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_up24x25__xop_mul32(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__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 >= 24; c -= 24) {
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 __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
174 __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
175
176
177 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
178 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t))));
179 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 4));
180 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 4 * sizeof(int8_t))));
181 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 8));
182 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t))));
183 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 12));
184 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 12 * sizeof(int8_t))));
185 const __m128i vi0xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 16));
186 const __m128i vk0xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t))));
187 const __m128i vi0xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 20));
188 const __m128i vk0xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 20 * sizeof(int8_t))));
189 i0 += 24;
190
191 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
192 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
193 vacc89AB = _mm_macc_epi32(vi0x89AB, vk0x89AB, vacc89AB);
194 vaccCDEF = _mm_macc_epi32(vi0xCDEF, vk0xCDEF, vaccCDEF);
195 vaccGHIJ = _mm_macc_epi32(vi0xGHIJ, vk0xGHIJ, vaccGHIJ);
196 vaccKLMN = _mm_macc_epi32(vi0xKLMN, vk0xKLMN, vaccKLMN);
197
198 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
199 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t))));
200 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 4));
201 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 28 * sizeof(int8_t))));
202 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 8));
203 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t))));
204 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 12));
205 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 36 * sizeof(int8_t))));
206 const __m128i vi1xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 16));
207 const __m128i vk1xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t))));
208 const __m128i vi1xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 20));
209 const __m128i vk1xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 44 * sizeof(int8_t))));
210 i1 += 24;
211
212 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
213 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
214 vacc89AB = _mm_macc_epi32(vi1x89AB, vk1x89AB, vacc89AB);
215 vaccCDEF = _mm_macc_epi32(vi1xCDEF, vk1xCDEF, vaccCDEF);
216 vaccGHIJ = _mm_macc_epi32(vi1xGHIJ, vk1xGHIJ, vaccGHIJ);
217 vaccKLMN = _mm_macc_epi32(vi1xKLMN, vk1xKLMN, vaccKLMN);
218
219 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
220 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t))));
221 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 4));
222 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 52 * sizeof(int8_t))));
223 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 8));
224 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t))));
225 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 12));
226 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 60 * sizeof(int8_t))));
227 const __m128i vi2xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 16));
228 const __m128i vk2xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t))));
229 const __m128i vi2xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 20));
230 const __m128i vk2xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 68 * sizeof(int8_t))));
231 i2 += 24;
232
233 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
234 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
235 vacc89AB = _mm_macc_epi32(vi2x89AB, vk2x89AB, vacc89AB);
236 vaccCDEF = _mm_macc_epi32(vi2xCDEF, vk2xCDEF, vaccCDEF);
237 vaccGHIJ = _mm_macc_epi32(vi2xGHIJ, vk2xGHIJ, vaccGHIJ);
238 vaccKLMN = _mm_macc_epi32(vi2xKLMN, vk2xKLMN, vaccKLMN);
239
240 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
241 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t))));
242 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 4));
243 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 76 * sizeof(int8_t))));
244 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 8));
245 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t))));
246 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 12));
247 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 84 * sizeof(int8_t))));
248 const __m128i vi3xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 16));
249 const __m128i vk3xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t))));
250 const __m128i vi3xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 20));
251 const __m128i vk3xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 92 * sizeof(int8_t))));
252 i3 += 24;
253
254 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
255 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
256 vacc89AB = _mm_macc_epi32(vi3x89AB, vk3x89AB, vacc89AB);
257 vaccCDEF = _mm_macc_epi32(vi3xCDEF, vk3xCDEF, vaccCDEF);
258 vaccGHIJ = _mm_macc_epi32(vi3xGHIJ, vk3xGHIJ, vaccGHIJ);
259 vaccKLMN = _mm_macc_epi32(vi3xKLMN, vk3xKLMN, vaccKLMN);
260
261 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
262 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t))));
263 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 4));
264 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 100 * sizeof(int8_t))));
265 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 8));
266 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t))));
267 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 12));
268 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 108 * sizeof(int8_t))));
269 const __m128i vi4xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 16));
270 const __m128i vk4xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t))));
271 const __m128i vi4xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 20));
272 const __m128i vk4xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 116 * sizeof(int8_t))));
273 i4 += 24;
274
275 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
276 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
277 vacc89AB = _mm_macc_epi32(vi4x89AB, vk4x89AB, vacc89AB);
278 vaccCDEF = _mm_macc_epi32(vi4xCDEF, vk4xCDEF, vaccCDEF);
279 vaccGHIJ = _mm_macc_epi32(vi4xGHIJ, vk4xGHIJ, vaccGHIJ);
280 vaccKLMN = _mm_macc_epi32(vi4xKLMN, vk4xKLMN, vaccKLMN);
281
282 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
283 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t))));
284 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 4));
285 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 124 * sizeof(int8_t))));
286 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 8));
287 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t))));
288 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 12));
289 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 132 * sizeof(int8_t))));
290 const __m128i vi5xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 16));
291 const __m128i vk5xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t))));
292 const __m128i vi5xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 20));
293 const __m128i vk5xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 140 * sizeof(int8_t))));
294 i5 += 24;
295
296 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
297 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
298 vacc89AB = _mm_macc_epi32(vi5x89AB, vk5x89AB, vacc89AB);
299 vaccCDEF = _mm_macc_epi32(vi5xCDEF, vk5xCDEF, vaccCDEF);
300 vaccGHIJ = _mm_macc_epi32(vi5xGHIJ, vk5xGHIJ, vaccGHIJ);
301 vaccKLMN = _mm_macc_epi32(vi5xKLMN, vk5xKLMN, vaccKLMN);
302
303 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
304 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t))));
305 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 4));
306 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 148 * sizeof(int8_t))));
307 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 8));
308 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t))));
309 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 12));
310 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 156 * sizeof(int8_t))));
311 const __m128i vi6xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 16));
312 const __m128i vk6xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t))));
313 const __m128i vi6xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 20));
314 const __m128i vk6xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 164 * sizeof(int8_t))));
315 i6 += 24;
316
317 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
318 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
319 vacc89AB = _mm_macc_epi32(vi6x89AB, vk6x89AB, vacc89AB);
320 vaccCDEF = _mm_macc_epi32(vi6xCDEF, vk6xCDEF, vaccCDEF);
321 vaccGHIJ = _mm_macc_epi32(vi6xGHIJ, vk6xGHIJ, vaccGHIJ);
322 vaccKLMN = _mm_macc_epi32(vi6xKLMN, vk6xKLMN, vaccKLMN);
323
324 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
325 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t))));
326 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 4));
327 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 172 * sizeof(int8_t))));
328 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 8));
329 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t))));
330 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 12));
331 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 180 * sizeof(int8_t))));
332 const __m128i vi7xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 16));
333 const __m128i vk7xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t))));
334 const __m128i vi7xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 20));
335 const __m128i vk7xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 188 * sizeof(int8_t))));
336 i7 += 24;
337
338 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
339 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
340 vacc89AB = _mm_macc_epi32(vi7x89AB, vk7x89AB, vacc89AB);
341 vaccCDEF = _mm_macc_epi32(vi7xCDEF, vk7xCDEF, vaccCDEF);
342 vaccGHIJ = _mm_macc_epi32(vi7xGHIJ, vk7xGHIJ, vaccGHIJ);
343 vaccKLMN = _mm_macc_epi32(vi7xKLMN, vk7xKLMN, vaccKLMN);
344
345 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
346 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t))));
347 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 4));
348 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 196 * sizeof(int8_t))));
349 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 8));
350 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t))));
351 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 12));
352 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 204 * sizeof(int8_t))));
353 const __m128i vi8xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 16));
354 const __m128i vk8xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t))));
355 const __m128i vi8xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 20));
356 const __m128i vk8xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 212 * sizeof(int8_t))));
357 i8 += 24;
358
359 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
360 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
361 vacc89AB = _mm_macc_epi32(vi8x89AB, vk8x89AB, vacc89AB);
362 vaccCDEF = _mm_macc_epi32(vi8xCDEF, vk8xCDEF, vaccCDEF);
363 vaccGHIJ = _mm_macc_epi32(vi8xGHIJ, vk8xGHIJ, vaccGHIJ);
364 vaccKLMN = _mm_macc_epi32(vi8xKLMN, vk8xKLMN, vaccKLMN);
365
366 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9));
367 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t))));
368 const __m128i vi9x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 4));
369 const __m128i vk9x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 220 * sizeof(int8_t))));
370 const __m128i vi9x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 8));
371 const __m128i vk9x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 224 * sizeof(int8_t))));
372 const __m128i vi9xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 12));
373 const __m128i vk9xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 228 * sizeof(int8_t))));
374 const __m128i vi9xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 16));
375 const __m128i vk9xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 232 * sizeof(int8_t))));
376 const __m128i vi9xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 20));
377 const __m128i vk9xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 236 * sizeof(int8_t))));
378 i9 += 24;
379
380 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
381 vacc4567 = _mm_macc_epi32(vi9x4567, vk9x4567, vacc4567);
382 vacc89AB = _mm_macc_epi32(vi9x89AB, vk9x89AB, vacc89AB);
383 vaccCDEF = _mm_macc_epi32(vi9xCDEF, vk9xCDEF, vaccCDEF);
384 vaccGHIJ = _mm_macc_epi32(vi9xGHIJ, vk9xGHIJ, vaccGHIJ);
385 vaccKLMN = _mm_macc_epi32(vi9xKLMN, vk9xKLMN, vaccKLMN);
386
387 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10));
388 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 240 * sizeof(int8_t))));
389 const __m128i vi10x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 4));
390 const __m128i vk10x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 244 * sizeof(int8_t))));
391 const __m128i vi10x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 8));
392 const __m128i vk10x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 248 * sizeof(int8_t))));
393 const __m128i vi10xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 12));
394 const __m128i vk10xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 252 * sizeof(int8_t))));
395 const __m128i vi10xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 16));
396 const __m128i vk10xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 256 * sizeof(int8_t))));
397 const __m128i vi10xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 20));
398 const __m128i vk10xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 260 * sizeof(int8_t))));
399 i10 += 24;
400
401 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
402 vacc4567 = _mm_macc_epi32(vi10x4567, vk10x4567, vacc4567);
403 vacc89AB = _mm_macc_epi32(vi10x89AB, vk10x89AB, vacc89AB);
404 vaccCDEF = _mm_macc_epi32(vi10xCDEF, vk10xCDEF, vaccCDEF);
405 vaccGHIJ = _mm_macc_epi32(vi10xGHIJ, vk10xGHIJ, vaccGHIJ);
406 vaccKLMN = _mm_macc_epi32(vi10xKLMN, vk10xKLMN, vaccKLMN);
407
408 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11));
409 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 264 * sizeof(int8_t))));
410 const __m128i vi11x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 4));
411 const __m128i vk11x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 268 * sizeof(int8_t))));
412 const __m128i vi11x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 8));
413 const __m128i vk11x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 272 * sizeof(int8_t))));
414 const __m128i vi11xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 12));
415 const __m128i vk11xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 276 * sizeof(int8_t))));
416 const __m128i vi11xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 16));
417 const __m128i vk11xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 280 * sizeof(int8_t))));
418 const __m128i vi11xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 20));
419 const __m128i vk11xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 284 * sizeof(int8_t))));
420 i11 += 24;
421
422 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
423 vacc4567 = _mm_macc_epi32(vi11x4567, vk11x4567, vacc4567);
424 vacc89AB = _mm_macc_epi32(vi11x89AB, vk11x89AB, vacc89AB);
425 vaccCDEF = _mm_macc_epi32(vi11xCDEF, vk11xCDEF, vaccCDEF);
426 vaccGHIJ = _mm_macc_epi32(vi11xGHIJ, vk11xGHIJ, vaccGHIJ);
427 vaccKLMN = _mm_macc_epi32(vi11xKLMN, vk11xKLMN, vaccKLMN);
428
429 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12));
430 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 288 * sizeof(int8_t))));
431 const __m128i vi12x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 4));
432 const __m128i vk12x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 292 * sizeof(int8_t))));
433 const __m128i vi12x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 8));
434 const __m128i vk12x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 296 * sizeof(int8_t))));
435 const __m128i vi12xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 12));
436 const __m128i vk12xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 300 * sizeof(int8_t))));
437 const __m128i vi12xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 16));
438 const __m128i vk12xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 304 * sizeof(int8_t))));
439 const __m128i vi12xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 20));
440 const __m128i vk12xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 308 * sizeof(int8_t))));
441 i12 += 24;
442
443 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
444 vacc4567 = _mm_macc_epi32(vi12x4567, vk12x4567, vacc4567);
445 vacc89AB = _mm_macc_epi32(vi12x89AB, vk12x89AB, vacc89AB);
446 vaccCDEF = _mm_macc_epi32(vi12xCDEF, vk12xCDEF, vaccCDEF);
447 vaccGHIJ = _mm_macc_epi32(vi12xGHIJ, vk12xGHIJ, vaccGHIJ);
448 vaccKLMN = _mm_macc_epi32(vi12xKLMN, vk12xKLMN, vaccKLMN);
449
450 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13));
451 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 312 * sizeof(int8_t))));
452 const __m128i vi13x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 4));
453 const __m128i vk13x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 316 * sizeof(int8_t))));
454 const __m128i vi13x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 8));
455 const __m128i vk13x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 320 * sizeof(int8_t))));
456 const __m128i vi13xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 12));
457 const __m128i vk13xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 324 * sizeof(int8_t))));
458 const __m128i vi13xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 16));
459 const __m128i vk13xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 328 * sizeof(int8_t))));
460 const __m128i vi13xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 20));
461 const __m128i vk13xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 332 * sizeof(int8_t))));
462 i13 += 24;
463
464 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
465 vacc4567 = _mm_macc_epi32(vi13x4567, vk13x4567, vacc4567);
466 vacc89AB = _mm_macc_epi32(vi13x89AB, vk13x89AB, vacc89AB);
467 vaccCDEF = _mm_macc_epi32(vi13xCDEF, vk13xCDEF, vaccCDEF);
468 vaccGHIJ = _mm_macc_epi32(vi13xGHIJ, vk13xGHIJ, vaccGHIJ);
469 vaccKLMN = _mm_macc_epi32(vi13xKLMN, vk13xKLMN, vaccKLMN);
470
471 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14));
472 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 336 * sizeof(int8_t))));
473 const __m128i vi14x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 4));
474 const __m128i vk14x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 340 * sizeof(int8_t))));
475 const __m128i vi14x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 8));
476 const __m128i vk14x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 344 * sizeof(int8_t))));
477 const __m128i vi14xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 12));
478 const __m128i vk14xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 348 * sizeof(int8_t))));
479 const __m128i vi14xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 16));
480 const __m128i vk14xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 352 * sizeof(int8_t))));
481 const __m128i vi14xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 20));
482 const __m128i vk14xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 356 * sizeof(int8_t))));
483 i14 += 24;
484
485 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
486 vacc4567 = _mm_macc_epi32(vi14x4567, vk14x4567, vacc4567);
487 vacc89AB = _mm_macc_epi32(vi14x89AB, vk14x89AB, vacc89AB);
488 vaccCDEF = _mm_macc_epi32(vi14xCDEF, vk14xCDEF, vaccCDEF);
489 vaccGHIJ = _mm_macc_epi32(vi14xGHIJ, vk14xGHIJ, vaccGHIJ);
490 vaccKLMN = _mm_macc_epi32(vi14xKLMN, vk14xKLMN, vaccKLMN);
491
492 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15));
493 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 360 * sizeof(int8_t))));
494 const __m128i vi15x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 4));
495 const __m128i vk15x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 364 * sizeof(int8_t))));
496 const __m128i vi15x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 8));
497 const __m128i vk15x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 368 * sizeof(int8_t))));
498 const __m128i vi15xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 12));
499 const __m128i vk15xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 372 * sizeof(int8_t))));
500 const __m128i vi15xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 16));
501 const __m128i vk15xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 376 * sizeof(int8_t))));
502 const __m128i vi15xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 20));
503 const __m128i vk15xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 380 * sizeof(int8_t))));
504 i15 += 24;
505
506 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
507 vacc4567 = _mm_macc_epi32(vi15x4567, vk15x4567, vacc4567);
508 vacc89AB = _mm_macc_epi32(vi15x89AB, vk15x89AB, vacc89AB);
509 vaccCDEF = _mm_macc_epi32(vi15xCDEF, vk15xCDEF, vaccCDEF);
510 vaccGHIJ = _mm_macc_epi32(vi15xGHIJ, vk15xGHIJ, vaccGHIJ);
511 vaccKLMN = _mm_macc_epi32(vi15xKLMN, vk15xKLMN, vaccKLMN);
512
513 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16));
514 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 384 * sizeof(int8_t))));
515 const __m128i vi16x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 4));
516 const __m128i vk16x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 388 * sizeof(int8_t))));
517 const __m128i vi16x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 8));
518 const __m128i vk16x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 392 * sizeof(int8_t))));
519 const __m128i vi16xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 12));
520 const __m128i vk16xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 396 * sizeof(int8_t))));
521 const __m128i vi16xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 16));
522 const __m128i vk16xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 400 * sizeof(int8_t))));
523 const __m128i vi16xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 20));
524 const __m128i vk16xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 404 * sizeof(int8_t))));
525 i16 += 24;
526
527 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
528 vacc4567 = _mm_macc_epi32(vi16x4567, vk16x4567, vacc4567);
529 vacc89AB = _mm_macc_epi32(vi16x89AB, vk16x89AB, vacc89AB);
530 vaccCDEF = _mm_macc_epi32(vi16xCDEF, vk16xCDEF, vaccCDEF);
531 vaccGHIJ = _mm_macc_epi32(vi16xGHIJ, vk16xGHIJ, vaccGHIJ);
532 vaccKLMN = _mm_macc_epi32(vi16xKLMN, vk16xKLMN, vaccKLMN);
533
534 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17));
535 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 408 * sizeof(int8_t))));
536 const __m128i vi17x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 4));
537 const __m128i vk17x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 412 * sizeof(int8_t))));
538 const __m128i vi17x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 8));
539 const __m128i vk17x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 416 * sizeof(int8_t))));
540 const __m128i vi17xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 12));
541 const __m128i vk17xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 420 * sizeof(int8_t))));
542 const __m128i vi17xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 16));
543 const __m128i vk17xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 424 * sizeof(int8_t))));
544 const __m128i vi17xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 20));
545 const __m128i vk17xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 428 * sizeof(int8_t))));
546 i17 += 24;
547
548 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
549 vacc4567 = _mm_macc_epi32(vi17x4567, vk17x4567, vacc4567);
550 vacc89AB = _mm_macc_epi32(vi17x89AB, vk17x89AB, vacc89AB);
551 vaccCDEF = _mm_macc_epi32(vi17xCDEF, vk17xCDEF, vaccCDEF);
552 vaccGHIJ = _mm_macc_epi32(vi17xGHIJ, vk17xGHIJ, vaccGHIJ);
553 vaccKLMN = _mm_macc_epi32(vi17xKLMN, vk17xKLMN, vaccKLMN);
554
555 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18));
556 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 432 * sizeof(int8_t))));
557 const __m128i vi18x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 4));
558 const __m128i vk18x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 436 * sizeof(int8_t))));
559 const __m128i vi18x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 8));
560 const __m128i vk18x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 440 * sizeof(int8_t))));
561 const __m128i vi18xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 12));
562 const __m128i vk18xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 444 * sizeof(int8_t))));
563 const __m128i vi18xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 16));
564 const __m128i vk18xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 448 * sizeof(int8_t))));
565 const __m128i vi18xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 20));
566 const __m128i vk18xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 452 * sizeof(int8_t))));
567 i18 += 24;
568
569 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
570 vacc4567 = _mm_macc_epi32(vi18x4567, vk18x4567, vacc4567);
571 vacc89AB = _mm_macc_epi32(vi18x89AB, vk18x89AB, vacc89AB);
572 vaccCDEF = _mm_macc_epi32(vi18xCDEF, vk18xCDEF, vaccCDEF);
573 vaccGHIJ = _mm_macc_epi32(vi18xGHIJ, vk18xGHIJ, vaccGHIJ);
574 vaccKLMN = _mm_macc_epi32(vi18xKLMN, vk18xKLMN, vaccKLMN);
575
576 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19));
577 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 456 * sizeof(int8_t))));
578 const __m128i vi19x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 4));
579 const __m128i vk19x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 460 * sizeof(int8_t))));
580 const __m128i vi19x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 8));
581 const __m128i vk19x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 464 * sizeof(int8_t))));
582 const __m128i vi19xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 12));
583 const __m128i vk19xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 468 * sizeof(int8_t))));
584 const __m128i vi19xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 16));
585 const __m128i vk19xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 472 * sizeof(int8_t))));
586 const __m128i vi19xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 20));
587 const __m128i vk19xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 476 * sizeof(int8_t))));
588 i19 += 24;
589
590 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
591 vacc4567 = _mm_macc_epi32(vi19x4567, vk19x4567, vacc4567);
592 vacc89AB = _mm_macc_epi32(vi19x89AB, vk19x89AB, vacc89AB);
593 vaccCDEF = _mm_macc_epi32(vi19xCDEF, vk19xCDEF, vaccCDEF);
594 vaccGHIJ = _mm_macc_epi32(vi19xGHIJ, vk19xGHIJ, vaccGHIJ);
595 vaccKLMN = _mm_macc_epi32(vi19xKLMN, vk19xKLMN, vaccKLMN);
596
597 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20));
598 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 480 * sizeof(int8_t))));
599 const __m128i vi20x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 4));
600 const __m128i vk20x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 484 * sizeof(int8_t))));
601 const __m128i vi20x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 8));
602 const __m128i vk20x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 488 * sizeof(int8_t))));
603 const __m128i vi20xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 12));
604 const __m128i vk20xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 492 * sizeof(int8_t))));
605 const __m128i vi20xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 16));
606 const __m128i vk20xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 496 * sizeof(int8_t))));
607 const __m128i vi20xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 20));
608 const __m128i vk20xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 500 * sizeof(int8_t))));
609 i20 += 24;
610
611 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
612 vacc4567 = _mm_macc_epi32(vi20x4567, vk20x4567, vacc4567);
613 vacc89AB = _mm_macc_epi32(vi20x89AB, vk20x89AB, vacc89AB);
614 vaccCDEF = _mm_macc_epi32(vi20xCDEF, vk20xCDEF, vaccCDEF);
615 vaccGHIJ = _mm_macc_epi32(vi20xGHIJ, vk20xGHIJ, vaccGHIJ);
616 vaccKLMN = _mm_macc_epi32(vi20xKLMN, vk20xKLMN, vaccKLMN);
617
618 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21));
619 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 504 * sizeof(int8_t))));
620 const __m128i vi21x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 4));
621 const __m128i vk21x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 508 * sizeof(int8_t))));
622 const __m128i vi21x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 8));
623 const __m128i vk21x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 512 * sizeof(int8_t))));
624 const __m128i vi21xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 12));
625 const __m128i vk21xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 516 * sizeof(int8_t))));
626 const __m128i vi21xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 16));
627 const __m128i vk21xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 520 * sizeof(int8_t))));
628 const __m128i vi21xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 20));
629 const __m128i vk21xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 524 * sizeof(int8_t))));
630 i21 += 24;
631
632 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
633 vacc4567 = _mm_macc_epi32(vi21x4567, vk21x4567, vacc4567);
634 vacc89AB = _mm_macc_epi32(vi21x89AB, vk21x89AB, vacc89AB);
635 vaccCDEF = _mm_macc_epi32(vi21xCDEF, vk21xCDEF, vaccCDEF);
636 vaccGHIJ = _mm_macc_epi32(vi21xGHIJ, vk21xGHIJ, vaccGHIJ);
637 vaccKLMN = _mm_macc_epi32(vi21xKLMN, vk21xKLMN, vaccKLMN);
638
639 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22));
640 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 528 * sizeof(int8_t))));
641 const __m128i vi22x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 4));
642 const __m128i vk22x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 532 * sizeof(int8_t))));
643 const __m128i vi22x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 8));
644 const __m128i vk22x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 536 * sizeof(int8_t))));
645 const __m128i vi22xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 12));
646 const __m128i vk22xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 540 * sizeof(int8_t))));
647 const __m128i vi22xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 16));
648 const __m128i vk22xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 544 * sizeof(int8_t))));
649 const __m128i vi22xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 20));
650 const __m128i vk22xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 548 * sizeof(int8_t))));
651 i22 += 24;
652
653 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
654 vacc4567 = _mm_macc_epi32(vi22x4567, vk22x4567, vacc4567);
655 vacc89AB = _mm_macc_epi32(vi22x89AB, vk22x89AB, vacc89AB);
656 vaccCDEF = _mm_macc_epi32(vi22xCDEF, vk22xCDEF, vaccCDEF);
657 vaccGHIJ = _mm_macc_epi32(vi22xGHIJ, vk22xGHIJ, vaccGHIJ);
658 vaccKLMN = _mm_macc_epi32(vi22xKLMN, vk22xKLMN, vaccKLMN);
659
660 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23));
661 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 552 * sizeof(int8_t))));
662 const __m128i vi23x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 4));
663 const __m128i vk23x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 556 * sizeof(int8_t))));
664 const __m128i vi23x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 8));
665 const __m128i vk23x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 560 * sizeof(int8_t))));
666 const __m128i vi23xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 12));
667 const __m128i vk23xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 564 * sizeof(int8_t))));
668 const __m128i vi23xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 16));
669 const __m128i vk23xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 568 * sizeof(int8_t))));
670 const __m128i vi23xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 20));
671 const __m128i vk23xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 572 * sizeof(int8_t))));
672 i23 += 24;
673
674 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
675 vacc4567 = _mm_macc_epi32(vi23x4567, vk23x4567, vacc4567);
676 vacc89AB = _mm_macc_epi32(vi23x89AB, vk23x89AB, vacc89AB);
677 vaccCDEF = _mm_macc_epi32(vi23xCDEF, vk23xCDEF, vaccCDEF);
678 vaccGHIJ = _mm_macc_epi32(vi23xGHIJ, vk23xGHIJ, vaccGHIJ);
679 vaccKLMN = _mm_macc_epi32(vi23xKLMN, vk23xKLMN, vaccKLMN);
680
681 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24));
682 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 576 * sizeof(int8_t))));
683 const __m128i vi24x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 4));
684 const __m128i vk24x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 580 * sizeof(int8_t))));
685 const __m128i vi24x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 8));
686 const __m128i vk24x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 584 * sizeof(int8_t))));
687 const __m128i vi24xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 12));
688 const __m128i vk24xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 588 * sizeof(int8_t))));
689 const __m128i vi24xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 16));
690 const __m128i vk24xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 592 * sizeof(int8_t))));
691 const __m128i vi24xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 20));
692 const __m128i vk24xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 596 * sizeof(int8_t))));
693 i24 += 24;
694
695 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
696 vacc4567 = _mm_macc_epi32(vi24x4567, vk24x4567, vacc4567);
697 vacc89AB = _mm_macc_epi32(vi24x89AB, vk24x89AB, vacc89AB);
698 vaccCDEF = _mm_macc_epi32(vi24xCDEF, vk24xCDEF, vaccCDEF);
699 vaccGHIJ = _mm_macc_epi32(vi24xGHIJ, vk24xGHIJ, vaccGHIJ);
700 vaccKLMN = _mm_macc_epi32(vi24xKLMN, vk24xKLMN, vaccKLMN);
701
702 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t));
703
704 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
705 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
706 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
707 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
708 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
709 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
710
711 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
712 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
713 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
714 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
715 const __m128 vscaleGHIJ = _mm_loadu_ps((const float*) w + 16);
716 const __m128 vscaleKLMN = _mm_loadu_ps((const float*) w + 20);
717 w = (const void*) ((const float*) w + 24);
718 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
719 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
720 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
721 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
722 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscaleGHIJ);
723 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscaleKLMN);
724
725 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
726 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
727 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
728 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
729 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
730 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
731 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
732
733 vacc0123 = _mm_cvtps_epi32(vscaled0123);
734 vacc4567 = _mm_cvtps_epi32(vscaled4567);
735 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
736 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
737 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
738 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
739
740 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
741 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
742 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
743 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
744
745 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
746 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
747 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
748 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
749 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
750
751 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
752 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
753 output += 24;
754 }
755 if XNN_UNLIKELY(c != 0) {
756 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
757 do {
758 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
759
760 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
761 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(k));
762 i0 += 4;
763
764 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
765 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
766 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 24)));
767 i1 += 4;
768
769 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
770 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
771 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 48)));
772 i2 += 4;
773
774 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
775 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
776 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 72)));
777 i3 += 4;
778
779 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
780 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
781 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 96)));
782 i4 += 4;
783
784 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
785 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
786 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 120)));
787 i5 += 4;
788
789 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
790 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
791 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 144)));
792 i6 += 4;
793
794 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
795 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
796 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 168)));
797 i7 += 4;
798
799 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
800 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
801 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 192)));
802 i8 += 4;
803
804 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
805 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9));
806 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 216)));
807 i9 += 4;
808
809 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
810 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10));
811 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 240)));
812 i10 += 4;
813
814 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
815 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11));
816 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 264)));
817 i11 += 4;
818
819 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
820 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12));
821 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 288)));
822 i12 += 4;
823
824 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
825 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13));
826 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 312)));
827 i13 += 4;
828
829 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
830 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14));
831 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 336)));
832 i14 += 4;
833
834 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
835 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15));
836 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 360)));
837 i15 += 4;
838
839 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
840 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16));
841 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 384)));
842 i16 += 4;
843
844 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
845 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17));
846 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 408)));
847 i17 += 4;
848
849 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
850 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18));
851 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 432)));
852 i18 += 4;
853
854 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
855 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19));
856 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 456)));
857 i19 += 4;
858
859 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
860 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20));
861 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 480)));
862 i20 += 4;
863
864 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
865 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21));
866 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 504)));
867 i21 += 4;
868
869 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
870 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22));
871 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 528)));
872 i22 += 4;
873
874 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
875 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23));
876 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 552)));
877 i23 += 4;
878
879 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
880 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24));
881 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 576)));
882 i24 += 4;
883
884 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
885
886 k += 4;
887
888 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
889 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t)));
890 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
891 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->sse4.output_max_less_zero_point));
892 vacc0123 = _mm_cvtps_epi32(vscaled0123);
893
894 w = (const void*) ((const int32_t*) w + 4);
895
896 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
897 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
898
899 vout0123 = _mm_packs_epi16(vout0123, vout0123);
900 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->sse4.output_min));
901
902 if XNN_LIKELY(c >= 4) {
903 _mm_storeu_si32(output, vout0123);
904 output += 4;
905 c -= 4;
906 } else {
907 if (c & 2) {
908 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
909 vout0123 = _mm_srli_epi32(vout0123, 16);
910 output += 2;
911 }
912 if (c & 1) {
913 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
914 output += 1;
915 }
916 c = 0;
917 }
918 } while (c != 0);
919 }
920
921 output = (int8_t*) ((uintptr_t) output + output_increment);
922 } while (--output_width != 0);
923 }
924