Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l8c4s4r-minmax-fp32-avx-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_8f8m9l8c4s4r__avx_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 > 8);
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 8 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 const int8_t* i6 = input[6];
86 assert(i6 != NULL);
87 if XNN_UNPREDICTABLE(i6 != zero) {
88 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
89 }
90 const int8_t* i7 = input[7];
91 assert(i7 != NULL);
92 if XNN_UNPREDICTABLE(i7 != zero) {
93 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
94 }
95 input += 8;
96
97 size_t c = round_up_po2(channels, 4);
98
99 for (; c >= 8; c -= 8) {
100 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
101 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
102
103
104 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
105 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
106 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
107 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
108 i0 += 8;
109
110 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
111 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
112
113 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
114 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
115 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
116 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
117 i1 += 8;
118
119 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
120 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
121
122 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
123 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
124 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
125 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
126 i2 += 8;
127
128 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
129 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
130
131 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
132 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
133 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
134 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
135 i3 += 8;
136
137 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
138 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
139
140 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
141 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
142 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
143 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
144 i4 += 8;
145
146 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
147 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
148
149 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
150 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
151 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
152 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
153 i5 += 8;
154
155 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
156 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
157
158 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
159 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)))));
160 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
161 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 52 * sizeof(int8_t)))));
162 i6 += 8;
163
164 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
165 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
166
167 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
168 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)))));
169 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
170 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 60 * sizeof(int8_t)))));
171 i7 += 8;
172
173 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
174 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
175
176 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t));
177
178 _mm_storeu_si128((__m128i*) b, vacc0123);
179 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
180 b += 8;
181 }
182
183 if XNN_UNLIKELY(c != 0) {
184 do {
185 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
186
187 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
188 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))));
189 i0 += 4;
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
192
193 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
194 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
195 i1 += 4;
196
197 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
198
199 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
200 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
201 i2 += 4;
202
203 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
204
205 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
206 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
207 i3 += 4;
208
209 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
210
211 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
212 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
213 i4 += 4;
214
215 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
216
217 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
218 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
219 i5 += 4;
220
221 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
222
223 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
224 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
225 i6 += 4;
226
227 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
228
229 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
230 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
231 i7 += 4;
232
233 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
234
235 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t) + 32 * 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 8 inputs in each iteration.
245 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
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 const int8_t* i6 = input[6];
278 assert(i6 != NULL);
279 if XNN_UNPREDICTABLE(i6 != zero) {
280 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
281 }
282 const int8_t* i7 = input[7];
283 assert(i7 != NULL);
284 if XNN_UNPREDICTABLE(i7 != zero) {
285 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
286 }
287 input += 8;
288
289 size_t c = round_up_po2(channels, 4);
290
291 for (; c >= 8; c -= 8) {
292 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
293 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
294
295
296 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
297 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 0 * sizeof(int8_t)))));
298 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
299 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
300 i0 += 8;
301
302 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
303 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
304
305 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
306 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
307 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
308 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
309 i1 += 8;
310
311 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
312 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
313
314 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
315 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
316 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
317 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
318 i2 += 8;
319
320 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
321 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
322
323 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
324 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int8_t)))));
325 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
326 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 28 * sizeof(int8_t)))));
327 i3 += 8;
328
329 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
330 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
331
332 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
333 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 32 * sizeof(int8_t)))));
334 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
335 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 36 * sizeof(int8_t)))));
336 i4 += 8;
337
338 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
339 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
340
341 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
342 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 40 * sizeof(int8_t)))));
343 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
344 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 44 * sizeof(int8_t)))));
345 i5 += 8;
346
347 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
348 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
349
350 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
351 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 48 * sizeof(int8_t)))));
352 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
353 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 52 * sizeof(int8_t)))));
354 i6 += 8;
355
356 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
357 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
358
359 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
360 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 56 * sizeof(int8_t)))));
361 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
362 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 60 * sizeof(int8_t)))));
363 i7 += 8;
364
365 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
366 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
367
368 w = (const void*) ((uintptr_t) w + 64 * sizeof(int8_t));
369
370 _mm_storeu_si128((__m128i*) b, vacc0123);
371 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
372 b += 8;
373 }
374
375 if XNN_UNLIKELY(c != 0) {
376 do {
377 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
378
379 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
380 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w))));
381 i0 += 4;
382
383 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
384
385 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
386 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
387 i1 += 4;
388
389 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
390
391 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
392 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
393 i2 += 4;
394
395 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
396
397 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
398 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
399 i3 += 4;
400
401 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
402
403 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
404 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
405 i4 += 4;
406
407 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
408
409 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
410 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
411 i5 += 4;
412
413 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
414
415 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
416 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 24 * sizeof(int8_t)))));
417 i6 += 4;
418
419 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
420
421 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
422 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 28 * sizeof(int8_t)))));
423 i7 += 4;
424
425 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
426
427 w = (const void*) ((uintptr_t) w + 32 * 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 9 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 const int8_t* i7 = input[7];
475 assert(i7 != NULL);
476 if XNN_UNPREDICTABLE(i7 != zero) {
477 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
478 }
479 const int8_t* i8 = input[8];
480 assert(i8 != NULL);
481 if XNN_UNPREDICTABLE(i8 != zero) {
482 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
483 }
484
485 size_t c = channels;
486
487 for (; c >= 8; c -= 8) {
488 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
489 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
490 b += 8;
491
492
493 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
494 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 0 * sizeof(int8_t)))));
495 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
496 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
497 i0 += 8;
498
499 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
500 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
501
502 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
503 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
504 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
505 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
506 i1 += 8;
507
508 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
509 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
510
511 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
512 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
513 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
514 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
515 i2 += 8;
516
517 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
518 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
519
520 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
521 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int8_t)))));
522 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
523 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 28 * sizeof(int8_t)))));
524 i3 += 8;
525
526 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
527 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
528
529 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
530 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 32 * sizeof(int8_t)))));
531 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
532 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 36 * sizeof(int8_t)))));
533 i4 += 8;
534
535 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
536 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
537
538 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
539 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 40 * sizeof(int8_t)))));
540 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
541 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 44 * sizeof(int8_t)))));
542 i5 += 8;
543
544 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
545 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
546
547 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
548 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 48 * sizeof(int8_t)))));
549 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
550 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 52 * sizeof(int8_t)))));
551 i6 += 8;
552
553 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
554 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
555
556 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
557 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 56 * sizeof(int8_t)))));
558 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
559 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 60 * sizeof(int8_t)))));
560 i7 += 8;
561
562 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
563 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
564
565 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
566 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 64 * sizeof(int8_t)))));
567 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
568 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 68 * sizeof(int8_t)))));
569 i8 += 8;
570
571 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
572 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
573
574 w = (const void*) ((uintptr_t) w + 72 * sizeof(int8_t));
575
576 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
577 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
578
579 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
580 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
581
582 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
583 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
584
585 vacc0123 = _mm_cvtps_epi32(vscaled0123);
586 vacc4567 = _mm_cvtps_epi32(vscaled4567);
587
588 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
589
590 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
591 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
592
593 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
594 output += 8;
595 }
596
597 if XNN_UNLIKELY(c != 0) {
598 do {
599 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
600 b += 4;
601
602 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
603 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 0 * sizeof(int8_t)))));
604 i0 += 4;
605
606 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
607
608 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
609 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
610 i1 += 4;
611
612 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
613
614 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
615 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
616 i2 += 4;
617
618 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
619
620 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
621 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
622 i3 += 4;
623
624 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
625
626 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
627 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
628 i4 += 4;
629
630 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
631
632 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
633 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
634 i5 += 4;
635
636 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
637
638 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
639 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int8_t)))));
640 i6 += 4;
641
642 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
643
644 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
645 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 28 * sizeof(int8_t)))));
646 i7 += 4;
647
648 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
649
650 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
651 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 32 * sizeof(int8_t)))));
652 i8 += 4;
653
654 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
655
656 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
657 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
658 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
659 vacc0123 = _mm_cvtps_epi32(vscaled0123);
660
661 w = (void*) ((uintptr_t) w + 36 * sizeof(int8_t));
662
663 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
664
665 vout0123 = _mm_packs_epi16(vout0123, vout0123);
666 vout0123 = _mm_max_epi8(vout0123, voutput_min);
667
668 if XNN_LIKELY(c >= 4) {
669 _mm_storeu_si32(output, vout0123);
670 output += 4;
671 c -= 4;
672 } else {
673 if (c & 2) {
674 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
675 vout0123 = _mm_srli_epi32(vout0123, 16);
676 output += 2;
677 }
678 if (c & 1) {
679 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
680 output += 1;
681 }
682 c = 0;
683 }
684 } while (c != 0);
685 }
686 }
687
688 input = (const int8_t**) ((uintptr_t) input + input_stride);
689 output = (int8_t*) ((uintptr_t) output + output_increment);
690 } while (--output_width != 0);
691 }