1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <emmintrin.h>
9
10 #include <xnnpack/argmaxpool.h>
11
12
xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4(size_t output_pixels,size_t pooling_elements,size_t channels,const float ** input,size_t input_offset,float * accumulation_buffer,uint32_t * index_buffer,float * output,uint32_t * index,size_t input_increment,size_t output_increment)13 void xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4(
14 size_t output_pixels,
15 size_t pooling_elements,
16 size_t channels,
17 const float** input,
18 size_t input_offset,
19 float* accumulation_buffer,
20 uint32_t* index_buffer,
21 float* output,
22 uint32_t* index,
23 size_t input_increment,
24 size_t output_increment) XNN_DISABLE_TSAN
25 {
26 assert(output_pixels != 0);
27 assert(pooling_elements != 0);
28 assert(pooling_elements > 9);
29 assert(channels != 0);
30
31 do {
32 {
33 float* ab = accumulation_buffer;
34 uint32_t* ib = index_buffer;
35
36 const float* i0 = *input++;
37 const float* i1 = *input++;
38 const float* i2 = *input++;
39 const float* i3 = *input++;
40 const float* i4 = *input++;
41 const float* i5 = *input++;
42 const float* i6 = *input++;
43 const float* i7 = *input++;
44 const float* i8 = *input++;
45 i0 = (const float*) ((uintptr_t) i0 + input_offset);
46 i1 = (const float*) ((uintptr_t) i1 + input_offset);
47 i2 = (const float*) ((uintptr_t) i2 + input_offset);
48 i3 = (const float*) ((uintptr_t) i3 + input_offset);
49 i4 = (const float*) ((uintptr_t) i4 + input_offset);
50 i5 = (const float*) ((uintptr_t) i5 + input_offset);
51 i6 = (const float*) ((uintptr_t) i6 + input_offset);
52 i7 = (const float*) ((uintptr_t) i7 + input_offset);
53 i8 = (const float*) ((uintptr_t) i8 + input_offset);
54
55 for (size_t c = 0; c < channels; c += 4) {
56 const __m128 vi0 = _mm_loadu_ps(i0);
57 i0 += 4;
58 const __m128 vi1 = _mm_loadu_ps(i1);
59 i1 += 4;
60 const __m128 vi2 = _mm_loadu_ps(i2);
61 i2 += 4;
62 const __m128 vi3 = _mm_loadu_ps(i3);
63 i3 += 4;
64 const __m128 vi4 = _mm_loadu_ps(i4);
65 i4 += 4;
66 const __m128 vi5 = _mm_loadu_ps(i5);
67 i5 += 4;
68 const __m128 vi6 = _mm_loadu_ps(i6);
69 i6 += 4;
70 const __m128 vi7 = _mm_loadu_ps(i7);
71 i7 += 4;
72 const __m128 vi8 = _mm_loadu_ps(i8);
73 i8 += 4;
74
75 __m128 vmax = vi0;
76 __m128i vidx = _mm_setzero_si128();
77
78 const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax));
79 vmax = _mm_max_ps(vi1, vmax);
80 vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1)));
81
82 const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax));
83 vmax = _mm_max_ps(vi2, vmax);
84 vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2)));
85
86 const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax));
87 vmax = _mm_max_ps(vi3, vmax);
88 vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3)));
89
90 const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax));
91 vmax = _mm_max_ps(vi4, vmax);
92 vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, _mm_set1_epi32(4)));
93
94 const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax));
95 vmax = _mm_max_ps(vi5, vmax);
96 vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, _mm_set1_epi32(5)));
97
98 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax));
99 vmax = _mm_max_ps(vi6, vmax);
100 vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, _mm_set1_epi32(6)));
101
102 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax));
103 vmax = _mm_max_ps(vi7, vmax);
104 vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, _mm_set1_epi32(7)));
105
106 const __m128i vm8 = _mm_castps_si128(_mm_cmpgt_ps(vi8, vmax));
107 vmax = _mm_max_ps(vi8, vmax);
108 vidx = _mm_or_si128(_mm_andnot_si128(vm8, vidx), _mm_and_si128(vm8, _mm_set1_epi32(8)));
109
110 _mm_store_ps(ab, vmax);
111 ab += 4;
112 _mm_store_si128((__m128i*) ib, vidx);
113 ib += 4;
114 }
115 }
116 const __m128i v1 = _mm_set1_epi32(1);
117 const __m128i v8 = _mm_set1_epi32(8);
118 __m128i vidx0 = _mm_add_epi32(v1, v8);
119
120 size_t k = pooling_elements;
121 for (k -= 9; k > 8; k -= 8) {
122 const float* i0 = *input++;
123 const float* i1 = *input++;
124 const float* i2 = *input++;
125 const float* i3 = *input++;
126 const float* i4 = *input++;
127 const float* i5 = *input++;
128 const float* i6 = *input++;
129 const float* i7 = *input++;
130 i0 = (const float*) ((uintptr_t) i0 + input_offset);
131 i1 = (const float*) ((uintptr_t) i1 + input_offset);
132 i2 = (const float*) ((uintptr_t) i2 + input_offset);
133 i3 = (const float*) ((uintptr_t) i3 + input_offset);
134 i4 = (const float*) ((uintptr_t) i4 + input_offset);
135 i5 = (const float*) ((uintptr_t) i5 + input_offset);
136 i6 = (const float*) ((uintptr_t) i6 + input_offset);
137 i7 = (const float*) ((uintptr_t) i7 + input_offset);
138
139 float* ab = accumulation_buffer;
140 uint32_t* ib = index_buffer;
141
142 for (size_t c = 0; c < channels; c += 4) {
143 const __m128 vi0 = _mm_loadu_ps(i0);
144 i0 += 4;
145 const __m128 vi1 = _mm_loadu_ps(i1);
146 i1 += 4;
147 const __m128 vi2 = _mm_loadu_ps(i2);
148 i2 += 4;
149 const __m128 vi3 = _mm_loadu_ps(i3);
150 i3 += 4;
151 const __m128 vi4 = _mm_loadu_ps(i4);
152 i4 += 4;
153 const __m128 vi5 = _mm_loadu_ps(i5);
154 i5 += 4;
155 const __m128 vi6 = _mm_loadu_ps(i6);
156 i6 += 4;
157 const __m128 vi7 = _mm_loadu_ps(i7);
158 i7 += 4;
159
160 __m128 vmax = _mm_load_ps(ab);
161 __m128i vidx = _mm_load_si128((const __m128i*) ib);
162
163 const __m128i vm0 = _mm_castps_si128(_mm_cmpgt_ps(vi0, vmax));
164 vmax = _mm_max_ps(vi0, vmax);
165 vidx = _mm_or_si128(_mm_andnot_si128(vm0, vidx), _mm_and_si128(vm0, vidx0));
166
167 const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax));
168 const __m128i vidx1 = _mm_add_epi32(vidx0, v1);
169 vmax = _mm_max_ps(vi1, vmax);
170 vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, vidx1));
171
172 const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax));
173 const __m128i vidx2 = _mm_add_epi32(vidx1, v1);
174 vmax = _mm_max_ps(vi2, vmax);
175 vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, vidx2));
176
177 const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax));
178 const __m128i vidx3 = _mm_add_epi32(vidx2, v1);
179 vmax = _mm_max_ps(vi3, vmax);
180 vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, vidx3));
181
182 const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax));
183 const __m128i vidx4 = _mm_add_epi32(vidx3, v1);
184 vmax = _mm_max_ps(vi4, vmax);
185 vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, vidx4));
186
187 const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax));
188 const __m128i vidx5 = _mm_add_epi32(vidx4, v1);
189 vmax = _mm_max_ps(vi5, vmax);
190 vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, vidx5));
191
192 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax));
193 const __m128i vidx6 = _mm_add_epi32(vidx5, v1);
194 vmax = _mm_max_ps(vi6, vmax);
195 vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, vidx6));
196
197 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax));
198 const __m128i vidx7 = _mm_add_epi32(vidx6, v1);
199 vmax = _mm_max_ps(vi7, vmax);
200 vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, vidx7));
201
202 _mm_store_ps(ab, vmax);
203 ab += 4;
204 _mm_store_si128((__m128i*) ib, vidx);
205 ib += 4;
206 }
207 vidx0 = _mm_add_epi32(vidx0, v8);
208 }
209
210 float* o = output;
211 uint32_t* i = index;
212 {
213 const float* i0 = input[0];
214 const float* i1 = input[1];
215 const float* i2 = input[2];
216 const float* i3 = input[3];
217 const float* i4 = input[4];
218 const float* i5 = input[5];
219 const float* i6 = input[6];
220 const float* i7 = input[7];
221 i0 = (const float*) ((uintptr_t) i0 + input_offset);
222 i1 = (const float*) ((uintptr_t) i1 + input_offset);
223 i2 = (const float*) ((uintptr_t) i2 + input_offset);
224 i3 = (const float*) ((uintptr_t) i3 + input_offset);
225 i4 = (const float*) ((uintptr_t) i4 + input_offset);
226 i5 = (const float*) ((uintptr_t) i5 + input_offset);
227 i6 = (const float*) ((uintptr_t) i6 + input_offset);
228 i7 = (const float*) ((uintptr_t) i7 + input_offset);
229 input = (const float**) ((uintptr_t) input + input_increment);
230 if (k < 2) {
231 i1 = i0;
232 }
233 if (k <= 2) {
234 i2 = i0;
235 }
236 if (k < 4) {
237 i3 = i0;
238 }
239 if (k <= 4) {
240 i4 = i0;
241 }
242 if (k < 6) {
243 i5 = i0;
244 }
245 if (k <= 6) {
246 i6 = i0;
247 }
248 if (k != 8) {
249 i7 = i0;
250 }
251
252 size_t c = channels;
253 float* ab = accumulation_buffer;
254 uint32_t* ib = index_buffer;
255 for (; c >= 4; c -= 4) {
256 const __m128 vi0 = _mm_loadu_ps(i0);
257 i0 += 4;
258 const __m128 vi1 = _mm_loadu_ps(i1);
259 i1 += 4;
260 const __m128 vi2 = _mm_loadu_ps(i2);
261 i2 += 4;
262 const __m128 vi3 = _mm_loadu_ps(i3);
263 i3 += 4;
264 const __m128 vi4 = _mm_loadu_ps(i4);
265 i4 += 4;
266 const __m128 vi5 = _mm_loadu_ps(i5);
267 i5 += 4;
268 const __m128 vi6 = _mm_loadu_ps(i6);
269 i6 += 4;
270 const __m128 vi7 = _mm_loadu_ps(i7);
271 i7 += 4;
272
273 __m128 vmax = _mm_load_ps(ab);
274 ab += 4;
275 __m128i vidx = _mm_load_si128((const __m128i*) ib);
276 ib += 4;
277
278 const __m128i vm0 = _mm_castps_si128(_mm_cmpgt_ps(vi0, vmax));
279 vmax = _mm_max_ps(vi0, vmax);
280 vidx = _mm_or_si128(_mm_andnot_si128(vm0, vidx), _mm_and_si128(vm0, vidx0));
281
282 const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax));
283 const __m128i vidx1 = _mm_add_epi32(vidx0, v1);
284 vmax = _mm_max_ps(vi1, vmax);
285 vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, vidx1));
286
287 const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax));
288 const __m128i vidx2 = _mm_add_epi32(vidx1, v1);
289 vmax = _mm_max_ps(vi2, vmax);
290 vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, vidx2));
291
292 const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax));
293 const __m128i vidx3 = _mm_add_epi32(vidx2, v1);
294 vmax = _mm_max_ps(vi3, vmax);
295 vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, vidx3));
296
297 const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax));
298 const __m128i vidx4 = _mm_add_epi32(vidx3, v1);
299 vmax = _mm_max_ps(vi4, vmax);
300 vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, vidx4));
301
302 const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax));
303 const __m128i vidx5 = _mm_add_epi32(vidx4, v1);
304 vmax = _mm_max_ps(vi5, vmax);
305 vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, vidx5));
306
307 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax));
308 const __m128i vidx6 = _mm_add_epi32(vidx5, v1);
309 vmax = _mm_max_ps(vi6, vmax);
310 vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, vidx6));
311
312 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax));
313 const __m128i vidx7 = _mm_add_epi32(vidx6, v1);
314 vmax = _mm_max_ps(vi7, vmax);
315 vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, vidx7));
316
317 _mm_storeu_ps(o, vmax);
318 o += 4;
319 _mm_storeu_si128((__m128i*) i, vidx);
320 i += 4;
321 }
322 if (c != 0) {
323 const __m128 vi0 = _mm_loadu_ps(i0);
324 const __m128 vi1 = _mm_loadu_ps(i1);
325 const __m128 vi2 = _mm_loadu_ps(i2);
326 const __m128 vi3 = _mm_loadu_ps(i3);
327 const __m128 vi4 = _mm_loadu_ps(i4);
328 const __m128 vi5 = _mm_loadu_ps(i5);
329 const __m128 vi6 = _mm_loadu_ps(i6);
330 const __m128 vi7 = _mm_loadu_ps(i7);
331
332 __m128 vmax = _mm_load_ps(ab);
333 __m128i vidx = _mm_load_si128((const __m128i*) ib);
334
335 const __m128i vm0 = _mm_castps_si128(_mm_cmpgt_ps(vi0, vmax));
336 vmax = _mm_max_ps(vi0, vmax);
337 vidx = _mm_or_si128(_mm_andnot_si128(vm0, vidx), _mm_and_si128(vm0, vidx0));
338
339 const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax));
340 const __m128i vidx1 = _mm_add_epi32(vidx0, v1);
341 vmax = _mm_max_ps(vi1, vmax);
342 vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, vidx1));
343
344 const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax));
345 const __m128i vidx2 = _mm_add_epi32(vidx1, v1);
346 vmax = _mm_max_ps(vi2, vmax);
347 vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, vidx2));
348
349 const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax));
350 const __m128i vidx3 = _mm_add_epi32(vidx2, v1);
351 vmax = _mm_max_ps(vi3, vmax);
352 vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, vidx3));
353
354 const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax));
355 const __m128i vidx4 = _mm_add_epi32(vidx3, v1);
356 vmax = _mm_max_ps(vi4, vmax);
357 vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, vidx4));
358
359 const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax));
360 const __m128i vidx5 = _mm_add_epi32(vidx4, v1);
361 vmax = _mm_max_ps(vi5, vmax);
362 vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, vidx5));
363
364 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax));
365 const __m128i vidx6 = _mm_add_epi32(vidx5, v1);
366 vmax = _mm_max_ps(vi6, vmax);
367 vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, vidx6));
368
369 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax));
370 const __m128i vidx7 = _mm_add_epi32(vidx6, v1);
371 vmax = _mm_max_ps(vi7, vmax);
372 vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, vidx7));
373
374 if (c & 2) {
375 _mm_storel_pi((__m64*) o, vmax);
376 _mm_storel_epi64((__m128i*) i, vidx);
377 vmax = _mm_movehl_ps(vmax, vmax);
378 vidx = _mm_unpackhi_epi64(vidx, vidx);
379 o += 2;
380 i += 2;
381 }
382 if (c & 1) {
383 _mm_store_ss(o, vmax);
384 *i = (uint32_t) _mm_cvtsi128_si32(vidx);
385 o += 1;
386 i += 1;
387 }
388 }
389 }
390
391 output = (float*) ((uintptr_t) o + output_increment);
392 index = (uint32_t*) i;
393 } while (--output_pixels != 0);
394 }
395