1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/multipass-sse2.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 <emmintrin.h>
13
14 #include <xnnpack/gavgpool.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16(size_t rows,size_t channels,const uint8_t * input,size_t input_stride,const uint8_t * zero,int32_t * buffer,uint8_t * output,const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16(
20 size_t rows,
21 size_t channels,
22 const uint8_t* input,
23 size_t input_stride,
24 const uint8_t* zero,
25 int32_t* buffer,
26 uint8_t* output,
27 const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(rows > 7);
30 assert(channels != 0);
31
32 const uint8_t* i0 = input;
33 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
34 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
35 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
36 const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
37 const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
38 const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
39 const size_t input_increment = 7 * input_stride - round_up_po2(channels, 16) * sizeof(uint8_t);
40
41 const __m128i vinit_bias = _mm_load_si128((const __m128i*) params->fp32_sse2.init_bias);
42 const __m128i vzero = _mm_setzero_si128();
43 int32_t* b = buffer;
44 size_t c = channels;
45 for (; c != 0; c = doz(c, 16)) {
46
47 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
48 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
49 i0 += 16;
50
51 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
52 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
53 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
54 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
55 i1 += 16;
56
57 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
58 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
59 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
60 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
61 i2 += 16;
62
63 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
64 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
65 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
66 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
67 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
68 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
69 i3 += 16;
70
71 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
72 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
73 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
74 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
75 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
76 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
77 i4 += 16;
78
79 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
80 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
81 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
82 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
83 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
84 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
85 i5 += 16;
86
87 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
88 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
89 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
90 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
91 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
92 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
93 i6 += 16;
94
95 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
96 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
97 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
98 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
99
100 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
101 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
102
103 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
104 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
105 __m128i vacc89AB = _mm_unpacklo_epi16(vacc89ABCDEF, vzero);
106 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
107
108 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
109 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
110 vacc89AB = _mm_add_epi32(vacc89AB, vinit_bias);
111 vaccCDEF = _mm_add_epi32(vaccCDEF, vinit_bias);
112
113 _mm_store_si128((__m128i*) b, vacc0123);
114 _mm_store_si128((__m128i*) (b + 4), vacc4567);
115 _mm_store_si128((__m128i*) (b + 8), vacc89AB);
116 _mm_store_si128((__m128i*) (b + 12), vaccCDEF);
117 b += 16;
118 }
119
120 for (rows -= 7; rows > 7; rows -= 7) {
121 i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment);
122 i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment);
123 i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment);
124 i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment);
125 i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment);
126 i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment);
127 i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment);
128
129 int32_t* b = buffer;
130 size_t c = channels;
131 for (; c != 0; c = doz(c, 16)) {
132
133 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
134 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
135 i0 += 16;
136
137 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
138 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
139 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
140 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
141 i1 += 16;
142
143 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
144 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
145 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
146 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
147 i2 += 16;
148
149 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
150 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
151 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
152 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
153 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
154 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
155 i3 += 16;
156
157 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
158 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
159 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
160 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
161 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
162 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
163 i4 += 16;
164
165 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
166 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
167 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
168 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
169 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
170 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
171 i5 += 16;
172
173 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
174 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
175 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
176 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
177 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
178 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
179 i6 += 16;
180
181 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
182 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
183 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
184 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
185
186 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
187 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
188
189 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
190 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
191 __m128i vacc89AB = _mm_unpacklo_epi16(vacc89ABCDEF, vzero);
192 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
193
194 vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) b));
195 vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (b + 4)));
196 vacc89AB = _mm_add_epi32(vacc89AB, _mm_load_si128((const __m128i*) (b + 8)));
197 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_load_si128((const __m128i*) (b + 12)));
198
199 _mm_store_si128((__m128i*) b, vacc0123);
200 _mm_store_si128((__m128i*) (b + 4), vacc4567);
201 _mm_store_si128((__m128i*) (b + 8), vacc89AB);
202 _mm_store_si128((__m128i*) (b + 12), vaccCDEF);
203 b += 16;
204 }
205 }
206
207 i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment);
208 i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment);
209 if XNN_UNPREDICTABLE(rows < 2) {
210 i1 = zero;
211 }
212 i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment);
213 if XNN_UNPREDICTABLE(rows <= 2) {
214 i2 = zero;
215 }
216 i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment);
217 if XNN_UNPREDICTABLE(rows < 4) {
218 i3 = zero;
219 }
220 i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment);
221 if XNN_UNPREDICTABLE(rows <= 4) {
222 i4 = zero;
223 }
224 i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment);
225 if XNN_UNPREDICTABLE(rows < 6) {
226 i5 = zero;
227 }
228 i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment);
229 if XNN_UNPREDICTABLE(rows <= 6) {
230 i6 = zero;
231 }
232
233 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
234 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
235 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
236 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
237 for (; channels >= 16; channels -= 16) {
238
239 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
240 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
241 i0 += 16;
242
243 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
244 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
245 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
246 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
247 i1 += 16;
248
249 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
250 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
251 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
252 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
253 i2 += 16;
254
255 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
256 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
257 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
258 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
259 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
260 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
261 i3 += 16;
262
263 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
264 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
265 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
266 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
267 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
268 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
269 i4 += 16;
270
271 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
272 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
273 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
274 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
275 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
276 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
277 i5 += 16;
278
279 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
280 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
281 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
282 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
283 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
284 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
285 i6 += 16;
286
287 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
288 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
289 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
290 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
291
292 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
293 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
294
295 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
296 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
297 __m128i vacc89AB = _mm_unpacklo_epi16(vacc89ABCDEF, vzero);
298 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
299
300 vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) buffer));
301 vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (buffer + 4)));
302 vacc89AB = _mm_add_epi32(vacc89AB, _mm_load_si128((const __m128i*) (buffer + 8)));
303 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_load_si128((const __m128i*) (buffer + 12)));
304 buffer += 16;
305
306 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
307 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
308 __m128 vfpacc89AB = _mm_cvtepi32_ps(vacc89AB);
309 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vaccCDEF);
310
311 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
312 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
313 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
314 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
315
316 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
317 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
318 vfpacc89AB = _mm_min_ps(vfpacc89AB, voutput_max_less_zero_point);
319 vfpaccCDEF = _mm_min_ps(vfpaccCDEF, voutput_max_less_zero_point);
320
321 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
322 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
323 vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
324 vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
325
326 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
327 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
328
329
330 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
331
332 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
333
334 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
335 output += 16;
336 }
337 if XNN_UNLIKELY(channels != 0) {
338 do {
339
340 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
341 i0 += 8;
342
343 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
344 i1 += 8;
345
346 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
347 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
348 i2 += 8;
349
350 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
351 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
352 i3 += 8;
353
354 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
355 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
356 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
357 i4 += 8;
358
359 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
360 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
361 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
362 i5 += 8;
363
364 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
365 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
366 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
367 i6 += 8;
368
369 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
370 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
371
372 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
373 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
374
375 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
376
377 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
378 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
379
380 vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) buffer));
381 vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (buffer + 4)));
382 buffer += 8;
383
384 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
385 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
386
387 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
388 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
389
390 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
391 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
392
393 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
394 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
395
396 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
397
398 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
399 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
400
401 if XNN_LIKELY(channels >= 8) {
402 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
403 output += 8;
404 channels -= 8;
405 } else {
406 if (channels & 4) {
407 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
408 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
409 output += 4;
410 }
411 uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
412 if (channels & 2) {
413 unaligned_store_u16(output, (uint16_t) vout0123);
414 vout0123 >>= 16;
415 output += 2;
416 }
417 if (channels & 1) {
418 *output = (uint8_t) vout0123;
419 output += 1;
420 }
421 channels = 0;
422 }
423 } while (channels != 0);
424 }
425 }
426