Merge commit 'd7f4f5c4a18a0c9e62635cfa6fe8a9302b413c01'
[ffmpeg.git] / tests / checkasm / h264dsp.c
1 /*
2  * Copyright (c) 2016 Martin Storsjo
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 <string.h>
22 #include "checkasm.h"
23 #include "libavcodec/avcodec.h"
24 #include "libavcodec/h264dsp.h"
25 #include "libavcodec/h264data.h"
26 #include "libavutil/common.h"
27 #include "libavutil/internal.h"
28 #include "libavutil/intreadwrite.h"
29
30 static const uint32_t pixel_mask[3] = { 0xffffffff, 0x01ff01ff, 0x03ff03ff };
31 static const uint32_t pixel_mask_lf[3] = { 0xff0fff0f, 0x01ff000f, 0x03ff000f };
32
33 #define SIZEOF_PIXEL ((bit_depth + 7) / 8)
34 #define SIZEOF_COEF  (2 * ((bit_depth + 7) / 8))
35 #define PIXEL_STRIDE 16
36
37 #define randomize_buffers()                                                  \
38     do {                                                                     \
39         int x, y;                                                            \
40         uint32_t mask = pixel_mask[bit_depth - 8];                           \
41         for (y = 0; y < sz; y++) {                                           \
42             for (x = 0; x < PIXEL_STRIDE; x += 4) {                          \
43                 AV_WN32A(src + y * PIXEL_STRIDE + x, rnd() & mask);          \
44                 AV_WN32A(dst + y * PIXEL_STRIDE + x, rnd() & mask);          \
45             }                                                                \
46             for (x = 0; x < sz; x++) {                                       \
47                 if (bit_depth == 8) {                                        \
48                     coef[y * sz + x] = src[y * PIXEL_STRIDE + x] -           \
49                                        dst[y * PIXEL_STRIDE + x];            \
50                 } else {                                                     \
51                     ((int32_t *)coef)[y * sz + x] =                          \
52                         ((uint16_t *)src)[y * (PIXEL_STRIDE/2) + x] -        \
53                         ((uint16_t *)dst)[y * (PIXEL_STRIDE/2) + x];         \
54                 }                                                            \
55             }                                                                \
56         }                                                                    \
57     } while (0)
58
59 #define dct4x4_impl(size, dctcoef)                                           \
60 static void dct4x4_##size(dctcoef *coef)                                     \
61 {                                                                            \
62     int i, y, x;                                                             \
63     dctcoef tmp[16];                                                         \
64     for (i = 0; i < 4; i++) {                                                \
65         const int z0 = coef[i*4 + 0] + coef[i*4 + 3];                        \
66         const int z1 = coef[i*4 + 1] + coef[i*4 + 2];                        \
67         const int z2 = coef[i*4 + 0] - coef[i*4 + 3];                        \
68         const int z3 = coef[i*4 + 1] - coef[i*4 + 2];                        \
69         tmp[i + 4*0] =   z0 +   z1;                                          \
70         tmp[i + 4*1] = 2*z2 +   z3;                                          \
71         tmp[i + 4*2] =   z0 -   z1;                                          \
72         tmp[i + 4*3] =   z2 - 2*z3;                                          \
73     }                                                                        \
74     for (i = 0; i < 4; i++) {                                                \
75         const int z0 = tmp[i*4 + 0] + tmp[i*4 + 3];                          \
76         const int z1 = tmp[i*4 + 1] + tmp[i*4 + 2];                          \
77         const int z2 = tmp[i*4 + 0] - tmp[i*4 + 3];                          \
78         const int z3 = tmp[i*4 + 1] - tmp[i*4 + 2];                          \
79         coef[i*4 + 0] =   z0 +   z1;                                         \
80         coef[i*4 + 1] = 2*z2 +   z3;                                         \
81         coef[i*4 + 2] =   z0 -   z1;                                         \
82         coef[i*4 + 3] =   z2 - 2*z3;                                         \
83     }                                                                        \
84     for (y = 0; y < 4; y++) {                                                \
85         for (x = 0; x < 4; x++) {                                            \
86             static const int scale[] = { 13107 * 10, 8066 * 13, 5243 * 16 }; \
87             const int idx = (y & 1) + (x & 1);                               \
88             coef[y*4 + x] = (coef[y*4 + x] * scale[idx] + (1 << 14)) >> 15;  \
89         }                                                                    \
90     }                                                                        \
91 }
92
93 #define DCT8_1D(src, srcstride, dst, dststride) do {                         \
94     const int a0 = (src)[srcstride * 0] + (src)[srcstride * 7];              \
95     const int a1 = (src)[srcstride * 0] - (src)[srcstride * 7];              \
96     const int a2 = (src)[srcstride * 1] + (src)[srcstride * 6];              \
97     const int a3 = (src)[srcstride * 1] - (src)[srcstride * 6];              \
98     const int a4 = (src)[srcstride * 2] + (src)[srcstride * 5];              \
99     const int a5 = (src)[srcstride * 2] - (src)[srcstride * 5];              \
100     const int a6 = (src)[srcstride * 3] + (src)[srcstride * 4];              \
101     const int a7 = (src)[srcstride * 3] - (src)[srcstride * 4];              \
102     const int b0 = a0 + a6;                                                  \
103     const int b1 = a2 + a4;                                                  \
104     const int b2 = a0 - a6;                                                  \
105     const int b3 = a2 - a4;                                                  \
106     const int b4 = a3 + a5 + (a1 + (a1 >> 1));                               \
107     const int b5 = a1 - a7 - (a5 + (a5 >> 1));                               \
108     const int b6 = a1 + a7 - (a3 + (a3 >> 1));                               \
109     const int b7 = a3 - a5 + (a7 + (a7 >> 1));                               \
110     (dst)[dststride * 0] =  b0 +  b1;                                        \
111     (dst)[dststride * 1] =  b4 + (b7 >> 2);                                  \
112     (dst)[dststride * 2] =  b2 + (b3 >> 1);                                  \
113     (dst)[dststride * 3] =  b5 + (b6 >> 2);                                  \
114     (dst)[dststride * 4] =  b0  - b1;                                        \
115     (dst)[dststride * 5] =  b6 - (b5 >> 2);                                  \
116     (dst)[dststride * 6] = (b2 >> 1) - b3;                                   \
117     (dst)[dststride * 7] = (b4 >> 2) - b7;                                   \
118 } while (0)
119
120 #define dct8x8_impl(size, dctcoef)                                           \
121 static void dct8x8_##size(dctcoef *coef)                                     \
122 {                                                                            \
123     int i, x, y;                                                             \
124     dctcoef tmp[64];                                                         \
125     for (i = 0; i < 8; i++)                                                  \
126         DCT8_1D(coef + i, 8, tmp + i, 8);                                    \
127                                                                              \
128     for (i = 0; i < 8; i++)                                                  \
129         DCT8_1D(tmp + 8*i, 1, coef + i, 8);                                  \
130                                                                              \
131     for (y = 0; y < 8; y++) {                                                \
132         for (x = 0; x < 8; x++) {                                            \
133             static const int scale[] = {                                     \
134                 13107 * 20, 11428 * 18, 20972 * 32,                          \
135                 12222 * 19, 16777 * 25, 15481 * 24,                          \
136             };                                                               \
137             static const int idxmap[] = {                                    \
138                 0, 3, 4, 3,                                                  \
139                 3, 1, 5, 1,                                                  \
140                 4, 5, 2, 5,                                                  \
141                 3, 1, 5, 1,                                                  \
142             };                                                               \
143             const int idx = idxmap[(y & 3) * 4 + (x & 3)];                   \
144             coef[y*8 + x] = ((int64_t)coef[y*8 + x] *                        \
145                              scale[idx] + (1 << 17)) >> 18;                  \
146         }                                                                    \
147     }                                                                        \
148 }
149
150 dct4x4_impl(16, int16_t)
151 dct4x4_impl(32, int32_t)
152
153 dct8x8_impl(16, int16_t)
154 dct8x8_impl(32, int32_t)
155
156 static void dct4x4(int16_t *coef, int bit_depth)
157 {
158     if (bit_depth == 8)
159         dct4x4_16(coef);
160     else
161         dct4x4_32((int32_t *) coef);
162 }
163
164 static void dct8x8(int16_t *coef, int bit_depth)
165 {
166     if (bit_depth == 8) {
167         dct8x8_16(coef);
168     } else {
169         dct8x8_32((int32_t *) coef);
170     }
171 }
172
173
174 static void check_idct(void)
175 {
176     LOCAL_ALIGNED_16(uint8_t, src,  [8 * 8 * 2]);
177     LOCAL_ALIGNED_16(uint8_t, dst,  [8 * 8 * 2]);
178     LOCAL_ALIGNED_16(uint8_t, dst0, [8 * 8 * 2]);
179     LOCAL_ALIGNED_16(uint8_t, dst1_base, [8 * 8 * 2 + 32]);
180     LOCAL_ALIGNED_16(int16_t, coef, [8 * 8 * 2]);
181     LOCAL_ALIGNED_16(int16_t, subcoef0, [8 * 8 * 2]);
182     LOCAL_ALIGNED_16(int16_t, subcoef1, [8 * 8 * 2]);
183     H264DSPContext h;
184     int bit_depth, sz, align, dc;
185     declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t *block, int stride);
186
187     for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
188         ff_h264dsp_init(&h, bit_depth, 1);
189         for (sz = 4; sz <= 8; sz += 4) {
190             randomize_buffers();
191
192             if (sz == 4)
193                 dct4x4(coef, bit_depth);
194             else
195                 dct8x8(coef, bit_depth);
196
197             for (dc = 0; dc <= 1; dc++) {
198                 void (*idct)(uint8_t *, int16_t *, int) = NULL;
199                 switch ((sz << 1) | dc) {
200                 case (4 << 1) | 0: idct = h.h264_idct_add; break;
201                 case (4 << 1) | 1: idct = h.h264_idct_dc_add; break;
202                 case (8 << 1) | 0: idct = h.h264_idct8_add; break;
203                 case (8 << 1) | 1: idct = h.h264_idct8_dc_add; break;
204                 }
205                 if (check_func(idct, "h264_idct%d_add%s_%dbpp", sz, dc ? "_dc" : "", bit_depth)) {
206                     for (align = 0; align < 16; align += sz * SIZEOF_PIXEL) {
207                         uint8_t *dst1 = dst1_base + align;
208                         if (dc) {
209                             memset(subcoef0, 0, sz * sz * SIZEOF_COEF);
210                             memcpy(subcoef0, coef, SIZEOF_COEF);
211                         } else {
212                             memcpy(subcoef0, coef, sz * sz * SIZEOF_COEF);
213                         }
214                         memcpy(dst0, dst, sz * PIXEL_STRIDE);
215                         memcpy(dst1, dst, sz * PIXEL_STRIDE);
216                         memcpy(subcoef1, subcoef0, sz * sz * SIZEOF_COEF);
217                         call_ref(dst0, subcoef0, PIXEL_STRIDE);
218                         call_new(dst1, subcoef1, PIXEL_STRIDE);
219                         if (memcmp(dst0, dst1, sz * PIXEL_STRIDE) ||
220                             memcmp(subcoef0, subcoef1, sz * sz * SIZEOF_COEF))
221                             fail();
222                         bench_new(dst1, subcoef1, sz * SIZEOF_PIXEL);
223                     }
224                 }
225             }
226         }
227     }
228 }
229
230 static void check_idct_multiple(void)
231 {
232     LOCAL_ALIGNED_16(uint8_t, dst_full,  [16 * 16 * 2]);
233     LOCAL_ALIGNED_16(int16_t, coef_full, [16 * 16 * 2]);
234     LOCAL_ALIGNED_16(uint8_t, dst0,  [16 * 16 * 2]);
235     LOCAL_ALIGNED_16(uint8_t, dst1,  [16 * 16 * 2]);
236     LOCAL_ALIGNED_16(int16_t, coef0, [16 * 16 * 2]);
237     LOCAL_ALIGNED_16(int16_t, coef1, [16 * 16 * 2]);
238     LOCAL_ALIGNED_16(uint8_t, nnzc,  [15 * 8]);
239     H264DSPContext h;
240     int bit_depth, i, y, func;
241     declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, const int *block_offset, int16_t *block, int stride, const uint8_t nnzc[15*8]);
242
243     for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
244         ff_h264dsp_init(&h, bit_depth, 1);
245         for (func = 0; func < 3; func++) {
246             void (*idct)(uint8_t *, const int *, int16_t *, int, const uint8_t[]) = NULL;
247             const char *name;
248             int sz = 4, intra = 0;
249             int block_offset[16] = { 0 };
250             switch (func) {
251             case 0:
252                 idct = h.h264_idct_add16;
253                 name = "h264_idct_add16";
254                 break;
255             case 1:
256                 idct = h.h264_idct_add16intra;
257                 name = "h264_idct_add16intra";
258                 intra = 1;
259                 break;
260             case 2:
261                 idct = h.h264_idct8_add4;
262                 name = "h264_idct8_add4";
263                 sz = 8;
264                 break;
265             }
266             memset(nnzc, 0, 15 * 8);
267             memset(coef_full, 0, 16 * 16 * SIZEOF_COEF);
268             for (i = 0; i < 16 * 16; i += sz * sz) {
269                 uint8_t src[8 * 8 * 2];
270                 uint8_t dst[8 * 8 * 2];
271                 int16_t coef[8 * 8 * 2];
272                 int index = i / sz;
273                 int block_y = (index / 16) * sz;
274                 int block_x = index % 16;
275                 int offset = (block_y * 16 + block_x) * SIZEOF_PIXEL;
276                 int nnz = rnd() % 3;
277
278                 randomize_buffers();
279                 if (sz == 4)
280                     dct4x4(coef, bit_depth);
281                 else
282                     dct8x8(coef, bit_depth);
283
284                 for (y = 0; y < sz; y++)
285                     memcpy(&dst_full[offset + y * 16 * SIZEOF_PIXEL],
286                            &dst[PIXEL_STRIDE * y], sz * SIZEOF_PIXEL);
287
288                 if (nnz > 1)
289                     nnz = sz * sz;
290                 memcpy(&coef_full[i * SIZEOF_COEF/sizeof(coef[0])],
291                        coef, nnz * SIZEOF_COEF);
292
293                 if (intra && nnz == 1)
294                     nnz = 0;
295
296                 nnzc[scan8[i / 16]] = nnz;
297                 block_offset[i / 16] = offset;
298             }
299
300             if (check_func(idct, "%s_%dbpp", name, bit_depth)) {
301                 memcpy(coef0, coef_full, 16 * 16 * SIZEOF_COEF);
302                 memcpy(coef1, coef_full, 16 * 16 * SIZEOF_COEF);
303                 memcpy(dst0, dst_full, 16 * 16 * SIZEOF_PIXEL);
304                 memcpy(dst1, dst_full, 16 * 16 * SIZEOF_PIXEL);
305                 call_ref(dst0, block_offset, coef0, 16 * SIZEOF_PIXEL, nnzc);
306                 call_new(dst1, block_offset, coef1, 16 * SIZEOF_PIXEL, nnzc);
307                 if (memcmp(dst0, dst1, 16 * 16 * SIZEOF_PIXEL) ||
308                     memcmp(coef0, coef1, 16 * 16 * SIZEOF_COEF))
309                     fail();
310                 bench_new(dst1, block_offset, coef1, 16 * SIZEOF_PIXEL, nnzc);
311             }
312         }
313     }
314 }
315
316
317 static void check_loop_filter(void)
318 {
319     LOCAL_ALIGNED_16(uint8_t, dst, [32 * 16 * 2]);
320     LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 16 * 2]);
321     LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 16 * 2]);
322     H264DSPContext h;
323     int bit_depth;
324     int alphas[36], betas[36];
325     int8_t tc0[36][4];
326
327     declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *pix, ptrdiff_t stride,
328                       int alpha, int beta, int8_t *tc0);
329
330     for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
331         int i, j, a, c;
332         uint32_t mask = pixel_mask_lf[bit_depth - 8];
333         ff_h264dsp_init(&h, bit_depth, 1);
334         for (i = 35, a = 255, c = 250; i >= 0; i--) {
335             alphas[i] = a << (bit_depth - 8);
336             betas[i]  = (i + 1) / 2 << (bit_depth - 8);
337             tc0[i][0] = tc0[i][3] = (c + 6) / 10;
338             tc0[i][1] = (c + 7) / 15;
339             tc0[i][2] = (c + 9) / 20;
340             a = a*9/10;
341             c = c*9/10;
342         }
343
344 #define CHECK_LOOP_FILTER(name, align, ...)                             \
345         do {                                                            \
346             if (check_func(h.name, #name "_%dbpp", bit_depth)) {        \
347                 for (j = 0; j < 36; j++) {                              \
348                     intptr_t off = 8 * 32 + (j & 15) * 4 * !align;      \
349                     for (i = 0; i < 1024; i+=4) {                       \
350                         AV_WN32A(dst + i, rnd() & mask);                \
351                     }                                                   \
352                     memcpy(dst0, dst, 32 * 16 * 2);                     \
353                     memcpy(dst1, dst, 32 * 16 * 2);                     \
354                                                                         \
355                     call_ref(dst0 + off, 32, alphas[j], betas[j], tc0[j]); \
356                     call_new(dst1 + off, 32, alphas[j], betas[j], tc0[j]); \
357                     if (memcmp(dst0, dst1, 32 * 16 * SIZEOF_PIXEL)) {   \
358                         fprintf(stderr, #name ": j:%d, alpha:%d beta:%d " \
359                                 "tc0:{%d,%d,%d,%d}\n", j, alphas[j], betas[j], \
360                                 tc0[j][0], tc0[j][1], tc0[j][2], tc0[j][3]); \
361                         fail();                                         \
362                     }                                                   \
363                     bench_new(dst1, 32, alphas[j], betas[j], tc0[j]);   \
364                 }                                                       \
365             }                                                           \
366         } while (0)
367
368         CHECK_LOOP_FILTER(h264_v_loop_filter_luma, 1);
369         CHECK_LOOP_FILTER(h264_h_loop_filter_luma, 0);
370         CHECK_LOOP_FILTER(h264_h_loop_filter_luma_mbaff, 0);
371         CHECK_LOOP_FILTER(h264_v_loop_filter_chroma, 1);
372         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma, 0);
373         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff, 0);
374 #undef CHECK_LOOP_FILTER
375     }
376 }
377
378 static void check_loop_filter_intra(void)
379 {
380     LOCAL_ALIGNED_16(uint8_t, dst, [32 * 16 * 2]);
381     LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 16 * 2]);
382     LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 16 * 2]);
383     H264DSPContext h;
384     int bit_depth;
385     int alphas[36], betas[36];
386
387     declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *pix, ptrdiff_t stride,
388                       int alpha, int beta);
389
390     for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
391         int i, j, a;
392         uint32_t mask = pixel_mask_lf[bit_depth - 8];
393         ff_h264dsp_init(&h, bit_depth, 1);
394         for (i = 35, a = 255; i >= 0; i--) {
395             alphas[i] = a << (bit_depth - 8);
396             betas[i]  = (i + 1) / 2 << (bit_depth - 8);
397             a = a*9/10;
398         }
399
400 #define CHECK_LOOP_FILTER(name, align)                                  \
401         do {                                                            \
402             if (check_func(h.name, #name "_%dbpp", bit_depth)) {        \
403                 for (j = 0; j < 36; j++) {                              \
404                     intptr_t off = 8 * 32 + (j & 15) * 4 * !align;      \
405                     for (i = 0; i < 1024; i+=4) {                       \
406                         AV_WN32A(dst + i, rnd() & mask);                \
407                     }                                                   \
408                     memcpy(dst0, dst, 32 * 16 * 2);                     \
409                     memcpy(dst1, dst, 32 * 16 * 2);                     \
410                                                                         \
411                     call_ref(dst0 + off, 32, alphas[j], betas[j]);      \
412                     call_new(dst1 + off, 32, alphas[j], betas[j]);      \
413                     if (memcmp(dst0, dst1, 32 * 16 * SIZEOF_PIXEL)) {   \
414                         fprintf(stderr, #name ": j:%d, alpha:%d beta:%d\n", \
415                                 j, alphas[j], betas[j]);                \
416                         fail();                                         \
417                     }                                                   \
418                     bench_new(dst1, 32, alphas[j], betas[j]);           \
419                 }                                                       \
420             }                                                           \
421         } while (0)
422
423         CHECK_LOOP_FILTER(h264_v_loop_filter_luma_intra, 1);
424         CHECK_LOOP_FILTER(h264_h_loop_filter_luma_intra, 0);
425         CHECK_LOOP_FILTER(h264_h_loop_filter_luma_mbaff_intra, 0);
426         CHECK_LOOP_FILTER(h264_v_loop_filter_chroma_intra, 1);
427         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_intra, 0);
428         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff_intra, 0);
429 #undef CHECK_LOOP_FILTER
430     }
431 }
432
433 void checkasm_check_h264dsp(void)
434 {
435     check_idct();
436     check_idct_multiple();
437     report("idct");
438
439     check_loop_filter();
440     report("loop_filter");
441
442     check_loop_filter_intra();
443     report("loop_filter_intra");
444 }