• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2022 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 <immintrin.h>
9 
10 #include <xnnpack/intrinsics-polyfill.h>
11 #include <xnnpack/pavgpool.h>
12 
13 
xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8(size_t output_pixels,size_t kernel_elements,size_t channels,const void ** input,size_t input_offset,const void * zero,const void * multiplier,void * buffer,void * output,size_t input_increment,size_t output_increment,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])14 void xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8(
15     size_t output_pixels,
16     size_t kernel_elements,
17     size_t channels,
18     const void** input,
19     size_t input_offset,
20     const void* zero,
21     const void* multiplier,
22     void* buffer,
23     void* output,
24     size_t input_increment,
25     size_t output_increment,
26     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28   assert(output_pixels != 0);
29   assert(kernel_elements > 9);
30   assert(channels != 0);
31 
32   const __m256 voutput_min = _mm256_load_ps(params->avx.min);
33   const __m256 voutput_max = _mm256_load_ps(params->avx.max);
34 
35   uint16_t* o = (uint16_t*) output;
36   do {
37     {
38       const uint16_t* i0 = (const uint16_t*) *input++;
39       assert(i0 != NULL);
40       if XNN_UNPREDICTABLE(i0 != zero) {
41         i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
42       }
43       const uint16_t* i1 = (const uint16_t*) *input++;
44       assert(i1 != NULL);
45       if XNN_UNPREDICTABLE(i1 != zero) {
46         i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
47       }
48       const uint16_t* i2 = (const uint16_t*) *input++;
49       assert(i2 != NULL);
50       if XNN_UNPREDICTABLE(i2 != zero) {
51         i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
52       }
53       const uint16_t* i3 = (const uint16_t*) *input++;
54       assert(i3 != NULL);
55       if XNN_UNPREDICTABLE(i3 != zero) {
56         i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
57       }
58       const uint16_t* i4 = (const uint16_t*) *input++;
59       assert(i4 != NULL);
60       if XNN_UNPREDICTABLE(i4 != zero) {
61         i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
62       }
63       const uint16_t* i5 = (const uint16_t*) *input++;
64       assert(i5 != NULL);
65       if XNN_UNPREDICTABLE(i5 != zero) {
66         i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
67       }
68       const uint16_t* i6 = (const uint16_t*) *input++;
69       assert(i6 != NULL);
70       if XNN_UNPREDICTABLE(i6 != zero) {
71         i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
72       }
73       const uint16_t* i7 = (const uint16_t*) *input++;
74       assert(i7 != NULL);
75       if XNN_UNPREDICTABLE(i7 != zero) {
76         i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
77       }
78       const uint16_t* i8 = (const uint16_t*) *input++;
79       assert(i8 != NULL);
80       if XNN_UNPREDICTABLE(i8 != zero) {
81         i8 = (const uint16_t*) ((uintptr_t) i8 + input_offset);
82       }
83 
84       uint16_t* b = (uint16_t*) buffer;
85       for (size_t c = 0; c < channels; c += 8) {
86         const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
87         i0 += 8;
88         const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
89         i1 += 8;
90         const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
91         i2 += 8;
92         const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
93         i3 += 8;
94         const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
95         i4 += 8;
96         const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
97         i5 += 8;
98         const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
99         i6 += 8;
100         const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
101         i7 += 8;
102         const __m256 vi8 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
103         i8 += 8;
104 
105         const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC));
106         const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC));
107         const __m256 vsum45 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi4, vi5), _MM_FROUND_NO_EXC));
108         const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC));
109         const __m256 vsum018 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01, vi8), _MM_FROUND_NO_EXC));
110         const __m256 vsum2345 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum23, vsum45), _MM_FROUND_NO_EXC));
111         const __m256 vsum01678 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum018, vsum67), _MM_FROUND_NO_EXC));
112         const __m256 vsum = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum2345, vsum01678), _MM_FROUND_NO_EXC));
113 
114         _mm_storeu_si128((__m128i*) b, _mm256_cvtps_ph(vsum, _MM_FROUND_NO_EXC));
115         b += 8;
116       }
117     }
118 
119     size_t k = kernel_elements;
120     for (k -= 9; k > 8; k -= 8) {
121       const uint16_t* i0 = (const uint16_t*) *input++;
122       assert(i0 != NULL);
123       if XNN_UNPREDICTABLE(i0 != zero) {
124         i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
125       }
126       const uint16_t* i1 = (const uint16_t*) *input++;
127       assert(i1 != NULL);
128       if XNN_UNPREDICTABLE(i1 != zero) {
129         i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
130       }
131       const uint16_t* i2 = (const uint16_t*) *input++;
132       assert(i2 != NULL);
133       if XNN_UNPREDICTABLE(i2 != zero) {
134         i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
135       }
136       const uint16_t* i3 = (const uint16_t*) *input++;
137       assert(i3 != NULL);
138       if XNN_UNPREDICTABLE(i3 != zero) {
139         i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
140       }
141       const uint16_t* i4 = (const uint16_t*) *input++;
142       assert(i4 != NULL);
143       if XNN_UNPREDICTABLE(i4 != zero) {
144         i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
145       }
146       const uint16_t* i5 = (const uint16_t*) *input++;
147       assert(i5 != NULL);
148       if XNN_UNPREDICTABLE(i5 != zero) {
149         i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
150       }
151       const uint16_t* i6 = (const uint16_t*) *input++;
152       assert(i6 != NULL);
153       if XNN_UNPREDICTABLE(i6 != zero) {
154         i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
155       }
156       const uint16_t* i7 = (const uint16_t*) *input++;
157       assert(i7 != NULL);
158       if XNN_UNPREDICTABLE(i7 != zero) {
159         i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
160       }
161 
162       uint16_t* b = (uint16_t*) buffer;
163       for (size_t c = 0; c < channels; c += 8) {
164         const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
165         i0 += 8;
166         const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
167         i1 += 8;
168         const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
169         i2 += 8;
170         const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
171         i3 += 8;
172         const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
173         i4 += 8;
174         const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
175         i5 += 8;
176         const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
177         i6 += 8;
178         const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
179         i7 += 8;
180         const __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
181 
182         const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC));
183         const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC));
184         const __m256 vsum45 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi4, vi5), _MM_FROUND_NO_EXC));
185         const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC));
186         const __m256 vsum01a = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01, vacc), _MM_FROUND_NO_EXC));
187         const __m256 vsum2345 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum23, vsum45), _MM_FROUND_NO_EXC));
188         const __m256 vsum0167a = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01a, vsum67), _MM_FROUND_NO_EXC));
189         const __m256 vsum = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum2345, vsum0167a), _MM_FROUND_NO_EXC));
190 
191         _mm_storeu_si128((__m128i*) b, _mm256_cvtps_ph(vsum, _MM_FROUND_NO_EXC));
192         b += 8;
193       }
194     }
195 
196     {
197       const uint16_t* i0 = (const uint16_t*) input[0];
198       assert(i0 != NULL);
199       const uint16_t* i1 = (const uint16_t*) input[1];
200       const uint16_t* i2 = (const uint16_t*) input[2];
201       const uint16_t* i3 = (const uint16_t*) input[3];
202       const uint16_t* i4 = (const uint16_t*) input[4];
203       const uint16_t* i5 = (const uint16_t*) input[5];
204       const uint16_t* i6 = (const uint16_t*) input[6];
205       const uint16_t* i7 = (const uint16_t*) input[7];
206       input = (const void**) ((uintptr_t) input + input_increment);
207       if (k < 2) {
208         i1 = (const uint16_t*) zero;
209       }
210       assert(i1 != NULL);
211       if (k <= 2) {
212         i2 = (const uint16_t*) zero;
213       }
214       assert(i2 != NULL);
215       if (k < 4) {
216         i3 = (const uint16_t*) zero;
217       }
218       assert(i3 != NULL);
219       if (k <= 4) {
220         i4 = (const uint16_t*) zero;
221       }
222       assert(i4 != NULL);
223       if (k < 6) {
224         i5 = (const uint16_t*) zero;
225       }
226       assert(i5 != NULL);
227       if (k <= 6) {
228         i6 = (const uint16_t*) zero;
229       }
230       assert(i6 != NULL);
231       if (k < 8) {
232         i7 = (const uint16_t*) zero;
233       }
234       assert(i7 != NULL);
235       if XNN_UNPREDICTABLE(i0 != zero) {
236         i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
237       }
238       if XNN_UNPREDICTABLE(i1 != zero) {
239         i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
240       }
241       if XNN_UNPREDICTABLE(i2 != zero) {
242         i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
243       }
244       if XNN_UNPREDICTABLE(i3 != zero) {
245         i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
246       }
247       if XNN_UNPREDICTABLE(i4 != zero) {
248         i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
249       }
250       if XNN_UNPREDICTABLE(i5 != zero) {
251         i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
252       }
253       if XNN_UNPREDICTABLE(i6 != zero) {
254         i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
255       }
256       if XNN_UNPREDICTABLE(i7 != zero) {
257         i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
258       }
259 
260       const __m256 vmultiplier = _mm256_cvtph_ps(_mm_set1_epi16((short) *((const uint16_t*) multiplier)));
261       multiplier = (const uint16_t*) multiplier + 1;
262 
263       size_t c = channels;
264       const uint16_t* b = (const uint16_t*) buffer;
265       while (c >= 8) {
266         const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
267         i0 += 8;
268         const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
269         i1 += 8;
270         const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
271         i2 += 8;
272         const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
273         i3 += 8;
274         const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
275         i4 += 8;
276         const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
277         i5 += 8;
278         const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
279         i6 += 8;
280         const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
281         i7 += 8;
282         const __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
283         b += 8;
284 
285         const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC));
286         const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC));
287         const __m256 vsum45 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi4, vi5), _MM_FROUND_NO_EXC));
288         const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC));
289         const __m256 vsum01a = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01, vacc), _MM_FROUND_NO_EXC));
290         const __m256 vsum2345 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum23, vsum45), _MM_FROUND_NO_EXC));
291         const __m256 vsum0167a = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01a, vsum67), _MM_FROUND_NO_EXC));
292         const __m256 vsum = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum2345, vsum0167a), _MM_FROUND_NO_EXC));
293 
294         __m256 vout = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vsum, vmultiplier), _MM_FROUND_NO_EXC));
295         vout = _mm256_max_ps(vout, voutput_min);
296         vout = _mm256_min_ps(vout, voutput_max);
297 
298         _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vout, _MM_FROUND_NO_EXC));
299         o += 8;
300 
301         c -= 8;
302       }
303       if (c != 0) {
304         const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
305         const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
306         const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
307         const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
308         const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
309         const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
310         const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
311         const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
312         const __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
313 
314         const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC));
315         const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC));
316         const __m256 vsum45 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi4, vi5), _MM_FROUND_NO_EXC));
317         const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC));
318         const __m256 vsum01a = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01, vacc), _MM_FROUND_NO_EXC));
319         const __m256 vsum2345 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum23, vsum45), _MM_FROUND_NO_EXC));
320         const __m256 vsum0167a = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01a, vsum67), _MM_FROUND_NO_EXC));
321         const __m256 vsum = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum2345, vsum0167a), _MM_FROUND_NO_EXC));
322 
323         __m256 vout = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vsum, vmultiplier), _MM_FROUND_NO_EXC));
324         vout = _mm256_max_ps(vout, voutput_min);
325         vout = _mm256_min_ps(vout, voutput_max);
326 
327         __m128i vh = _mm256_cvtps_ph(vout, _MM_FROUND_NO_EXC);
328         if (c & 4) {
329           _mm_storel_epi64((__m128i*) o, vh);
330           vh = _mm_unpackhi_epi64(vh, vh);
331           o += 4;
332         }
333         if (c & 2) {
334           _mm_storeu_si32(o, vh);
335           vh = _mm_srli_epi64(vh, 32);
336           o += 2;
337         }
338         if (c & 1) {
339           *o = (uint16_t) _mm_extract_epi16(vh, 0);
340           o += 1;
341         }
342       }
343     }
344     o = (uint16_t*) ((uintptr_t) o + output_increment);
345   } while (--output_pixels != 0);
346 }
347