Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l16c8s8r-minmax-fp32-sse2-mul16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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 <emmintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_dwconv_minmax_fp32_ukernel_8f8m9l16c8s8r__sse2_mul16(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 intptr_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 size_t kernel_size,
30 int32_t* buffer,
31 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(channels != 0);
34 assert(output_width != 0);
35 assert(kernel_size > 8);
36
37
38 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
39 XNN_FORCE_REALIZATION(vscale);
40
41 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
42 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
43 const __m128i voutput_min = _mm_set1_epi16(params->fp32_scalar.output_min);
44 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
45 XNN_FORCE_REALIZATION(voutput_zero_point);
46 XNN_FORCE_REALIZATION(voutput_min);
47
48 do {
49 const void* w = weights;
50
51 // First pass to process 8 inputs.
52 {
53 int32_t* b = buffer;
54 const int8_t* i0 = input[0];
55 assert(i0 != NULL);
56 if XNN_UNPREDICTABLE(i0 != zero) {
57 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
58 }
59 const int8_t* i1 = input[1];
60 assert(i1 != NULL);
61 if XNN_UNPREDICTABLE(i1 != zero) {
62 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
63 }
64 const int8_t* i2 = input[2];
65 assert(i2 != NULL);
66 if XNN_UNPREDICTABLE(i2 != zero) {
67 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
68 }
69 const int8_t* i3 = input[3];
70 assert(i3 != NULL);
71 if XNN_UNPREDICTABLE(i3 != zero) {
72 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
73 }
74 const int8_t* i4 = input[4];
75 assert(i4 != NULL);
76 if XNN_UNPREDICTABLE(i4 != zero) {
77 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
78 }
79 const int8_t* i5 = input[5];
80 assert(i5 != NULL);
81 if XNN_UNPREDICTABLE(i5 != zero) {
82 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
83 }
84 const int8_t* i6 = input[6];
85 assert(i6 != NULL);
86 if XNN_UNPREDICTABLE(i6 != zero) {
87 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
88 }
89 const int8_t* i7 = input[7];
90 assert(i7 != NULL);
91 if XNN_UNPREDICTABLE(i7 != zero) {
92 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
93 }
94 input += 8;
95
96 size_t c = round_up_po2(channels, 8);
97
98 for (; c >= 16; c -= 16) {
99 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
100 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
101 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
102 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
103
104
105
106 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
107 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
108 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
109 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
110 i0 += 16;
111
112 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
113 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
114 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
115 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
116
117 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
118 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
119 const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
120 const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
121
122 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
123 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
124 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
125 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
126
127 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
128 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
129 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
130 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
131 i1 += 16;
132
133 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
134 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
135 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
136 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
137
138 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
139 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
140 const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
141 const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
142
143 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
144 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
145 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
146 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
147
148 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
149 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
150 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
151 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
152 i2 += 16;
153
154 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
155 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
156 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
157 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
158
159 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
160 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
161 const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
162 const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
163
164 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
165 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
166 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
167 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
168
169 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
170 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
171 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
172 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
173 i3 += 16;
174
175 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
176 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
177 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
178 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
179
180 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
181 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
182 const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
183 const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
184
185 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
186 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
187 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
188 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
189
190 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
191 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
192 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
193 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
194 i4 += 16;
195
196 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
197 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
198 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
199 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
200
201 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
202 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
203 const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
204 const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
205
206 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
207 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
208 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
209 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
210
211 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
212 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
213 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
214 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
215 i5 += 16;
216
217 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
218 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
219 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
220 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
221
222 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
223 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
224 const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
225 const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
226
227 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
228 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
229 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
230 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
231
232 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
233 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
234 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
235 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
236 i6 += 16;
237
238 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
239 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
240 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
241 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
242
243 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
244 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
245 const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
246 const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
247
248 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
249 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
250 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
251 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
252
253 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
254 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
255 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
256 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
257 i7 += 16;
258
259 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
260 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
261 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
262 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
263
264 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
265 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
266 const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
267 const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
268
269 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
270 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
271 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
272 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
273
274 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t));
275
276 _mm_storeu_si128((__m128i*) b, vacc0123);
277 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
278 _mm_storeu_si128((__m128i*) (b + 8), vacc89AB);
279 _mm_storeu_si128((__m128i*) (b + 12), vaccCDEF);
280 b += 16;
281 }
282
283 if XNN_UNLIKELY(c != 0) {
284 do {
285 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
286 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
287
288
289 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
290 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 * sizeof(int8_t)));
291 i0 += 8;
292
293 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
294 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
295
296 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
297 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
298
299 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
300 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
301
302 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
303 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 8 * sizeof(int8_t)));
304 i1 += 8;
305
306 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
307 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
308
309 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
310 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
311
312 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
313 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
314
315 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
316 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 16 * sizeof(int8_t)));
317 i2 += 8;
318
319 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
320 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
321
322 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
323 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
324
325 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
326 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
327
328 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
329 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 24 * sizeof(int8_t)));
330 i3 += 8;
331
332 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
333 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
334
335 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
336 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
337
338 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
339 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
340
341 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
342 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 32 * sizeof(int8_t)));
343 i4 += 8;
344
345 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
346 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
347
348 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
349 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
350
351 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
352 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
353
354 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
355 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 40 * sizeof(int8_t)));
356 i5 += 8;
357
358 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
359 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
360
361 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
362 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
363
364 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
365 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
366
367 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
368 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 48 * sizeof(int8_t)));
369 i6 += 8;
370
371 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
372 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
373
374 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
375 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
376
377 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
378 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
379
380 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
381 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 56 * sizeof(int8_t)));
382 i7 += 8;
383
384 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
385 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
386
387 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
388 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
389
390 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
391 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
392
393 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t));
394
395 _mm_storeu_si128((__m128i*) (b), vacc0123);
396 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
397 b += 8;
398 c -= 8;
399 } while (c != 0);
400 }
401 }
402
403 // Middle pass to process 8 inputs in each iteration.
404 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
405 int32_t* b = buffer;
406 const int8_t* i0 = input[0];
407 assert(i0 != NULL);
408 if XNN_UNPREDICTABLE(i0 != zero) {
409 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
410 }
411 const int8_t* i1 = input[1];
412 assert(i1 != NULL);
413 if XNN_UNPREDICTABLE(i1 != zero) {
414 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
415 }
416 const int8_t* i2 = input[2];
417 assert(i2 != NULL);
418 if XNN_UNPREDICTABLE(i2 != zero) {
419 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
420 }
421 const int8_t* i3 = input[3];
422 assert(i3 != NULL);
423 if XNN_UNPREDICTABLE(i3 != zero) {
424 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
425 }
426 const int8_t* i4 = input[4];
427 assert(i4 != NULL);
428 if XNN_UNPREDICTABLE(i4 != zero) {
429 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
430 }
431 const int8_t* i5 = input[5];
432 assert(i5 != NULL);
433 if XNN_UNPREDICTABLE(i5 != zero) {
434 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
435 }
436 const int8_t* i6 = input[6];
437 assert(i6 != NULL);
438 if XNN_UNPREDICTABLE(i6 != zero) {
439 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
440 }
441 const int8_t* i7 = input[7];
442 assert(i7 != NULL);
443 if XNN_UNPREDICTABLE(i7 != zero) {
444 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
445 }
446 input += 8;
447
448 size_t c = round_up_po2(channels, 8);
449
450 for (; c >= 16; c -= 16) {
451 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
452 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
453 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) (b + 8));
454 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) (b + 12));
455
456
457
458 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
459 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
460 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
461 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
462 i0 += 16;
463
464 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
465 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
466 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
467 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
468
469 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
470 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
471 const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
472 const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
473
474 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
475 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
476 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
477 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
478
479 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
480 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
481 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
482 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
483 i1 += 16;
484
485 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
486 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
487 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
488 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
489
490 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
491 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
492 const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
493 const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
494
495 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
496 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
497 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
498 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
499
500 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
501 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
502 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
503 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
504 i2 += 16;
505
506 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
507 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
508 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
509 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
510
511 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
512 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
513 const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
514 const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
515
516 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
517 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
518 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
519 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
520
521 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
522 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
523 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
524 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
525 i3 += 16;
526
527 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
528 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
529 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
530 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
531
532 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
533 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
534 const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
535 const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
536
537 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
538 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
539 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
540 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
541
542 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
543 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
544 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
545 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72 * sizeof(int8_t)));
546 i4 += 16;
547
548 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
549 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
550 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
551 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
552
553 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
554 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
555 const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
556 const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
557
558 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
559 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
560 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
561 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
562
563 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
564 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t)));
565 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
566 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 88 * sizeof(int8_t)));
567 i5 += 16;
568
569 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
570 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
571 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
572 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
573
574 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
575 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
576 const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
577 const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
578
579 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
580 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
581 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
582 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
583
584 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
585 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t)));
586 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
587 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 104 * sizeof(int8_t)));
588 i6 += 16;
589
590 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
591 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
592 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
593 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
594
595 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
596 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
597 const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
598 const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
599
600 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
601 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
602 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
603 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
604
605 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
606 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t)));
607 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
608 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 120 * sizeof(int8_t)));
609 i7 += 16;
610
611 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
612 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
613 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
614 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
615
616 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
617 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
618 const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
619 const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
620
621 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
622 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
623 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
624 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
625
626 w = (const void*) ((uintptr_t) w + 128 * sizeof(int8_t));
627
628 _mm_storeu_si128((__m128i*) (b), vacc0123);
629 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
630 _mm_storeu_si128((__m128i*) (b + 8), vacc89AB);
631 _mm_storeu_si128((__m128i*) (b + 12), vaccCDEF);
632 b += 16;
633 }
634
635 if XNN_UNLIKELY(c != 0) {
636 do {
637 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
638 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
639
640
641 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
642 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) w);
643 i0 += 8;
644
645 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
646 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
647
648 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
649 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
650
651 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
652 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
653
654 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
655 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
656 i1 += 8;
657
658 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
659 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
660
661 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
662 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
663
664 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
665 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
666
667 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
668 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
669 i2 += 8;
670
671 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
672 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
673
674 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
675 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
676
677 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
678 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
679
680 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
681 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
682 i3 += 8;
683
684 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
685 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
686
687 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
688 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
689
690 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
691 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
692
693 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
694 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
695 i4 += 8;
696
697 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
698 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
699
700 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
701 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
702
703 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
704 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
705
706 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
707 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
708 i5 += 8;
709
710 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
711 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
712
713 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
714 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
715
716 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
717 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
718
719 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
720 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
721 i6 += 8;
722
723 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
724 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
725
726 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
727 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
728
729 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
730 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
731
732 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
733 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
734 i7 += 8;
735
736 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
737 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
738
739 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
740 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
741
742 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
743 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
744
745 w = (const void*) ((uintptr_t) w + 64 * sizeof(int8_t));
746
747 _mm_storeu_si128((__m128i*) (b), vacc0123);
748 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
749 b += 8;
750 c -= 8;
751 } while (c != 0);
752 }
753 }
754
755 // Last pass to process up to 9 inputs.
756 {
757 const int32_t* b = buffer;
758 const int8_t* i0 = input[0];
759 assert(i0 != NULL);
760 if XNN_UNPREDICTABLE(i0 != zero) {
761 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
762 }
763 const int8_t* i1 = input[1];
764 assert(i1 != NULL);
765 if XNN_UNPREDICTABLE(i1 != zero) {
766 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
767 }
768 const int8_t* i2 = input[2];
769 assert(i2 != NULL);
770 if XNN_UNPREDICTABLE(i2 != zero) {
771 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
772 }
773 const int8_t* i3 = input[3];
774 assert(i3 != NULL);
775 if XNN_UNPREDICTABLE(i3 != zero) {
776 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
777 }
778 const int8_t* i4 = input[4];
779 assert(i4 != NULL);
780 if XNN_UNPREDICTABLE(i4 != zero) {
781 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
782 }
783 const int8_t* i5 = input[5];
784 assert(i5 != NULL);
785 if XNN_UNPREDICTABLE(i5 != zero) {
786 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
787 }
788 const int8_t* i6 = input[6];
789 assert(i6 != NULL);
790 if XNN_UNPREDICTABLE(i6 != zero) {
791 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
792 }
793 const int8_t* i7 = input[7];
794 assert(i7 != NULL);
795 if XNN_UNPREDICTABLE(i7 != zero) {
796 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
797 }
798 const int8_t* i8 = input[8];
799 assert(i8 != NULL);
800 if XNN_UNPREDICTABLE(i8 != zero) {
801 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
802 }
803
804 size_t c = channels;
805
806 for (; c >= 16; c -= 16) {
807 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
808 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
809 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) (b + 8));
810 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) (b + 12));
811 b += 16;
812
813
814 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
815 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
816 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
817 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
818 i0 += 16;
819
820 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
821 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
822 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
823 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
824
825 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
826 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
827 const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
828 const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
829
830 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
831 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
832 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
833 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
834
835 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
836 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
837 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
838 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
839 i1 += 16;
840
841 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
842 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
843 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
844 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
845
846 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
847 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
848 const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
849 const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
850
851 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
852 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
853 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
854 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
855
856 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
857 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
858 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
859 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
860 i2 += 16;
861
862 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
863 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
864 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
865 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
866
867 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
868 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
869 const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
870 const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
871
872 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
873 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
874 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
875 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
876
877 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
878 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
879 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
880 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
881 i3 += 16;
882
883 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
884 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
885 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
886 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
887
888 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
889 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
890 const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
891 const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
892
893 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
894 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
895 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
896 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
897
898 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
899 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
900 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
901 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72 * sizeof(int8_t)));
902 i4 += 16;
903
904 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
905 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
906 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
907 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
908
909 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
910 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
911 const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
912 const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
913
914 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
915 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
916 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
917 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
918
919 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
920 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t)));
921 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
922 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 88 * sizeof(int8_t)));
923 i5 += 16;
924
925 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
926 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
927 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
928 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
929
930 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
931 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
932 const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
933 const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
934
935 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
936 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
937 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
938 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
939
940 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
941 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t)));
942 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
943 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 104 * sizeof(int8_t)));
944 i6 += 16;
945
946 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
947 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
948 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
949 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
950
951 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
952 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
953 const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
954 const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
955
956 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
957 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
958 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
959 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
960
961 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
962 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t)));
963 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
964 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 120 * sizeof(int8_t)));
965 i7 += 16;
966
967 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
968 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
969 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
970 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
971
972 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
973 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
974 const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
975 const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
976
977 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
978 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
979 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
980 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
981
982 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
983 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t)));
984 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
985 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 136 * sizeof(int8_t)));
986 i8 += 16;
987
988 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
989 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
990 const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
991 const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
992
993 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
994 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
995 const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
996 const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
997
998 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
999 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
1000 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
1001 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
1002
1003 w = (const void*) ((uintptr_t) w + 144 * sizeof(int8_t));
1004
1005 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
1006 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
1007 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
1008 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
1009
1010 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
1011 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
1012 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
1013 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
1014
1015 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
1016 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1017 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
1018 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
1019
1020 vacc0123 = _mm_cvtps_epi32(vscaled0123);
1021 vacc4567 = _mm_cvtps_epi32(vscaled4567);
1022 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
1023 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
1024
1025 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1026 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
1027
1028 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
1029 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
1030
1031 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
1032
1033
1034 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
1035 output += 16;
1036 }
1037
1038 if XNN_UNLIKELY(c != 0) {
1039 do {
1040 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
1041 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
1042 b += 8;
1043
1044
1045 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
1046 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) w);
1047 i0 += 8;
1048
1049 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
1050 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
1051
1052 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
1053 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
1054
1055 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
1056 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
1057
1058 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
1059 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
1060 i1 += 8;
1061
1062 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
1063 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
1064
1065 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
1066 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
1067
1068 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
1069 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
1070
1071 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
1072 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
1073 i2 += 8;
1074
1075 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
1076 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
1077
1078 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
1079 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
1080
1081 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
1082 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
1083
1084 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
1085 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
1086 i3 += 8;
1087
1088 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
1089 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
1090
1091 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
1092 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
1093
1094 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
1095 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
1096
1097 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
1098 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
1099 i4 += 8;
1100
1101 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
1102 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
1103
1104 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
1105 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
1106
1107 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
1108 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
1109
1110 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
1111 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
1112 i5 += 8;
1113
1114 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
1115 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
1116
1117 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
1118 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
1119
1120 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
1121 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
1122
1123 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
1124 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
1125 i6 += 8;
1126
1127 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
1128 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
1129
1130 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
1131 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
1132
1133 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
1134 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
1135
1136 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
1137 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
1138 i7 += 8;
1139
1140 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
1141 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
1142
1143 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
1144 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
1145
1146 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
1147 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
1148
1149 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
1150 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
1151 i8 += 8;
1152
1153 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
1154 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
1155
1156 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
1157 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
1158
1159 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
1160 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
1161
1162 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
1163 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
1164
1165 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
1166 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
1167
1168 w = (void*) ((uintptr_t) w + 72);
1169
1170 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
1171 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1172
1173 vacc0123 = _mm_cvtps_epi32(vscaled0123);
1174 vacc4567 = _mm_cvtps_epi32(vscaled4567);
1175
1176 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1177
1178 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
1179
1180 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
1181
1182
1183 if XNN_LIKELY(c >= 8) {
1184 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
1185 output += 8;
1186 c -= 8;
1187 } else {
1188 if (c & 4) {
1189 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
1190 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
1191 output += 4;
1192 }
1193 if (c & 2) {
1194 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
1195 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1196 output += 2;
1197 }
1198 if (c & 1) {
1199 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
1200 output += 1;
1201 }
1202 c = 0;
1203 }
1204 } while (c != 0);
1205 }
1206 }
1207
1208 input = (const int8_t**) ((uintptr_t) input + input_stride);
1209 output = (int8_t*) ((uintptr_t) output + output_increment);
1210 } while (--output_width != 0);
1211 }