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_add16_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_add16_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 __m256i vacc0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
182 __m256i vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV);
183
184
185 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
186 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
187 const __m256i vi1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i1 + 16)));
188 const __m256i vk1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
189 i1 += 32;
190
191 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
192 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV));
193
194 __m256i vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
195 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
196 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
197 __m256i vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
198 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
199 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
200
201 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
202 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
203 const __m256i vi2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i2 + 16)));
204 const __m256i vk2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
205 i2 += 32;
206
207 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
208 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV);
209
210
211 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
212 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
213 const __m256i vi3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i3 + 16)));
214 const __m256i vk3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
215 i3 += 32;
216
217 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
218 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV));
219
220 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
221 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
222 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
223 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
224 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
225 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
226
227 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
228 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
229 const __m256i vi4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i4 + 16)));
230 const __m256i vk4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
231 i4 += 32;
232
233 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
234 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV);
235
236
237 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
238 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
239 const __m256i vi5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i5 + 16)));
240 const __m256i vk5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
241 i5 += 32;
242
243 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
244 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV));
245
246 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
247 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
248 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
249 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
250 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
251 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
252
253 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
254 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
255 const __m256i vi6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i6 + 16)));
256 const __m256i vk6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
257 i6 += 32;
258
259 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
260 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV);
261
262
263 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
264 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
265 const __m256i vi7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i7 + 16)));
266 const __m256i vk7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
267 i7 += 32;
268
269 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
270 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV));
271
272 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
273 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
274 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
275 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
276 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
277 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
278
279 const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
280 const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t))));
281 const __m256i vi8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i8 + 16)));
282 const __m256i vk8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t))));
283 i8 += 32;
284
285 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
286 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV);
287
288
289 const __m256i vi9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i9));
290 const __m256i vk9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t))));
291 const __m256i vi9xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i9 + 16)));
292 const __m256i vk9xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 304 * sizeof(int8_t))));
293 i9 += 32;
294
295 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF));
296 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi9xGHIJKLMNOPQRSTUV, vk9xGHIJKLMNOPQRSTUV));
297
298 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
299 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
300 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
301 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
302 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
303 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
304
305 const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10));
306 const __m256i vk10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 320 * sizeof(int8_t))));
307 const __m256i vi10xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i10 + 16)));
308 const __m256i vk10xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 336 * sizeof(int8_t))));
309 i10 += 32;
310
311 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF);
312 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi10xGHIJKLMNOPQRSTUV, vk10xGHIJKLMNOPQRSTUV);
313
314
315 const __m256i vi11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i11));
316 const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 352 * sizeof(int8_t))));
317 const __m256i vi11xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i11 + 16)));
318 const __m256i vk11xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 368 * sizeof(int8_t))));
319 i11 += 32;
320
321 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF));
322 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi11xGHIJKLMNOPQRSTUV, vk11xGHIJKLMNOPQRSTUV));
323
324 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
325 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
326 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
327 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
328 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
329 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
330
331 const __m256i vi12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i12));
332 const __m256i vk12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 384 * sizeof(int8_t))));
333 const __m256i vi12xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i12 + 16)));
334 const __m256i vk12xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 400 * sizeof(int8_t))));
335 i12 += 32;
336
337 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF);
338 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi12xGHIJKLMNOPQRSTUV, vk12xGHIJKLMNOPQRSTUV);
339
340
341 const __m256i vi13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i13));
342 const __m256i vk13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 416 * sizeof(int8_t))));
343 const __m256i vi13xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i13 + 16)));
344 const __m256i vk13xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 432 * sizeof(int8_t))));
345 i13 += 32;
346
347 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF));
348 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi13xGHIJKLMNOPQRSTUV, vk13xGHIJKLMNOPQRSTUV));
349
350 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
351 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
352 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
353 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
354 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
355 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
356
357 const __m256i vi14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i14));
358 const __m256i vk14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 448 * sizeof(int8_t))));
359 const __m256i vi14xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i14 + 16)));
360 const __m256i vk14xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 464 * sizeof(int8_t))));
361 i14 += 32;
362
363 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF);
364 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi14xGHIJKLMNOPQRSTUV, vk14xGHIJKLMNOPQRSTUV);
365
366
367 const __m256i vi15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i15));
368 const __m256i vk15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 480 * sizeof(int8_t))));
369 const __m256i vi15xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i15 + 16)));
370 const __m256i vk15xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 496 * sizeof(int8_t))));
371 i15 += 32;
372
373 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF));
374 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi15xGHIJKLMNOPQRSTUV, vk15xGHIJKLMNOPQRSTUV));
375
376 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
377 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
378 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
379 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
380 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
381 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
382
383 const __m256i vi16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i16));
384 const __m256i vk16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 512 * sizeof(int8_t))));
385 const __m256i vi16xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i16 + 16)));
386 const __m256i vk16xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 528 * sizeof(int8_t))));
387 i16 += 32;
388
389 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF);
390 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi16xGHIJKLMNOPQRSTUV, vk16xGHIJKLMNOPQRSTUV);
391
392
393 const __m256i vi17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i17));
394 const __m256i vk17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 544 * sizeof(int8_t))));
395 const __m256i vi17xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i17 + 16)));
396 const __m256i vk17xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 560 * sizeof(int8_t))));
397 i17 += 32;
398
399 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF));
400 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi17xGHIJKLMNOPQRSTUV, vk17xGHIJKLMNOPQRSTUV));
401
402 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
403 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
404 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
405 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
406 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
407 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
408
409 const __m256i vi18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i18));
410 const __m256i vk18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 576 * sizeof(int8_t))));
411 const __m256i vi18xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i18 + 16)));
412 const __m256i vk18xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 592 * sizeof(int8_t))));
413 i18 += 32;
414
415 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF);
416 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi18xGHIJKLMNOPQRSTUV, vk18xGHIJKLMNOPQRSTUV);
417
418
419 const __m256i vi19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i19));
420 const __m256i vk19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 608 * sizeof(int8_t))));
421 const __m256i vi19xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i19 + 16)));
422 const __m256i vk19xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 624 * sizeof(int8_t))));
423 i19 += 32;
424
425 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF));
426 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi19xGHIJKLMNOPQRSTUV, vk19xGHIJKLMNOPQRSTUV));
427
428 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
429 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
430 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
431 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
432 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
433 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
434
435 const __m256i vi20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i20));
436 const __m256i vk20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 640 * sizeof(int8_t))));
437 const __m256i vi20xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i20 + 16)));
438 const __m256i vk20xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 656 * sizeof(int8_t))));
439 i20 += 32;
440
441 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF);
442 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi20xGHIJKLMNOPQRSTUV, vk20xGHIJKLMNOPQRSTUV);
443
444
445 const __m256i vi21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i21));
446 const __m256i vk21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 672 * sizeof(int8_t))));
447 const __m256i vi21xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i21 + 16)));
448 const __m256i vk21xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 688 * sizeof(int8_t))));
449 i21 += 32;
450
451 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF));
452 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi21xGHIJKLMNOPQRSTUV, vk21xGHIJKLMNOPQRSTUV));
453
454 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
455 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
456 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
457 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
458 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
459 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
460
461 const __m256i vi22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i22));
462 const __m256i vk22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 704 * sizeof(int8_t))));
463 const __m256i vi22xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i22 + 16)));
464 const __m256i vk22xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 720 * sizeof(int8_t))));
465 i22 += 32;
466
467 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF);
468 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi22xGHIJKLMNOPQRSTUV, vk22xGHIJKLMNOPQRSTUV);
469
470
471 const __m256i vi23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i23));
472 const __m256i vk23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 736 * sizeof(int8_t))));
473 const __m256i vi23xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i23 + 16)));
474 const __m256i vk23xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 752 * sizeof(int8_t))));
475 i23 += 32;
476
477 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF));
478 vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi23xGHIJKLMNOPQRSTUV, vk23xGHIJKLMNOPQRSTUV));
479
480 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
481 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
482 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
483 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
484 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
485 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
486
487 const __m256i vi24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i24));
488 const __m256i vk24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 768 * sizeof(int8_t))));
489 const __m256i vi24xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i24 + 16)));
490 const __m256i vk24xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 784 * sizeof(int8_t))));
491 i24 += 32;
492
493 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF);
494 vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi24xGHIJKLMNOPQRSTUV, vk24xGHIJKLMNOPQRSTUV);
495
496 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
497 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
498 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
499 vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
500 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
501 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
502
503 w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t));
504
505 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
506 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
507 vaccGHIJKLMN = _mm256_inserti128_si256(vaccGHIJOPQR, _mm256_castsi256_si128(vaccKLMNSTUV), 1);
508 vaccOPQRSTUV = _mm256_permute2x128_si256(vaccGHIJOPQR, vaccKLMNSTUV, 0x31);
509
510 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
511 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
512 __m256 vfpaccGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN);
513 __m256 vfpaccOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV);
514
515 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w);
516 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float)));
517 const __m256 vscaleGHIJKLMN = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(float)));
518 const __m256 vscaleOPQRSTUV = _mm256_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(float)));
519 w = (const void*) ((uintptr_t) w + 32 * sizeof(float));
520 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
521 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
522 vfpaccGHIJKLMN = _mm256_mul_ps(vfpaccGHIJKLMN, vscaleGHIJKLMN);
523 vfpaccOPQRSTUV = _mm256_mul_ps(vfpaccOPQRSTUV, vscaleOPQRSTUV);
524
525 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
526 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
527 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
528 vfpaccGHIJKLMN = _mm256_min_ps(vfpaccGHIJKLMN, voutput_max_less_zero_point);
529 vfpaccOPQRSTUV = _mm256_min_ps(vfpaccOPQRSTUV, voutput_max_less_zero_point);
530
531 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
532 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
533 vaccGHIJKLMN = _mm256_cvtps_epi32(vfpaccGHIJKLMN);
534 vaccOPQRSTUV = _mm256_cvtps_epi32(vfpaccOPQRSTUV);
535
536 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
537 const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
538 const __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
539
540 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
541 __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
542
543 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
544 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
545 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
546
547 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
548 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
549 output += 32;
550 }
551 if XNN_UNLIKELY(c != 0) {
552 const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t));
553 do {
554 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
555 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
556
557 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
558 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
559
560
561 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
562 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) k));
563 i0 += 16;
564
565 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
566 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
567
568 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
569 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
570
571 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
572 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 32)));
573 i1 += 16;
574
575 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
576 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
577
578 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
579 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
580
581 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
582 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 64)));
583 i2 += 16;
584
585 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
586 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
587
588 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
589 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
590
591 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
592 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 96)));
593 i3 += 16;
594
595 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
596 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
597
598 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
599 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
600
601 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
602 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 128)));
603 i4 += 16;
604
605 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
606 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
607
608 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
609 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
610
611 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
612 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 160)));
613 i5 += 16;
614
615 const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
616 const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
617
618 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
619 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
620
621 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
622 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 192)));
623 i6 += 16;
624
625 const __m256i vprod6x0123456789ABCDEFlo = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
626 const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
627
628 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
629 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
630
631 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
632 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 224)));
633 i7 += 16;
634
635 const __m256i vprod7x0123456789ABCDEFlo = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
636 const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15);
637
638 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
639 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
640
641 const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
642 const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 256)));
643 i8 += 16;
644
645 const __m256i vprod8x0123456789ABCDEFlo = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
646 const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15);
647
648 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
649 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
650
651 const __m256i vi9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i9));
652 const __m256i vk9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 288)));
653 i9 += 16;
654
655 const __m256i vprod9x0123456789ABCDEFlo = _mm256_mullo_epi16(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF);
656 const __m256i vprod9x0123456789ABCDEFhi = _mm256_srai_epi16(vprod9x0123456789ABCDEFlo, 15);
657
658 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod9x0123456789ABCDEFlo, vprod9x0123456789ABCDEFhi));
659 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod9x0123456789ABCDEFlo, vprod9x0123456789ABCDEFhi));
660
661 const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10));
662 const __m256i vk10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 320)));
663 i10 += 16;
664
665 const __m256i vprod10x0123456789ABCDEFlo = _mm256_mullo_epi16(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF);
666 const __m256i vprod10x0123456789ABCDEFhi = _mm256_srai_epi16(vprod10x0123456789ABCDEFlo, 15);
667
668 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod10x0123456789ABCDEFlo, vprod10x0123456789ABCDEFhi));
669 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod10x0123456789ABCDEFlo, vprod10x0123456789ABCDEFhi));
670
671 const __m256i vi11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i11));
672 const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 352)));
673 i11 += 16;
674
675 const __m256i vprod11x0123456789ABCDEFlo = _mm256_mullo_epi16(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF);
676 const __m256i vprod11x0123456789ABCDEFhi = _mm256_srai_epi16(vprod11x0123456789ABCDEFlo, 15);
677
678 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod11x0123456789ABCDEFlo, vprod11x0123456789ABCDEFhi));
679 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod11x0123456789ABCDEFlo, vprod11x0123456789ABCDEFhi));
680
681 const __m256i vi12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i12));
682 const __m256i vk12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 384)));
683 i12 += 16;
684
685 const __m256i vprod12x0123456789ABCDEFlo = _mm256_mullo_epi16(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF);
686 const __m256i vprod12x0123456789ABCDEFhi = _mm256_srai_epi16(vprod12x0123456789ABCDEFlo, 15);
687
688 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod12x0123456789ABCDEFlo, vprod12x0123456789ABCDEFhi));
689 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod12x0123456789ABCDEFlo, vprod12x0123456789ABCDEFhi));
690
691 const __m256i vi13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i13));
692 const __m256i vk13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 416)));
693 i13 += 16;
694
695 const __m256i vprod13x0123456789ABCDEFlo = _mm256_mullo_epi16(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF);
696 const __m256i vprod13x0123456789ABCDEFhi = _mm256_srai_epi16(vprod13x0123456789ABCDEFlo, 15);
697
698 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod13x0123456789ABCDEFlo, vprod13x0123456789ABCDEFhi));
699 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod13x0123456789ABCDEFlo, vprod13x0123456789ABCDEFhi));
700
701 const __m256i vi14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i14));
702 const __m256i vk14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 448)));
703 i14 += 16;
704
705 const __m256i vprod14x0123456789ABCDEFlo = _mm256_mullo_epi16(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF);
706 const __m256i vprod14x0123456789ABCDEFhi = _mm256_srai_epi16(vprod14x0123456789ABCDEFlo, 15);
707
708 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod14x0123456789ABCDEFlo, vprod14x0123456789ABCDEFhi));
709 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod14x0123456789ABCDEFlo, vprod14x0123456789ABCDEFhi));
710
711 const __m256i vi15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i15));
712 const __m256i vk15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 480)));
713 i15 += 16;
714
715 const __m256i vprod15x0123456789ABCDEFlo = _mm256_mullo_epi16(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF);
716 const __m256i vprod15x0123456789ABCDEFhi = _mm256_srai_epi16(vprod15x0123456789ABCDEFlo, 15);
717
718 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod15x0123456789ABCDEFlo, vprod15x0123456789ABCDEFhi));
719 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod15x0123456789ABCDEFlo, vprod15x0123456789ABCDEFhi));
720
721 const __m256i vi16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i16));
722 const __m256i vk16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 512)));
723 i16 += 16;
724
725 const __m256i vprod16x0123456789ABCDEFlo = _mm256_mullo_epi16(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF);
726 const __m256i vprod16x0123456789ABCDEFhi = _mm256_srai_epi16(vprod16x0123456789ABCDEFlo, 15);
727
728 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod16x0123456789ABCDEFlo, vprod16x0123456789ABCDEFhi));
729 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod16x0123456789ABCDEFlo, vprod16x0123456789ABCDEFhi));
730
731 const __m256i vi17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i17));
732 const __m256i vk17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 544)));
733 i17 += 16;
734
735 const __m256i vprod17x0123456789ABCDEFlo = _mm256_mullo_epi16(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF);
736 const __m256i vprod17x0123456789ABCDEFhi = _mm256_srai_epi16(vprod17x0123456789ABCDEFlo, 15);
737
738 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod17x0123456789ABCDEFlo, vprod17x0123456789ABCDEFhi));
739 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod17x0123456789ABCDEFlo, vprod17x0123456789ABCDEFhi));
740
741 const __m256i vi18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i18));
742 const __m256i vk18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 576)));
743 i18 += 16;
744
745 const __m256i vprod18x0123456789ABCDEFlo = _mm256_mullo_epi16(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF);
746 const __m256i vprod18x0123456789ABCDEFhi = _mm256_srai_epi16(vprod18x0123456789ABCDEFlo, 15);
747
748 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod18x0123456789ABCDEFlo, vprod18x0123456789ABCDEFhi));
749 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod18x0123456789ABCDEFlo, vprod18x0123456789ABCDEFhi));
750
751 const __m256i vi19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i19));
752 const __m256i vk19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 608)));
753 i19 += 16;
754
755 const __m256i vprod19x0123456789ABCDEFlo = _mm256_mullo_epi16(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF);
756 const __m256i vprod19x0123456789ABCDEFhi = _mm256_srai_epi16(vprod19x0123456789ABCDEFlo, 15);
757
758 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod19x0123456789ABCDEFlo, vprod19x0123456789ABCDEFhi));
759 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod19x0123456789ABCDEFlo, vprod19x0123456789ABCDEFhi));
760
761 const __m256i vi20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i20));
762 const __m256i vk20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 640)));
763 i20 += 16;
764
765 const __m256i vprod20x0123456789ABCDEFlo = _mm256_mullo_epi16(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF);
766 const __m256i vprod20x0123456789ABCDEFhi = _mm256_srai_epi16(vprod20x0123456789ABCDEFlo, 15);
767
768 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod20x0123456789ABCDEFlo, vprod20x0123456789ABCDEFhi));
769 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod20x0123456789ABCDEFlo, vprod20x0123456789ABCDEFhi));
770
771 const __m256i vi21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i21));
772 const __m256i vk21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 672)));
773 i21 += 16;
774
775 const __m256i vprod21x0123456789ABCDEFlo = _mm256_mullo_epi16(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF);
776 const __m256i vprod21x0123456789ABCDEFhi = _mm256_srai_epi16(vprod21x0123456789ABCDEFlo, 15);
777
778 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod21x0123456789ABCDEFlo, vprod21x0123456789ABCDEFhi));
779 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod21x0123456789ABCDEFlo, vprod21x0123456789ABCDEFhi));
780
781 const __m256i vi22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i22));
782 const __m256i vk22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 704)));
783 i22 += 16;
784
785 const __m256i vprod22x0123456789ABCDEFlo = _mm256_mullo_epi16(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF);
786 const __m256i vprod22x0123456789ABCDEFhi = _mm256_srai_epi16(vprod22x0123456789ABCDEFlo, 15);
787
788 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod22x0123456789ABCDEFlo, vprod22x0123456789ABCDEFhi));
789 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod22x0123456789ABCDEFlo, vprod22x0123456789ABCDEFhi));
790
791 const __m256i vi23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i23));
792 const __m256i vk23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 736)));
793 i23 += 16;
794
795 const __m256i vprod23x0123456789ABCDEFlo = _mm256_mullo_epi16(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF);
796 const __m256i vprod23x0123456789ABCDEFhi = _mm256_srai_epi16(vprod23x0123456789ABCDEFlo, 15);
797
798 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod23x0123456789ABCDEFlo, vprod23x0123456789ABCDEFhi));
799 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod23x0123456789ABCDEFlo, vprod23x0123456789ABCDEFhi));
800
801 const __m256i vi24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i24));
802 const __m256i vk24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 768)));
803 i24 += 16;
804
805 const __m256i vprod24x0123456789ABCDEFlo = _mm256_mullo_epi16(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF);
806 const __m256i vprod24x0123456789ABCDEFhi = _mm256_srai_epi16(vprod24x0123456789ABCDEFlo, 15);
807
808 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod24x0123456789ABCDEFlo, vprod24x0123456789ABCDEFhi));
809 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod24x0123456789ABCDEFlo, vprod24x0123456789ABCDEFhi));
810
811 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
812 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
813
814 k += 16;
815
816 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
817 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
818
819 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t)));
820 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t) + 8 * sizeof(float)));
821 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
822 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
823
824 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
825 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
826 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
827
828 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
829 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
830
831 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
832
833 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
834 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
835 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), voutput_zero_point);
836
837 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
838
839 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
840 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
841
842 if XNN_LIKELY(c >= 16) {
843 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
844 output += 16;
845 c -= 16;
846 } else {
847 if (c & 8) {
848 _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
849 vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
850 output += 8;
851 }
852 if (c & 4) {
853 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
854 vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
855 output += 4;
856 }
857 if (c & 2) {
858 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
859 vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
860 output += 2;
861 }
862 if (c & 1) {
863 *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
864 output += 1;
865 }
866 c = 0;
867 }
868 } while (c != 0);
869 }
870
871 output = (int8_t*) ((uintptr_t) output + output_increment);
872 } while (--output_width != 0);
873 }
874