• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include "libavutil/mem_internal.h"
30 
31 static const uint32_t pixel_mask[3] = { 0xffffffff, 0x01ff01ff, 0x03ff03ff };
32 static const uint32_t pixel_mask_lf[3] = { 0xff0fff0f, 0x01ff000f, 0x03ff000f };
33 
34 #define SIZEOF_PIXEL ((bit_depth + 7) / 8)
35 #define SIZEOF_COEF  (2 * ((bit_depth + 7) / 8))
36 #define PIXEL_STRIDE 16
37 
38 #define randomize_buffers()                                                  \
39     do {                                                                     \
40         int x, y;                                                            \
41         uint32_t mask = pixel_mask[bit_depth - 8];                           \
42         for (y = 0; y < sz; y++) {                                           \
43             for (x = 0; x < PIXEL_STRIDE; x += 4) {                          \
44                 AV_WN32A(src + y * PIXEL_STRIDE + x, rnd() & mask);          \
45                 AV_WN32A(dst + y * PIXEL_STRIDE + x, rnd() & mask);          \
46             }                                                                \
47             for (x = 0; x < sz; x++) {                                       \
48                 if (bit_depth == 8) {                                        \
49                     coef[y * sz + x] = src[y * PIXEL_STRIDE + x] -           \
50                                        dst[y * PIXEL_STRIDE + x];            \
51                 } else {                                                     \
52                     ((int32_t *)coef)[y * sz + x] =                          \
53                         ((uint16_t *)src)[y * (PIXEL_STRIDE/2) + x] -        \
54                         ((uint16_t *)dst)[y * (PIXEL_STRIDE/2) + x];         \
55                 }                                                            \
56             }                                                                \
57         }                                                                    \
58     } while (0)
59 
60 #define dct4x4_impl(size, dctcoef)                                           \
61 static void dct4x4_##size(dctcoef *coef)                                     \
62 {                                                                            \
63     int i, y, x;                                                             \
64     dctcoef tmp[16];                                                         \
65     for (i = 0; i < 4; i++) {                                                \
66         const int z0 = coef[i*4 + 0] + coef[i*4 + 3];                        \
67         const int z1 = coef[i*4 + 1] + coef[i*4 + 2];                        \
68         const int z2 = coef[i*4 + 0] - coef[i*4 + 3];                        \
69         const int z3 = coef[i*4 + 1] - coef[i*4 + 2];                        \
70         tmp[i + 4*0] =   z0 +   z1;                                          \
71         tmp[i + 4*1] = 2*z2 +   z3;                                          \
72         tmp[i + 4*2] =   z0 -   z1;                                          \
73         tmp[i + 4*3] =   z2 - 2*z3;                                          \
74     }                                                                        \
75     for (i = 0; i < 4; i++) {                                                \
76         const int z0 = tmp[i*4 + 0] + tmp[i*4 + 3];                          \
77         const int z1 = tmp[i*4 + 1] + tmp[i*4 + 2];                          \
78         const int z2 = tmp[i*4 + 0] - tmp[i*4 + 3];                          \
79         const int z3 = tmp[i*4 + 1] - tmp[i*4 + 2];                          \
80         coef[i*4 + 0] =   z0 +   z1;                                         \
81         coef[i*4 + 1] = 2*z2 +   z3;                                         \
82         coef[i*4 + 2] =   z0 -   z1;                                         \
83         coef[i*4 + 3] =   z2 - 2*z3;                                         \
84     }                                                                        \
85     for (y = 0; y < 4; y++) {                                                \
86         for (x = 0; x < 4; x++) {                                            \
87             static const int scale[] = { 13107 * 10, 8066 * 13, 5243 * 16 }; \
88             const int idx = (y & 1) + (x & 1);                               \
89             coef[y*4 + x] = (coef[y*4 + x] * scale[idx] + (1 << 14)) >> 15;  \
90         }                                                                    \
91     }                                                                        \
92 }
93 
94 #define DCT8_1D(src, srcstride, dst, dststride) do {                         \
95     const int a0 = (src)[srcstride * 0] + (src)[srcstride * 7];              \
96     const int a1 = (src)[srcstride * 0] - (src)[srcstride * 7];              \
97     const int a2 = (src)[srcstride * 1] + (src)[srcstride * 6];              \
98     const int a3 = (src)[srcstride * 1] - (src)[srcstride * 6];              \
99     const int a4 = (src)[srcstride * 2] + (src)[srcstride * 5];              \
100     const int a5 = (src)[srcstride * 2] - (src)[srcstride * 5];              \
101     const int a6 = (src)[srcstride * 3] + (src)[srcstride * 4];              \
102     const int a7 = (src)[srcstride * 3] - (src)[srcstride * 4];              \
103     const int b0 = a0 + a6;                                                  \
104     const int b1 = a2 + a4;                                                  \
105     const int b2 = a0 - a6;                                                  \
106     const int b3 = a2 - a4;                                                  \
107     const int b4 = a3 + a5 + (a1 + (a1 >> 1));                               \
108     const int b5 = a1 - a7 - (a5 + (a5 >> 1));                               \
109     const int b6 = a1 + a7 - (a3 + (a3 >> 1));                               \
110     const int b7 = a3 - a5 + (a7 + (a7 >> 1));                               \
111     (dst)[dststride * 0] =  b0 +  b1;                                        \
112     (dst)[dststride * 1] =  b4 + (b7 >> 2);                                  \
113     (dst)[dststride * 2] =  b2 + (b3 >> 1);                                  \
114     (dst)[dststride * 3] =  b5 + (b6 >> 2);                                  \
115     (dst)[dststride * 4] =  b0  - b1;                                        \
116     (dst)[dststride * 5] =  b6 - (b5 >> 2);                                  \
117     (dst)[dststride * 6] = (b2 >> 1) - b3;                                   \
118     (dst)[dststride * 7] = (b4 >> 2) - b7;                                   \
119 } while (0)
120 
121 #define dct8x8_impl(size, dctcoef)                                           \
122 static void dct8x8_##size(dctcoef *coef)                                     \
123 {                                                                            \
124     int i, x, y;                                                             \
125     dctcoef tmp[64];                                                         \
126     for (i = 0; i < 8; i++)                                                  \
127         DCT8_1D(coef + i, 8, tmp + i, 8);                                    \
128                                                                              \
129     for (i = 0; i < 8; i++)                                                  \
130         DCT8_1D(tmp + 8*i, 1, coef + i, 8);                                  \
131                                                                              \
132     for (y = 0; y < 8; y++) {                                                \
133         for (x = 0; x < 8; x++) {                                            \
134             static const int scale[] = {                                     \
135                 13107 * 20, 11428 * 18, 20972 * 32,                          \
136                 12222 * 19, 16777 * 25, 15481 * 24,                          \
137             };                                                               \
138             static const int idxmap[] = {                                    \
139                 0, 3, 4, 3,                                                  \
140                 3, 1, 5, 1,                                                  \
141                 4, 5, 2, 5,                                                  \
142                 3, 1, 5, 1,                                                  \
143             };                                                               \
144             const int idx = idxmap[(y & 3) * 4 + (x & 3)];                   \
145             coef[y*8 + x] = ((int64_t)coef[y*8 + x] *                        \
146                              scale[idx] + (1 << 17)) >> 18;                  \
147         }                                                                    \
148     }                                                                        \
149 }
150 
151 dct4x4_impl(16, int16_t)
152 dct4x4_impl(32, int32_t)
153 
154 dct8x8_impl(16, int16_t)
155 dct8x8_impl(32, int32_t)
156 
dct4x4(int16_t * coef,int bit_depth)157 static void dct4x4(int16_t *coef, int bit_depth)
158 {
159     if (bit_depth == 8)
160         dct4x4_16(coef);
161     else
162         dct4x4_32((int32_t *) coef);
163 }
164 
dct8x8(int16_t * coef,int bit_depth)165 static void dct8x8(int16_t *coef, int bit_depth)
166 {
167     if (bit_depth == 8) {
168         dct8x8_16(coef);
169     } else {
170         dct8x8_32((int32_t *) coef);
171     }
172 }
173 
174 
check_idct(void)175 static void check_idct(void)
176 {
177     LOCAL_ALIGNED_16(uint8_t, src,  [8 * 8 * 2]);
178     LOCAL_ALIGNED_16(uint8_t, dst,  [8 * 8 * 2]);
179     LOCAL_ALIGNED_16(uint8_t, dst0, [8 * 8 * 2]);
180     LOCAL_ALIGNED_16(uint8_t, dst1_base, [8 * 8 * 2 + 32]);
181     LOCAL_ALIGNED_16(int16_t, coef, [8 * 8 * 2]);
182     LOCAL_ALIGNED_16(int16_t, subcoef0, [8 * 8 * 2]);
183     LOCAL_ALIGNED_16(int16_t, subcoef1, [8 * 8 * 2]);
184     H264DSPContext h;
185     int bit_depth, sz, align, dc;
186     declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t *block, int stride);
187 
188     for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
189         ff_h264dsp_init(&h, bit_depth, 1);
190         for (sz = 4; sz <= 8; sz += 4) {
191             randomize_buffers();
192 
193             if (sz == 4)
194                 dct4x4(coef, bit_depth);
195             else
196                 dct8x8(coef, bit_depth);
197 
198             for (dc = 0; dc <= 1; dc++) {
199                 void (*idct)(uint8_t *, int16_t *, int) = NULL;
200                 switch ((sz << 1) | dc) {
201                 case (4 << 1) | 0: idct = h.h264_idct_add; break;
202                 case (4 << 1) | 1: idct = h.h264_idct_dc_add; break;
203                 case (8 << 1) | 0: idct = h.h264_idct8_add; break;
204                 case (8 << 1) | 1: idct = h.h264_idct8_dc_add; break;
205                 }
206                 if (check_func(idct, "h264_idct%d_add%s_%dbpp", sz, dc ? "_dc" : "", bit_depth)) {
207                     for (align = 0; align < 16; align += sz * SIZEOF_PIXEL) {
208                         uint8_t *dst1 = dst1_base + align;
209                         if (dc) {
210                             memset(subcoef0, 0, sz * sz * SIZEOF_COEF);
211                             memcpy(subcoef0, coef, SIZEOF_COEF);
212                         } else {
213                             memcpy(subcoef0, coef, sz * sz * SIZEOF_COEF);
214                         }
215                         memcpy(dst0, dst, sz * PIXEL_STRIDE);
216                         memcpy(dst1, dst, sz * PIXEL_STRIDE);
217                         memcpy(subcoef1, subcoef0, sz * sz * SIZEOF_COEF);
218                         call_ref(dst0, subcoef0, PIXEL_STRIDE);
219                         call_new(dst1, subcoef1, PIXEL_STRIDE);
220                         if (memcmp(dst0, dst1, sz * PIXEL_STRIDE) ||
221                             memcmp(subcoef0, subcoef1, sz * sz * SIZEOF_COEF))
222                             fail();
223                         bench_new(dst1, subcoef1, sz * SIZEOF_PIXEL);
224                     }
225                 }
226             }
227         }
228     }
229 }
230 
check_idct_multiple(void)231 static void check_idct_multiple(void)
232 {
233     LOCAL_ALIGNED_16(uint8_t, dst_full,  [16 * 16 * 2]);
234     LOCAL_ALIGNED_16(int16_t, coef_full, [16 * 16 * 2]);
235     LOCAL_ALIGNED_16(uint8_t, dst0,  [16 * 16 * 2]);
236     LOCAL_ALIGNED_16(uint8_t, dst1,  [16 * 16 * 2]);
237     LOCAL_ALIGNED_16(int16_t, coef0, [16 * 16 * 2]);
238     LOCAL_ALIGNED_16(int16_t, coef1, [16 * 16 * 2]);
239     LOCAL_ALIGNED_16(uint8_t, nnzc,  [15 * 8]);
240     H264DSPContext h;
241     int bit_depth, i, y, func;
242     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]);
243 
244     for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
245         ff_h264dsp_init(&h, bit_depth, 1);
246         for (func = 0; func < 3; func++) {
247             void (*idct)(uint8_t *, const int *, int16_t *, int, const uint8_t[]) = NULL;
248             const char *name;
249             int sz = 4, intra = 0;
250             int block_offset[16] = { 0 };
251             switch (func) {
252             case 0:
253                 idct = h.h264_idct_add16;
254                 name = "h264_idct_add16";
255                 break;
256             case 1:
257                 idct = h.h264_idct_add16intra;
258                 name = "h264_idct_add16intra";
259                 intra = 1;
260                 break;
261             case 2:
262                 idct = h.h264_idct8_add4;
263                 name = "h264_idct8_add4";
264                 sz = 8;
265                 break;
266             }
267             memset(nnzc, 0, 15 * 8);
268             memset(coef_full, 0, 16 * 16 * SIZEOF_COEF);
269             for (i = 0; i < 16 * 16; i += sz * sz) {
270                 uint8_t src[8 * 8 * 2];
271                 uint8_t dst[8 * 8 * 2];
272                 int16_t coef[8 * 8 * 2];
273                 int index = i / sz;
274                 int block_y = (index / 16) * sz;
275                 int block_x = index % 16;
276                 int offset = (block_y * 16 + block_x) * SIZEOF_PIXEL;
277                 int nnz = rnd() % 3;
278 
279                 randomize_buffers();
280                 if (sz == 4)
281                     dct4x4(coef, bit_depth);
282                 else
283                     dct8x8(coef, bit_depth);
284 
285                 for (y = 0; y < sz; y++)
286                     memcpy(&dst_full[offset + y * 16 * SIZEOF_PIXEL],
287                            &dst[PIXEL_STRIDE * y], sz * SIZEOF_PIXEL);
288 
289                 if (nnz > 1)
290                     nnz = sz * sz;
291                 memcpy(&coef_full[i * SIZEOF_COEF/sizeof(coef[0])],
292                        coef, nnz * SIZEOF_COEF);
293 
294                 if (intra && nnz == 1)
295                     nnz = 0;
296 
297                 nnzc[scan8[i / 16]] = nnz;
298                 block_offset[i / 16] = offset;
299             }
300 
301             if (check_func(idct, "%s_%dbpp", name, bit_depth)) {
302                 memcpy(coef0, coef_full, 16 * 16 * SIZEOF_COEF);
303                 memcpy(coef1, coef_full, 16 * 16 * SIZEOF_COEF);
304                 memcpy(dst0, dst_full, 16 * 16 * SIZEOF_PIXEL);
305                 memcpy(dst1, dst_full, 16 * 16 * SIZEOF_PIXEL);
306                 call_ref(dst0, block_offset, coef0, 16 * SIZEOF_PIXEL, nnzc);
307                 call_new(dst1, block_offset, coef1, 16 * SIZEOF_PIXEL, nnzc);
308                 if (memcmp(dst0, dst1, 16 * 16 * SIZEOF_PIXEL) ||
309                     memcmp(coef0, coef1, 16 * 16 * SIZEOF_COEF))
310                     fail();
311                 bench_new(dst1, block_offset, coef1, 16 * SIZEOF_PIXEL, nnzc);
312             }
313         }
314     }
315 }
316 
317 
check_loop_filter(void)318 static void check_loop_filter(void)
319 {
320     LOCAL_ALIGNED_16(uint8_t, dst, [32 * 16 * 2]);
321     LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 16 * 2]);
322     LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 16 * 2]);
323     H264DSPContext h;
324     int bit_depth;
325     int alphas[36], betas[36];
326     int8_t tc0[36][4];
327 
328     declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *pix, ptrdiff_t stride,
329                       int alpha, int beta, int8_t *tc0);
330 
331     for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
332         int i, j, a, c;
333         uint32_t mask = pixel_mask_lf[bit_depth - 8];
334         ff_h264dsp_init(&h, bit_depth, 1);
335         for (i = 35, a = 255, c = 250; i >= 0; i--) {
336             alphas[i] = a << (bit_depth - 8);
337             betas[i]  = (i + 1) / 2 << (bit_depth - 8);
338             tc0[i][0] = tc0[i][3] = (c + 6) / 10;
339             tc0[i][1] = (c + 7) / 15;
340             tc0[i][2] = (c + 9) / 20;
341             a = a*9/10;
342             c = c*9/10;
343         }
344 
345 #define CHECK_LOOP_FILTER(name, align, idc)                             \
346         do {                                                            \
347             if (check_func(h.name, #name #idc "_%dbpp", bit_depth)) {   \
348                 for (j = 0; j < 36; j++) {                              \
349                     intptr_t off = 8 * 32 + (j & 15) * 4 * !align;      \
350                     for (i = 0; i < 1024; i+=4) {                       \
351                         AV_WN32A(dst + i, rnd() & mask);                \
352                     }                                                   \
353                     memcpy(dst0, dst, 32 * 16 * 2);                     \
354                     memcpy(dst1, dst, 32 * 16 * 2);                     \
355                                                                         \
356                     call_ref(dst0 + off, 32, alphas[j], betas[j], tc0[j]); \
357                     call_new(dst1 + off, 32, alphas[j], betas[j], tc0[j]); \
358                     if (memcmp(dst0, dst1, 32 * 16 * SIZEOF_PIXEL)) {   \
359                         fprintf(stderr, #name #idc ": j:%d, alpha:%d beta:%d " \
360                                 "tc0:{%d,%d,%d,%d}\n", j, alphas[j], betas[j], \
361                                 tc0[j][0], tc0[j][1], tc0[j][2], tc0[j][3]); \
362                         fail();                                         \
363                     }                                                   \
364                     bench_new(dst1, 32, alphas[j], betas[j], tc0[j]);   \
365                 }                                                       \
366             }                                                           \
367         } while (0)
368 
369         CHECK_LOOP_FILTER(h264_v_loop_filter_luma, 1,);
370         CHECK_LOOP_FILTER(h264_h_loop_filter_luma, 0,);
371         CHECK_LOOP_FILTER(h264_h_loop_filter_luma_mbaff, 0,);
372         CHECK_LOOP_FILTER(h264_v_loop_filter_chroma, 1,);
373         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma, 0,);
374         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff, 0,);
375 
376         ff_h264dsp_init(&h, bit_depth, 2);
377         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma, 0, 422);
378         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff, 0, 422);
379 #undef CHECK_LOOP_FILTER
380     }
381 }
382 
check_loop_filter_intra(void)383 static void check_loop_filter_intra(void)
384 {
385     LOCAL_ALIGNED_16(uint8_t, dst, [32 * 16 * 2]);
386     LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 16 * 2]);
387     LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 16 * 2]);
388     H264DSPContext h;
389     int bit_depth;
390     int alphas[36], betas[36];
391 
392     declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *pix, ptrdiff_t stride,
393                       int alpha, int beta);
394 
395     for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
396         int i, j, a;
397         uint32_t mask = pixel_mask_lf[bit_depth - 8];
398         ff_h264dsp_init(&h, bit_depth, 1);
399         for (i = 35, a = 255; i >= 0; i--) {
400             alphas[i] = a << (bit_depth - 8);
401             betas[i]  = (i + 1) / 2 << (bit_depth - 8);
402             a = a*9/10;
403         }
404 
405 #define CHECK_LOOP_FILTER(name, align, idc)                             \
406         do {                                                            \
407             if (check_func(h.name, #name #idc "_%dbpp", bit_depth)) {   \
408                 for (j = 0; j < 36; j++) {                              \
409                     intptr_t off = 8 * 32 + (j & 15) * 4 * !align;      \
410                     for (i = 0; i < 1024; i+=4) {                       \
411                         AV_WN32A(dst + i, rnd() & mask);                \
412                     }                                                   \
413                     memcpy(dst0, dst, 32 * 16 * 2);                     \
414                     memcpy(dst1, dst, 32 * 16 * 2);                     \
415                                                                         \
416                     call_ref(dst0 + off, 32, alphas[j], betas[j]);      \
417                     call_new(dst1 + off, 32, alphas[j], betas[j]);      \
418                     if (memcmp(dst0, dst1, 32 * 16 * SIZEOF_PIXEL)) {   \
419                         fprintf(stderr, #name #idc ": j:%d, alpha:%d beta:%d\n", \
420                                 j, alphas[j], betas[j]);                \
421                         fail();                                         \
422                     }                                                   \
423                     bench_new(dst1, 32, alphas[j], betas[j]);           \
424                 }                                                       \
425             }                                                           \
426         } while (0)
427 
428         CHECK_LOOP_FILTER(h264_v_loop_filter_luma_intra, 1,);
429         CHECK_LOOP_FILTER(h264_h_loop_filter_luma_intra, 0,);
430         CHECK_LOOP_FILTER(h264_h_loop_filter_luma_mbaff_intra, 0,);
431         CHECK_LOOP_FILTER(h264_v_loop_filter_chroma_intra, 1,);
432         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_intra, 0,);
433         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff_intra, 0,);
434 
435         ff_h264dsp_init(&h, bit_depth, 2);
436         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_intra, 0, 422);
437         CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff_intra, 0, 422);
438 #undef CHECK_LOOP_FILTER
439     }
440 }
441 
checkasm_check_h264dsp(void)442 void checkasm_check_h264dsp(void)
443 {
444     check_idct();
445     check_idct_multiple();
446     report("idct");
447 
448     check_loop_filter();
449     report("loop_filter");
450 
451     check_loop_filter_intra();
452     report("loop_filter_intra");
453 }
454