1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/dwconv.h>
20 #include <xnnpack/intrinsics-polyfill.h>
21 #include <xnnpack/unaligned.h>
22
23
xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__xop_mul32(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__xop_mul32(
25 size_t channels,
26 size_t output_width,
27 const int8_t** input,
28 const void* weights,
29 int8_t* output,
30 size_t input_stride,
31 size_t output_increment,
32 size_t input_offset,
33 const int8_t* zero,
34 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(channels != 0);
37 assert(output_width != 0);
38
39 do {
40 const int8_t* i0 = input[0];
41 assert(i0 != NULL);
42 if XNN_UNPREDICTABLE(i0 != zero) {
43 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
44 }
45 const int8_t* i1 = input[1];
46 assert(i1 != NULL);
47 if XNN_UNPREDICTABLE(i1 != zero) {
48 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
49 }
50 const int8_t* i2 = input[2];
51 assert(i2 != NULL);
52 if XNN_UNPREDICTABLE(i2 != zero) {
53 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
54 }
55 const int8_t* i3 = input[3];
56 assert(i3 != NULL);
57 if XNN_UNPREDICTABLE(i3 != zero) {
58 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
59 }
60 const int8_t* i4 = input[4];
61 assert(i4 != NULL);
62 if XNN_UNPREDICTABLE(i4 != zero) {
63 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
64 }
65 const int8_t* i5 = input[5];
66 assert(i5 != NULL);
67 if XNN_UNPREDICTABLE(i5 != zero) {
68 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
69 }
70 const int8_t* i6 = input[6];
71 assert(i6 != NULL);
72 if XNN_UNPREDICTABLE(i6 != zero) {
73 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
74 }
75 const int8_t* i7 = input[7];
76 assert(i7 != NULL);
77 if XNN_UNPREDICTABLE(i7 != zero) {
78 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
79 }
80 const int8_t* i8 = input[8];
81 assert(i8 != NULL);
82 if XNN_UNPREDICTABLE(i8 != zero) {
83 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
84 }
85 const int8_t* i9 = input[9];
86 assert(i9 != NULL);
87 if XNN_UNPREDICTABLE(i9 != zero) {
88 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
89 }
90 const int8_t* i10 = input[10];
91 assert(i10 != NULL);
92 if XNN_UNPREDICTABLE(i10 != zero) {
93 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
94 }
95 const int8_t* i11 = input[11];
96 assert(i11 != NULL);
97 if XNN_UNPREDICTABLE(i11 != zero) {
98 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
99 }
100 const int8_t* i12 = input[12];
101 assert(i12 != NULL);
102 if XNN_UNPREDICTABLE(i12 != zero) {
103 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
104 }
105 const int8_t* i13 = input[13];
106 assert(i13 != NULL);
107 if XNN_UNPREDICTABLE(i13 != zero) {
108 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
109 }
110 const int8_t* i14 = input[14];
111 assert(i14 != NULL);
112 if XNN_UNPREDICTABLE(i14 != zero) {
113 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
114 }
115 const int8_t* i15 = input[15];
116 assert(i15 != NULL);
117 if XNN_UNPREDICTABLE(i15 != zero) {
118 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
119 }
120 const int8_t* i16 = input[16];
121 assert(i16 != NULL);
122 if XNN_UNPREDICTABLE(i16 != zero) {
123 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
124 }
125 const int8_t* i17 = input[17];
126 assert(i17 != NULL);
127 if XNN_UNPREDICTABLE(i17 != zero) {
128 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
129 }
130 const int8_t* i18 = input[18];
131 assert(i18 != NULL);
132 if XNN_UNPREDICTABLE(i18 != zero) {
133 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
134 }
135 const int8_t* i19 = input[19];
136 assert(i19 != NULL);
137 if XNN_UNPREDICTABLE(i19 != zero) {
138 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
139 }
140 const int8_t* i20 = input[20];
141 assert(i20 != NULL);
142 if XNN_UNPREDICTABLE(i20 != zero) {
143 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
144 }
145 const int8_t* i21 = input[21];
146 assert(i21 != NULL);
147 if XNN_UNPREDICTABLE(i21 != zero) {
148 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
149 }
150 const int8_t* i22 = input[22];
151 assert(i22 != NULL);
152 if XNN_UNPREDICTABLE(i22 != zero) {
153 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
154 }
155 const int8_t* i23 = input[23];
156 assert(i23 != NULL);
157 if XNN_UNPREDICTABLE(i23 != zero) {
158 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
159 }
160 const int8_t* i24 = input[24];
161 assert(i24 != NULL);
162 if XNN_UNPREDICTABLE(i24 != zero) {
163 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
164 }
165 input = (const int8_t**) ((uintptr_t) input + input_stride);
166
167 size_t c = channels;
168 const void* w = weights;
169 for (; c >= 16; c -= 16) {
170 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
171 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
172 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
173 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
174
175
176 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
177 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
178 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
179 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
180 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
181 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
182 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
183 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
184 i0 += 16;
185
186 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
187 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
188 vacc89AB = _mm_macc_epi32(vi0x89AB, vk0x89AB, vacc89AB);
189 vaccCDEF = _mm_macc_epi32(vi0xCDEF, vk0xCDEF, vaccCDEF);
190
191 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
192 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
193 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
194 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
195 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
196 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
197 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
198 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
199 i1 += 16;
200
201 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
202 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
203 vacc89AB = _mm_macc_epi32(vi1x89AB, vk1x89AB, vacc89AB);
204 vaccCDEF = _mm_macc_epi32(vi1xCDEF, vk1xCDEF, vaccCDEF);
205
206 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
207 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
208 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
209 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
210 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
211 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
212 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
213 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
214 i2 += 16;
215
216 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
217 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
218 vacc89AB = _mm_macc_epi32(vi2x89AB, vk2x89AB, vacc89AB);
219 vaccCDEF = _mm_macc_epi32(vi2xCDEF, vk2xCDEF, vaccCDEF);
220
221 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
222 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)))));
223 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
224 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(int8_t)))));
225 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
226 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)))));
227 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
228 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(int8_t)))));
229 i3 += 16;
230
231 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
232 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
233 vacc89AB = _mm_macc_epi32(vi3x89AB, vk3x89AB, vacc89AB);
234 vaccCDEF = _mm_macc_epi32(vi3xCDEF, vk3xCDEF, vaccCDEF);
235
236 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
237 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)))));
238 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
239 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(int8_t)))));
240 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
241 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)))));
242 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
243 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(int8_t)))));
244 i4 += 16;
245
246 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
247 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
248 vacc89AB = _mm_macc_epi32(vi4x89AB, vk4x89AB, vacc89AB);
249 vaccCDEF = _mm_macc_epi32(vi4xCDEF, vk4xCDEF, vaccCDEF);
250
251 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
252 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)))));
253 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
254 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(int8_t)))));
255 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
256 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)))));
257 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
258 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(int8_t)))));
259 i5 += 16;
260
261 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
262 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
263 vacc89AB = _mm_macc_epi32(vi5x89AB, vk5x89AB, vacc89AB);
264 vaccCDEF = _mm_macc_epi32(vi5xCDEF, vk5xCDEF, vaccCDEF);
265
266 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
267 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)))));
268 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
269 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 100 * sizeof(int8_t)))));
270 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
271 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)))));
272 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
273 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 108 * sizeof(int8_t)))));
274 i6 += 16;
275
276 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
277 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
278 vacc89AB = _mm_macc_epi32(vi6x89AB, vk6x89AB, vacc89AB);
279 vaccCDEF = _mm_macc_epi32(vi6xCDEF, vk6xCDEF, vaccCDEF);
280
281 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
282 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)))));
283 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
284 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 116 * sizeof(int8_t)))));
285 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 8)));
286 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)))));
287 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 12)));
288 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 124 * sizeof(int8_t)))));
289 i7 += 16;
290
291 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
292 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
293 vacc89AB = _mm_macc_epi32(vi7x89AB, vk7x89AB, vacc89AB);
294 vaccCDEF = _mm_macc_epi32(vi7xCDEF, vk7xCDEF, vaccCDEF);
295
296 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
297 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)))));
298 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
299 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 132 * sizeof(int8_t)))));
300 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 8)));
301 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)))));
302 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 12)));
303 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 140 * sizeof(int8_t)))));
304 i8 += 16;
305
306 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
307 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
308 vacc89AB = _mm_macc_epi32(vi8x89AB, vk8x89AB, vacc89AB);
309 vaccCDEF = _mm_macc_epi32(vi8xCDEF, vk8xCDEF, vaccCDEF);
310
311 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
312 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)))));
313 const __m128i vi9x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 4)));
314 const __m128i vk9x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 148 * sizeof(int8_t)))));
315 const __m128i vi9x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 8)));
316 const __m128i vk9x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)))));
317 const __m128i vi9xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 12)));
318 const __m128i vk9xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 156 * sizeof(int8_t)))));
319 i9 += 16;
320
321 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
322 vacc4567 = _mm_macc_epi32(vi9x4567, vk9x4567, vacc4567);
323 vacc89AB = _mm_macc_epi32(vi9x89AB, vk9x89AB, vacc89AB);
324 vaccCDEF = _mm_macc_epi32(vi9xCDEF, vk9xCDEF, vaccCDEF);
325
326 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
327 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)))));
328 const __m128i vi10x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 4)));
329 const __m128i vk10x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 164 * sizeof(int8_t)))));
330 const __m128i vi10x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 8)));
331 const __m128i vk10x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)))));
332 const __m128i vi10xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 12)));
333 const __m128i vk10xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 172 * sizeof(int8_t)))));
334 i10 += 16;
335
336 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
337 vacc4567 = _mm_macc_epi32(vi10x4567, vk10x4567, vacc4567);
338 vacc89AB = _mm_macc_epi32(vi10x89AB, vk10x89AB, vacc89AB);
339 vaccCDEF = _mm_macc_epi32(vi10xCDEF, vk10xCDEF, vaccCDEF);
340
341 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
342 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)))));
343 const __m128i vi11x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 4)));
344 const __m128i vk11x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 180 * sizeof(int8_t)))));
345 const __m128i vi11x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 8)));
346 const __m128i vk11x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)))));
347 const __m128i vi11xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 12)));
348 const __m128i vk11xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 188 * sizeof(int8_t)))));
349 i11 += 16;
350
351 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
352 vacc4567 = _mm_macc_epi32(vi11x4567, vk11x4567, vacc4567);
353 vacc89AB = _mm_macc_epi32(vi11x89AB, vk11x89AB, vacc89AB);
354 vaccCDEF = _mm_macc_epi32(vi11xCDEF, vk11xCDEF, vaccCDEF);
355
356 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
357 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)))));
358 const __m128i vi12x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 4)));
359 const __m128i vk12x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 196 * sizeof(int8_t)))));
360 const __m128i vi12x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 8)));
361 const __m128i vk12x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)))));
362 const __m128i vi12xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 12)));
363 const __m128i vk12xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 204 * sizeof(int8_t)))));
364 i12 += 16;
365
366 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
367 vacc4567 = _mm_macc_epi32(vi12x4567, vk12x4567, vacc4567);
368 vacc89AB = _mm_macc_epi32(vi12x89AB, vk12x89AB, vacc89AB);
369 vaccCDEF = _mm_macc_epi32(vi12xCDEF, vk12xCDEF, vaccCDEF);
370
371 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
372 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)))));
373 const __m128i vi13x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 4)));
374 const __m128i vk13x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 212 * sizeof(int8_t)))));
375 const __m128i vi13x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 8)));
376 const __m128i vk13x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)))));
377 const __m128i vi13xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 12)));
378 const __m128i vk13xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 220 * sizeof(int8_t)))));
379 i13 += 16;
380
381 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
382 vacc4567 = _mm_macc_epi32(vi13x4567, vk13x4567, vacc4567);
383 vacc89AB = _mm_macc_epi32(vi13x89AB, vk13x89AB, vacc89AB);
384 vaccCDEF = _mm_macc_epi32(vi13xCDEF, vk13xCDEF, vaccCDEF);
385
386 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
387 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)))));
388 const __m128i vi14x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 4)));
389 const __m128i vk14x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 228 * sizeof(int8_t)))));
390 const __m128i vi14x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 8)));
391 const __m128i vk14x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)))));
392 const __m128i vi14xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 12)));
393 const __m128i vk14xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 236 * sizeof(int8_t)))));
394 i14 += 16;
395
396 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
397 vacc4567 = _mm_macc_epi32(vi14x4567, vk14x4567, vacc4567);
398 vacc89AB = _mm_macc_epi32(vi14x89AB, vk14x89AB, vacc89AB);
399 vaccCDEF = _mm_macc_epi32(vi14xCDEF, vk14xCDEF, vaccCDEF);
400
401 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
402 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)))));
403 const __m128i vi15x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 4)));
404 const __m128i vk15x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 244 * sizeof(int8_t)))));
405 const __m128i vi15x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 8)));
406 const __m128i vk15x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)))));
407 const __m128i vi15xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 12)));
408 const __m128i vk15xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 252 * sizeof(int8_t)))));
409 i15 += 16;
410
411 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
412 vacc4567 = _mm_macc_epi32(vi15x4567, vk15x4567, vacc4567);
413 vacc89AB = _mm_macc_epi32(vi15x89AB, vk15x89AB, vacc89AB);
414 vaccCDEF = _mm_macc_epi32(vi15xCDEF, vk15xCDEF, vaccCDEF);
415
416 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
417 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)))));
418 const __m128i vi16x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 4)));
419 const __m128i vk16x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 260 * sizeof(int8_t)))));
420 const __m128i vi16x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 8)));
421 const __m128i vk16x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)))));
422 const __m128i vi16xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 12)));
423 const __m128i vk16xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 268 * sizeof(int8_t)))));
424 i16 += 16;
425
426 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
427 vacc4567 = _mm_macc_epi32(vi16x4567, vk16x4567, vacc4567);
428 vacc89AB = _mm_macc_epi32(vi16x89AB, vk16x89AB, vacc89AB);
429 vaccCDEF = _mm_macc_epi32(vi16xCDEF, vk16xCDEF, vaccCDEF);
430
431 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
432 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)))));
433 const __m128i vi17x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 4)));
434 const __m128i vk17x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 276 * sizeof(int8_t)))));
435 const __m128i vi17x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 8)));
436 const __m128i vk17x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)))));
437 const __m128i vi17xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 12)));
438 const __m128i vk17xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 284 * sizeof(int8_t)))));
439 i17 += 16;
440
441 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
442 vacc4567 = _mm_macc_epi32(vi17x4567, vk17x4567, vacc4567);
443 vacc89AB = _mm_macc_epi32(vi17x89AB, vk17x89AB, vacc89AB);
444 vaccCDEF = _mm_macc_epi32(vi17xCDEF, vk17xCDEF, vaccCDEF);
445
446 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
447 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)))));
448 const __m128i vi18x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 4)));
449 const __m128i vk18x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 292 * sizeof(int8_t)))));
450 const __m128i vi18x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 8)));
451 const __m128i vk18x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)))));
452 const __m128i vi18xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 12)));
453 const __m128i vk18xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 300 * sizeof(int8_t)))));
454 i18 += 16;
455
456 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
457 vacc4567 = _mm_macc_epi32(vi18x4567, vk18x4567, vacc4567);
458 vacc89AB = _mm_macc_epi32(vi18x89AB, vk18x89AB, vacc89AB);
459 vaccCDEF = _mm_macc_epi32(vi18xCDEF, vk18xCDEF, vaccCDEF);
460
461 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
462 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)))));
463 const __m128i vi19x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 4)));
464 const __m128i vk19x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 308 * sizeof(int8_t)))));
465 const __m128i vi19x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 8)));
466 const __m128i vk19x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)))));
467 const __m128i vi19xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 12)));
468 const __m128i vk19xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 316 * sizeof(int8_t)))));
469 i19 += 16;
470
471 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
472 vacc4567 = _mm_macc_epi32(vi19x4567, vk19x4567, vacc4567);
473 vacc89AB = _mm_macc_epi32(vi19x89AB, vk19x89AB, vacc89AB);
474 vaccCDEF = _mm_macc_epi32(vi19xCDEF, vk19xCDEF, vaccCDEF);
475
476 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
477 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)))));
478 const __m128i vi20x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 4)));
479 const __m128i vk20x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 324 * sizeof(int8_t)))));
480 const __m128i vi20x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 8)));
481 const __m128i vk20x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)))));
482 const __m128i vi20xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 12)));
483 const __m128i vk20xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 332 * sizeof(int8_t)))));
484 i20 += 16;
485
486 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
487 vacc4567 = _mm_macc_epi32(vi20x4567, vk20x4567, vacc4567);
488 vacc89AB = _mm_macc_epi32(vi20x89AB, vk20x89AB, vacc89AB);
489 vaccCDEF = _mm_macc_epi32(vi20xCDEF, vk20xCDEF, vaccCDEF);
490
491 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
492 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)))));
493 const __m128i vi21x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 4)));
494 const __m128i vk21x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 340 * sizeof(int8_t)))));
495 const __m128i vi21x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 8)));
496 const __m128i vk21x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)))));
497 const __m128i vi21xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 12)));
498 const __m128i vk21xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 348 * sizeof(int8_t)))));
499 i21 += 16;
500
501 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
502 vacc4567 = _mm_macc_epi32(vi21x4567, vk21x4567, vacc4567);
503 vacc89AB = _mm_macc_epi32(vi21x89AB, vk21x89AB, vacc89AB);
504 vaccCDEF = _mm_macc_epi32(vi21xCDEF, vk21xCDEF, vaccCDEF);
505
506 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
507 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)))));
508 const __m128i vi22x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 4)));
509 const __m128i vk22x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 356 * sizeof(int8_t)))));
510 const __m128i vi22x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 8)));
511 const __m128i vk22x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)))));
512 const __m128i vi22xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 12)));
513 const __m128i vk22xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 364 * sizeof(int8_t)))));
514 i22 += 16;
515
516 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
517 vacc4567 = _mm_macc_epi32(vi22x4567, vk22x4567, vacc4567);
518 vacc89AB = _mm_macc_epi32(vi22x89AB, vk22x89AB, vacc89AB);
519 vaccCDEF = _mm_macc_epi32(vi22xCDEF, vk22xCDEF, vaccCDEF);
520
521 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
522 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)))));
523 const __m128i vi23x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 4)));
524 const __m128i vk23x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 372 * sizeof(int8_t)))));
525 const __m128i vi23x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 8)));
526 const __m128i vk23x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)))));
527 const __m128i vi23xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 12)));
528 const __m128i vk23xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 380 * sizeof(int8_t)))));
529 i23 += 16;
530
531 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
532 vacc4567 = _mm_macc_epi32(vi23x4567, vk23x4567, vacc4567);
533 vacc89AB = _mm_macc_epi32(vi23x89AB, vk23x89AB, vacc89AB);
534 vaccCDEF = _mm_macc_epi32(vi23xCDEF, vk23xCDEF, vaccCDEF);
535
536 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
537 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)))));
538 const __m128i vi24x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 4)));
539 const __m128i vk24x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 388 * sizeof(int8_t)))));
540 const __m128i vi24x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 8)));
541 const __m128i vk24x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)))));
542 const __m128i vi24xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 12)));
543 const __m128i vk24xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 396 * sizeof(int8_t)))));
544 i24 += 16;
545
546 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
547 vacc4567 = _mm_macc_epi32(vi24x4567, vk24x4567, vacc4567);
548 vacc89AB = _mm_macc_epi32(vi24x89AB, vk24x89AB, vacc89AB);
549 vaccCDEF = _mm_macc_epi32(vi24xCDEF, vk24xCDEF, vaccCDEF);
550
551 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
552
553 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
554 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
555 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
556 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
557
558 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
559 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
560 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
561 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
562 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
563
564 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
565 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
566 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
567 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
568 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
569
570 vacc0123 = _mm_cvtps_epi32(vscaled0123);
571 vacc4567 = _mm_cvtps_epi32(vscaled4567);
572 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
573 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
574
575 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
576 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
577 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
578
579 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
580 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
581 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
582
583 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
584 output += 16;
585 }
586 if XNN_UNLIKELY(c != 0) {
587 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
588 do {
589 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
590
591 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
592 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) k)));
593 i0 += 4;
594
595 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
596 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
597 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16))));
598 i1 += 4;
599
600 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
601 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
602 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32))));
603 i2 += 4;
604
605 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
606 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
607 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48))));
608 i3 += 4;
609
610 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
611 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
612 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64))));
613 i4 += 4;
614
615 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
616 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
617 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 80))));
618 i5 += 4;
619
620 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
621 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
622 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96))));
623 i6 += 4;
624
625 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
626 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
627 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 112))));
628 i7 += 4;
629
630 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
631 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
632 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 128))));
633 i8 += 4;
634
635 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
636 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
637 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 144))));
638 i9 += 4;
639
640 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
641 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
642 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 160))));
643 i10 += 4;
644
645 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
646 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
647 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 176))));
648 i11 += 4;
649
650 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
651 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
652 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 192))));
653 i12 += 4;
654
655 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
656 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
657 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 208))));
658 i13 += 4;
659
660 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
661 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
662 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 224))));
663 i14 += 4;
664
665 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
666 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
667 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 240))));
668 i15 += 4;
669
670 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
671 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
672 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 256))));
673 i16 += 4;
674
675 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
676 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
677 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 272))));
678 i17 += 4;
679
680 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
681 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
682 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 288))));
683 i18 += 4;
684
685 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
686 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
687 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 304))));
688 i19 += 4;
689
690 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
691 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
692 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 320))));
693 i20 += 4;
694
695 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
696 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
697 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 336))));
698 i21 += 4;
699
700 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
701 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
702 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 352))));
703 i22 += 4;
704
705 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
706 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
707 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 368))));
708 i23 += 4;
709
710 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
711 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
712 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 384))));
713 i24 += 4;
714
715 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
716
717 k += 4;
718
719 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
720 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
721 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
722 vacc0123 = _mm_cvtps_epi32(vscaled0123);
723
724 w = (const void*) ((const int32_t*) w + 4);
725
726 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
727 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
728
729 vout0123 = _mm_packs_epi16(vout0123, vout0123);
730 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
731
732 if XNN_LIKELY(c >= 4) {
733 _mm_storeu_si32(output, vout0123);
734 output += 4;
735 c -= 4;
736 } else {
737 if (c & 2) {
738 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
739 vout0123 = _mm_srli_epi32(vout0123, 16);
740 output += 2;
741 }
742 if (c & 1) {
743 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
744 output += 1;
745 }
746 c = 0;
747 }
748 } while (c != 0);
749 }
750
751 output = (int8_t*) ((uintptr_t) output + output_increment);
752 } while (--output_width != 0);
753 }
754