1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-avx2-mul16-vpunpck.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/unaligned.h>
16
17
xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpunpck(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpunpck(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 do {
34 const int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const int8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const int8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const int8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const int8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const int8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const int8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 const int8_t* i9 = input[9];
80 assert(i9 != NULL);
81 if XNN_UNPREDICTABLE(i9 != zero) {
82 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
83 }
84 const int8_t* i10 = input[10];
85 assert(i10 != NULL);
86 if XNN_UNPREDICTABLE(i10 != zero) {
87 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
88 }
89 const int8_t* i11 = input[11];
90 assert(i11 != NULL);
91 if XNN_UNPREDICTABLE(i11 != zero) {
92 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
93 }
94 const int8_t* i12 = input[12];
95 assert(i12 != NULL);
96 if XNN_UNPREDICTABLE(i12 != zero) {
97 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
98 }
99 const int8_t* i13 = input[13];
100 assert(i13 != NULL);
101 if XNN_UNPREDICTABLE(i13 != zero) {
102 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
103 }
104 const int8_t* i14 = input[14];
105 assert(i14 != NULL);
106 if XNN_UNPREDICTABLE(i14 != zero) {
107 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
108 }
109 const int8_t* i15 = input[15];
110 assert(i15 != NULL);
111 if XNN_UNPREDICTABLE(i15 != zero) {
112 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
113 }
114 const int8_t* i16 = input[16];
115 assert(i16 != NULL);
116 if XNN_UNPREDICTABLE(i16 != zero) {
117 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
118 }
119 const int8_t* i17 = input[17];
120 assert(i17 != NULL);
121 if XNN_UNPREDICTABLE(i17 != zero) {
122 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
123 }
124 const int8_t* i18 = input[18];
125 assert(i18 != NULL);
126 if XNN_UNPREDICTABLE(i18 != zero) {
127 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
128 }
129 const int8_t* i19 = input[19];
130 assert(i19 != NULL);
131 if XNN_UNPREDICTABLE(i19 != zero) {
132 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
133 }
134 const int8_t* i20 = input[20];
135 assert(i20 != NULL);
136 if XNN_UNPREDICTABLE(i20 != zero) {
137 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
138 }
139 const int8_t* i21 = input[21];
140 assert(i21 != NULL);
141 if XNN_UNPREDICTABLE(i21 != zero) {
142 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
143 }
144 const int8_t* i22 = input[22];
145 assert(i22 != NULL);
146 if XNN_UNPREDICTABLE(i22 != zero) {
147 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
148 }
149 const int8_t* i23 = input[23];
150 assert(i23 != NULL);
151 if XNN_UNPREDICTABLE(i23 != zero) {
152 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
153 }
154 const int8_t* i24 = input[24];
155 assert(i24 != NULL);
156 if XNN_UNPREDICTABLE(i24 != zero) {
157 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
158 }
159 input = (const int8_t**) ((uintptr_t) input + input_stride);
160
161 size_t c = channels;
162 const void* w = weights;
163 for (; c >= 32; c -= 32) {
164 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
165 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
166 __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 16 * sizeof(int32_t)));
167 __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 24 * sizeof(int32_t)));
168
169 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
170 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
171 __m256i vaccGHIJOPQR = _mm256_inserti128_si256(vaccGHIJKLMN, _mm256_castsi256_si128(vaccOPQRSTUV), 1);
172 __m256i vaccKLMNSTUV = _mm256_permute2x128_si256(vaccGHIJKLMN, vaccOPQRSTUV, 0x31);
173
174
175 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
176 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t))));
177 const __m256i vi0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i0 + 16)));
178 const __m256i vk0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t))));
179 i0 += 32;
180
181 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
182 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
183 const __m256i vprod0xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV);
184 const __m256i vprod0xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod0xGHIJKLMNOPQRSTUVlo, 15);
185
186 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
187 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
188 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod0xGHIJKLMNOPQRSTUVlo, vprod0xGHIJKLMNOPQRSTUVhi));
189 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod0xGHIJKLMNOPQRSTUVlo, vprod0xGHIJKLMNOPQRSTUVhi));
190
191 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
192 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
193 const __m256i vi1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i1 + 16)));
194 const __m256i vk1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
195 i1 += 32;
196
197 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
198 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
199 const __m256i vprod1xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV);
200 const __m256i vprod1xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod1xGHIJKLMNOPQRSTUVlo, 15);
201
202 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
203 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
204 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod1xGHIJKLMNOPQRSTUVlo, vprod1xGHIJKLMNOPQRSTUVhi));
205 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod1xGHIJKLMNOPQRSTUVlo, vprod1xGHIJKLMNOPQRSTUVhi));
206
207 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
208 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
209 const __m256i vi2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i2 + 16)));
210 const __m256i vk2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
211 i2 += 32;
212
213 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
214 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
215 const __m256i vprod2xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV);
216 const __m256i vprod2xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod2xGHIJKLMNOPQRSTUVlo, 15);
217
218 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
219 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
220 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod2xGHIJKLMNOPQRSTUVlo, vprod2xGHIJKLMNOPQRSTUVhi));
221 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod2xGHIJKLMNOPQRSTUVlo, vprod2xGHIJKLMNOPQRSTUVhi));
222
223 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
224 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
225 const __m256i vi3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i3 + 16)));
226 const __m256i vk3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
227 i3 += 32;
228
229 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
230 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
231 const __m256i vprod3xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV);
232 const __m256i vprod3xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod3xGHIJKLMNOPQRSTUVlo, 15);
233
234 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
235 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
236 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod3xGHIJKLMNOPQRSTUVlo, vprod3xGHIJKLMNOPQRSTUVhi));
237 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod3xGHIJKLMNOPQRSTUVlo, vprod3xGHIJKLMNOPQRSTUVhi));
238
239 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
240 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
241 const __m256i vi4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i4 + 16)));
242 const __m256i vk4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
243 i4 += 32;
244
245 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
246 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
247 const __m256i vprod4xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV);
248 const __m256i vprod4xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod4xGHIJKLMNOPQRSTUVlo, 15);
249
250 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
251 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
252 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod4xGHIJKLMNOPQRSTUVlo, vprod4xGHIJKLMNOPQRSTUVhi));
253 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod4xGHIJKLMNOPQRSTUVlo, vprod4xGHIJKLMNOPQRSTUVhi));
254
255 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
256 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
257 const __m256i vi5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i5 + 16)));
258 const __m256i vk5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
259 i5 += 32;
260
261 const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
262 const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
263 const __m256i vprod5xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV);
264 const __m256i vprod5xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod5xGHIJKLMNOPQRSTUVlo, 15);
265
266 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
267 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
268 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod5xGHIJKLMNOPQRSTUVlo, vprod5xGHIJKLMNOPQRSTUVhi));
269 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod5xGHIJKLMNOPQRSTUVlo, vprod5xGHIJKLMNOPQRSTUVhi));
270
271 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
272 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
273 const __m256i vi6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i6 + 16)));
274 const __m256i vk6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
275 i6 += 32;
276
277 const __m256i vprod6x0123456789ABCDEFlo = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
278 const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
279 const __m256i vprod6xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV);
280 const __m256i vprod6xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod6xGHIJKLMNOPQRSTUVlo, 15);
281
282 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
283 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
284 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod6xGHIJKLMNOPQRSTUVlo, vprod6xGHIJKLMNOPQRSTUVhi));
285 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod6xGHIJKLMNOPQRSTUVlo, vprod6xGHIJKLMNOPQRSTUVhi));
286
287 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
288 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
289 const __m256i vi7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i7 + 16)));
290 const __m256i vk7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
291 i7 += 32;
292
293 const __m256i vprod7x0123456789ABCDEFlo = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
294 const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15);
295 const __m256i vprod7xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV);
296 const __m256i vprod7xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod7xGHIJKLMNOPQRSTUVlo, 15);
297
298 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
299 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
300 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod7xGHIJKLMNOPQRSTUVlo, vprod7xGHIJKLMNOPQRSTUVhi));
301 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod7xGHIJKLMNOPQRSTUVlo, vprod7xGHIJKLMNOPQRSTUVhi));
302
303 const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
304 const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t))));
305 const __m256i vi8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i8 + 16)));
306 const __m256i vk8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t))));
307 i8 += 32;
308
309 const __m256i vprod8x0123456789ABCDEFlo = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
310 const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15);
311 const __m256i vprod8xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV);
312 const __m256i vprod8xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod8xGHIJKLMNOPQRSTUVlo, 15);
313
314 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
315 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
316 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod8xGHIJKLMNOPQRSTUVlo, vprod8xGHIJKLMNOPQRSTUVhi));
317 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod8xGHIJKLMNOPQRSTUVlo, vprod8xGHIJKLMNOPQRSTUVhi));
318
319 const __m256i vi9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i9));
320 const __m256i vk9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t))));
321 const __m256i vi9xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i9 + 16)));
322 const __m256i vk9xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 304 * sizeof(int8_t))));
323 i9 += 32;
324
325 const __m256i vprod9x0123456789ABCDEFlo = _mm256_mullo_epi16(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF);
326 const __m256i vprod9x0123456789ABCDEFhi = _mm256_srai_epi16(vprod9x0123456789ABCDEFlo, 15);
327 const __m256i vprod9xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi9xGHIJKLMNOPQRSTUV, vk9xGHIJKLMNOPQRSTUV);
328 const __m256i vprod9xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod9xGHIJKLMNOPQRSTUVlo, 15);
329
330 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod9x0123456789ABCDEFlo, vprod9x0123456789ABCDEFhi));
331 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod9x0123456789ABCDEFlo, vprod9x0123456789ABCDEFhi));
332 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod9xGHIJKLMNOPQRSTUVlo, vprod9xGHIJKLMNOPQRSTUVhi));
333 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod9xGHIJKLMNOPQRSTUVlo, vprod9xGHIJKLMNOPQRSTUVhi));
334
335 const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10));
336 const __m256i vk10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 320 * sizeof(int8_t))));
337 const __m256i vi10xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i10 + 16)));
338 const __m256i vk10xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 336 * sizeof(int8_t))));
339 i10 += 32;
340
341 const __m256i vprod10x0123456789ABCDEFlo = _mm256_mullo_epi16(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF);
342 const __m256i vprod10x0123456789ABCDEFhi = _mm256_srai_epi16(vprod10x0123456789ABCDEFlo, 15);
343 const __m256i vprod10xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi10xGHIJKLMNOPQRSTUV, vk10xGHIJKLMNOPQRSTUV);
344 const __m256i vprod10xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod10xGHIJKLMNOPQRSTUVlo, 15);
345
346 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod10x0123456789ABCDEFlo, vprod10x0123456789ABCDEFhi));
347 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod10x0123456789ABCDEFlo, vprod10x0123456789ABCDEFhi));
348 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod10xGHIJKLMNOPQRSTUVlo, vprod10xGHIJKLMNOPQRSTUVhi));
349 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod10xGHIJKLMNOPQRSTUVlo, vprod10xGHIJKLMNOPQRSTUVhi));
350
351 const __m256i vi11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i11));
352 const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 352 * sizeof(int8_t))));
353 const __m256i vi11xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i11 + 16)));
354 const __m256i vk11xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 368 * sizeof(int8_t))));
355 i11 += 32;
356
357 const __m256i vprod11x0123456789ABCDEFlo = _mm256_mullo_epi16(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF);
358 const __m256i vprod11x0123456789ABCDEFhi = _mm256_srai_epi16(vprod11x0123456789ABCDEFlo, 15);
359 const __m256i vprod11xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi11xGHIJKLMNOPQRSTUV, vk11xGHIJKLMNOPQRSTUV);
360 const __m256i vprod11xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod11xGHIJKLMNOPQRSTUVlo, 15);
361
362 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod11x0123456789ABCDEFlo, vprod11x0123456789ABCDEFhi));
363 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod11x0123456789ABCDEFlo, vprod11x0123456789ABCDEFhi));
364 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod11xGHIJKLMNOPQRSTUVlo, vprod11xGHIJKLMNOPQRSTUVhi));
365 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod11xGHIJKLMNOPQRSTUVlo, vprod11xGHIJKLMNOPQRSTUVhi));
366
367 const __m256i vi12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i12));
368 const __m256i vk12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 384 * sizeof(int8_t))));
369 const __m256i vi12xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i12 + 16)));
370 const __m256i vk12xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 400 * sizeof(int8_t))));
371 i12 += 32;
372
373 const __m256i vprod12x0123456789ABCDEFlo = _mm256_mullo_epi16(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF);
374 const __m256i vprod12x0123456789ABCDEFhi = _mm256_srai_epi16(vprod12x0123456789ABCDEFlo, 15);
375 const __m256i vprod12xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi12xGHIJKLMNOPQRSTUV, vk12xGHIJKLMNOPQRSTUV);
376 const __m256i vprod12xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod12xGHIJKLMNOPQRSTUVlo, 15);
377
378 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod12x0123456789ABCDEFlo, vprod12x0123456789ABCDEFhi));
379 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod12x0123456789ABCDEFlo, vprod12x0123456789ABCDEFhi));
380 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod12xGHIJKLMNOPQRSTUVlo, vprod12xGHIJKLMNOPQRSTUVhi));
381 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod12xGHIJKLMNOPQRSTUVlo, vprod12xGHIJKLMNOPQRSTUVhi));
382
383 const __m256i vi13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i13));
384 const __m256i vk13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 416 * sizeof(int8_t))));
385 const __m256i vi13xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i13 + 16)));
386 const __m256i vk13xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 432 * sizeof(int8_t))));
387 i13 += 32;
388
389 const __m256i vprod13x0123456789ABCDEFlo = _mm256_mullo_epi16(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF);
390 const __m256i vprod13x0123456789ABCDEFhi = _mm256_srai_epi16(vprod13x0123456789ABCDEFlo, 15);
391 const __m256i vprod13xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi13xGHIJKLMNOPQRSTUV, vk13xGHIJKLMNOPQRSTUV);
392 const __m256i vprod13xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod13xGHIJKLMNOPQRSTUVlo, 15);
393
394 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod13x0123456789ABCDEFlo, vprod13x0123456789ABCDEFhi));
395 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod13x0123456789ABCDEFlo, vprod13x0123456789ABCDEFhi));
396 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod13xGHIJKLMNOPQRSTUVlo, vprod13xGHIJKLMNOPQRSTUVhi));
397 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod13xGHIJKLMNOPQRSTUVlo, vprod13xGHIJKLMNOPQRSTUVhi));
398
399 const __m256i vi14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i14));
400 const __m256i vk14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 448 * sizeof(int8_t))));
401 const __m256i vi14xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i14 + 16)));
402 const __m256i vk14xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 464 * sizeof(int8_t))));
403 i14 += 32;
404
405 const __m256i vprod14x0123456789ABCDEFlo = _mm256_mullo_epi16(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF);
406 const __m256i vprod14x0123456789ABCDEFhi = _mm256_srai_epi16(vprod14x0123456789ABCDEFlo, 15);
407 const __m256i vprod14xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi14xGHIJKLMNOPQRSTUV, vk14xGHIJKLMNOPQRSTUV);
408 const __m256i vprod14xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod14xGHIJKLMNOPQRSTUVlo, 15);
409
410 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod14x0123456789ABCDEFlo, vprod14x0123456789ABCDEFhi));
411 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod14x0123456789ABCDEFlo, vprod14x0123456789ABCDEFhi));
412 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod14xGHIJKLMNOPQRSTUVlo, vprod14xGHIJKLMNOPQRSTUVhi));
413 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod14xGHIJKLMNOPQRSTUVlo, vprod14xGHIJKLMNOPQRSTUVhi));
414
415 const __m256i vi15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i15));
416 const __m256i vk15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 480 * sizeof(int8_t))));
417 const __m256i vi15xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i15 + 16)));
418 const __m256i vk15xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 496 * sizeof(int8_t))));
419 i15 += 32;
420
421 const __m256i vprod15x0123456789ABCDEFlo = _mm256_mullo_epi16(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF);
422 const __m256i vprod15x0123456789ABCDEFhi = _mm256_srai_epi16(vprod15x0123456789ABCDEFlo, 15);
423 const __m256i vprod15xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi15xGHIJKLMNOPQRSTUV, vk15xGHIJKLMNOPQRSTUV);
424 const __m256i vprod15xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod15xGHIJKLMNOPQRSTUVlo, 15);
425
426 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod15x0123456789ABCDEFlo, vprod15x0123456789ABCDEFhi));
427 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod15x0123456789ABCDEFlo, vprod15x0123456789ABCDEFhi));
428 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod15xGHIJKLMNOPQRSTUVlo, vprod15xGHIJKLMNOPQRSTUVhi));
429 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod15xGHIJKLMNOPQRSTUVlo, vprod15xGHIJKLMNOPQRSTUVhi));
430
431 const __m256i vi16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i16));
432 const __m256i vk16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 512 * sizeof(int8_t))));
433 const __m256i vi16xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i16 + 16)));
434 const __m256i vk16xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 528 * sizeof(int8_t))));
435 i16 += 32;
436
437 const __m256i vprod16x0123456789ABCDEFlo = _mm256_mullo_epi16(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF);
438 const __m256i vprod16x0123456789ABCDEFhi = _mm256_srai_epi16(vprod16x0123456789ABCDEFlo, 15);
439 const __m256i vprod16xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi16xGHIJKLMNOPQRSTUV, vk16xGHIJKLMNOPQRSTUV);
440 const __m256i vprod16xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod16xGHIJKLMNOPQRSTUVlo, 15);
441
442 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod16x0123456789ABCDEFlo, vprod16x0123456789ABCDEFhi));
443 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod16x0123456789ABCDEFlo, vprod16x0123456789ABCDEFhi));
444 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod16xGHIJKLMNOPQRSTUVlo, vprod16xGHIJKLMNOPQRSTUVhi));
445 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod16xGHIJKLMNOPQRSTUVlo, vprod16xGHIJKLMNOPQRSTUVhi));
446
447 const __m256i vi17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i17));
448 const __m256i vk17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 544 * sizeof(int8_t))));
449 const __m256i vi17xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i17 + 16)));
450 const __m256i vk17xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 560 * sizeof(int8_t))));
451 i17 += 32;
452
453 const __m256i vprod17x0123456789ABCDEFlo = _mm256_mullo_epi16(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF);
454 const __m256i vprod17x0123456789ABCDEFhi = _mm256_srai_epi16(vprod17x0123456789ABCDEFlo, 15);
455 const __m256i vprod17xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi17xGHIJKLMNOPQRSTUV, vk17xGHIJKLMNOPQRSTUV);
456 const __m256i vprod17xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod17xGHIJKLMNOPQRSTUVlo, 15);
457
458 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod17x0123456789ABCDEFlo, vprod17x0123456789ABCDEFhi));
459 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod17x0123456789ABCDEFlo, vprod17x0123456789ABCDEFhi));
460 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod17xGHIJKLMNOPQRSTUVlo, vprod17xGHIJKLMNOPQRSTUVhi));
461 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod17xGHIJKLMNOPQRSTUVlo, vprod17xGHIJKLMNOPQRSTUVhi));
462
463 const __m256i vi18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i18));
464 const __m256i vk18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 576 * sizeof(int8_t))));
465 const __m256i vi18xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i18 + 16)));
466 const __m256i vk18xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 592 * sizeof(int8_t))));
467 i18 += 32;
468
469 const __m256i vprod18x0123456789ABCDEFlo = _mm256_mullo_epi16(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF);
470 const __m256i vprod18x0123456789ABCDEFhi = _mm256_srai_epi16(vprod18x0123456789ABCDEFlo, 15);
471 const __m256i vprod18xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi18xGHIJKLMNOPQRSTUV, vk18xGHIJKLMNOPQRSTUV);
472 const __m256i vprod18xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod18xGHIJKLMNOPQRSTUVlo, 15);
473
474 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod18x0123456789ABCDEFlo, vprod18x0123456789ABCDEFhi));
475 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod18x0123456789ABCDEFlo, vprod18x0123456789ABCDEFhi));
476 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod18xGHIJKLMNOPQRSTUVlo, vprod18xGHIJKLMNOPQRSTUVhi));
477 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod18xGHIJKLMNOPQRSTUVlo, vprod18xGHIJKLMNOPQRSTUVhi));
478
479 const __m256i vi19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i19));
480 const __m256i vk19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 608 * sizeof(int8_t))));
481 const __m256i vi19xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i19 + 16)));
482 const __m256i vk19xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 624 * sizeof(int8_t))));
483 i19 += 32;
484
485 const __m256i vprod19x0123456789ABCDEFlo = _mm256_mullo_epi16(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF);
486 const __m256i vprod19x0123456789ABCDEFhi = _mm256_srai_epi16(vprod19x0123456789ABCDEFlo, 15);
487 const __m256i vprod19xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi19xGHIJKLMNOPQRSTUV, vk19xGHIJKLMNOPQRSTUV);
488 const __m256i vprod19xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod19xGHIJKLMNOPQRSTUVlo, 15);
489
490 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod19x0123456789ABCDEFlo, vprod19x0123456789ABCDEFhi));
491 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod19x0123456789ABCDEFlo, vprod19x0123456789ABCDEFhi));
492 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod19xGHIJKLMNOPQRSTUVlo, vprod19xGHIJKLMNOPQRSTUVhi));
493 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod19xGHIJKLMNOPQRSTUVlo, vprod19xGHIJKLMNOPQRSTUVhi));
494
495 const __m256i vi20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i20));
496 const __m256i vk20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 640 * sizeof(int8_t))));
497 const __m256i vi20xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i20 + 16)));
498 const __m256i vk20xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 656 * sizeof(int8_t))));
499 i20 += 32;
500
501 const __m256i vprod20x0123456789ABCDEFlo = _mm256_mullo_epi16(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF);
502 const __m256i vprod20x0123456789ABCDEFhi = _mm256_srai_epi16(vprod20x0123456789ABCDEFlo, 15);
503 const __m256i vprod20xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi20xGHIJKLMNOPQRSTUV, vk20xGHIJKLMNOPQRSTUV);
504 const __m256i vprod20xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod20xGHIJKLMNOPQRSTUVlo, 15);
505
506 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod20x0123456789ABCDEFlo, vprod20x0123456789ABCDEFhi));
507 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod20x0123456789ABCDEFlo, vprod20x0123456789ABCDEFhi));
508 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod20xGHIJKLMNOPQRSTUVlo, vprod20xGHIJKLMNOPQRSTUVhi));
509 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod20xGHIJKLMNOPQRSTUVlo, vprod20xGHIJKLMNOPQRSTUVhi));
510
511 const __m256i vi21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i21));
512 const __m256i vk21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 672 * sizeof(int8_t))));
513 const __m256i vi21xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i21 + 16)));
514 const __m256i vk21xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 688 * sizeof(int8_t))));
515 i21 += 32;
516
517 const __m256i vprod21x0123456789ABCDEFlo = _mm256_mullo_epi16(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF);
518 const __m256i vprod21x0123456789ABCDEFhi = _mm256_srai_epi16(vprod21x0123456789ABCDEFlo, 15);
519 const __m256i vprod21xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi21xGHIJKLMNOPQRSTUV, vk21xGHIJKLMNOPQRSTUV);
520 const __m256i vprod21xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod21xGHIJKLMNOPQRSTUVlo, 15);
521
522 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod21x0123456789ABCDEFlo, vprod21x0123456789ABCDEFhi));
523 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod21x0123456789ABCDEFlo, vprod21x0123456789ABCDEFhi));
524 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod21xGHIJKLMNOPQRSTUVlo, vprod21xGHIJKLMNOPQRSTUVhi));
525 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod21xGHIJKLMNOPQRSTUVlo, vprod21xGHIJKLMNOPQRSTUVhi));
526
527 const __m256i vi22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i22));
528 const __m256i vk22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 704 * sizeof(int8_t))));
529 const __m256i vi22xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i22 + 16)));
530 const __m256i vk22xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 720 * sizeof(int8_t))));
531 i22 += 32;
532
533 const __m256i vprod22x0123456789ABCDEFlo = _mm256_mullo_epi16(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF);
534 const __m256i vprod22x0123456789ABCDEFhi = _mm256_srai_epi16(vprod22x0123456789ABCDEFlo, 15);
535 const __m256i vprod22xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi22xGHIJKLMNOPQRSTUV, vk22xGHIJKLMNOPQRSTUV);
536 const __m256i vprod22xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod22xGHIJKLMNOPQRSTUVlo, 15);
537
538 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod22x0123456789ABCDEFlo, vprod22x0123456789ABCDEFhi));
539 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod22x0123456789ABCDEFlo, vprod22x0123456789ABCDEFhi));
540 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod22xGHIJKLMNOPQRSTUVlo, vprod22xGHIJKLMNOPQRSTUVhi));
541 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod22xGHIJKLMNOPQRSTUVlo, vprod22xGHIJKLMNOPQRSTUVhi));
542
543 const __m256i vi23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i23));
544 const __m256i vk23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 736 * sizeof(int8_t))));
545 const __m256i vi23xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i23 + 16)));
546 const __m256i vk23xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 752 * sizeof(int8_t))));
547 i23 += 32;
548
549 const __m256i vprod23x0123456789ABCDEFlo = _mm256_mullo_epi16(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF);
550 const __m256i vprod23x0123456789ABCDEFhi = _mm256_srai_epi16(vprod23x0123456789ABCDEFlo, 15);
551 const __m256i vprod23xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi23xGHIJKLMNOPQRSTUV, vk23xGHIJKLMNOPQRSTUV);
552 const __m256i vprod23xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod23xGHIJKLMNOPQRSTUVlo, 15);
553
554 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod23x0123456789ABCDEFlo, vprod23x0123456789ABCDEFhi));
555 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod23x0123456789ABCDEFlo, vprod23x0123456789ABCDEFhi));
556 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod23xGHIJKLMNOPQRSTUVlo, vprod23xGHIJKLMNOPQRSTUVhi));
557 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod23xGHIJKLMNOPQRSTUVlo, vprod23xGHIJKLMNOPQRSTUVhi));
558
559 const __m256i vi24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i24));
560 const __m256i vk24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 768 * sizeof(int8_t))));
561 const __m256i vi24xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i24 + 16)));
562 const __m256i vk24xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 784 * sizeof(int8_t))));
563 i24 += 32;
564
565 const __m256i vprod24x0123456789ABCDEFlo = _mm256_mullo_epi16(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF);
566 const __m256i vprod24x0123456789ABCDEFhi = _mm256_srai_epi16(vprod24x0123456789ABCDEFlo, 15);
567 const __m256i vprod24xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi24xGHIJKLMNOPQRSTUV, vk24xGHIJKLMNOPQRSTUV);
568 const __m256i vprod24xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod24xGHIJKLMNOPQRSTUVlo, 15);
569
570 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod24x0123456789ABCDEFlo, vprod24x0123456789ABCDEFhi));
571 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod24x0123456789ABCDEFlo, vprod24x0123456789ABCDEFhi));
572 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod24xGHIJKLMNOPQRSTUVlo, vprod24xGHIJKLMNOPQRSTUVhi));
573 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod24xGHIJKLMNOPQRSTUVlo, vprod24xGHIJKLMNOPQRSTUVhi));
574
575 w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t));
576
577 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
578 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
579 vaccGHIJKLMN = _mm256_inserti128_si256(vaccGHIJOPQR, _mm256_castsi256_si128(vaccKLMNSTUV), 1);
580 vaccOPQRSTUV = _mm256_permute2x128_si256(vaccGHIJOPQR, vaccKLMNSTUV, 0x31);
581
582 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
583 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
584 __m256 vfpaccGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN);
585 __m256 vfpaccOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV);
586
587 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w);
588 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float)));
589 const __m256 vscaleGHIJKLMN = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(float)));
590 const __m256 vscaleOPQRSTUV = _mm256_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(float)));
591 w = (const void*) ((uintptr_t) w + 32 * sizeof(float));
592 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
593 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
594 vfpaccGHIJKLMN = _mm256_mul_ps(vfpaccGHIJKLMN, vscaleGHIJKLMN);
595 vfpaccOPQRSTUV = _mm256_mul_ps(vfpaccOPQRSTUV, vscaleOPQRSTUV);
596
597 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
598 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
599 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
600 vfpaccGHIJKLMN = _mm256_min_ps(vfpaccGHIJKLMN, voutput_max_less_zero_point);
601 vfpaccOPQRSTUV = _mm256_min_ps(vfpaccOPQRSTUV, voutput_max_less_zero_point);
602
603 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
604 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
605 vaccGHIJKLMN = _mm256_cvtps_epi32(vfpaccGHIJKLMN);
606 vaccOPQRSTUV = _mm256_cvtps_epi32(vfpaccOPQRSTUV);
607
608 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
609 const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
610 const __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
611
612 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
613 __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
614
615 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
616 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
617 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
618
619 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
620 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
621 output += 32;
622 }
623 if XNN_UNLIKELY(c != 0) {
624 const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t));
625 do {
626 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
627 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
628
629 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
630 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
631
632
633 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
634 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) k));
635 i0 += 16;
636
637 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
638 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
639
640 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
641 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
642
643 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
644 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 32)));
645 i1 += 16;
646
647 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
648 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
649
650 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
651 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
652
653 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
654 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 64)));
655 i2 += 16;
656
657 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
658 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
659
660 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
661 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
662
663 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
664 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 96)));
665 i3 += 16;
666
667 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
668 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
669
670 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
671 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
672
673 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
674 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 128)));
675 i4 += 16;
676
677 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
678 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
679
680 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
681 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
682
683 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
684 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 160)));
685 i5 += 16;
686
687 const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
688 const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
689
690 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
691 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
692
693 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
694 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 192)));
695 i6 += 16;
696
697 const __m256i vprod6x0123456789ABCDEFlo = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
698 const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
699
700 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
701 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
702
703 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
704 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 224)));
705 i7 += 16;
706
707 const __m256i vprod7x0123456789ABCDEFlo = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
708 const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15);
709
710 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
711 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
712
713 const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
714 const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 256)));
715 i8 += 16;
716
717 const __m256i vprod8x0123456789ABCDEFlo = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
718 const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15);
719
720 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
721 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
722
723 const __m256i vi9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i9));
724 const __m256i vk9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 288)));
725 i9 += 16;
726
727 const __m256i vprod9x0123456789ABCDEFlo = _mm256_mullo_epi16(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF);
728 const __m256i vprod9x0123456789ABCDEFhi = _mm256_srai_epi16(vprod9x0123456789ABCDEFlo, 15);
729
730 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod9x0123456789ABCDEFlo, vprod9x0123456789ABCDEFhi));
731 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod9x0123456789ABCDEFlo, vprod9x0123456789ABCDEFhi));
732
733 const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10));
734 const __m256i vk10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 320)));
735 i10 += 16;
736
737 const __m256i vprod10x0123456789ABCDEFlo = _mm256_mullo_epi16(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF);
738 const __m256i vprod10x0123456789ABCDEFhi = _mm256_srai_epi16(vprod10x0123456789ABCDEFlo, 15);
739
740 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod10x0123456789ABCDEFlo, vprod10x0123456789ABCDEFhi));
741 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod10x0123456789ABCDEFlo, vprod10x0123456789ABCDEFhi));
742
743 const __m256i vi11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i11));
744 const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 352)));
745 i11 += 16;
746
747 const __m256i vprod11x0123456789ABCDEFlo = _mm256_mullo_epi16(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF);
748 const __m256i vprod11x0123456789ABCDEFhi = _mm256_srai_epi16(vprod11x0123456789ABCDEFlo, 15);
749
750 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod11x0123456789ABCDEFlo, vprod11x0123456789ABCDEFhi));
751 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod11x0123456789ABCDEFlo, vprod11x0123456789ABCDEFhi));
752
753 const __m256i vi12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i12));
754 const __m256i vk12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 384)));
755 i12 += 16;
756
757 const __m256i vprod12x0123456789ABCDEFlo = _mm256_mullo_epi16(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF);
758 const __m256i vprod12x0123456789ABCDEFhi = _mm256_srai_epi16(vprod12x0123456789ABCDEFlo, 15);
759
760 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod12x0123456789ABCDEFlo, vprod12x0123456789ABCDEFhi));
761 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod12x0123456789ABCDEFlo, vprod12x0123456789ABCDEFhi));
762
763 const __m256i vi13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i13));
764 const __m256i vk13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 416)));
765 i13 += 16;
766
767 const __m256i vprod13x0123456789ABCDEFlo = _mm256_mullo_epi16(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF);
768 const __m256i vprod13x0123456789ABCDEFhi = _mm256_srai_epi16(vprod13x0123456789ABCDEFlo, 15);
769
770 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod13x0123456789ABCDEFlo, vprod13x0123456789ABCDEFhi));
771 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod13x0123456789ABCDEFlo, vprod13x0123456789ABCDEFhi));
772
773 const __m256i vi14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i14));
774 const __m256i vk14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 448)));
775 i14 += 16;
776
777 const __m256i vprod14x0123456789ABCDEFlo = _mm256_mullo_epi16(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF);
778 const __m256i vprod14x0123456789ABCDEFhi = _mm256_srai_epi16(vprod14x0123456789ABCDEFlo, 15);
779
780 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod14x0123456789ABCDEFlo, vprod14x0123456789ABCDEFhi));
781 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod14x0123456789ABCDEFlo, vprod14x0123456789ABCDEFhi));
782
783 const __m256i vi15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i15));
784 const __m256i vk15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 480)));
785 i15 += 16;
786
787 const __m256i vprod15x0123456789ABCDEFlo = _mm256_mullo_epi16(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF);
788 const __m256i vprod15x0123456789ABCDEFhi = _mm256_srai_epi16(vprod15x0123456789ABCDEFlo, 15);
789
790 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod15x0123456789ABCDEFlo, vprod15x0123456789ABCDEFhi));
791 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod15x0123456789ABCDEFlo, vprod15x0123456789ABCDEFhi));
792
793 const __m256i vi16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i16));
794 const __m256i vk16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 512)));
795 i16 += 16;
796
797 const __m256i vprod16x0123456789ABCDEFlo = _mm256_mullo_epi16(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF);
798 const __m256i vprod16x0123456789ABCDEFhi = _mm256_srai_epi16(vprod16x0123456789ABCDEFlo, 15);
799
800 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod16x0123456789ABCDEFlo, vprod16x0123456789ABCDEFhi));
801 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod16x0123456789ABCDEFlo, vprod16x0123456789ABCDEFhi));
802
803 const __m256i vi17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i17));
804 const __m256i vk17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 544)));
805 i17 += 16;
806
807 const __m256i vprod17x0123456789ABCDEFlo = _mm256_mullo_epi16(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF);
808 const __m256i vprod17x0123456789ABCDEFhi = _mm256_srai_epi16(vprod17x0123456789ABCDEFlo, 15);
809
810 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod17x0123456789ABCDEFlo, vprod17x0123456789ABCDEFhi));
811 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod17x0123456789ABCDEFlo, vprod17x0123456789ABCDEFhi));
812
813 const __m256i vi18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i18));
814 const __m256i vk18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 576)));
815 i18 += 16;
816
817 const __m256i vprod18x0123456789ABCDEFlo = _mm256_mullo_epi16(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF);
818 const __m256i vprod18x0123456789ABCDEFhi = _mm256_srai_epi16(vprod18x0123456789ABCDEFlo, 15);
819
820 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod18x0123456789ABCDEFlo, vprod18x0123456789ABCDEFhi));
821 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod18x0123456789ABCDEFlo, vprod18x0123456789ABCDEFhi));
822
823 const __m256i vi19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i19));
824 const __m256i vk19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 608)));
825 i19 += 16;
826
827 const __m256i vprod19x0123456789ABCDEFlo = _mm256_mullo_epi16(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF);
828 const __m256i vprod19x0123456789ABCDEFhi = _mm256_srai_epi16(vprod19x0123456789ABCDEFlo, 15);
829
830 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod19x0123456789ABCDEFlo, vprod19x0123456789ABCDEFhi));
831 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod19x0123456789ABCDEFlo, vprod19x0123456789ABCDEFhi));
832
833 const __m256i vi20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i20));
834 const __m256i vk20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 640)));
835 i20 += 16;
836
837 const __m256i vprod20x0123456789ABCDEFlo = _mm256_mullo_epi16(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF);
838 const __m256i vprod20x0123456789ABCDEFhi = _mm256_srai_epi16(vprod20x0123456789ABCDEFlo, 15);
839
840 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod20x0123456789ABCDEFlo, vprod20x0123456789ABCDEFhi));
841 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod20x0123456789ABCDEFlo, vprod20x0123456789ABCDEFhi));
842
843 const __m256i vi21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i21));
844 const __m256i vk21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 672)));
845 i21 += 16;
846
847 const __m256i vprod21x0123456789ABCDEFlo = _mm256_mullo_epi16(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF);
848 const __m256i vprod21x0123456789ABCDEFhi = _mm256_srai_epi16(vprod21x0123456789ABCDEFlo, 15);
849
850 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod21x0123456789ABCDEFlo, vprod21x0123456789ABCDEFhi));
851 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod21x0123456789ABCDEFlo, vprod21x0123456789ABCDEFhi));
852
853 const __m256i vi22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i22));
854 const __m256i vk22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 704)));
855 i22 += 16;
856
857 const __m256i vprod22x0123456789ABCDEFlo = _mm256_mullo_epi16(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF);
858 const __m256i vprod22x0123456789ABCDEFhi = _mm256_srai_epi16(vprod22x0123456789ABCDEFlo, 15);
859
860 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod22x0123456789ABCDEFlo, vprod22x0123456789ABCDEFhi));
861 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod22x0123456789ABCDEFlo, vprod22x0123456789ABCDEFhi));
862
863 const __m256i vi23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i23));
864 const __m256i vk23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 736)));
865 i23 += 16;
866
867 const __m256i vprod23x0123456789ABCDEFlo = _mm256_mullo_epi16(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF);
868 const __m256i vprod23x0123456789ABCDEFhi = _mm256_srai_epi16(vprod23x0123456789ABCDEFlo, 15);
869
870 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod23x0123456789ABCDEFlo, vprod23x0123456789ABCDEFhi));
871 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod23x0123456789ABCDEFlo, vprod23x0123456789ABCDEFhi));
872
873 const __m256i vi24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i24));
874 const __m256i vk24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 768)));
875 i24 += 16;
876
877 const __m256i vprod24x0123456789ABCDEFlo = _mm256_mullo_epi16(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF);
878 const __m256i vprod24x0123456789ABCDEFhi = _mm256_srai_epi16(vprod24x0123456789ABCDEFlo, 15);
879
880 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod24x0123456789ABCDEFlo, vprod24x0123456789ABCDEFhi));
881 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod24x0123456789ABCDEFlo, vprod24x0123456789ABCDEFhi));
882
883 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
884 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
885
886 k += 16;
887
888 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
889 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
890
891 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t)));
892 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t) + 8 * sizeof(float)));
893 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
894 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
895
896 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
897 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
898 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
899
900 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
901 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
902
903 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
904
905 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
906 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
907 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), voutput_zero_point);
908
909 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
910
911 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
912 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
913
914 if XNN_LIKELY(c >= 16) {
915 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
916 output += 16;
917 c -= 16;
918 } else {
919 if (c & 8) {
920 _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
921 vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
922 output += 8;
923 }
924 if (c & 4) {
925 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
926 vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
927 output += 4;
928 }
929 if (c & 2) {
930 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
931 vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
932 output += 2;
933 }
934 if (c & 1) {
935 *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
936 output += 1;
937 }
938 c = 0;
939 }
940 } while (c != 0);
941 }
942
943 output = (int8_t*) ((uintptr_t) output + output_increment);
944 } while (--output_width != 0);
945 }
946