Line | Branch | Exec | Source |
---|---|---|---|
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/h264dsp.h" | ||
24 | #include "libavcodec/h264data.h" | ||
25 | #include "libavcodec/h264idct.h" | ||
26 | #include "libavcodec/h264_parse.h" | ||
27 | #include "libavutil/common.h" | ||
28 | #include "libavutil/intreadwrite.h" | ||
29 | #include "libavutil/mem_internal.h" | ||
30 | |||
31 | static const uint32_t pixel_mask[5] = { 0xffffffff, 0x01ff01ff, 0x03ff03ff, 0x0fff0fff, 0x3fff3fff }; | ||
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(idx) \ | ||
39 | do { \ | ||
40 | int x, y; \ | ||
41 | uint32_t mask = pixel_mask[(idx)]; \ | ||
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 | const int64_t 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 |
8/8✓ Branch 0 taken 1820 times.
✓ Branch 1 taken 455 times.
✓ Branch 2 taken 1820 times.
✓ Branch 3 taken 455 times.
✓ Branch 4 taken 7280 times.
✓ Branch 5 taken 1820 times.
✓ Branch 6 taken 1820 times.
✓ Branch 7 taken 455 times.
|
13195 | dct4x4_impl(16, int16_t) |
152 |
8/8✓ Branch 0 taken 3952 times.
✓ Branch 1 taken 988 times.
✓ Branch 2 taken 3952 times.
✓ Branch 3 taken 988 times.
✓ Branch 4 taken 15808 times.
✓ Branch 5 taken 3952 times.
✓ Branch 6 taken 3952 times.
✓ Branch 7 taken 988 times.
|
28652 | dct4x4_impl(32, int32_t) |
153 | |||
154 |
8/8✓ Branch 0 taken 728 times.
✓ Branch 1 taken 91 times.
✓ Branch 2 taken 728 times.
✓ Branch 3 taken 91 times.
✓ Branch 4 taken 5824 times.
✓ Branch 5 taken 728 times.
✓ Branch 6 taken 728 times.
✓ Branch 7 taken 91 times.
|
8099 | dct8x8_impl(16, int16_t) |
155 |
8/8✓ Branch 0 taken 2080 times.
✓ Branch 1 taken 260 times.
✓ Branch 2 taken 2080 times.
✓ Branch 3 taken 260 times.
✓ Branch 4 taken 16640 times.
✓ Branch 5 taken 2080 times.
✓ Branch 6 taken 2080 times.
✓ Branch 7 taken 260 times.
|
23140 | dct8x8_impl(32, int32_t) |
156 | |||
157 | 1443 | static void dct4x4(int16_t *coef, int bit_depth) | |
158 | { | ||
159 |
2/2✓ Branch 0 taken 455 times.
✓ Branch 1 taken 988 times.
|
1443 | if (bit_depth == 8) |
160 | 455 | dct4x4_16(coef); | |
161 | else | ||
162 | 988 | dct4x4_32((int32_t *) coef); | |
163 | 1443 | } | |
164 | |||
165 | 351 | static void dct8x8(int16_t *coef, int bit_depth) | |
166 | { | ||
167 |
2/2✓ Branch 0 taken 91 times.
✓ Branch 1 taken 260 times.
|
351 | if (bit_depth == 8) { |
168 | 91 | dct8x8_16(coef); | |
169 | } else { | ||
170 | 260 | dct8x8_32((int32_t *) coef); | |
171 | } | ||
172 | 351 | } | |
173 | |||
174 | |||
175 | 13 | static void check_idct(void) | |
176 | { | ||
177 | static const int depths[5] = { 8, 9, 10, 12, 14 }; | ||
178 | 13 | LOCAL_ALIGNED_16(uint8_t, src, [8 * 8 * 2]); | |
179 | 13 | LOCAL_ALIGNED_16(uint8_t, dst, [8 * 8 * 2]); | |
180 | 13 | LOCAL_ALIGNED_16(uint8_t, dst0, [8 * 8 * 2]); | |
181 | 13 | LOCAL_ALIGNED_16(uint8_t, dst1_base, [8 * 8 * 2 + 32]); | |
182 | 13 | LOCAL_ALIGNED_16(int16_t, coef, [8 * 8 * 2]); | |
183 | 13 | LOCAL_ALIGNED_16(int16_t, subcoef0, [8 * 8 * 2]); | |
184 | 13 | LOCAL_ALIGNED_16(int16_t, subcoef1, [8 * 8 * 2]); | |
185 | H264DSPContext h; | ||
186 | int bit_depth, sz, align, dc, i; | ||
187 |
2/2✓ Branch 1 taken 12 times.
✓ Branch 2 taken 1 times.
|
13 | declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t *block, int stride); |
188 | |||
189 |
2/2✓ Branch 0 taken 65 times.
✓ Branch 1 taken 13 times.
|
78 | for (i = 0; i < FF_ARRAY_ELEMS(depths); i++) { |
190 | 65 | bit_depth = depths[i]; | |
191 | 65 | ff_h264dsp_init(&h, bit_depth, 1); | |
192 | |||
193 |
2/2✓ Branch 0 taken 195 times.
✓ Branch 1 taken 65 times.
|
260 | for (dc = 0; dc <= 2; dc++) { |
194 |
2/2✓ Branch 0 taken 390 times.
✓ Branch 1 taken 195 times.
|
585 | for (sz = 4; sz <= 8; sz += 4) { |
195 | 390 | void (*idct)(uint8_t *, int16_t *, int) = NULL; | |
196 | 390 | const char fmts[3][28] = { | |
197 | "h264_idct%d_add_%dbpp", "h264_idct%d_dc_add_%dbpp", | ||
198 | "h264_add_pixels%d_%dbpp", | ||
199 | }; | ||
200 | |||
201 |
8/8✓ Branch 2 taken 9360 times.
✓ Branch 3 taken 2340 times.
✓ Branch 4 taken 3120 times.
✓ Branch 5 taken 12480 times.
✓ Branch 6 taken 15600 times.
✓ Branch 7 taken 2340 times.
✓ Branch 8 taken 2340 times.
✓ Branch 9 taken 390 times.
|
27690 | randomize_buffers(i); |
202 | |||
203 |
2/2✓ Branch 0 taken 195 times.
✓ Branch 1 taken 195 times.
|
390 | if (sz == 4) |
204 | 195 | dct4x4(coef, bit_depth); | |
205 | else | ||
206 | 195 | dct8x8(coef, bit_depth); | |
207 | |||
208 |
6/7✓ Branch 0 taken 65 times.
✓ Branch 1 taken 65 times.
✓ Branch 2 taken 65 times.
✓ Branch 3 taken 65 times.
✓ Branch 4 taken 65 times.
✓ Branch 5 taken 65 times.
✗ Branch 6 not taken.
|
390 | switch ((sz << 2) | dc) { |
209 | 65 | case (4 << 2) | 0: idct = h.h264_idct_add; break; | |
210 | 65 | case (4 << 2) | 1: idct = h.h264_idct_dc_add; break; | |
211 | 65 | case (4 << 2) | 2: idct = h.h264_add_pixels4_clear; break; | |
212 | 65 | case (8 << 2) | 0: idct = h.h264_idct8_add; break; | |
213 | 65 | case (8 << 2) | 1: idct = h.h264_idct8_dc_add; break; | |
214 | 65 | case (8 << 2) | 2: idct = h.h264_add_pixels8_clear; break; | |
215 | } | ||
216 | |||
217 |
2/2✓ Branch 3 taken 44 times.
✓ Branch 4 taken 346 times.
|
390 | if (check_func(idct, fmts[dc], sz, bit_depth)) { |
218 |
2/2✓ Branch 0 taken 86 times.
✓ Branch 1 taken 44 times.
|
130 | for (align = 0; align < 16; align += sz * SIZEOF_PIXEL) { |
219 | 86 | uint8_t *dst1 = dst1_base + align; | |
220 |
2/2✓ Branch 0 taken 52 times.
✓ Branch 1 taken 34 times.
|
86 | if (dc) { |
221 | 52 | memset(subcoef0, 0, sz * sz * SIZEOF_COEF); | |
222 | 52 | memcpy(subcoef0, coef, SIZEOF_COEF); | |
223 | } else { | ||
224 | 34 | memcpy(subcoef0, coef, sz * sz * SIZEOF_COEF); | |
225 | } | ||
226 | 86 | memcpy(dst0, dst, sz * PIXEL_STRIDE); | |
227 | 86 | memcpy(dst1, dst, sz * PIXEL_STRIDE); | |
228 | 86 | memcpy(subcoef1, subcoef0, sz * sz * SIZEOF_COEF); | |
229 | 86 | call_ref(dst0, subcoef0, PIXEL_STRIDE); | |
230 | 86 | call_new(dst1, subcoef1, PIXEL_STRIDE); | |
231 |
1/2✓ Branch 0 taken 86 times.
✗ Branch 1 not taken.
|
86 | if (memcmp(dst0, dst1, sz * PIXEL_STRIDE) || |
232 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 86 times.
|
86 | memcmp(subcoef0, subcoef1, sz * sz * SIZEOF_COEF)) |
233 | ✗ | fail(); | |
234 |
1/8✗ Branch 1 not taken.
✓ Branch 2 taken 86 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
|
86 | bench_new(dst1, subcoef1, sz * SIZEOF_PIXEL); |
235 | } | ||
236 | } | ||
237 | } | ||
238 | } | ||
239 | } | ||
240 | 13 | } | |
241 | |||
242 | 13 | static void check_idct_multiple(void) | |
243 | { | ||
244 | 13 | LOCAL_ALIGNED_16(uint8_t, dst_full, [16 * 16 * 2]); | |
245 | 13 | LOCAL_ALIGNED_16(int16_t, coef_full, [16 * 16 * 2]); | |
246 | 13 | LOCAL_ALIGNED_16(uint8_t, dst0, [16 * 16 * 2]); | |
247 | 13 | LOCAL_ALIGNED_16(uint8_t, dst1, [16 * 16 * 2]); | |
248 | 13 | LOCAL_ALIGNED_16(int16_t, coef0, [16 * 16 * 2]); | |
249 | 13 | LOCAL_ALIGNED_16(int16_t, coef1, [16 * 16 * 2]); | |
250 | 13 | LOCAL_ALIGNED_16(uint8_t, nnzc, [15 * 8]); | |
251 | H264DSPContext h; | ||
252 | int bit_depth, i, y, func; | ||
253 |
2/2✓ Branch 1 taken 12 times.
✓ Branch 2 taken 1 times.
|
13 | 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]); |
254 | |||
255 |
2/2✓ Branch 0 taken 39 times.
✓ Branch 1 taken 13 times.
|
52 | for (bit_depth = 8; bit_depth <= 10; bit_depth++) { |
256 | 39 | ff_h264dsp_init(&h, bit_depth, 1); | |
257 |
2/2✓ Branch 0 taken 117 times.
✓ Branch 1 taken 39 times.
|
156 | for (func = 0; func < 3; func++) { |
258 | 117 | void (*idct)(uint8_t *, const int *, int16_t *, int, const uint8_t[]) = NULL; | |
259 | const char *name; | ||
260 | 117 | int sz = 4, intra = 0; | |
261 | 117 | int block_offset[16] = { 0 }; | |
262 |
3/4✓ Branch 0 taken 39 times.
✓ Branch 1 taken 39 times.
✓ Branch 2 taken 39 times.
✗ Branch 3 not taken.
|
117 | switch (func) { |
263 | 39 | case 0: | |
264 | 39 | idct = h.h264_idct_add16; | |
265 | 39 | name = "h264_idct_add16"; | |
266 | 39 | break; | |
267 | 39 | case 1: | |
268 | 39 | idct = h.h264_idct_add16intra; | |
269 | 39 | name = "h264_idct_add16intra"; | |
270 | 39 | intra = 1; | |
271 | 39 | break; | |
272 | 39 | case 2: | |
273 | 39 | idct = h.h264_idct8_add4; | |
274 | 39 | name = "h264_idct8_add4"; | |
275 | 39 | sz = 8; | |
276 | 39 | break; | |
277 | } | ||
278 | 117 | memset(nnzc, 0, 15 * 8); | |
279 | 117 | memset(coef_full, 0, 16 * 16 * SIZEOF_COEF); | |
280 |
2/2✓ Branch 0 taken 1404 times.
✓ Branch 1 taken 117 times.
|
1521 | for (i = 0; i < 16 * 16; i += sz * sz) { |
281 | uint8_t src[8 * 8 * 2]; | ||
282 | uint8_t dst[8 * 8 * 2]; | ||
283 | int16_t coef[8 * 8 * 2]; | ||
284 | 1404 | int index = i / sz; | |
285 | 1404 | int block_y = (index / 16) * sz; | |
286 | 1404 | int block_x = index % 16; | |
287 | 1404 | int offset = (block_y * 16 + block_x) * SIZEOF_PIXEL; | |
288 | 1404 | int nnz = rnd() % 3; | |
289 | |||
290 |
8/8✓ Branch 2 taken 24960 times.
✓ Branch 3 taken 6240 times.
✓ Branch 4 taken 9984 times.
✓ Branch 5 taken 19968 times.
✓ Branch 6 taken 29952 times.
✓ Branch 7 taken 6240 times.
✓ Branch 8 taken 6240 times.
✓ Branch 9 taken 1404 times.
|
62556 | randomize_buffers(bit_depth - 8); |
291 |
2/2✓ Branch 0 taken 1248 times.
✓ Branch 1 taken 156 times.
|
1404 | if (sz == 4) |
292 | 1248 | dct4x4(coef, bit_depth); | |
293 | else | ||
294 | 156 | dct8x8(coef, bit_depth); | |
295 | |||
296 |
2/2✓ Branch 0 taken 6240 times.
✓ Branch 1 taken 1404 times.
|
7644 | for (y = 0; y < sz; y++) |
297 | 6240 | memcpy(&dst_full[offset + y * 16 * SIZEOF_PIXEL], | |
298 | 6240 | &dst[PIXEL_STRIDE * y], sz * SIZEOF_PIXEL); | |
299 | |||
300 |
2/2✓ Branch 0 taken 465 times.
✓ Branch 1 taken 939 times.
|
1404 | if (nnz > 1) |
301 | 465 | nnz = sz * sz; | |
302 | 1404 | memcpy(&coef_full[i * SIZEOF_COEF/sizeof(coef[0])], | |
303 | 1404 | coef, nnz * SIZEOF_COEF); | |
304 | |||
305 |
4/4✓ Branch 0 taken 624 times.
✓ Branch 1 taken 780 times.
✓ Branch 2 taken 222 times.
✓ Branch 3 taken 402 times.
|
1404 | if (intra && nnz == 1) |
306 | 222 | nnz = 0; | |
307 | |||
308 | 1404 | nnzc[scan8[i / 16]] = nnz; | |
309 | 1404 | block_offset[i / 16] = offset; | |
310 | } | ||
311 | |||
312 |
2/2✓ Branch 3 taken 18 times.
✓ Branch 4 taken 99 times.
|
117 | if (check_func(idct, "%s_%dbpp", name, bit_depth)) { |
313 | 18 | memcpy(coef0, coef_full, 16 * 16 * SIZEOF_COEF); | |
314 | 18 | memcpy(coef1, coef_full, 16 * 16 * SIZEOF_COEF); | |
315 | 18 | memcpy(dst0, dst_full, 16 * 16 * SIZEOF_PIXEL); | |
316 | 18 | memcpy(dst1, dst_full, 16 * 16 * SIZEOF_PIXEL); | |
317 | 18 | call_ref(dst0, block_offset, coef0, 16 * SIZEOF_PIXEL, nnzc); | |
318 | 18 | call_new(dst1, block_offset, coef1, 16 * SIZEOF_PIXEL, nnzc); | |
319 |
1/2✓ Branch 0 taken 18 times.
✗ Branch 1 not taken.
|
18 | if (memcmp(dst0, dst1, 16 * 16 * SIZEOF_PIXEL) || |
320 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 18 times.
|
18 | memcmp(coef0, coef1, 16 * 16 * SIZEOF_COEF)) |
321 | ✗ | fail(); | |
322 |
1/8✗ Branch 1 not taken.
✓ Branch 2 taken 18 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
|
18 | bench_new(dst1, block_offset, coef1, 16 * SIZEOF_PIXEL, nnzc); |
323 | } | ||
324 | } | ||
325 | } | ||
326 | 13 | } | |
327 | |||
328 | 13 | static void check_idct_dequant(void) | |
329 | { | ||
330 | static const int depths[5] = { 8, 9, 10, 12, 14 }; | ||
331 | 13 | LOCAL_ALIGNED_16(int16_t, src, [16]); | |
332 | /* Ensure dst buffers are large enough to hold dctcoefs of all bit-depths. */ | ||
333 | 13 | LOCAL_ALIGNED_16(uint8_t, dst0, [16 * 16 * sizeof(int32_t)]); | |
334 | 13 | LOCAL_ALIGNED_16(uint8_t, dst1, [16 * 16 * sizeof(int32_t)]); | |
335 | 13 | int16_t *dst_ref = (int16_t *)dst0; | |
336 | 13 | int16_t *dst_new = (int16_t *)dst1; | |
337 | H264DSPContext h; | ||
338 | int bit_depth, i, qmul; | ||
339 |
2/2✓ Branch 1 taken 12 times.
✓ Branch 2 taken 1 times.
|
13 | declare_func_emms(AV_CPU_FLAG_MMX | AV_CPU_FLAG_SSE2, void, int16_t *output, int16_t *input, int qmul); |
340 | |||
341 |
2/2✓ Branch 0 taken 208 times.
✓ Branch 1 taken 13 times.
|
221 | for (int j = 0; j < 16; j++) |
342 | 208 | src[j] = (rnd() % 512) - 256; | |
343 | |||
344 | 13 | qmul = rnd() % 4096; | |
345 | |||
346 |
2/2✓ Branch 0 taken 65 times.
✓ Branch 1 taken 13 times.
|
78 | for (i = 0; i < FF_ARRAY_ELEMS(depths); i++) { |
347 | 65 | bit_depth = depths[i]; | |
348 | 65 | ff_h264dsp_init(&h, bit_depth, 1); | |
349 | |||
350 | 65 | memset(dst0, 0, 16 * 16 * SIZEOF_COEF); | |
351 | 65 | memset(dst1, 0, 16 * 16 * SIZEOF_COEF); | |
352 | |||
353 |
2/2✓ Branch 3 taken 6 times.
✓ Branch 4 taken 59 times.
|
65 | if (check_func(h.h264_luma_dc_dequant_idct, "h264_luma_dc_dequant_idct_%d", bit_depth)) { |
354 | |||
355 | 6 | call_ref(dst_ref, src, qmul); | |
356 | 6 | call_new(dst_new, src, qmul); | |
357 |
2/2✓ Branch 0 taken 4 times.
✓ Branch 1 taken 2 times.
|
6 | checkasm_check_dctcoef(dst_ref, 16*SIZEOF_COEF, dst_new, 16*SIZEOF_COEF, 16, 16, "dst"); |
358 |
1/8✗ Branch 1 not taken.
✓ Branch 2 taken 6 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
|
6 | bench_new(dst_new, src, qmul); |
359 | } | ||
360 | } | ||
361 | 13 | } | |
362 | |||
363 | |||
364 | 13 | static void check_loop_filter(void) | |
365 | { | ||
366 | 13 | LOCAL_ALIGNED_16(uint8_t, dst, [32 * 16 * 2]); | |
367 | 13 | LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 16 * 2]); | |
368 | 13 | LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 16 * 2]); | |
369 | H264DSPContext h; | ||
370 | int bit_depth; | ||
371 | int alphas[36], betas[36]; | ||
372 | int8_t tc0[36][4]; | ||
373 | |||
374 |
2/2✓ Branch 1 taken 12 times.
✓ Branch 2 taken 1 times.
|
13 | declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *pix, ptrdiff_t stride, |
375 | int alpha, int beta, int8_t *tc0); | ||
376 | |||
377 |
2/2✓ Branch 0 taken 39 times.
✓ Branch 1 taken 13 times.
|
52 | for (bit_depth = 8; bit_depth <= 10; bit_depth++) { |
378 | int i, j, a, c; | ||
379 | 39 | uint32_t mask = pixel_mask_lf[bit_depth - 8]; | |
380 | 39 | ff_h264dsp_init(&h, bit_depth, 1); | |
381 |
2/2✓ Branch 0 taken 1404 times.
✓ Branch 1 taken 39 times.
|
1443 | for (i = 35, a = 255, c = 250; i >= 0; i--) { |
382 | 1404 | alphas[i] = a << (bit_depth - 8); | |
383 | 1404 | betas[i] = (i + 1) / 2 << (bit_depth - 8); | |
384 | 1404 | tc0[i][0] = tc0[i][3] = (c + 6) / 10; | |
385 | 1404 | tc0[i][1] = (c + 7) / 15; | |
386 | 1404 | tc0[i][2] = (c + 9) / 20; | |
387 | 1404 | a = a*9/10; | |
388 | 1404 | c = c*9/10; | |
389 | } | ||
390 | |||
391 | #define CHECK_LOOP_FILTER(name, align, idc) \ | ||
392 | do { \ | ||
393 | if (check_func(h.name, #name #idc "_%dbpp", bit_depth)) { \ | ||
394 | for (j = 0; j < 36; j++) { \ | ||
395 | intptr_t off = 8 * 32 + (j & 15) * 4 * !align; \ | ||
396 | for (i = 0; i < 1024; i+=4) { \ | ||
397 | AV_WN32A(dst + i, rnd() & mask); \ | ||
398 | } \ | ||
399 | memcpy(dst0, dst, 32 * 16 * 2); \ | ||
400 | memcpy(dst1, dst, 32 * 16 * 2); \ | ||
401 | \ | ||
402 | call_ref(dst0 + off, 32, alphas[j], betas[j], tc0[j]); \ | ||
403 | call_new(dst1 + off, 32, alphas[j], betas[j], tc0[j]); \ | ||
404 | if (memcmp(dst0, dst1, 32 * 16 * SIZEOF_PIXEL)) { \ | ||
405 | fprintf(stderr, #name #idc ": j:%d, alpha:%d beta:%d " \ | ||
406 | "tc0:{%d,%d,%d,%d}\n", j, alphas[j], betas[j], \ | ||
407 | tc0[j][0], tc0[j][1], tc0[j][2], tc0[j][3]); \ | ||
408 | fail(); \ | ||
409 | } \ | ||
410 | bench_new(dst1 + off, 32, alphas[j], betas[j], tc0[j]);\ | ||
411 | } \ | ||
412 | } \ | ||
413 | } while (0) | ||
414 | |||
415 |
8/16✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
|
64803 | CHECK_LOOP_FILTER(h264_v_loop_filter_luma, 1,); |
416 |
8/16✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
|
64803 | CHECK_LOOP_FILTER(h264_h_loop_filter_luma, 0,); |
417 |
8/16✓ Branch 3 taken 5 times.
✓ Branch 4 taken 34 times.
✓ Branch 6 taken 46080 times.
✓ Branch 7 taken 180 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 180 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 180 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 180 times.
✓ Branch 40 taken 5 times.
|
46299 | CHECK_LOOP_FILTER(h264_h_loop_filter_luma_mbaff, 0,); |
418 |
8/16✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
|
64803 | CHECK_LOOP_FILTER(h264_v_loop_filter_chroma, 1,); |
419 |
8/16✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
|
64803 | CHECK_LOOP_FILTER(h264_h_loop_filter_chroma, 0,); |
420 |
8/16✓ Branch 3 taken 3 times.
✓ Branch 4 taken 36 times.
✓ Branch 6 taken 27648 times.
✓ Branch 7 taken 108 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 108 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 108 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 108 times.
✓ Branch 40 taken 3 times.
|
27795 | CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff, 0,); |
421 | |||
422 | 39 | ff_h264dsp_init(&h, bit_depth, 2); | |
423 |
8/16✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
|
64803 | CHECK_LOOP_FILTER(h264_h_loop_filter_chroma, 0, 422); |
424 |
8/16✓ Branch 3 taken 3 times.
✓ Branch 4 taken 36 times.
✓ Branch 6 taken 27648 times.
✓ Branch 7 taken 108 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 108 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 108 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 108 times.
✓ Branch 40 taken 3 times.
|
27795 | CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff, 0, 422); |
425 | #undef CHECK_LOOP_FILTER | ||
426 | } | ||
427 | 13 | } | |
428 | |||
429 | 13 | static void check_loop_filter_intra(void) | |
430 | { | ||
431 | 13 | LOCAL_ALIGNED_16(uint8_t, dst, [32 * 16 * 2]); | |
432 | 13 | LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 16 * 2]); | |
433 | 13 | LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 16 * 2]); | |
434 | H264DSPContext h; | ||
435 | int bit_depth; | ||
436 | int alphas[36], betas[36]; | ||
437 | |||
438 |
2/2✓ Branch 1 taken 12 times.
✓ Branch 2 taken 1 times.
|
13 | declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *pix, ptrdiff_t stride, |
439 | int alpha, int beta); | ||
440 | |||
441 |
2/2✓ Branch 0 taken 39 times.
✓ Branch 1 taken 13 times.
|
52 | for (bit_depth = 8; bit_depth <= 10; bit_depth++) { |
442 | int i, j, a; | ||
443 | 39 | uint32_t mask = pixel_mask_lf[bit_depth - 8]; | |
444 | 39 | ff_h264dsp_init(&h, bit_depth, 1); | |
445 |
2/2✓ Branch 0 taken 1404 times.
✓ Branch 1 taken 39 times.
|
1443 | for (i = 35, a = 255; i >= 0; i--) { |
446 | 1404 | alphas[i] = a << (bit_depth - 8); | |
447 | 1404 | betas[i] = (i + 1) / 2 << (bit_depth - 8); | |
448 | 1404 | a = a*9/10; | |
449 | } | ||
450 | |||
451 | #define CHECK_LOOP_FILTER(name, align, idc) \ | ||
452 | do { \ | ||
453 | if (check_func(h.name, #name #idc "_%dbpp", bit_depth)) { \ | ||
454 | for (j = 0; j < 36; j++) { \ | ||
455 | intptr_t off = 8 * 32 + (j & 15) * 4 * !align; \ | ||
456 | for (i = 0; i < 1024; i+=4) { \ | ||
457 | AV_WN32A(dst + i, rnd() & mask); \ | ||
458 | } \ | ||
459 | memcpy(dst0, dst, 32 * 16 * 2); \ | ||
460 | memcpy(dst1, dst, 32 * 16 * 2); \ | ||
461 | \ | ||
462 | call_ref(dst0 + off, 32, alphas[j], betas[j]); \ | ||
463 | call_new(dst1 + off, 32, alphas[j], betas[j]); \ | ||
464 | if (memcmp(dst0, dst1, 32 * 16 * SIZEOF_PIXEL)) { \ | ||
465 | fprintf(stderr, #name #idc ": j:%d, alpha:%d beta:%d\n", \ | ||
466 | j, alphas[j], betas[j]); \ | ||
467 | fail(); \ | ||
468 | } \ | ||
469 | bench_new(dst1 + off, 32, alphas[j], betas[j]); \ | ||
470 | } \ | ||
471 | } \ | ||
472 | } while (0) | ||
473 | |||
474 |
8/16✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
|
64803 | CHECK_LOOP_FILTER(h264_v_loop_filter_luma_intra, 1,); |
475 |
8/16✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
|
64803 | CHECK_LOOP_FILTER(h264_h_loop_filter_luma_intra, 0,); |
476 |
8/16✓ Branch 3 taken 3 times.
✓ Branch 4 taken 36 times.
✓ Branch 6 taken 27648 times.
✓ Branch 7 taken 108 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 108 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 108 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 108 times.
✓ Branch 40 taken 3 times.
|
27795 | CHECK_LOOP_FILTER(h264_h_loop_filter_luma_mbaff_intra, 0,); |
477 |
8/16✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
|
64803 | CHECK_LOOP_FILTER(h264_v_loop_filter_chroma_intra, 1,); |
478 |
8/16✓ Branch 3 taken 5 times.
✓ Branch 4 taken 34 times.
✓ Branch 6 taken 46080 times.
✓ Branch 7 taken 180 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 180 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 180 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 180 times.
✓ Branch 40 taken 5 times.
|
46299 | CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_intra, 0,); |
479 |
8/16✓ Branch 3 taken 3 times.
✓ Branch 4 taken 36 times.
✓ Branch 6 taken 27648 times.
✓ Branch 7 taken 108 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 108 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 108 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 108 times.
✓ Branch 40 taken 3 times.
|
27795 | CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff_intra, 0,); |
480 | |||
481 | 39 | ff_h264dsp_init(&h, bit_depth, 2); | |
482 |
8/16✓ Branch 3 taken 5 times.
✓ Branch 4 taken 34 times.
✓ Branch 6 taken 46080 times.
✓ Branch 7 taken 180 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 180 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 180 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 180 times.
✓ Branch 40 taken 5 times.
|
46299 | CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_intra, 0, 422); |
483 |
8/16✓ Branch 3 taken 3 times.
✓ Branch 4 taken 36 times.
✓ Branch 6 taken 27648 times.
✓ Branch 7 taken 108 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 108 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 108 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 108 times.
✓ Branch 40 taken 3 times.
|
27795 | CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff_intra, 0, 422); |
484 | #undef CHECK_LOOP_FILTER | ||
485 | } | ||
486 | 13 | } | |
487 | |||
488 | 13 | void checkasm_check_h264dsp(void) | |
489 | { | ||
490 | 13 | check_idct(); | |
491 | 13 | check_idct_multiple(); | |
492 | 13 | check_idct_dequant(); | |
493 | 13 | report("idct"); | |
494 | |||
495 | 13 | check_loop_filter(); | |
496 | 13 | report("loop_filter"); | |
497 | |||
498 | 13 | check_loop_filter_intra(); | |
499 | 13 | report("loop_filter_intra"); | |
500 | 13 | } | |
501 |