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_c24(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_c24(
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, 8) * 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 >= 24; c -= 24) {
46
47 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
48 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
49 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
50 i0 += 24;
51
52 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
53 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
54 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
55 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
56 const __m128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, vzero);
57 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
58 i1 += 24;
59
60 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
61 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
62 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
63 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
64 const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, vzero);
65 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
66 i2 += 24;
67
68 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
69 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
70 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
71 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
72 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
73 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
74 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN);
75 const __m128i vxi2xGHIJKLMN = _mm_unpacklo_epi8(vi2xGHIJKLMN, vzero);
76 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
77 i3 += 24;
78
79 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
80 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
81 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
82 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
83 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
84 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
85 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi2xGHIJKLMN);
86 const __m128i vxi3xGHIJKLMN = _mm_unpacklo_epi8(vi3xGHIJKLMN, vzero);
87 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
88 i4 += 24;
89
90 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
91 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
92 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
93 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
94 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
95 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
96 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi3xGHIJKLMN);
97 const __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, vzero);
98 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
99 i5 += 24;
100
101 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
102 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
103 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
104 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
105 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
106 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
107 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi4xGHIJKLMN);
108 const __m128i vxi5xGHIJKLMN = _mm_unpacklo_epi8(vi5xGHIJKLMN, vzero);
109 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
110 i6 += 24;
111
112 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
113 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
114 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
115 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
116 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi5xGHIJKLMN);
117 const __m128i vxi6xGHIJKLMN = _mm_unpacklo_epi8(vi6xGHIJKLMN, vzero);
118
119 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
120 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
121 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi6xGHIJKLMN);
122
123 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
124 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
125 __m128i vacc89AB = _mm_unpacklo_epi16(vacc89ABCDEF, vzero);
126 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
127 __m128i vaccGHIJ = _mm_unpacklo_epi16(vaccGHIJKLMN, vzero);
128 __m128i vaccKLMN = _mm_unpackhi_epi16(vaccGHIJKLMN, vzero);
129
130 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
131 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
132 vacc89AB = _mm_add_epi32(vacc89AB, vinit_bias);
133 vaccCDEF = _mm_add_epi32(vaccCDEF, vinit_bias);
134 vaccGHIJ = _mm_add_epi32(vaccGHIJ, vinit_bias);
135 vaccKLMN = _mm_add_epi32(vaccKLMN, vinit_bias);
136
137 _mm_store_si128((__m128i*) b, vacc0123);
138 _mm_store_si128((__m128i*) (b + 4), vacc4567);
139 _mm_store_si128((__m128i*) (b + 8), vacc89AB);
140 _mm_store_si128((__m128i*) (b + 12), vaccCDEF);
141 _mm_store_si128((__m128i*) (b + 16), vaccGHIJ);
142 _mm_store_si128((__m128i*) (b + 20), vaccKLMN);
143 b += 24;
144 }
145 if XNN_UNLIKELY(c != 0) {
146 do {
147
148 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
149 i0 += 8;
150
151 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
152 i1 += 8;
153
154 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
155 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
156 i2 += 8;
157
158 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
159 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
160 i3 += 8;
161
162 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
163 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
164 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
165 i4 += 8;
166
167 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
168 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
169 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
170 i5 += 8;
171
172 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
173 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
174 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
175 i6 += 8;
176
177 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
178 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
179
180 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
181 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
182
183 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
184
185 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
186 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
187
188 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
189 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
190
191 _mm_store_si128((__m128i*) b, vacc0123);
192 _mm_store_si128((__m128i*) (b + 4), vacc4567);
193 b += 8;
194
195 c = doz(c, 8);
196 } while (c != 0);
197 }
198
199 for (rows -= 7; rows > 7; rows -= 7) {
200 i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment);
201 i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment);
202 i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment);
203 i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment);
204 i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment);
205 i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment);
206 i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment);
207
208 int32_t* b = buffer;
209 size_t c = channels;
210 for (; c >= 24; c -= 24) {
211
212 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
213 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
214 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
215 i0 += 24;
216
217 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
218 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
219 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
220 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
221 const __m128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, vzero);
222 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
223 i1 += 24;
224
225 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
226 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
227 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
228 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
229 const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, vzero);
230 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
231 i2 += 24;
232
233 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
234 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
235 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
236 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
237 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
238 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
239 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN);
240 const __m128i vxi2xGHIJKLMN = _mm_unpacklo_epi8(vi2xGHIJKLMN, vzero);
241 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
242 i3 += 24;
243
244 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
245 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
246 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
247 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
248 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
249 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
250 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi2xGHIJKLMN);
251 const __m128i vxi3xGHIJKLMN = _mm_unpacklo_epi8(vi3xGHIJKLMN, vzero);
252 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
253 i4 += 24;
254
255 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
256 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
257 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
258 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
259 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
260 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
261 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi3xGHIJKLMN);
262 const __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, vzero);
263 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
264 i5 += 24;
265
266 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
267 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
268 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
269 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
270 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
271 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
272 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi4xGHIJKLMN);
273 const __m128i vxi5xGHIJKLMN = _mm_unpacklo_epi8(vi5xGHIJKLMN, vzero);
274 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
275 i6 += 24;
276
277 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
278 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
279 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
280 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
281 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi5xGHIJKLMN);
282 const __m128i vxi6xGHIJKLMN = _mm_unpacklo_epi8(vi6xGHIJKLMN, vzero);
283
284 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
285 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
286 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi6xGHIJKLMN);
287
288 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
289 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
290 __m128i vacc89AB = _mm_unpacklo_epi16(vacc89ABCDEF, vzero);
291 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
292 __m128i vaccGHIJ = _mm_unpacklo_epi16(vaccGHIJKLMN, vzero);
293 __m128i vaccKLMN = _mm_unpackhi_epi16(vaccGHIJKLMN, vzero);
294
295 vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) b));
296 vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (b + 4)));
297 vacc89AB = _mm_add_epi32(vacc89AB, _mm_load_si128((const __m128i*) (b + 8)));
298 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_load_si128((const __m128i*) (b + 12)));
299 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_load_si128((const __m128i*) (b + 16)));
300 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_load_si128((const __m128i*) (b + 20)));
301
302 _mm_store_si128((__m128i*) b, vacc0123);
303 _mm_store_si128((__m128i*) (b + 4), vacc4567);
304 _mm_store_si128((__m128i*) (b + 8), vacc89AB);
305 _mm_store_si128((__m128i*) (b + 12), vaccCDEF);
306 _mm_store_si128((__m128i*) (b + 16), vaccGHIJ);
307 _mm_store_si128((__m128i*) (b + 20), vaccKLMN);
308 b += 24;
309 }
310 if XNN_UNLIKELY(c != 0) {
311 do {
312
313 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
314 i0 += 8;
315
316 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
317 i1 += 8;
318
319 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
320 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
321 i2 += 8;
322
323 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
324 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
325 i3 += 8;
326
327 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
328 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
329 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
330 i4 += 8;
331
332 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
333 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
334 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
335 i5 += 8;
336
337 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
338 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
339 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
340 i6 += 8;
341
342 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
343 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
344
345 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
346 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
347
348 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
349
350 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
351 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
352
353 vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) b));
354 vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (b + 4)));
355
356 _mm_store_si128((__m128i*) b, vacc0123);
357 _mm_store_si128((__m128i*) (b + 4), vacc4567);
358 b += 8;
359
360 c = doz(c, 8);
361 } while (c != 0);
362 }
363 }
364
365 i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment);
366 i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment);
367 if XNN_UNPREDICTABLE(rows < 2) {
368 i1 = zero;
369 }
370 i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment);
371 if XNN_UNPREDICTABLE(rows <= 2) {
372 i2 = zero;
373 }
374 i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment);
375 if XNN_UNPREDICTABLE(rows < 4) {
376 i3 = zero;
377 }
378 i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment);
379 if XNN_UNPREDICTABLE(rows <= 4) {
380 i4 = zero;
381 }
382 i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment);
383 if XNN_UNPREDICTABLE(rows < 6) {
384 i5 = zero;
385 }
386 i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment);
387 if XNN_UNPREDICTABLE(rows <= 6) {
388 i6 = zero;
389 }
390
391 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
392 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
393 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
394 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
395 for (; channels >= 24; channels -= 24) {
396
397 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
398 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
399 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
400 i0 += 24;
401
402 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
403 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
404 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
405 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
406 const __m128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, vzero);
407 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
408 i1 += 24;
409
410 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
411 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
412 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
413 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
414 const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, vzero);
415 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
416 i2 += 24;
417
418 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
419 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
420 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
421 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
422 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
423 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
424 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN);
425 const __m128i vxi2xGHIJKLMN = _mm_unpacklo_epi8(vi2xGHIJKLMN, vzero);
426 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
427 i3 += 24;
428
429 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
430 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
431 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
432 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
433 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
434 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
435 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi2xGHIJKLMN);
436 const __m128i vxi3xGHIJKLMN = _mm_unpacklo_epi8(vi3xGHIJKLMN, vzero);
437 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
438 i4 += 24;
439
440 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
441 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
442 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
443 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
444 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
445 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
446 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi3xGHIJKLMN);
447 const __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, vzero);
448 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
449 i5 += 24;
450
451 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
452 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
453 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
454 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
455 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
456 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
457 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi4xGHIJKLMN);
458 const __m128i vxi5xGHIJKLMN = _mm_unpacklo_epi8(vi5xGHIJKLMN, vzero);
459 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
460 i6 += 24;
461
462 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
463 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
464 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
465 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
466 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi5xGHIJKLMN);
467 const __m128i vxi6xGHIJKLMN = _mm_unpacklo_epi8(vi6xGHIJKLMN, vzero);
468
469 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
470 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
471 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi6xGHIJKLMN);
472
473 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
474 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
475 __m128i vacc89AB = _mm_unpacklo_epi16(vacc89ABCDEF, vzero);
476 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
477 __m128i vaccGHIJ = _mm_unpacklo_epi16(vaccGHIJKLMN, vzero);
478 __m128i vaccKLMN = _mm_unpackhi_epi16(vaccGHIJKLMN, vzero);
479
480 vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) buffer));
481 vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (buffer + 4)));
482 vacc89AB = _mm_add_epi32(vacc89AB, _mm_load_si128((const __m128i*) (buffer + 8)));
483 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_load_si128((const __m128i*) (buffer + 12)));
484 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_load_si128((const __m128i*) (buffer + 16)));
485 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_load_si128((const __m128i*) (buffer + 20)));
486 buffer += 24;
487
488 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
489 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
490 __m128 vfpacc89AB = _mm_cvtepi32_ps(vacc89AB);
491 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vaccCDEF);
492 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
493 __m128 vfpaccKLMN = _mm_cvtepi32_ps(vaccKLMN);
494
495 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
496 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
497 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
498 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
499 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale);
500 vfpaccKLMN = _mm_mul_ps(vfpaccKLMN, vscale);
501
502 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
503 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
504 vfpacc89AB = _mm_min_ps(vfpacc89AB, voutput_max_less_zero_point);
505 vfpaccCDEF = _mm_min_ps(vfpaccCDEF, voutput_max_less_zero_point);
506 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point);
507 vfpaccKLMN = _mm_min_ps(vfpaccKLMN, voutput_max_less_zero_point);
508
509 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
510 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
511 vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
512 vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
513 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ);
514 vaccKLMN = _mm_cvtps_epi32(vfpaccKLMN);
515
516 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
517 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
518 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
519
520
521 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
522 __m128i voutGHIJKLMNGHIJKLMN = _mm_packus_epi16(voutGHIJKLMN, voutGHIJKLMN);
523
524 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
525 voutGHIJKLMNGHIJKLMN = _mm_max_epu8(voutGHIJKLMNGHIJKLMN, voutput_min);
526
527 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
528 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
529 output += 24;
530 }
531 if XNN_UNLIKELY(channels != 0) {
532 do {
533
534 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
535 i0 += 8;
536
537 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
538 i1 += 8;
539
540 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
541 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
542 i2 += 8;
543
544 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
545 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
546 i3 += 8;
547
548 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
549 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
550 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
551 i4 += 8;
552
553 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
554 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
555 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
556 i5 += 8;
557
558 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
559 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
560 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
561 i6 += 8;
562
563 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
564 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
565
566 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
567 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
568
569 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
570
571 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
572 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
573
574 vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) buffer));
575 vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (buffer + 4)));
576 buffer += 8;
577
578 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
579 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
580
581 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
582 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
583
584 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
585 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
586
587 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
588 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
589
590 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
591
592 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
593 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
594
595 if XNN_LIKELY(channels >= 8) {
596 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
597 output += 8;
598 channels -= 8;
599 } else {
600 if (channels & 4) {
601 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
602 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
603 output += 4;
604 }
605 uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
606 if (channels & 2) {
607 unaligned_store_u16(output, (uint16_t) vout0123);
608 vout0123 >>= 16;
609 output += 2;
610 }
611 if (channels & 1) {
612 *output = (uint8_t) vout0123;
613 output += 1;
614 }
615 channels = 0;
616 }
617 } while (channels != 0);
618 }
619 }
620