Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p16c-minmax-fp32-sse41-mul16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <smmintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/unaligned.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_9p16c__sse41_mul16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 intptr_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
34 XNN_FORCE_REALIZATION(vscale);
35
36 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);
37 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
38 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
39 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
40 XNN_FORCE_REALIZATION(voutput_zero_point);
41 XNN_FORCE_REALIZATION(voutput_min);
42
43 do {
44 const int8_t* i0 = input[0];
45 assert(i0 != NULL);
46 if XNN_UNPREDICTABLE(i0 != zero) {
47 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
48 }
49 const int8_t* i1 = input[1];
50 assert(i1 != NULL);
51 if XNN_UNPREDICTABLE(i1 != zero) {
52 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
53 }
54 const int8_t* i2 = input[2];
55 assert(i2 != NULL);
56 if XNN_UNPREDICTABLE(i2 != zero) {
57 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
58 }
59 const int8_t* i3 = input[3];
60 assert(i3 != NULL);
61 if XNN_UNPREDICTABLE(i3 != zero) {
62 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
63 }
64 const int8_t* i4 = input[4];
65 assert(i4 != NULL);
66 if XNN_UNPREDICTABLE(i4 != zero) {
67 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
68 }
69 const int8_t* i5 = input[5];
70 assert(i5 != NULL);
71 if XNN_UNPREDICTABLE(i5 != zero) {
72 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
73 }
74 const int8_t* i6 = input[6];
75 assert(i6 != NULL);
76 if XNN_UNPREDICTABLE(i6 != zero) {
77 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
78 }
79 const int8_t* i7 = input[7];
80 assert(i7 != NULL);
81 if XNN_UNPREDICTABLE(i7 != zero) {
82 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
83 }
84 const int8_t* i8 = input[8];
85 assert(i8 != NULL);
86 if XNN_UNPREDICTABLE(i8 != zero) {
87 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
88 }
89 input = (const int8_t**) ((uintptr_t) input + input_stride);
90
91 size_t c = channels;
92 const void* w = weights;
93 for (; c >= 16; c -= 16) {
94 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
95 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
96 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
97 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
98
99
100 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
101 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
102 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
103 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
104 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
105 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
106 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
107 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
108 i0 += 16;
109
110
111 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
112 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
113
114 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
115 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
116 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
117 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
118
119 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
120 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
121 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
122 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
123 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
124 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
125 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
126 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
127 i1 += 16;
128
129
130 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
131 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
132
133 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
134 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
135 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
136 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
137
138 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
139 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
140 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
141 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
142 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
143 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
144 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
145 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
146 i2 += 16;
147
148
149 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
150 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
151
152 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
153 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
154 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
155 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
156
157 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
158 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
159 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
160 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
161 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
162 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
163 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
164 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
165 i3 += 16;
166
167
168 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
169 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
170
171 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
172 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
173 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
174 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
175
176 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
177 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
178 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
179 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
180 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
181 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
182 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
183 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
184 i4 += 16;
185
186
187 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
188 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
189
190 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
191 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
192 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
193 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
194
195 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
196 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
197 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
198 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
199 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
200 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
201 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
202 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
203 i5 += 16;
204
205
206 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
207 vprod89ABCDEF = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
208
209 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
210 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
211 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
212 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
213
214 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
215 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
216 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
217 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
218 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
219 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
220 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
221 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
222 i6 += 16;
223
224
225 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
226 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
227
228 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
229 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
230 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
231 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
232
233 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
234 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
235 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
236 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
237 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
238 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
239 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
240 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
241 i7 += 16;
242
243
244 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
245 vprod89ABCDEF = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
246
247 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
248 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
249 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
250 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
251
252 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
253 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
254 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
255 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
256 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
257 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
258 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
259 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
260 i8 += 16;
261
262
263 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
264 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
265
266 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
267 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
268 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
269 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
270
271 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
272
273 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
274 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
275 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
276 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
277
278 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
279 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
280 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
281 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
282
283 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
284 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
285 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
286 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
287
288 vacc0123 = _mm_cvtps_epi32(vscaled0123);
289 vacc4567 = _mm_cvtps_epi32(vscaled4567);
290 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
291 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
292
293 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
294 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
295
296
297 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
298
299 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
300
301 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
302 output += 16;
303 }
304 if XNN_UNLIKELY(c != 0) {
305 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
306 do {
307 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
308 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
309
310
311 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
312 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
313 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
314 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
315 i0 += 8;
316
317
318 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
319
320 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
321 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
322
323 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
324 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
325 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
326 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
327 i1 += 8;
328
329
330 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
331
332 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
333 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
334
335 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
336 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
337 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
338 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
339 i2 += 8;
340
341
342 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
343
344 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
345 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
346
347 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
348 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
349 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
350 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
351 i3 += 8;
352
353
354 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
355
356 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
357 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
358
359 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
360 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
361 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
362 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
363 i4 += 8;
364
365
366 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
367
368 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
369 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
370
371 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
372 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
373 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
374 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
375 i5 += 8;
376
377
378 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
379
380 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
381 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
382
383 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
384 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
385 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
386 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
387 i6 += 8;
388
389
390 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
391
392 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
393 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
394
395 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
396 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
397 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
398 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
399 i7 += 8;
400
401
402 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
403
404 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
405 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
406
407 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
408 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
409 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
410 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
411 i8 += 8;
412
413
414 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
415
416 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
417 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
418
419 k += 8;
420
421 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
422 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
423
424 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
425 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
426
427 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
428 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
429
430 vacc0123 = _mm_cvtps_epi32(vscaled0123);
431 vacc4567 = _mm_cvtps_epi32(vscaled4567);
432
433 w = (const void*) ((const int32_t*) w + 8);
434
435 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
436
437
438 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
439
440 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
441
442 if XNN_LIKELY(c >= 8) {
443 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
444 output += 8;
445 c -= 8;
446 } else {
447 if (c & 4) {
448 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
449 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
450 output += 4;
451 }
452 if (c & 2) {
453 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
454 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
455 output += 2;
456 }
457 if (c & 1) {
458 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
459 output += 1;
460 }
461 c = 0;
462 }
463 } while (c != 0);
464 }
465
466 output = (int8_t*) ((uintptr_t) output + output_increment);
467 } while (--output_width != 0);
468 }