343 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi10x89AB, vk10x89AB));
344 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi10xCDEF, vk10xCDEF));
345
346 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
347 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)))));
348 const __m128i vi11x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 4)));
349 const __m128i vk11x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 180 * sizeof(int8_t)))));
350 const __m128i vi11x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 8)));
351 const __m128i vk11x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)))));
352 const __m128i vi11xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 12)));
353 const __m128i vk11xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 188 * sizeof(int8_t)))));
354 i11 += 16;
355
356 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
357 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi11x4567, vk11x4567));
358 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi11x89AB, vk11x89AB));
359 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi11xCDEF, vk11xCDEF));
360
361 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
362 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)))));
363 const __m128i vi12x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 4)));
364 const __m128i vk12x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 196 * sizeof(int8_t)))));
365 const __m128i vi12x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 8)));
366 const __m128i vk12x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)))));
367 const __m128i vi12xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 12)));
368 const __m128i vk12xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 204 * sizeof(int8_t)))));
369 i12 += 16;
370
371 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
372 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi12x4567, vk12x4567));
373 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi12x89AB, vk12x89AB));
374 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi12xCDEF, vk12xCDEF));
375
376 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
377 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)))));
378 const __m128i vi13x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 4)));
379 const __m128i vk13x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 212 * sizeof(int8_t)))));
380 const __m128i vi13x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 8)));
381 const __m128i vk13x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)))));
382 const __m128i vi13xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 12)));
383 const __m128i vk13xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 220 * sizeof(int8_t)))));
384 i13 += 16;
385
386 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
387 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi13x4567, vk13x4567));
388 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi13x89AB, vk13x89AB));
389 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi13xCDEF, vk13xCDEF));
390
391 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
392 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)))));
393 const __m128i vi14x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 4)));
394 const __m128i vk14x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 228 * sizeof(int8_t)))));
395 const __m128i vi14x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 8)));
396 const __m128i vk14x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)))));
397 const __m128i vi14xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 12)));
398 const __m128i vk14xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 236 * sizeof(int8_t)))));
399 i14 += 16;
400
401 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
402 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi14x4567, vk14x4567));
403 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi14x89AB, vk14x89AB));
404 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi14xCDEF, vk14xCDEF));
405
406 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
407 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)))));
408 const __m128i vi15x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 4)));
409 const __m128i vk15x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 244 * sizeof(int8_t)))));
410 const __m128i vi15x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 8)));
411 const __m128i vk15x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)))));
412 const __m128i vi15xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 12)));
413 const __m128i vk15xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 252 * sizeof(int8_t)))));
414 i15 += 16;
415
416 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
417 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi15x4567, vk15x4567));
418 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi15x89AB, vk15x89AB));
419 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi15xCDEF, vk15xCDEF));
420
421 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
422 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)))));
423 const __m128i vi16x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 4)));
424 const __m128i vk16x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 260 * sizeof(int8_t)))));
425 const __m128i vi16x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 8)));
426 const __m128i vk16x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)))));
427 const __m128i vi16xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 12)));
428 const __m128i vk16xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 268 * sizeof(int8_t)))));
429 i16 += 16;
430
431 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
432 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi16x4567, vk16x4567));
433 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi16x89AB, vk16x89AB));
434 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi16xCDEF, vk16xCDEF));
435
436 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
437 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)))));
438 const __m128i vi17x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 4)));
439 const __m128i vk17x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 276 * sizeof(int8_t)))));
440 const __m128i vi17x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 8)));
441 const __m128i vk17x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)))));
442 const __m128i vi17xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 12)));
443 const __m128i vk17xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 284 * sizeof(int8_t)))));
444 i17 += 16;
445
446 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
447 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi17x4567, vk17x4567));
448 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi17x89AB, vk17x89AB));
449 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi17xCDEF, vk17xCDEF));
450
451 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
452 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)))));
453 const __m128i vi18x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 4)));
454 const __m128i vk18x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 292 * sizeof(int8_t)))));
455 const __m128i vi18x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 8)));
456 const __m128i vk18x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)))));
457 const __m128i vi18xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 12)));
458 const __m128i vk18xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 300 * sizeof(int8_t)))));
459 i18 += 16;
460
461 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
462 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi18x4567, vk18x4567));
463 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi18x89AB, vk18x89AB));
464 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi18xCDEF, vk18xCDEF));
465
466 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
467 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)))));
468 const __m128i vi19x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 4)));
469 const __m128i vk19x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 308 * sizeof(int8_t)))));
470 const __m128i vi19x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 8)));
471 const __m128i vk19x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)))));
472 const __m128i vi19xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 12)));
473 const __m128i vk19xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 316 * sizeof(int8_t)))));
474 i19 += 16;
475
476 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
477 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi19x4567, vk19x4567));
478 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi19x89AB, vk19x89AB));
479 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi19xCDEF, vk19xCDEF));
480
481 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
482 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)))));
483 const __m128i vi20x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 4)));
484 const __m128i vk20x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 324 * sizeof(int8_t)))));
485 const __m128i vi20x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 8)));
486 const __m128i vk20x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)))));
487 const __m128i vi20xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 12)));
488 const __m128i vk20xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 332 * sizeof(int8_t)))));
489 i20 += 16;
490
491 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
492 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi20x4567, vk20x4567));
493 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi20x89AB, vk20x89AB));
494 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi20xCDEF, vk20xCDEF));
495
496 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
497 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)))));
498 const __m128i vi21x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 4)));
499 const __m128i vk21x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 340 * sizeof(int8_t)))));
500 const __m128i vi21x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 8)));
501 const __m128i vk21x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)))));
502 const __m128i vi21xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 12)));
503 const __m128i vk21xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 348 * sizeof(int8_t)))));
504 i21 += 16;
505
506 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
507 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi21x4567, vk21x4567));
508 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi21x89AB, vk21x89AB));
509 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi21xCDEF, vk21xCDEF));
510
511 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
512 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)))));
513 const __m128i vi22x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 4)));
514 const __m128i vk22x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 356 * sizeof(int8_t)))));
515 const __m128i vi22x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 8)));
516 const __m128i vk22x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)))));
517 const __m128i vi22xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 12)));
518 const __m128i vk22xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 364 * sizeof(int8_t)))));
519 i22 += 16;
520
521 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
522 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi22x4567, vk22x4567));
523 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi22x89AB, vk22x89AB));
524 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi22xCDEF, vk22xCDEF));
525
526 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
527 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)))));
528 const __m128i vi23x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 4)));
529 const __m128i vk23x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 372 * sizeof(int8_t)))));
530 const __m128i vi23x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 8)));
531 const __m128i vk23x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)))));
532 const __m128i vi23xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 12)));
533 const __m128i vk23xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 380 * sizeof(int8_t)))));
534 i23 += 16;
535
536 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
537 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi23x4567, vk23x4567));
538 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi23x89AB, vk23x89AB));
539 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi23xCDEF, vk23xCDEF));
540
541 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
542 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)))));
543 const __m128i vi24x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 4)));
544 const __m128i vk24x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 388 * sizeof(int8_t)))));
545 const __m128i vi24x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 8)));
546 const __m128i vk24x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)))));
547 const __m128i vi24xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 12)));
548 const __m128i vk24xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 396 * sizeof(int8_t)))));
549 i24 += 16;
550
551 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
552 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi24x4567, vk24x4567));
553 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi24x89AB, vk24x89AB));
554 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi24xCDEF, vk24xCDEF));
555
556 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
557
558 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
559 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
560 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
561 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
562
563 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
564 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
565 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
566 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
567
568 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
569 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
570 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
571 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
572
573 vacc0123 = _mm_cvtps_epi32(vscaled0123);
574 vacc4567 = _mm_cvtps_epi32(vscaled4567);
575 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
576 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
577
578 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
579 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
580
581 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
582 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
583
584 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
585 output += 16;
586 }
587 if XNN_UNLIKELY(c != 0) {
588 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
589 do {
590 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
591
592 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
593 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) k)));
594 i0 += 4;
595
596 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
597 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
598 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16))));
599 i1 += 4;
600
601 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
602 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
603 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32))));
604 i2 += 4;
605
606 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
607 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
608 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48))));
609 i3 += 4;
610
611 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
612 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
613 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64))));
614 i4 += 4;
615
616 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
617 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
618 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 80))));
619 i5 += 4;
620
621 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
622 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
623 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96))));
624 i6 += 4;
625
626 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
627 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
628 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 112))));
629 i7 += 4;
630
631 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
632 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
633 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 128))));
634 i8 += 4;
635
636 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
637 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
638 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 144))));
639 i9 += 4;
640
641 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
642 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
643 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 160))));
644 i10 += 4;
645
646 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
647 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
648 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 176))));
649 i11 += 4;
650
651 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
652 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
653 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 192))));
654 i12 += 4;
655
656 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
657 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
658 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 208))));
659 i13 += 4;
660
661 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
662 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
663 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 224))));
664 i14 += 4;
665
666 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
667 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
668 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 240))));
669 i15 += 4;
670
671 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
672 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
673 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 256))));
674 i16 += 4;
675
676 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
677 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
678 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 272))));
679 i17 += 4;
680
681 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
682 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
683 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 288))));
684 i18 += 4;
685
686 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
687 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
688 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 304))));
689 i19 += 4;
690
691 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
692 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
693 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 320))));
694 i20 += 4;
695
696 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
697 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
698 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 336))));
699 i21 += 4;
700
701 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
702 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
703 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 352))));
704 i22 += 4;
705
706 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
707 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
708 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 368))));
709 i23 += 4;
710
711 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
712 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
713 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 384))));
714 i24 += 4;
715
716 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
717
718 k += 4;
719
720 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
721 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
722 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
723 vacc0123 = _mm_cvtps_epi32(vscaled0123);
724
725 w = (const void*) ((const int32_t*) w + 4);
726
727 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
728
729 vout0123 = _mm_packs_epi16(vout0123, vout0123);
730 vout0123 = _mm_max_epi8(vout0123, voutput_min);
731
732 if XNN_LIKELY(c >= 4) {
733 _mm_storeu_si32(output, vout0123);
734 output += 4;
735 c -= 4;
736 } else {
737 if (c & 2) {
738 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
739 vout0123 = _mm_srli_epi32(vout0123, 16);
740 output += 2;
741 }
742 if (c & 1) {
743 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
744 output += 1;
745 }
746 c = 0;
747 }
748 } while (c != 0);
749 }
750
751 output = (int8_t*) ((uintptr_t) output + output_increment);
752 } while (--output_width != 0);
753 }