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#include "./vpx_dsp_rtcd.h"
15
16#include "vpx_ports/mem.h"
17
18typedef void (*getNxMvar_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 vpx_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
41static void 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
67void vpx_get8x8var_sse2(const uint8_t *src, int src_stride,
68                        const uint8_t *ref, int ref_stride,
69                        unsigned int *sse, int *sum) {
70  const __m128i zero = _mm_setzero_si128();
71  __m128i vsum = _mm_setzero_si128();
72  __m128i vsse = _mm_setzero_si128();
73  int i;
74
75  for (i = 0; i < 8; i += 2) {
76    const __m128i src0 = _mm_unpacklo_epi8(_mm_loadl_epi64(
77        (const __m128i *)(src + i * src_stride)), zero);
78    const __m128i ref0 = _mm_unpacklo_epi8(_mm_loadl_epi64(
79        (const __m128i *)(ref + i * ref_stride)), zero);
80    const __m128i diff0 = _mm_sub_epi16(src0, ref0);
81
82    const __m128i src1 = _mm_unpacklo_epi8(_mm_loadl_epi64(
83        (const __m128i *)(src + (i + 1) * src_stride)), zero);
84    const __m128i ref1 = _mm_unpacklo_epi8(_mm_loadl_epi64(
85        (const __m128i *)(ref + (i + 1) * ref_stride)), zero);
86    const __m128i diff1 = _mm_sub_epi16(src1, ref1);
87
88    vsum = _mm_add_epi16(vsum, diff0);
89    vsum = _mm_add_epi16(vsum, diff1);
90    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
91    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
92  }
93
94  // sum
95  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
96  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
97  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
98  *sum = (int16_t)_mm_extract_epi16(vsum, 0);
99
100  // sse
101  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
102  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
103  *sse = _mm_cvtsi128_si32(vsse);
104}
105
106void vpx_get16x16var_sse2(const uint8_t *src, int src_stride,
107                          const uint8_t *ref, int ref_stride,
108                          unsigned int *sse, int *sum) {
109  const __m128i zero = _mm_setzero_si128();
110  __m128i vsum = _mm_setzero_si128();
111  __m128i vsse = _mm_setzero_si128();
112  int i;
113
114  for (i = 0; i < 16; ++i) {
115    const __m128i s = _mm_loadu_si128((const __m128i *)src);
116    const __m128i r = _mm_loadu_si128((const __m128i *)ref);
117
118    const __m128i src0 = _mm_unpacklo_epi8(s, zero);
119    const __m128i ref0 = _mm_unpacklo_epi8(r, zero);
120    const __m128i diff0 = _mm_sub_epi16(src0, ref0);
121
122    const __m128i src1 = _mm_unpackhi_epi8(s, zero);
123    const __m128i ref1 = _mm_unpackhi_epi8(r, zero);
124    const __m128i diff1 = _mm_sub_epi16(src1, ref1);
125
126    vsum = _mm_add_epi16(vsum, diff0);
127    vsum = _mm_add_epi16(vsum, diff1);
128    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
129    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
130
131    src += src_stride;
132    ref += ref_stride;
133  }
134
135  // sum
136  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
137  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
138  *sum = (int16_t)_mm_extract_epi16(vsum, 0) +
139             (int16_t)_mm_extract_epi16(vsum, 1);
140
141  // sse
142  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
143  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
144  *sse = _mm_cvtsi128_si32(vsse);
145}
146
147
148static void variance_sse2(const unsigned char *src, int src_stride,
149                          const unsigned char *ref, int ref_stride,
150                          int w, int h, unsigned int *sse, int *sum,
151                          getNxMvar_fn_t var_fn, int block_size) {
152  int i, j;
153
154  *sse = 0;
155  *sum = 0;
156
157  for (i = 0; i < h; i += block_size) {
158    for (j = 0; j < w; j += block_size) {
159      unsigned int sse0;
160      int sum0;
161      var_fn(src + src_stride * i + j, src_stride,
162             ref + ref_stride * i + j, ref_stride, &sse0, &sum0);
163      *sse += sse0;
164      *sum += sum0;
165    }
166  }
167}
168
169unsigned int vpx_variance4x4_sse2(const unsigned char *src, int src_stride,
170                                  const unsigned char *ref, int ref_stride,
171                                  unsigned int *sse) {
172  int sum;
173  get4x4var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
174  return *sse - (((unsigned int)sum * sum) >> 4);
175}
176
177unsigned int vpx_variance8x4_sse2(const uint8_t *src, int src_stride,
178                                  const uint8_t *ref, int ref_stride,
179                                  unsigned int *sse) {
180  int sum;
181  variance_sse2(src, src_stride, ref, ref_stride, 8, 4,
182                sse, &sum, get4x4var_sse2, 4);
183  return *sse - (((unsigned int)sum * sum) >> 5);
184}
185
186unsigned int vpx_variance4x8_sse2(const uint8_t *src, int src_stride,
187                                  const uint8_t *ref, int ref_stride,
188                                  unsigned int *sse) {
189  int sum;
190  variance_sse2(src, src_stride, ref, ref_stride, 4, 8,
191                sse, &sum, get4x4var_sse2, 4);
192  return *sse - (((unsigned int)sum * sum) >> 5);
193}
194
195unsigned int vpx_variance8x8_sse2(const unsigned char *src, int src_stride,
196                                  const unsigned char *ref, int ref_stride,
197                                  unsigned int *sse) {
198  int sum;
199  vpx_get8x8var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
200  return *sse - (((unsigned int)sum * sum) >> 6);
201}
202
203unsigned int vpx_variance16x8_sse2(const unsigned char *src, int src_stride,
204                                   const unsigned char *ref, int ref_stride,
205                                   unsigned int *sse) {
206  int sum;
207  variance_sse2(src, src_stride, ref, ref_stride, 16, 8,
208                sse, &sum, vpx_get8x8var_sse2, 8);
209  return *sse - (((unsigned int)sum * sum) >> 7);
210}
211
212unsigned int vpx_variance8x16_sse2(const unsigned char *src, int src_stride,
213                                   const unsigned char *ref, int ref_stride,
214                                   unsigned int *sse) {
215  int sum;
216  variance_sse2(src, src_stride, ref, ref_stride, 8, 16,
217                sse, &sum, vpx_get8x8var_sse2, 8);
218  return *sse - (((unsigned int)sum * sum) >> 7);
219}
220
221unsigned int vpx_variance16x16_sse2(const unsigned char *src, int src_stride,
222                                    const unsigned char *ref, int ref_stride,
223                                    unsigned int *sse) {
224  int sum;
225  vpx_get16x16var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
226  return *sse - (((unsigned int)sum * sum) >> 8);
227}
228
229unsigned int vpx_variance32x32_sse2(const uint8_t *src, int src_stride,
230                                    const uint8_t *ref, int ref_stride,
231                                    unsigned int *sse) {
232  int sum;
233  variance_sse2(src, src_stride, ref, ref_stride, 32, 32,
234                sse, &sum, vpx_get16x16var_sse2, 16);
235  return *sse - (((int64_t)sum * sum) >> 10);
236}
237
238unsigned int vpx_variance32x16_sse2(const uint8_t *src, int src_stride,
239                                    const uint8_t *ref, int ref_stride,
240                                    unsigned int *sse) {
241  int sum;
242  variance_sse2(src, src_stride, ref, ref_stride, 32, 16,
243                sse, &sum, vpx_get16x16var_sse2, 16);
244  return *sse - (((int64_t)sum * sum) >> 9);
245}
246
247unsigned int vpx_variance16x32_sse2(const uint8_t *src, int src_stride,
248                                    const uint8_t *ref, int ref_stride,
249                                    unsigned int *sse) {
250  int sum;
251  variance_sse2(src, src_stride, ref, ref_stride, 16, 32,
252                sse, &sum, vpx_get16x16var_sse2, 16);
253  return *sse - (((int64_t)sum * sum) >> 9);
254}
255
256unsigned int vpx_variance64x64_sse2(const uint8_t *src, int src_stride,
257                                    const uint8_t *ref, int ref_stride,
258                                    unsigned int *sse) {
259  int sum;
260  variance_sse2(src, src_stride, ref, ref_stride, 64, 64,
261                sse, &sum, vpx_get16x16var_sse2, 16);
262  return *sse - (((int64_t)sum * sum) >> 12);
263}
264
265unsigned int vpx_variance64x32_sse2(const uint8_t *src, int src_stride,
266                                    const uint8_t *ref, int ref_stride,
267                                    unsigned int *sse) {
268  int sum;
269  variance_sse2(src, src_stride, ref, ref_stride, 64, 32,
270                sse, &sum, vpx_get16x16var_sse2, 16);
271  return *sse - (((int64_t)sum * sum) >> 11);
272}
273
274unsigned int vpx_variance32x64_sse2(const uint8_t *src, int src_stride,
275                                    const uint8_t *ref, int ref_stride,
276                                    unsigned int *sse) {
277  int sum;
278  variance_sse2(src, src_stride, ref, ref_stride, 32, 64,
279                sse, &sum, vpx_get16x16var_sse2, 16);
280  return *sse - (((int64_t)sum * sum) >> 11);
281}
282
283unsigned int vpx_mse8x8_sse2(const uint8_t *src, int src_stride,
284                             const uint8_t *ref, int ref_stride,
285                             unsigned int *sse) {
286  vpx_variance8x8_sse2(src, src_stride, ref, ref_stride, sse);
287  return *sse;
288}
289
290unsigned int vpx_mse8x16_sse2(const uint8_t *src, int src_stride,
291                              const uint8_t *ref, int ref_stride,
292                              unsigned int *sse) {
293  vpx_variance8x16_sse2(src, src_stride, ref, ref_stride, sse);
294  return *sse;
295}
296
297unsigned int vpx_mse16x8_sse2(const uint8_t *src, int src_stride,
298                              const uint8_t *ref, int ref_stride,
299                              unsigned int *sse) {
300  vpx_variance16x8_sse2(src, src_stride, ref, ref_stride, sse);
301  return *sse;
302}
303
304unsigned int vpx_mse16x16_sse2(const uint8_t *src, int src_stride,
305                               const uint8_t *ref, int ref_stride,
306                               unsigned int *sse) {
307  vpx_variance16x16_sse2(src, src_stride, ref, ref_stride, sse);
308  return *sse;
309}
310
311#if CONFIG_USE_X86INC
312// The 2 unused parameters are place holders for PIC enabled build.
313// These definitions are for functions defined in subpel_variance.asm
314#define DECL(w, opt) \
315  int vpx_sub_pixel_variance##w##xh_##opt(const uint8_t *src, \
316                                          ptrdiff_t src_stride, \
317                                          int x_offset, int y_offset, \
318                                          const uint8_t *dst, \
319                                          ptrdiff_t dst_stride, \
320                                          int height, unsigned int *sse, \
321                                          void *unused0, void *unused)
322#define DECLS(opt1, opt2) \
323  DECL(4, opt2); \
324  DECL(8, opt1); \
325  DECL(16, opt1)
326
327DECLS(sse2, sse);
328DECLS(ssse3, ssse3);
329#undef DECLS
330#undef DECL
331
332#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
333unsigned int vpx_sub_pixel_variance##w##x##h##_##opt(const uint8_t *src, \
334                                                     int src_stride, \
335                                                     int x_offset, \
336                                                     int y_offset, \
337                                                     const uint8_t *dst, \
338                                                     int dst_stride, \
339                                                     unsigned int *sse_ptr) { \
340  unsigned int sse; \
341  int se = vpx_sub_pixel_variance##wf##xh_##opt(src, src_stride, x_offset, \
342                                                y_offset, dst, dst_stride, \
343                                                h, &sse, NULL, NULL); \
344  if (w > wf) { \
345    unsigned int sse2; \
346    int se2 = vpx_sub_pixel_variance##wf##xh_##opt(src + 16, src_stride, \
347                                                   x_offset, y_offset, \
348                                                   dst + 16, dst_stride, \
349                                                   h, &sse2, NULL, NULL); \
350    se += se2; \
351    sse += sse2; \
352    if (w > wf * 2) { \
353      se2 = vpx_sub_pixel_variance##wf##xh_##opt(src + 32, src_stride, \
354                                                 x_offset, y_offset, \
355                                                 dst + 32, dst_stride, \
356                                                 h, &sse2, NULL, NULL); \
357      se += se2; \
358      sse += sse2; \
359      se2 = vpx_sub_pixel_variance##wf##xh_##opt(src + 48, src_stride, \
360                                                 x_offset, y_offset, \
361                                                 dst + 48, dst_stride, \
362                                                 h, &sse2, NULL, NULL); \
363      se += se2; \
364      sse += sse2; \
365    } \
366  } \
367  *sse_ptr = sse; \
368  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
369}
370
371#define FNS(opt1, opt2) \
372FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
373FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
374FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
375FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
376FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
377FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
378FN(16, 16, 16, 4, 4, opt1, (uint32_t)); \
379FN(16,  8, 16, 4, 3, opt1, (uint32_t)); \
380FN(8,  16,  8, 3, 4, opt1, (uint32_t)); \
381FN(8,   8,  8, 3, 3, opt1, (uint32_t)); \
382FN(8,   4,  8, 3, 2, opt1, (uint32_t)); \
383FN(4,   8,  4, 2, 3, opt2, (uint32_t)); \
384FN(4,   4,  4, 2, 2, opt2, (uint32_t))
385
386FNS(sse2, sse);
387FNS(ssse3, ssse3);
388
389#undef FNS
390#undef FN
391
392// The 2 unused parameters are place holders for PIC enabled build.
393#define DECL(w, opt) \
394int vpx_sub_pixel_avg_variance##w##xh_##opt(const uint8_t *src, \
395                                            ptrdiff_t src_stride, \
396                                            int x_offset, int y_offset, \
397                                            const uint8_t *dst, \
398                                            ptrdiff_t dst_stride, \
399                                            const uint8_t *sec, \
400                                            ptrdiff_t sec_stride, \
401                                            int height, unsigned int *sse, \
402                                            void *unused0, void *unused)
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 vpx_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 = vpx_sub_pixel_avg_variance##wf##xh_##opt(src, src_stride, x_offset, \
424                                                    y_offset, dst, dst_stride, \
425                                                    sec, w, h, &sse, NULL, \
426                                                    NULL); \
427  if (w > wf) { \
428    unsigned int sse2; \
429    int se2 = vpx_sub_pixel_avg_variance##wf##xh_##opt(src + 16, src_stride, \
430                                                       x_offset, y_offset, \
431                                                       dst + 16, dst_stride, \
432                                                       sec + 16, w, h, &sse2, \
433                                                       NULL, NULL); \
434    se += se2; \
435    sse += sse2; \
436    if (w > wf * 2) { \
437      se2 = vpx_sub_pixel_avg_variance##wf##xh_##opt(src + 32, src_stride, \
438                                                     x_offset, y_offset, \
439                                                     dst + 32, dst_stride, \
440                                                     sec + 32, w, h, &sse2, \
441                                                     NULL, NULL); \
442      se += se2; \
443      sse += sse2; \
444      se2 = vpx_sub_pixel_avg_variance##wf##xh_##opt(src + 48, src_stride, \
445                                                     x_offset, y_offset, \
446                                                     dst + 48, dst_stride, \
447                                                     sec + 48, w, h, &sse2, \
448                                                     NULL, NULL); \
449      se += se2; \
450      sse += sse2; \
451    } \
452  } \
453  *sseptr = sse; \
454  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
455}
456
457#define FNS(opt1, opt2) \
458FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
459FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
460FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
461FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
462FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
463FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
464FN(16, 16, 16, 4, 4, opt1, (uint32_t)); \
465FN(16,  8, 16, 4, 3, opt1, (uint32_t)); \
466FN(8,  16,  8, 3, 4, opt1, (uint32_t)); \
467FN(8,   8,  8, 3, 3, opt1, (uint32_t)); \
468FN(8,   4,  8, 3, 2, opt1, (uint32_t)); \
469FN(4,   8,  4, 2, 3, opt2, (uint32_t)); \
470FN(4,   4,  4, 2, 2, opt2, (uint32_t))
471
472FNS(sse2, sse);
473FNS(ssse3, ssse3);
474
475#undef FNS
476#undef FN
477#endif  // CONFIG_USE_X86INC
478