Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l8c8s8r-minmax-fp32-sse2-mul16-add16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <emmintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_dwconv_minmax_fp32_ukernel_5f5m5l8c8s8r__sse2_mul16_add16(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 intptr_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 size_t kernel_size,
30 int32_t* buffer,
31 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(channels != 0);
34 assert(output_width != 0);
35 assert(kernel_size > 5);
36
37
38 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
39 XNN_FORCE_REALIZATION(vscale);
40
41 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
42 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
43 const __m128i voutput_min = _mm_set1_epi16(params->fp32_scalar.output_min);
44 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
45 XNN_FORCE_REALIZATION(voutput_zero_point);
46 XNN_FORCE_REALIZATION(voutput_min);
47
48 do {
49 const void* w = weights;
50
51 // First pass to process 5 inputs.
52 {
53 int32_t* b = buffer;
54 const int8_t* i0 = input[0];
55 assert(i0 != NULL);
56 if XNN_UNPREDICTABLE(i0 != zero) {
57 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
58 }
59 const int8_t* i1 = input[1];
60 assert(i1 != NULL);
61 if XNN_UNPREDICTABLE(i1 != zero) {
62 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
63 }
64 const int8_t* i2 = input[2];
65 assert(i2 != NULL);
66 if XNN_UNPREDICTABLE(i2 != zero) {
67 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
68 }
69 const int8_t* i3 = input[3];
70 assert(i3 != NULL);
71 if XNN_UNPREDICTABLE(i3 != zero) {
72 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
73 }
74 const int8_t* i4 = input[4];
75 assert(i4 != NULL);
76 if XNN_UNPREDICTABLE(i4 != zero) {
77 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
78 }
79 input += 5;
80
81 size_t c = round_up_po2(channels, 8);
82
83 for (; c >= 8; c -= 8) {
84 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
85 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
86
87
88
89 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
90 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
91 i0 += 8;
92
93 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
94 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
95
96 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
97
98
99 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
100 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
101 i1 += 8;
102
103 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
104 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
105
106 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
107
108 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
109 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
110 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
111
112 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
113 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
114 i2 += 8;
115
116 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
117 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
118
119 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
120
121
122 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
123 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
124 i3 += 8;
125
126 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
127 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
128
129 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
130
131 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
132 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
133 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
134
135 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
136 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
137 i4 += 8;
138
139 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
140 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
141
142 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
143
144 const __m128i vsignprod4x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
145 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod4x01234567));
146 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod4x01234567));
147
148 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * 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 assert(c == 0);
156 }
157
158 // Middle pass to process 5 inputs in each iteration.
159 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
160 int32_t* b = buffer;
161 const int8_t* i0 = input[0];
162 assert(i0 != NULL);
163 if XNN_UNPREDICTABLE(i0 != zero) {
164 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
165 }
166 const int8_t* i1 = input[1];
167 assert(i1 != NULL);
168 if XNN_UNPREDICTABLE(i1 != zero) {
169 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
170 }
171 const int8_t* i2 = input[2];
172 assert(i2 != NULL);
173 if XNN_UNPREDICTABLE(i2 != zero) {
174 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
175 }
176 const int8_t* i3 = input[3];
177 assert(i3 != NULL);
178 if XNN_UNPREDICTABLE(i3 != zero) {
179 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
180 }
181 const int8_t* i4 = input[4];
182 assert(i4 != NULL);
183 if XNN_UNPREDICTABLE(i4 != zero) {
184 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
185 }
186 input += 5;
187
188 size_t c = round_up_po2(channels, 8);
189
190 for (; c >= 8; c -= 8) {
191 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
192 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
193
194
195
196 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
197 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
198 i0 += 8;
199
200 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
201 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
202
203 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
204
205
206 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
207 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
208 i1 += 8;
209
210 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
211 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
212
213 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
214
215 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
216 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
217 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
218
219 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
220 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
221 i2 += 8;
222
223 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
224 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
225
226 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
227
228
229 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
230 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
231 i3 += 8;
232
233 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
234 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
235
236 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
237
238 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
239 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
240 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
241
242 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
243 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
244 i4 += 8;
245
246 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
247 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
248
249 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
250
251 const __m128i vsignprod4x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
252 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod4x01234567));
253 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod4x01234567));
254
255 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
256
257 _mm_storeu_si128((__m128i*) (b), vacc0123);
258 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
259 b += 8;
260 }
261
262 assert(c == 0);
263 }
264
265 // Last pass to process up to 5 inputs.
266 {
267 const int32_t* b = buffer;
268 const int8_t* i0 = input[0];
269 assert(i0 != NULL);
270 if XNN_UNPREDICTABLE(i0 != zero) {
271 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
272 }
273 const int8_t* i1 = input[1];
274 assert(i1 != NULL);
275 if XNN_UNPREDICTABLE(i1 != zero) {
276 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
277 }
278 const int8_t* i2 = input[2];
279 assert(i2 != NULL);
280 if XNN_UNPREDICTABLE(i2 != zero) {
281 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
282 }
283 const int8_t* i3 = input[3];
284 assert(i3 != NULL);
285 if XNN_UNPREDICTABLE(i3 != zero) {
286 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
287 }
288 const int8_t* i4 = input[4];
289 assert(i4 != NULL);
290 if XNN_UNPREDICTABLE(i4 != zero) {
291 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
292 }
293
294 size_t c = channels;
295
296 for (; c >= 8; c -= 8) {
297 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
298 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
299 b += 8;
300
301
302 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
303 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
304 i0 += 8;
305
306 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
307 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
308
309 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
310
311
312 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
313 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
314 i1 += 8;
315
316 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
317 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
318
319 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
320
321 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
322 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
323 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
324
325 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
326 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
327 i2 += 8;
328
329 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
330 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
331
332 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
333
334
335 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
336 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
337 i3 += 8;
338
339 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
340 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
341
342 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
343
344 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
345 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
346 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
347
348 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
349 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
350 i4 += 8;
351
352 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
353 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
354
355 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
356
357 const __m128i vsignprod4x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
358 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod4x01234567));
359 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod4x01234567));
360
361 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
362
363 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
364 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
365
366 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
367 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
368
369 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
370 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
371
372 vacc0123 = _mm_cvtps_epi32(vscaled0123);
373 vacc4567 = _mm_cvtps_epi32(vscaled4567);
374
375 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
376
377 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
378
379 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
380
381
382 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
383 output += 8;
384 }
385
386 if XNN_UNLIKELY(c != 0) {
387 {
388 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
389 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
390 b += 8;
391
392
393 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
394 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
395
396 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
397 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
398
399 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
400
401
402 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
403 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
404
405 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
406 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
407
408 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
409
410 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
411 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
412 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
413
414 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
415 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
416
417 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
418 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
419
420 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
421
422
423 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
424 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
425
426 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
427 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
428
429 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
430
431 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
432 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
433 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
434
435 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
436 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
437
438 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
439 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
440
441 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
442
443 const __m128i vsignprod4x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
444 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod4x01234567));
445 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod4x01234567));
446
447 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
448 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
449
450 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
451 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
452
453
454 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
455 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
456
457 vacc0123 = _mm_cvtps_epi32(vscaled0123);
458 vacc4567 = _mm_cvtps_epi32(vscaled4567);
459
460 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
461
462 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
463
464 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
465
466
467 if (c & 4) {
468 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
469 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
470 output += 4;
471 }
472 if (c & 2) {
473 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
474 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
475 output += 2;
476 }
477 if (c & 1) {
478 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
479 output += 1;
480 }
481 }
482 }
483 }
484
485 input = (const int8_t**) ((uintptr_t) input + input_stride);
486 output = (int8_t*) ((uintptr_t) output + output_increment);
487 } while (--output_width != 0);
488 }