Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l8c4s4r-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_6f6m7l8c4s4r__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 >= 8; c -= 8) {
90 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
91 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
92
93
94 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
95 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
96 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
97 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
98 i0 += 8;
99
100 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
101 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
102
103 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
104 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
105 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
106 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
107 i1 += 8;
108
109 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
110 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
111
112 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
113 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
114 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
115 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
116 i2 += 8;
117
118 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
119 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
120
121 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
122 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
123 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
124 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
125 i3 += 8;
126
127 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
128 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
129
130 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
131 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
132 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
133 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
134 i4 += 8;
135
136 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
137 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
138
139 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
140 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
141 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
142 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
143 i5 += 8;
144
145 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
146 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
147
148 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t));
149
150 _mm_storeu_si128((__m128i*) b, vacc0123);
151 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
152 b += 8;
153 }
154
155 if XNN_UNLIKELY(c != 0) {
156 do {
157 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
158
159 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
160 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))));
161 i0 += 4;
162
163 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
164
165 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
166 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
167 i1 += 4;
168
169 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
170
171 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
172 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
173 i2 += 4;
174
175 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
176
177 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
178 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
179 i3 += 4;
180
181 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
182
183 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
184 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
185 i4 += 4;
186
187 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
188
189 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
190 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
191 i5 += 4;
192
193 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
194
195 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t) + 24 * sizeof(int8_t));
196
197 _mm_storeu_si128((__m128i*) b, vacc0123);
198 b += 4;
199 c -= 4;
200 } while (c != 0);
201 }
202 }
203
204 // Middle pass to process 6 inputs in each iteration.
205 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
206 int32_t* b = buffer;
207 const int8_t* i0 = input[0];
208 assert(i0 != NULL);
209 if XNN_UNPREDICTABLE(i0 != zero) {
210 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
211 }
212 const int8_t* i1 = input[1];
213 assert(i1 != NULL);
214 if XNN_UNPREDICTABLE(i1 != zero) {
215 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
216 }
217 const int8_t* i2 = input[2];
218 assert(i2 != NULL);
219 if XNN_UNPREDICTABLE(i2 != zero) {
220 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
221 }
222 const int8_t* i3 = input[3];
223 assert(i3 != NULL);
224 if XNN_UNPREDICTABLE(i3 != zero) {
225 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
226 }
227 const int8_t* i4 = input[4];
228 assert(i4 != NULL);
229 if XNN_UNPREDICTABLE(i4 != zero) {
230 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
231 }
232 const int8_t* i5 = input[5];
233 assert(i5 != NULL);
234 if XNN_UNPREDICTABLE(i5 != zero) {
235 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
236 }
237 input += 6;
238
239 size_t c = round_up_po2(channels, 4);
240
241 for (; c >= 8; c -= 8) {
242 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
243 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
244
245
246 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
247 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 0 * sizeof(int8_t)))));
248 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
249 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
250 i0 += 8;
251
252 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
253 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
254
255 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
256 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
257 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
258 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
259 i1 += 8;
260
261 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
262 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
263
264 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
265 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
266 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
267 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
268 i2 += 8;
269
270 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
271 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
272
273 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
274 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int8_t)))));
275 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
276 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 28 * sizeof(int8_t)))));
277 i3 += 8;
278
279 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
280 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
281
282 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
283 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 32 * sizeof(int8_t)))));
284 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
285 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 36 * sizeof(int8_t)))));
286 i4 += 8;
287
288 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
289 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
290
291 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
292 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 40 * sizeof(int8_t)))));
293 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
294 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 44 * sizeof(int8_t)))));
295 i5 += 8;
296
297 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
298 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
299
300 w = (const void*) ((uintptr_t) w + 48 * sizeof(int8_t));
301
302 _mm_storeu_si128((__m128i*) b, vacc0123);
303 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
304 b += 8;
305 }
306
307 if XNN_UNLIKELY(c != 0) {
308 do {
309 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
310
311 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
312 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w))));
313 i0 += 4;
314
315 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
316
317 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
318 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
319 i1 += 4;
320
321 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
322
323 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
324 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
325 i2 += 4;
326
327 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
328
329 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
330 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
331 i3 += 4;
332
333 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
334
335 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
336 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int32_t*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
337 i4 += 4;
338
339 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
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 int32_t*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
343 i5 += 4;
344
345 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
346
347 w = (const void*) ((uintptr_t) w + 24 * sizeof(int8_t));
348
349 _mm_storeu_si128((__m128i*) b, vacc0123);
350 b += 4;
351 c -= 4;
352 } while (c != 0);
353 }
354 }
355
356 // Last pass to process up to 7 inputs.
357 {
358 const int32_t* b = buffer;
359 const int8_t* i0 = input[0];
360 assert(i0 != NULL);
361 if XNN_UNPREDICTABLE(i0 != zero) {
362 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
363 }
364 const int8_t* i1 = input[1];
365 assert(i1 != NULL);
366 if XNN_UNPREDICTABLE(i1 != zero) {
367 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
368 }
369 const int8_t* i2 = input[2];
370 assert(i2 != NULL);
371 if XNN_UNPREDICTABLE(i2 != zero) {
372 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
373 }
374 const int8_t* i3 = input[3];
375 assert(i3 != NULL);
376 if XNN_UNPREDICTABLE(i3 != zero) {
377 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
378 }
379 const int8_t* i4 = input[4];
380 assert(i4 != NULL);
381 if XNN_UNPREDICTABLE(i4 != zero) {
382 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
383 }
384 const int8_t* i5 = input[5];
385 assert(i5 != NULL);
386 if XNN_UNPREDICTABLE(i5 != zero) {
387 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
388 }
389 const int8_t* i6 = input[6];
390 assert(i6 != NULL);
391 if XNN_UNPREDICTABLE(i6 != zero) {
392 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
393 }
394
395 size_t c = channels;
396
397 for (; c >= 8; c -= 8) {
398 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
399 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
400 b += 8;
401
402
403 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
404 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 0 * sizeof(int8_t)))));
405 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
406 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 4 * sizeof(int8_t)))));
407 i0 += 8;
408
409 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
410 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
411
412 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
413 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
414 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
415 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
416 i1 += 8;
417
418 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
419 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
420
421 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
422 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
423 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
424 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
425 i2 += 8;
426
427 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
428 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
429
430 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
431 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int8_t)))));
432 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
433 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 28 * sizeof(int8_t)))));
434 i3 += 8;
435
436 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
437 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
438
439 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
440 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 32 * sizeof(int8_t)))));
441 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
442 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 36 * sizeof(int8_t)))));
443 i4 += 8;
444
445 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
446 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
447
448 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
449 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 40 * sizeof(int8_t)))));
450 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
451 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 44 * sizeof(int8_t)))));
452 i5 += 8;
453
454 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
455 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
456
457 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
458 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 48 * sizeof(int8_t)))));
459 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
460 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 52 * sizeof(int8_t)))));
461 i6 += 8;
462
463 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
464 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
465
466 w = (const void*) ((uintptr_t) w + 56 * sizeof(int8_t));
467
468 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
469 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
470
471 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
472 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
473
474 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
475 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
476
477 vacc0123 = _mm_cvtps_epi32(vscaled0123);
478 vacc4567 = _mm_cvtps_epi32(vscaled4567);
479
480 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
481
482 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
483 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
484
485 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
486 output += 8;
487 }
488
489 if XNN_UNLIKELY(c != 0) {
490 do {
491 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
492 b += 4;
493
494 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
495 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 0 * sizeof(int8_t)))));
496 i0 += 4;
497
498 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
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 + 4 * sizeof(int8_t)))));
502 i1 += 4;
503
504 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
505
506 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
507 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int8_t)))));
508 i2 += 4;
509
510 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
511
512 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
513 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 12 * sizeof(int8_t)))));
514 i3 += 4;
515
516 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
517
518 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
519 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int8_t)))));
520 i4 += 4;
521
522 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
523
524 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
525 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 20 * sizeof(int8_t)))));
526 i5 += 4;
527
528 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
529
530 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
531 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int8_t)))));
532 i6 += 4;
533
534 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
535
536 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
537 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
538 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
539 vacc0123 = _mm_cvtps_epi32(vscaled0123);
540
541 w = (void*) ((uintptr_t) w + 28 * sizeof(int8_t));
542
543 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
544
545 vout0123 = _mm_packs_epi16(vout0123, vout0123);
546 vout0123 = _mm_max_epi8(vout0123, voutput_min);
547
548 if XNN_LIKELY(c >= 4) {
549 _mm_storeu_si32(output, vout0123);
550 output += 4;
551 c -= 4;
552 } else {
553 if (c & 2) {
554 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
555 vout0123 = _mm_srli_epi32(vout0123, 16);
556 output += 2;
557 }
558 if (c & 1) {
559 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
560 output += 1;
561 }
562 c = 0;
563 }
564 } while (c != 0);
565 }
566 }
567
568 input = (const int8_t**) ((uintptr_t) input + input_stride);
569 output = (int8_t*) ((uintptr_t) output + output_increment);
570 } while (--output_width != 0);
571 }