1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_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)])19 void xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__sse41_mul32(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 size_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(channels != 0);
32 assert(output_width != 0);
33
34 do {
35 const int8_t* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
39 }
40 const int8_t* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
44 }
45 const int8_t* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
49 }
50 const int8_t* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
54 }
55 const int8_t* i4 = input[4];
56 assert(i4 != NULL);
57 if XNN_UNPREDICTABLE(i4 != zero) {
58 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
59 }
60 const int8_t* i5 = input[5];
61 assert(i5 != NULL);
62 if XNN_UNPREDICTABLE(i5 != zero) {
63 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
64 }
65 const int8_t* i6 = input[6];
66 assert(i6 != NULL);
67 if XNN_UNPREDICTABLE(i6 != zero) {
68 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
69 }
70 const int8_t* i7 = input[7];
71 assert(i7 != NULL);
72 if XNN_UNPREDICTABLE(i7 != zero) {
73 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
74 }
75 const int8_t* i8 = input[8];
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i8 != zero) {
78 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
79 }
80 const int8_t* i9 = input[9];
81 assert(i9 != NULL);
82 if XNN_UNPREDICTABLE(i9 != zero) {
83 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
84 }
85 const int8_t* i10 = input[10];
86 assert(i10 != NULL);
87 if XNN_UNPREDICTABLE(i10 != zero) {
88 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
89 }
90 const int8_t* i11 = input[11];
91 assert(i11 != NULL);
92 if XNN_UNPREDICTABLE(i11 != zero) {
93 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
94 }
95 const int8_t* i12 = input[12];
96 assert(i12 != NULL);
97 if XNN_UNPREDICTABLE(i12 != zero) {
98 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
99 }
100 const int8_t* i13 = input[13];
101 assert(i13 != NULL);
102 if XNN_UNPREDICTABLE(i13 != zero) {
103 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
104 }
105 const int8_t* i14 = input[14];
106 assert(i14 != NULL);
107 if XNN_UNPREDICTABLE(i14 != zero) {
108 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
109 }
110 const int8_t* i15 = input[15];
111 assert(i15 != NULL);
112 if XNN_UNPREDICTABLE(i15 != zero) {
113 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
114 }
115 const int8_t* i16 = input[16];
116 assert(i16 != NULL);
117 if XNN_UNPREDICTABLE(i16 != zero) {
118 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
119 }
120 const int8_t* i17 = input[17];
121 assert(i17 != NULL);
122 if XNN_UNPREDICTABLE(i17 != zero) {
123 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
124 }
125 const int8_t* i18 = input[18];
126 assert(i18 != NULL);
127 if XNN_UNPREDICTABLE(i18 != zero) {
128 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
129 }
130 const int8_t* i19 = input[19];
131 assert(i19 != NULL);
132 if XNN_UNPREDICTABLE(i19 != zero) {
133 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
134 }
135 const int8_t* i20 = input[20];
136 assert(i20 != NULL);
137 if XNN_UNPREDICTABLE(i20 != zero) {
138 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
139 }
140 const int8_t* i21 = input[21];
141 assert(i21 != NULL);
142 if XNN_UNPREDICTABLE(i21 != zero) {
143 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
144 }
145 const int8_t* i22 = input[22];
146 assert(i22 != NULL);
147 if XNN_UNPREDICTABLE(i22 != zero) {
148 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
149 }
150 const int8_t* i23 = input[23];
151 assert(i23 != NULL);
152 if XNN_UNPREDICTABLE(i23 != zero) {
153 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
154 }
155 const int8_t* i24 = input[24];
156 assert(i24 != NULL);
157 if XNN_UNPREDICTABLE(i24 != zero) {
158 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
159 }
160 input = (const int8_t**) ((uintptr_t) input + input_stride);
161
162 size_t c = channels;
163 const void* w = weights;
164 for (; c >= 24; c -= 24) {
165 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
166 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
167 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
168 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
169 __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
170 __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
171
172
173 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
174 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
175 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
176 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
177 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
178 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
179 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
180 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
181 const __m128i vi0xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 16)));
182 const __m128i vk0xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
183 const __m128i vi0xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 20)));
184 const __m128i vk0xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
185 i0 += 24;
186
187 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
188 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
189 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
190 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
191 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ));
192 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN));
193
194 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
195 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
196 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
197 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
198 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
199 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
200 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
201 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
202 const __m128i vi1xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 16)));
203 const __m128i vk1xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
204 const __m128i vi1xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 20)));
205 const __m128i vk1xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
206 i1 += 24;
207
208 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
209 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
210 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
211 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
212 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi1xGHIJ, vk1xGHIJ));
213 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi1xKLMN, vk1xKLMN));
214
215 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
216 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)))));
217 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
218 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 52 * sizeof(int8_t)))));
219 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
220 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)))));
221 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
222 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 60 * sizeof(int8_t)))));
223 const __m128i vi2xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 16)));
224 const __m128i vk2xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)))));
225 const __m128i vi2xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 20)));
226 const __m128i vk2xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 68 * sizeof(int8_t)))));
227 i2 += 24;
228
229 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
230 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
231 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
232 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
233 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi2xGHIJ, vk2xGHIJ));
234 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi2xKLMN, vk2xKLMN));
235
236 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
237 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)))));
238 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
239 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 76 * sizeof(int8_t)))));
240 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
241 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)))));
242 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
243 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 84 * sizeof(int8_t)))));
244 const __m128i vi3xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 16)));
245 const __m128i vk3xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)))));
246 const __m128i vi3xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 20)));
247 const __m128i vk3xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 92 * sizeof(int8_t)))));
248 i3 += 24;
249
250 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
251 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
252 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
253 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
254 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi3xGHIJ, vk3xGHIJ));
255 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi3xKLMN, vk3xKLMN));
256
257 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
258 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)))));
259 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
260 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 100 * sizeof(int8_t)))));
261 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
262 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)))));
263 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
264 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 108 * sizeof(int8_t)))));
265 const __m128i vi4xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 16)));
266 const __m128i vk4xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)))));
267 const __m128i vi4xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 20)));
268 const __m128i vk4xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 116 * sizeof(int8_t)))));
269 i4 += 24;
270
271 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
272 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
273 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
274 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
275 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi4xGHIJ, vk4xGHIJ));
276 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi4xKLMN, vk4xKLMN));
277
278 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
279 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)))));
280 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
281 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 124 * sizeof(int8_t)))));
282 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
283 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)))));
284 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
285 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 132 * sizeof(int8_t)))));
286 const __m128i vi5xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 16)));
287 const __m128i vk5xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)))));
288 const __m128i vi5xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 20)));
289 const __m128i vk5xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 140 * sizeof(int8_t)))));
290 i5 += 24;
291
292 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
293 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
294 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
295 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
296 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi5xGHIJ, vk5xGHIJ));
297 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi5xKLMN, vk5xKLMN));
298
299 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
300 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)))));
301 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
302 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 148 * sizeof(int8_t)))));
303 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
304 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)))));
305 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
306 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 156 * sizeof(int8_t)))));
307 const __m128i vi6xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 16)));
308 const __m128i vk6xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)))));
309 const __m128i vi6xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 20)));
310 const __m128i vk6xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 164 * sizeof(int8_t)))));
311 i6 += 24;
312
313 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
314 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
315 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi6x89AB, vk6x89AB));
316 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi6xCDEF, vk6xCDEF));
317 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi6xGHIJ, vk6xGHIJ));
318 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi6xKLMN, vk6xKLMN));
319
320 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
321 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)))));
322 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
323 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 172 * sizeof(int8_t)))));
324 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 8)));
325 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)))));
326 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 12)));
327 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 180 * sizeof(int8_t)))));
328 const __m128i vi7xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 16)));
329 const __m128i vk7xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)))));
330 const __m128i vi7xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 20)));
331 const __m128i vk7xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 188 * sizeof(int8_t)))));
332 i7 += 24;
333
334 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
335 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
336 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi7x89AB, vk7x89AB));
337 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi7xCDEF, vk7xCDEF));
338 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi7xGHIJ, vk7xGHIJ));
339 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi7xKLMN, vk7xKLMN));
340
341 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
342 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)))));
343 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
344 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 196 * sizeof(int8_t)))));
345 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 8)));
346 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)))));
347 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 12)));
348 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 204 * sizeof(int8_t)))));
349 const __m128i vi8xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 16)));
350 const __m128i vk8xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)))));
351 const __m128i vi8xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 20)));
352 const __m128i vk8xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 212 * sizeof(int8_t)))));
353 i8 += 24;
354
355 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
356 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
357 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi8x89AB, vk8x89AB));
358 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi8xCDEF, vk8xCDEF));
359 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi8xGHIJ, vk8xGHIJ));
360 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi8xKLMN, vk8xKLMN));
361
362 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
363 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)))));
364 const __m128i vi9x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 4)));
365 const __m128i vk9x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 220 * sizeof(int8_t)))));
366 const __m128i vi9x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 8)));
367 const __m128i vk9x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 224 * sizeof(int8_t)))));
368 const __m128i vi9xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 12)));
369 const __m128i vk9xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 228 * sizeof(int8_t)))));
370 const __m128i vi9xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 16)));
371 const __m128i vk9xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 232 * sizeof(int8_t)))));
372 const __m128i vi9xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 20)));
373 const __m128i vk9xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 236 * sizeof(int8_t)))));
374 i9 += 24;
375
376 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
377 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi9x4567, vk9x4567));
378 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi9x89AB, vk9x89AB));
379 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi9xCDEF, vk9xCDEF));
380 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi9xGHIJ, vk9xGHIJ));
381 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi9xKLMN, vk9xKLMN));
382
383 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
384 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 240 * sizeof(int8_t)))));
385 const __m128i vi10x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 4)));
386 const __m128i vk10x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 244 * sizeof(int8_t)))));
387 const __m128i vi10x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 8)));
388 const __m128i vk10x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 248 * sizeof(int8_t)))));
389 const __m128i vi10xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 12)));
390 const __m128i vk10xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 252 * sizeof(int8_t)))));
391 const __m128i vi10xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 16)));
392 const __m128i vk10xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 256 * sizeof(int8_t)))));
393 const __m128i vi10xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 20)));
394 const __m128i vk10xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 260 * sizeof(int8_t)))));
395 i10 += 24;
396
397 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
398 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi10x4567, vk10x4567));
399 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi10x89AB, vk10x89AB));
400 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi10xCDEF, vk10xCDEF));
401 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi10xGHIJ, vk10xGHIJ));
402 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi10xKLMN, vk10xKLMN));
403
404 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
405 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 264 * sizeof(int8_t)))));
406 const __m128i vi11x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 4)));
407 const __m128i vk11x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 268 * sizeof(int8_t)))));
408 const __m128i vi11x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 8)));
409 const __m128i vk11x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 272 * sizeof(int8_t)))));
410 const __m128i vi11xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 12)));
411 const __m128i vk11xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 276 * sizeof(int8_t)))));
412 const __m128i vi11xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 16)));
413 const __m128i vk11xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 280 * sizeof(int8_t)))));
414 const __m128i vi11xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 20)));
415 const __m128i vk11xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 284 * sizeof(int8_t)))));
416 i11 += 24;
417
418 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
419 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi11x4567, vk11x4567));
420 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi11x89AB, vk11x89AB));
421 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi11xCDEF, vk11xCDEF));
422 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi11xGHIJ, vk11xGHIJ));
423 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi11xKLMN, vk11xKLMN));
424
425 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
426 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 288 * sizeof(int8_t)))));
427 const __m128i vi12x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 4)));
428 const __m128i vk12x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 292 * sizeof(int8_t)))));
429 const __m128i vi12x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 8)));
430 const __m128i vk12x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 296 * sizeof(int8_t)))));
431 const __m128i vi12xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 12)));
432 const __m128i vk12xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 300 * sizeof(int8_t)))));
433 const __m128i vi12xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 16)));
434 const __m128i vk12xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 304 * sizeof(int8_t)))));
435 const __m128i vi12xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 20)));
436 const __m128i vk12xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 308 * sizeof(int8_t)))));
437 i12 += 24;
438
439 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
440 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi12x4567, vk12x4567));
441 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi12x89AB, vk12x89AB));
442 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi12xCDEF, vk12xCDEF));
443 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi12xGHIJ, vk12xGHIJ));
444 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi12xKLMN, vk12xKLMN));
445
446 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
447 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 312 * sizeof(int8_t)))));
448 const __m128i vi13x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 4)));
449 const __m128i vk13x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 316 * sizeof(int8_t)))));
450 const __m128i vi13x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 8)));
451 const __m128i vk13x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 320 * sizeof(int8_t)))));
452 const __m128i vi13xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 12)));
453 const __m128i vk13xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 324 * sizeof(int8_t)))));
454 const __m128i vi13xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 16)));
455 const __m128i vk13xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 328 * sizeof(int8_t)))));
456 const __m128i vi13xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 20)));
457 const __m128i vk13xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 332 * sizeof(int8_t)))));
458 i13 += 24;
459
460 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
461 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi13x4567, vk13x4567));
462 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi13x89AB, vk13x89AB));
463 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi13xCDEF, vk13xCDEF));
464 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi13xGHIJ, vk13xGHIJ));
465 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi13xKLMN, vk13xKLMN));
466
467 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
468 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 336 * sizeof(int8_t)))));
469 const __m128i vi14x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 4)));
470 const __m128i vk14x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 340 * sizeof(int8_t)))));
471 const __m128i vi14x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 8)));
472 const __m128i vk14x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 344 * sizeof(int8_t)))));
473 const __m128i vi14xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 12)));
474 const __m128i vk14xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 348 * sizeof(int8_t)))));
475 const __m128i vi14xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 16)));
476 const __m128i vk14xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 352 * sizeof(int8_t)))));
477 const __m128i vi14xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 20)));
478 const __m128i vk14xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 356 * sizeof(int8_t)))));
479 i14 += 24;
480
481 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
482 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi14x4567, vk14x4567));
483 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi14x89AB, vk14x89AB));
484 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi14xCDEF, vk14xCDEF));
485 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi14xGHIJ, vk14xGHIJ));
486 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi14xKLMN, vk14xKLMN));
487
488 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
489 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 360 * sizeof(int8_t)))));
490 const __m128i vi15x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 4)));
491 const __m128i vk15x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 364 * sizeof(int8_t)))));
492 const __m128i vi15x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 8)));
493 const __m128i vk15x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 368 * sizeof(int8_t)))));
494 const __m128i vi15xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 12)));
495 const __m128i vk15xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 372 * sizeof(int8_t)))));
496 const __m128i vi15xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 16)));
497 const __m128i vk15xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 376 * sizeof(int8_t)))));
498 const __m128i vi15xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 20)));
499 const __m128i vk15xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 380 * sizeof(int8_t)))));
500 i15 += 24;
501
502 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
503 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi15x4567, vk15x4567));
504 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi15x89AB, vk15x89AB));
505 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi15xCDEF, vk15xCDEF));
506 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi15xGHIJ, vk15xGHIJ));
507 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi15xKLMN, vk15xKLMN));
508
509 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
510 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 384 * sizeof(int8_t)))));
511 const __m128i vi16x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 4)));
512 const __m128i vk16x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 388 * sizeof(int8_t)))));
513 const __m128i vi16x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 8)));
514 const __m128i vk16x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 392 * sizeof(int8_t)))));
515 const __m128i vi16xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 12)));
516 const __m128i vk16xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 396 * sizeof(int8_t)))));
517 const __m128i vi16xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 16)));
518 const __m128i vk16xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 400 * sizeof(int8_t)))));
519 const __m128i vi16xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 20)));
520 const __m128i vk16xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 404 * sizeof(int8_t)))));
521 i16 += 24;
522
523 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
524 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi16x4567, vk16x4567));
525 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi16x89AB, vk16x89AB));
526 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi16xCDEF, vk16xCDEF));
527 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi16xGHIJ, vk16xGHIJ));
528 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi16xKLMN, vk16xKLMN));
529
530 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
531 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 408 * sizeof(int8_t)))));
532 const __m128i vi17x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 4)));
533 const __m128i vk17x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 412 * sizeof(int8_t)))));
534 const __m128i vi17x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 8)));
535 const __m128i vk17x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 416 * sizeof(int8_t)))));
536 const __m128i vi17xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 12)));
537 const __m128i vk17xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 420 * sizeof(int8_t)))));
538 const __m128i vi17xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 16)));
539 const __m128i vk17xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 424 * sizeof(int8_t)))));
540 const __m128i vi17xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 20)));
541 const __m128i vk17xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 428 * sizeof(int8_t)))));
542 i17 += 24;
543
544 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
545 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi17x4567, vk17x4567));
546 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi17x89AB, vk17x89AB));
547 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi17xCDEF, vk17xCDEF));
548 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi17xGHIJ, vk17xGHIJ));
549 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi17xKLMN, vk17xKLMN));
550
551 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
552 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 432 * sizeof(int8_t)))));
553 const __m128i vi18x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 4)));
554 const __m128i vk18x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 436 * sizeof(int8_t)))));
555 const __m128i vi18x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 8)));
556 const __m128i vk18x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 440 * sizeof(int8_t)))));
557 const __m128i vi18xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 12)));
558 const __m128i vk18xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 444 * sizeof(int8_t)))));
559 const __m128i vi18xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 16)));
560 const __m128i vk18xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 448 * sizeof(int8_t)))));
561 const __m128i vi18xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 20)));
562 const __m128i vk18xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 452 * sizeof(int8_t)))));
563 i18 += 24;
564
565 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
566 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi18x4567, vk18x4567));
567 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi18x89AB, vk18x89AB));
568 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi18xCDEF, vk18xCDEF));
569 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi18xGHIJ, vk18xGHIJ));
570 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi18xKLMN, vk18xKLMN));
571
572 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
573 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 456 * sizeof(int8_t)))));
574 const __m128i vi19x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 4)));
575 const __m128i vk19x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 460 * sizeof(int8_t)))));
576 const __m128i vi19x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 8)));
577 const __m128i vk19x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 464 * sizeof(int8_t)))));
578 const __m128i vi19xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 12)));
579 const __m128i vk19xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 468 * sizeof(int8_t)))));
580 const __m128i vi19xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 16)));
581 const __m128i vk19xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 472 * sizeof(int8_t)))));
582 const __m128i vi19xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 20)));
583 const __m128i vk19xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 476 * sizeof(int8_t)))));
584 i19 += 24;
585
586 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
587 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi19x4567, vk19x4567));
588 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi19x89AB, vk19x89AB));
589 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi19xCDEF, vk19xCDEF));
590 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi19xGHIJ, vk19xGHIJ));
591 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi19xKLMN, vk19xKLMN));
592
593 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
594 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 480 * sizeof(int8_t)))));
595 const __m128i vi20x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 4)));
596 const __m128i vk20x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 484 * sizeof(int8_t)))));
597 const __m128i vi20x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 8)));
598 const __m128i vk20x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 488 * sizeof(int8_t)))));
599 const __m128i vi20xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 12)));
600 const __m128i vk20xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 492 * sizeof(int8_t)))));
601 const __m128i vi20xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 16)));
602 const __m128i vk20xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 496 * sizeof(int8_t)))));
603 const __m128i vi20xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 20)));
604 const __m128i vk20xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 500 * sizeof(int8_t)))));
605 i20 += 24;
606
607 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
608 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi20x4567, vk20x4567));
609 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi20x89AB, vk20x89AB));
610 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi20xCDEF, vk20xCDEF));
611 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi20xGHIJ, vk20xGHIJ));
612 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi20xKLMN, vk20xKLMN));
613
614 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
615 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 504 * sizeof(int8_t)))));
616 const __m128i vi21x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 4)));
617 const __m128i vk21x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 508 * sizeof(int8_t)))));
618 const __m128i vi21x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 8)));
619 const __m128i vk21x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 512 * sizeof(int8_t)))));
620 const __m128i vi21xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 12)));
621 const __m128i vk21xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 516 * sizeof(int8_t)))));
622 const __m128i vi21xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 16)));
623 const __m128i vk21xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 520 * sizeof(int8_t)))));
624 const __m128i vi21xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 20)));
625 const __m128i vk21xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 524 * sizeof(int8_t)))));
626 i21 += 24;
627
628 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
629 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi21x4567, vk21x4567));
630 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi21x89AB, vk21x89AB));
631 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi21xCDEF, vk21xCDEF));
632 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi21xGHIJ, vk21xGHIJ));
633 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi21xKLMN, vk21xKLMN));
634
635 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
636 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 528 * sizeof(int8_t)))));
637 const __m128i vi22x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 4)));
638 const __m128i vk22x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 532 * sizeof(int8_t)))));
639 const __m128i vi22x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 8)));
640 const __m128i vk22x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 536 * sizeof(int8_t)))));
641 const __m128i vi22xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 12)));
642 const __m128i vk22xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 540 * sizeof(int8_t)))));
643 const __m128i vi22xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 16)));
644 const __m128i vk22xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 544 * sizeof(int8_t)))));
645 const __m128i vi22xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 20)));
646 const __m128i vk22xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 548 * sizeof(int8_t)))));
647 i22 += 24;
648
649 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
650 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi22x4567, vk22x4567));
651 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi22x89AB, vk22x89AB));
652 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi22xCDEF, vk22xCDEF));
653 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi22xGHIJ, vk22xGHIJ));
654 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi22xKLMN, vk22xKLMN));
655
656 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
657 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 552 * sizeof(int8_t)))));
658 const __m128i vi23x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 4)));
659 const __m128i vk23x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 556 * sizeof(int8_t)))));
660 const __m128i vi23x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 8)));
661 const __m128i vk23x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 560 * sizeof(int8_t)))));
662 const __m128i vi23xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 12)));
663 const __m128i vk23xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 564 * sizeof(int8_t)))));
664 const __m128i vi23xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 16)));
665 const __m128i vk23xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 568 * sizeof(int8_t)))));
666 const __m128i vi23xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 20)));
667 const __m128i vk23xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 572 * sizeof(int8_t)))));
668 i23 += 24;
669
670 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
671 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi23x4567, vk23x4567));
672 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi23x89AB, vk23x89AB));
673 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi23xCDEF, vk23xCDEF));
674 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi23xGHIJ, vk23xGHIJ));
675 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi23xKLMN, vk23xKLMN));
676
677 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
678 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 576 * sizeof(int8_t)))));
679 const __m128i vi24x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 4)));
680 const __m128i vk24x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 580 * sizeof(int8_t)))));
681 const __m128i vi24x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 8)));
682 const __m128i vk24x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 584 * sizeof(int8_t)))));
683 const __m128i vi24xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 12)));
684 const __m128i vk24xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 588 * sizeof(int8_t)))));
685 const __m128i vi24xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 16)));
686 const __m128i vk24xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 592 * sizeof(int8_t)))));
687 const __m128i vi24xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 20)));
688 const __m128i vk24xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 596 * sizeof(int8_t)))));
689 i24 += 24;
690
691 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
692 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi24x4567, vk24x4567));
693 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi24x89AB, vk24x89AB));
694 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi24xCDEF, vk24xCDEF));
695 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi24xGHIJ, vk24xGHIJ));
696 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi24xKLMN, vk24xKLMN));
697
698 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t));
699
700 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
701 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
702 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
703 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
704 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
705 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
706
707 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
708 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
709 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
710 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
711 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
712 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscale);
713 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscale);
714
715 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
716 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
717 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
718 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
719 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
720 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
721 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
722
723 vacc0123 = _mm_cvtps_epi32(vscaled0123);
724 vacc4567 = _mm_cvtps_epi32(vscaled4567);
725 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
726 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
727 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
728 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
729
730 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
731 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
732 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
733 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
734
735 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
736 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
737 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
738 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
739 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
740
741 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
742 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
743 output += 24;
744 }
745 if XNN_UNLIKELY(c != 0) {
746 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
747 do {
748 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
749
750 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
751 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) k)));
752 i0 += 4;
753
754 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
755 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
756 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 24))));
757 i1 += 4;
758
759 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
760 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
761 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48))));
762 i2 += 4;
763
764 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
765 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
766 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 72))));
767 i3 += 4;
768
769 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
770 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
771 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96))));
772 i4 += 4;
773
774 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
775 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
776 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 120))));
777 i5 += 4;
778
779 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
780 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
781 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 144))));
782 i6 += 4;
783
784 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
785 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
786 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 168))));
787 i7 += 4;
788
789 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
790 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
791 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 192))));
792 i8 += 4;
793
794 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
795 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
796 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 216))));
797 i9 += 4;
798
799 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
800 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
801 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 240))));
802 i10 += 4;
803
804 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
805 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
806 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 264))));
807 i11 += 4;
808
809 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
810 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
811 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 288))));
812 i12 += 4;
813
814 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
815 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
816 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 312))));
817 i13 += 4;
818
819 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
820 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
821 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 336))));
822 i14 += 4;
823
824 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
825 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
826 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 360))));
827 i15 += 4;
828
829 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
830 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
831 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 384))));
832 i16 += 4;
833
834 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
835 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
836 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 408))));
837 i17 += 4;
838
839 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
840 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
841 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 432))));
842 i18 += 4;
843
844 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
845 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
846 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 456))));
847 i19 += 4;
848
849 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
850 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
851 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 480))));
852 i20 += 4;
853
854 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
855 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
856 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 504))));
857 i21 += 4;
858
859 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
860 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
861 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 528))));
862 i22 += 4;
863
864 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
865 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
866 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 552))));
867 i23 += 4;
868
869 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
870 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
871 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 576))));
872 i24 += 4;
873
874 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
875
876 k += 4;
877
878 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
879 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
880 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
881 vacc0123 = _mm_cvtps_epi32(vscaled0123);
882
883 w = (const void*) ((const int32_t*) w + 4);
884
885 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
886 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
887
888 vout0123 = _mm_packs_epi16(vout0123, vout0123);
889 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
890
891 if XNN_LIKELY(c >= 4) {
892 _mm_storeu_si32(output, vout0123);
893 output += 4;
894 c -= 4;
895 } else {
896 if (c & 2) {
897 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
898 vout0123 = _mm_srli_epi32(vout0123, 16);
899 output += 2;
900 }
901 if (c & 1) {
902 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
903 output += 1;
904 }
905 c = 0;
906 }
907 } while (c != 0);
908 }
909
910 output = (int8_t*) ((uintptr_t) output + output_increment);
911 } while (--output_width != 0);
912 }
913