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
22
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__xop_mul32(
24 size_t channels,
25 size_t output_width,
26 const int8_t** input,
27 const void* weights,
28 int8_t* output,
29 size_t input_stride,
30 size_t output_increment,
31 size_t input_offset,
32 const int8_t* zero,
33 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(channels != 0);
36 assert(output_width != 0);
37
38 do {
39 const int8_t* i0 = input[0];
40 assert(i0 != NULL);
41 if XNN_UNPREDICTABLE(i0 != zero) {
42 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43 }
44 const int8_t* i1 = input[1];
45 assert(i1 != NULL);
46 if XNN_UNPREDICTABLE(i1 != zero) {
47 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
48 }
49 const int8_t* i2 = input[2];
50 assert(i2 != NULL);
51 if XNN_UNPREDICTABLE(i2 != zero) {
52 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
53 }
54 const int8_t* i3 = input[3];
55 assert(i3 != NULL);
56 if XNN_UNPREDICTABLE(i3 != zero) {
57 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
58 }
59 const int8_t* i4 = input[4];
60 assert(i4 != NULL);
61 if XNN_UNPREDICTABLE(i4 != zero) {
62 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
63 }
64 const int8_t* i5 = input[5];
65 assert(i5 != NULL);
66 if XNN_UNPREDICTABLE(i5 != zero) {
67 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
68 }
69 const int8_t* i6 = input[6];
70 assert(i6 != NULL);
71 if XNN_UNPREDICTABLE(i6 != zero) {
72 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
73 }
74 const int8_t* i7 = input[7];
75 assert(i7 != NULL);
76 if XNN_UNPREDICTABLE(i7 != zero) {
77 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
78 }
79 const int8_t* i8 = input[8];
80 assert(i8 != NULL);
81 if XNN_UNPREDICTABLE(i8 != zero) {
82 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
83 }
84 const int8_t* i9 = input[9];
85 assert(i9 != NULL);
86 if XNN_UNPREDICTABLE(i9 != zero) {
87 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
88 }
89 const int8_t* i10 = input[10];
90 assert(i10 != NULL);
91 if XNN_UNPREDICTABLE(i10 != zero) {
92 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
93 }
94 const int8_t* i11 = input[11];
95 assert(i11 != NULL);
96 if XNN_UNPREDICTABLE(i11 != zero) {
97 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
98 }
99 const int8_t* i12 = input[12];
100 assert(i12 != NULL);
101 if XNN_UNPREDICTABLE(i12 != zero) {
102 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
103 }
104 const int8_t* i13 = input[13];
105 assert(i13 != NULL);
106 if XNN_UNPREDICTABLE(i13 != zero) {
107 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
108 }
109 const int8_t* i14 = input[14];
110 assert(i14 != NULL);
111 if XNN_UNPREDICTABLE(i14 != zero) {
112 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
113 }
114 const int8_t* i15 = input[15];
115 assert(i15 != NULL);
116 if XNN_UNPREDICTABLE(i15 != zero) {
117 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
118 }
119 const int8_t* i16 = input[16];
120 assert(i16 != NULL);
121 if XNN_UNPREDICTABLE(i16 != zero) {
122 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
123 }
124 const int8_t* i17 = input[17];
125 assert(i17 != NULL);
126 if XNN_UNPREDICTABLE(i17 != zero) {
127 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
128 }
129 const int8_t* i18 = input[18];
130 assert(i18 != NULL);
131 if XNN_UNPREDICTABLE(i18 != zero) {
132 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
133 }
134 const int8_t* i19 = input[19];
135 assert(i19 != NULL);
136 if XNN_UNPREDICTABLE(i19 != zero) {
137 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
138 }
139 const int8_t* i20 = input[20];
140 assert(i20 != NULL);
141 if XNN_UNPREDICTABLE(i20 != zero) {
142 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
143 }
144 const int8_t* i21 = input[21];
145 assert(i21 != NULL);
146 if XNN_UNPREDICTABLE(i21 != zero) {
147 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
148 }
149 const int8_t* i22 = input[22];
150 assert(i22 != NULL);
151 if XNN_UNPREDICTABLE(i22 != zero) {
152 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
153 }
154 const int8_t* i23 = input[23];
155 assert(i23 != NULL);
156 if XNN_UNPREDICTABLE(i23 != zero) {
157 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
158 }
159 const int8_t* i24 = input[24];
160 assert(i24 != NULL);
161 if XNN_UNPREDICTABLE(i24 != zero) {
162 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
163 }
164 input = (const int8_t**) ((uintptr_t) input + input_stride);
165
166 size_t c = channels;
167 const void* w = weights;
168 for (; c >= 16; c -= 16) {
169 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
170 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
171 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
172 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
173
174
175 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
176 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
177 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 4));
178 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(int8_t))));
179 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 8));
180 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t))));
181 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 12));
182 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(int8_t))));
183 i0 += 16;
184
185 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
186 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
187 vacc89AB = _mm_macc_epi32(vi0x89AB, vk0x89AB, vacc89AB);
188 vaccCDEF = _mm_macc_epi32(vi0xCDEF, vk0xCDEF, vaccCDEF);
189
190 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
191 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
192 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 4));
193 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(int8_t))));
194 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 8));
195 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t))));
196 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 12));
197 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(int8_t))));
198 i1 += 16;
199
200 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
201 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
202 vacc89AB = _mm_macc_epi32(vi1x89AB, vk1x89AB, vacc89AB);
203 vaccCDEF = _mm_macc_epi32(vi1xCDEF, vk1xCDEF, vaccCDEF);
204
205 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
206 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
207 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 4));
208 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(int8_t))));
209 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 8));
210 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t))));
211 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 12));
212 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(int8_t))));
213 i2 += 16;
214
215 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
216 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
217 vacc89AB = _mm_macc_epi32(vi2x89AB, vk2x89AB, vacc89AB);
218 vaccCDEF = _mm_macc_epi32(vi2xCDEF, vk2xCDEF, vaccCDEF);
219
220 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
221 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
222 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 4));
223 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(int8_t))));
224 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 8));
225 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t))));
226 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 12));
227 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(int8_t))));
228 i3 += 16;
229
230 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
231 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
232 vacc89AB = _mm_macc_epi32(vi3x89AB, vk3x89AB, vacc89AB);
233 vaccCDEF = _mm_macc_epi32(vi3xCDEF, vk3xCDEF, vaccCDEF);
234
235 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
236 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
237 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 4));
238 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(int8_t))));
239 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 8));
240 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t))));
241 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 12));
242 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(int8_t))));
243 i4 += 16;
244
245 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
246 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
247 vacc89AB = _mm_macc_epi32(vi4x89AB, vk4x89AB, vacc89AB);
248 vaccCDEF = _mm_macc_epi32(vi4xCDEF, vk4xCDEF, vaccCDEF);
249
250 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
251 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
252 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 4));
253 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(int8_t))));
254 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 8));
255 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t))));
256 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 12));
257 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(int8_t))));
258 i5 += 16;
259
260 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
261 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
262 vacc89AB = _mm_macc_epi32(vi5x89AB, vk5x89AB, vacc89AB);
263 vaccCDEF = _mm_macc_epi32(vi5xCDEF, vk5xCDEF, vaccCDEF);
264
265 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
266 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
267 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 4));
268 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 100 * sizeof(int8_t))));
269 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 8));
270 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t))));
271 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 12));
272 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 108 * sizeof(int8_t))));
273 i6 += 16;
274
275 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
276 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
277 vacc89AB = _mm_macc_epi32(vi6x89AB, vk6x89AB, vacc89AB);
278 vaccCDEF = _mm_macc_epi32(vi6xCDEF, vk6xCDEF, vaccCDEF);
279
280 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
281 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
282 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 4));
283 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 116 * sizeof(int8_t))));
284 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 8));
285 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t))));
286 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 12));
287 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 124 * sizeof(int8_t))));
288 i7 += 16;
289
290 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
291 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
292 vacc89AB = _mm_macc_epi32(vi7x89AB, vk7x89AB, vacc89AB);
293 vaccCDEF = _mm_macc_epi32(vi7xCDEF, vk7xCDEF, vaccCDEF);
294
295 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
296 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
297 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 4));
298 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 132 * sizeof(int8_t))));
299 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 8));
300 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t))));
301 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 12));
302 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 140 * sizeof(int8_t))));
303 i8 += 16;
304
305 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
306 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
307 vacc89AB = _mm_macc_epi32(vi8x89AB, vk8x89AB, vacc89AB);
308 vaccCDEF = _mm_macc_epi32(vi8xCDEF, vk8xCDEF, vaccCDEF);
309
310 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9));
311 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t))));
312 const __m128i vi9x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 4));
313 const __m128i vk9x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 148 * sizeof(int8_t))));
314 const __m128i vi9x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 8));
315 const __m128i vk9x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t))));
316 const __m128i vi9xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 12));
317 const __m128i vk9xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 156 * sizeof(int8_t))));
318 i9 += 16;
319
320 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
321 vacc4567 = _mm_macc_epi32(vi9x4567, vk9x4567, vacc4567);
322 vacc89AB = _mm_macc_epi32(vi9x89AB, vk9x89AB, vacc89AB);
323 vaccCDEF = _mm_macc_epi32(vi9xCDEF, vk9xCDEF, vaccCDEF);
324
325 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10));
326 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t))));
327 const __m128i vi10x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 4));
328 const __m128i vk10x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 164 * sizeof(int8_t))));
329 const __m128i vi10x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 8));
330 const __m128i vk10x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t))));
331 const __m128i vi10xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 12));
332 const __m128i vk10xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 172 * sizeof(int8_t))));
333 i10 += 16;
334
335 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
336 vacc4567 = _mm_macc_epi32(vi10x4567, vk10x4567, vacc4567);
337 vacc89AB = _mm_macc_epi32(vi10x89AB, vk10x89AB, vacc89AB);
338 vaccCDEF = _mm_macc_epi32(vi10xCDEF, vk10xCDEF, vaccCDEF);
339
340 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11));
341 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t))));
342 const __m128i vi11x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 4));
343 const __m128i vk11x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 180 * sizeof(int8_t))));
344 const __m128i vi11x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 8));
345 const __m128i vk11x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t))));
346 const __m128i vi11xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 12));
347 const __m128i vk11xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 188 * sizeof(int8_t))));
348 i11 += 16;
349
350 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
351 vacc4567 = _mm_macc_epi32(vi11x4567, vk11x4567, vacc4567);
352 vacc89AB = _mm_macc_epi32(vi11x89AB, vk11x89AB, vacc89AB);
353 vaccCDEF = _mm_macc_epi32(vi11xCDEF, vk11xCDEF, vaccCDEF);
354
355 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12));
356 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t))));
357 const __m128i vi12x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 4));
358 const __m128i vk12x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 196 * sizeof(int8_t))));
359 const __m128i vi12x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 8));
360 const __m128i vk12x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t))));
361 const __m128i vi12xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 12));
362 const __m128i vk12xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 204 * sizeof(int8_t))));
363 i12 += 16;
364
365 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
366 vacc4567 = _mm_macc_epi32(vi12x4567, vk12x4567, vacc4567);
367 vacc89AB = _mm_macc_epi32(vi12x89AB, vk12x89AB, vacc89AB);
368 vaccCDEF = _mm_macc_epi32(vi12xCDEF, vk12xCDEF, vaccCDEF);
369
370 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13));
371 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t))));
372 const __m128i vi13x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 4));
373 const __m128i vk13x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 212 * sizeof(int8_t))));
374 const __m128i vi13x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 8));
375 const __m128i vk13x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t))));
376 const __m128i vi13xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 12));
377 const __m128i vk13xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 220 * sizeof(int8_t))));
378 i13 += 16;
379
380 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
381 vacc4567 = _mm_macc_epi32(vi13x4567, vk13x4567, vacc4567);
382 vacc89AB = _mm_macc_epi32(vi13x89AB, vk13x89AB, vacc89AB);
383 vaccCDEF = _mm_macc_epi32(vi13xCDEF, vk13xCDEF, vaccCDEF);
384
385 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14));
386 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t))));
387 const __m128i vi14x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 4));
388 const __m128i vk14x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 228 * sizeof(int8_t))));
389 const __m128i vi14x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 8));
390 const __m128i vk14x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t))));
391 const __m128i vi14xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 12));
392 const __m128i vk14xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 236 * sizeof(int8_t))));
393 i14 += 16;
394
395 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
396 vacc4567 = _mm_macc_epi32(vi14x4567, vk14x4567, vacc4567);
397 vacc89AB = _mm_macc_epi32(vi14x89AB, vk14x89AB, vacc89AB);
398 vaccCDEF = _mm_macc_epi32(vi14xCDEF, vk14xCDEF, vaccCDEF);
399
400 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15));
401 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t))));
402 const __m128i vi15x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 4));
403 const __m128i vk15x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 244 * sizeof(int8_t))));
404 const __m128i vi15x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 8));
405 const __m128i vk15x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t))));
406 const __m128i vi15xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 12));
407 const __m128i vk15xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 252 * sizeof(int8_t))));
408 i15 += 16;
409
410 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
411 vacc4567 = _mm_macc_epi32(vi15x4567, vk15x4567, vacc4567);
412 vacc89AB = _mm_macc_epi32(vi15x89AB, vk15x89AB, vacc89AB);
413 vaccCDEF = _mm_macc_epi32(vi15xCDEF, vk15xCDEF, vaccCDEF);
414
415 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16));
416 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t))));
417 const __m128i vi16x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 4));
418 const __m128i vk16x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 260 * sizeof(int8_t))));
419 const __m128i vi16x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 8));
420 const __m128i vk16x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t))));
421 const __m128i vi16xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 12));
422 const __m128i vk16xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 268 * sizeof(int8_t))));
423 i16 += 16;
424
425 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
426 vacc4567 = _mm_macc_epi32(vi16x4567, vk16x4567, vacc4567);
427 vacc89AB = _mm_macc_epi32(vi16x89AB, vk16x89AB, vacc89AB);
428 vaccCDEF = _mm_macc_epi32(vi16xCDEF, vk16xCDEF, vaccCDEF);
429
430 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17));
431 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t))));
432 const __m128i vi17x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 4));
433 const __m128i vk17x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 276 * sizeof(int8_t))));
434 const __m128i vi17x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 8));
435 const __m128i vk17x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t))));
436 const __m128i vi17xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 12));
437 const __m128i vk17xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 284 * sizeof(int8_t))));
438 i17 += 16;
439
440 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
441 vacc4567 = _mm_macc_epi32(vi17x4567, vk17x4567, vacc4567);
442 vacc89AB = _mm_macc_epi32(vi17x89AB, vk17x89AB, vacc89AB);
443 vaccCDEF = _mm_macc_epi32(vi17xCDEF, vk17xCDEF, vaccCDEF);
444
445 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18));
446 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t))));
447 const __m128i vi18x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 4));
448 const __m128i vk18x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 292 * sizeof(int8_t))));
449 const __m128i vi18x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 8));
450 const __m128i vk18x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t))));
451 const __m128i vi18xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 12));
452 const __m128i vk18xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 300 * sizeof(int8_t))));
453 i18 += 16;
454
455 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
456 vacc4567 = _mm_macc_epi32(vi18x4567, vk18x4567, vacc4567);
457 vacc89AB = _mm_macc_epi32(vi18x89AB, vk18x89AB, vacc89AB);
458 vaccCDEF = _mm_macc_epi32(vi18xCDEF, vk18xCDEF, vaccCDEF);
459
460 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19));
461 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t))));
462 const __m128i vi19x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 4));
463 const __m128i vk19x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 308 * sizeof(int8_t))));
464 const __m128i vi19x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 8));
465 const __m128i vk19x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t))));
466 const __m128i vi19xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 12));
467 const __m128i vk19xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 316 * sizeof(int8_t))));
468 i19 += 16;
469
470 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
471 vacc4567 = _mm_macc_epi32(vi19x4567, vk19x4567, vacc4567);
472 vacc89AB = _mm_macc_epi32(vi19x89AB, vk19x89AB, vacc89AB);
473 vaccCDEF = _mm_macc_epi32(vi19xCDEF, vk19xCDEF, vaccCDEF);
474
475 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20));
476 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t))));
477 const __m128i vi20x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 4));
478 const __m128i vk20x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 324 * sizeof(int8_t))));
479 const __m128i vi20x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 8));
480 const __m128i vk20x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t))));
481 const __m128i vi20xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 12));
482 const __m128i vk20xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 332 * sizeof(int8_t))));
483 i20 += 16;
484
485 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
486 vacc4567 = _mm_macc_epi32(vi20x4567, vk20x4567, vacc4567);
487 vacc89AB = _mm_macc_epi32(vi20x89AB, vk20x89AB, vacc89AB);
488 vaccCDEF = _mm_macc_epi32(vi20xCDEF, vk20xCDEF, vaccCDEF);
489
490 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21));
491 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t))));
492 const __m128i vi21x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 4));
493 const __m128i vk21x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 340 * sizeof(int8_t))));
494 const __m128i vi21x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 8));
495 const __m128i vk21x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t))));
496 const __m128i vi21xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 12));
497 const __m128i vk21xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 348 * sizeof(int8_t))));
498 i21 += 16;
499
500 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
501 vacc4567 = _mm_macc_epi32(vi21x4567, vk21x4567, vacc4567);
502 vacc89AB = _mm_macc_epi32(vi21x89AB, vk21x89AB, vacc89AB);
503 vaccCDEF = _mm_macc_epi32(vi21xCDEF, vk21xCDEF, vaccCDEF);
504
505 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22));
506 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t))));
507 const __m128i vi22x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 4));
508 const __m128i vk22x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 356 * sizeof(int8_t))));
509 const __m128i vi22x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 8));
510 const __m128i vk22x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t))));
511 const __m128i vi22xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 12));
512 const __m128i vk22xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 364 * sizeof(int8_t))));
513 i22 += 16;
514
515 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
516 vacc4567 = _mm_macc_epi32(vi22x4567, vk22x4567, vacc4567);
517 vacc89AB = _mm_macc_epi32(vi22x89AB, vk22x89AB, vacc89AB);
518 vaccCDEF = _mm_macc_epi32(vi22xCDEF, vk22xCDEF, vaccCDEF);
519
520 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23));
521 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t))));
522 const __m128i vi23x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 4));
523 const __m128i vk23x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 372 * sizeof(int8_t))));
524 const __m128i vi23x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 8));
525 const __m128i vk23x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t))));
526 const __m128i vi23xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 12));
527 const __m128i vk23xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 380 * sizeof(int8_t))));
528 i23 += 16;
529
530 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
531 vacc4567 = _mm_macc_epi32(vi23x4567, vk23x4567, vacc4567);
532 vacc89AB = _mm_macc_epi32(vi23x89AB, vk23x89AB, vacc89AB);
533 vaccCDEF = _mm_macc_epi32(vi23xCDEF, vk23xCDEF, vaccCDEF);
534
535 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24));
536 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t))));
537 const __m128i vi24x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 4));
538 const __m128i vk24x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 388 * sizeof(int8_t))));
539 const __m128i vi24x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 8));
540 const __m128i vk24x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t))));
541 const __m128i vi24xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 12));
542 const __m128i vk24xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 396 * sizeof(int8_t))));
543 i24 += 16;
544
545 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
546 vacc4567 = _mm_macc_epi32(vi24x4567, vk24x4567, vacc4567);
547 vacc89AB = _mm_macc_epi32(vi24x89AB, vk24x89AB, vacc89AB);
548 vaccCDEF = _mm_macc_epi32(vi24xCDEF, vk24xCDEF, vaccCDEF);
549
550 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
551
552 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
553 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
554 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
555 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
556
557 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
558 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
559 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
560 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
561 w = (const void*) ((const float*) w + 16);
562 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
563 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
564 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
565 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
566
567 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
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 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
579 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
580 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
581
582 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
583 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
584 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
585
586 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
587 output += 16;
588 }
589 if XNN_UNLIKELY(c != 0) {
590 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
591 do {
592 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
593
594 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
595 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(k));
596 i0 += 4;
597
598 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
599 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
600 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 16)));
601 i1 += 4;
602
603 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
604 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
605 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 32)));
606 i2 += 4;
607
608 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
609 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
610 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 48)));
611 i3 += 4;
612
613 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
614 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
615 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 64)));
616 i4 += 4;
617
618 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
619 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
620 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 80)));
621 i5 += 4;
622
623 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
624 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
625 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 96)));
626 i6 += 4;
627
628 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
629 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
630 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 112)));
631 i7 += 4;
632
633 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
634 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
635 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 128)));
636 i8 += 4;
637
638 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
639 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9));
640 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 144)));
641 i9 += 4;
642
643 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
644 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10));
645 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 160)));
646 i10 += 4;
647
648 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
649 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11));
650 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 176)));
651 i11 += 4;
652
653 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
654 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12));
655 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 192)));
656 i12 += 4;
657
658 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
659 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13));
660 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 208)));
661 i13 += 4;
662
663 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
664 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14));
665 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 224)));
666 i14 += 4;
667
668 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
669 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15));
670 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 240)));
671 i15 += 4;
672
673 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
674 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16));
675 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 256)));
676 i16 += 4;
677
678 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
679 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17));
680 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 272)));
681 i17 += 4;
682
683 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
684 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18));
685 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 288)));
686 i18 += 4;
687
688 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
689 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19));
690 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 304)));
691 i19 += 4;
692
693 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
694 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20));
695 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 320)));
696 i20 += 4;
697
698 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
699 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21));
700 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 336)));
701 i21 += 4;
702
703 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
704 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22));
705 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 352)));
706 i22 += 4;
707
708 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
709 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23));
710 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 368)));
711 i23 += 4;
712
713 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
714 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24));
715 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 384)));
716 i24 += 4;
717
718 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
719
720 k += 4;
721
722 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
723 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t)));
724 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
725 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->sse4.output_max_less_zero_point));
726 vacc0123 = _mm_cvtps_epi32(vscaled0123);
727
728 w = (const void*) ((const int32_t*) w + 4);
729
730 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
731 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
732
733 vout0123 = _mm_packs_epi16(vout0123, vout0123);
734 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->sse4.output_min));
735
736 if XNN_LIKELY(c >= 4) {
737 _mm_storeu_si32(output, vout0123);
738 output += 4;
739 c -= 4;
740 } else {
741 if (c & 2) {
742 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
743 vout0123 = _mm_srli_epi32(vout0123, 16);
744 output += 2;
745 }
746 if (c & 1) {
747 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
748 output += 1;
749 }
750 c = 0;
751 }
752 } while (c != 0);
753 }
754
755 output = (int8_t*) ((uintptr_t) output + output_increment);
756 } while (--output_width != 0);
757 }
758