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