1/*
2 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 *
4 *  Use of this source code is governed by a BSD-style license
5 *  that can be found in the LICENSE file in the root of the source
6 *  tree. An additional intellectual property rights grant can be found
7 *  in the file PATENTS.  All contributing project authors may
8 *  be found in the AUTHORS file in the root of the source tree.
9 */
10
11#include <emmintrin.h>  // SSE2
12
13#include "./vpx_config.h"
14
15#include "vp9/encoder/vp9_variance.h"
16#include "vpx_ports/mem.h"
17
18typedef unsigned int (*variance_fn_t) (const unsigned char *src, int src_stride,
19                                       const unsigned char *ref, int ref_stride,
20                                       unsigned int *sse, int *sum);
21
22unsigned int vp9_get_mb_ss_sse2(const int16_t *src) {
23  __m128i vsum = _mm_setzero_si128();
24  int i;
25
26  for (i = 0; i < 32; ++i) {
27    const __m128i v = _mm_loadu_si128((const __m128i *)src);
28    vsum = _mm_add_epi32(vsum, _mm_madd_epi16(v, v));
29    src += 8;
30  }
31
32  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8));
33  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4));
34  return  _mm_cvtsi128_si32(vsum);
35}
36
37#define READ64(p, stride, i) \
38  _mm_unpacklo_epi8(_mm_cvtsi32_si128(*(const uint32_t *)(p + i * stride)), \
39      _mm_cvtsi32_si128(*(const uint32_t *)(p + (i + 1) * stride)))
40
41unsigned int vp9_get4x4var_sse2(const uint8_t *src, int src_stride,
42                                const uint8_t *ref, int ref_stride,
43                                unsigned int *sse, int *sum) {
44  const __m128i zero = _mm_setzero_si128();
45  const __m128i src0 = _mm_unpacklo_epi8(READ64(src, src_stride, 0), zero);
46  const __m128i src1 = _mm_unpacklo_epi8(READ64(src, src_stride, 2), zero);
47  const __m128i ref0 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 0), zero);
48  const __m128i ref1 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 2), zero);
49  const __m128i diff0 = _mm_sub_epi16(src0, ref0);
50  const __m128i diff1 = _mm_sub_epi16(src1, ref1);
51
52  // sum
53  __m128i vsum = _mm_add_epi16(diff0, diff1);
54  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
55  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
56  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
57  *sum = (int16_t)_mm_extract_epi16(vsum, 0);
58
59  // sse
60  vsum = _mm_add_epi32(_mm_madd_epi16(diff0, diff0),
61                       _mm_madd_epi16(diff1, diff1));
62  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8));
63  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4));
64  *sse = _mm_cvtsi128_si32(vsum);
65
66  return 0;
67}
68
69unsigned int vp9_get8x8var_sse2(const uint8_t *src, int src_stride,
70                                const uint8_t *ref, int ref_stride,
71                                unsigned int *sse, int *sum) {
72  const __m128i zero = _mm_setzero_si128();
73  __m128i vsum = _mm_setzero_si128();
74  __m128i vsse = _mm_setzero_si128();
75  int i;
76
77  for (i = 0; i < 8; i += 2) {
78    const __m128i src0 = _mm_unpacklo_epi8(_mm_loadl_epi64(
79        (const __m128i *)(src + i * src_stride)), zero);
80    const __m128i ref0 = _mm_unpacklo_epi8(_mm_loadl_epi64(
81        (const __m128i *)(ref + i * ref_stride)), zero);
82    const __m128i diff0 = _mm_sub_epi16(src0, ref0);
83
84    const __m128i src1 = _mm_unpacklo_epi8(_mm_loadl_epi64(
85        (const __m128i *)(src + (i + 1) * src_stride)), zero);
86    const __m128i ref1 = _mm_unpacklo_epi8(_mm_loadl_epi64(
87        (const __m128i *)(ref + (i + 1) * ref_stride)), zero);
88    const __m128i diff1 = _mm_sub_epi16(src1, ref1);
89
90    vsum = _mm_add_epi16(vsum, diff0);
91    vsum = _mm_add_epi16(vsum, diff1);
92    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
93    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
94  }
95
96  // sum
97  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
98  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
99  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
100  *sum = (int16_t)_mm_extract_epi16(vsum, 0);
101
102  // sse
103  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
104  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
105  *sse = _mm_cvtsi128_si32(vsse);
106
107  return 0;
108}
109
110unsigned int vp9_get16x16var_sse2(const uint8_t *src, int src_stride,
111                                  const uint8_t *ref, int ref_stride,
112                                  unsigned int *sse, int *sum) {
113  const __m128i zero = _mm_setzero_si128();
114  __m128i vsum = _mm_setzero_si128();
115  __m128i vsse = _mm_setzero_si128();
116  int i;
117
118  for (i = 0; i < 16; ++i) {
119    const __m128i s = _mm_loadu_si128((const __m128i *)src);
120    const __m128i r = _mm_loadu_si128((const __m128i *)ref);
121
122    const __m128i src0 = _mm_unpacklo_epi8(s, zero);
123    const __m128i ref0 = _mm_unpacklo_epi8(r, zero);
124    const __m128i diff0 = _mm_sub_epi16(src0, ref0);
125
126    const __m128i src1 = _mm_unpackhi_epi8(s, zero);
127    const __m128i ref1 = _mm_unpackhi_epi8(r, zero);
128    const __m128i diff1 = _mm_sub_epi16(src1, ref1);
129
130    vsum = _mm_add_epi16(vsum, diff0);
131    vsum = _mm_add_epi16(vsum, diff1);
132    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
133    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
134
135    src += src_stride;
136    ref += ref_stride;
137  }
138
139  // sum
140  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
141  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
142  *sum = (int16_t)_mm_extract_epi16(vsum, 0) +
143             (int16_t)_mm_extract_epi16(vsum, 1);
144
145  // sse
146  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
147  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
148  *sse = _mm_cvtsi128_si32(vsse);
149
150  return 0;
151}
152
153
154static void variance_sse2(const unsigned char *src, int src_stride,
155                          const unsigned char *ref, int ref_stride,
156                          int w, int h, unsigned int *sse, int *sum,
157                          variance_fn_t var_fn, int block_size) {
158  int i, j;
159
160  *sse = 0;
161  *sum = 0;
162
163  for (i = 0; i < h; i += block_size) {
164    for (j = 0; j < w; j += block_size) {
165      unsigned int sse0;
166      int sum0;
167      var_fn(src + src_stride * i + j, src_stride,
168             ref + ref_stride * i + j, ref_stride, &sse0, &sum0);
169      *sse += sse0;
170      *sum += sum0;
171    }
172  }
173}
174
175unsigned int vp9_variance4x4_sse2(const unsigned char *src, int src_stride,
176                                  const unsigned char *ref, int ref_stride,
177                                  unsigned int *sse) {
178  int sum;
179  vp9_get4x4var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
180  return *sse - (((unsigned int)sum * sum) >> 4);
181}
182
183unsigned int vp9_variance8x4_sse2(const uint8_t *src, int src_stride,
184                                  const uint8_t *ref, int ref_stride,
185                                  unsigned int *sse) {
186  int sum;
187  variance_sse2(src, src_stride, ref, ref_stride, 8, 4,
188                sse, &sum, vp9_get4x4var_sse2, 4);
189  return *sse - (((unsigned int)sum * sum) >> 5);
190}
191
192unsigned int vp9_variance4x8_sse2(const uint8_t *src, int src_stride,
193                                  const uint8_t *ref, int ref_stride,
194                                  unsigned int *sse) {
195  int sum;
196  variance_sse2(src, src_stride, ref, ref_stride, 4, 8,
197                sse, &sum, vp9_get4x4var_sse2, 4);
198  return *sse - (((unsigned int)sum * sum) >> 5);
199}
200
201unsigned int vp9_variance8x8_sse2(const unsigned char *src, int src_stride,
202                                  const unsigned char *ref, int ref_stride,
203                                  unsigned int *sse) {
204  int sum;
205  vp9_get8x8var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
206  return *sse - (((unsigned int)sum * sum) >> 6);
207}
208
209unsigned int vp9_variance16x8_sse2(const unsigned char *src, int src_stride,
210                                   const unsigned char *ref, int ref_stride,
211                                   unsigned int *sse) {
212  int sum;
213  variance_sse2(src, src_stride, ref, ref_stride, 16, 8,
214                sse, &sum, vp9_get8x8var_sse2, 8);
215  return *sse - (((unsigned int)sum * sum) >> 7);
216}
217
218unsigned int vp9_variance8x16_sse2(const unsigned char *src, int src_stride,
219                                   const unsigned char *ref, int ref_stride,
220                                   unsigned int *sse) {
221  int sum;
222  variance_sse2(src, src_stride, ref, ref_stride, 8, 16,
223                sse, &sum, vp9_get8x8var_sse2, 8);
224  return *sse - (((unsigned int)sum * sum) >> 7);
225}
226
227unsigned int vp9_variance16x16_sse2(const unsigned char *src, int src_stride,
228                                    const unsigned char *ref, int ref_stride,
229                                    unsigned int *sse) {
230  int sum;
231  vp9_get16x16var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
232  return *sse - (((unsigned int)sum * sum) >> 8);
233}
234
235unsigned int vp9_variance32x32_sse2(const uint8_t *src, int src_stride,
236                                    const uint8_t *ref, int ref_stride,
237                                    unsigned int *sse) {
238  int sum;
239  variance_sse2(src, src_stride, ref, ref_stride, 32, 32,
240                sse, &sum, vp9_get16x16var_sse2, 16);
241  return *sse - (((int64_t)sum * sum) >> 10);
242}
243
244unsigned int vp9_variance32x16_sse2(const uint8_t *src, int src_stride,
245                                    const uint8_t *ref, int ref_stride,
246                                    unsigned int *sse) {
247  int sum;
248  variance_sse2(src, src_stride, ref, ref_stride, 32, 16,
249                sse, &sum, vp9_get16x16var_sse2, 16);
250  return *sse - (((int64_t)sum * sum) >> 9);
251}
252
253unsigned int vp9_variance16x32_sse2(const uint8_t *src, int src_stride,
254                                    const uint8_t *ref, int ref_stride,
255                                    unsigned int *sse) {
256  int sum;
257  variance_sse2(src, src_stride, ref, ref_stride, 16, 32,
258                sse, &sum, vp9_get16x16var_sse2, 16);
259  return *sse - (((int64_t)sum * sum) >> 9);
260}
261
262unsigned int vp9_variance64x64_sse2(const uint8_t *src, int src_stride,
263                                    const uint8_t *ref, int ref_stride,
264                                    unsigned int *sse) {
265  int sum;
266  variance_sse2(src, src_stride, ref, ref_stride, 64, 64,
267                sse, &sum, vp9_get16x16var_sse2, 16);
268  return *sse - (((int64_t)sum * sum) >> 12);
269}
270
271unsigned int vp9_variance64x32_sse2(const uint8_t *src, int src_stride,
272                                    const uint8_t *ref, int ref_stride,
273                                    unsigned int *sse) {
274  int sum;
275  variance_sse2(src, src_stride, ref, ref_stride, 64, 32,
276                sse, &sum, vp9_get16x16var_sse2, 16);
277  return *sse - (((int64_t)sum * sum) >> 11);
278}
279
280unsigned int vp9_variance32x64_sse2(const uint8_t *src, int src_stride,
281                                    const uint8_t *ref, int ref_stride,
282                                    unsigned int *sse) {
283  int sum;
284  variance_sse2(src, src_stride, ref, ref_stride, 32, 64,
285                sse, &sum, vp9_get16x16var_sse2, 16);
286  return *sse - (((int64_t)sum * sum) >> 11);
287}
288
289unsigned int vp9_mse8x8_sse2(const uint8_t *src, int src_stride,
290                             const uint8_t *ref, int ref_stride,
291                             unsigned int *sse) {
292  vp9_variance8x8_sse2(src, src_stride, ref, ref_stride, sse);
293  return *sse;
294}
295
296unsigned int vp9_mse8x16_sse2(const uint8_t *src, int src_stride,
297                              const uint8_t *ref, int ref_stride,
298                              unsigned int *sse) {
299  vp9_variance8x16_sse2(src, src_stride, ref, ref_stride, sse);
300  return *sse;
301}
302
303unsigned int vp9_mse16x8_sse2(const uint8_t *src, int src_stride,
304                              const uint8_t *ref, int ref_stride,
305                              unsigned int *sse) {
306  vp9_variance16x8_sse2(src, src_stride, ref, ref_stride, sse);
307  return *sse;
308}
309
310unsigned int vp9_mse16x16_sse2(const uint8_t *src, int src_stride,
311                               const uint8_t *ref, int ref_stride,
312                               unsigned int *sse) {
313  vp9_variance16x16_sse2(src, src_stride, ref, ref_stride, sse);
314  return *sse;
315}
316
317#define DECL(w, opt) \
318int vp9_sub_pixel_variance##w##xh_##opt(const uint8_t *src, \
319                                        ptrdiff_t src_stride, \
320                                        int x_offset, int y_offset, \
321                                        const uint8_t *dst, \
322                                        ptrdiff_t dst_stride, \
323                                        int height, unsigned int *sse)
324#define DECLS(opt1, opt2) \
325DECL(4, opt2); \
326DECL(8, opt1); \
327DECL(16, opt1)
328
329DECLS(sse2, sse);
330DECLS(ssse3, ssse3);
331#undef DECLS
332#undef DECL
333
334#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
335unsigned int vp9_sub_pixel_variance##w##x##h##_##opt(const uint8_t *src, \
336                                                     int src_stride, \
337                                                     int x_offset, \
338                                                     int y_offset, \
339                                                     const uint8_t *dst, \
340                                                     int dst_stride, \
341                                                     unsigned int *sse_ptr) { \
342  unsigned int sse; \
343  int se = vp9_sub_pixel_variance##wf##xh_##opt(src, src_stride, x_offset, \
344                                                y_offset, dst, dst_stride, \
345                                                h, &sse); \
346  if (w > wf) { \
347    unsigned int sse2; \
348    int se2 = vp9_sub_pixel_variance##wf##xh_##opt(src + 16, src_stride, \
349                                                   x_offset, y_offset, \
350                                                   dst + 16, dst_stride, \
351                                                   h, &sse2); \
352    se += se2; \
353    sse += sse2; \
354    if (w > wf * 2) { \
355      se2 = vp9_sub_pixel_variance##wf##xh_##opt(src + 32, src_stride, \
356                                                 x_offset, y_offset, \
357                                                 dst + 32, dst_stride, \
358                                                 h, &sse2); \
359      se += se2; \
360      sse += sse2; \
361      se2 = vp9_sub_pixel_variance##wf##xh_##opt(src + 48, src_stride, \
362                                                 x_offset, y_offset, \
363                                                 dst + 48, dst_stride, \
364                                                 h, &sse2); \
365      se += se2; \
366      sse += sse2; \
367    } \
368  } \
369  *sse_ptr = sse; \
370  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
371}
372
373#define FNS(opt1, opt2) \
374FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
375FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
376FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
377FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
378FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
379FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
380FN(16, 16, 16, 4, 4, opt1, (unsigned int)); \
381FN(16,  8, 16, 4, 3, opt1, (unsigned int)); \
382FN(8,  16,  8, 3, 4, opt1, (unsigned int)); \
383FN(8,   8,  8, 3, 3, opt1, (unsigned int)); \
384FN(8,   4,  8, 3, 2, opt1, (unsigned int)); \
385FN(4,   8,  4, 2, 3, opt2, (unsigned int)); \
386FN(4,   4,  4, 2, 2, opt2, (unsigned int))
387
388FNS(sse2, sse);
389FNS(ssse3, ssse3);
390
391#undef FNS
392#undef FN
393
394#define DECL(w, opt) \
395int vp9_sub_pixel_avg_variance##w##xh_##opt(const uint8_t *src, \
396                                            ptrdiff_t src_stride, \
397                                            int x_offset, int y_offset, \
398                                            const uint8_t *dst, \
399                                            ptrdiff_t dst_stride, \
400                                            const uint8_t *sec, \
401                                            ptrdiff_t sec_stride, \
402                                            int height, unsigned int *sse)
403#define DECLS(opt1, opt2) \
404DECL(4, opt2); \
405DECL(8, opt1); \
406DECL(16, opt1)
407
408DECLS(sse2, sse);
409DECLS(ssse3, ssse3);
410#undef DECL
411#undef DECLS
412
413#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
414unsigned int vp9_sub_pixel_avg_variance##w##x##h##_##opt(const uint8_t *src, \
415                                                         int src_stride, \
416                                                         int x_offset, \
417                                                         int y_offset, \
418                                                         const uint8_t *dst, \
419                                                         int dst_stride, \
420                                                         unsigned int *sseptr, \
421                                                         const uint8_t *sec) { \
422  unsigned int sse; \
423  int se = vp9_sub_pixel_avg_variance##wf##xh_##opt(src, src_stride, x_offset, \
424                                                    y_offset, dst, dst_stride, \
425                                                    sec, w, h, &sse); \
426  if (w > wf) { \
427    unsigned int sse2; \
428    int se2 = vp9_sub_pixel_avg_variance##wf##xh_##opt(src + 16, src_stride, \
429                                                       x_offset, y_offset, \
430                                                       dst + 16, dst_stride, \
431                                                       sec + 16, w, h, &sse2); \
432    se += se2; \
433    sse += sse2; \
434    if (w > wf * 2) { \
435      se2 = vp9_sub_pixel_avg_variance##wf##xh_##opt(src + 32, src_stride, \
436                                                     x_offset, y_offset, \
437                                                     dst + 32, dst_stride, \
438                                                     sec + 32, w, h, &sse2); \
439      se += se2; \
440      sse += sse2; \
441      se2 = vp9_sub_pixel_avg_variance##wf##xh_##opt(src + 48, src_stride, \
442                                                     x_offset, y_offset, \
443                                                     dst + 48, dst_stride, \
444                                                     sec + 48, w, h, &sse2); \
445      se += se2; \
446      sse += sse2; \
447    } \
448  } \
449  *sseptr = sse; \
450  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
451}
452
453#define FNS(opt1, opt2) \
454FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
455FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
456FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
457FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
458FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
459FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
460FN(16, 16, 16, 4, 4, opt1, (unsigned int)); \
461FN(16,  8, 16, 4, 3, opt1, (unsigned int)); \
462FN(8,  16,  8, 3, 4, opt1, (unsigned int)); \
463FN(8,   8,  8, 3, 3, opt1, (unsigned int)); \
464FN(8,   4,  8, 3, 2, opt1, (unsigned int)); \
465FN(4,   8,  4, 2, 3, opt2, (unsigned int)); \
466FN(4,   4,  4, 2, 2, opt2, (unsigned int))
467
468FNS(sse2, sse);
469FNS(ssse3, ssse3);
470
471#undef FNS
472#undef FN
473