Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p16c-minmax-fp32-sse41-mul16-add16.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_add16(
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
115 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
116 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
117 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
118 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
119 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
120 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
121 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
122 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
123 i1 += 16;
124
125
126 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
127 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
128
129 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
130 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
131 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
132 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
133
134 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
135 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
136 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
137 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
138 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
139 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
140 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
141 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
142 i2 += 16;
143
144
145 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
146 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
147
148
149 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
150 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
151 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
152 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
153 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
154 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
155 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
156 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
157 i3 += 16;
158
159
160 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
161 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
162
163 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
164 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
165 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
166 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
167
168 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
169 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
170 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
171 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
172 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
173 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
174 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
175 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
176 i4 += 16;
177
178
179 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
180 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
181
182
183 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
184 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
185 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
186 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
187 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
188 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
189 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
190 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
191 i5 += 16;
192
193
194 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
195 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
196
197 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
198 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
199 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
200 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
201
202 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
203 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
204 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
205 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
206 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
207 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
208 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
209 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
210 i6 += 16;
211
212
213 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
214 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
215
216
217 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
218 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
219 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
220 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
221 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
222 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
223 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
224 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
225 i7 += 16;
226
227
228 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
229 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
230
231 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
232 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
233 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
234 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
235
236 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
237 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
238 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
239 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
240 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
241 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
242 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
243 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
244 i8 += 16;
245
246
247 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
248 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
249
250 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
251 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
252 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
253 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
254
255 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
256
257 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
258 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
259 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
260 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
261
262 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
263 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
264 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
265 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
266
267 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
268 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
269 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
270 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
271
272 vacc0123 = _mm_cvtps_epi32(vscaled0123);
273 vacc4567 = _mm_cvtps_epi32(vscaled4567);
274 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
275 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
276
277 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
278 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
279
280
281 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
282
283 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
284
285 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
286 output += 16;
287 }
288 if XNN_UNLIKELY(c != 0) {
289 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
290 do {
291 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
292 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
293
294
295 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
296 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
297 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
298 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
299 i0 += 8;
300
301
302 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
303
304
305 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
306 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
307 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
308 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
309 i1 += 8;
310
311
312 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
313
314 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
315 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
316
317 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
318 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
319 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
320 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
321 i2 += 8;
322
323
324 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
325
326
327 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
328 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
329 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
330 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
331 i3 += 8;
332
333
334 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
335
336 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
337 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
338
339 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
340 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
341 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
342 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
343 i4 += 8;
344
345
346 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
347
348
349 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
350 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
351 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
352 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
353 i5 += 8;
354
355
356 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
357
358 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
359 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
360
361 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
362 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
363 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
364 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
365 i6 += 8;
366
367
368 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
369
370
371 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
372 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
373 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
374 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
375 i7 += 8;
376
377
378 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
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 vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
384 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
385 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
386 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
387 i8 += 8;
388
389
390 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
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 k += 8;
396
397 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
398 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
399
400 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
401 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
402
403 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
404 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
405
406 vacc0123 = _mm_cvtps_epi32(vscaled0123);
407 vacc4567 = _mm_cvtps_epi32(vscaled4567);
408
409 w = (const void*) ((const int32_t*) w + 8);
410
411 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
412
413
414 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
415
416 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
417
418 if XNN_LIKELY(c >= 8) {
419 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
420 output += 8;
421 c -= 8;
422 } else {
423 if (c & 4) {
424 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
425 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
426 output += 4;
427 }
428 if (c & 2) {
429 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
430 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
431 output += 2;
432 }
433 if (c & 1) {
434 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
435 output += 1;
436 }
437 c = 0;
438 }
439 } while (c != 0);
440 }
441
442 output = (int8_t*) ((uintptr_t) output + output_increment);
443 } while (--output_width != 0);
444 }