1 /*****************************************************************************
2 * checkasm.c: assembly check tool
3 *****************************************************************************
4 * Copyright (C) 2003-2011 x264 project
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
7 * Laurent Aimar <fenrir@via.ecp.fr>
8 * Fiona Glaser <fiona@x264.com>
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.
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.
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.
24 * This program is also available under a commercial proprietary license.
25 * For more information, contact us at licensing@x264.com.
26 *****************************************************************************/
29 #include "common/common.h"
30 #include "common/cpu.h"
32 // GCC doesn't align stack variables on ARM, so use .bss
35 #define ALIGNED_16( var ) DECLARE_ALIGNED( static var, 16 )
38 /* buf1, buf2: initialised to random data and shouldn't write into them */
40 /* buf3, buf4: used to store output */
42 /* pbuf1, pbuf2: initialised to random pixel data and shouldn't write into them. */
44 /* pbuf3, pbuf4: point to buf3, buf4, just for type convenience */
49 #define report( name ) { \
50 if( used_asm && !quiet ) \
51 fprintf( stderr, " - %-21s [%s]\n", name, ok ? "OK" : "FAILED" ); \
55 #define BENCH_RUNS 100 // tradeoff between accuracy and speed
56 #define BENCH_ALIGNS 16 // number of stack+heap data alignments (another accuracy vs speed tradeoff)
57 #define MAX_FUNCS 1000 // just has to be big enough to hold all the existing functions
58 #define MAX_CPUS 10 // number of different combinations of cpu flags
62 void *pointer; // just for detecting duplicates
71 bench_t vers[MAX_CPUS];
75 int bench_pattern_len = 0;
76 const char *bench_pattern = "";
78 static bench_func_t benchs[MAX_FUNCS];
80 static const char *pixel_names[10] = { "16x16", "16x8", "8x16", "8x8", "8x4", "4x8", "4x4", "4x2", "2x4", "2x2" };
81 static const char *intra_predict_16x16_names[7] = { "v", "h", "dc", "p", "dcl", "dct", "dc8" };
82 static const char *intra_predict_8x8c_names[7] = { "dc", "h", "v", "p", "dcl", "dct", "dc8" };
83 static const char *intra_predict_4x4_names[12] = { "v", "h", "dc", "ddl", "ddr", "vr", "hd", "vl", "hu", "dcl", "dct", "dc8" };
84 static const char **intra_predict_8x8_names = intra_predict_4x4_names;
86 #define set_func_name(...) snprintf( func_name, sizeof(func_name), __VA_ARGS__ )
88 static inline uint32_t read_time(void)
91 #if HAVE_X86_INLINE_ASM
92 asm volatile( "rdtsc" :"=a"(a) ::"edx" );
94 asm volatile( "mftb %0" : "=r" (a) );
95 #elif ARCH_ARM // ARMv7 only
96 asm volatile( "mrc p15, 0, %0, c9, c13, 0" : "=r"(a) );
101 static bench_t* get_bench( const char *name, int cpu )
104 for( i = 0; benchs[i].name && strcmp(name, benchs[i].name); i++ )
105 assert( i < MAX_FUNCS );
106 if( !benchs[i].name )
107 benchs[i].name = strdup( name );
109 return &benchs[i].vers[0];
110 for( j = 1; benchs[i].vers[j].cpu && benchs[i].vers[j].cpu != cpu; j++ )
111 assert( j < MAX_CPUS );
112 benchs[i].vers[j].cpu = cpu;
113 return &benchs[i].vers[j];
116 static int cmp_nop( const void *a, const void *b )
118 return *(uint16_t*)a - *(uint16_t*)b;
121 static int cmp_bench( const void *a, const void *b )
123 // asciibetical sort except preserving numbers
124 const char *sa = ((bench_func_t*)a)->name;
125 const char *sb = ((bench_func_t*)b)->name;
130 if( isdigit( *sa ) && isdigit( *sb ) && isdigit( sa[1] ) != isdigit( sb[1] ) )
131 return isdigit( sa[1] ) - isdigit( sb[1] );
137 static void print_bench(void)
139 uint16_t nops[10000] = {0};
140 int nfuncs, nop_time=0;
142 for( int i = 0; i < 10000; i++ )
145 nops[i] = read_time() - t;
147 qsort( nops, 10000, sizeof(uint16_t), cmp_nop );
148 for( int i = 500; i < 9500; i++ )
151 printf( "nop: %d\n", nop_time );
153 for( nfuncs = 0; nfuncs < MAX_FUNCS && benchs[nfuncs].name; nfuncs++ );
154 qsort( benchs, nfuncs, sizeof(bench_func_t), cmp_bench );
155 for( int i = 0; i < nfuncs; i++ )
156 for( int j = 0; j < MAX_CPUS && (!j || benchs[i].vers[j].cpu); j++ )
159 bench_t *b = &benchs[i].vers[j];
162 for( k = 0; k < j && benchs[i].vers[k].pointer != b->pointer; k++ );
165 printf( "%s_%s%s: %"PRId64"\n", benchs[i].name,
166 b->cpu&X264_CPU_AVX ? "avx" :
167 b->cpu&X264_CPU_SSE4 ? "sse4" :
168 b->cpu&X264_CPU_SHUFFLE_IS_FAST ? "fastshuffle" :
169 b->cpu&X264_CPU_SSSE3 ? "ssse3" :
170 b->cpu&X264_CPU_SSE3 ? "sse3" :
171 /* print sse2slow only if there's also a sse2fast version of the same func */
172 b->cpu&X264_CPU_SSE2_IS_SLOW && j<MAX_CPUS && b[1].cpu&X264_CPU_SSE2_IS_FAST && !(b[1].cpu&X264_CPU_SSE3) ? "sse2slow" :
173 b->cpu&X264_CPU_SSE2 ? "sse2" :
174 b->cpu&X264_CPU_MMX ? "mmx" :
175 b->cpu&X264_CPU_ALTIVEC ? "altivec" :
176 b->cpu&X264_CPU_NEON ? "neon" :
177 b->cpu&X264_CPU_ARMV6 ? "armv6" : "c",
178 b->cpu&X264_CPU_CACHELINE_32 ? "_c32" :
179 b->cpu&X264_CPU_CACHELINE_64 ? "_c64" :
180 b->cpu&X264_CPU_SSE_MISALIGN ? "_misalign" :
181 b->cpu&X264_CPU_LZCNT ? "_lzcnt" :
182 b->cpu&X264_CPU_FAST_NEON_MRC ? "_fast_mrc" :
183 b->cpu&X264_CPU_SLOW_CTZ ? "_slow_ctz" :
184 b->cpu&X264_CPU_SLOW_ATOM ? "_slow_atom" : "",
185 ((int64_t)10*b->cycles/b->den - nop_time)/4 );
189 #if ARCH_X86 || ARCH_X86_64
190 int x264_stack_pagealign( int (*func)(), int align );
192 #define x264_stack_pagealign( func, align ) func()
195 #define call_c1(func,...) func(__VA_ARGS__)
197 #if ARCH_X86 || defined(_WIN64)
198 /* detect when callee-saved regs aren't saved.
199 * needs an explicit asm check because it only sometimes crashes in normal use. */
200 intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
201 #define call_a1(func,...) x264_checkasm_call((intptr_t(*)())func, &ok, __VA_ARGS__)
203 #define call_a1 call_c1
206 #define call_bench(func,cpu,...)\
207 if( do_bench && !strncmp(func_name, bench_pattern, bench_pattern_len) )\
211 call_a1(func, __VA_ARGS__);\
212 for( int ti = 0; ti < (cpu?BENCH_RUNS:BENCH_RUNS/4); ti++ )\
214 uint32_t t = read_time();\
219 t = read_time() - t;\
220 if( t*tcount <= tsum*4 && ti > 0 )\
226 bench_t *b = get_bench( func_name, cpu );\
232 /* for most functions, run benchmark and correctness test at the same time.
233 * for those that modify their inputs, run the above macros separately */
234 #define call_a(func,...) ({ call_a2(func,__VA_ARGS__); call_a1(func,__VA_ARGS__); })
235 #define call_c(func,...) ({ call_c2(func,__VA_ARGS__); call_c1(func,__VA_ARGS__); })
236 #define call_a2(func,...) ({ call_bench(func,cpu_new,__VA_ARGS__); })
237 #define call_c2(func,...) ({ call_bench(func,0,__VA_ARGS__); })
240 static int check_pixel( int cpu_ref, int cpu_new )
242 x264_pixel_function_t pixel_c;
243 x264_pixel_function_t pixel_ref;
244 x264_pixel_function_t pixel_asm;
245 x264_predict8x8_t predict_8x8[9+3];
246 x264_predict_8x8_filter_t predict_8x8_filter;
247 ALIGNED_16( pixel edge[33] );
248 uint16_t cost_mv[32];
249 int ret = 0, ok, used_asm;
251 x264_pixel_init( 0, &pixel_c );
252 x264_pixel_init( cpu_ref, &pixel_ref );
253 x264_pixel_init( cpu_new, &pixel_asm );
254 x264_predict_8x8_init( 0, predict_8x8, &predict_8x8_filter );
255 predict_8x8_filter( pbuf2+40, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
258 for( int i = 0; i < 256; i++ )
263 pbuf4[i] = -(z&1) & PIXEL_MAX;
264 pbuf3[i] = ~pbuf4[i] & PIXEL_MAX;
266 // random pattern made of maxed pixel differences, in case an intermediate value overflows
267 for( int i = 256; i < 0x1000; i++ )
269 pbuf4[i] = -(pbuf1[i&~0x88]&1) & PIXEL_MAX;
270 pbuf3[i] = ~(pbuf4[i]) & PIXEL_MAX;
273 #define TEST_PIXEL( name, align ) \
274 ok = 1, used_asm = 0; \
275 for( int i = 0; i < 7; i++ ) \
277 int res_c, res_asm; \
278 if( pixel_asm.name[i] != pixel_ref.name[i] ) \
280 set_func_name( "%s_%s", #name, pixel_names[i] ); \
282 for( int j = 0; j < 64; j++ ) \
284 res_c = call_c( pixel_c.name[i], pbuf1, 16, pbuf2+j*!align, 64 ); \
285 res_asm = call_a( pixel_asm.name[i], pbuf1, 16, pbuf2+j*!align, 64 ); \
286 if( res_c != res_asm ) \
289 fprintf( stderr, #name "[%d]: %d != %d [FAILED]\n", i, res_c, res_asm ); \
293 for( int j = 0; j < 0x1000 && ok; j += 256 ) \
295 res_c = pixel_c .name[i]( pbuf3+j, 16, pbuf4+j, 16 ); \
296 res_asm = pixel_asm.name[i]( pbuf3+j, 16, pbuf4+j, 16 ); \
297 if( res_c != res_asm ) \
300 fprintf( stderr, #name "[%d]: overflow %d != %d\n", i, res_c, res_asm ); \
305 report( "pixel " #name " :" );
307 TEST_PIXEL( sad, 0 );
308 TEST_PIXEL( sad_aligned, 1 );
309 TEST_PIXEL( ssd, 1 );
310 TEST_PIXEL( satd, 0 );
311 TEST_PIXEL( sa8d, 1 );
313 #define TEST_PIXEL_X( N ) \
314 ok = 1; used_asm = 0; \
315 for( int i = 0; i < 7; i++ ) \
317 int res_c[4]={0}, res_asm[4]={0}; \
318 if( pixel_asm.sad_x##N[i] && pixel_asm.sad_x##N[i] != pixel_ref.sad_x##N[i] ) \
320 set_func_name( "sad_x%d_%s", N, pixel_names[i] ); \
322 for( int j = 0; j < 64; j++ ) \
324 pixel *pix2 = pbuf2+j; \
325 res_c[0] = pixel_c.sad[i]( pbuf1, 16, pix2, 64 ); \
326 res_c[1] = pixel_c.sad[i]( pbuf1, 16, pix2+6, 64 ); \
327 res_c[2] = pixel_c.sad[i]( pbuf1, 16, pix2+1, 64 ); \
330 res_c[3] = pixel_c.sad[i]( pbuf1, 16, pix2+10, 64 ); \
331 call_a( pixel_asm.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, 64, res_asm ); \
334 call_a( pixel_asm.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, 64, res_asm ); \
335 if( memcmp(res_c, res_asm, sizeof(res_c)) ) \
338 fprintf( stderr, "sad_x"#N"[%d]: %d,%d,%d,%d != %d,%d,%d,%d [FAILED]\n", \
339 i, res_c[0], res_c[1], res_c[2], res_c[3], \
340 res_asm[0], res_asm[1], res_asm[2], res_asm[3] ); \
343 call_c2( pixel_c.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, 64, res_asm ); \
345 call_c2( pixel_c.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, 64, res_asm ); \
349 report( "pixel sad_x"#N" :" );
354 #define TEST_PIXEL_VAR( i ) \
355 if( pixel_asm.var[i] != pixel_ref.var[i] ) \
357 set_func_name( "%s_%s", "var", pixel_names[i] ); \
359 /* abi-check wrapper can't return uint64_t, so separate it from return value check */ \
360 call_c1( pixel_c.var[i], pbuf1, 16 ); \
361 call_a1( pixel_asm.var[i], pbuf1, 16 ); \
362 uint64_t res_c = pixel_c.var[i]( pbuf1, 16 ); \
363 uint64_t res_asm = pixel_asm.var[i]( pbuf1, 16 ); \
364 if( res_c != res_asm ) \
367 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) ); \
369 call_c2( pixel_c.var[i], pbuf1, 16 ); \
370 call_a2( pixel_asm.var[i], pbuf1, 16 ); \
373 ok = 1; used_asm = 0;
374 TEST_PIXEL_VAR( PIXEL_16x16 );
375 TEST_PIXEL_VAR( PIXEL_8x8 );
376 report( "pixel var :" );
378 ok = 1; used_asm = 0;
379 if( pixel_asm.var2_8x8 != pixel_ref.var2_8x8 )
381 int res_c, res_asm, ssd_c, ssd_asm;
382 set_func_name( "var2_8x8" );
384 res_c = call_c( pixel_c.var2_8x8, pbuf1, 16, pbuf2, 16, &ssd_c );
385 res_asm = call_a( pixel_asm.var2_8x8, pbuf1, 16, pbuf2, 16, &ssd_asm );
386 if( res_c != res_asm || ssd_c != ssd_asm )
389 fprintf( stderr, "var2_8x8: %d != %d or %d != %d [FAILED]\n", res_c, res_asm, ssd_c, ssd_asm );
393 report( "pixel var2 :" );
395 ok = 1; used_asm = 0;
396 for( int i = 0; i < 4; i++ )
397 if( pixel_asm.hadamard_ac[i] != pixel_ref.hadamard_ac[i] )
399 set_func_name( "hadamard_ac_%s", pixel_names[i] );
401 for( int j = 0; j < 32; j++ )
403 pixel *pix = (j&16 ? pbuf1 : pbuf3) + (j&15)*256;
404 call_c1( pixel_c.hadamard_ac[i], pbuf1, 16 );
405 call_a1( pixel_asm.hadamard_ac[i], pbuf1, 16 );
406 uint64_t rc = pixel_c.hadamard_ac[i]( pix, 16 );
407 uint64_t ra = pixel_asm.hadamard_ac[i]( pix, 16 );
411 fprintf( stderr, "hadamard_ac[%d]: %d,%d != %d,%d\n", i, (int)rc, (int)(rc>>32), (int)ra, (int)(ra>>32) );
415 call_c2( pixel_c.hadamard_ac[i], pbuf1, 16 );
416 call_a2( pixel_asm.hadamard_ac[i], pbuf1, 16 );
418 report( "pixel hadamard_ac :" );
420 ok = 1; used_asm = 0;
421 if( pixel_asm.vsad != pixel_ref.vsad )
423 for( int h = 2; h <= 32; h += 2 )
426 set_func_name( "vsad" );
428 res_c = call_c( pixel_c.vsad, pbuf1, 16, h );
429 res_asm = call_a( pixel_asm.vsad, pbuf1, 16, h );
430 if( res_c != res_asm )
433 fprintf( stderr, "vsad: height=%d, %d != %d\n", h, res_c, res_asm );
438 report( "pixel vsad :" );
440 #define TEST_INTRA_MBCMP( name, pred, satd, i8x8, ... ) \
441 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
443 int res_c[3], res_asm[3]; \
444 set_func_name( #name ); \
446 call_c( pixel_c.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_c ); \
447 call_a( pixel_asm.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_asm ); \
448 if( memcmp(res_c, res_asm, sizeof(res_c)) ) \
451 fprintf( stderr, #name": %d,%d,%d != %d,%d,%d [FAILED]\n", \
452 res_c[0], res_c[1], res_c[2], \
453 res_asm[0], res_asm[1], res_asm[2] ); \
457 ok = 1; used_asm = 0;
458 TEST_INTRA_MBCMP( intra_satd_x3_16x16, predict_16x16, satd[PIXEL_16x16], 0 );
459 TEST_INTRA_MBCMP( intra_satd_x3_8x8c , predict_8x8c , satd[PIXEL_8x8] , 0 );
460 TEST_INTRA_MBCMP( intra_satd_x3_4x4 , predict_4x4 , satd[PIXEL_4x4] , 0 );
461 TEST_INTRA_MBCMP( intra_sa8d_x3_8x8 , predict_8x8 , sa8d[PIXEL_8x8] , 1, edge );
462 report( "intra satd_x3 :" );
463 TEST_INTRA_MBCMP( intra_sad_x3_16x16 , predict_16x16, sad [PIXEL_16x16], 0 );
464 TEST_INTRA_MBCMP( intra_sad_x3_8x8c , predict_8x8c , sad [PIXEL_8x8] , 0 );
465 TEST_INTRA_MBCMP( intra_sad_x3_8x8 , predict_8x8 , sad [PIXEL_8x8] , 1, edge );
466 TEST_INTRA_MBCMP( intra_sad_x3_4x4 , predict_4x4 , sad [PIXEL_4x4] , 0 );
467 report( "intra sad_x3 :" );
469 ok = 1; used_asm = 0;
470 if( pixel_asm.ssd_nv12_core != pixel_ref.ssd_nv12_core )
473 set_func_name( "ssd_nv12" );
474 uint64_t res_u_c, res_v_c, res_u_a, res_v_a;
475 pixel_c.ssd_nv12_core( pbuf1, 368, pbuf2, 368, 360, 8, &res_u_c, &res_v_c );
476 pixel_asm.ssd_nv12_core( pbuf1, 368, pbuf2, 368, 360, 8, &res_u_a, &res_v_a );
477 if( res_u_c != res_u_a || res_v_c != res_v_a )
480 fprintf( stderr, "ssd_nv12: %"PRIu64",%"PRIu64" != %"PRIu64",%"PRIu64"\n",
481 res_u_c, res_v_c, res_u_a, res_v_a );
483 call_c( pixel_c.ssd_nv12_core, pbuf1, 368, pbuf2, 368, 360, 8, &res_u_c, &res_v_c );
484 call_a( pixel_asm.ssd_nv12_core, pbuf1, 368, pbuf2, 368, 360, 8, &res_u_a, &res_v_a );
486 report( "ssd_nv12 :" );
488 if( pixel_asm.ssim_4x4x2_core != pixel_ref.ssim_4x4x2_core ||
489 pixel_asm.ssim_end4 != pixel_ref.ssim_end4 )
493 ALIGNED_16( int sums[5][4] ) = {{0}};
496 res_c = x264_pixel_ssim_wxh( &pixel_c, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
497 res_a = x264_pixel_ssim_wxh( &pixel_asm, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
498 if( fabs( res_c - res_a ) > 1e-6 )
501 fprintf( stderr, "ssim: %.7f != %.7f [FAILED]\n", res_c, res_a );
503 set_func_name( "ssim_core" );
504 call_c2( pixel_c.ssim_4x4x2_core, pbuf1+2, 32, pbuf2+2, 32, sums );
505 call_a2( pixel_asm.ssim_4x4x2_core, pbuf1+2, 32, pbuf2+2, 32, sums );
506 set_func_name( "ssim_end" );
507 call_c2( pixel_c.ssim_end4, sums, sums, 4 );
508 call_a2( pixel_asm.ssim_end4, sums, sums, 4 );
512 ok = 1; used_asm = 0;
513 for( int i = 0; i < 32; i++ )
515 for( int i = 0; i < 100 && ok; i++ )
516 if( pixel_asm.ads[i&3] != pixel_ref.ads[i&3] )
518 ALIGNED_16( uint16_t sums[72] );
519 ALIGNED_16( int dc[4] );
520 int16_t mvs_a[32], mvs_c[32];
522 int thresh = rand() & 0x3fff;
523 set_func_name( "esa_ads" );
524 for( int j = 0; j < 72; j++ )
525 sums[j] = rand() & 0x3fff;
526 for( int j = 0; j < 4; j++ )
527 dc[j] = rand() & 0x3fff;
529 mvn_c = call_c( pixel_c.ads[i&3], dc, sums, 32, cost_mv, mvs_c, 28, thresh );
530 mvn_a = call_a( pixel_asm.ads[i&3], dc, sums, 32, cost_mv, mvs_a, 28, thresh );
531 if( mvn_c != mvn_a || memcmp( mvs_c, mvs_a, mvn_c*sizeof(*mvs_c) ) )
534 printf( "c%d: ", i&3 );
535 for( int j = 0; j < mvn_c; j++ )
536 printf( "%d ", mvs_c[j] );
537 printf( "\na%d: ", i&3 );
538 for( int j = 0; j < mvn_a; j++ )
539 printf( "%d ", mvs_a[j] );
543 report( "esa ads:" );
548 static int check_dct( int cpu_ref, int cpu_new )
550 x264_dct_function_t dct_c;
551 x264_dct_function_t dct_ref;
552 x264_dct_function_t dct_asm;
553 x264_quant_function_t qf;
554 int ret = 0, ok, used_asm, interlace = 0;
555 ALIGNED_16( dctcoef dct1[16][16] );
556 ALIGNED_16( dctcoef dct2[16][16] );
557 ALIGNED_16( dctcoef dct4[16][16] );
558 ALIGNED_16( dctcoef dct8[4][64] );
559 ALIGNED_16( dctcoef dctdc[2][4] );
563 x264_dct_init( 0, &dct_c );
564 x264_dct_init( cpu_ref, &dct_ref);
565 x264_dct_init( cpu_new, &dct_asm );
567 memset( h, 0, sizeof(*h) );
568 x264_param_default( &h->param );
569 h->sps->i_chroma_format_idc = 1;
570 h->chroma_qp_table = i_chroma_qp_table + 12;
571 h->param.analyse.i_luma_deadzone[0] = 0;
572 h->param.analyse.i_luma_deadzone[1] = 0;
573 h->param.analyse.b_transform_8x8 = 1;
574 for( int i = 0; i < 6; i++ )
575 h->pps->scaling_list[i] = x264_cqm_flat16;
577 x264_quant_init( h, 0, &qf );
579 /* overflow test cases */
580 for( int i = 0; i < 5; i++ )
582 pixel *enc = &pbuf3[16*i*FENC_STRIDE];
583 pixel *dec = &pbuf4[16*i*FDEC_STRIDE];
585 for( int j = 0; j < 16; j++ )
587 int cond_a = (i < 2) ? 1 : ((j&3) == 0 || (j&3) == (i-1));
588 int cond_b = (i == 0) ? 1 : !cond_a;
589 enc[0] = enc[1] = cond_a ? PIXEL_MAX : 0;
590 enc[2] = enc[3] = cond_b ? PIXEL_MAX : 0;
592 for( int k = 0; k < 4; k++ )
593 dec[k] = PIXEL_MAX - enc[k];
600 #define TEST_DCT( name, t1, t2, size ) \
601 if( dct_asm.name != dct_ref.name ) \
603 set_func_name( #name ); \
605 pixel *enc = pbuf3; \
606 pixel *dec = pbuf4; \
607 for( int j = 0; j < 5; j++) \
609 call_c( dct_c.name, t1, &pbuf1[j*64], &pbuf2[j*64] ); \
610 call_a( dct_asm.name, t2, &pbuf1[j*64], &pbuf2[j*64] ); \
611 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
614 fprintf( stderr, #name " [FAILED]\n" ); \
617 call_c( dct_c.name, t1, enc, dec ); \
618 call_a( dct_asm.name, t2, enc, dec ); \
619 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
622 fprintf( stderr, #name " [FAILED] (overflow)\n" ); \
625 enc += 16*FENC_STRIDE; \
626 dec += 16*FDEC_STRIDE; \
629 ok = 1; used_asm = 0;
630 TEST_DCT( sub4x4_dct, dct1[0], dct2[0], 16 );
631 TEST_DCT( sub8x8_dct, dct1, dct2, 16*4 );
632 TEST_DCT( sub8x8_dct_dc, dctdc[0], dctdc[1], 4 );
633 TEST_DCT( sub16x16_dct, dct1, dct2, 16*16 );
634 report( "sub_dct4 :" );
636 ok = 1; used_asm = 0;
637 TEST_DCT( sub8x8_dct8, (void*)dct1[0], (void*)dct2[0], 64 );
638 TEST_DCT( sub16x16_dct8, (void*)dct1, (void*)dct2, 64*4 );
639 report( "sub_dct8 :" );
642 // fdct and idct are denormalized by different factors, so quant/dequant
643 // is needed to force the coefs into the right range.
644 dct_c.sub16x16_dct( dct4, pbuf1, pbuf2 );
645 dct_c.sub16x16_dct8( dct8, pbuf1, pbuf2 );
646 for( int i = 0; i < 16; i++ )
648 qf.quant_4x4( dct4[i], h->quant4_mf[CQM_4IY][20], h->quant4_bias[CQM_4IY][20] );
649 qf.dequant_4x4( dct4[i], h->dequant4_mf[CQM_4IY], 20 );
651 for( int i = 0; i < 4; i++ )
653 qf.quant_8x8( dct8[i], h->quant8_mf[CQM_8IY][20], h->quant8_bias[CQM_8IY][20] );
654 qf.dequant_8x8( dct8[i], h->dequant8_mf[CQM_8IY], 20 );
656 x264_cqm_delete( h );
658 #define TEST_IDCT( name, src ) \
659 if( dct_asm.name != dct_ref.name ) \
661 set_func_name( #name ); \
663 memcpy( pbuf3, pbuf1, 32*32 * sizeof(pixel) ); \
664 memcpy( pbuf4, pbuf1, 32*32 * sizeof(pixel) ); \
665 memcpy( dct1, src, 256 * sizeof(dctcoef) ); \
666 memcpy( dct2, src, 256 * sizeof(dctcoef) ); \
667 call_c1( dct_c.name, pbuf3, (void*)dct1 ); \
668 call_a1( dct_asm.name, pbuf4, (void*)dct2 ); \
669 if( memcmp( pbuf3, pbuf4, 32*32 * sizeof(pixel) ) ) \
672 fprintf( stderr, #name " [FAILED]\n" ); \
674 call_c2( dct_c.name, pbuf3, (void*)dct1 ); \
675 call_a2( dct_asm.name, pbuf4, (void*)dct2 ); \
677 ok = 1; used_asm = 0;
678 TEST_IDCT( add4x4_idct, dct4 );
679 TEST_IDCT( add8x8_idct, dct4 );
680 TEST_IDCT( add8x8_idct_dc, dct4 );
681 TEST_IDCT( add16x16_idct, dct4 );
682 TEST_IDCT( add16x16_idct_dc, dct4 );
683 report( "add_idct4 :" );
685 ok = 1; used_asm = 0;
686 TEST_IDCT( add8x8_idct8, dct8 );
687 TEST_IDCT( add16x16_idct8, dct8 );
688 report( "add_idct8 :" );
691 #define TEST_DCTDC( name )\
692 ok = 1; used_asm = 0;\
693 if( dct_asm.name != dct_ref.name )\
695 set_func_name( #name );\
697 uint16_t *p = (uint16_t*)buf1;\
698 for( int i = 0; i < 16 && ok; i++ )\
700 for( int j = 0; j < 16; j++ )\
701 dct1[0][j] = !i ? (j^j>>1^j>>2^j>>3)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
702 : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
703 : ((*p++)&0x1fff)-0x1000; /* general case */\
704 memcpy( dct2, dct1, 16 * sizeof(dctcoef) );\
705 call_c1( dct_c.name, dct1[0] );\
706 call_a1( dct_asm.name, dct2[0] );\
707 if( memcmp( dct1, dct2, 16 * sizeof(dctcoef) ) )\
710 call_c2( dct_c.name, dct1[0] );\
711 call_a2( dct_asm.name, dct2[0] );\
713 report( #name " :" );
715 TEST_DCTDC( dct4x4dc );
716 TEST_DCTDC( idct4x4dc );
719 x264_zigzag_function_t zigzag_c[2];
720 x264_zigzag_function_t zigzag_ref[2];
721 x264_zigzag_function_t zigzag_asm[2];
723 ALIGNED_16( dctcoef level1[64] );
724 ALIGNED_16( dctcoef level2[64] );
726 #define TEST_ZIGZAG_SCAN( name, t1, t2, dct, size ) \
727 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
729 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
731 memcpy(dct, buf1, size*sizeof(dctcoef)); \
732 call_c( zigzag_c[interlace].name, t1, dct ); \
733 call_a( zigzag_asm[interlace].name, t2, dct ); \
734 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
737 fprintf( stderr, #name " [FAILED]\n" ); \
741 #define TEST_ZIGZAG_SUB( name, t1, t2, size ) \
742 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
745 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
747 memcpy( pbuf3, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
748 memcpy( pbuf4, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
749 nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
750 nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
751 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( pbuf3, pbuf4, 16*FDEC_STRIDE*sizeof(pixel) ) || nz_c != nz_a ) \
754 fprintf( stderr, #name " [FAILED]\n" ); \
756 call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
757 call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
760 #define TEST_ZIGZAG_SUBAC( name, t1, t2 ) \
761 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
764 dctcoef dc_a, dc_c; \
765 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
767 for( int i = 0; i < 2; i++ ) \
769 memcpy( pbuf3, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
770 memcpy( pbuf4, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
771 for( int j = 0; j < 4; j++ ) \
773 memcpy( pbuf3 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
774 memcpy( pbuf4 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
776 nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
777 nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
778 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 ) \
781 fprintf( stderr, #name " [FAILED]\n" ); \
785 call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
786 call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
789 #define TEST_INTERLEAVE( name, t1, t2, dct, size ) \
790 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
792 for( int j = 0; j < 100; j++ ) \
794 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
796 memcpy(dct, buf1, size*sizeof(dctcoef)); \
797 for( int i = 0; i < size; i++ ) \
798 dct[i] = rand()&0x1F ? 0 : dct[i]; \
799 memcpy(buf3, buf4, 10); \
800 call_c( zigzag_c[interlace].name, t1, dct, buf3 ); \
801 call_a( zigzag_asm[interlace].name, t2, dct, buf4 ); \
802 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( buf3, buf4, 10 ) ) \
809 x264_zigzag_init( 0, &zigzag_c[0], &zigzag_c[1] );
810 x264_zigzag_init( cpu_ref, &zigzag_ref[0], &zigzag_ref[1] );
811 x264_zigzag_init( cpu_new, &zigzag_asm[0], &zigzag_asm[1] );
813 ok = 1; used_asm = 0;
814 TEST_INTERLEAVE( interleave_8x8_cavlc, level1, level2, dct1[0], 64 );
815 report( "zigzag_interleave :" );
817 for( interlace = 0; interlace <= 1; interlace++ )
819 ok = 1; used_asm = 0;
820 TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, (void*)dct1, 64 );
821 TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 16 );
822 TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
823 TEST_ZIGZAG_SUBAC( sub_4x4ac, level1, level2 );
824 report( interlace ? "zigzag_field :" : "zigzag_frame :" );
826 #undef TEST_ZIGZAG_SCAN
827 #undef TEST_ZIGZAG_SUB
832 static int check_mc( int cpu_ref, int cpu_new )
834 x264_mc_functions_t mc_c;
835 x264_mc_functions_t mc_ref;
836 x264_mc_functions_t mc_a;
837 x264_pixel_function_t pixf;
839 pixel *src = &(pbuf1)[2*64+2];
840 pixel *src2[4] = { &(pbuf1)[3*64+2], &(pbuf1)[5*64+2],
841 &(pbuf1)[7*64+2], &(pbuf1)[9*64+2] };
845 int ret = 0, ok, used_asm;
847 x264_mc_init( 0, &mc_c );
848 x264_mc_init( cpu_ref, &mc_ref );
849 x264_mc_init( cpu_new, &mc_a );
850 x264_pixel_init( 0, &pixf );
852 #define MC_TEST_LUMA( w, h ) \
853 if( mc_a.mc_luma != mc_ref.mc_luma && !(w&(w-1)) && h<=16 ) \
855 const x264_weight_t *weight = weight_none; \
856 set_func_name( "mc_luma_%dx%d", w, h ); \
858 for( int i = 0; i < 1024; i++ ) \
859 pbuf3[i] = pbuf4[i] = 0xCD; \
860 call_c( mc_c.mc_luma, dst1, 32, src2, 64, dx, dy, w, h, weight ); \
861 call_a( mc_a.mc_luma, dst2, 32, src2, 64, dx, dy, w, h, weight ); \
862 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
864 fprintf( stderr, "mc_luma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
868 if( mc_a.get_ref != mc_ref.get_ref ) \
871 int ref_stride = 32; \
872 int w_checked = ( ( sizeof(pixel) == 2 && (w == 12 || w == 20)) ? w-2 : w ); \
873 const x264_weight_t *weight = weight_none; \
874 set_func_name( "get_ref_%dx%d", w_checked, h ); \
876 for( int i = 0; i < 1024; i++ ) \
877 pbuf3[i] = pbuf4[i] = 0xCD; \
878 call_c( mc_c.mc_luma, dst1, 32, src2, 64, dx, dy, w, h, weight ); \
879 ref = (pixel*)call_a( mc_a.get_ref, ref, &ref_stride, src2, 64, dx, dy, w, h, weight ); \
880 for( int i = 0; i < h; i++ ) \
881 if( memcmp( dst1+i*32, ref+i*ref_stride, w_checked * sizeof(pixel) ) ) \
883 fprintf( stderr, "get_ref[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w_checked, h ); \
889 #define MC_TEST_CHROMA( w, h ) \
890 if( mc_a.mc_chroma != mc_ref.mc_chroma ) \
892 set_func_name( "mc_chroma_%dx%d", w, h ); \
894 for( int i = 0; i < 1024; i++ ) \
895 pbuf3[i] = pbuf4[i] = 0xCD; \
896 call_c( mc_c.mc_chroma, dst1, dst1+8, 16, src, 64, dx, dy, w, h ); \
897 call_a( mc_a.mc_chroma, dst2, dst2+8, 16, src, 64, dx, dy, w, h ); \
898 /* mc_chroma width=2 may write garbage to the right of dst. ignore that. */ \
899 for( int j = 0; j < h; j++ ) \
900 for( int i = w; i < 8; i++ ) \
902 dst2[i+j*16+8] = dst1[i+j*16+8]; \
903 dst2[i+j*16] = dst1[i+j*16]; \
905 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
907 fprintf( stderr, "mc_chroma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
911 ok = 1; used_asm = 0;
912 for( int dy = -8; dy < 8; dy++ )
913 for( int dx = -128; dx < 128; dx++ )
915 if( rand()&15 ) continue; // running all of them is too slow
916 MC_TEST_LUMA( 20, 18 );
917 MC_TEST_LUMA( 16, 16 );
918 MC_TEST_LUMA( 16, 8 );
919 MC_TEST_LUMA( 12, 10 );
920 MC_TEST_LUMA( 8, 16 );
921 MC_TEST_LUMA( 8, 8 );
922 MC_TEST_LUMA( 8, 4 );
923 MC_TEST_LUMA( 4, 8 );
924 MC_TEST_LUMA( 4, 4 );
926 report( "mc luma :" );
928 ok = 1; used_asm = 0;
929 for( int dy = -1; dy < 9; dy++ )
930 for( int dx = -128; dx < 128; dx++ )
932 if( rand()&15 ) continue;
933 MC_TEST_CHROMA( 8, 8 );
934 MC_TEST_CHROMA( 8, 4 );
935 MC_TEST_CHROMA( 4, 8 );
936 MC_TEST_CHROMA( 4, 4 );
937 MC_TEST_CHROMA( 4, 2 );
938 MC_TEST_CHROMA( 2, 4 );
939 MC_TEST_CHROMA( 2, 2 );
941 report( "mc chroma :" );
943 #undef MC_TEST_CHROMA
945 #define MC_TEST_AVG( name, weight ) \
947 ok = 1, used_asm = 0; \
948 for( int i = 0; i < 10; i++ ) \
950 memcpy( pbuf3, pbuf1+320, 320 * sizeof(pixel) ); \
951 memcpy( pbuf4, pbuf1+320, 320 * sizeof(pixel) ); \
952 if( mc_a.name[i] != mc_ref.name[i] ) \
954 set_func_name( "%s_%s", #name, pixel_names[i] ); \
956 call_c1( mc_c.name[i], pbuf3, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
957 call_a1( mc_a.name[i], pbuf4, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
958 if( memcmp( pbuf3, pbuf4, 320 * sizeof(pixel) ) ) \
961 fprintf( stderr, #name "[%d]: [FAILED]\n", i ); \
963 call_c2( mc_c.name[i], pbuf3, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
964 call_a2( mc_a.name[i], pbuf4, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
969 for( int w = -63; w <= 127 && ok; w++ )
970 MC_TEST_AVG( avg, w );
971 report( "mc wpredb :" );
973 #define MC_TEST_WEIGHT( name, weight, aligned ) \
974 int align_off = (aligned ? 0 : rand()%16); \
975 ok = 1, used_asm = 0; \
976 for( int i = 1; i <= 5; i++ ) \
978 ALIGNED_16( pixel buffC[640] ); \
979 ALIGNED_16( pixel buffA[640] ); \
980 int j = X264_MAX( i*4, 2 ); \
981 memset( buffC, 0, 640 * sizeof(pixel) ); \
982 memset( buffA, 0, 640 * sizeof(pixel) ); \
985 /* w12 is the same as w16 in some cases */ \
986 if( i == 3 && mc_a.name[i] == mc_a.name[i+1] ) \
988 if( mc_a.name[i] != mc_ref.name[i] ) \
990 set_func_name( "%s_w%d", #name, j ); \
992 call_c1( mc_c.weight[i], buffC, 32, pbuf2+align_off, 32, &weight, 16 ); \
993 mc_a.weight_cache(&ha, &weight); \
994 call_a1( weight.weightfn[i], buffA, 32, pbuf2+align_off, 32, &weight, 16 ); \
995 for( int k = 0; k < 16; k++ ) \
996 if( memcmp( &buffC[k*32], &buffA[k*32], j * sizeof(pixel) ) ) \
999 fprintf( stderr, #name "[%d]: [FAILED] s:%d o:%d d%d\n", i, s, o, d ); \
1002 call_c2( mc_c.weight[i], buffC, 32, pbuf2+align_off, 32, &weight, 16 ); \
1003 call_a2( weight.weightfn[i], buffA, 32, pbuf2+align_off, 32, &weight, 16 ); \
1007 ok = 1; used_asm = 0;
1010 for( int s = 0; s <= 127 && ok; s++ )
1012 for( int o = -128; o <= 127 && ok; o++ )
1014 if( rand() & 2047 ) continue;
1015 for( int d = 0; d <= 7 && ok; d++ )
1019 x264_weight_t weight = { .i_scale = s, .i_denom = d, .i_offset = o };
1020 MC_TEST_WEIGHT( weight, weight, (align_cnt++ % 4) );
1025 report( "mc weight :" );
1027 ok = 1; used_asm = 0;
1028 for( int o = 0; o <= 127 && ok; o++ )
1031 if( rand() & 15 ) continue;
1032 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1033 MC_TEST_WEIGHT( offsetadd, weight, (align_cnt++ % 4) );
1035 report( "mc offsetadd :" );
1036 ok = 1; used_asm = 0;
1037 for( int o = -128; o < 0 && ok; o++ )
1040 if( rand() & 15 ) continue;
1041 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1042 MC_TEST_WEIGHT( offsetsub, weight, (align_cnt++ % 4) );
1044 report( "mc offsetsub :" );
1046 ok = 1; used_asm = 0;
1047 if( mc_a.store_interleave_8x8x2 != mc_ref.store_interleave_8x8x2 )
1049 set_func_name( "store_interleave_8x8x2" );
1051 memset( pbuf3, 0, 64*8 );
1052 memset( pbuf4, 0, 64*8 );
1053 call_c( mc_c.store_interleave_8x8x2, pbuf3, 64, pbuf1, pbuf1+16 );
1054 call_a( mc_a.store_interleave_8x8x2, pbuf4, 64, pbuf1, pbuf1+16 );
1055 if( memcmp( pbuf3, pbuf4, 64*8 ) )
1058 if( mc_a.load_deinterleave_8x8x2_fenc != mc_ref.load_deinterleave_8x8x2_fenc )
1060 set_func_name( "load_deinterleave_8x8x2_fenc" );
1062 call_c( mc_c.load_deinterleave_8x8x2_fenc, pbuf3, pbuf1, 64 );
1063 call_a( mc_a.load_deinterleave_8x8x2_fenc, pbuf4, pbuf1, 64 );
1064 if( memcmp( pbuf3, pbuf4, FENC_STRIDE*8 ) )
1067 if( mc_a.load_deinterleave_8x8x2_fdec != mc_ref.load_deinterleave_8x8x2_fdec )
1069 set_func_name( "load_deinterleave_8x8x2_fdec" );
1071 call_c( mc_c.load_deinterleave_8x8x2_fdec, pbuf3, pbuf1, 64 );
1072 call_a( mc_a.load_deinterleave_8x8x2_fdec, pbuf4, pbuf1, 64 );
1073 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*8 ) )
1076 report( "store_interleave :" );
1079 int w, h, src_stride;
1080 } 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} };
1081 ok = 1; used_asm = 0;
1082 if( mc_a.plane_copy != mc_ref.plane_copy )
1084 set_func_name( "plane_copy" );
1086 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1088 int w = plane_specs[i].w;
1089 int h = plane_specs[i].h;
1090 int src_stride = plane_specs[i].src_stride;
1091 int dst_stride = (w + 127) & ~63;
1092 assert( dst_stride * h <= 0x1000 );
1093 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1094 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1095 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1096 call_c( mc_c.plane_copy, pbuf3, dst_stride, src1, src_stride, w, h );
1097 call_a( mc_a.plane_copy, pbuf4, dst_stride, src1, src_stride, w, h );
1098 for( int y = 0; y < h; y++ )
1099 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(pixel) ) )
1102 fprintf( stderr, "plane_copy FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
1108 if( mc_a.plane_copy_interleave != mc_ref.plane_copy_interleave )
1110 set_func_name( "plane_copy_interleave" );
1112 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1114 int w = (plane_specs[i].w + 1) >> 1;
1115 int h = plane_specs[i].h;
1116 int src_stride = (plane_specs[i].src_stride + 1) >> 1;
1117 int dst_stride = (2*w + 127) & ~63;
1118 assert( dst_stride * h <= 0x1000 );
1119 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1120 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1121 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1122 call_c( mc_c.plane_copy_interleave, pbuf3, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1123 call_a( mc_a.plane_copy_interleave, pbuf4, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1124 for( int y = 0; y < h; y++ )
1125 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
1128 fprintf( stderr, "plane_copy_interleave FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
1134 if( mc_a.plane_copy_deinterleave != mc_ref.plane_copy_deinterleave )
1136 set_func_name( "plane_copy_deinterleave" );
1138 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1140 int w = (plane_specs[i].w + 1) >> 1;
1141 int h = plane_specs[i].h;
1143 int src_stride = (2*w + 127) & ~63;
1144 int offv = (dst_stride*h + 31) & ~15;
1145 memset( pbuf3, 0, 0x1000 );
1146 memset( pbuf4, 0, 0x1000 );
1147 call_c( mc_c.plane_copy_deinterleave, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf1, src_stride, w, h );
1148 call_a( mc_a.plane_copy_deinterleave, pbuf4, dst_stride, pbuf4+offv, dst_stride, pbuf1, src_stride, w, h );
1149 for( int y = 0; y < h; y++ )
1150 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w ) ||
1151 memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w ) )
1154 fprintf( stderr, "plane_copy_deinterleave FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
1159 report( "plane_copy :" );
1161 if( mc_a.hpel_filter != mc_ref.hpel_filter )
1163 pixel *srchpel = pbuf1+8+2*64;
1164 pixel *dstc[3] = { pbuf3+8, pbuf3+8+16*64, pbuf3+8+32*64 };
1165 pixel *dsta[3] = { pbuf4+8, pbuf4+8+16*64, pbuf4+8+32*64 };
1166 void *tmp = pbuf3+49*64;
1167 set_func_name( "hpel_filter" );
1168 ok = 1; used_asm = 1;
1169 memset( pbuf3, 0, 4096 * sizeof(pixel) );
1170 memset( pbuf4, 0, 4096 * sizeof(pixel) );
1171 call_c( mc_c.hpel_filter, dstc[0], dstc[1], dstc[2], srchpel, 64, 48, 10, tmp );
1172 call_a( mc_a.hpel_filter, dsta[0], dsta[1], dsta[2], srchpel, 64, 48, 10, tmp );
1173 for( int i = 0; i < 3; i++ )
1174 for( int j = 0; j < 10; j++ )
1175 //FIXME ideally the first pixels would match too, but they aren't actually used
1176 if( memcmp( dstc[i]+j*64+2, dsta[i]+j*64+2, 43 * sizeof(pixel) ) )
1179 fprintf( stderr, "hpel filter differs at plane %c line %d\n", "hvc"[i], j );
1180 for( int k = 0; k < 48; k++ )
1181 printf( "%02x%s", dstc[i][j*64+k], (k+1)&3 ? "" : " " );
1183 for( int k = 0; k < 48; k++ )
1184 printf( "%02x%s", dsta[i][j*64+k], (k+1)&3 ? "" : " " );
1188 report( "hpel filter :" );
1191 if( mc_a.frame_init_lowres_core != mc_ref.frame_init_lowres_core )
1193 pixel *dstc[4] = { pbuf3, pbuf3+1024, pbuf3+2048, pbuf3+3072 };
1194 pixel *dsta[4] = { pbuf4, pbuf4+1024, pbuf4+2048, pbuf4+3072 };
1195 set_func_name( "lowres_init" );
1196 ok = 1; used_asm = 1;
1197 for( int w = 40; w <= 48; w += 8 )
1199 int stride = (w+8)&~15;
1200 call_c( mc_c.frame_init_lowres_core, pbuf1, dstc[0], dstc[1], dstc[2], dstc[3], w*2, stride, w, 16 );
1201 call_a( mc_a.frame_init_lowres_core, pbuf1, dsta[0], dsta[1], dsta[2], dsta[3], w*2, stride, w, 16 );
1202 for( int i = 0; i < 16; i++ )
1204 for( int j = 0; j < 4; j++ )
1205 if( memcmp( dstc[j]+i*stride, dsta[j]+i*stride, w * sizeof(pixel) ) )
1208 fprintf( stderr, "frame_init_lowres differs at plane %d line %d\n", j, i );
1209 for( int k = 0; k < w; k++ )
1210 printf( "%d ", dstc[j][k+i*stride] );
1212 for( int k = 0; k < w; k++ )
1213 printf( "%d ", dsta[j][k+i*stride] );
1219 report( "lowres init :" );
1222 #define INTEGRAL_INIT( name, size, ... )\
1223 if( mc_a.name != mc_ref.name )\
1226 set_func_name( #name );\
1228 memcpy( buf3, buf1, size*2*stride );\
1229 memcpy( buf4, buf1, size*2*stride );\
1230 uint16_t *sum = (uint16_t*)buf3;\
1231 call_c1( mc_c.name, __VA_ARGS__ );\
1232 sum = (uint16_t*)buf4;\
1233 call_a1( mc_a.name, __VA_ARGS__ );\
1234 if( memcmp( buf3, buf4, (stride-8)*2 ) \
1235 || (size>9 && memcmp( buf3+18*stride, buf4+18*stride, (stride-8)*2 )))\
1237 call_c2( mc_c.name, __VA_ARGS__ );\
1238 call_a2( mc_a.name, __VA_ARGS__ );\
1240 ok = 1; used_asm = 0;
1241 INTEGRAL_INIT( integral_init4h, 2, sum+stride, pbuf2, stride );
1242 INTEGRAL_INIT( integral_init8h, 2, sum+stride, pbuf2, stride );
1243 INTEGRAL_INIT( integral_init4v, 14, sum, sum+9*stride, stride );
1244 INTEGRAL_INIT( integral_init8v, 9, sum, stride );
1245 report( "integral init :" );
1247 if( mc_a.mbtree_propagate_cost != mc_ref.mbtree_propagate_cost )
1250 for( int i = 0; i < 10; i++ )
1252 float fps_factor = (rand()&65535) / 256.;
1253 ok = 1; used_asm = 1;
1254 set_func_name( "mbtree_propagate" );
1255 int *dsta = (int*)buf3;
1256 int *dstc = dsta+400;
1257 uint16_t *prop = (uint16_t*)buf1;
1258 uint16_t *intra = (uint16_t*)buf4;
1259 uint16_t *inter = intra+128;
1260 uint16_t *qscale = inter+128;
1261 uint16_t *rnd = (uint16_t*)buf2;
1263 for( int j = 0; j < 100; j++ )
1265 intra[j] = *rnd++ & 0x7fff;
1266 intra[j] += !intra[j];
1267 inter[j] = *rnd++ & 0x7fff;
1268 qscale[j] = *rnd++ & 0x7fff;
1270 call_c( mc_c.mbtree_propagate_cost, dstc, prop, intra, inter, qscale, &fps_factor, 100 );
1271 call_a( mc_a.mbtree_propagate_cost, dsta, prop, intra, inter, qscale, &fps_factor, 100 );
1272 // I don't care about exact rounding, this is just how close the floating-point implementation happens to be
1274 for( int j = 0; j < 100; j++ )
1275 ok &= abs( dstc[j]-dsta[j] ) <= 1 || fabs( (double)dstc[j]/dsta[j]-1 ) < 1e-4;
1277 report( "mbtree propagate :" );
1280 if( mc_a.memcpy_aligned != mc_ref.memcpy_aligned )
1282 set_func_name( "memcpy_aligned" );
1283 ok = 1; used_asm = 1;
1284 for( int size = 16; size < 256; size += 16 )
1286 memset( buf4, 0xAA, size + 1 );
1287 call_c( mc_c.memcpy_aligned, buf3, buf1, size );
1288 call_a( mc_a.memcpy_aligned, buf4, buf1, size );
1289 if( memcmp( buf3, buf4, size ) || buf4[size] != 0xAA )
1292 fprintf( stderr, "memcpy_aligned FAILED: size=%d\n", size );
1296 report( "memcpy aligned :" );
1299 if( mc_a.memzero_aligned != mc_ref.memzero_aligned )
1301 set_func_name( "memzero_aligned" );
1302 ok = 1; used_asm = 1;
1303 for( int size = 128; size < 1024; size += 128 )
1305 memset( buf4, 0xAA, size + 1 );
1306 call_c( mc_c.memzero_aligned, buf3, size );
1307 call_a( mc_a.memzero_aligned, buf4, size );
1308 if( memcmp( buf3, buf4, size ) || buf4[size] != 0xAA )
1311 fprintf( stderr, "memzero_aligned FAILED: size=%d\n", size );
1315 report( "memzero aligned :" );
1321 static int check_deblock( int cpu_ref, int cpu_new )
1323 x264_deblock_function_t db_c;
1324 x264_deblock_function_t db_ref;
1325 x264_deblock_function_t db_a;
1326 int ret = 0, ok = 1, used_asm = 0;
1327 int alphas[36], betas[36];
1330 x264_deblock_init( 0, &db_c, 0 );
1331 x264_deblock_init( cpu_ref, &db_ref, 0 );
1332 x264_deblock_init( cpu_new, &db_a, 0 );
1334 /* not exactly the real values of a,b,tc but close enough */
1335 for( int i = 35, a = 255, c = 250; i >= 0; i-- )
1337 alphas[i] = a << (BIT_DEPTH-8);
1338 betas[i] = (i+1)/2 << (BIT_DEPTH-8);
1339 tcs[i][0] = tcs[i][3] = (c+6)/10 << (BIT_DEPTH-8);
1340 tcs[i][1] = (c+7)/15 << (BIT_DEPTH-8);
1341 tcs[i][2] = (c+9)/20 << (BIT_DEPTH-8);
1346 #define TEST_DEBLOCK( name, align, ... ) \
1347 for( int i = 0; i < 36; i++ ) \
1349 int off = 8*32 + (i&15)*4*!align; /* benchmark various alignments of h filter */ \
1350 for( int j = 0; j < 1024; j++ ) \
1351 /* two distributions of random to excersize different failure modes */ \
1352 pbuf3[j] = rand() & (i&1 ? 0xf : PIXEL_MAX ); \
1353 memcpy( pbuf4, pbuf3, 1024 * sizeof(pixel) ); \
1354 if( db_a.name != db_ref.name ) \
1356 set_func_name( #name ); \
1358 call_c1( db_c.name, pbuf3+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1359 call_a1( db_a.name, pbuf4+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1360 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1363 fprintf( stderr, #name "(a=%d, b=%d): [FAILED]\n", alphas[i], betas[i] ); \
1366 call_c2( db_c.name, pbuf3+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1367 call_a2( db_a.name, pbuf4+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1371 TEST_DEBLOCK( deblock_luma[0], 0, tcs[i] );
1372 TEST_DEBLOCK( deblock_luma[1], 1, tcs[i] );
1373 TEST_DEBLOCK( deblock_chroma[0], 0, tcs[i] );
1374 TEST_DEBLOCK( deblock_chroma[1], 1, tcs[i] );
1375 TEST_DEBLOCK( deblock_luma_intra[0], 0 );
1376 TEST_DEBLOCK( deblock_luma_intra[1], 1 );
1377 TEST_DEBLOCK( deblock_chroma_intra[0], 0 );
1378 TEST_DEBLOCK( deblock_chroma_intra[1], 1 );
1380 if( db_a.deblock_strength != db_ref.deblock_strength )
1382 for( int i = 0; i < 100; i++ )
1384 ALIGNED_ARRAY_16( uint8_t, nnz, [X264_SCAN8_SIZE] );
1385 ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );
1386 ALIGNED_ARRAY_16( int16_t, mv, [2],[X264_SCAN8_LUMA_SIZE][2] );
1387 ALIGNED_ARRAY_16( uint8_t, bs, [2],[2][8][4] );
1388 memset( bs, 99, sizeof(bs) );
1389 for( int j = 0; j < X264_SCAN8_SIZE; j++ )
1390 nnz[j] = ((rand()&7) == 7) * rand() & 0xf;
1391 for( int j = 0; j < 2; j++ )
1392 for( int k = 0; k < X264_SCAN8_LUMA_SIZE; k++ )
1394 ref[j][k] = ((rand()&3) != 3) ? 0 : (rand() & 31) - 2;
1395 for( int l = 0; l < 2; l++ )
1396 mv[j][k][l] = ((rand()&7) != 7) ? (rand()&7) - 3 : (rand()&1023) - 512;
1398 set_func_name( "deblock_strength" );
1399 call_c( db_c.deblock_strength, nnz, ref, mv, bs[0], 2<<(i&1), ((i>>1)&1), NULL );
1400 call_a( db_a.deblock_strength, nnz, ref, mv, bs[1], 2<<(i&1), ((i>>1)&1), NULL );
1401 if( memcmp( bs[0], bs[1], sizeof(bs[0]) ) )
1404 fprintf( stderr, "deblock_strength: [FAILED]\n" );
1405 for( int j = 0; j < 2; j++ )
1407 for( int k = 0; k < 2; k++ )
1408 for( int l = 0; l < 4; l++ )
1410 for( int m = 0; m < 4; m++ )
1411 printf("%d ",bs[j][k][l][m]);
1421 report( "deblock :" );
1426 static int check_quant( int cpu_ref, int cpu_new )
1428 x264_quant_function_t qf_c;
1429 x264_quant_function_t qf_ref;
1430 x264_quant_function_t qf_a;
1431 ALIGNED_16( dctcoef dct1[64] );
1432 ALIGNED_16( dctcoef dct2[64] );
1433 ALIGNED_16( uint8_t cqm_buf[64] );
1434 int ret = 0, ok, used_asm;
1435 int oks[3] = {1,1,1}, used_asms[3] = {0,0,0};
1438 memset( h, 0, sizeof(*h) );
1439 h->sps->i_chroma_format_idc = 1;
1440 x264_param_default( &h->param );
1441 h->chroma_qp_table = i_chroma_qp_table + 12;
1442 h->param.analyse.b_transform_8x8 = 1;
1444 for( int i_cqm = 0; i_cqm < 4; i_cqm++ )
1448 for( int i = 0; i < 6; i++ )
1449 h->pps->scaling_list[i] = x264_cqm_flat16;
1450 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_FLAT;
1452 else if( i_cqm == 1 )
1454 for( int i = 0; i < 6; i++ )
1455 h->pps->scaling_list[i] = x264_cqm_jvt[i];
1456 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_JVT;
1460 int max_scale = BIT_DEPTH < 10 ? 255 : 228;
1462 for( int i = 0; i < 64; i++ )
1463 cqm_buf[i] = 10 + rand() % (max_scale - 9);
1465 for( int i = 0; i < 64; i++ )
1467 for( int i = 0; i < 6; i++ )
1468 h->pps->scaling_list[i] = cqm_buf;
1469 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_CUSTOM;
1472 h->param.rc.i_qp_min = 0;
1473 h->param.rc.i_qp_max = QP_MAX;
1475 x264_quant_init( h, 0, &qf_c );
1476 x264_quant_init( h, cpu_ref, &qf_ref );
1477 x264_quant_init( h, cpu_new, &qf_a );
1479 #define INIT_QUANT8(j) \
1481 static const int scale1d[8] = {32,31,24,31,32,31,24,31}; \
1482 for( int i = 0; i < 64; i++ ) \
1484 unsigned int scale = (255*scale1d[i>>3]*scale1d[i&7])/16; \
1485 dct1[i] = dct2[i] = j ? (rand()%(2*scale+1))-scale : 0; \
1489 #define INIT_QUANT4(j) \
1491 static const int scale1d[4] = {4,6,4,6}; \
1492 for( int i = 0; i < 16; i++ ) \
1494 unsigned int scale = 255*scale1d[i>>2]*scale1d[i&3]; \
1495 dct1[i] = dct2[i] = j ? (rand()%(2*scale+1))-scale : 0; \
1499 #define TEST_QUANT_DC( name, cqm ) \
1500 if( qf_a.name != qf_ref.name ) \
1502 set_func_name( #name ); \
1504 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1506 for( int j = 0; j < 2; j++ ) \
1508 int result_c, result_a; \
1509 for( int i = 0; i < 16; i++ ) \
1510 dct1[i] = dct2[i] = j ? (rand() & 0x1fff) - 0xfff : 0; \
1511 result_c = call_c1( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1512 result_a = call_a1( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1513 if( memcmp( dct1, dct2, 16*sizeof(dctcoef) ) || result_c != result_a ) \
1516 fprintf( stderr, #name "(cqm=%d): [FAILED]\n", i_cqm ); \
1519 call_c2( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1520 call_a2( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1525 #define TEST_QUANT( qname, block, w ) \
1526 if( qf_a.qname != qf_ref.qname ) \
1528 set_func_name( #qname ); \
1530 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1532 for( int j = 0; j < 2; j++ ) \
1535 int result_c = call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1536 int result_a = call_a1( qf_a.qname, dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1537 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) || result_c != result_a ) \
1540 fprintf( stderr, #qname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1543 call_c2( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1544 call_a2( qf_a.qname, dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1549 TEST_QUANT( quant_8x8, CQM_8IY, 8 );
1550 TEST_QUANT( quant_8x8, CQM_8PY, 8 );
1551 TEST_QUANT( quant_4x4, CQM_4IY, 4 );
1552 TEST_QUANT( quant_4x4, CQM_4PY, 4 );
1553 TEST_QUANT_DC( quant_4x4_dc, **h->quant4_mf[CQM_4IY] );
1554 TEST_QUANT_DC( quant_2x2_dc, **h->quant4_mf[CQM_4IC] );
1556 #define TEST_DEQUANT( qname, dqname, block, w ) \
1557 if( qf_a.dqname != qf_ref.dqname ) \
1559 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
1561 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1564 call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1565 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1566 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1567 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1568 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1571 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1574 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1575 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1579 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8IY, 8 );
1580 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8PY, 8 );
1581 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4IY, 4 );
1582 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4PY, 4 );
1584 #define TEST_DEQUANT_DC( qname, dqname, block, w ) \
1585 if( qf_a.dqname != qf_ref.dqname ) \
1587 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
1589 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1591 for( int i = 0; i < 16; i++ ) \
1592 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16; \
1593 call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp][0]>>1, h->quant##w##_bias[block][qp][0]>>1 ); \
1594 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1595 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1596 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1597 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1600 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1602 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1603 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1607 TEST_DEQUANT_DC( quant_4x4_dc, dequant_4x4_dc, CQM_4IY, 4 );
1609 #define TEST_OPTIMIZE_CHROMA_DC( qname, optname, w ) \
1610 if( qf_a.optname != qf_ref.optname ) \
1612 set_func_name( #optname ); \
1614 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1616 int dmf = h->dequant4_mf[CQM_4IC][qp%6][0] << qp/6; \
1619 for( int i = 16; ; i <<= 1 )\
1621 int res_c, res_asm; \
1622 int max = X264_MIN( i, PIXEL_MAX*16 ); \
1623 for( int j = 0; j < w*w; j++ ) \
1624 dct1[j] = rand()%(max*2+1) - max; \
1625 call_c1( qf_c.qname, dct1, h->quant4_mf[CQM_4IC][qp][0]>>1, h->quant4_bias[CQM_4IC][qp][0]>>1 ); \
1626 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1627 res_c = call_c1( qf_c.optname, dct1, dmf ); \
1628 res_asm = call_a1( qf_a.optname, dct2, dmf ); \
1629 if( res_c != res_asm || memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1632 fprintf( stderr, #optname "(qp=%d, res_c=%d, res_asm=%d): [FAILED]\n", qp, res_c, res_asm ); \
1634 call_c2( qf_c.optname, dct1, dmf ); \
1635 call_a2( qf_a.optname, dct2, dmf ); \
1636 if( i >= PIXEL_MAX*16 ) \
1642 TEST_OPTIMIZE_CHROMA_DC( quant_2x2_dc, optimize_chroma_dc, 2 );
1644 x264_cqm_delete( h );
1647 ok = oks[0]; used_asm = used_asms[0];
1648 report( "quant :" );
1650 ok = oks[1]; used_asm = used_asms[1];
1651 report( "dequant :" );
1653 ok = oks[2]; used_asm = used_asms[2];
1654 report( "optimize chroma dc :" );
1656 ok = 1; used_asm = 0;
1657 if( qf_a.denoise_dct != qf_ref.denoise_dct )
1660 for( int size = 16; size <= 64; size += 48 )
1662 set_func_name( "denoise_dct" );
1663 memcpy( dct1, buf1, size*sizeof(dctcoef) );
1664 memcpy( dct2, buf1, size*sizeof(dctcoef) );
1665 memcpy( buf3+256, buf3, 256 );
1666 call_c1( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
1667 call_a1( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
1668 if( memcmp( dct1, dct2, size*sizeof(dctcoef) ) || memcmp( buf3+4, buf3+256+4, (size-1)*sizeof(uint32_t) ) )
1670 call_c2( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
1671 call_a2( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
1674 report( "denoise dct :" );
1676 #define TEST_DECIMATE( decname, w, ac, thresh ) \
1677 if( qf_a.decname != qf_ref.decname ) \
1679 set_func_name( #decname ); \
1681 for( int i = 0; i < 100; i++ ) \
1683 static const int distrib[16] = {1,1,1,1,1,1,1,1,1,1,1,1,2,3,4};\
1684 static const int zerorate_lut[4] = {3,7,15,31};\
1685 int zero_rate = zerorate_lut[i&3];\
1686 for( int idx = 0; idx < w*w; idx++ ) \
1688 int sign = (rand()&1) ? -1 : 1; \
1689 int abs_level = distrib[rand()&15]; \
1690 if( abs_level == 4 ) abs_level = rand()&0x3fff; \
1691 int zero = !(rand()&zero_rate); \
1692 dct1[idx] = zero * abs_level * sign; \
1696 int result_c = call_c( qf_c.decname, dct1 ); \
1697 int result_a = call_a( qf_a.decname, dct1 ); \
1698 if( X264_MIN(result_c,thresh) != X264_MIN(result_a,thresh) ) \
1701 fprintf( stderr, #decname ": [FAILED]\n" ); \
1707 ok = 1; used_asm = 0;
1708 TEST_DECIMATE( decimate_score64, 8, 0, 6 );
1709 TEST_DECIMATE( decimate_score16, 4, 0, 6 );
1710 TEST_DECIMATE( decimate_score15, 4, 1, 7 );
1711 report( "decimate_score :" );
1713 #define TEST_LAST( last, lastname, w, ac ) \
1714 if( qf_a.last != qf_ref.last ) \
1716 set_func_name( #lastname ); \
1718 for( int i = 0; i < 100; i++ ) \
1721 int max = rand() & (w*w-1); \
1722 memset( dct1, 0, w*w*sizeof(dctcoef) ); \
1723 for( int idx = ac; idx < max; idx++ ) \
1724 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
1727 int result_c = call_c( qf_c.last, dct1+ac ); \
1728 int result_a = call_a( qf_a.last, dct1+ac ); \
1729 if( result_c != result_a ) \
1732 fprintf( stderr, #lastname ": [FAILED]\n" ); \
1738 ok = 1; used_asm = 0;
1739 TEST_LAST( coeff_last[DCT_CHROMA_DC], coeff_last4, 2, 0 );
1740 TEST_LAST( coeff_last[ DCT_LUMA_AC], coeff_last15, 4, 1 );
1741 TEST_LAST( coeff_last[ DCT_LUMA_4x4], coeff_last16, 4, 0 );
1742 TEST_LAST( coeff_last[ DCT_LUMA_8x8], coeff_last64, 8, 0 );
1743 report( "coeff_last :" );
1745 #define TEST_LEVELRUN( lastname, name, w, ac ) \
1746 if( qf_a.lastname != qf_ref.lastname ) \
1748 set_func_name( #name ); \
1750 for( int i = 0; i < 100; i++ ) \
1752 x264_run_level_t runlevel_c, runlevel_a; \
1754 int max = rand() & (w*w-1); \
1755 memset( dct1, 0, w*w*sizeof(dctcoef) ); \
1756 memcpy( &runlevel_a, buf1+i, sizeof(x264_run_level_t) ); \
1757 memcpy( &runlevel_c, buf1+i, sizeof(x264_run_level_t) ); \
1758 for( int idx = ac; idx < max; idx++ ) \
1759 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
1762 int result_c = call_c( qf_c.lastname, dct1+ac, &runlevel_c ); \
1763 int result_a = call_a( qf_a.lastname, dct1+ac, &runlevel_a ); \
1764 if( result_c != result_a || runlevel_c.last != runlevel_a.last || \
1765 memcmp(runlevel_c.level, runlevel_a.level, sizeof(dctcoef)*result_c) || \
1766 memcmp(runlevel_c.run, runlevel_a.run, sizeof(uint8_t)*(result_c-1)) ) \
1769 fprintf( stderr, #name ": [FAILED]\n" ); \
1775 ok = 1; used_asm = 0;
1776 TEST_LEVELRUN( coeff_level_run[DCT_CHROMA_DC], coeff_level_run4, 2, 0 );
1777 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_AC], coeff_level_run15, 4, 1 );
1778 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_4x4], coeff_level_run16, 4, 0 );
1779 report( "coeff_level_run :" );
1784 static int check_intra( int cpu_ref, int cpu_new )
1786 int ret = 0, ok = 1, used_asm = 0;
1787 ALIGNED_16( pixel edge[33] );
1788 ALIGNED_16( pixel edge2[33] );
1789 ALIGNED_16( pixel fdec[FDEC_STRIDE*20] );
1792 x264_predict_t predict_16x16[4+3];
1793 x264_predict_t predict_8x8c[4+3];
1794 x264_predict8x8_t predict_8x8[9+3];
1795 x264_predict_t predict_4x4[9+3];
1796 x264_predict_8x8_filter_t predict_8x8_filter;
1797 } ip_c, ip_ref, ip_a;
1799 x264_predict_16x16_init( 0, ip_c.predict_16x16 );
1800 x264_predict_8x8c_init( 0, ip_c.predict_8x8c );
1801 x264_predict_8x8_init( 0, ip_c.predict_8x8, &ip_c.predict_8x8_filter );
1802 x264_predict_4x4_init( 0, ip_c.predict_4x4 );
1804 x264_predict_16x16_init( cpu_ref, ip_ref.predict_16x16 );
1805 x264_predict_8x8c_init( cpu_ref, ip_ref.predict_8x8c );
1806 x264_predict_8x8_init( cpu_ref, ip_ref.predict_8x8, &ip_ref.predict_8x8_filter );
1807 x264_predict_4x4_init( cpu_ref, ip_ref.predict_4x4 );
1809 x264_predict_16x16_init( cpu_new, ip_a.predict_16x16 );
1810 x264_predict_8x8c_init( cpu_new, ip_a.predict_8x8c );
1811 x264_predict_8x8_init( cpu_new, ip_a.predict_8x8, &ip_a.predict_8x8_filter );
1812 x264_predict_4x4_init( cpu_new, ip_a.predict_4x4 );
1814 memcpy( fdec, pbuf1, 32*20 * sizeof(pixel) );\
1816 ip_c.predict_8x8_filter( fdec+48, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
1818 #define INTRA_TEST( name, dir, w, bench, ... )\
1819 if( ip_a.name[dir] != ip_ref.name[dir] )\
1821 set_func_name( "intra_%s_%s", #name, intra_##name##_names[dir] );\
1823 memcpy( pbuf3, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
1824 memcpy( pbuf4, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
1825 call_c##bench( ip_c.name[dir], pbuf3+48, ##__VA_ARGS__ );\
1826 call_a##bench( ip_a.name[dir], pbuf4+48, ##__VA_ARGS__ );\
1827 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*20 * sizeof(pixel) ) )\
1829 fprintf( stderr, #name "[%d] : [FAILED]\n", dir );\
1831 for( int k = -1; k < 16; k++ )\
1832 printf( "%2x ", edge[16+k] );\
1834 for( int j = 0; j < w; j++ )\
1836 printf( "%2x ", edge[14-j] );\
1837 for( int k = 0; k < w; k++ )\
1838 printf( "%2x ", pbuf4[48+k+j*FDEC_STRIDE] );\
1842 for( int j = 0; j < w; j++ )\
1845 for( int k = 0; k < w; k++ )\
1846 printf( "%2x ", pbuf3[48+k+j*FDEC_STRIDE] );\
1852 for( int i = 0; i < 12; i++ )
1853 INTRA_TEST( predict_4x4, i, 4, );
1854 for( int i = 0; i < 7; i++ )
1855 INTRA_TEST( predict_8x8c, i, 8, );
1856 for( int i = 0; i < 7; i++ )
1857 INTRA_TEST( predict_16x16, i, 16, );
1858 for( int i = 0; i < 12; i++ )
1859 INTRA_TEST( predict_8x8, i, 8, , edge );
1861 set_func_name("intra_predict_8x8_filter");
1862 if( ip_a.predict_8x8_filter != ip_ref.predict_8x8_filter )
1865 for( int i = 0; i < 32; i++ )
1867 memcpy( edge2, edge, 33 * sizeof(pixel) );
1868 call_c(ip_c.predict_8x8_filter, pbuf1+48, edge, (i&24)>>1, i&7);
1869 call_a(ip_a.predict_8x8_filter, pbuf1+48, edge2, (i&24)>>1, i&7);
1870 if( memcmp( edge, edge2, 33 * sizeof(pixel) ) )
1872 fprintf( stderr, "predict_8x8_filter : [FAILED] %d %d\n", (i&24)>>1, i&7);
1878 #define EXTREMAL_PLANE(size) \
1881 for( int j = 0; j < 7; j++ ) \
1882 max[j] = test ? rand()&PIXEL_MAX : PIXEL_MAX; \
1883 fdec[48-1-FDEC_STRIDE] = (i&1)*max[0]; \
1884 for( int j = 0; j < size/2; j++ ) \
1885 fdec[48+j-FDEC_STRIDE] = (!!(i&2))*max[1]; \
1886 for( int j = size/2; j < size-1; j++ ) \
1887 fdec[48+j-FDEC_STRIDE] = (!!(i&4))*max[2]; \
1888 fdec[48+(size-1)-FDEC_STRIDE] = (!!(i&8))*max[3]; \
1889 for( int j = 0; j < size/2; j++ ) \
1890 fdec[48+j*FDEC_STRIDE-1] = (!!(i&16))*max[4]; \
1891 for( int j = size/2; j < size-1; j++ ) \
1892 fdec[48+j*FDEC_STRIDE-1] = (!!(i&32))*max[5]; \
1893 fdec[48+(size-1)*FDEC_STRIDE-1] = (!!(i&64))*max[6]; \
1895 /* Extremal test case for planar prediction. */
1896 for( int test = 0; test < 100 && ok; test++ )
1897 for( int i = 0; i < 128 && ok; i++ )
1899 EXTREMAL_PLANE( 8 );
1900 INTRA_TEST( predict_8x8c, I_PRED_CHROMA_P, 8, 1 );
1901 EXTREMAL_PLANE( 16 );
1902 INTRA_TEST( predict_16x16, I_PRED_16x16_P, 16, 1 );
1904 report( "intra pred :" );
1908 #define DECL_CABAC(cpu) \
1909 static void run_cabac_decision_##cpu( x264_t *h, uint8_t *dst )\
1912 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
1913 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
1914 for( int i = 0; i < 0x1000; i++ )\
1915 x264_cabac_encode_decision_##cpu( &cb, buf1[i]>>1, buf1[i]&1 );\
1917 static void run_cabac_bypass_##cpu( x264_t *h, uint8_t *dst )\
1920 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
1921 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
1922 for( int i = 0; i < 0x1000; i++ )\
1923 x264_cabac_encode_bypass_##cpu( &cb, buf1[i]&1 );\
1925 static void run_cabac_terminal_##cpu( x264_t *h, uint8_t *dst )\
1928 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
1929 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
1930 for( int i = 0; i < 0x1000; i++ )\
1931 x264_cabac_encode_terminal_##cpu( &cb );\
1937 #define run_cabac_decision_asm run_cabac_decision_c
1938 #define run_cabac_bypass_asm run_cabac_bypass_c
1939 #define run_cabac_terminal_asm run_cabac_terminal_c
1942 static int check_cabac( int cpu_ref, int cpu_new )
1944 int ret = 0, ok, used_asm = 1;
1946 h.sps->i_chroma_format_idc = 3;
1947 if( cpu_ref || run_cabac_decision_c == run_cabac_decision_asm )
1949 x264_cabac_init( &h );
1951 set_func_name( "cabac_encode_decision" );
1952 memcpy( buf4, buf3, 0x1000 );
1953 call_c( run_cabac_decision_c, &h, buf3 );
1954 call_a( run_cabac_decision_asm, &h, buf4 );
1955 ok = !memcmp( buf3, buf4, 0x1000 );
1956 report( "cabac decision:" );
1958 set_func_name( "cabac_encode_bypass" );
1959 memcpy( buf4, buf3, 0x1000 );
1960 call_c( run_cabac_bypass_c, &h, buf3 );
1961 call_a( run_cabac_bypass_asm, &h, buf4 );
1962 ok = !memcmp( buf3, buf4, 0x1000 );
1963 report( "cabac bypass:" );
1965 set_func_name( "cabac_encode_terminal" );
1966 memcpy( buf4, buf3, 0x1000 );
1967 call_c( run_cabac_terminal_c, &h, buf3 );
1968 call_a( run_cabac_terminal_asm, &h, buf4 );
1969 ok = !memcmp( buf3, buf4, 0x1000 );
1970 report( "cabac terminal:" );
1975 static int check_bitstream( int cpu_ref, int cpu_new )
1977 x264_bitstream_function_t bs_c;
1978 x264_bitstream_function_t bs_ref;
1979 x264_bitstream_function_t bs_a;
1981 int ret = 0, ok = 1, used_asm = 0;
1983 x264_bitstream_init( 0, &bs_c );
1984 x264_bitstream_init( cpu_ref, &bs_ref );
1985 x264_bitstream_init( cpu_new, &bs_a );
1986 if( bs_a.nal_escape != bs_ref.nal_escape )
1989 uint8_t *input = malloc(size+100);
1990 uint8_t *output1 = malloc(size*2);
1991 uint8_t *output2 = malloc(size*2);
1993 set_func_name( "nal_escape" );
1994 for( int i = 0; i < 100; i++ )
1996 /* Test corner-case sizes */
1997 int test_size = i < 10 ? i+1 : rand() & 0x3fff;
1998 /* Test 8 different probability distributions of zeros */
1999 for( int j = 0; j < test_size; j++ )
2000 input[j] = (rand()&((1 << ((i&7)+1)) - 1)) * rand();
2001 uint8_t *end_c = (uint8_t*)call_c1( bs_c.nal_escape, output1, input, input+test_size );
2002 uint8_t *end_a = (uint8_t*)call_a1( bs_a.nal_escape, output2, input, input+test_size );
2003 int size_c = end_c-output1;
2004 int size_a = end_a-output2;
2005 if( size_c != size_a || memcmp( output1, output2, size_c ) )
2007 fprintf( stderr, "nal_escape : [FAILED] %d %d\n", size_c, size_a );
2012 for( int j = 0; j < size; j++ )
2014 call_c2( bs_c.nal_escape, output1, input, input+size );
2015 call_a2( bs_a.nal_escape, output2, input, input+size );
2020 report( "nal escape:" );
2025 static int check_all_funcs( int cpu_ref, int cpu_new )
2027 return check_pixel( cpu_ref, cpu_new )
2028 + check_dct( cpu_ref, cpu_new )
2029 + check_mc( cpu_ref, cpu_new )
2030 + check_intra( cpu_ref, cpu_new )
2031 + check_deblock( cpu_ref, cpu_new )
2032 + check_quant( cpu_ref, cpu_new )
2033 + check_cabac( cpu_ref, cpu_new )
2034 + check_bitstream( cpu_ref, cpu_new );
2037 static int add_flags( int *cpu_ref, int *cpu_new, int flags, const char *name )
2039 *cpu_ref = *cpu_new;
2041 if( *cpu_new & X264_CPU_SSE2_IS_FAST )
2042 *cpu_new &= ~X264_CPU_SSE2_IS_SLOW;
2044 fprintf( stderr, "x264: %s\n", name );
2045 return check_all_funcs( *cpu_ref, *cpu_new );
2048 static int check_all_flags( void )
2051 int cpu0 = 0, cpu1 = 0;
2053 if( x264_cpu_detect() & X264_CPU_MMXEXT )
2055 ret |= add_flags( &cpu0, &cpu1, X264_CPU_MMX | X264_CPU_MMXEXT, "MMX" );
2056 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "MMX Cache64" );
2057 cpu1 &= ~X264_CPU_CACHELINE_64;
2059 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_32, "MMX Cache32" );
2060 cpu1 &= ~X264_CPU_CACHELINE_32;
2062 if( x264_cpu_detect() & X264_CPU_LZCNT )
2064 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "MMX_LZCNT" );
2065 cpu1 &= ~X264_CPU_LZCNT;
2067 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "MMX SlowCTZ" );
2068 cpu1 &= ~X264_CPU_SLOW_CTZ;
2070 if( x264_cpu_detect() & X264_CPU_SSE2 )
2072 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE | X264_CPU_SSE2 | X264_CPU_SSE2_IS_SLOW, "SSE2Slow" );
2073 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2_IS_FAST, "SSE2Fast" );
2074 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSE2Fast Cache64" );
2075 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SHUFFLE_IS_FAST, "SSE2 FastShuffle" );
2076 cpu1 &= ~X264_CPU_SHUFFLE_IS_FAST;
2077 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSE2 SlowCTZ" );
2078 cpu1 &= ~X264_CPU_SLOW_CTZ;
2079 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSE2 SlowAtom" );
2080 cpu1 &= ~X264_CPU_SLOW_ATOM;
2082 if( x264_cpu_detect() & X264_CPU_SSE_MISALIGN )
2084 cpu1 &= ~X264_CPU_CACHELINE_64;
2085 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE_MISALIGN, "SSE_Misalign" );
2086 cpu1 &= ~X264_CPU_SSE_MISALIGN;
2088 if( x264_cpu_detect() & X264_CPU_LZCNT )
2090 cpu1 &= ~X264_CPU_CACHELINE_64;
2091 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSE_LZCNT" );
2092 cpu1 &= ~X264_CPU_LZCNT;
2094 if( x264_cpu_detect() & X264_CPU_SSE3 )
2095 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE3 | X264_CPU_CACHELINE_64, "SSE3" );
2096 if( x264_cpu_detect() & X264_CPU_SSSE3 )
2098 cpu1 &= ~X264_CPU_CACHELINE_64;
2099 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSSE3, "SSSE3" );
2100 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64" );
2101 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SHUFFLE_IS_FAST, "SSSE3 FastShuffle" );
2102 cpu1 &= ~X264_CPU_SHUFFLE_IS_FAST;
2103 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSSE3 SlowCTZ" );
2104 cpu1 &= ~X264_CPU_SLOW_CTZ;
2105 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSSE3 SlowAtom" );
2106 cpu1 &= ~X264_CPU_SLOW_ATOM;
2108 if( x264_cpu_detect() & X264_CPU_SSE4 )
2110 cpu1 &= ~X264_CPU_CACHELINE_64;
2111 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE4, "SSE4" );
2113 if( x264_cpu_detect() & X264_CPU_AVX )
2114 ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX, "AVX" );
2116 if( x264_cpu_detect() & X264_CPU_ALTIVEC )
2118 fprintf( stderr, "x264: ALTIVEC against C\n" );
2119 ret = check_all_funcs( 0, X264_CPU_ALTIVEC );
2122 if( x264_cpu_detect() & X264_CPU_ARMV6 )
2123 ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV6, "ARMv6" );
2124 if( x264_cpu_detect() & X264_CPU_NEON )
2125 ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
2126 if( x264_cpu_detect() & X264_CPU_FAST_NEON_MRC )
2127 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FAST_NEON_MRC, "Fast NEON MRC" );
2132 int main(int argc, char *argv[])
2136 if( argc > 1 && !strncmp( argv[1], "--bench", 7 ) )
2138 #if !ARCH_X86 && !ARCH_X86_64 && !ARCH_PPC && !ARCH_ARM
2139 fprintf( stderr, "no --bench for your cpu until you port rdtsc\n" );
2143 if( argv[1][7] == '=' )
2145 bench_pattern = argv[1]+8;
2146 bench_pattern_len = strlen(bench_pattern);
2152 int seed = ( argc > 1 ) ? atoi(argv[1]) : x264_mdate();
2153 fprintf( stderr, "x264: using random seed %u\n", seed );
2156 buf1 = x264_malloc( 0x1e00 + 0x2000*sizeof(pixel) + 16*BENCH_ALIGNS );
2157 pbuf1 = x264_malloc( 0x1e00*sizeof(pixel) + 16*BENCH_ALIGNS );
2158 if( !buf1 || !pbuf1 )
2160 fprintf( stderr, "malloc failed, unable to initiate tests!\n" );
2163 #define INIT_POINTER_OFFSETS\
2164 buf2 = buf1 + 0xf00;\
2165 buf3 = buf2 + 0xf00;\
2166 buf4 = buf3 + 0x1000*sizeof(pixel);\
2167 pbuf2 = pbuf1 + 0xf00;\
2168 pbuf3 = (pixel*)buf3;\
2169 pbuf4 = (pixel*)buf4;
2170 INIT_POINTER_OFFSETS;
2171 for( int i = 0; i < 0x1e00; i++ )
2173 buf1[i] = rand() & 0xFF;
2174 pbuf1[i] = rand() & PIXEL_MAX;
2176 memset( buf1+0x1e00, 0, 0x2000*sizeof(pixel) );
2178 /* 16-byte alignment is guaranteed whenever it's useful, but some functions also vary in speed depending on %64 */
2180 for( int i = 0; i < BENCH_ALIGNS && !ret; i++ )
2182 INIT_POINTER_OFFSETS;
2183 ret |= x264_stack_pagealign( check_all_flags, i*16 );
2187 fprintf( stderr, "%d/%d\r", i+1, BENCH_ALIGNS );
2190 ret = check_all_flags();
2194 fprintf( stderr, "x264: at least one test has failed. Go and fix that Right Now!\n" );
2197 fprintf( stderr, "x264: All tests passed Yeah :)\n" );