1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul32(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 do {
34 const int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const int8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const int8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const int8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const int8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const int8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const int8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 const int8_t* i9 = input[9];
80 assert(i9 != NULL);
81 if XNN_UNPREDICTABLE(i9 != zero) {
82 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
83 }
84 const int8_t* i10 = input[10];
85 assert(i10 != NULL);
86 if XNN_UNPREDICTABLE(i10 != zero) {
87 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
88 }
89 const int8_t* i11 = input[11];
90 assert(i11 != NULL);
91 if XNN_UNPREDICTABLE(i11 != zero) {
92 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
93 }
94 const int8_t* i12 = input[12];
95 assert(i12 != NULL);
96 if XNN_UNPREDICTABLE(i12 != zero) {
97 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
98 }
99 const int8_t* i13 = input[13];
100 assert(i13 != NULL);
101 if XNN_UNPREDICTABLE(i13 != zero) {
102 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
103 }
104 const int8_t* i14 = input[14];
105 assert(i14 != NULL);
106 if XNN_UNPREDICTABLE(i14 != zero) {
107 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
108 }
109 const int8_t* i15 = input[15];
110 assert(i15 != NULL);
111 if XNN_UNPREDICTABLE(i15 != zero) {
112 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
113 }
114 const int8_t* i16 = input[16];
115 assert(i16 != NULL);
116 if XNN_UNPREDICTABLE(i16 != zero) {
117 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
118 }
119 const int8_t* i17 = input[17];
120 assert(i17 != NULL);
121 if XNN_UNPREDICTABLE(i17 != zero) {
122 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
123 }
124 const int8_t* i18 = input[18];
125 assert(i18 != NULL);
126 if XNN_UNPREDICTABLE(i18 != zero) {
127 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
128 }
129 const int8_t* i19 = input[19];
130 assert(i19 != NULL);
131 if XNN_UNPREDICTABLE(i19 != zero) {
132 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
133 }
134 const int8_t* i20 = input[20];
135 assert(i20 != NULL);
136 if XNN_UNPREDICTABLE(i20 != zero) {
137 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
138 }
139 const int8_t* i21 = input[21];
140 assert(i21 != NULL);
141 if XNN_UNPREDICTABLE(i21 != zero) {
142 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
143 }
144 const int8_t* i22 = input[22];
145 assert(i22 != NULL);
146 if XNN_UNPREDICTABLE(i22 != zero) {
147 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
148 }
149 const int8_t* i23 = input[23];
150 assert(i23 != NULL);
151 if XNN_UNPREDICTABLE(i23 != zero) {
152 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
153 }
154 const int8_t* i24 = input[24];
155 assert(i24 != NULL);
156 if XNN_UNPREDICTABLE(i24 != zero) {
157 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
158 }
159 input = (const int8_t**) ((uintptr_t) input + input_stride);
160
161 size_t c = channels;
162 const void* w = weights;
163 for (; c >= 16; c -= 16) {
164 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
165 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
166 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
167 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
168
169
170 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
171 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
172 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 4));
173 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(int8_t))));
174 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 8));
175 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t))));
176 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 12));
177 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(int8_t))));
178 i0 += 16;
179
180 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
181 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
182 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
183 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
184
185 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
186 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
187 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 4));
188 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(int8_t))));
189 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 8));
190 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t))));
191 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 12));
192 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(int8_t))));
193 i1 += 16;
194
195 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
196 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
197 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
198 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
199
200 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
201 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
202 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 4));
203 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(int8_t))));
204 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 8));
205 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t))));
206 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 12));
207 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(int8_t))));
208 i2 += 16;
209
210 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
211 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
212 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
213 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
214
215 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
216 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
217 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 4));
218 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(int8_t))));
219 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 8));
220 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t))));
221 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 12));
222 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(int8_t))));
223 i3 += 16;
224
225 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
226 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
227 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
228 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
229
230 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
231 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
232 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 4));
233 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(int8_t))));
234 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 8));
235 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t))));
236 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 12));
237 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(int8_t))));
238 i4 += 16;
239
240 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
241 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
242 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
243 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
244
245 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
246 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
247 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 4));
248 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(int8_t))));
249 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 8));
250 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t))));
251 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 12));
252 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(int8_t))));
253 i5 += 16;
254
255 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
256 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
257 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
258 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
259
260 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
261 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
262 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 4));
263 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 100 * sizeof(int8_t))));
264 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 8));
265 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t))));
266 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 12));
267 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 108 * sizeof(int8_t))));
268 i6 += 16;
269
270 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
271 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
272 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi6x89AB, vk6x89AB));
273 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi6xCDEF, vk6xCDEF));
274
275 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
276 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
277 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 4));
278 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 116 * sizeof(int8_t))));
279 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 8));
280 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t))));
281 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 12));
282 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 124 * sizeof(int8_t))));
283 i7 += 16;
284
285 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
286 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
287 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi7x89AB, vk7x89AB));
288 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi7xCDEF, vk7xCDEF));
289
290 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
291 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
292 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 4));
293 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 132 * sizeof(int8_t))));
294 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 8));
295 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t))));
296 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 12));
297 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 140 * sizeof(int8_t))));
298 i8 += 16;
299
300 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
301 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
302 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi8x89AB, vk8x89AB));
303 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi8xCDEF, vk8xCDEF));
304
305 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9));
306 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t))));
307 const __m128i vi9x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 4));
308 const __m128i vk9x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 148 * sizeof(int8_t))));
309 const __m128i vi9x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 8));
310 const __m128i vk9x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t))));
311 const __m128i vi9xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i9 + 12));
312 const __m128i vk9xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 156 * sizeof(int8_t))));
313 i9 += 16;
314
315 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
316 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi9x4567, vk9x4567));
317 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi9x89AB, vk9x89AB));
318 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi9xCDEF, vk9xCDEF));
319
320 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10));
321 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t))));
322 const __m128i vi10x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 4));
323 const __m128i vk10x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 164 * sizeof(int8_t))));
324 const __m128i vi10x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 8));
325 const __m128i vk10x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t))));
326 const __m128i vi10xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i10 + 12));
327 const __m128i vk10xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 172 * sizeof(int8_t))));
328 i10 += 16;
329
330 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
331 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi10x4567, vk10x4567));
332 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi10x89AB, vk10x89AB));
333 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi10xCDEF, vk10xCDEF));
334
335 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11));
336 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t))));
337 const __m128i vi11x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 4));
338 const __m128i vk11x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 180 * sizeof(int8_t))));
339 const __m128i vi11x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 8));
340 const __m128i vk11x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t))));
341 const __m128i vi11xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i11 + 12));
342 const __m128i vk11xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 188 * sizeof(int8_t))));
343 i11 += 16;
344
345 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
346 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi11x4567, vk11x4567));
347 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi11x89AB, vk11x89AB));
348 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi11xCDEF, vk11xCDEF));
349
350 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12));
351 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t))));
352 const __m128i vi12x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 4));
353 const __m128i vk12x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 196 * sizeof(int8_t))));
354 const __m128i vi12x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 8));
355 const __m128i vk12x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t))));
356 const __m128i vi12xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i12 + 12));
357 const __m128i vk12xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 204 * sizeof(int8_t))));
358 i12 += 16;
359
360 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
361 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi12x4567, vk12x4567));
362 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi12x89AB, vk12x89AB));
363 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi12xCDEF, vk12xCDEF));
364
365 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13));
366 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t))));
367 const __m128i vi13x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 4));
368 const __m128i vk13x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 212 * sizeof(int8_t))));
369 const __m128i vi13x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 8));
370 const __m128i vk13x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t))));
371 const __m128i vi13xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i13 + 12));
372 const __m128i vk13xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 220 * sizeof(int8_t))));
373 i13 += 16;
374
375 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
376 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi13x4567, vk13x4567));
377 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi13x89AB, vk13x89AB));
378 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi13xCDEF, vk13xCDEF));
379
380 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14));
381 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t))));
382 const __m128i vi14x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 4));
383 const __m128i vk14x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 228 * sizeof(int8_t))));
384 const __m128i vi14x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 8));
385 const __m128i vk14x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t))));
386 const __m128i vi14xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i14 + 12));
387 const __m128i vk14xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 236 * sizeof(int8_t))));
388 i14 += 16;
389
390 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
391 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi14x4567, vk14x4567));
392 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi14x89AB, vk14x89AB));
393 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi14xCDEF, vk14xCDEF));
394
395 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15));
396 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t))));
397 const __m128i vi15x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 4));
398 const __m128i vk15x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 244 * sizeof(int8_t))));
399 const __m128i vi15x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 8));
400 const __m128i vk15x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t))));
401 const __m128i vi15xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i15 + 12));
402 const __m128i vk15xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 252 * sizeof(int8_t))));
403 i15 += 16;
404
405 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
406 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi15x4567, vk15x4567));
407 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi15x89AB, vk15x89AB));
408 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi15xCDEF, vk15xCDEF));
409
410 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16));
411 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t))));
412 const __m128i vi16x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 4));
413 const __m128i vk16x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 260 * sizeof(int8_t))));
414 const __m128i vi16x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 8));
415 const __m128i vk16x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t))));
416 const __m128i vi16xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i16 + 12));
417 const __m128i vk16xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 268 * sizeof(int8_t))));
418 i16 += 16;
419
420 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
421 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi16x4567, vk16x4567));
422 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi16x89AB, vk16x89AB));
423 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi16xCDEF, vk16xCDEF));
424
425 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17));
426 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t))));
427 const __m128i vi17x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 4));
428 const __m128i vk17x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 276 * sizeof(int8_t))));
429 const __m128i vi17x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 8));
430 const __m128i vk17x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t))));
431 const __m128i vi17xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i17 + 12));
432 const __m128i vk17xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 284 * sizeof(int8_t))));
433 i17 += 16;
434
435 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
436 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi17x4567, vk17x4567));
437 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi17x89AB, vk17x89AB));
438 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi17xCDEF, vk17xCDEF));
439
440 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18));
441 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t))));
442 const __m128i vi18x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 4));
443 const __m128i vk18x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 292 * sizeof(int8_t))));
444 const __m128i vi18x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 8));
445 const __m128i vk18x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t))));
446 const __m128i vi18xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i18 + 12));
447 const __m128i vk18xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 300 * sizeof(int8_t))));
448 i18 += 16;
449
450 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
451 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi18x4567, vk18x4567));
452 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi18x89AB, vk18x89AB));
453 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi18xCDEF, vk18xCDEF));
454
455 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19));
456 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t))));
457 const __m128i vi19x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 4));
458 const __m128i vk19x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 308 * sizeof(int8_t))));
459 const __m128i vi19x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 8));
460 const __m128i vk19x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t))));
461 const __m128i vi19xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i19 + 12));
462 const __m128i vk19xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 316 * sizeof(int8_t))));
463 i19 += 16;
464
465 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
466 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi19x4567, vk19x4567));
467 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi19x89AB, vk19x89AB));
468 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi19xCDEF, vk19xCDEF));
469
470 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20));
471 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t))));
472 const __m128i vi20x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 4));
473 const __m128i vk20x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 324 * sizeof(int8_t))));
474 const __m128i vi20x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 8));
475 const __m128i vk20x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t))));
476 const __m128i vi20xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i20 + 12));
477 const __m128i vk20xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 332 * sizeof(int8_t))));
478 i20 += 16;
479
480 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
481 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi20x4567, vk20x4567));
482 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi20x89AB, vk20x89AB));
483 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi20xCDEF, vk20xCDEF));
484
485 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21));
486 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t))));
487 const __m128i vi21x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 4));
488 const __m128i vk21x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 340 * sizeof(int8_t))));
489 const __m128i vi21x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 8));
490 const __m128i vk21x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t))));
491 const __m128i vi21xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i21 + 12));
492 const __m128i vk21xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 348 * sizeof(int8_t))));
493 i21 += 16;
494
495 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
496 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi21x4567, vk21x4567));
497 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi21x89AB, vk21x89AB));
498 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi21xCDEF, vk21xCDEF));
499
500 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22));
501 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t))));
502 const __m128i vi22x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 4));
503 const __m128i vk22x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 356 * sizeof(int8_t))));
504 const __m128i vi22x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 8));
505 const __m128i vk22x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t))));
506 const __m128i vi22xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i22 + 12));
507 const __m128i vk22xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 364 * sizeof(int8_t))));
508 i22 += 16;
509
510 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
511 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi22x4567, vk22x4567));
512 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi22x89AB, vk22x89AB));
513 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi22xCDEF, vk22xCDEF));
514
515 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23));
516 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t))));
517 const __m128i vi23x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 4));
518 const __m128i vk23x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 372 * sizeof(int8_t))));
519 const __m128i vi23x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 8));
520 const __m128i vk23x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t))));
521 const __m128i vi23xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i23 + 12));
522 const __m128i vk23xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 380 * sizeof(int8_t))));
523 i23 += 16;
524
525 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
526 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi23x4567, vk23x4567));
527 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi23x89AB, vk23x89AB));
528 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi23xCDEF, vk23xCDEF));
529
530 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24));
531 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t))));
532 const __m128i vi24x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 4));
533 const __m128i vk24x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 388 * sizeof(int8_t))));
534 const __m128i vi24x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 8));
535 const __m128i vk24x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t))));
536 const __m128i vi24xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i24 + 12));
537 const __m128i vk24xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 396 * sizeof(int8_t))));
538 i24 += 16;
539
540 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
541 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi24x4567, vk24x4567));
542 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi24x89AB, vk24x89AB));
543 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi24xCDEF, vk24xCDEF));
544
545 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
546
547 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
548 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
549 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
550 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
551
552 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
553 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
554 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
555 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
556 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
557
558 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
559 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
560 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
561 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
562 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
563
564 vacc0123 = _mm_cvtps_epi32(vscaled0123);
565 vacc4567 = _mm_cvtps_epi32(vscaled4567);
566 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
567 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
568
569 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
570 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
571 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
572
573 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
574 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
575 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
576
577 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
578 output += 16;
579 }
580 if XNN_UNLIKELY(c != 0) {
581 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
582 do {
583 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
584
585 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
586 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(k));
587 i0 += 4;
588
589 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
590 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
591 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 16)));
592 i1 += 4;
593
594 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
595 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
596 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 32)));
597 i2 += 4;
598
599 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
600 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
601 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 48)));
602 i3 += 4;
603
604 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
605 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
606 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 64)));
607 i4 += 4;
608
609 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
610 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
611 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 80)));
612 i5 += 4;
613
614 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
615 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
616 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 96)));
617 i6 += 4;
618
619 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
620 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
621 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 112)));
622 i7 += 4;
623
624 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
625 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
626 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 128)));
627 i8 += 4;
628
629 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
630 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i9));
631 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 144)));
632 i9 += 4;
633
634 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
635 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i10));
636 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 160)));
637 i10 += 4;
638
639 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
640 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i11));
641 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 176)));
642 i11 += 4;
643
644 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
645 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i12));
646 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 192)));
647 i12 += 4;
648
649 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
650 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i13));
651 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 208)));
652 i13 += 4;
653
654 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
655 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i14));
656 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 224)));
657 i14 += 4;
658
659 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
660 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i15));
661 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 240)));
662 i15 += 4;
663
664 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
665 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i16));
666 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 256)));
667 i16 += 4;
668
669 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
670 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i17));
671 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 272)));
672 i17 += 4;
673
674 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
675 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i18));
676 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 288)));
677 i18 += 4;
678
679 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
680 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i19));
681 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 304)));
682 i19 += 4;
683
684 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
685 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i20));
686 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 320)));
687 i20 += 4;
688
689 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
690 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i21));
691 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 336)));
692 i21 += 4;
693
694 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
695 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i22));
696 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 352)));
697 i22 += 4;
698
699 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
700 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i23));
701 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 368)));
702 i23 += 4;
703
704 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
705 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i24));
706 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 384)));
707 i24 += 4;
708
709 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
710
711 k += 4;
712
713 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
714 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
715 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
716 vacc0123 = _mm_cvtps_epi32(vscaled0123);
717
718 w = (const void*) ((const int32_t*) w + 4);
719
720 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
721 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
722
723 vout0123 = _mm_packs_epi16(vout0123, vout0123);
724 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
725
726 if XNN_LIKELY(c >= 4) {
727 _mm_storeu_si32(output, vout0123);
728 output += 4;
729 c -= 4;
730 } else {
731 if (c & 2) {
732 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
733 vout0123 = _mm_srli_epi32(vout0123, 16);
734 output += 2;
735 }
736 if (c & 1) {
737 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
738 output += 1;
739 }
740 c = 0;
741 }
742 } while (c != 0);
743 }
744
745 output = (int8_t*) ((uintptr_t) output + output_increment);
746 } while (--output_width != 0);
747 }
748