xref: /third_party/ffmpeg/tests/checkasm/vp9dsp.c (revision cabdff1a)
1/*
2 * Copyright (c) 2015 Ronald S. Bultje <rsbultje@gmail.com>
3 *
4 * This file is part of FFmpeg.
5 *
6 * FFmpeg is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License along
17 * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
19 */
20
21#include <math.h>
22#include <string.h>
23#include "checkasm.h"
24#include "libavcodec/vp9data.h"
25#include "libavcodec/vp9.h"
26#include "libavutil/common.h"
27#include "libavutil/internal.h"
28#include "libavutil/intreadwrite.h"
29#include "libavutil/mathematics.h"
30#include "libavutil/mem_internal.h"
31
32static const uint32_t pixel_mask[3] = { 0xffffffff, 0x03ff03ff, 0x0fff0fff };
33#define SIZEOF_PIXEL ((bit_depth + 7) / 8)
34
35#define randomize_buffers()                                        \
36    do {                                                           \
37        uint32_t mask = pixel_mask[(bit_depth - 8) >> 1];          \
38        int k;                                                     \
39        for (k = -4;  k < SIZEOF_PIXEL * FFMAX(8, size); k += 4) { \
40            uint32_t r = rnd() & mask;                             \
41            AV_WN32A(a + k, r);                                    \
42        }                                                          \
43        for (k = 0; k < size * SIZEOF_PIXEL; k += 4) {             \
44            uint32_t r = rnd() & mask;                             \
45            AV_WN32A(l + k, r);                                    \
46        }                                                          \
47    } while (0)
48
49static void check_ipred(void)
50{
51    LOCAL_ALIGNED_32(uint8_t, a_buf, [64 * 2]);
52    uint8_t *a = &a_buf[32 * 2];
53    LOCAL_ALIGNED_32(uint8_t, l, [32 * 2]);
54    LOCAL_ALIGNED_32(uint8_t, dst0, [32 * 32 * 2]);
55    LOCAL_ALIGNED_32(uint8_t, dst1, [32 * 32 * 2]);
56    VP9DSPContext dsp;
57    int tx, mode, bit_depth;
58    declare_func_emms(AV_CPU_FLAG_MMX | AV_CPU_FLAG_MMXEXT, void, uint8_t *dst, ptrdiff_t stride,
59                      const uint8_t *left, const uint8_t *top);
60    static const char *const mode_names[N_INTRA_PRED_MODES] = {
61        [VERT_PRED] = "vert",
62        [HOR_PRED] = "hor",
63        [DC_PRED] = "dc",
64        [DIAG_DOWN_LEFT_PRED] = "diag_downleft",
65        [DIAG_DOWN_RIGHT_PRED] = "diag_downright",
66        [VERT_RIGHT_PRED] = "vert_right",
67        [HOR_DOWN_PRED] = "hor_down",
68        [VERT_LEFT_PRED] = "vert_left",
69        [HOR_UP_PRED] = "hor_up",
70        [TM_VP8_PRED] = "tm",
71        [LEFT_DC_PRED] = "dc_left",
72        [TOP_DC_PRED] = "dc_top",
73        [DC_128_PRED] = "dc_128",
74        [DC_127_PRED] = "dc_127",
75        [DC_129_PRED] = "dc_129",
76    };
77
78    for (bit_depth = 8; bit_depth <= 12; bit_depth += 2) {
79        ff_vp9dsp_init(&dsp, bit_depth, 0);
80        for (tx = 0; tx < 4; tx++) {
81            int size = 4 << tx;
82
83            for (mode = 0; mode < N_INTRA_PRED_MODES; mode++) {
84                if (check_func(dsp.intra_pred[tx][mode], "vp9_%s_%dx%d_%dbpp",
85                               mode_names[mode], size, size, bit_depth)) {
86                    randomize_buffers();
87                    call_ref(dst0, size * SIZEOF_PIXEL, l, a);
88                    call_new(dst1, size * SIZEOF_PIXEL, l, a);
89                    if (memcmp(dst0, dst1, size * size * SIZEOF_PIXEL))
90                        fail();
91                    bench_new(dst1, size * SIZEOF_PIXEL,l, a);
92                }
93            }
94        }
95    }
96    report("ipred");
97}
98
99#undef randomize_buffers
100
101#define randomize_buffers() \
102    do { \
103        uint32_t mask = pixel_mask[(bit_depth - 8) >> 1];                  \
104        for (y = 0; y < sz; y++) {                                         \
105            for (x = 0; x < sz * SIZEOF_PIXEL; x += 4) {                   \
106                uint32_t r = rnd() & mask;                                 \
107                AV_WN32A(dst + y * sz * SIZEOF_PIXEL + x, r);              \
108                AV_WN32A(src + y * sz * SIZEOF_PIXEL + x, rnd() & mask);   \
109            }                                                              \
110            for (x = 0; x < sz; x++) {                                     \
111                if (bit_depth == 8) {                                      \
112                    coef[y * sz + x] = src[y * sz + x] - dst[y * sz + x];  \
113                } else {                                                   \
114                    ((int32_t *) coef)[y * sz + x] =                       \
115                        ((uint16_t *) src)[y * sz + x] -                   \
116                        ((uint16_t *) dst)[y * sz + x];                    \
117                }                                                          \
118            }                                                              \
119        }                                                                  \
120    } while(0)
121
122// wht function copied from libvpx
123static void fwht_1d(double *out, const double *in, int sz)
124{
125    double t0 = in[0] + in[1];
126    double t3 = in[3] - in[2];
127    double t4 = trunc((t0 - t3) * 0.5);
128    double t1 = t4 - in[1];
129    double t2 = t4 - in[2];
130
131    out[0] = t0 - t2;
132    out[1] = t2;
133    out[2] = t3 + t1;
134    out[3] = t1;
135}
136
137// standard DCT-II
138static void fdct_1d(double *out, const double *in, int sz)
139{
140    int k, n;
141
142    for (k = 0; k < sz; k++) {
143        out[k] = 0.0;
144        for (n = 0; n < sz; n++)
145            out[k] += in[n] * cos(M_PI * (2 * n + 1) * k / (sz * 2.0));
146    }
147    out[0] *= M_SQRT1_2;
148}
149
150// see "Towards jointly optimal spatial prediction and adaptive transform in
151// video/image coding", by J. Han, A. Saxena, and K. Rose
152// IEEE Proc. ICASSP, pp. 726-729, Mar. 2010.
153static void fadst4_1d(double *out, const double *in, int sz)
154{
155    int k, n;
156
157    for (k = 0; k < sz; k++) {
158        out[k] = 0.0;
159        for (n = 0; n < sz; n++)
160            out[k] += in[n] * sin(M_PI * (n + 1) * (2 * k + 1) / (sz * 2.0 + 1.0));
161    }
162}
163
164// see "A Butterfly Structured Design of The Hybrid Transform Coding Scheme",
165// by Jingning Han, Yaowu Xu, and Debargha Mukherjee
166// http://static.googleusercontent.com/media/research.google.com/en//pubs/archive/41418.pdf
167static void fadst_1d(double *out, const double *in, int sz)
168{
169    int k, n;
170
171    for (k = 0; k < sz; k++) {
172        out[k] = 0.0;
173        for (n = 0; n < sz; n++)
174            out[k] += in[n] * sin(M_PI * (2 * n + 1) * (2 * k + 1) / (sz * 4.0));
175    }
176}
177
178typedef void (*ftx1d_fn)(double *out, const double *in, int sz);
179static void ftx_2d(double *out, const double *in, enum TxfmMode tx,
180                   enum TxfmType txtp, int sz)
181{
182    static const double scaling_factors[5][4] = {
183        { 4.0, 16.0 * M_SQRT1_2 / 3.0, 16.0 * M_SQRT1_2 / 3.0, 32.0 / 9.0 },
184        { 2.0, 2.0, 2.0, 2.0 },
185        { 1.0, 1.0, 1.0, 1.0 },
186        { 0.25 },
187        { 4.0 }
188    };
189    static const ftx1d_fn ftx1d_tbl[5][4][2] = {
190        {
191            { fdct_1d, fdct_1d },
192            { fadst4_1d, fdct_1d },
193            { fdct_1d, fadst4_1d },
194            { fadst4_1d, fadst4_1d },
195        }, {
196            { fdct_1d, fdct_1d },
197            { fadst_1d, fdct_1d },
198            { fdct_1d, fadst_1d },
199            { fadst_1d, fadst_1d },
200        }, {
201            { fdct_1d, fdct_1d },
202            { fadst_1d, fdct_1d },
203            { fdct_1d, fadst_1d },
204            { fadst_1d, fadst_1d },
205        }, {
206            { fdct_1d, fdct_1d },
207        }, {
208            { fwht_1d, fwht_1d },
209        },
210    };
211    double temp[1024];
212    double scaling_factor = scaling_factors[tx][txtp];
213    int i, j;
214
215    // cols
216    for (i = 0; i < sz; ++i) {
217        double temp_out[32];
218
219        ftx1d_tbl[tx][txtp][0](temp_out, &in[i * sz], sz);
220        // scale and transpose
221        for (j = 0; j < sz; ++j)
222            temp[j * sz + i] = temp_out[j] * scaling_factor;
223    }
224
225    // rows
226    for (i = 0; i < sz; i++)
227        ftx1d_tbl[tx][txtp][1](&out[i * sz], &temp[i * sz], sz);
228}
229
230static void ftx(int16_t *buf, enum TxfmMode tx,
231                enum TxfmType txtp, int sz, int bit_depth)
232{
233    double ind[1024], outd[1024];
234    int n;
235
236    emms_c();
237    for (n = 0; n < sz * sz; n++) {
238        if (bit_depth == 8)
239            ind[n] = buf[n];
240        else
241            ind[n] = ((int32_t *) buf)[n];
242    }
243    ftx_2d(outd, ind, tx, txtp, sz);
244    for (n = 0; n < sz * sz; n++) {
245        if (bit_depth == 8)
246            buf[n] = lrint(outd[n]);
247        else
248            ((int32_t *) buf)[n] = lrint(outd[n]);
249    }
250}
251
252static int copy_subcoefs(int16_t *out, const int16_t *in, enum TxfmMode tx,
253                         enum TxfmType txtp, int sz, int sub, int bit_depth)
254{
255    // copy the topleft coefficients such that the return value (being the
256    // coefficient scantable index for the eob token) guarantees that only
257    // the topleft $sub out of $sz (where $sz >= $sub) coefficients in both
258    // dimensions are non-zero. This leads to braching to specific optimized
259    // simd versions (e.g. dc-only) so that we get full asm coverage in this
260    // test
261
262    int n;
263    const int16_t *scan = ff_vp9_scans[tx][txtp];
264    int eob;
265
266    for (n = 0; n < sz * sz; n++) {
267        int rc = scan[n], rcx = rc % sz, rcy = rc / sz;
268
269        // find eob for this sub-idct
270        if (rcx >= sub || rcy >= sub)
271            break;
272
273        // copy coef
274        if (bit_depth == 8) {
275            out[rc] = in[rc];
276        } else {
277            AV_COPY32(&out[rc * 2], &in[rc * 2]);
278        }
279    }
280
281    eob = n;
282
283    for (; n < sz * sz; n++) {
284        int rc = scan[n];
285
286        // zero
287        if (bit_depth == 8) {
288            out[rc] = 0;
289        } else {
290            AV_ZERO32(&out[rc * 2]);
291        }
292    }
293
294    return eob;
295}
296
297static int is_zero(const int16_t *c, int sz)
298{
299    int n;
300
301    for (n = 0; n < sz / sizeof(int16_t); n += 2)
302        if (AV_RN32A(&c[n]))
303            return 0;
304
305    return 1;
306}
307
308#define SIZEOF_COEF (2 * ((bit_depth + 7) / 8))
309
310static void check_itxfm(void)
311{
312    LOCAL_ALIGNED_32(uint8_t, src, [32 * 32 * 2]);
313    LOCAL_ALIGNED_32(uint8_t, dst, [32 * 32 * 2]);
314    LOCAL_ALIGNED_32(uint8_t, dst0, [32 * 32 * 2]);
315    LOCAL_ALIGNED_32(uint8_t, dst1, [32 * 32 * 2]);
316    LOCAL_ALIGNED_32(int16_t, coef, [32 * 32 * 2]);
317    LOCAL_ALIGNED_32(int16_t, subcoef0, [32 * 32 * 2]);
318    LOCAL_ALIGNED_32(int16_t, subcoef1, [32 * 32 * 2]);
319    declare_func_emms(AV_CPU_FLAG_MMX | AV_CPU_FLAG_MMXEXT, void, uint8_t *dst, ptrdiff_t stride, int16_t *block, int eob);
320    VP9DSPContext dsp;
321    int y, x, tx, txtp, bit_depth, sub;
322    static const char *const txtp_types[N_TXFM_TYPES] = {
323        [DCT_DCT] = "dct_dct", [DCT_ADST] = "adst_dct",
324        [ADST_DCT] = "dct_adst", [ADST_ADST] = "adst_adst"
325    };
326
327    for (bit_depth = 8; bit_depth <= 12; bit_depth += 2) {
328        ff_vp9dsp_init(&dsp, bit_depth, 0);
329
330        for (tx = TX_4X4; tx <= N_TXFM_SIZES /* 4 = lossless */; tx++) {
331            int sz = 4 << (tx & 3);
332            int n_txtps = tx < TX_32X32 ? N_TXFM_TYPES : 1;
333
334            for (txtp = 0; txtp < n_txtps; txtp++) {
335                // skip testing sub-IDCTs for WHT or ADST since they don't
336                // implement it in any of the SIMD functions. If they do,
337                // consider changing this to ensure we have complete test
338                // coverage. Test sub=1 for dc-only, then 2, 4, 8, 12, etc,
339                // since the arm version can distinguish them at that level.
340                for (sub = (txtp == 0 && tx < 4) ? 1 : sz; sub <= sz;
341                     sub < 4 ? (sub <<= 1) : (sub += 4)) {
342                    if (check_func(dsp.itxfm_add[tx][txtp],
343                                   "vp9_inv_%s_%dx%d_sub%d_add_%d",
344                                   tx == 4 ? "wht_wht" : txtp_types[txtp],
345                                   sz, sz, sub, bit_depth)) {
346                        int eob;
347
348                        randomize_buffers();
349                        ftx(coef, tx, txtp, sz, bit_depth);
350
351                        if (sub < sz) {
352                            eob = copy_subcoefs(subcoef0, coef, tx, txtp,
353                                                sz, sub, bit_depth);
354                        } else {
355                            eob = sz * sz;
356                            memcpy(subcoef0, coef, sz * sz * SIZEOF_COEF);
357                        }
358
359                        memcpy(dst0, dst, sz * sz * SIZEOF_PIXEL);
360                        memcpy(dst1, dst, sz * sz * SIZEOF_PIXEL);
361                        memcpy(subcoef1, subcoef0, sz * sz * SIZEOF_COEF);
362                        call_ref(dst0, sz * SIZEOF_PIXEL, subcoef0, eob);
363                        call_new(dst1, sz * SIZEOF_PIXEL, subcoef1, eob);
364                        if (memcmp(dst0, dst1, sz * sz * SIZEOF_PIXEL) ||
365                            !is_zero(subcoef0, sz * sz * SIZEOF_COEF) ||
366                            !is_zero(subcoef1, sz * sz * SIZEOF_COEF))
367                            fail();
368
369                        bench_new(dst, sz * SIZEOF_PIXEL, coef, eob);
370                    }
371                }
372            }
373        }
374    }
375    report("itxfm");
376}
377
378#undef randomize_buffers
379
380#define setpx(a,b,c) \
381    do { \
382        if (SIZEOF_PIXEL == 1) { \
383            buf0[(a) + (b) * jstride] = av_clip_uint8(c); \
384        } else { \
385            ((uint16_t *)buf0)[(a) + (b) * jstride] = av_clip_uintp2(c, bit_depth); \
386        } \
387    } while (0)
388
389// c can be an assignment and must not be put under ()
390#define setdx(a,b,c,d) setpx(a,b,c-(d)+(rnd()%((d)*2+1)))
391#define setsx(a,b,c,d) setdx(a,b,c,(d) << (bit_depth - 8))
392static void randomize_loopfilter_buffers(int bidx, int lineoff, int str,
393                                         int bit_depth, int dir, const int *E,
394                                         const int *F, const int *H, const int *I,
395                                         uint8_t *buf0, uint8_t *buf1)
396{
397    uint32_t mask = (1 << bit_depth) - 1;
398    int off = dir ? lineoff : lineoff * 16;
399    int istride = dir ? 1 : 16;
400    int jstride = dir ? str : 1;
401    int i, j;
402    for (i = 0; i < 2; i++) /* flat16 */ {
403        int idx = off + i * istride, p0, q0;
404        setpx(idx,  0, q0 = rnd() & mask);
405        setsx(idx, -1, p0 = q0, E[bidx] >> 2);
406        for (j = 1; j < 8; j++) {
407            setsx(idx, -1 - j, p0, F[bidx]);
408            setsx(idx, j, q0, F[bidx]);
409        }
410    }
411    for (i = 2; i < 4; i++) /* flat8 */ {
412        int idx = off + i * istride, p0, q0;
413        setpx(idx,  0, q0 = rnd() & mask);
414        setsx(idx, -1, p0 = q0, E[bidx] >> 2);
415        for (j = 1; j < 4; j++) {
416            setsx(idx, -1 - j, p0, F[bidx]);
417            setsx(idx, j, q0, F[bidx]);
418        }
419        for (j = 4; j < 8; j++) {
420            setpx(idx, -1 - j, rnd() & mask);
421            setpx(idx, j, rnd() & mask);
422        }
423    }
424    for (i = 4; i < 6; i++) /* regular */ {
425        int idx = off + i * istride, p2, p1, p0, q0, q1, q2;
426        setpx(idx,  0, q0 = rnd() & mask);
427        setsx(idx,  1, q1 = q0, I[bidx]);
428        setsx(idx,  2, q2 = q1, I[bidx]);
429        setsx(idx,  3, q2,      I[bidx]);
430        setsx(idx, -1, p0 = q0, E[bidx] >> 2);
431        setsx(idx, -2, p1 = p0, I[bidx]);
432        setsx(idx, -3, p2 = p1, I[bidx]);
433        setsx(idx, -4, p2,      I[bidx]);
434        for (j = 4; j < 8; j++) {
435            setpx(idx, -1 - j, rnd() & mask);
436            setpx(idx, j, rnd() & mask);
437        }
438    }
439    for (i = 6; i < 8; i++) /* off */ {
440        int idx = off + i * istride;
441        for (j = 0; j < 8; j++) {
442            setpx(idx, -1 - j, rnd() & mask);
443            setpx(idx, j, rnd() & mask);
444        }
445    }
446}
447#define randomize_buffers(bidx, lineoff, str) \
448        randomize_loopfilter_buffers(bidx, lineoff, str, bit_depth, dir, \
449                                     E, F, H, I, buf0, buf1)
450
451static void check_loopfilter(void)
452{
453    LOCAL_ALIGNED_32(uint8_t, base0, [32 + 16 * 16 * 2]);
454    LOCAL_ALIGNED_32(uint8_t, base1, [32 + 16 * 16 * 2]);
455    VP9DSPContext dsp;
456    int dir, wd, wd2, bit_depth;
457    static const char *const dir_name[2] = { "h", "v" };
458    static const int E[2] = { 20, 28 }, I[2] = { 10, 16 };
459    static const int H[2] = { 7, 11 }, F[2] = { 1, 1 };
460    declare_func_emms(AV_CPU_FLAG_MMX | AV_CPU_FLAG_MMXEXT, void, uint8_t *dst, ptrdiff_t stride, int E, int I, int H);
461
462    for (bit_depth = 8; bit_depth <= 12; bit_depth += 2) {
463        ff_vp9dsp_init(&dsp, bit_depth, 0);
464
465        for (dir = 0; dir < 2; dir++) {
466            int midoff = (dir ? 8 * 8 : 8) * SIZEOF_PIXEL;
467            int midoff_aligned = (dir ? 8 * 8 : 16) * SIZEOF_PIXEL;
468            uint8_t *buf0 = base0 + midoff_aligned;
469            uint8_t *buf1 = base1 + midoff_aligned;
470
471            for (wd = 0; wd < 3; wd++) {
472                // 4/8/16wd_8px
473                if (check_func(dsp.loop_filter_8[wd][dir],
474                               "vp9_loop_filter_%s_%d_8_%dbpp",
475                               dir_name[dir], 4 << wd, bit_depth)) {
476                    randomize_buffers(0, 0, 8);
477                    memcpy(buf1 - midoff, buf0 - midoff,
478                           16 * 8 * SIZEOF_PIXEL);
479                    call_ref(buf0, 16 * SIZEOF_PIXEL >> dir, E[0], I[0], H[0]);
480                    call_new(buf1, 16 * SIZEOF_PIXEL >> dir, E[0], I[0], H[0]);
481                    if (memcmp(buf0 - midoff, buf1 - midoff, 16 * 8 * SIZEOF_PIXEL))
482                        fail();
483                    bench_new(buf1, 16 * SIZEOF_PIXEL >> dir, E[0], I[0], H[0]);
484                }
485            }
486
487            midoff = (dir ? 16 * 8 : 8) * SIZEOF_PIXEL;
488            midoff_aligned = (dir ? 16 * 8 : 16) * SIZEOF_PIXEL;
489
490            buf0 = base0 + midoff_aligned;
491            buf1 = base1 + midoff_aligned;
492
493            // 16wd_16px loopfilter
494            if (check_func(dsp.loop_filter_16[dir],
495                           "vp9_loop_filter_%s_16_16_%dbpp",
496                           dir_name[dir], bit_depth)) {
497                randomize_buffers(0, 0, 16);
498                randomize_buffers(0, 8, 16);
499                memcpy(buf1 - midoff, buf0 - midoff, 16 * 16 * SIZEOF_PIXEL);
500                call_ref(buf0, 16 * SIZEOF_PIXEL, E[0], I[0], H[0]);
501                call_new(buf1, 16 * SIZEOF_PIXEL, E[0], I[0], H[0]);
502                if (memcmp(buf0 - midoff, buf1 - midoff, 16 * 16 * SIZEOF_PIXEL))
503                    fail();
504                bench_new(buf1, 16 * SIZEOF_PIXEL, E[0], I[0], H[0]);
505            }
506
507            for (wd = 0; wd < 2; wd++) {
508                for (wd2 = 0; wd2 < 2; wd2++) {
509                    // mix2 loopfilter
510                    if (check_func(dsp.loop_filter_mix2[wd][wd2][dir],
511                                   "vp9_loop_filter_mix2_%s_%d%d_16_%dbpp",
512                                   dir_name[dir], 4 << wd, 4 << wd2, bit_depth)) {
513                        randomize_buffers(0, 0, 16);
514                        randomize_buffers(1, 8, 16);
515                        memcpy(buf1 - midoff, buf0 - midoff, 16 * 16 * SIZEOF_PIXEL);
516#define M(a) (((a)[1] << 8) | (a)[0])
517                        call_ref(buf0, 16 * SIZEOF_PIXEL, M(E), M(I), M(H));
518                        call_new(buf1, 16 * SIZEOF_PIXEL, M(E), M(I), M(H));
519                        if (memcmp(buf0 - midoff, buf1 - midoff, 16 * 16 * SIZEOF_PIXEL))
520                            fail();
521                        bench_new(buf1, 16 * SIZEOF_PIXEL, M(E), M(I), M(H));
522#undef M
523                    }
524                }
525            }
526        }
527    }
528    report("loopfilter");
529}
530
531#undef setsx
532#undef setpx
533#undef setdx
534#undef randomize_buffers
535
536#define DST_BUF_SIZE (size * size * SIZEOF_PIXEL)
537#define SRC_BUF_STRIDE 72
538#define SRC_BUF_SIZE ((size + 7) * SRC_BUF_STRIDE * SIZEOF_PIXEL)
539#define src (buf + 3 * SIZEOF_PIXEL * (SRC_BUF_STRIDE + 1))
540
541#define randomize_buffers()                               \
542    do {                                                  \
543        uint32_t mask = pixel_mask[(bit_depth - 8) >> 1]; \
544        int k;                                            \
545        for (k = 0; k < SRC_BUF_SIZE; k += 4) {           \
546            uint32_t r = rnd() & mask;                    \
547            AV_WN32A(buf + k, r);                         \
548        }                                                 \
549        if (op == 1) {                                    \
550            for (k = 0; k < DST_BUF_SIZE; k += 4) {       \
551                uint32_t r = rnd() & mask;                \
552                AV_WN32A(dst0 + k, r);                    \
553                AV_WN32A(dst1 + k, r);                    \
554            }                                             \
555        }                                                 \
556    } while (0)
557
558static void check_mc(void)
559{
560    LOCAL_ALIGNED_32(uint8_t, buf, [72 * 72 * 2]);
561    LOCAL_ALIGNED_32(uint8_t, dst0, [64 * 64 * 2]);
562    LOCAL_ALIGNED_32(uint8_t, dst1, [64 * 64 * 2]);
563    VP9DSPContext dsp;
564    int op, hsize, bit_depth, filter, dx, dy;
565    declare_func_emms(AV_CPU_FLAG_MMX | AV_CPU_FLAG_MMXEXT, void, uint8_t *dst, ptrdiff_t dst_stride,
566                      const uint8_t *ref, ptrdiff_t ref_stride,
567                 int h, int mx, int my);
568    static const char *const filter_names[4] = {
569        "8tap_smooth", "8tap_regular", "8tap_sharp", "bilin"
570    };
571    static const char *const subpel_names[2][2] = { { "", "h" }, { "v", "hv" } };
572    static const char *const op_names[2] = { "put", "avg" };
573    char str[256];
574
575    for (op = 0; op < 2; op++) {
576        for (bit_depth = 8; bit_depth <= 12; bit_depth += 2) {
577            ff_vp9dsp_init(&dsp, bit_depth, 0);
578            for (hsize = 0; hsize < 5; hsize++) {
579                int size = 64 >> hsize;
580
581                for (filter = 0; filter < 4; filter++) {
582                    for (dx = 0; dx < 2; dx++) {
583                        for (dy = 0; dy < 2; dy++) {
584                            if (dx || dy) {
585                                snprintf(str, sizeof(str),
586                                         "%s_%s_%d%s", op_names[op],
587                                         filter_names[filter], size,
588                                         subpel_names[dy][dx]);
589                            } else {
590                                snprintf(str, sizeof(str),
591                                         "%s%d", op_names[op], size);
592                            }
593                            if (check_func(dsp.mc[hsize][filter][op][dx][dy],
594                                           "vp9_%s_%dbpp", str, bit_depth)) {
595                                int mx = dx ? 1 + (rnd() % 14) : 0;
596                                int my = dy ? 1 + (rnd() % 14) : 0;
597                                randomize_buffers();
598                                call_ref(dst0, size * SIZEOF_PIXEL,
599                                         src, SRC_BUF_STRIDE * SIZEOF_PIXEL,
600                                         size, mx, my);
601                                call_new(dst1, size * SIZEOF_PIXEL,
602                                         src, SRC_BUF_STRIDE * SIZEOF_PIXEL,
603                                         size, mx, my);
604                                if (memcmp(dst0, dst1, DST_BUF_SIZE))
605                                    fail();
606
607                                // simd implementations for each filter of subpel
608                                // functions are identical
609                                if (filter >= 1 && filter <= 2) continue;
610                                // 10/12 bpp for bilin are identical
611                                if (bit_depth == 12 && filter == 3) continue;
612
613                                bench_new(dst1, size * SIZEOF_PIXEL,
614                                          src, SRC_BUF_STRIDE * SIZEOF_PIXEL,
615                                          size, mx, my);
616                            }
617                        }
618                    }
619                }
620            }
621        }
622    }
623    report("mc");
624}
625
626void checkasm_check_vp9dsp(void)
627{
628    check_ipred();
629    check_itxfm();
630    check_loopfilter();
631    check_mc();
632}
633