?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Remove compatibility workarounds
[x264.git] / tools / checkasm.c
blobeaa66da8538755b83d53942704a35204240a6303
1 /*****************************************************************************
2 * checkasm.c: assembly check tool
3 *****************************************************************************
4 * Copyright (C) 2003-2019 x264 project
5 *
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
7 * Laurent Aimar <fenrir@via.ecp.fr>
8 * Fiona Glaser <fiona@x264.com>
9 *
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; either version 2 of the License, or
13 * (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
23 *
24 * This program is also available under a commercial proprietary license.
25 * For more information, contact us at licensing@x264.com.
26 *****************************************************************************/
27
28 #include <ctype.h>
29 #include "common/common.h"
30 #include "encoder/macroblock.h"
31
32 #ifdef _WIN32
33 #include <windows.h>
34 #endif
35
36 // GCC doesn't align stack variables on ARM, so use .bss
37 #if ARCH_ARM
38 #undef ALIGNED_16
39 #define ALIGNED_16( var ) DECLARE_ALIGNED( static var, 16 )
40 #endif
41
42 /* buf1, buf2: initialised to random data and shouldn't write into them */
43 static uint8_t *buf1, *buf2;
44 /* buf3, buf4: used to store output */
45 static uint8_t *buf3, *buf4;
46 /* pbuf1, pbuf2: initialised to random pixel data and shouldn't write into them. */
47 static pixel *pbuf1, *pbuf2;
48 /* pbuf3, pbuf4: point to buf3, buf4, just for type convenience */
49 static pixel *pbuf3, *pbuf4;
50
51 static int quiet = 0;
52
53 #define report( name ) { \
54 if( used_asm && !quiet ) \
55 fprintf( stderr, " - %-21s [%s]\n", name, ok ? "OK" : "FAILED" ); \
56 if( !ok ) ret = -1; \
57 }
58
59 #define BENCH_RUNS 2000 // tradeoff between accuracy and speed
60 #define MAX_FUNCS 1000 // just has to be big enough to hold all the existing functions
61 #define MAX_CPUS 30 // number of different combinations of cpu flags
62
63 typedef struct
64 {
65 void *pointer; // just for detecting duplicates
66 uint32_t cpu;
67 uint64_t cycles;
68 uint32_t den;
69 } bench_t;
70
71 typedef struct
72 {
73 char *name;
74 bench_t vers[MAX_CPUS];
75 } bench_func_t;
76
77 static int do_bench = 0;
78 static int bench_pattern_len = 0;
79 static const char *bench_pattern = "";
80 static char func_name[100];
81 static bench_func_t benchs[MAX_FUNCS];
82
83 static const char *pixel_names[12] = { "16x16", "16x8", "8x16", "8x8", "8x4", "4x8", "4x4", "4x16", "4x2", "2x8", "2x4", "2x2" };
84 static const char *intra_predict_16x16_names[7] = { "v", "h", "dc", "p", "dcl", "dct", "dc8" };
85 static const char *intra_predict_8x8c_names[7] = { "dc", "h", "v", "p", "dcl", "dct", "dc8" };
86 static const char *intra_predict_4x4_names[12] = { "v", "h", "dc", "ddl", "ddr", "vr", "hd", "vl", "hu", "dcl", "dct", "dc8" };
87 static const char **intra_predict_8x8_names = intra_predict_4x4_names;
88 static const char **intra_predict_8x16c_names = intra_predict_8x8c_names;
89
90 #define set_func_name(...) snprintf( func_name, sizeof(func_name), __VA_ARGS__ )
91
92 static inline uint32_t read_time(void)
93 {
94 uint32_t a = 0;
95 #if HAVE_X86_INLINE_ASM
96 asm volatile( "lfence \n"
97 "rdtsc \n"
98 : "=a"(a) :: "edx", "memory" );
99 #elif ARCH_PPC
100 asm volatile( "mftb %0" : "=r"(a) :: "memory" );
101 #elif HAVE_ARM_INLINE_ASM // ARMv7 only
102 asm volatile( "mrc p15, 0, %0, c9, c13, 0" : "=r"(a) :: "memory" );
103 #elif ARCH_AARCH64
104 uint64_t b = 0;
105 asm volatile( "mrs %0, pmccntr_el0" : "=r"(b) :: "memory" );
106 a = b;
107 #elif ARCH_MIPS
108 asm volatile( "rdhwr %0, $2" : "=r"(a) :: "memory" );
109 #endif
110 return a;
111 }
112
113 static bench_t* get_bench( const char *name, int cpu )
114 {
115 int i, j;
116 for( i = 0; benchs[i].name && strcmp(name, benchs[i].name); i++ )
117 assert( i < MAX_FUNCS );
118 if( !benchs[i].name )
119 benchs[i].name = strdup( name );
120 if( !cpu )
121 return &benchs[i].vers[0];
122 for( j = 1; benchs[i].vers[j].cpu && benchs[i].vers[j].cpu != cpu; j++ )
123 assert( j < MAX_CPUS );
124 benchs[i].vers[j].cpu = cpu;
125 return &benchs[i].vers[j];
126 }
127
128 static int cmp_nop( const void *a, const void *b )
129 {
130 return *(uint16_t*)a - *(uint16_t*)b;
131 }
132
133 static int cmp_bench( const void *a, const void *b )
134 {
135 // asciibetical sort except preserving numbers
136 const char *sa = ((bench_func_t*)a)->name;
137 const char *sb = ((bench_func_t*)b)->name;
138 for( ;; sa++, sb++ )
139 {
140 if( !*sa && !*sb )
141 return 0;
142 if( isdigit( *sa ) && isdigit( *sb ) && isdigit( sa[1] ) != isdigit( sb[1] ) )
143 return isdigit( sa[1] ) - isdigit( sb[1] );
144 if( *sa != *sb )
145 return *sa - *sb;
146 }
147 }
148
149 static void print_bench(void)
150 {
151 uint16_t nops[10000];
152 int nfuncs, nop_time=0;
153
154 for( int i = 0; i < 10000; i++ )
155 {
156 uint32_t t = read_time();
157 nops[i] = read_time() - t;
158 }
159 qsort( nops, 10000, sizeof(uint16_t), cmp_nop );
160 for( int i = 500; i < 9500; i++ )
161 nop_time += nops[i];
162 nop_time /= 900;
163 printf( "nop: %d\n", nop_time );
164
165 for( nfuncs = 0; nfuncs < MAX_FUNCS && benchs[nfuncs].name; nfuncs++ );
166 qsort( benchs, nfuncs, sizeof(bench_func_t), cmp_bench );
167 for( int i = 0; i < nfuncs; i++ )
168 for( int j = 0; j < MAX_CPUS && (!j || benchs[i].vers[j].cpu); j++ )
169 {
170 int k;
171 bench_t *b = &benchs[i].vers[j];
172 if( !b->den )
173 continue;
174 for( k = 0; k < j && benchs[i].vers[k].pointer != b->pointer; k++ );
175 if( k < j )
176 continue;
177 printf( "%s_%s%s: %"PRId64"\n", benchs[i].name,
178 #if HAVE_MMX
179 b->cpu&X264_CPU_AVX512 ? "avx512" :
180 b->cpu&X264_CPU_AVX2 ? "avx2" :
181 b->cpu&X264_CPU_BMI2 ? "bmi2" :
182 b->cpu&X264_CPU_BMI1 ? "bmi1" :
183 b->cpu&X264_CPU_FMA3 ? "fma3" :
184 b->cpu&X264_CPU_FMA4 ? "fma4" :
185 b->cpu&X264_CPU_XOP ? "xop" :
186 b->cpu&X264_CPU_AVX ? "avx" :
187 b->cpu&X264_CPU_SSE42 ? "sse42" :
188 b->cpu&X264_CPU_SSE4 ? "sse4" :
189 b->cpu&X264_CPU_SSSE3 ? "ssse3" :
190 b->cpu&X264_CPU_SSE3 ? "sse3" :
191 b->cpu&X264_CPU_LZCNT ? "lzcnt" :
192 /* print sse2slow only if there's also a sse2fast version of the same func */
193 b->cpu&X264_CPU_SSE2_IS_SLOW && j<MAX_CPUS-1 && b[1].cpu&X264_CPU_SSE2_IS_FAST && !(b[1].cpu&X264_CPU_SSE3) ? "sse2slow" :
194 b->cpu&X264_CPU_SSE2 ? "sse2" :
195 b->cpu&X264_CPU_SSE ? "sse" :
196 b->cpu&X264_CPU_MMX ? "mmx" :
197 #elif ARCH_PPC
198 b->cpu&X264_CPU_ALTIVEC ? "altivec" :
199 #elif ARCH_ARM
200 b->cpu&X264_CPU_NEON ? "neon" :
201 b->cpu&X264_CPU_ARMV6 ? "armv6" :
202 #elif ARCH_AARCH64
203 b->cpu&X264_CPU_NEON ? "neon" :
204 b->cpu&X264_CPU_ARMV8 ? "armv8" :
205 #elif ARCH_MIPS
206 b->cpu&X264_CPU_MSA ? "msa" :
207 #endif
208 "c",
209 #if HAVE_MMX
210 b->cpu&X264_CPU_CACHELINE_32 ? "_c32" :
211 b->cpu&X264_CPU_SLOW_ATOM && b->cpu&X264_CPU_CACHELINE_64 ? "_c64_atom" :
212 b->cpu&X264_CPU_CACHELINE_64 ? "_c64" :
213 b->cpu&X264_CPU_SLOW_SHUFFLE ? "_slowshuffle" :
214 b->cpu&X264_CPU_LZCNT && b->cpu&X264_CPU_SSE3 && !(b->cpu&X264_CPU_BMI1) ? "_lzcnt" :
215 b->cpu&X264_CPU_SLOW_ATOM ? "_atom" :
216 #elif ARCH_ARM
217 b->cpu&X264_CPU_FAST_NEON_MRC ? "_fast_mrc" :
218 #endif
219 "",
220 (int64_t)(10*b->cycles/b->den - nop_time)/4 );
221 }
222 }
223
224 /* YMM and ZMM registers on x86 are turned off to save power when they haven't been
225 * used for some period of time. When they are used there will be a "warmup" period
226 * during which performance will be reduced and inconsistent which is problematic when
227 * trying to benchmark individual functions. We can work around this by periodically
228 * issuing "dummy" instructions that uses those registers to keep them powered on. */
229 static void (*simd_warmup_func)( void ) = NULL;
230 #define simd_warmup() do { if( simd_warmup_func ) simd_warmup_func(); } while( 0 )
231
232 #if ARCH_X86 || ARCH_X86_64
233 int x264_stack_pagealign( int (*func)(), int align );
234 void x264_checkasm_warmup_avx( void );
235 void x264_checkasm_warmup_avx512( void );
236
237 /* detect when callee-saved regs aren't saved
238 * needs an explicit asm check because it only sometimes crashes in normal use. */
239 intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
240 #else
241 #define x264_stack_pagealign( func, align ) func()
242 #endif
243
244 #if ARCH_AARCH64
245 intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
246 #endif
247
248 #if ARCH_ARM
249 intptr_t x264_checkasm_call_neon( intptr_t (*func)(), int *ok, ... );
250 intptr_t x264_checkasm_call_noneon( intptr_t (*func)(), int *ok, ... );
251 intptr_t (*x264_checkasm_call)( intptr_t (*func)(), int *ok, ... ) = x264_checkasm_call_noneon;
252 #endif
253
254 #define call_c1(func,...) func(__VA_ARGS__)
255
256 #if ARCH_X86_64
257 /* Evil hack: detect incorrect assumptions that 32-bit ints are zero-extended to 64-bit.
258 * This is done by clobbering the stack with junk around the stack pointer and calling the
259 * assembly function through x264_checkasm_call with added dummy arguments which forces all
260 * real arguments to be passed on the stack and not in registers. For 32-bit argument the
261 * upper half of the 64-bit register location on the stack will now contain junk. Note that
262 * this is dependant on compiler behaviour and that interrupts etc. at the wrong time may
263 * overwrite the junk written to the stack so there's no guarantee that it will always
264 * detect all functions that assumes zero-extension.
265 */
266 void x264_checkasm_stack_clobber( uint64_t clobber, ... );
267 #define call_a1(func,...) ({ \
268 uint64_t r = (rand() & 0xffff) * 0x0001000100010001ULL; \
269 x264_checkasm_stack_clobber( r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r ); /* max_args+6 */ \
270 simd_warmup(); \
271 x264_checkasm_call(( intptr_t(*)())func, &ok, 0, 0, 0, 0, __VA_ARGS__ ); })
272 #elif ARCH_AARCH64 && !defined(__APPLE__)
273 void x264_checkasm_stack_clobber( uint64_t clobber, ... );
274 #define call_a1(func,...) ({ \
275 uint64_t r = (rand() & 0xffff) * 0x0001000100010001ULL; \
276 x264_checkasm_stack_clobber( r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r ); /* max_args+8 */ \
277 x264_checkasm_call(( intptr_t(*)())func, &ok, 0, 0, 0, 0, 0, 0, __VA_ARGS__ ); })
278 #elif ARCH_X86 || ARCH_ARM
279 #define call_a1(func,...) x264_checkasm_call( (intptr_t(*)())func, &ok, __VA_ARGS__ )
280 #else
281 #define call_a1 call_c1
282 #endif
283
284 #if ARCH_ARM
285 #define call_a1_64(func,...) ((uint64_t (*)(intptr_t(*)(), int*, ...))x264_checkasm_call)( (intptr_t(*)())func, &ok, __VA_ARGS__ )
286 #else
287 #define call_a1_64 call_a1
288 #endif
289
290 #define call_bench(func,cpu,...)\
291 if( do_bench && !strncmp(func_name, bench_pattern, bench_pattern_len) )\
292 {\
293 uint64_t tsum = 0;\
294 int tcount = 0;\
295 call_a1(func, __VA_ARGS__);\
296 for( int ti = 0; ti < (cpu?BENCH_RUNS:BENCH_RUNS/4); ti++ )\
297 {\
298 simd_warmup();\
299 uint32_t t = read_time();\
300 func(__VA_ARGS__);\
301 func(__VA_ARGS__);\
302 func(__VA_ARGS__);\
303 func(__VA_ARGS__);\
304 t = read_time() - t;\
305 if( (uint64_t)t*tcount <= tsum*4 && ti > 0 )\
306 {\
307 tsum += t;\
308 tcount++;\
309 }\
310 }\
311 bench_t *b = get_bench( func_name, cpu );\
312 b->cycles += tsum;\
313 b->den += tcount;\
314 b->pointer = func;\
315 }
316
317 /* for most functions, run benchmark and correctness test at the same time.
318 * for those that modify their inputs, run the above macros separately */
319 #define call_a(func,...) ({ call_a2(func,__VA_ARGS__); call_a1(func,__VA_ARGS__); })
320 #define call_c(func,...) ({ call_c2(func,__VA_ARGS__); call_c1(func,__VA_ARGS__); })
321 #define call_a2(func,...) ({ call_bench(func,cpu_new,__VA_ARGS__); })
322 #define call_c2(func,...) ({ call_bench(func,0,__VA_ARGS__); })
323 #define call_a64(func,...) ({ call_a2(func,__VA_ARGS__); call_a1_64(func,__VA_ARGS__); })
324
325
326 static int check_pixel( int cpu_ref, int cpu_new )
327 {
328 x264_pixel_function_t pixel_c;
329 x264_pixel_function_t pixel_ref;
330 x264_pixel_function_t pixel_asm;
331 x264_predict_t predict_4x4[12];
332 x264_predict8x8_t predict_8x8[12];
333 x264_predict_8x8_filter_t predict_8x8_filter;
334 ALIGNED_16( pixel edge[36] );
335 uint16_t cost_mv[32];
336 int ret = 0, ok, used_asm;
337
338 x264_pixel_init( 0, &pixel_c );
339 x264_pixel_init( cpu_ref, &pixel_ref );
340 x264_pixel_init( cpu_new, &pixel_asm );
341 x264_predict_4x4_init( 0, predict_4x4 );
342 x264_predict_8x8_init( 0, predict_8x8, &predict_8x8_filter );
343 predict_8x8_filter( pbuf2+40, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
344
345 // maximize sum
346 for( int i = 0; i < 256; i++ )
347 {
348 int z = i|(i>>4);
349 z ^= z>>2;
350 z ^= z>>1;
351 pbuf4[i] = -(z&1) & PIXEL_MAX;
352 pbuf3[i] = ~pbuf4[i] & PIXEL_MAX;
353 }
354 // random pattern made of maxed pixel differences, in case an intermediate value overflows
355 for( int i = 256; i < 0x1000; i++ )
356 {
357 pbuf4[i] = -(pbuf1[i&~0x88]&1) & PIXEL_MAX;
358 pbuf3[i] = ~(pbuf4[i]) & PIXEL_MAX;
359 }
360
361 #define TEST_PIXEL( name, align ) \
362 ok = 1, used_asm = 0; \
363 for( int i = 0; i < ARRAY_ELEMS(pixel_c.name); i++ ) \
364 { \
365 int res_c, res_asm; \
366 if( pixel_asm.name[i] != pixel_ref.name[i] ) \
367 { \
368 set_func_name( "%s_%s", #name, pixel_names[i] ); \
369 used_asm = 1; \
370 for( int j = 0; j < 64; j++ ) \
371 { \
372 intptr_t stride1 = (j&31) == 31 ? 32 : FENC_STRIDE; \
373 res_c = call_c( pixel_c.name[i], pbuf1, stride1, pbuf2+j*!align, (intptr_t)64 ); \
374 res_asm = call_a( pixel_asm.name[i], pbuf1, stride1, pbuf2+j*!align, (intptr_t)64 ); \
375 if( res_c != res_asm ) \
376 { \
377 ok = 0; \
378 fprintf( stderr, #name "[%d]: %d != %d [FAILED]\n", i, res_c, res_asm ); \
379 break; \
380 } \
381 } \
382 for( int j = 0; j < 0x1000 && ok; j += 256 ) \
383 { \
384 res_c = pixel_c .name[i]( pbuf3+j, 16, pbuf4+j, 16 ); \
385 res_asm = pixel_asm.name[i]( pbuf3+j, 16, pbuf4+j, 16 ); \
386 if( res_c != res_asm ) \
387 { \
388 ok = 0; \
389 fprintf( stderr, #name "[%d]: overflow %d != %d\n", i, res_c, res_asm ); \
390 } \
391 } \
392 } \
393 } \
394 report( "pixel " #name " :" );
395
396 TEST_PIXEL( sad, 0 );
397 TEST_PIXEL( sad_aligned, 1 );
398 TEST_PIXEL( ssd, 1 );
399 TEST_PIXEL( satd, 0 );
400 TEST_PIXEL( sa8d, 1 );
401
402 ok = 1, used_asm = 0;
403 if( pixel_asm.sa8d_satd[PIXEL_16x16] != pixel_ref.sa8d_satd[PIXEL_16x16] )
404 {
405 set_func_name( "sa8d_satd_%s", pixel_names[PIXEL_16x16] );
406 used_asm = 1;
407 for( int j = 0; j < 64; j++ )
408 {
409 uint32_t cost8_c = pixel_c.sa8d[PIXEL_16x16]( pbuf1, 16, pbuf2, 64 );
410 uint32_t cost4_c = pixel_c.satd[PIXEL_16x16]( pbuf1, 16, pbuf2, 64 );
411 uint64_t res_a = call_a64( pixel_asm.sa8d_satd[PIXEL_16x16], pbuf1, (intptr_t)16, pbuf2, (intptr_t)64 );
412 uint32_t cost8_a = res_a;
413 uint32_t cost4_a = res_a >> 32;
414 if( cost8_a != cost8_c || cost4_a != cost4_c )
415 {
416 ok = 0;
417 fprintf( stderr, "sa8d_satd [%d]: (%d,%d) != (%d,%d) [FAILED]\n", PIXEL_16x16,
418 cost8_c, cost4_c, cost8_a, cost4_a );
419 break;
420 }
421 }
422 for( int j = 0; j < 0x1000 && ok; j += 256 ) \
423 {
424 uint32_t cost8_c = pixel_c.sa8d[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
425 uint32_t cost4_c = pixel_c.satd[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
426 uint64_t res_a = pixel_asm.sa8d_satd[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
427 uint32_t cost8_a = res_a;
428 uint32_t cost4_a = res_a >> 32;
429 if( cost8_a != cost8_c || cost4_a != cost4_c )
430 {
431 ok = 0;
432 fprintf( stderr, "sa8d_satd [%d]: overflow (%d,%d) != (%d,%d) [FAILED]\n", PIXEL_16x16,
433 cost8_c, cost4_c, cost8_a, cost4_a );
434 }
435 }
436 }
437 report( "pixel sa8d_satd :" );
438
439 #define TEST_PIXEL_X( N ) \
440 ok = 1; used_asm = 0; \
441 for( int i = 0; i < 7; i++ ) \
442 { \
443 ALIGNED_16( int res_c[4] ) = {0}; \
444 ALIGNED_16( int res_asm[4] ) = {0}; \
445 if( pixel_asm.sad_x##N[i] && pixel_asm.sad_x##N[i] != pixel_ref.sad_x##N[i] ) \
446 { \
447 set_func_name( "sad_x%d_%s", N, pixel_names[i] ); \
448 used_asm = 1; \
449 for( int j = 0; j < 64; j++ ) \
450 { \
451 pixel *pix2 = pbuf2+j; \
452 res_c[0] = pixel_c.sad[i]( pbuf1, 16, pix2, 64 ); \
453 res_c[1] = pixel_c.sad[i]( pbuf1, 16, pix2+6, 64 ); \
454 res_c[2] = pixel_c.sad[i]( pbuf1, 16, pix2+1, 64 ); \
455 if( N == 4 ) \
456 { \
457 res_c[3] = pixel_c.sad[i]( pbuf1, 16, pix2+10, 64 ); \
458 call_a( pixel_asm.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, (intptr_t)64, res_asm ); \
459 } \
460 else \
461 call_a( pixel_asm.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, (intptr_t)64, res_asm ); \
462 if( memcmp(res_c, res_asm, N*sizeof(int)) ) \
463 { \
464 ok = 0; \
465 fprintf( stderr, "sad_x"#N"[%d]: %d,%d,%d,%d != %d,%d,%d,%d [FAILED]\n", \
466 i, res_c[0], res_c[1], res_c[2], res_c[3], \
467 res_asm[0], res_asm[1], res_asm[2], res_asm[3] ); \
468 } \
469 if( N == 4 ) \
470 call_c2( pixel_c.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, (intptr_t)64, res_asm ); \
471 else \
472 call_c2( pixel_c.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, (intptr_t)64, res_asm ); \
473 } \
474 } \
475 } \
476 report( "pixel sad_x"#N" :" );
477
478 TEST_PIXEL_X(3);
479 TEST_PIXEL_X(4);
480
481 #define TEST_PIXEL_VAR( i ) \
482 if( pixel_asm.var[i] != pixel_ref.var[i] ) \
483 { \
484 set_func_name( "%s_%s", "var", pixel_names[i] ); \
485 used_asm = 1; \
486 /* abi-check wrapper can't return uint64_t, so separate it from return value check */ \
487 call_c1( pixel_c.var[i], pbuf1, 16 ); \
488 call_a1( pixel_asm.var[i], pbuf1, (intptr_t)16 ); \
489 uint64_t res_c = pixel_c.var[i]( pbuf1, 16 ); \
490 uint64_t res_asm = pixel_asm.var[i]( pbuf1, 16 ); \
491 if( res_c != res_asm ) \
492 { \
493 ok = 0; \
494 fprintf( stderr, "var[%d]: %d %d != %d %d [FAILED]\n", i, (int)res_c, (int)(res_c>>32), (int)res_asm, (int)(res_asm>>32) ); \
495 } \
496 call_c2( pixel_c.var[i], pbuf1, (intptr_t)16 ); \
497 call_a2( pixel_asm.var[i], pbuf1, (intptr_t)16 ); \
498 }
499
500 ok = 1; used_asm = 0;
501 TEST_PIXEL_VAR( PIXEL_16x16 );
502 TEST_PIXEL_VAR( PIXEL_8x16 );
503 TEST_PIXEL_VAR( PIXEL_8x8 );
504 report( "pixel var :" );
505
506 #define TEST_PIXEL_VAR2( i ) \
507 if( pixel_asm.var2[i] != pixel_ref.var2[i] ) \
508 { \
509 int res_c, res_asm; \
510 ALIGNED_ARRAY_8( int, ssd_c, [2] ); \
511 ALIGNED_ARRAY_8( int, ssd_asm,[2] ); \
512 set_func_name( "%s_%s", "var2", pixel_names[i] ); \
513 used_asm = 1; \
514 res_c = call_c( pixel_c.var2[i], pbuf1, pbuf2, ssd_c ); \
515 res_asm = call_a( pixel_asm.var2[i], pbuf1, pbuf2, ssd_asm ); \
516 if( res_c != res_asm || memcmp( ssd_c, ssd_asm, 2*sizeof(int) ) ) \
517 { \
518 ok = 0; \
519 fprintf( stderr, "var2[%d]: {%d, %d, %d} != {%d, %d, %d} [FAILED]\n", i, res_c, ssd_c[0], ssd_c[1], res_asm, ssd_asm[0], ssd_asm[1] ); \
520 } \
521 }
522
523 ok = 1; used_asm = 0;
524 TEST_PIXEL_VAR2( PIXEL_8x16 );
525 TEST_PIXEL_VAR2( PIXEL_8x8 );
526 report( "pixel var2 :" );
527
528 ok = 1; used_asm = 0;
529 for( int i = 0; i < 4; i++ )
530 if( pixel_asm.hadamard_ac[i] != pixel_ref.hadamard_ac[i] )
531 {
532 set_func_name( "hadamard_ac_%s", pixel_names[i] );
533 used_asm = 1;
534 for( int j = 0; j < 32; j++ )
535 {
536 pixel *pix = (j&16 ? pbuf1 : pbuf3) + (j&15)*256;
537 call_c1( pixel_c.hadamard_ac[i], pbuf1, (intptr_t)16 );
538 call_a1( pixel_asm.hadamard_ac[i], pbuf1, (intptr_t)16 );
539 uint64_t rc = pixel_c.hadamard_ac[i]( pix, 16 );
540 uint64_t ra = pixel_asm.hadamard_ac[i]( pix, 16 );
541 if( rc != ra )
542 {
543 ok = 0;
544 fprintf( stderr, "hadamard_ac[%d]: %d,%d != %d,%d\n", i, (int)rc, (int)(rc>>32), (int)ra, (int)(ra>>32) );
545 break;
546 }
547 }
548 call_c2( pixel_c.hadamard_ac[i], pbuf1, (intptr_t)16 );
549 call_a2( pixel_asm.hadamard_ac[i], pbuf1, (intptr_t)16 );
550 }
551 report( "pixel hadamard_ac :" );
552
553 // maximize sum
554 for( int i = 0; i < 32; i++ )
555 for( int j = 0; j < 16; j++ )
556 pbuf4[16*i+j] = -((i+j)&1) & PIXEL_MAX;
557 ok = 1; used_asm = 0;
558 if( pixel_asm.vsad != pixel_ref.vsad )
559 {
560 for( int h = 2; h <= 32; h += 2 )
561 {
562 int res_c, res_asm;
563 set_func_name( "vsad" );
564 used_asm = 1;
565 for( int j = 0; j < 2 && ok; j++ )
566 {
567 pixel *p = j ? pbuf4 : pbuf1;
568 res_c = call_c( pixel_c.vsad, p, (intptr_t)16, h );
569 res_asm = call_a( pixel_asm.vsad, p, (intptr_t)16, h );
570 if( res_c != res_asm )
571 {
572 ok = 0;
573 fprintf( stderr, "vsad: height=%d, %d != %d\n", h, res_c, res_asm );
574 break;
575 }
576 }
577 }
578 }
579 report( "pixel vsad :" );
580
581 ok = 1; used_asm = 0;
582 if( pixel_asm.asd8 != pixel_ref.asd8 )
583 {
584 set_func_name( "asd8" );
585 used_asm = 1;
586 int res_c = call_c( pixel_c.asd8, pbuf1, (intptr_t)8, pbuf2, (intptr_t)8, 16 );
587 int res_a = call_a( pixel_asm.asd8, pbuf1, (intptr_t)8, pbuf2, (intptr_t)8, 16 );
588 if( res_c != res_a )
589 {
590 ok = 0;
591 fprintf( stderr, "asd: %d != %d\n", res_c, res_a );
592 }
593 }
594 report( "pixel asd :" );
595
596 #define TEST_INTRA_X3( name, i8x8, ... ) \
597 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
598 { \
599 ALIGNED_16( int res_c[4] ); \
600 ALIGNED_16( int res_asm[4] ); \
601 set_func_name( #name ); \
602 used_asm = 1; \
603 call_c( pixel_c.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_c ); \
604 call_a( pixel_asm.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_asm ); \
605 if( memcmp(res_c, res_asm, 3 * sizeof(*res_c)) ) \
606 { \
607 ok = 0; \
608 fprintf( stderr, #name": %d,%d,%d != %d,%d,%d [FAILED]\n", \
609 res_c[0], res_c[1], res_c[2], \
610 res_asm[0], res_asm[1], res_asm[2] ); \
611 } \
612 }
613
614 #define TEST_INTRA_X9( name, cmp ) \
615 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
616 { \
617 set_func_name( #name ); \
618 used_asm = 1; \
619 ALIGNED_ARRAY_64( uint16_t, bitcosts,[17] ); \
620 for( int i=0; i<17; i++ ) \
621 bitcosts[i] = 9*(i!=8); \
622 memcpy( pbuf3, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) ); \
623 memcpy( pbuf4, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) ); \
624 for( int i=0; i<32; i++ ) \
625 { \
626 pixel *fenc = pbuf1+48+i*12; \
627 pixel *fdec1 = pbuf3+48+i*12; \
628 pixel *fdec2 = pbuf4+48+i*12; \
629 int pred_mode = i%9; \
630 int res_c = INT_MAX; \
631 for( int j=0; j<9; j++ ) \
632 { \
633 predict_4x4[j]( fdec1 ); \
634 int cost = pixel_c.cmp[PIXEL_4x4]( fenc, FENC_STRIDE, fdec1, FDEC_STRIDE ) + 9*(j!=pred_mode); \
635 if( cost < (uint16_t)res_c ) \
636 res_c = cost + (j<<16); \
637 } \
638 predict_4x4[res_c>>16]( fdec1 ); \
639 int res_a = call_a( pixel_asm.name, fenc, fdec2, bitcosts+8-pred_mode ); \
640 if( res_c != res_a ) \
641 { \
642 ok = 0; \
643 fprintf( stderr, #name": %d,%d != %d,%d [FAILED]\n", res_c>>16, res_c&0xffff, res_a>>16, res_a&0xffff ); \
644 break; \
645 } \
646 if( memcmp(fdec1, fdec2, 4*FDEC_STRIDE*sizeof(pixel)) ) \
647 { \
648 ok = 0; \
649 fprintf( stderr, #name" [FAILED]\n" ); \
650 for( int j=0; j<16; j++ ) \
651 fprintf( stderr, "%02x ", fdec1[(j&3)+(j>>2)*FDEC_STRIDE] ); \
652 fprintf( stderr, "\n" ); \
653 for( int j=0; j<16; j++ ) \
654 fprintf( stderr, "%02x ", fdec2[(j&3)+(j>>2)*FDEC_STRIDE] ); \
655 fprintf( stderr, "\n" ); \
656 break; \
657 } \
658 } \
659 }
660
661 #define TEST_INTRA8_X9( name, cmp ) \
662 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
663 { \
664 set_func_name( #name ); \
665 used_asm = 1; \
666 ALIGNED_ARRAY_64( uint16_t, bitcosts,[17] ); \
667 ALIGNED_ARRAY_16( uint16_t, satds_c,[16] ); \
668 ALIGNED_ARRAY_16( uint16_t, satds_a,[16] ); \
669 memset( satds_c, 0, 16 * sizeof(*satds_c) ); \
670 memset( satds_a, 0, 16 * sizeof(*satds_a) ); \
671 for( int i=0; i<17; i++ ) \
672 bitcosts[i] = 9*(i!=8); \
673 for( int i=0; i<32; i++ ) \
674 { \
675 pixel *fenc = pbuf1+48+i*12; \
676 pixel *fdec1 = pbuf3+48+i*12; \
677 pixel *fdec2 = pbuf4+48+i*12; \
678 int pred_mode = i%9; \
679 int res_c = INT_MAX; \
680 predict_8x8_filter( fdec1, edge, ALL_NEIGHBORS, ALL_NEIGHBORS ); \
681 for( int j=0; j<9; j++ ) \
682 { \
683 predict_8x8[j]( fdec1, edge ); \
684 satds_c[j] = pixel_c.cmp[PIXEL_8x8]( fenc, FENC_STRIDE, fdec1, FDEC_STRIDE ) + 9*(j!=pred_mode); \
685 if( satds_c[j] < (uint16_t)res_c ) \
686 res_c = satds_c[j] + (j<<16); \
687 } \
688 predict_8x8[res_c>>16]( fdec1, edge ); \
689 int res_a = call_a( pixel_asm.name, fenc, fdec2, edge, bitcosts+8-pred_mode, satds_a ); \
690 if( res_c != res_a || memcmp(satds_c, satds_a, 16 * sizeof(*satds_c)) ) \
691 { \
692 ok = 0; \
693 fprintf( stderr, #name": %d,%d != %d,%d [FAILED]\n", res_c>>16, res_c&0xffff, res_a>>16, res_a&0xffff ); \
694 for( int j = 0; j < 9; j++ ) \
695 fprintf( stderr, "%5d ", satds_c[j]); \
696 fprintf( stderr, "\n" ); \
697 for( int j = 0; j < 9; j++ ) \
698 fprintf( stderr, "%5d ", satds_a[j]); \
699 fprintf( stderr, "\n" ); \
700 break; \
701 } \
702 for( int j=0; j<8; j++ ) \
703 if( memcmp(fdec1+j*FDEC_STRIDE, fdec2+j*FDEC_STRIDE, 8*sizeof(pixel)) ) \
704 ok = 0; \
705 if( !ok ) \
706 { \
707 fprintf( stderr, #name" [FAILED]\n" ); \
708 for( int j=0; j<8; j++ ) \
709 { \
710 for( int k=0; k<8; k++ ) \
711 fprintf( stderr, "%02x ", fdec1[k+j*FDEC_STRIDE] ); \
712 fprintf( stderr, "\n" ); \
713 } \
714 fprintf( stderr, "\n" ); \
715 for( int j=0; j<8; j++ ) \
716 { \
717 for( int k=0; k<8; k++ ) \
718 fprintf( stderr, "%02x ", fdec2[k+j*FDEC_STRIDE] ); \
719 fprintf( stderr, "\n" ); \
720 } \
721 fprintf( stderr, "\n" ); \
722 break; \
723 } \
724 } \
725 }
726
727 memcpy( pbuf3, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) );
728 ok = 1; used_asm = 0;
729 TEST_INTRA_X3( intra_satd_x3_16x16, 0 );
730 TEST_INTRA_X3( intra_satd_x3_8x16c, 0 );
731 TEST_INTRA_X3( intra_satd_x3_8x8c, 0 );
732 TEST_INTRA_X3( intra_sa8d_x3_8x8, 1, edge );
733 TEST_INTRA_X3( intra_satd_x3_4x4, 0 );
734 report( "intra satd_x3 :" );
735 ok = 1; used_asm = 0;
736 TEST_INTRA_X3( intra_sad_x3_16x16, 0 );
737 TEST_INTRA_X3( intra_sad_x3_8x16c, 0 );
738 TEST_INTRA_X3( intra_sad_x3_8x8c, 0 );
739 TEST_INTRA_X3( intra_sad_x3_8x8, 1, edge );
740 TEST_INTRA_X3( intra_sad_x3_4x4, 0 );
741 report( "intra sad_x3 :" );
742 ok = 1; used_asm = 0;
743 TEST_INTRA_X9( intra_satd_x9_4x4, satd );
744 TEST_INTRA8_X9( intra_sa8d_x9_8x8, sa8d );
745 report( "intra satd_x9 :" );
746 ok = 1; used_asm = 0;
747 TEST_INTRA_X9( intra_sad_x9_4x4, sad );
748 TEST_INTRA8_X9( intra_sad_x9_8x8, sad );
749 report( "intra sad_x9 :" );
750
751 ok = 1; used_asm = 0;
752 if( pixel_asm.ssd_nv12_core != pixel_ref.ssd_nv12_core )
753 {
754 used_asm = 1;
755 set_func_name( "ssd_nv12" );
756 uint64_t res_u_c, res_v_c, res_u_a, res_v_a;
757 for( int w = 8; w <= 360; w += 8 )
758 {
759 pixel_c.ssd_nv12_core( pbuf1, 368, pbuf2, 368, w, 8, &res_u_c, &res_v_c );
760 pixel_asm.ssd_nv12_core( pbuf1, 368, pbuf2, 368, w, 8, &res_u_a, &res_v_a );
761 if( res_u_c != res_u_a || res_v_c != res_v_a )
762 {
763 ok = 0;
764 fprintf( stderr, "ssd_nv12: %"PRIu64",%"PRIu64" != %"PRIu64",%"PRIu64"\n",
765 res_u_c, res_v_c, res_u_a, res_v_a );
766 }
767 }
768 call_c( pixel_c.ssd_nv12_core, pbuf1, (intptr_t)368, pbuf2, (intptr_t)368, 360, 8, &res_u_c, &res_v_c );
769 call_a( pixel_asm.ssd_nv12_core, pbuf1, (intptr_t)368, pbuf2, (intptr_t)368, 360, 8, &res_u_a, &res_v_a );
770 }
771 report( "ssd_nv12 :" );
772
773 if( pixel_asm.ssim_4x4x2_core != pixel_ref.ssim_4x4x2_core ||
774 pixel_asm.ssim_end4 != pixel_ref.ssim_end4 )
775 {
776 int cnt;
777 float res_c, res_a;
778 ALIGNED_16( int sums[5][4] ) = {{0}};
779 used_asm = ok = 1;
780 x264_emms();
781 res_c = x264_pixel_ssim_wxh( &pixel_c, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
782 res_a = x264_pixel_ssim_wxh( &pixel_asm, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
783 if( fabs( res_c - res_a ) > 1e-6 )
784 {
785 ok = 0;
786 fprintf( stderr, "ssim: %.7f != %.7f [FAILED]\n", res_c, res_a );
787 }
788 set_func_name( "ssim_core" );
789 call_c( pixel_c.ssim_4x4x2_core, pbuf1+2, (intptr_t)32, pbuf2+2, (intptr_t)32, sums );
790 call_a( pixel_asm.ssim_4x4x2_core, pbuf1+2, (intptr_t)32, pbuf2+2, (intptr_t)32, sums );
791 set_func_name( "ssim_end" );
792 call_c2( pixel_c.ssim_end4, sums, sums, 4 );
793 call_a2( pixel_asm.ssim_end4, sums, sums, 4 );
794 /* check incorrect assumptions that 32-bit ints are zero-extended to 64-bit */
795 call_c1( pixel_c.ssim_end4, sums, sums, 3 );
796 call_a1( pixel_asm.ssim_end4, sums, sums, 3 );
797 report( "ssim :" );
798 }
799
800 ok = 1; used_asm = 0;
801 for( int i = 0; i < 32; i++ )
802 cost_mv[i] = i*10;
803 for( int i = 0; i < 100 && ok; i++ )
804 if( pixel_asm.ads[i&3] != pixel_ref.ads[i&3] )
805 {
806 ALIGNED_16( uint16_t sums[72] );
807 ALIGNED_16( int dc[4] );
808 ALIGNED_16( int16_t mvs_a[48] );
809 ALIGNED_16( int16_t mvs_c[48] );
810 int mvn_a, mvn_c;
811 int thresh = rand() & 0x3fff;
812 set_func_name( "esa_ads" );
813 for( int j = 0; j < 72; j++ )
814 sums[j] = rand() & 0x3fff;
815 for( int j = 0; j < 4; j++ )
816 dc[j] = rand() & 0x3fff;
817 used_asm = 1;
818 mvn_c = call_c( pixel_c.ads[i&3], dc, sums, 32, cost_mv, mvs_c, 28, thresh );
819 mvn_a = call_a( pixel_asm.ads[i&3], dc, sums, 32, cost_mv, mvs_a, 28, thresh );
820 if( mvn_c != mvn_a || memcmp( mvs_c, mvs_a, mvn_c*sizeof(*mvs_c) ) )
821 {
822 ok = 0;
823 printf( "c%d: ", i&3 );
824 for( int j = 0; j < mvn_c; j++ )
825 printf( "%d ", mvs_c[j] );
826 printf( "\na%d: ", i&3 );
827 for( int j = 0; j < mvn_a; j++ )
828 printf( "%d ", mvs_a[j] );
829 printf( "\n\n" );
830 }
831 }
832 report( "esa ads:" );
833
834 return ret;
835 }
836
837 static int check_dct( int cpu_ref, int cpu_new )
838 {
839 x264_dct_function_t dct_c;
840 x264_dct_function_t dct_ref;
841 x264_dct_function_t dct_asm;
842 x264_quant_function_t qf;
843 int ret = 0, ok, used_asm, interlace = 0;
844 ALIGNED_ARRAY_64( dctcoef, dct1, [16],[16] );
845 ALIGNED_ARRAY_64( dctcoef, dct2, [16],[16] );
846 ALIGNED_ARRAY_64( dctcoef, dct4, [16],[16] );
847 ALIGNED_ARRAY_64( dctcoef, dct8, [4],[64] );
848 ALIGNED_16( dctcoef dctdc[2][8] );
849 x264_t h_buf;
850 x264_t *h = &h_buf;
851
852 x264_dct_init( 0, &dct_c );
853 x264_dct_init( cpu_ref, &dct_ref);
854 x264_dct_init( cpu_new, &dct_asm );
855
856 memset( h, 0, sizeof(*h) );
857 x264_param_default( &h->param );
858 h->sps->i_chroma_format_idc = 1;
859 h->chroma_qp_table = i_chroma_qp_table + 12;
860 h->param.analyse.i_luma_deadzone[0] = 0;
861 h->param.analyse.i_luma_deadzone[1] = 0;
862 h->param.analyse.b_transform_8x8 = 1;
863 for( int i = 0; i < 6; i++ )
864 h->sps->scaling_list[i] = x264_cqm_flat16;
865 x264_cqm_init( h );
866 x264_quant_init( h, 0, &qf );
867
868 /* overflow test cases */
869 for( int i = 0; i < 5; i++ )
870 {
871 pixel *enc = &pbuf3[16*i*FENC_STRIDE];
872 pixel *dec = &pbuf4[16*i*FDEC_STRIDE];
873
874 for( int j = 0; j < 16; j++ )
875 {
876 int cond_a = (i < 2) ? 1 : ((j&3) == 0 || (j&3) == (i-1));
877 int cond_b = (i == 0) ? 1 : !cond_a;
878 enc[0] = enc[1] = enc[4] = enc[5] = enc[8] = enc[9] = enc[12] = enc[13] = cond_a ? PIXEL_MAX : 0;
879 enc[2] = enc[3] = enc[6] = enc[7] = enc[10] = enc[11] = enc[14] = enc[15] = cond_b ? PIXEL_MAX : 0;
880
881 for( int k = 0; k < 4; k++ )
882 dec[k] = PIXEL_MAX - enc[k];
883
884 enc += FENC_STRIDE;
885 dec += FDEC_STRIDE;
886 }
887 }
888
889 #define TEST_DCT( name, t1, t2, size ) \
890 if( dct_asm.name != dct_ref.name ) \
891 { \
892 set_func_name( #name ); \
893 used_asm = 1; \
894 pixel *enc = pbuf3; \
895 pixel *dec = pbuf4; \
896 for( int j = 0; j < 5; j++) \
897 { \
898 call_c( dct_c.name, t1, &pbuf1[j*64], &pbuf2[j*64] ); \
899 call_a( dct_asm.name, t2, &pbuf1[j*64], &pbuf2[j*64] ); \
900 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
901 { \
902 ok = 0; \
903 fprintf( stderr, #name " [FAILED]\n" ); \
904 for( int k = 0; k < size; k++ )\
905 printf( "%d ", ((dctcoef*)t1)[k] );\
906 printf("\n");\
907 for( int k = 0; k < size; k++ )\
908 printf( "%d ", ((dctcoef*)t2)[k] );\
909 printf("\n");\
910 break; \
911 } \
912 call_c( dct_c.name, t1, enc, dec ); \
913 call_a( dct_asm.name, t2, enc, dec ); \
914 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
915 { \
916 ok = 0; \
917 fprintf( stderr, #name " [FAILED] (overflow)\n" ); \
918 break; \
919 } \
920 enc += 16*FENC_STRIDE; \
921 dec += 16*FDEC_STRIDE; \
922 } \
923 }
924 ok = 1; used_asm = 0;
925 TEST_DCT( sub4x4_dct, dct1[0], dct2[0], 16 );
926 TEST_DCT( sub8x8_dct, dct1, dct2, 16*4 );
927 TEST_DCT( sub8x8_dct_dc, dctdc[0], dctdc[1], 4 );
928 TEST_DCT( sub8x16_dct_dc, dctdc[0], dctdc[1], 8 );
929 TEST_DCT( sub16x16_dct, dct1, dct2, 16*16 );
930 report( "sub_dct4 :" );
931
932 ok = 1; used_asm = 0;
933 TEST_DCT( sub8x8_dct8, (void*)dct1[0], (void*)dct2[0], 64 );
934 TEST_DCT( sub16x16_dct8, (void*)dct1, (void*)dct2, 64*4 );
935 report( "sub_dct8 :" );
936 #undef TEST_DCT
937
938 // fdct and idct are denormalized by different factors, so quant/dequant
939 // is needed to force the coefs into the right range.
940 dct_c.sub16x16_dct( dct4, pbuf1, pbuf2 );
941 dct_c.sub16x16_dct8( dct8, pbuf1, pbuf2 );
942 for( int i = 0; i < 16; i++ )
943 {
944 qf.quant_4x4( dct4[i], h->quant4_mf[CQM_4IY][20], h->quant4_bias[CQM_4IY][20] );
945 qf.dequant_4x4( dct4[i], h->dequant4_mf[CQM_4IY], 20 );
946 }
947 for( int i = 0; i < 4; i++ )
948 {
949 qf.quant_8x8( dct8[i], h->quant8_mf[CQM_8IY][20], h->quant8_bias[CQM_8IY][20] );
950 qf.dequant_8x8( dct8[i], h->dequant8_mf[CQM_8IY], 20 );
951 }
952 x264_cqm_delete( h );
953
954 #define TEST_IDCT( name, src ) \
955 if( dct_asm.name != dct_ref.name ) \
956 { \
957 set_func_name( #name ); \
958 used_asm = 1; \
959 memcpy( pbuf3, pbuf1, 32*32 * sizeof(pixel) ); \
960 memcpy( pbuf4, pbuf1, 32*32 * sizeof(pixel) ); \
961 memcpy( dct1, src, 256 * sizeof(dctcoef) ); \
962 memcpy( dct2, src, 256 * sizeof(dctcoef) ); \
963 call_c1( dct_c.name, pbuf3, (void*)dct1 ); \
964 call_a1( dct_asm.name, pbuf4, (void*)dct2 ); \
965 if( memcmp( pbuf3, pbuf4, 32*32 * sizeof(pixel) ) ) \
966 { \
967 ok = 0; \
968 fprintf( stderr, #name " [FAILED]\n" ); \
969 } \
970 call_c2( dct_c.name, pbuf3, (void*)dct1 ); \
971 call_a2( dct_asm.name, pbuf4, (void*)dct2 ); \
972 }
973 ok = 1; used_asm = 0;
974 TEST_IDCT( add4x4_idct, dct4 );
975 TEST_IDCT( add8x8_idct, dct4 );
976 TEST_IDCT( add8x8_idct_dc, dct4 );
977 TEST_IDCT( add16x16_idct, dct4 );
978 TEST_IDCT( add16x16_idct_dc, dct4 );
979 report( "add_idct4 :" );
980
981 ok = 1; used_asm = 0;
982 TEST_IDCT( add8x8_idct8, dct8 );
983 TEST_IDCT( add16x16_idct8, dct8 );
984 report( "add_idct8 :" );
985 #undef TEST_IDCT
986
987 #define TEST_DCTDC( name )\
988 ok = 1; used_asm = 0;\
989 if( dct_asm.name != dct_ref.name )\
990 {\
991 set_func_name( #name );\
992 used_asm = 1;\
993 uint16_t *p = (uint16_t*)buf1;\
994 for( int i = 0; i < 16 && ok; i++ )\
995 {\
996 for( int j = 0; j < 16; j++ )\
997 dct1[0][j] = !i ? (j^j>>1^j>>2^j>>3)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
998 : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
999 : ((*p++)&0x1fff)-0x1000; /* general case */\
1000 memcpy( dct2, dct1, 16 * sizeof(dctcoef) );\
1001 call_c1( dct_c.name, dct1[0] );\
1002 call_a1( dct_asm.name, dct2[0] );\
1003 if( memcmp( dct1, dct2, 16 * sizeof(dctcoef) ) )\
1004 ok = 0;\
1005 }\
1006 call_c2( dct_c.name, dct1[0] );\
1007 call_a2( dct_asm.name, dct2[0] );\
1008 }\
1009 report( #name " :" );
1010
1011 TEST_DCTDC( dct4x4dc );
1012 TEST_DCTDC( idct4x4dc );
1013 #undef TEST_DCTDC
1014
1015 #define TEST_DCTDC_CHROMA( name )\
1016 ok = 1; used_asm = 0;\
1017 if( dct_asm.name != dct_ref.name )\
1018 {\
1019 set_func_name( #name );\
1020 used_asm = 1;\
1021 uint16_t *p = (uint16_t*)buf1;\
1022 for( int i = 0; i < 16 && ok; i++ )\
1023 {\
1024 for( int j = 0; j < 8; j++ )\
1025 dct1[j][0] = !i ? (j^j>>1^j>>2)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
1026 : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
1027 : ((*p++)&0x1fff)-0x1000; /* general case */\
1028 memcpy( dct2, dct1, 8*16 * sizeof(dctcoef) );\
1029 call_c1( dct_c.name, dctdc[0], dct1 );\
1030 call_a1( dct_asm.name, dctdc[1], dct2 );\
1031 if( memcmp( dctdc[0], dctdc[1], 8 * sizeof(dctcoef) ) || memcmp( dct1, dct2, 8*16 * sizeof(dctcoef) ) )\
1032 {\
1033 ok = 0;\
1034 fprintf( stderr, #name " [FAILED]\n" ); \
1035 }\
1036 }\
1037 call_c2( dct_c.name, dctdc[0], dct1 );\
1038 call_a2( dct_asm.name, dctdc[1], dct2 );\
1039 }\
1040 report( #name " :" );
1041
1042 TEST_DCTDC_CHROMA( dct2x4dc );
1043 #undef TEST_DCTDC_CHROMA
1044
1045 x264_zigzag_function_t zigzag_c[2];
1046 x264_zigzag_function_t zigzag_ref[2];
1047 x264_zigzag_function_t zigzag_asm[2];
1048
1049 ALIGNED_ARRAY_64( dctcoef, level1,[64] );
1050 ALIGNED_ARRAY_64( dctcoef, level2,[64] );
1051
1052 #define TEST_ZIGZAG_SCAN( name, t1, t2, dct, size ) \
1053 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1054 { \
1055 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1056 used_asm = 1; \
1057 for( int i = 0; i < size*size; i++ ) \
1058 dct[i] = i; \
1059 call_c( zigzag_c[interlace].name, t1, dct ); \
1060 call_a( zigzag_asm[interlace].name, t2, dct ); \
1061 if( memcmp( t1, t2, size*size*sizeof(dctcoef) ) ) \
1062 { \
1063 ok = 0; \
1064 for( int i = 0; i < 2; i++ ) \
1065 { \
1066 dctcoef *d = (dctcoef*)(i ? t2 : t1); \
1067 for( int j = 0; j < size; j++ ) \
1068 { \
1069 for( int k = 0; k < size; k++ ) \
1070 fprintf( stderr, "%2d ", d[k+j*8] ); \
1071 fprintf( stderr, "\n" ); \
1072 } \
1073 fprintf( stderr, "\n" ); \
1074 } \
1075 fprintf( stderr, #name " [FAILED]\n" ); \
1076 } \
1077 }
1078
1079 #define TEST_ZIGZAG_SUB( name, t1, t2, size ) \
1080 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1081 { \
1082 int nz_a, nz_c; \
1083 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1084 used_asm = 1; \
1085 memcpy( pbuf3, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
1086 memcpy( pbuf4, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
1087 nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
1088 nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
1089 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( pbuf3, pbuf4, 16*FDEC_STRIDE*sizeof(pixel) ) || nz_c != nz_a ) \
1090 { \
1091 ok = 0; \
1092 fprintf( stderr, #name " [FAILED]\n" ); \
1093 } \
1094 call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
1095 call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
1096 }
1097
1098 #define TEST_ZIGZAG_SUBAC( name, t1, t2 ) \
1099 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1100 { \
1101 int nz_a, nz_c; \
1102 dctcoef dc_a, dc_c; \
1103 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1104 used_asm = 1; \
1105 for( int i = 0; i < 2; i++ ) \
1106 { \
1107 memcpy( pbuf3, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
1108 memcpy( pbuf4, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
1109 for( int j = 0; j < 4; j++ ) \
1110 { \
1111 memcpy( pbuf3 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
1112 memcpy( pbuf4 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
1113 } \
1114 nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
1115 nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
1116 if( memcmp( t1+1, t2+1, 15*sizeof(dctcoef) ) || memcmp( pbuf3, pbuf4, 16*FDEC_STRIDE * sizeof(pixel) ) || nz_c != nz_a || dc_c != dc_a ) \
1117 { \
1118 ok = 0; \
1119 fprintf( stderr, #name " [FAILED]\n" ); \
1120 break; \
1121 } \
1122 } \
1123 call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
1124 call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
1125 }
1126
1127 #define TEST_INTERLEAVE( name, t1, t2, dct, size ) \
1128 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1129 { \
1130 for( int j = 0; j < 100; j++ ) \
1131 { \
1132 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1133 used_asm = 1; \
1134 memcpy(dct, buf1, size*sizeof(dctcoef)); \
1135 for( int i = 0; i < size; i++ ) \
1136 dct[i] = rand()&0x1F ? 0 : dct[i]; \
1137 memcpy(buf3, buf4, 10); \
1138 call_c( zigzag_c[interlace].name, t1, dct, buf3 ); \
1139 call_a( zigzag_asm[interlace].name, t2, dct, buf4 ); \
1140 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( buf3, buf4, 10 ) ) \
1141 { \
1142 ok = 0; printf("%d: %d %d %d %d\n%d %d %d %d\n\n",memcmp( t1, t2, size*sizeof(dctcoef) ),buf3[0], buf3[1], buf3[8], buf3[9], buf4[0], buf4[1], buf4[8], buf4[9]);break;\
1143 } \
1144 } \
1145 }
1146
1147 x264_zigzag_init( 0, &zigzag_c[0], &zigzag_c[1] );
1148 x264_zigzag_init( cpu_ref, &zigzag_ref[0], &zigzag_ref[1] );
1149 x264_zigzag_init( cpu_new, &zigzag_asm[0], &zigzag_asm[1] );
1150
1151 ok = 1; used_asm = 0;
1152 TEST_INTERLEAVE( interleave_8x8_cavlc, level1, level2, dct8[0], 64 );
1153 report( "zigzag_interleave :" );
1154
1155 for( interlace = 0; interlace <= 1; interlace++ )
1156 {
1157 ok = 1; used_asm = 0;
1158 TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, dct8[0], 8 );
1159 TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 4 );
1160 TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
1161 TEST_ZIGZAG_SUB( sub_8x8, level1, level2, 64 );
1162 TEST_ZIGZAG_SUBAC( sub_4x4ac, level1, level2 );
1163 report( interlace ? "zigzag_field :" : "zigzag_frame :" );
1164 }
1165 #undef TEST_ZIGZAG_SCAN
1166 #undef TEST_ZIGZAG_SUB
1167
1168 return ret;
1169 }
1170
1171 static int check_mc( int cpu_ref, int cpu_new )
1172 {
1173 x264_mc_functions_t mc_c;
1174 x264_mc_functions_t mc_ref;
1175 x264_mc_functions_t mc_a;
1176 x264_pixel_function_t pixf;
1177
1178 pixel *src = &(pbuf1)[2*64+2];
1179 pixel *src2[4] = { &(pbuf1)[3*64+2], &(pbuf1)[5*64+2],
1180 &(pbuf1)[7*64+2], &(pbuf1)[9*64+2] };
1181 pixel *dst1 = pbuf3;
1182 pixel *dst2 = pbuf4;
1183
1184 int ret = 0, ok, used_asm;
1185
1186 x264_mc_init( 0, &mc_c, 0 );
1187 x264_mc_init( cpu_ref, &mc_ref, 0 );
1188 x264_mc_init( cpu_new, &mc_a, 0 );
1189 x264_pixel_init( 0, &pixf );
1190
1191 #define MC_TEST_LUMA( w, h ) \
1192 if( mc_a.mc_luma != mc_ref.mc_luma && !(w&(w-1)) && h<=16 ) \
1193 { \
1194 const x264_weight_t *weight = x264_weight_none; \
1195 set_func_name( "mc_luma_%dx%d", w, h ); \
1196 used_asm = 1; \
1197 for( int i = 0; i < 1024; i++ ) \
1198 pbuf3[i] = pbuf4[i] = 0xCD; \
1199 call_c( mc_c.mc_luma, dst1, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1200 call_a( mc_a.mc_luma, dst2, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1201 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1202 { \
1203 fprintf( stderr, "mc_luma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
1204 ok = 0; \
1205 } \
1206 } \
1207 if( mc_a.get_ref != mc_ref.get_ref ) \
1208 { \
1209 pixel *ref = dst2; \
1210 intptr_t ref_stride = 32; \
1211 int w_checked = ( ( sizeof(pixel) == 2 && (w == 12 || w == 20)) ? w-2 : w ); \
1212 const x264_weight_t *weight = x264_weight_none; \
1213 set_func_name( "get_ref_%dx%d", w_checked, h ); \
1214 used_asm = 1; \
1215 for( int i = 0; i < 1024; i++ ) \
1216 pbuf3[i] = pbuf4[i] = 0xCD; \
1217 call_c( mc_c.mc_luma, dst1, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1218 ref = (pixel*)call_a( mc_a.get_ref, ref, &ref_stride, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1219 for( int i = 0; i < h; i++ ) \
1220 if( memcmp( dst1+i*32, ref+i*ref_stride, w_checked * sizeof(pixel) ) ) \
1221 { \
1222 fprintf( stderr, "get_ref[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w_checked, h ); \
1223 ok = 0; \
1224 break; \
1225 } \
1226 }
1227
1228 #define MC_TEST_CHROMA( w, h ) \
1229 if( mc_a.mc_chroma != mc_ref.mc_chroma ) \
1230 { \
1231 set_func_name( "mc_chroma_%dx%d", w, h ); \
1232 used_asm = 1; \
1233 for( int i = 0; i < 1024; i++ ) \
1234 pbuf3[i] = pbuf4[i] = 0xCD; \
1235 call_c( mc_c.mc_chroma, dst1, dst1+8, (intptr_t)16, src, (intptr_t)64, dx, dy, w, h ); \
1236 call_a( mc_a.mc_chroma, dst2, dst2+8, (intptr_t)16, src, (intptr_t)64, dx, dy, w, h ); \
1237 /* mc_chroma width=2 may write garbage to the right of dst. ignore that. */ \
1238 for( int j = 0; j < h; j++ ) \
1239 for( int i = w; i < 8; i++ ) \
1240 { \
1241 dst2[i+j*16+8] = dst1[i+j*16+8]; \
1242 dst2[i+j*16 ] = dst1[i+j*16 ]; \
1243 } \
1244 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1245 { \
1246 fprintf( stderr, "mc_chroma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
1247 ok = 0; \
1248 } \
1249 }
1250 ok = 1; used_asm = 0;
1251 for( int dy = -8; dy < 8; dy++ )
1252 for( int dx = -128; dx < 128; dx++ )
1253 {
1254 if( rand()&15 ) continue; // running all of them is too slow
1255 MC_TEST_LUMA( 20, 18 );
1256 MC_TEST_LUMA( 16, 16 );
1257 MC_TEST_LUMA( 16, 8 );
1258 MC_TEST_LUMA( 12, 10 );
1259 MC_TEST_LUMA( 8, 16 );
1260 MC_TEST_LUMA( 8, 8 );
1261 MC_TEST_LUMA( 8, 4 );
1262 MC_TEST_LUMA( 4, 8 );
1263 MC_TEST_LUMA( 4, 4 );
1264 }
1265 report( "mc luma :" );
1266
1267 ok = 1; used_asm = 0;
1268 for( int dy = -1; dy < 9; dy++ )
1269 for( int dx = -128; dx < 128; dx++ )
1270 {
1271 if( rand()&15 ) continue;
1272 MC_TEST_CHROMA( 8, 8 );
1273 MC_TEST_CHROMA( 8, 4 );
1274 MC_TEST_CHROMA( 4, 8 );
1275 MC_TEST_CHROMA( 4, 4 );
1276 MC_TEST_CHROMA( 4, 2 );
1277 MC_TEST_CHROMA( 2, 4 );
1278 MC_TEST_CHROMA( 2, 2 );
1279 }
1280 report( "mc chroma :" );
1281 #undef MC_TEST_LUMA
1282 #undef MC_TEST_CHROMA
1283
1284 #define MC_TEST_AVG( name, weight ) \
1285 { \
1286 for( int i = 0; i < 12; i++ ) \
1287 { \
1288 memcpy( pbuf3, pbuf1+320, 320 * sizeof(pixel) ); \
1289 memcpy( pbuf4, pbuf1+320, 320 * sizeof(pixel) ); \
1290 if( mc_a.name[i] != mc_ref.name[i] ) \
1291 { \
1292 set_func_name( "%s_%s", #name, pixel_names[i] ); \
1293 used_asm = 1; \
1294 call_c1( mc_c.name[i], pbuf3, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1295 call_a1( mc_a.name[i], pbuf4, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1296 if( memcmp( pbuf3, pbuf4, 320 * sizeof(pixel) ) ) \
1297 { \
1298 ok = 0; \
1299 fprintf( stderr, #name "[%d]: [FAILED]\n", i ); \
1300 } \
1301 call_c2( mc_c.name[i], pbuf3, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1302 call_a2( mc_a.name[i], pbuf4, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1303 } \
1304 } \
1305 }
1306
1307 ok = 1, used_asm = 0;
1308 for( int w = -63; w <= 127 && ok; w++ )
1309 MC_TEST_AVG( avg, w );
1310 report( "mc wpredb :" );
1311
1312 #define MC_TEST_WEIGHT( name, weight, aligned ) \
1313 int align_off = (aligned ? 0 : rand()%16); \
1314 for( int i = 1; i <= 5; i++ ) \
1315 { \
1316 ALIGNED_16( pixel buffC[640] ); \
1317 ALIGNED_16( pixel buffA[640] ); \
1318 int j = X264_MAX( i*4, 2 ); \
1319 memset( buffC, 0, 640 * sizeof(pixel) ); \
1320 memset( buffA, 0, 640 * sizeof(pixel) ); \
1321 x264_t ha; \
1322 ha.mc = mc_a; \
1323 /* w12 is the same as w16 in some cases */ \
1324 if( i == 3 && mc_a.name[i] == mc_a.name[i+1] ) \
1325 continue; \
1326 if( mc_a.name[i] != mc_ref.name[i] ) \
1327 { \
1328 set_func_name( "%s_w%d", #name, j ); \
1329 used_asm = 1; \
1330 call_c1( mc_c.weight[i], buffC, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1331 mc_a.weight_cache(&ha, &weight); \
1332 call_a1( weight.weightfn[i], buffA, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1333 for( int k = 0; k < 16; k++ ) \
1334 if( memcmp( &buffC[k*32], &buffA[k*32], j * sizeof(pixel) ) ) \
1335 { \
1336 ok = 0; \
1337 fprintf( stderr, #name "[%d]: [FAILED] s:%d o:%d d%d\n", i, s, o, d ); \
1338 break; \
1339 } \
1340 /* omit unlikely high scales for benchmarking */ \
1341 if( (s << (8-d)) < 512 ) \
1342 { \
1343 call_c2( mc_c.weight[i], buffC, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1344 call_a2( weight.weightfn[i], buffA, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1345 } \
1346 } \
1347 }
1348
1349 ok = 1; used_asm = 0;
1350
1351 int align_cnt = 0;
1352 for( int s = 0; s <= 127 && ok; s++ )
1353 {
1354 for( int o = -128; o <= 127 && ok; o++ )
1355 {
1356 if( rand() & 2047 ) continue;
1357 for( int d = 0; d <= 7 && ok; d++ )
1358 {
1359 if( s == 1<<d )
1360 continue;
1361 x264_weight_t weight = { .i_scale = s, .i_denom = d, .i_offset = o };
1362 MC_TEST_WEIGHT( weight, weight, (align_cnt++ % 4) );
1363 }
1364 }
1365
1366 }
1367 report( "mc weight :" );
1368
1369 ok = 1; used_asm = 0;
1370 for( int o = 0; o <= 127 && ok; o++ )
1371 {
1372 int s = 1, d = 0;
1373 if( rand() & 15 ) continue;
1374 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1375 MC_TEST_WEIGHT( offsetadd, weight, (align_cnt++ % 4) );
1376 }
1377 report( "mc offsetadd :" );
1378 ok = 1; used_asm = 0;
1379 for( int o = -128; o < 0 && ok; o++ )
1380 {
1381 int s = 1, d = 0;
1382 if( rand() & 15 ) continue;
1383 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1384 MC_TEST_WEIGHT( offsetsub, weight, (align_cnt++ % 4) );
1385 }
1386 report( "mc offsetsub :" );
1387
1388 memset( pbuf3, 0, 64*16 );
1389 memset( pbuf4, 0, 64*16 );
1390 ok = 1; used_asm = 0;
1391 for( int height = 8; height <= 16; height += 8 )
1392 {
1393 if( mc_a.store_interleave_chroma != mc_ref.store_interleave_chroma )
1394 {
1395 set_func_name( "store_interleave_chroma" );
1396 used_asm = 1;
1397 call_c( mc_c.store_interleave_chroma, pbuf3, (intptr_t)64, pbuf1, pbuf1+16, height );
1398 call_a( mc_a.store_interleave_chroma, pbuf4, (intptr_t)64, pbuf1, pbuf1+16, height );
1399 if( memcmp( pbuf3, pbuf4, 64*height ) )
1400 {
1401 ok = 0;
1402 fprintf( stderr, "store_interleave_chroma FAILED: h=%d\n", height );
1403 break;
1404 }
1405 }
1406 if( mc_a.load_deinterleave_chroma_fenc != mc_ref.load_deinterleave_chroma_fenc )
1407 {
1408 set_func_name( "load_deinterleave_chroma_fenc" );
1409 used_asm = 1;
1410 call_c( mc_c.load_deinterleave_chroma_fenc, pbuf3, pbuf1, (intptr_t)64, height );
1411 call_a( mc_a.load_deinterleave_chroma_fenc, pbuf4, pbuf1, (intptr_t)64, height );
1412 if( memcmp( pbuf3, pbuf4, FENC_STRIDE*height ) )
1413 {
1414 ok = 0;
1415 fprintf( stderr, "load_deinterleave_chroma_fenc FAILED: h=%d\n", height );
1416 break;
1417 }
1418 }
1419 if( mc_a.load_deinterleave_chroma_fdec != mc_ref.load_deinterleave_chroma_fdec )
1420 {
1421 set_func_name( "load_deinterleave_chroma_fdec" );
1422 used_asm = 1;
1423 call_c( mc_c.load_deinterleave_chroma_fdec, pbuf3, pbuf1, (intptr_t)64, height );
1424 call_a( mc_a.load_deinterleave_chroma_fdec, pbuf4, pbuf1, (intptr_t)64, height );
1425 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*height ) )
1426 {
1427 ok = 0;
1428 fprintf( stderr, "load_deinterleave_chroma_fdec FAILED: h=%d\n", height );
1429 break;
1430 }
1431 }
1432 }
1433 report( "store_interleave :" );
1434
1435 struct plane_spec {
1436 int w, h, src_stride;
1437 } plane_specs[] = { {2,2,2}, {8,6,8}, {20,31,24}, {32,8,40}, {256,10,272}, {504,7,505}, {528,6,528}, {256,10,-256}, {263,9,-264}, {1904,1,0} };
1438 ok = 1; used_asm = 0;
1439 if( mc_a.plane_copy != mc_ref.plane_copy )
1440 {
1441 set_func_name( "plane_copy" );
1442 used_asm = 1;
1443 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1444 {
1445 int w = plane_specs[i].w;
1446 int h = plane_specs[i].h;
1447 intptr_t src_stride = plane_specs[i].src_stride;
1448 intptr_t dst_stride = (w + 127) & ~63;
1449 assert( dst_stride * h <= 0x1000 );
1450 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1451 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1452 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1453 call_c( mc_c.plane_copy, pbuf3, dst_stride, src1, src_stride, w, h );
1454 call_a( mc_a.plane_copy, pbuf4, dst_stride, src1, src_stride, w, h );
1455 for( int y = 0; y < h; y++ )
1456 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(pixel) ) )
1457 {
1458 ok = 0;
1459 fprintf( stderr, "plane_copy FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1460 break;
1461 }
1462 }
1463 }
1464
1465 if( mc_a.plane_copy_swap != mc_ref.plane_copy_swap )
1466 {
1467 set_func_name( "plane_copy_swap" );
1468 used_asm = 1;
1469 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1470 {
1471 int w = (plane_specs[i].w + 1) >> 1;
1472 int h = plane_specs[i].h;
1473 intptr_t src_stride = plane_specs[i].src_stride;
1474 intptr_t dst_stride = (2*w + 127) & ~63;
1475 assert( dst_stride * h <= 0x1000 );
1476 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1477 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1478 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1479 call_c( mc_c.plane_copy_swap, pbuf3, dst_stride, src1, src_stride, w, h );
1480 call_a( mc_a.plane_copy_swap, pbuf4, dst_stride, src1, src_stride, w, h );
1481 for( int y = 0; y < h; y++ )
1482 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
1483 {
1484 ok = 0;
1485 fprintf( stderr, "plane_copy_swap FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1486 break;
1487 }
1488 }
1489 }
1490
1491 if( mc_a.plane_copy_interleave != mc_ref.plane_copy_interleave )
1492 {
1493 set_func_name( "plane_copy_interleave" );
1494 used_asm = 1;
1495 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1496 {
1497 int w = (plane_specs[i].w + 1) >> 1;
1498 int h = plane_specs[i].h;
1499 intptr_t src_stride = (plane_specs[i].src_stride + 1) >> 1;
1500 intptr_t dst_stride = (2*w + 127) & ~63;
1501 assert( dst_stride * h <= 0x1000 );
1502 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1503 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1504 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1505 call_c( mc_c.plane_copy_interleave, pbuf3, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1506 call_a( mc_a.plane_copy_interleave, pbuf4, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1507 for( int y = 0; y < h; y++ )
1508 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
1509 {
1510 ok = 0;
1511 fprintf( stderr, "plane_copy_interleave FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1512 break;
1513 }
1514 }
1515 }
1516
1517 if( mc_a.plane_copy_deinterleave != mc_ref.plane_copy_deinterleave )
1518 {
1519 set_func_name( "plane_copy_deinterleave" );
1520 used_asm = 1;
1521 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1522 {
1523 int w = (plane_specs[i].w + 1) >> 1;
1524 int h = plane_specs[i].h;
1525 intptr_t dst_stride = w;
1526 intptr_t src_stride = (2*w + 127) & ~63;
1527 intptr_t offv = (dst_stride*h + 63) & ~31;
1528 memset( pbuf3, 0, 0x1000 );
1529 memset( pbuf4, 0, 0x1000 );
1530 call_c( mc_c.plane_copy_deinterleave, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf1, src_stride, w, h );
1531 call_a( mc_a.plane_copy_deinterleave, pbuf4, dst_stride, pbuf4+offv, dst_stride, pbuf1, src_stride, w, h );
1532 for( int y = 0; y < h; y++ )
1533 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w ) ||
1534 memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w ) )
1535 {
1536 ok = 0;
1537 fprintf( stderr, "plane_copy_deinterleave FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1538 break;
1539 }
1540 }
1541 }
1542
1543 if( mc_a.plane_copy_deinterleave_yuyv != mc_ref.plane_copy_deinterleave_yuyv )
1544 {
1545 set_func_name( "plane_copy_deinterleave_yuyv" );
1546 used_asm = 1;
1547 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1548 {
1549 int w = (plane_specs[i].w + 1) >> 1;
1550 int h = plane_specs[i].h;
1551 intptr_t dst_stride = ALIGN( w, 32/sizeof(pixel) );
1552 intptr_t src_stride = (plane_specs[i].src_stride + 1) >> 1;
1553 intptr_t offv = dst_stride*h;
1554 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1555 memset( pbuf3, 0, 0x1000 );
1556 memset( pbuf4, 0, 0x1000 );
1557 /* Skip benchmarking since it's the same as plane_copy_deinterleave(), just verify correctness. */
1558 call_c1( mc_c.plane_copy_deinterleave_yuyv, pbuf3, dst_stride, pbuf3+offv, dst_stride, src1, src_stride, w, h );
1559 call_a1( mc_a.plane_copy_deinterleave_yuyv, pbuf4, dst_stride, pbuf4+offv, dst_stride, src1, src_stride, w, h );
1560 for( int y = 0; y < h; y++ )
1561 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(pixel) ) ||
1562 memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w*sizeof(pixel) ) )
1563 {
1564 fprintf( stderr, "plane_copy_deinterleave_yuyv FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1565 break;
1566 }
1567 }
1568 }
1569
1570 if( mc_a.plane_copy_deinterleave_rgb != mc_ref.plane_copy_deinterleave_rgb )
1571 {
1572 set_func_name( "plane_copy_deinterleave_rgb" );
1573 used_asm = 1;
1574 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1575 {
1576 int w = (plane_specs[i].w + 2) >> 2;
1577 int h = plane_specs[i].h;
1578 intptr_t src_stride = plane_specs[i].src_stride;
1579 intptr_t dst_stride = ALIGN( w, 16 );
1580 intptr_t offv = dst_stride*h + 16;
1581
1582 for( int pw = 3; pw <= 4; pw++ )
1583 {
1584 memset( pbuf3, 0, 0x1000 );
1585 memset( pbuf4, 0, 0x1000 );
1586 call_c( mc_c.plane_copy_deinterleave_rgb, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf3+2*offv, dst_stride, pbuf1, src_stride, pw, w, h );
1587 call_a( mc_a.plane_copy_deinterleave_rgb, pbuf4, dst_stride, pbuf4+offv, dst_stride, pbuf4+2*offv, dst_stride, pbuf1, src_stride, pw, w, h );
1588 for( int y = 0; y < h; y++ )
1589 if( memcmp( pbuf3+y*dst_stride+0*offv, pbuf4+y*dst_stride+0*offv, w ) ||
1590 memcmp( pbuf3+y*dst_stride+1*offv, pbuf4+y*dst_stride+1*offv, w ) ||
1591 memcmp( pbuf3+y*dst_stride+2*offv, pbuf4+y*dst_stride+2*offv, w ) )
1592 {
1593 ok = 0;
1594 fprintf( stderr, "plane_copy_deinterleave_rgb FAILED: w=%d h=%d stride=%d pw=%d\n", w, h, (int)src_stride, pw );
1595 break;
1596 }
1597 }
1598 }
1599 }
1600 report( "plane_copy :" );
1601
1602 if( mc_a.plane_copy_deinterleave_v210 != mc_ref.plane_copy_deinterleave_v210 )
1603 {
1604 set_func_name( "plane_copy_deinterleave_v210" );
1605 ok = 1; used_asm = 1;
1606 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1607 {
1608 int w = (plane_specs[i].w + 1) >> 1;
1609 int h = plane_specs[i].h;
1610 intptr_t dst_stride = ALIGN( w, 32 );
1611 intptr_t src_stride = (w + 47) / 48 * 128 / sizeof(uint32_t);
1612 intptr_t offv = dst_stride*h + 32;
1613 memset( pbuf3, 0, 0x1000 );
1614 memset( pbuf4, 0, 0x1000 );
1615 call_c( mc_c.plane_copy_deinterleave_v210, pbuf3, dst_stride, pbuf3+offv, dst_stride, (uint32_t *)buf1, src_stride, w, h );
1616 call_a( mc_a.plane_copy_deinterleave_v210, pbuf4, dst_stride, pbuf4+offv, dst_stride, (uint32_t *)buf1, src_stride, w, h );
1617 for( int y = 0; y < h; y++ )
1618 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(uint16_t) ) ||
1619 memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w*sizeof(uint16_t) ) )
1620 {
1621 ok = 0;
1622 fprintf( stderr, "plane_copy_deinterleave_v210 FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1623 break;
1624 }
1625 }
1626 report( "v210 :" );
1627 }
1628
1629 if( mc_a.hpel_filter != mc_ref.hpel_filter )
1630 {
1631 pixel *srchpel = pbuf1+8+2*64;
1632 pixel *dstc[3] = { pbuf3+8, pbuf3+8+16*64, pbuf3+8+32*64 };
1633 pixel *dsta[3] = { pbuf4+8, pbuf4+8+16*64, pbuf4+8+32*64 };
1634 void *tmp = pbuf3+49*64;
1635 set_func_name( "hpel_filter" );
1636 ok = 1; used_asm = 1;
1637 memset( pbuf3, 0, 4096 * sizeof(pixel) );
1638 memset( pbuf4, 0, 4096 * sizeof(pixel) );
1639 call_c( mc_c.hpel_filter, dstc[0], dstc[1], dstc[2], srchpel, (intptr_t)64, 48, 10, tmp );
1640 call_a( mc_a.hpel_filter, dsta[0], dsta[1], dsta[2], srchpel, (intptr_t)64, 48, 10, tmp );
1641 for( int i = 0; i < 3; i++ )
1642 for( int j = 0; j < 10; j++ )
1643 //FIXME ideally the first pixels would match too, but they aren't actually used
1644 if( memcmp( dstc[i]+j*64+2, dsta[i]+j*64+2, 43 * sizeof(pixel) ) )
1645 {
1646 ok = 0;
1647 fprintf( stderr, "hpel filter differs at plane %c line %d\n", "hvc"[i], j );
1648 for( int k = 0; k < 48; k++ )
1649 printf( "%02x%s", dstc[i][j*64+k], (k+1)&3 ? "" : " " );
1650 printf( "\n" );
1651 for( int k = 0; k < 48; k++ )
1652 printf( "%02x%s", dsta[i][j*64+k], (k+1)&3 ? "" : " " );
1653 printf( "\n" );
1654 break;
1655 }
1656 report( "hpel filter :" );
1657 }
1658
1659 if( mc_a.frame_init_lowres_core != mc_ref.frame_init_lowres_core )
1660 {
1661 pixel *dstc[4] = { pbuf3, pbuf3+1024, pbuf3+2048, pbuf3+3072 };
1662 pixel *dsta[4] = { pbuf4, pbuf4+1024, pbuf4+2048, pbuf4+3072 };
1663 set_func_name( "lowres_init" );
1664 ok = 1; used_asm = 1;
1665 for( int w = 96; w <= 96+24; w += 8 )
1666 {
1667 intptr_t stride = (w*2+31)&~31;
1668 intptr_t stride_lowres = (w+31)&~31;
1669 call_c( mc_c.frame_init_lowres_core, pbuf1, dstc[0], dstc[1], dstc[2], dstc[3], stride, stride_lowres, w, 8 );
1670 call_a( mc_a.frame_init_lowres_core, pbuf1, dsta[0], dsta[1], dsta[2], dsta[3], stride, stride_lowres, w, 8 );
1671 for( int i = 0; i < 8; i++ )
1672 {
1673 for( int j = 0; j < 4; j++ )
1674 if( memcmp( dstc[j]+i*stride_lowres, dsta[j]+i*stride_lowres, w * sizeof(pixel) ) )
1675 {
1676 ok = 0;
1677 fprintf( stderr, "frame_init_lowres differs at plane %d line %d\n", j, i );
1678 for( int k = 0; k < w; k++ )
1679 printf( "%d ", dstc[j][k+i*stride_lowres] );
1680 printf( "\n" );
1681 for( int k = 0; k < w; k++ )
1682 printf( "%d ", dsta[j][k+i*stride_lowres] );
1683 printf( "\n" );
1684 break;
1685 }
1686 }
1687 }
1688 report( "lowres init :" );
1689 }
1690
1691 #define INTEGRAL_INIT( name, size, offset, cmp_len, ... )\
1692 if( mc_a.name != mc_ref.name )\
1693 {\
1694 intptr_t stride = 96;\
1695 set_func_name( #name );\
1696 used_asm = 1;\
1697 memcpy( buf3, buf1, size*2*stride );\
1698 memcpy( buf4, buf1, size*2*stride );\
1699 uint16_t *sum = (uint16_t*)buf3;\
1700 call_c1( mc_c.name, sum+offset, __VA_ARGS__ );\
1701 sum = (uint16_t*)buf4;\
1702 call_a1( mc_a.name, sum+offset, __VA_ARGS__ );\
1703 if( memcmp( buf3+2*offset, buf4+2*offset, cmp_len*2 )\
1704 || (size>9 && memcmp( buf3+18*stride, buf4+18*stride, (stride-8)*2 )))\
1705 ok = 0;\
1706 call_c2( mc_c.name, sum+offset, __VA_ARGS__ );\
1707 call_a2( mc_a.name, sum+offset, __VA_ARGS__ );\
1708 }
1709 ok = 1; used_asm = 0;
1710 INTEGRAL_INIT( integral_init4h, 2, stride, stride-4, pbuf2, stride );
1711 INTEGRAL_INIT( integral_init8h, 2, stride, stride-8, pbuf2, stride );
1712 INTEGRAL_INIT( integral_init4v, 14, 0, stride-8, sum+9*stride, stride );
1713 INTEGRAL_INIT( integral_init8v, 9, 0, stride-8, stride );
1714 report( "integral init :" );
1715
1716 ok = 1; used_asm = 0;
1717 if( mc_a.mbtree_propagate_cost != mc_ref.mbtree_propagate_cost )
1718 {
1719 used_asm = 1;
1720 x264_emms();
1721 for( int i = 0; i < 10; i++ )
1722 {
1723 float fps_factor = (rand()&65535) / 65535.0f;
1724 set_func_name( "mbtree_propagate_cost" );
1725 int16_t *dsta = (int16_t*)buf3;
1726 int16_t *dstc = dsta+400;
1727 uint16_t *prop = (uint16_t*)buf1;
1728 uint16_t *intra = (uint16_t*)buf4;
1729 uint16_t *inter = intra+128;
1730 uint16_t *qscale = inter+128;
1731 uint16_t *rnd = (uint16_t*)buf2;
1732 x264_emms();
1733 for( int j = 0; j < 100; j++ )
1734 {
1735 intra[j] = *rnd++ & 0x7fff;
1736 intra[j] += !intra[j];
1737 inter[j] = *rnd++ & 0x7fff;
1738 qscale[j] = *rnd++ & 0x7fff;
1739 }
1740 call_c( mc_c.mbtree_propagate_cost, dstc, prop, intra, inter, qscale, &fps_factor, 100 );
1741 call_a( mc_a.mbtree_propagate_cost, dsta, prop, intra, inter, qscale, &fps_factor, 100 );
1742 // I don't care about exact rounding, this is just how close the floating-point implementation happens to be
1743 x264_emms();
1744 for( int j = 0; j < 100 && ok; j++ )
1745 {
1746 ok &= abs( dstc[j]-dsta[j] ) <= 1 || fabs( (double)dstc[j]/dsta[j]-1 ) < 1e-4;
1747 if( !ok )
1748 fprintf( stderr, "mbtree_propagate_cost FAILED: %d !~= %d\n", dstc[j], dsta[j] );
1749 }
1750 }
1751 }
1752
1753 if( mc_a.mbtree_propagate_list != mc_ref.mbtree_propagate_list )
1754 {
1755 used_asm = 1;
1756 for( int i = 0; i < 8; i++ )
1757 {
1758 set_func_name( "mbtree_propagate_list" );
1759 x264_t h;
1760 int height = 4;
1761 int width = 128;
1762 int size = width*height;
1763 h.mb.i_mb_stride = width;
1764 h.mb.i_mb_width = width;
1765 h.mb.i_mb_height = height;
1766
1767 uint16_t *ref_costsc = (uint16_t*)buf3 + width;
1768 uint16_t *ref_costsa = (uint16_t*)buf4 + width;
1769 int16_t (*mvs)[2] = (int16_t(*)[2])(ref_costsc + width + size);
1770 int16_t *propagate_amount = (int16_t*)(mvs + width);
1771 uint16_t *lowres_costs = (uint16_t*)(propagate_amount + width);
1772 h.scratch_buffer2 = (uint8_t*)(ref_costsa + width + size);
1773 int bipred_weight = (rand()%63)+1;
1774 int mb_y = rand()&3;
1775 int list = i&1;
1776 for( int j = -width; j < size+width; j++ )
1777 ref_costsc[j] = ref_costsa[j] = rand()&32767;
1778 for( int j = 0; j < width; j++ )
1779 {
1780 static const uint8_t list_dist[2][8] = {{0,1,1,1,1,1,1,1},{1,1,3,3,3,3,3,2}};
1781 for( int k = 0; k < 2; k++ )
1782 mvs[j][k] = (rand()&127) - 64;
1783 propagate_amount[j] = rand()&32767;
1784 lowres_costs[j] = list_dist[list][rand()&7] << LOWRES_COST_SHIFT;
1785 }
1786
1787 call_c1( mc_c.mbtree_propagate_list, &h, ref_costsc, mvs, propagate_amount, lowres_costs, bipred_weight, mb_y, width, list );
1788 call_a1( mc_a.mbtree_propagate_list, &h, ref_costsa, mvs, propagate_amount, lowres_costs, bipred_weight, mb_y, width, list );
1789
1790 for( int j = -width; j < size+width && ok; j++ )
1791 {
1792 ok &= abs(ref_costsa[j] - ref_costsc[j]) <= 1;
1793 if( !ok )
1794 fprintf( stderr, "mbtree_propagate_list FAILED at %d: %d !~= %d\n", j, ref_costsc[j], ref_costsa[j] );
1795 }
1796
1797 call_c2( mc_c.mbtree_propagate_list, &h, ref_costsc, mvs, propagate_amount, lowres_costs, bipred_weight, mb_y, width, list );
1798 call_a2( mc_a.mbtree_propagate_list, &h, ref_costsa, mvs, propagate_amount, lowres_costs, bipred_weight, mb_y, width, list );
1799 }
1800 }
1801
1802 static const uint16_t mbtree_fix8_counts[] = { 5, 384, 392, 400, 415 };
1803
1804 if( mc_a.mbtree_fix8_pack != mc_ref.mbtree_fix8_pack )
1805 {
1806 set_func_name( "mbtree_fix8_pack" );
1807 used_asm = 1;
1808 float *fix8_src = (float*)(buf3 + 0x800);
1809 uint16_t *dstc = (uint16_t*)buf3;
1810 uint16_t *dsta = (uint16_t*)buf4;
1811 for( int i = 0; i < ARRAY_ELEMS(mbtree_fix8_counts); i++ )
1812 {
1813 int count = mbtree_fix8_counts[i];
1814
1815 for( int j = 0; j < count; j++ )
1816 fix8_src[j] = (int16_t)(rand()) / 256.0f;
1817 dsta[count] = 0xAAAA;
1818
1819 call_c( mc_c.mbtree_fix8_pack, dstc, fix8_src, count );
1820 call_a( mc_a.mbtree_fix8_pack, dsta, fix8_src, count );
1821
1822 if( memcmp( dsta, dstc, count * sizeof(uint16_t) ) || dsta[count] != 0xAAAA )
1823 {
1824 ok = 0;
1825 fprintf( stderr, "mbtree_fix8_pack FAILED\n" );
1826 break;
1827 }
1828 }
1829 }
1830
1831 if( mc_a.mbtree_fix8_unpack != mc_ref.mbtree_fix8_unpack )
1832 {
1833 set_func_name( "mbtree_fix8_unpack" );
1834 used_asm = 1;
1835 uint16_t *fix8_src = (uint16_t*)(buf3 + 0x800);
1836 float *dstc = (float*)buf3;
1837 float *dsta = (float*)buf4;
1838 for( int i = 0; i < ARRAY_ELEMS(mbtree_fix8_counts); i++ )
1839 {
1840 int count = mbtree_fix8_counts[i];
1841
1842 for( int j = 0; j < count; j++ )
1843 fix8_src[j] = rand();
1844 M32( &dsta[count] ) = 0xAAAAAAAA;
1845
1846 call_c( mc_c.mbtree_fix8_unpack, dstc, fix8_src, count );
1847 call_a( mc_a.mbtree_fix8_unpack, dsta, fix8_src, count );
1848
1849 if( memcmp( dsta, dstc, count * sizeof(float) ) || M32( &dsta[count] ) != 0xAAAAAAAA )
1850 {
1851 ok = 0;
1852 fprintf( stderr, "mbtree_fix8_unpack FAILED\n" );
1853 break;
1854 }
1855 }
1856 }
1857 report( "mbtree :" );
1858
1859 if( mc_a.memcpy_aligned != mc_ref.memcpy_aligned )
1860 {
1861 set_func_name( "memcpy_aligned" );
1862 ok = 1; used_asm = 1;
1863 for( size_t size = 16; size < 512; size += 16 )
1864 {
1865 for( int i = 0; i < size; i++ )
1866 buf1[i] = rand();
1867 memset( buf4-1, 0xAA, size + 2 );
1868 call_c( mc_c.memcpy_aligned, buf3, buf1, size );
1869 call_a( mc_a.memcpy_aligned, buf4, buf1, size );
1870 if( memcmp( buf3, buf4, size ) || buf4[-1] != 0xAA || buf4[size] != 0xAA )
1871 {
1872 ok = 0;
1873 fprintf( stderr, "memcpy_aligned FAILED: size=%d\n", (int)size );
1874 break;
1875 }
1876 }
1877 report( "memcpy aligned :" );
1878 }
1879
1880 if( mc_a.memzero_aligned != mc_ref.memzero_aligned )
1881 {
1882 set_func_name( "memzero_aligned" );
1883 ok = 1; used_asm = 1;
1884 for( size_t size = 128; size < 1024; size += 128 )
1885 {
1886 memset( buf4-1, 0xAA, size + 2 );
1887 call_c( mc_c.memzero_aligned, buf3, size );
1888 call_a( mc_a.memzero_aligned, buf4, size );
1889 if( memcmp( buf3, buf4, size ) || buf4[-1] != 0xAA || buf4[size] != 0xAA )
1890 {
1891 ok = 0;
1892 fprintf( stderr, "memzero_aligned FAILED: size=%d\n", (int)size );
1893 break;
1894 }
1895 }
1896 report( "memzero aligned :" );
1897 }
1898
1899 return ret;
1900 }
1901
1902 static int check_deblock( int cpu_ref, int cpu_new )
1903 {
1904 x264_deblock_function_t db_c;
1905 x264_deblock_function_t db_ref;
1906 x264_deblock_function_t db_a;
1907 int ret = 0, ok = 1, used_asm = 0;
1908 int alphas[36], betas[36];
1909 int8_t tcs[36][4];
1910
1911 x264_deblock_init( 0, &db_c, 0 );
1912 x264_deblock_init( cpu_ref, &db_ref, 0 );
1913 x264_deblock_init( cpu_new, &db_a, 0 );
1914
1915 /* not exactly the real values of a,b,tc but close enough */
1916 for( int i = 35, a = 255, c = 250; i >= 0; i-- )
1917 {
1918 alphas[i] = a << (BIT_DEPTH-8);
1919 betas[i] = (i+1)/2 << (BIT_DEPTH-8);
1920 tcs[i][0] = tcs[i][3] = (c+6)/10 << (BIT_DEPTH-8);
1921 tcs[i][1] = (c+7)/15 << (BIT_DEPTH-8);
1922 tcs[i][2] = (c+9)/20 << (BIT_DEPTH-8);
1923 a = a*9/10;
1924 c = c*9/10;
1925 }
1926
1927 #define TEST_DEBLOCK( name, align, ... ) \
1928 for( int i = 0; i < 36; i++ ) \
1929 { \
1930 intptr_t off = 8*32 + (i&15)*4*!align; /* benchmark various alignments of h filter */ \
1931 for( int j = 0; j < 1024; j++ ) \
1932 /* two distributions of random to excersize different failure modes */ \
1933 pbuf3[j] = rand() & (i&1 ? 0xf : PIXEL_MAX ); \
1934 memcpy( pbuf4, pbuf3, 1024 * sizeof(pixel) ); \
1935 if( db_a.name != db_ref.name ) \
1936 { \
1937 set_func_name( #name ); \
1938 used_asm = 1; \
1939 call_c1( db_c.name, pbuf3+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1940 call_a1( db_a.name, pbuf4+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1941 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1942 { \
1943 ok = 0; \
1944 fprintf( stderr, #name "(a=%d, b=%d): [FAILED]\n", alphas[i], betas[i] ); \
1945 break; \
1946 } \
1947 call_c2( db_c.name, pbuf3+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1948 call_a2( db_a.name, pbuf4+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1949 } \
1950 }
1951
1952 TEST_DEBLOCK( deblock_luma[0], 0, tcs[i] );
1953 TEST_DEBLOCK( deblock_luma[1], 1, tcs[i] );
1954 TEST_DEBLOCK( deblock_h_chroma_420, 0, tcs[i] );
1955 TEST_DEBLOCK( deblock_h_chroma_422, 0, tcs[i] );
1956 TEST_DEBLOCK( deblock_chroma_420_mbaff, 0, tcs[i] );
1957 TEST_DEBLOCK( deblock_chroma_422_mbaff, 0, tcs[i] );
1958 TEST_DEBLOCK( deblock_chroma[1], 1, tcs[i] );
1959 TEST_DEBLOCK( deblock_luma_intra[0], 0 );
1960 TEST_DEBLOCK( deblock_luma_intra[1], 1 );
1961 TEST_DEBLOCK( deblock_h_chroma_420_intra, 0 );
1962 TEST_DEBLOCK( deblock_h_chroma_422_intra, 0 );
1963 TEST_DEBLOCK( deblock_chroma_420_intra_mbaff, 0 );
1964 TEST_DEBLOCK( deblock_chroma_422_intra_mbaff, 0 );
1965 TEST_DEBLOCK( deblock_chroma_intra[1], 1 );
1966
1967 if( db_a.deblock_strength != db_ref.deblock_strength )
1968 {
1969 set_func_name( "deblock_strength" );
1970 used_asm = 1;
1971 for( int i = 0; i < 100; i++ )
1972 {
1973 ALIGNED_ARRAY_16( uint8_t, nnz_buf, [X264_SCAN8_SIZE+8] );
1974 uint8_t *nnz = &nnz_buf[8];
1975 ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );
1976 ALIGNED_ARRAY_16( int16_t, mv, [2],[X264_SCAN8_LUMA_SIZE][2] );
1977 ALIGNED_ARRAY_32( uint8_t, bs, [2],[2][8][4] );
1978 memset( bs, 99, sizeof(uint8_t)*2*4*8*2 );
1979 for( int j = 0; j < X264_SCAN8_SIZE; j++ )
1980 nnz[j] = ((rand()&7) == 7) * rand() & 0xf;
1981 for( int j = 0; j < 2; j++ )
1982 for( int k = 0; k < X264_SCAN8_LUMA_SIZE; k++ )
1983 {
1984 ref[j][k] = ((rand()&3) != 3) ? 0 : (rand() & 31) - 2;
1985 for( int l = 0; l < 2; l++ )
1986 mv[j][k][l] = ((rand()&7) != 7) ? (rand()&7) - 3 : (rand()&16383) - 8192;
1987 }
1988 call_c( db_c.deblock_strength, nnz, ref, mv, bs[0], 2<<(i&1), ((i>>1)&1) );
1989 call_a( db_a.deblock_strength, nnz, ref, mv, bs[1], 2<<(i&1), ((i>>1)&1) );
1990 if( memcmp( bs[0], bs[1], sizeof(uint8_t)*2*4*8 ) )
1991 {
1992 ok = 0;
1993 fprintf( stderr, "deblock_strength: [FAILED]\n" );
1994 for( int j = 0; j < 2; j++ )
1995 {
1996 for( int k = 0; k < 2; k++ )
1997 for( int l = 0; l < 4; l++ )
1998 {
1999 for( int m = 0; m < 4; m++ )
2000 printf("%d ",bs[j][k][l][m]);
2001 printf("\n");
2002 }
2003 printf("\n");
2004 }
2005 break;
2006 }
2007 }
2008 }
2009
2010 report( "deblock :" );
2011
2012 return ret;
2013 }
2014
2015 static int check_quant( int cpu_ref, int cpu_new )
2016 {
2017 x264_quant_function_t qf_c;
2018 x264_quant_function_t qf_ref;
2019 x264_quant_function_t qf_a;
2020 ALIGNED_ARRAY_64( dctcoef, dct1,[64] );
2021 ALIGNED_ARRAY_64( dctcoef, dct2,[64] );
2022 ALIGNED_ARRAY_32( dctcoef, dct3,[8],[16] );
2023 ALIGNED_ARRAY_32( dctcoef, dct4,[8],[16] );
2024 ALIGNED_ARRAY_32( uint8_t, cqm_buf,[64] );
2025 int ret = 0, ok, used_asm;
2026 int oks[3] = {1,1,1}, used_asms[3] = {0,0,0};
2027 x264_t h_buf;
2028 x264_t *h = &h_buf;
2029 memset( h, 0, sizeof(*h) );
2030 h->sps->i_chroma_format_idc = 1;
2031 x264_param_default( &h->param );
2032 h->chroma_qp_table = i_chroma_qp_table + 12;
2033 h->param.analyse.b_transform_8x8 = 1;
2034
2035 for( int i_cqm = 0; i_cqm < 4; i_cqm++ )
2036 {
2037 if( i_cqm == 0 )
2038 {
2039 for( int i = 0; i < 6; i++ )
2040 h->sps->scaling_list[i] = x264_cqm_flat16;
2041 h->param.i_cqm_preset = h->sps->i_cqm_preset = X264_CQM_FLAT;
2042 }
2043 else if( i_cqm == 1 )
2044 {
2045 for( int i = 0; i < 6; i++ )
2046 h->sps->scaling_list[i] = x264_cqm_jvt[i];
2047 h->param.i_cqm_preset = h->sps->i_cqm_preset = X264_CQM_JVT;
2048 }
2049 else
2050 {
2051 int max_scale = BIT_DEPTH < 10 ? 255 : 228;
2052 if( i_cqm == 2 )
2053 for( int i = 0; i < 64; i++ )
2054 cqm_buf[i] = 10 + rand() % (max_scale - 9);
2055 else
2056 for( int i = 0; i < 64; i++ )
2057 cqm_buf[i] = 1;
2058 for( int i = 0; i < 6; i++ )
2059 h->sps->scaling_list[i] = cqm_buf;
2060 h->param.i_cqm_preset = h->sps->i_cqm_preset = X264_CQM_CUSTOM;
2061 }
2062
2063 h->param.rc.i_qp_min = 0;
2064 h->param.rc.i_qp_max = QP_MAX_SPEC;
2065 x264_cqm_init( h );
2066 x264_quant_init( h, 0, &qf_c );
2067 x264_quant_init( h, cpu_ref, &qf_ref );
2068 x264_quant_init( h, cpu_new, &qf_a );
2069
2070 #define INIT_QUANT8(j,max) \
2071 { \
2072 static const int scale1d[8] = {32,31,24,31,32,31,24,31}; \
2073 for( int i = 0; i < max; i++ ) \
2074 { \
2075 unsigned int scale = (255*scale1d[(i>>3)&7]*scale1d[i&7])/16; \
2076 dct1[i] = dct2[i] = (j>>(i>>6))&1 ? (rand()%(2*scale+1))-scale : 0; \
2077 } \
2078 }
2079
2080 #define INIT_QUANT4(j,max) \
2081 { \
2082 static const int scale1d[4] = {4,6,4,6}; \
2083 for( int i = 0; i < max; i++ ) \
2084 { \
2085 unsigned int scale = 255*scale1d[(i>>2)&3]*scale1d[i&3]; \
2086 dct1[i] = dct2[i] = (j>>(i>>4))&1 ? (rand()%(2*scale+1))-scale : 0; \
2087 } \
2088 }
2089
2090 #define TEST_QUANT_DC( name, cqm ) \
2091 if( qf_a.name != qf_ref.name ) \
2092 { \
2093 set_func_name( #name ); \
2094 used_asms[0] = 1; \
2095 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2096 { \
2097 for( int j = 0; j < 2; j++ ) \
2098 { \
2099 int result_c, result_a; \
2100 for( int i = 0; i < 16; i++ ) \
2101 dct1[i] = dct2[i] = j ? (rand() & 0x1fff) - 0xfff : 0; \
2102 result_c = call_c1( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
2103 result_a = call_a1( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
2104 if( memcmp( dct1, dct2, 16*sizeof(dctcoef) ) || result_c != result_a ) \
2105 { \
2106 oks[0] = 0; \
2107 fprintf( stderr, #name "(cqm=%d): [FAILED]\n", i_cqm ); \
2108 break; \
2109 } \
2110 call_c2( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
2111 call_a2( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
2112 } \
2113 } \
2114 }
2115
2116 #define TEST_QUANT( qname, block, type, w, maxj ) \
2117 if( qf_a.qname != qf_ref.qname ) \
2118 { \
2119 set_func_name( #qname ); \
2120 used_asms[0] = 1; \
2121 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2122 { \
2123 for( int j = 0; j < maxj; j++ ) \
2124 { \
2125 INIT_QUANT##type(j, w*w) \
2126 int result_c = call_c1( qf_c.qname, (void*)dct1, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
2127 int result_a = call_a1( qf_a.qname, (void*)dct2, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
2128 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) || result_c != result_a ) \
2129 { \
2130 oks[0] = 0; \
2131 fprintf( stderr, #qname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
2132 break; \
2133 } \
2134 call_c2( qf_c.qname, (void*)dct1, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
2135 call_a2( qf_a.qname, (void*)dct2, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
2136 } \
2137 } \
2138 }
2139
2140 TEST_QUANT( quant_8x8, CQM_8IY, 8, 8, 2 );
2141 TEST_QUANT( quant_8x8, CQM_8PY, 8, 8, 2 );
2142 TEST_QUANT( quant_4x4, CQM_4IY, 4, 4, 2 );
2143 TEST_QUANT( quant_4x4, CQM_4PY, 4, 4, 2 );
2144 TEST_QUANT( quant_4x4x4, CQM_4IY, 4, 8, 16 );
2145 TEST_QUANT( quant_4x4x4, CQM_4PY, 4, 8, 16 );
2146 TEST_QUANT_DC( quant_4x4_dc, **h->quant4_mf[CQM_4IY] );
2147 TEST_QUANT_DC( quant_2x2_dc, **h->quant4_mf[CQM_4IC] );
2148
2149 #define TEST_DEQUANT( qname, dqname, block, w ) \
2150 if( qf_a.dqname != qf_ref.dqname ) \
2151 { \
2152 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
2153 used_asms[1] = 1; \
2154 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2155 { \
2156 INIT_QUANT##w(1, w*w) \
2157 qf_c.qname( dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
2158 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
2159 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2160 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2161 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
2162 { \
2163 oks[1] = 0; \
2164 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
2165 break; \
2166 } \
2167 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2168 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2169 } \
2170 }
2171
2172 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8IY, 8 );
2173 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8PY, 8 );
2174 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4IY, 4 );
2175 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4PY, 4 );
2176
2177 #define TEST_DEQUANT_DC( qname, dqname, block, w ) \
2178 if( qf_a.dqname != qf_ref.dqname ) \
2179 { \
2180 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
2181 used_asms[1] = 1; \
2182 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2183 { \
2184 for( int i = 0; i < 16; i++ ) \
2185 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16; \
2186 qf_c.qname( dct1, h->quant##w##_mf[block][qp][0]>>1, h->quant##w##_bias[block][qp][0]>>1 ); \
2187 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
2188 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2189 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2190 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
2191 { \
2192 oks[1] = 0; \
2193 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
2194 } \
2195 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2196 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2197 } \
2198 }
2199
2200 TEST_DEQUANT_DC( quant_4x4_dc, dequant_4x4_dc, CQM_4IY, 4 );
2201
2202 if( qf_a.idct_dequant_2x4_dc != qf_ref.idct_dequant_2x4_dc )
2203 {
2204 set_func_name( "idct_dequant_2x4_dc_%s", i_cqm?"cqm":"flat" );
2205 used_asms[1] = 1;
2206 for( int qp = h->chroma_qp_table[h->param.rc.i_qp_max]; qp >= h->chroma_qp_table[h->param.rc.i_qp_min]; qp-- )
2207 {
2208 for( int i = 0; i < 8; i++ )
2209 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16;
2210 qf_c.quant_2x2_dc( &dct1[0], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
2211 qf_c.quant_2x2_dc( &dct1[4], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
2212 call_c( qf_c.idct_dequant_2x4_dc, dct1, dct3, h->dequant4_mf[CQM_4IC], qp+3 );
2213 call_a( qf_a.idct_dequant_2x4_dc, dct1, dct4, h->dequant4_mf[CQM_4IC], qp+3 );
2214 for( int i = 0; i < 8; i++ )
2215 if( dct3[i][0] != dct4[i][0] )
2216 {
2217 oks[1] = 0;
2218 fprintf( stderr, "idct_dequant_2x4_dc (qp=%d, cqm=%d): [FAILED]\n", qp, i_cqm );
2219 break;
2220 }
2221 }
2222 }
2223
2224 if( qf_a.idct_dequant_2x4_dconly != qf_ref.idct_dequant_2x4_dconly )
2225 {
2226 set_func_name( "idct_dequant_2x4_dconly_%s", i_cqm?"cqm":"flat" );
2227 used_asms[1] = 1;
2228 for( int qp = h->chroma_qp_table[h->param.rc.i_qp_max]; qp >= h->chroma_qp_table[h->param.rc.i_qp_min]; qp-- )
2229 {
2230 for( int i = 0; i < 8; i++ )
2231 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16;
2232 qf_c.quant_2x2_dc( &dct1[0], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
2233 qf_c.quant_2x2_dc( &dct1[4], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
2234 memcpy( dct2, dct1, 8*sizeof(dctcoef) );
2235 call_c1( qf_c.idct_dequant_2x4_dconly, dct1, h->dequant4_mf[CQM_4IC], qp+3 );
2236 call_a1( qf_a.idct_dequant_2x4_dconly, dct2, h->dequant4_mf[CQM_4IC], qp+3 );
2237 if( memcmp( dct1, dct2, 8*sizeof(dctcoef) ) )
2238 {
2239 oks[1] = 0;
2240 fprintf( stderr, "idct_dequant_2x4_dconly (qp=%d, cqm=%d): [FAILED]\n", qp, i_cqm );
2241 break;
2242 }
2243 call_c2( qf_c.idct_dequant_2x4_dconly, dct1, h->dequant4_mf[CQM_4IC], qp+3 );
2244 call_a2( qf_a.idct_dequant_2x4_dconly, dct2, h->dequant4_mf[CQM_4IC], qp+3 );
2245 }
2246 }
2247
2248 #define TEST_OPTIMIZE_CHROMA_DC( optname, size ) \
2249 if( qf_a.optname != qf_ref.optname ) \
2250 { \
2251 set_func_name( #optname ); \
2252 used_asms[2] = 1; \
2253 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2254 { \
2255 int qpdc = qp + (size == 8 ? 3 : 0); \
2256 int dmf = h->dequant4_mf[CQM_4IC][qpdc%6][0] << qpdc/6; \
2257 if( dmf > 32*64 ) \
2258 continue; \
2259 for( int i = 16;; i <<= 1 ) \
2260 { \
2261 int res_c, res_asm; \
2262 int max = X264_MIN( i, PIXEL_MAX*16 ); \
2263 for( int j = 0; j < size; j++ ) \
2264 dct1[j] = rand()%(max*2+1) - max; \
2265 for( int j = 0; j <= size; j += 4 ) \
2266 qf_c.quant_2x2_dc( &dct1[j], h->quant4_mf[CQM_4IC][qpdc][0]>>1, h->quant4_bias[CQM_4IC][qpdc][0]>>1 ); \
2267 memcpy( dct2, dct1, size*sizeof(dctcoef) ); \
2268 res_c = call_c1( qf_c.optname, dct1, dmf ); \
2269 res_asm = call_a1( qf_a.optname, dct2, dmf ); \
2270 if( res_c != res_asm || memcmp( dct1, dct2, size*sizeof(dctcoef) ) ) \
2271 { \
2272 oks[2] = 0; \
2273 fprintf( stderr, #optname "(qp=%d, res_c=%d, res_asm=%d): [FAILED]\n", qp, res_c, res_asm ); \
2274 } \
2275 call_c2( qf_c.optname, dct1, dmf ); \
2276 call_a2( qf_a.optname, dct2, dmf ); \
2277 if( i >= PIXEL_MAX*16 ) \
2278 break; \
2279 } \
2280 } \
2281 }
2282
2283 TEST_OPTIMIZE_CHROMA_DC( optimize_chroma_2x2_dc, 4 );
2284 TEST_OPTIMIZE_CHROMA_DC( optimize_chroma_2x4_dc, 8 );
2285
2286 x264_cqm_delete( h );
2287 }
2288
2289 ok = oks[0]; used_asm = used_asms[0];
2290 report( "quant :" );
2291
2292 ok = oks[1]; used_asm = used_asms[1];
2293 report( "dequant :" );
2294
2295 ok = oks[2]; used_asm = used_asms[2];
2296 report( "optimize chroma dc :" );
2297
2298 ok = 1; used_asm = 0;
2299 if( qf_a.denoise_dct != qf_ref.denoise_dct )
2300 {
2301 used_asm = 1;
2302 for( int size = 16; size <= 64; size += 48 )
2303 {
2304 set_func_name( "denoise_dct" );
2305 memcpy( dct1, buf1, size*sizeof(dctcoef) );
2306 memcpy( dct2, buf1, size*sizeof(dctcoef) );
2307 memcpy( buf3+256, buf3, 256 );
2308 call_c1( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
2309 call_a1( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
2310 if( memcmp( dct1, dct2, size*sizeof(dctcoef) ) || memcmp( buf3+4, buf3+256+4, (size-1)*sizeof(uint32_t) ) )
2311 ok = 0;
2312 call_c2( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
2313 call_a2( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
2314 }
2315 }
2316 report( "denoise dct :" );
2317
2318 #define TEST_DECIMATE( decname, w, ac, thresh ) \
2319 if( qf_a.decname != qf_ref.decname ) \
2320 { \
2321 set_func_name( #decname ); \
2322 used_asm = 1; \
2323 for( int i = 0; i < 100; i++ ) \
2324 { \
2325 static const int distrib[16] = {1,1,1,1,1,1,1,1,1,1,1,1,2,3,4};\
2326 static const int zerorate_lut[4] = {3,7,15,31};\
2327 int zero_rate = zerorate_lut[i&3];\
2328 for( int idx = 0; idx < w*w; idx++ ) \
2329 { \
2330 int sign = (rand()&1) ? -1 : 1; \
2331 int abs_level = distrib[rand()&15]; \
2332 if( abs_level == 4 ) abs_level = rand()&0x3fff; \
2333 int zero = !(rand()&zero_rate); \
2334 dct1[idx] = zero * abs_level * sign; \
2335 } \
2336 if( ac ) \
2337 dct1[0] = 0; \
2338 int result_c = call_c( qf_c.decname, dct1 ); \
2339 int result_a = call_a( qf_a.decname, dct1 ); \
2340 if( X264_MIN(result_c,thresh) != X264_MIN(result_a,thresh) ) \
2341 { \
2342 ok = 0; \
2343 fprintf( stderr, #decname ": [FAILED]\n" ); \
2344 break; \
2345 } \
2346 } \
2347 }
2348
2349 ok = 1; used_asm = 0;
2350 TEST_DECIMATE( decimate_score64, 8, 0, 6 );
2351 TEST_DECIMATE( decimate_score16, 4, 0, 6 );
2352 TEST_DECIMATE( decimate_score15, 4, 1, 7 );
2353 report( "decimate_score :" );
2354
2355 #define TEST_LAST( last, lastname, size, ac ) \
2356 if( qf_a.last != qf_ref.last ) \
2357 { \
2358 set_func_name( #lastname ); \
2359 used_asm = 1; \
2360 for( int i = 0; i < 100; i++ ) \
2361 { \
2362 int nnz = 0; \
2363 int max = rand() & (size-1); \
2364 memset( dct1, 0, size*sizeof(dctcoef) ); \
2365 for( int idx = ac; idx < max; idx++ ) \
2366 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
2367 if( !nnz ) \
2368 dct1[ac] = 1; \
2369 int result_c = call_c( qf_c.last, dct1+ac ); \
2370 int result_a = call_a( qf_a.last, dct1+ac ); \
2371 if( result_c != result_a ) \
2372 { \
2373 ok = 0; \
2374 fprintf( stderr, #lastname ": [FAILED]\n" ); \
2375 break; \
2376 } \
2377 } \
2378 }
2379
2380 ok = 1; used_asm = 0;
2381 TEST_LAST( coeff_last4 , coeff_last4, 4, 0 );
2382 TEST_LAST( coeff_last8 , coeff_last8, 8, 0 );
2383 TEST_LAST( coeff_last[ DCT_LUMA_AC], coeff_last15, 16, 1 );
2384 TEST_LAST( coeff_last[ DCT_LUMA_4x4], coeff_last16, 16, 0 );
2385 TEST_LAST( coeff_last[ DCT_LUMA_8x8], coeff_last64, 64, 0 );
2386 report( "coeff_last :" );
2387
2388 #define TEST_LEVELRUN( lastname, name, size, ac ) \
2389 if( qf_a.lastname != qf_ref.lastname ) \
2390 { \
2391 set_func_name( #name ); \
2392 used_asm = 1; \
2393 for( int i = 0; i < 100; i++ ) \
2394 { \
2395 x264_run_level_t runlevel_c, runlevel_a; \
2396 int nnz = 0; \
2397 int max = rand() & (size-1); \
2398 memset( dct1, 0, size*sizeof(dctcoef) ); \
2399 memcpy( &runlevel_a, buf1+i, sizeof(x264_run_level_t) ); \
2400 memcpy( &runlevel_c, buf1+i, sizeof(x264_run_level_t) ); \
2401 for( int idx = ac; idx < max; idx++ ) \
2402 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
2403 if( !nnz ) \
2404 dct1[ac] = 1; \
2405 int result_c = call_c( qf_c.lastname, dct1+ac, &runlevel_c ); \
2406 int result_a = call_a( qf_a.lastname, dct1+ac, &runlevel_a ); \
2407 if( result_c != result_a || runlevel_c.last != runlevel_a.last || \
2408 runlevel_c.mask != runlevel_a.mask || \
2409 memcmp(runlevel_c.level, runlevel_a.level, sizeof(dctcoef)*result_c)) \
2410 { \
2411 ok = 0; \
2412 fprintf( stderr, #name ": [FAILED]\n" ); \
2413 break; \
2414 } \
2415 } \
2416 }
2417
2418 ok = 1; used_asm = 0;
2419 TEST_LEVELRUN( coeff_level_run4 , coeff_level_run4, 4, 0 );
2420 TEST_LEVELRUN( coeff_level_run8 , coeff_level_run8, 8, 0 );
2421 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_AC], coeff_level_run15, 16, 1 );
2422 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_4x4], coeff_level_run16, 16, 0 );
2423 report( "coeff_level_run :" );
2424
2425 return ret;
2426 }
2427
2428 static int check_intra( int cpu_ref, int cpu_new )
2429 {
2430 int ret = 0, ok = 1, used_asm = 0;
2431 ALIGNED_ARRAY_32( pixel, edge,[36] );
2432 ALIGNED_ARRAY_32( pixel, edge2,[36] );
2433 ALIGNED_ARRAY_32( pixel, fdec,[FDEC_STRIDE*20] );
2434 struct
2435 {
2436 x264_predict_t predict_16x16[4+3];
2437 x264_predict_t predict_8x8c[4+3];
2438 x264_predict_t predict_8x16c[4+3];
2439 x264_predict8x8_t predict_8x8[9+3];
2440 x264_predict_t predict_4x4[9+3];
2441 x264_predict_8x8_filter_t predict_8x8_filter;
2442 } ip_c, ip_ref, ip_a;
2443
2444 x264_predict_16x16_init( 0, ip_c.predict_16x16 );
2445 x264_predict_8x8c_init( 0, ip_c.predict_8x8c );
2446 x264_predict_8x16c_init( 0, ip_c.predict_8x16c );
2447 x264_predict_8x8_init( 0, ip_c.predict_8x8, &ip_c.predict_8x8_filter );
2448 x264_predict_4x4_init( 0, ip_c.predict_4x4 );
2449
2450 x264_predict_16x16_init( cpu_ref, ip_ref.predict_16x16 );
2451 x264_predict_8x8c_init( cpu_ref, ip_ref.predict_8x8c );
2452 x264_predict_8x16c_init( cpu_ref, ip_ref.predict_8x16c );
2453 x264_predict_8x8_init( cpu_ref, ip_ref.predict_8x8, &ip_ref.predict_8x8_filter );
2454 x264_predict_4x4_init( cpu_ref, ip_ref.predict_4x4 );
2455
2456 x264_predict_16x16_init( cpu_new, ip_a.predict_16x16 );
2457 x264_predict_8x8c_init( cpu_new, ip_a.predict_8x8c );
2458 x264_predict_8x16c_init( cpu_new, ip_a.predict_8x16c );
2459 x264_predict_8x8_init( cpu_new, ip_a.predict_8x8, &ip_a.predict_8x8_filter );
2460 x264_predict_4x4_init( cpu_new, ip_a.predict_4x4 );
2461
2462 memcpy( fdec, pbuf1, 32*20 * sizeof(pixel) );\
2463
2464 ip_c.predict_8x8_filter( fdec+48, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
2465
2466 #define INTRA_TEST( name, dir, w, h, align, bench, ... )\
2467 if( ip_a.name[dir] != ip_ref.name[dir] )\
2468 {\
2469 set_func_name( "intra_%s_%s", #name, intra_##name##_names[dir] );\
2470 used_asm = 1;\
2471 memcpy( pbuf3, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
2472 memcpy( pbuf4, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
2473 for( int a = 0; a < (do_bench ? 64/sizeof(pixel) : 1); a += align )\
2474 {\
2475 call_c##bench( ip_c.name[dir], pbuf3+48+a, ##__VA_ARGS__ );\
2476 call_a##bench( ip_a.name[dir], pbuf4+48+a, ##__VA_ARGS__ );\
2477 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*20 * sizeof(pixel) ) )\
2478 {\
2479 fprintf( stderr, #name "[%d] : [FAILED]\n", dir );\
2480 ok = 0;\
2481 if( ip_c.name == (void *)ip_c.predict_8x8 )\
2482 {\
2483 for( int k = -1; k < 16; k++ )\
2484 printf( "%2x ", edge[16+k] );\
2485 printf( "\n" );\
2486 }\
2487 for( int j = 0; j < h; j++ )\
2488 {\
2489 if( ip_c.name == (void *)ip_c.predict_8x8 )\
2490 printf( "%2x ", edge[14-j] );\
2491 for( int k = 0; k < w; k++ )\
2492 printf( "%2x ", pbuf4[48+k+j*FDEC_STRIDE] );\
2493 printf( "\n" );\
2494 }\
2495 printf( "\n" );\
2496 for( int j = 0; j < h; j++ )\
2497 {\
2498 if( ip_c.name == (void *)ip_c.predict_8x8 )\
2499 printf( " " );\
2500 for( int k = 0; k < w; k++ )\
2501 printf( "%2x ", pbuf3[48+k+j*FDEC_STRIDE] );\
2502 printf( "\n" );\
2503 }\
2504 break;\
2505 }\
2506 }\
2507 }
2508
2509 for( int i = 0; i < 12; i++ )
2510 INTRA_TEST( predict_4x4, i, 4, 4, 4, );
2511 for( int i = 0; i < 7; i++ )
2512 INTRA_TEST( predict_8x8c, i, 8, 8, 16, );
2513 for( int i = 0; i < 7; i++ )
2514 INTRA_TEST( predict_8x16c, i, 8, 16, 16, );
2515 for( int i = 0; i < 7; i++ )
2516 INTRA_TEST( predict_16x16, i, 16, 16, 16, );
2517 for( int i = 0; i < 12; i++ )
2518 INTRA_TEST( predict_8x8, i, 8, 8, 8, , edge );
2519
2520 set_func_name("intra_predict_8x8_filter");
2521 if( ip_a.predict_8x8_filter != ip_ref.predict_8x8_filter )
2522 {
2523 used_asm = 1;
2524 for( int i = 0; i < 32; i++ )
2525 {
2526 if( !(i&7) || ((i&MB_TOPRIGHT) && !(i&MB_TOP)) )
2527 continue;
2528 int neighbor = (i&24)>>1;
2529 memset( edge, 0, 36*sizeof(pixel) );
2530 memset( edge2, 0, 36*sizeof(pixel) );
2531 call_c( ip_c.predict_8x8_filter, pbuf1+48, edge, neighbor, i&7 );
2532 call_a( ip_a.predict_8x8_filter, pbuf1+48, edge2, neighbor, i&7 );
2533 if( !(neighbor&MB_TOPLEFT) )
2534 edge[15] = edge2[15] = 0;
2535 if( memcmp( edge+7, edge2+7, (i&MB_TOPRIGHT ? 26 : i&MB_TOP ? 17 : 8) * sizeof(pixel) ) )
2536 {
2537 fprintf( stderr, "predict_8x8_filter : [FAILED] %d %d\n", (i&24)>>1, i&7);
2538 ok = 0;
2539 }
2540 }
2541 }
2542
2543 #define EXTREMAL_PLANE( w, h ) \
2544 { \
2545 int max[7]; \
2546 for( int j = 0; j < 7; j++ ) \
2547 max[j] = test ? rand()&PIXEL_MAX : PIXEL_MAX; \
2548 fdec[48-1-FDEC_STRIDE] = (i&1)*max[0]; \
2549 for( int j = 0; j < w/2; j++ ) \
2550 fdec[48+j-FDEC_STRIDE] = (!!(i&2))*max[1]; \
2551 for( int j = w/2; j < w-1; j++ ) \
2552 fdec[48+j-FDEC_STRIDE] = (!!(i&4))*max[2]; \
2553 fdec[48+(w-1)-FDEC_STRIDE] = (!!(i&8))*max[3]; \
2554 for( int j = 0; j < h/2; j++ ) \
2555 fdec[48+j*FDEC_STRIDE-1] = (!!(i&16))*max[4]; \
2556 for( int j = h/2; j < h-1; j++ ) \
2557 fdec[48+j*FDEC_STRIDE-1] = (!!(i&32))*max[5]; \
2558 fdec[48+(h-1)*FDEC_STRIDE-1] = (!!(i&64))*max[6]; \
2559 }
2560 /* Extremal test case for planar prediction. */
2561 for( int test = 0; test < 100 && ok; test++ )
2562 for( int i = 0; i < 128 && ok; i++ )
2563 {
2564 EXTREMAL_PLANE( 8, 8 );
2565 INTRA_TEST( predict_8x8c, I_PRED_CHROMA_P, 8, 8, 64, 1 );
2566 EXTREMAL_PLANE( 8, 16 );
2567 INTRA_TEST( predict_8x16c, I_PRED_CHROMA_P, 8, 16, 64, 1 );
2568 EXTREMAL_PLANE( 16, 16 );
2569 INTRA_TEST( predict_16x16, I_PRED_16x16_P, 16, 16, 64, 1 );
2570 }
2571 report( "intra pred :" );
2572 return ret;
2573 }
2574
2575 #define DECL_CABAC(cpu) \
2576 static void run_cabac_decision_##cpu( x264_t *h, uint8_t *dst )\
2577 {\
2578 x264_cabac_t cb;\
2579 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2580 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2581 for( int i = 0; i < 0x1000; i++ )\
2582 x264_cabac_encode_decision_##cpu( &cb, buf1[i]>>1, buf1[i]&1 );\
2583 }\
2584 static void run_cabac_bypass_##cpu( x264_t *h, uint8_t *dst )\
2585 {\
2586 x264_cabac_t cb;\
2587 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2588 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2589 for( int i = 0; i < 0x1000; i++ )\
2590 x264_cabac_encode_bypass_##cpu( &cb, buf1[i]&1 );\
2591 }\
2592 static void run_cabac_terminal_##cpu( x264_t *h, uint8_t *dst )\
2593 {\
2594 x264_cabac_t cb;\
2595 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2596 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2597 for( int i = 0; i < 0x1000; i++ )\
2598 x264_cabac_encode_terminal_##cpu( &cb );\
2599 }
2600 DECL_CABAC(c)
2601 #if HAVE_MMX
2602 DECL_CABAC(asm)
2603 #elif defined(ARCH_AARCH64)
2604 DECL_CABAC(asm)
2605 #else
2606 #define run_cabac_decision_asm run_cabac_decision_c
2607 #define run_cabac_bypass_asm run_cabac_bypass_c
2608 #define run_cabac_terminal_asm run_cabac_terminal_c
2609 #endif
2610
2611 extern const uint8_t x264_count_cat_m1[14];
2612
2613 static int check_cabac( int cpu_ref, int cpu_new )
2614 {
2615 int ret = 0, ok = 1, used_asm = 0;
2616 x264_t h;
2617 h.sps->i_chroma_format_idc = 3;
2618
2619 x264_bitstream_function_t bs_ref;
2620 x264_bitstream_function_t bs_a;
2621 x264_bitstream_init( cpu_ref, &bs_ref );
2622 x264_bitstream_init( cpu_new, &bs_a );
2623 x264_quant_init( &h, cpu_new, &h.quantf );
2624 h.quantf.coeff_last[DCT_CHROMA_DC] = h.quantf.coeff_last4;
2625
2626 /* Reset cabac state to avoid buffer overruns in do_bench() with large BENCH_RUNS values. */
2627 #define GET_CB( i ) (\
2628 x264_cabac_encode_init( &cb[i], bitstream[i], bitstream[i]+0xfff0 ),\
2629 cb[i].f8_bits_encoded = 0, &cb[i] )
2630
2631 #define CABAC_RESIDUAL(name, start, end, rd)\
2632 {\
2633 if( bs_a.name##_internal && (bs_a.name##_internal != bs_ref.name##_internal || (cpu_new&X264_CPU_SSE2_IS_SLOW)) )\
2634 {\
2635 used_asm = 1;\
2636 set_func_name( #name );\
2637 for( int i = 0; i < 2; i++ )\
2638 {\
2639 for( intptr_t ctx_block_cat = start; ctx_block_cat <= end; ctx_block_cat++ )\
2640 {\
2641 for( int j = 0; j < 256; j++ )\
2642 {\
2643 ALIGNED_ARRAY_64( dctcoef, dct, [2],[64] );\
2644 uint8_t bitstream[2][1<<16];\
2645 static const uint8_t ctx_ac[14] = {0,1,0,0,1,0,0,1,0,0,0,1,0,0};\
2646 int ac = ctx_ac[ctx_block_cat];\
2647 int nz = 0;\
2648 while( !nz )\
2649 {\
2650 for( int k = 0; k <= x264_count_cat_m1[ctx_block_cat]; k++ )\
2651 {\
2652 /* Very rough distribution that covers possible inputs */\
2653 int rnd = rand();\
2654 int coef = !(rnd&3);\
2655 coef += !(rnd& 15) * (rand()&0x0006);\
2656 coef += !(rnd& 63) * (rand()&0x0008);\
2657 coef += !(rnd& 255) * (rand()&0x00F0);\
2658 coef += !(rnd&1023) * (rand()&0x7F00);\
2659 nz |= dct[0][ac+k] = dct[1][ac+k] = coef * ((rand()&1) ? 1 : -1);\
2660 }\
2661 }\
2662 h.mb.b_interlaced = i;\
2663 x264_cabac_t cb[2];\
2664 x264_cabac_context_init( &h, &cb[0], SLICE_TYPE_P, 26, 0 );\
2665 x264_cabac_context_init( &h, &cb[1], SLICE_TYPE_P, 26, 0 );\
2666 if( !rd ) memcpy( bitstream[1], bitstream[0], 0x400 );\
2667 call_c1( x264_##name##_c, &h, GET_CB( 0 ), ctx_block_cat, dct[0]+ac );\
2668 call_a1( bs_a.name##_internal, dct[1]+ac, i, ctx_block_cat, GET_CB( 1 ) );\
2669 ok = cb[0].f8_bits_encoded == cb[1].f8_bits_encoded && !memcmp(cb[0].state, cb[1].state, 1024);\
2670 if( !rd ) ok |= !memcmp( bitstream[1], bitstream[0], 0x400 ) && !memcmp( &cb[1], &cb[0], offsetof(x264_cabac_t, p_start) );\
2671 if( !ok )\
2672 {\
2673 fprintf( stderr, #name " : [FAILED] ctx_block_cat %d", (int)ctx_block_cat );\
2674 if( rd && cb[0].f8_bits_encoded != cb[1].f8_bits_encoded )\
2675 fprintf( stderr, " (%d != %d)", cb[0].f8_bits_encoded, cb[1].f8_bits_encoded );\
2676 fprintf( stderr, "\n");\
2677 goto name##fail;\
2678 }\
2679 if( (j&15) == 0 )\
2680 {\
2681 call_c2( x264_##name##_c, &h, GET_CB( 0 ), ctx_block_cat, dct[0]+ac );\
2682 call_a2( bs_a.name##_internal, dct[1]+ac, i, ctx_block_cat, GET_CB( 1 ) );\
2683 }\
2684 }\
2685 }\
2686 }\
2687 }\
2688 }\
2689 name##fail:
2690
2691 CABAC_RESIDUAL( cabac_block_residual, 0, DCT_LUMA_8x8, 0 )
2692 report( "cabac residual:" );
2693
2694 ok = 1; used_asm = 0;
2695 CABAC_RESIDUAL( cabac_block_residual_rd, 0, DCT_LUMA_8x8-1, 1 )
2696 CABAC_RESIDUAL( cabac_block_residual_8x8_rd, DCT_LUMA_8x8, DCT_LUMA_8x8, 1 )
2697 report( "cabac residual rd:" );
2698
2699 if( cpu_ref || run_cabac_decision_c == run_cabac_decision_asm )
2700 return ret;
2701 ok = 1; used_asm = 0;
2702 x264_cabac_init( &h );
2703
2704 set_func_name( "cabac_encode_decision" );
2705 memcpy( buf4, buf3, 0x1000 );
2706 call_c( run_cabac_decision_c, &h, buf3 );
2707 call_a( run_cabac_decision_asm, &h, buf4 );
2708 ok = !memcmp( buf3, buf4, 0x1000 );
2709 report( "cabac decision:" );
2710
2711 set_func_name( "cabac_encode_bypass" );
2712 memcpy( buf4, buf3, 0x1000 );
2713 call_c( run_cabac_bypass_c, &h, buf3 );
2714 call_a( run_cabac_bypass_asm, &h, buf4 );
2715 ok = !memcmp( buf3, buf4, 0x1000 );
2716 report( "cabac bypass:" );
2717
2718 set_func_name( "cabac_encode_terminal" );
2719 memcpy( buf4, buf3, 0x1000 );
2720 call_c( run_cabac_terminal_c, &h, buf3 );
2721 call_a( run_cabac_terminal_asm, &h, buf4 );
2722 ok = !memcmp( buf3, buf4, 0x1000 );
2723 report( "cabac terminal:" );
2724
2725 return ret;
2726 }
2727
2728 static int check_bitstream( int cpu_ref, int cpu_new )
2729 {
2730 x264_bitstream_function_t bs_c;
2731 x264_bitstream_function_t bs_ref;
2732 x264_bitstream_function_t bs_a;
2733
2734 int ret = 0, ok = 1, used_asm = 0;
2735
2736 x264_bitstream_init( 0, &bs_c );
2737 x264_bitstream_init( cpu_ref, &bs_ref );
2738 x264_bitstream_init( cpu_new, &bs_a );
2739 if( bs_a.nal_escape != bs_ref.nal_escape )
2740 {
2741 int size = 0x4000;
2742 uint8_t *input = malloc(size+100);
2743 uint8_t *output1 = malloc(size*2);
2744 uint8_t *output2 = malloc(size*2);
2745 used_asm = 1;
2746 set_func_name( "nal_escape" );
2747 for( int i = 0; i < 100; i++ )
2748 {
2749 /* Test corner-case sizes */
2750 int test_size = i < 10 ? i+1 : rand() & 0x3fff;
2751 /* Test 8 different probability distributions of zeros */
2752 for( int j = 0; j < test_size+32; j++ )
2753 input[j] = (rand()&((1 << ((i&7)+1)) - 1)) * rand();
2754 uint8_t *end_c = (uint8_t*)call_c1( bs_c.nal_escape, output1, input, input+test_size );
2755 uint8_t *end_a = (uint8_t*)call_a1( bs_a.nal_escape, output2, input, input+test_size );
2756 int size_c = end_c-output1;
2757 int size_a = end_a-output2;
2758 if( size_c != size_a || memcmp( output1, output2, size_c ) )
2759 {
2760 fprintf( stderr, "nal_escape : [FAILED] %d %d\n", size_c, size_a );
2761 ok = 0;
2762 break;
2763 }
2764 }
2765 for( int j = 0; j < size+32; j++ )
2766 input[j] = rand();
2767 call_c2( bs_c.nal_escape, output1, input, input+size );
2768 call_a2( bs_a.nal_escape, output2, input, input+size );
2769 free(input);
2770 free(output1);
2771 free(output2);
2772 }
2773 report( "nal escape:" );
2774
2775 return ret;
2776 }
2777
2778 static int check_all_funcs( int cpu_ref, int cpu_new )
2779 {
2780 return check_pixel( cpu_ref, cpu_new )
2781 + check_dct( cpu_ref, cpu_new )
2782 + check_mc( cpu_ref, cpu_new )
2783 + check_intra( cpu_ref, cpu_new )
2784 + check_deblock( cpu_ref, cpu_new )
2785 + check_quant( cpu_ref, cpu_new )
2786 + check_cabac( cpu_ref, cpu_new )
2787 + check_bitstream( cpu_ref, cpu_new );
2788 }
2789
2790 static int add_flags( int *cpu_ref, int *cpu_new, int flags, const char *name )
2791 {
2792 *cpu_ref = *cpu_new;
2793 *cpu_new |= flags;
2794 #if STACK_ALIGNMENT < 16
2795 *cpu_new |= X264_CPU_STACK_MOD4;
2796 #endif
2797 if( *cpu_new & X264_CPU_SSE2_IS_FAST )
2798 *cpu_new &= ~X264_CPU_SSE2_IS_SLOW;
2799 if( !quiet )
2800 fprintf( stderr, "x264: %s\n", name );
2801 return check_all_funcs( *cpu_ref, *cpu_new );
2802 }
2803
2804 static int check_all_flags( void )
2805 {
2806 int ret = 0;
2807 int cpu0 = 0, cpu1 = 0;
2808 uint32_t cpu_detect = x264_cpu_detect();
2809 #if ARCH_X86 || ARCH_X86_64
2810 if( cpu_detect & X264_CPU_AVX512 )
2811 simd_warmup_func = x264_checkasm_warmup_avx512;
2812 else if( cpu_detect & X264_CPU_AVX )
2813 simd_warmup_func = x264_checkasm_warmup_avx;
2814 #endif
2815 simd_warmup();
2816
2817 #if HAVE_MMX
2818 if( cpu_detect & X264_CPU_MMX2 )
2819 {
2820 ret |= add_flags( &cpu0, &cpu1, X264_CPU_MMX | X264_CPU_MMX2, "MMX" );
2821 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "MMX Cache64" );
2822 cpu1 &= ~X264_CPU_CACHELINE_64;
2823 #if ARCH_X86
2824 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_32, "MMX Cache32" );
2825 cpu1 &= ~X264_CPU_CACHELINE_32;
2826 #endif
2827 }
2828 if( cpu_detect & X264_CPU_SSE )
2829 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE, "SSE" );
2830 if( cpu_detect & X264_CPU_SSE2 )
2831 {
2832 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2 | X264_CPU_SSE2_IS_SLOW, "SSE2Slow" );
2833 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2_IS_FAST, "SSE2Fast" );
2834 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSE2Fast Cache64" );
2835 cpu1 &= ~X264_CPU_CACHELINE_64;
2836 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_SHUFFLE, "SSE2 SlowShuffle" );
2837 cpu1 &= ~X264_CPU_SLOW_SHUFFLE;
2838 }
2839 if( cpu_detect & X264_CPU_LZCNT )
2840 {
2841 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "LZCNT" );
2842 cpu1 &= ~X264_CPU_LZCNT;
2843 }
2844 if( cpu_detect & X264_CPU_SSE3 )
2845 {
2846 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE3 | X264_CPU_CACHELINE_64, "SSE3" );
2847 cpu1 &= ~X264_CPU_CACHELINE_64;
2848 }
2849 if( cpu_detect & X264_CPU_SSSE3 )
2850 {
2851 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSSE3, "SSSE3" );
2852 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64" );
2853 cpu1 &= ~X264_CPU_CACHELINE_64;
2854 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_SHUFFLE, "SSSE3 SlowShuffle" );
2855 cpu1 &= ~X264_CPU_SLOW_SHUFFLE;
2856 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSSE3 SlowAtom" );
2857 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64 SlowAtom" );
2858 cpu1 &= ~X264_CPU_CACHELINE_64;
2859 cpu1 &= ~X264_CPU_SLOW_ATOM;
2860 if( cpu_detect & X264_CPU_LZCNT )
2861 {
2862 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSSE3 LZCNT" );
2863 cpu1 &= ~X264_CPU_LZCNT;
2864 }
2865 }
2866 if( cpu_detect & X264_CPU_SSE4 )
2867 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE4, "SSE4" );
2868 if( cpu_detect & X264_CPU_SSE42 )
2869 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE42, "SSE4.2" );
2870 if( cpu_detect & X264_CPU_AVX )
2871 ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX, "AVX" );
2872 if( cpu_detect & X264_CPU_XOP )
2873 ret |= add_flags( &cpu0, &cpu1, X264_CPU_XOP, "XOP" );
2874 if( cpu_detect & X264_CPU_FMA4 )
2875 {
2876 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA4, "FMA4" );
2877 cpu1 &= ~X264_CPU_FMA4;
2878 }
2879 if( cpu_detect & X264_CPU_FMA3 )
2880 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA3, "FMA3" );
2881 if( cpu_detect & X264_CPU_BMI1 )
2882 ret |= add_flags( &cpu0, &cpu1, X264_CPU_BMI1, "BMI1" );
2883 if( cpu_detect & X264_CPU_BMI2 )
2884 ret |= add_flags( &cpu0, &cpu1, X264_CPU_BMI2, "BMI2" );
2885 if( cpu_detect & X264_CPU_AVX2 )
2886 ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX2, "AVX2" );
2887 if( cpu_detect & X264_CPU_AVX512 )
2888 ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX512, "AVX512" );
2889 #elif ARCH_PPC
2890 if( cpu_detect & X264_CPU_ALTIVEC )
2891 {
2892 fprintf( stderr, "x264: ALTIVEC against C\n" );
2893 ret = check_all_funcs( 0, X264_CPU_ALTIVEC );
2894 }
2895 #elif ARCH_ARM
2896 if( cpu_detect & X264_CPU_NEON )
2897 x264_checkasm_call = x264_checkasm_call_neon;
2898 if( cpu_detect & X264_CPU_ARMV6 )
2899 ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV6, "ARMv6" );
2900 if( cpu_detect & X264_CPU_NEON )
2901 ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
2902 if( cpu_detect & X264_CPU_FAST_NEON_MRC )
2903 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FAST_NEON_MRC, "Fast NEON MRC" );
2904 #elif ARCH_AARCH64
2905 if( cpu_detect & X264_CPU_ARMV8 )
2906 ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV8, "ARMv8" );
2907 if( cpu_detect & X264_CPU_NEON )
2908 ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
2909 #elif ARCH_MIPS
2910 if( cpu_detect & X264_CPU_MSA )
2911 ret |= add_flags( &cpu0, &cpu1, X264_CPU_MSA, "MSA" );
2912 #endif
2913 return ret;
2914 }
2915
2916 static int main_internal( int argc, char **argv )
2917 {
2918 #ifdef _WIN32
2919 /* Disable the Windows Error Reporting dialog */
2920 SetErrorMode( SEM_NOGPFAULTERRORBOX );
2921 #endif
2922
2923 if( argc > 1 && !strncmp( argv[1], "--bench", 7 ) )
2924 {
2925 #if !ARCH_X86 && !ARCH_X86_64 && !ARCH_PPC && !ARCH_ARM && !ARCH_AARCH64 && !ARCH_MIPS
2926 fprintf( stderr, "no --bench for your cpu until you port rdtsc\n" );
2927 return 1;
2928 #endif
2929 do_bench = 1;
2930 if( argv[1][7] == '=' )
2931 {
2932 bench_pattern = argv[1]+8;
2933 bench_pattern_len = strlen(bench_pattern);
2934 }
2935 argc--;
2936 argv++;
2937 }
2938
2939 int seed = ( argc > 1 ) ? atoi(argv[1]) : x264_mdate();
2940 fprintf( stderr, "x264: using random seed %u\n", seed );
2941 srand( seed );
2942
2943 buf1 = x264_malloc( 0x1e00 + 0x2000*sizeof(pixel) );
2944 pbuf1 = x264_malloc( 0x1e00*sizeof(pixel) );
2945 if( !buf1 || !pbuf1 )
2946 {
2947 fprintf( stderr, "malloc failed, unable to initiate tests!\n" );
2948 return -1;
2949 }
2950 #define INIT_POINTER_OFFSETS\
2951 buf2 = buf1 + 0xf00;\
2952 buf3 = buf2 + 0xf00;\
2953 buf4 = buf3 + 0x1000*sizeof(pixel);\
2954 pbuf2 = pbuf1 + 0xf00;\
2955 pbuf3 = (pixel*)buf3;\
2956 pbuf4 = (pixel*)buf4;
2957 INIT_POINTER_OFFSETS;
2958 for( int i = 0; i < 0x1e00; i++ )
2959 {
2960 buf1[i] = rand() & 0xFF;
2961 pbuf1[i] = rand() & PIXEL_MAX;
2962 }
2963 memset( buf1+0x1e00, 0, 0x2000*sizeof(pixel) );
2964
2965 if( x264_stack_pagealign( check_all_flags, 0 ) )
2966 {
2967 fprintf( stderr, "x264: at least one test has failed. Go and fix that Right Now!\n" );
2968 return -1;
2969 }
2970 fprintf( stderr, "x264: All tests passed Yeah :)\n" );
2971 if( do_bench )
2972 print_bench();
2973 return 0;
2974 }
2975
2976 int main( int argc, char **argv )
2977 {
2978 return x264_stack_align( main_internal, argc, argv );
2979 }
Clone of VLC x264 git repo