Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l16c4s4r-minmax-fp32-sse41-mul32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-sse-mul32.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 <immintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qs8_dwconv_minmax_fp32_ukernel_6f6m7l16c4s4r__sse41_mul32(
21 size_t channels,
22 size_t output_width,
23 const int8_t** input,
24 const void* weights,
25 int8_t* output,
26 intptr_t input_stride,
27 size_t output_increment,
28 size_t input_offset,
29 const int8_t* zero,
30 size_t kernel_size,
31 int32_t* buffer,
32 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(channels != 0);
35 assert(output_width != 0);
36 assert(kernel_size > 6);
37
38
39 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
40 XNN_FORCE_REALIZATION(vscale);
41
42 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);
43 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
44 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
45 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
46 XNN_FORCE_REALIZATION(voutput_zero_point);
47 XNN_FORCE_REALIZATION(voutput_min);
48
49 do {
50 const void* w = weights;
51
52 // First pass to process 6 inputs.
53 {
54 int32_t* b = buffer;
55 const int8_t* i0 = input[0];
56 assert(i0 != NULL);
57 if XNN_UNPREDICTABLE(i0 != zero) {
58 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
59 }
60 const int8_t* i1 = input[1];
61 assert(i1 != NULL);
62 if XNN_UNPREDICTABLE(i1 != zero) {
63 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
64 }
65 const int8_t* i2 = input[2];
66 assert(i2 != NULL);
67 if XNN_UNPREDICTABLE(i2 != zero) {
68 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
69 }
70 const int8_t* i3 = input[3];
71 assert(i3 != NULL);
72 if XNN_UNPREDICTABLE(i3 != zero) {
73 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
74 }
75 const int8_t* i4 = input[4];
76 assert(i4 != NULL);
77 if XNN_UNPREDICTABLE(i4 != zero) {
78 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
79 }
80 const int8_t* i5 = input[5];
81 assert(i5 != NULL);
82 if XNN_UNPREDICTABLE(i5 != zero) {
83 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
84 }
85 input += 6;
86
87 size_t c = round_up_po2(channels, 4);
88
89 for (; c >= 16; c -= 16) {
90 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
91 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
92 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
93 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
94
95
96 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
97 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
98 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
99 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
100 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
101 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
102 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
103 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
104 i0 += 16;
105
106 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
107 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
108 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
109 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
110
111 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
112 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
113 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
114 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
115 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
116 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
117 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
118 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
119 i1 += 16;
120
121 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
122 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
123 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
124 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
125
126 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
127 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
128 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
129 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
130 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
131 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
132 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
133 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
134 i2 += 16;
135
136 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
137 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
138 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
139 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
140
141 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
142 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)))));
143 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
144 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(int8_t)))));
145 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
146 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)))));
147 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
148 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(int8_t)))));
149 i3 += 16;
150
151 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
152 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
153 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
154 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
155
156 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
157 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)))));
158 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
159 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(int8_t)))));
160 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
161 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)))));
162 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
163 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(int8_t)))));
164 i4 += 16;
165
166 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
167 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
168 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
169 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
170
171 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
172 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)))));
173 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
174 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(int8_t)))));
175 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
176 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)))));
177 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
178 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(int8_t)))));
179 i5 += 16;
180
181 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
182 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
183 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
184 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
185
186 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t));
187
188 _mm_storeu_si128((__m128i*) b, vacc0123);
189 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
190 _mm_storeu_si128((__m128i*) (b + 8), vacc89AB);
191 _mm_storeu_si128((__m128i*) (b + 12), vaccCDEF);
192 b += 16;
193 }
194
195 if XNN_UNLIKELY(c != 0) {
196 do {
197 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
198
199 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
200 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))));
201 i0 += 4;
202
203 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
204
205 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
206 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
207 i1 += 4;
208
209 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
210
211 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
212 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
213 i2 += 4;
214
215 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
216
217 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
218 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
219 i3 += 4;
220
221 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
222
223 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
224 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
225 i4 += 4;
226
227 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
228
229 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
230 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
231 i5 += 4;
232
233 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
234
235 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t) + 24 * sizeof(int8_t));
236
237 _mm_storeu_si128((__m128i*) b, vacc0123);
238 b += 4;
239 c -= 4;
240 } while (c != 0);
241 }
242 }
243
244 // Middle pass to process 6 inputs in each iteration.
245 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
246 int32_t* b = buffer;
247 const int8_t* i0 = input[0];
248 assert(i0 != NULL);
249 if XNN_UNPREDICTABLE(i0 != zero) {
250 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
251 }
252 const int8_t* i1 = input[1];
253 assert(i1 != NULL);
254 if XNN_UNPREDICTABLE(i1 != zero) {
255 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
256 }
257 const int8_t* i2 = input[2];
258 assert(i2 != NULL);
259 if XNN_UNPREDICTABLE(i2 != zero) {
260 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
261 }
262 const int8_t* i3 = input[3];
263 assert(i3 != NULL);
264 if XNN_UNPREDICTABLE(i3 != zero) {
265 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
266 }
267 const int8_t* i4 = input[4];
268 assert(i4 != NULL);
269 if XNN_UNPREDICTABLE(i4 != zero) {
270 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
271 }
272 const int8_t* i5 = input[5];
273 assert(i5 != NULL);
274 if XNN_UNPREDICTABLE(i5 != zero) {
275 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
276 }
277 input += 6;
278
279 size_t c = round_up_po2(channels, 4);
280
281 for (; c >= 16; c -= 16) {
282 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
283 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
284 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) (b + 8));
285 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) (b + 12));
286
287
288 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
289 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 0 * sizeof(int8_t)))));
290 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
291 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
292 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
293 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
294 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
295 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
296 i0 += 16;
297
298 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
299 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
300 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
301 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
302
303 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
304 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
305 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
306 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
307 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
308 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int8_t)))));
309 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
310 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 28 * sizeof(int8_t)))));
311 i1 += 16;
312
313 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
314 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
315 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
316 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
317
318 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
319 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 32 * sizeof(int8_t)))));
320 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
321 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 36 * sizeof(int8_t)))));
322 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
323 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 40 * sizeof(int8_t)))));
324 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
325 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 44 * sizeof(int8_t)))));
326 i2 += 16;
327
328 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
329 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
330 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
331 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
332
333 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
334 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 48 * sizeof(int8_t)))));
335 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
336 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 52 * sizeof(int8_t)))));
337 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
338 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 56 * sizeof(int8_t)))));
339 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
340 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 60 * sizeof(int8_t)))));
341 i3 += 16;
342
343 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
344 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
345 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
346 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
347
348 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
349 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 64 * sizeof(int8_t)))));
350 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
351 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 68 * sizeof(int8_t)))));
352 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
353 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 72 * sizeof(int8_t)))));
354 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
355 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 76 * sizeof(int8_t)))));
356 i4 += 16;
357
358 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
359 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
360 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
361 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
362
363 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
364 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 80 * sizeof(int8_t)))));
365 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
366 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 84 * sizeof(int8_t)))));
367 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
368 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 88 * sizeof(int8_t)))));
369 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
370 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 92 * sizeof(int8_t)))));
371 i5 += 16;
372
373 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
374 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
375 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
376 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
377
378 w = (const void*) ((uintptr_t) w + 96 * sizeof(int8_t));
379
380 _mm_storeu_si128((__m128i*) b, vacc0123);
381 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
382 _mm_storeu_si128((__m128i*) (b + 8), vacc89AB);
383 _mm_storeu_si128((__m128i*) (b + 12), vaccCDEF);
384 b += 16;
385 }
386
387 if XNN_UNLIKELY(c != 0) {
388 do {
389 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
390
391 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
392 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w))));
393 i0 += 4;
394
395 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
396
397 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
398 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
399 i1 += 4;
400
401 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
402
403 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
404 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
405 i2 += 4;
406
407 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
408
409 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
410 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
411 i3 += 4;
412
413 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
414
415 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
416 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
417 i4 += 4;
418
419 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
420
421 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
422 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
423 i5 += 4;
424
425 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
426
427 w = (const void*) ((uintptr_t) w + 24 * sizeof(int8_t));
428
429 _mm_storeu_si128((__m128i*) b, vacc0123);
430 b += 4;
431 c -= 4;
432 } while (c != 0);
433 }
434 }
435
436 // Last pass to process up to 7 inputs.
437 {
438 const int32_t* b = buffer;
439 const int8_t* i0 = input[0];
440 assert(i0 != NULL);
441 if XNN_UNPREDICTABLE(i0 != zero) {
442 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
443 }
444 const int8_t* i1 = input[1];
445 assert(i1 != NULL);
446 if XNN_UNPREDICTABLE(i1 != zero) {
447 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
448 }
449 const int8_t* i2 = input[2];
450 assert(i2 != NULL);
451 if XNN_UNPREDICTABLE(i2 != zero) {
452 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
453 }
454 const int8_t* i3 = input[3];
455 assert(i3 != NULL);
456 if XNN_UNPREDICTABLE(i3 != zero) {
457 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
458 }
459 const int8_t* i4 = input[4];
460 assert(i4 != NULL);
461 if XNN_UNPREDICTABLE(i4 != zero) {
462 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
463 }
464 const int8_t* i5 = input[5];
465 assert(i5 != NULL);
466 if XNN_UNPREDICTABLE(i5 != zero) {
467 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
468 }
469 const int8_t* i6 = input[6];
470 assert(i6 != NULL);
471 if XNN_UNPREDICTABLE(i6 != zero) {
472 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
473 }
474
475 size_t c = channels;
476
477 for (; c >= 16; c -= 16) {
478 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
479 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
480 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) (b + 8));
481 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) (b + 12));
482 b += 16;
483
484
485 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
486 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 0 * sizeof(int8_t)))));
487 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
488 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
489 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
490 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
491 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
492 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
493 i0 += 16;
494
495 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
496 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
497 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
498 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
499
500 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
501 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
502 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
503 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
504 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
505 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int8_t)))));
506 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
507 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 28 * sizeof(int8_t)))));
508 i1 += 16;
509
510 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
511 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
512 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
513 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
514
515 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
516 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 32 * sizeof(int8_t)))));
517 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
518 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 36 * sizeof(int8_t)))));
519 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
520 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 40 * sizeof(int8_t)))));
521 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
522 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 44 * sizeof(int8_t)))));
523 i2 += 16;
524
525 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
526 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
527 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
528 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
529
530 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
531 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 48 * sizeof(int8_t)))));
532 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
533 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 52 * sizeof(int8_t)))));
534 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
535 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 56 * sizeof(int8_t)))));
536 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
537 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 60 * sizeof(int8_t)))));
538 i3 += 16;
539
540 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
541 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
542 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
543 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
544
545 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
546 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 64 * sizeof(int8_t)))));
547 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
548 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 68 * sizeof(int8_t)))));
549 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
550 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 72 * sizeof(int8_t)))));
551 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
552 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 76 * sizeof(int8_t)))));
553 i4 += 16;
554
555 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
556 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
557 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
558 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
559
560 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
561 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 80 * sizeof(int8_t)))));
562 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
563 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 84 * sizeof(int8_t)))));
564 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
565 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 88 * sizeof(int8_t)))));
566 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
567 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 92 * sizeof(int8_t)))));
568 i5 += 16;
569
570 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
571 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
572 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
573 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
574
575 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
576 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 96 * sizeof(int8_t)))));
577 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
578 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 100 * sizeof(int8_t)))));
579 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
580 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 104 * sizeof(int8_t)))));
581 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
582 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 108 * sizeof(int8_t)))));
583 i6 += 16;
584
585 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
586 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
587 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi6x89AB, vk6x89AB));
588 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi6xCDEF, vk6xCDEF));
589
590 w = (const void*) ((uintptr_t) w + 112 * sizeof(int8_t));
591
592 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
593 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
594 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
595 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
596
597 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
598 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
599 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
600 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
601
602 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
603 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
604 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
605 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
606
607 vacc0123 = _mm_cvtps_epi32(vscaled0123);
608 vacc4567 = _mm_cvtps_epi32(vscaled4567);
609 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
610 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
611
612 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
613 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
614
615 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
616 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
617
618 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
619 output += 16;
620 }
621
622 if XNN_UNLIKELY(c != 0) {
623 do {
624 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
625 b += 4;
626
627 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
628 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 0 * sizeof(int8_t)))));
629 i0 += 4;
630
631 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
632
633 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
634 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
635 i1 += 4;
636
637 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
638
639 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
640 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
641 i2 += 4;
642
643 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
644
645 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
646 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
647 i3 += 4;
648
649 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
650
651 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
652 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
653 i4 += 4;
654
655 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
656
657 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
658 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
659 i5 += 4;
660
661 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
662
663 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
664 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int8_t)))));
665 i6 += 4;
666
667 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
668
669 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
670 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
671 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
672 vacc0123 = _mm_cvtps_epi32(vscaled0123);
673
674 w = (void*) ((uintptr_t) w + 28 * sizeof(int8_t));
675
676 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
677
678 vout0123 = _mm_packs_epi16(vout0123, vout0123);
679 vout0123 = _mm_max_epi8(vout0123, voutput_min);
680
681 if XNN_LIKELY(c >= 4) {
682 _mm_storeu_si32(output, vout0123);
683 output += 4;
684 c -= 4;
685 } else {
686 if (c & 2) {
687 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
688 vout0123 = _mm_srli_epi32(vout0123, 16);
689 output += 2;
690 }
691 if (c & 1) {
692 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
693 output += 1;
694 }
695 c = 0;
696 }
697 } while (c != 0);
698 }
699 }
700
701 input = (const int8_t**) ((uintptr_t) input + input_stride);
702 output = (int8_t*) ((uintptr_t) output + output_increment);
703 } while (--output_width != 0);
704 }