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 )
424 set_func_name( "vsad" );
426 res_c = call_c( pixel_c.vsad, pbuf1, 16 );
427 res_asm = call_a( pixel_asm.vsad, pbuf1, 16 );
428 if( res_c != res_asm )
431 fprintf( stderr, "vsad: %d != %d\n", res_c, res_asm );
434 report( "pixel vsad :" );
436 #define TEST_INTRA_MBCMP( name, pred, satd, i8x8, ... ) \
437 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
439 int res_c[3], res_asm[3]; \
440 set_func_name( #name ); \
442 call_c( pixel_c.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_c ); \
443 call_a( pixel_asm.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_asm ); \
444 if( memcmp(res_c, res_asm, sizeof(res_c)) ) \
447 fprintf( stderr, #name": %d,%d,%d != %d,%d,%d [FAILED]\n", \
448 res_c[0], res_c[1], res_c[2], \
449 res_asm[0], res_asm[1], res_asm[2] ); \
453 ok = 1; used_asm = 0;
454 TEST_INTRA_MBCMP( intra_satd_x3_16x16, predict_16x16, satd[PIXEL_16x16], 0 );
455 TEST_INTRA_MBCMP( intra_satd_x3_8x8c , predict_8x8c , satd[PIXEL_8x8] , 0 );
456 TEST_INTRA_MBCMP( intra_satd_x3_4x4 , predict_4x4 , satd[PIXEL_4x4] , 0 );
457 TEST_INTRA_MBCMP( intra_sa8d_x3_8x8 , predict_8x8 , sa8d[PIXEL_8x8] , 1, edge );
458 report( "intra satd_x3 :" );
459 TEST_INTRA_MBCMP( intra_sad_x3_16x16 , predict_16x16, sad [PIXEL_16x16], 0 );
460 TEST_INTRA_MBCMP( intra_sad_x3_8x8c , predict_8x8c , sad [PIXEL_8x8] , 0 );
461 TEST_INTRA_MBCMP( intra_sad_x3_8x8 , predict_8x8 , sad [PIXEL_8x8] , 1, edge );
462 TEST_INTRA_MBCMP( intra_sad_x3_4x4 , predict_4x4 , sad [PIXEL_4x4] , 0 );
463 report( "intra sad_x3 :" );
465 ok = 1; used_asm = 0;
466 if( pixel_asm.ssd_nv12_core != pixel_ref.ssd_nv12_core )
469 set_func_name( "ssd_nv12" );
470 uint64_t res_u_c, res_v_c, res_u_a, res_v_a;
471 pixel_c.ssd_nv12_core( pbuf1, 368, pbuf2, 368, 360, 8, &res_u_c, &res_v_c );
472 pixel_asm.ssd_nv12_core( pbuf1, 368, pbuf2, 368, 360, 8, &res_u_a, &res_v_a );
473 if( res_u_c != res_u_a || res_v_c != res_v_a )
476 fprintf( stderr, "ssd_nv12: %"PRIu64",%"PRIu64" != %"PRIu64",%"PRIu64"\n",
477 res_u_c, res_v_c, res_u_a, res_v_a );
479 call_c( pixel_c.ssd_nv12_core, pbuf1, 368, pbuf2, 368, 360, 8, &res_u_c, &res_v_c );
480 call_a( pixel_asm.ssd_nv12_core, pbuf1, 368, pbuf2, 368, 360, 8, &res_u_a, &res_v_a );
482 report( "ssd_nv12 :" );
484 if( pixel_asm.ssim_4x4x2_core != pixel_ref.ssim_4x4x2_core ||
485 pixel_asm.ssim_end4 != pixel_ref.ssim_end4 )
488 ALIGNED_16( int sums[5][4] ) = {{0}};
491 res_c = x264_pixel_ssim_wxh( &pixel_c, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3 );
492 res_a = x264_pixel_ssim_wxh( &pixel_asm, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3 );
493 if( fabs( res_c - res_a ) > 1e-6 )
496 fprintf( stderr, "ssim: %.7f != %.7f [FAILED]\n", res_c, res_a );
498 set_func_name( "ssim_core" );
499 call_c2( pixel_c.ssim_4x4x2_core, pbuf1+2, 32, pbuf2+2, 32, sums );
500 call_a2( pixel_asm.ssim_4x4x2_core, pbuf1+2, 32, pbuf2+2, 32, sums );
501 set_func_name( "ssim_end" );
502 call_c2( pixel_c.ssim_end4, sums, sums, 4 );
503 call_a2( pixel_asm.ssim_end4, sums, sums, 4 );
507 ok = 1; used_asm = 0;
508 for( int i = 0; i < 32; i++ )
510 for( int i = 0; i < 100 && ok; i++ )
511 if( pixel_asm.ads[i&3] != pixel_ref.ads[i&3] )
513 ALIGNED_16( uint16_t sums[72] );
514 ALIGNED_16( int dc[4] );
515 int16_t mvs_a[32], mvs_c[32];
517 int thresh = rand() & 0x3fff;
518 set_func_name( "esa_ads" );
519 for( int j = 0; j < 72; j++ )
520 sums[j] = rand() & 0x3fff;
521 for( int j = 0; j < 4; j++ )
522 dc[j] = rand() & 0x3fff;
524 mvn_c = call_c( pixel_c.ads[i&3], dc, sums, 32, cost_mv, mvs_c, 28, thresh );
525 mvn_a = call_a( pixel_asm.ads[i&3], dc, sums, 32, cost_mv, mvs_a, 28, thresh );
526 if( mvn_c != mvn_a || memcmp( mvs_c, mvs_a, mvn_c*sizeof(*mvs_c) ) )
529 printf( "c%d: ", i&3 );
530 for( int j = 0; j < mvn_c; j++ )
531 printf( "%d ", mvs_c[j] );
532 printf( "\na%d: ", i&3 );
533 for( int j = 0; j < mvn_a; j++ )
534 printf( "%d ", mvs_a[j] );
538 report( "esa ads:" );
543 static int check_dct( int cpu_ref, int cpu_new )
545 x264_dct_function_t dct_c;
546 x264_dct_function_t dct_ref;
547 x264_dct_function_t dct_asm;
548 x264_quant_function_t qf;
549 int ret = 0, ok, used_asm, interlace;
550 ALIGNED_16( dctcoef dct1[16][16] );
551 ALIGNED_16( dctcoef dct2[16][16] );
552 ALIGNED_16( dctcoef dct4[16][16] );
553 ALIGNED_16( dctcoef dct8[4][64] );
554 ALIGNED_16( dctcoef dctdc[2][4] );
558 x264_dct_init( 0, &dct_c );
559 x264_dct_init( cpu_ref, &dct_ref);
560 x264_dct_init( cpu_new, &dct_asm );
562 memset( h, 0, sizeof(*h) );
563 h->pps = h->pps_array;
564 x264_param_default( &h->param );
565 h->chroma_qp_table = i_chroma_qp_table + 12;
566 h->param.analyse.i_luma_deadzone[0] = 0;
567 h->param.analyse.i_luma_deadzone[1] = 0;
568 h->param.analyse.b_transform_8x8 = 1;
569 for( int i = 0; i < 6; i++ )
570 h->pps->scaling_list[i] = x264_cqm_flat16;
572 x264_quant_init( h, 0, &qf );
574 /* overflow test cases */
575 for( int i = 0; i < 5; i++ )
577 pixel *enc = &pbuf3[16*i*FENC_STRIDE];
578 pixel *dec = &pbuf4[16*i*FDEC_STRIDE];
580 for( int j = 0; j < 16; j++ )
582 int cond_a = (i < 2) ? 1 : ((j&3) == 0 || (j&3) == (i-1));
583 int cond_b = (i == 0) ? 1 : !cond_a;
584 enc[0] = enc[1] = cond_a ? PIXEL_MAX : 0;
585 enc[2] = enc[3] = cond_b ? PIXEL_MAX : 0;
587 for( int k = 0; k < 4; k++ )
588 dec[k] = PIXEL_MAX - enc[k];
595 #define TEST_DCT( name, t1, t2, size ) \
596 if( dct_asm.name != dct_ref.name ) \
598 set_func_name( #name ); \
600 pixel *enc = pbuf3; \
601 pixel *dec = pbuf4; \
602 for( int j = 0; j < 5; j++) \
604 call_c( dct_c.name, t1, &pbuf1[j*64], &pbuf2[j*64] ); \
605 call_a( dct_asm.name, t2, &pbuf1[j*64], &pbuf2[j*64] ); \
606 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
609 fprintf( stderr, #name " [FAILED]\n" ); \
612 call_c( dct_c.name, t1, enc, dec ); \
613 call_a( dct_asm.name, t2, enc, dec ); \
614 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
617 fprintf( stderr, #name " [FAILED] (overflow)\n" ); \
620 enc += 16*FENC_STRIDE; \
621 dec += 16*FDEC_STRIDE; \
624 ok = 1; used_asm = 0;
625 TEST_DCT( sub4x4_dct, dct1[0], dct2[0], 16 );
626 TEST_DCT( sub8x8_dct, dct1, dct2, 16*4 );
627 TEST_DCT( sub8x8_dct_dc, dctdc[0], dctdc[1], 4 );
628 TEST_DCT( sub16x16_dct, dct1, dct2, 16*16 );
629 report( "sub_dct4 :" );
631 ok = 1; used_asm = 0;
632 TEST_DCT( sub8x8_dct8, (void*)dct1[0], (void*)dct2[0], 64 );
633 TEST_DCT( sub16x16_dct8, (void*)dct1, (void*)dct2, 64*4 );
634 report( "sub_dct8 :" );
637 // fdct and idct are denormalized by different factors, so quant/dequant
638 // is needed to force the coefs into the right range.
639 dct_c.sub16x16_dct( dct4, pbuf1, pbuf2 );
640 dct_c.sub16x16_dct8( dct8, pbuf1, pbuf2 );
641 for( int i = 0; i < 16; i++ )
643 qf.quant_4x4( dct4[i], h->quant4_mf[CQM_4IY][20], h->quant4_bias[CQM_4IY][20] );
644 qf.dequant_4x4( dct4[i], h->dequant4_mf[CQM_4IY], 20 );
646 for( int i = 0; i < 4; i++ )
648 qf.quant_8x8( dct8[i], h->quant8_mf[CQM_8IY][20], h->quant8_bias[CQM_8IY][20] );
649 qf.dequant_8x8( dct8[i], h->dequant8_mf[CQM_8IY], 20 );
651 x264_cqm_delete( h );
653 #define TEST_IDCT( name, src ) \
654 if( dct_asm.name != dct_ref.name ) \
656 set_func_name( #name ); \
658 memcpy( pbuf3, pbuf1, 32*32 * sizeof(pixel) ); \
659 memcpy( pbuf4, pbuf1, 32*32 * sizeof(pixel) ); \
660 memcpy( dct1, src, 256 * sizeof(dctcoef) ); \
661 memcpy( dct2, src, 256 * sizeof(dctcoef) ); \
662 call_c1( dct_c.name, pbuf3, (void*)dct1 ); \
663 call_a1( dct_asm.name, pbuf4, (void*)dct2 ); \
664 if( memcmp( pbuf3, pbuf4, 32*32 * sizeof(pixel) ) ) \
667 fprintf( stderr, #name " [FAILED]\n" ); \
669 call_c2( dct_c.name, pbuf3, (void*)dct1 ); \
670 call_a2( dct_asm.name, pbuf4, (void*)dct2 ); \
672 ok = 1; used_asm = 0;
673 TEST_IDCT( add4x4_idct, dct4 );
674 TEST_IDCT( add8x8_idct, dct4 );
675 TEST_IDCT( add8x8_idct_dc, dct4 );
676 TEST_IDCT( add16x16_idct, dct4 );
677 TEST_IDCT( add16x16_idct_dc, dct4 );
678 report( "add_idct4 :" );
680 ok = 1; used_asm = 0;
681 TEST_IDCT( add8x8_idct8, dct8 );
682 TEST_IDCT( add16x16_idct8, dct8 );
683 report( "add_idct8 :" );
686 #define TEST_DCTDC( name )\
687 ok = 1; used_asm = 0;\
688 if( dct_asm.name != dct_ref.name )\
690 set_func_name( #name );\
692 uint16_t *p = (uint16_t*)buf1;\
693 for( int i = 0; i < 16 && ok; i++ )\
695 for( int j = 0; j < 16; j++ )\
696 dct1[0][j] = !i ? (j^j>>1^j>>2^j>>3)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
697 : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
698 : ((*p++)&0x1fff)-0x1000; /* general case */\
699 memcpy( dct2, dct1, 16 * sizeof(dctcoef) );\
700 call_c1( dct_c.name, dct1[0] );\
701 call_a1( dct_asm.name, dct2[0] );\
702 if( memcmp( dct1, dct2, 16 * sizeof(dctcoef) ) )\
705 call_c2( dct_c.name, dct1[0] );\
706 call_a2( dct_asm.name, dct2[0] );\
708 report( #name " :" );
710 TEST_DCTDC( dct4x4dc );
711 TEST_DCTDC( idct4x4dc );
714 x264_zigzag_function_t zigzag_c;
715 x264_zigzag_function_t zigzag_ref;
716 x264_zigzag_function_t zigzag_asm;
718 ALIGNED_16( dctcoef level1[64] );
719 ALIGNED_16( dctcoef level2[64] );
721 #define TEST_ZIGZAG_SCAN( name, t1, t2, dct, size ) \
722 if( zigzag_asm.name != zigzag_ref.name ) \
724 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
726 memcpy(dct, buf1, size*sizeof(dctcoef)); \
727 call_c( zigzag_c.name, t1, dct ); \
728 call_a( zigzag_asm.name, t2, dct ); \
729 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
732 fprintf( stderr, #name " [FAILED]\n" ); \
736 #define TEST_ZIGZAG_SUB( name, t1, t2, size ) \
737 if( zigzag_asm.name != zigzag_ref.name ) \
740 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
742 memcpy( pbuf3, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
743 memcpy( pbuf4, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
744 nz_c = call_c1( zigzag_c.name, t1, pbuf2, pbuf3 ); \
745 nz_a = call_a1( zigzag_asm.name, t2, pbuf2, pbuf4 ); \
746 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( pbuf3, pbuf4, 16*FDEC_STRIDE*sizeof(pixel) ) || nz_c != nz_a ) \
749 fprintf( stderr, #name " [FAILED]\n" ); \
751 call_c2( zigzag_c.name, t1, pbuf2, pbuf3 ); \
752 call_a2( zigzag_asm.name, t2, pbuf2, pbuf4 ); \
755 #define TEST_ZIGZAG_SUBAC( name, t1, t2 ) \
756 if( zigzag_asm.name != zigzag_ref.name ) \
759 dctcoef dc_a, dc_c; \
760 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
762 for( int i = 0; i < 2; i++ ) \
764 memcpy( pbuf3, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
765 memcpy( pbuf4, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
766 for( int j = 0; j < 4; j++ ) \
768 memcpy( pbuf3 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
769 memcpy( pbuf4 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
771 nz_c = call_c1( zigzag_c.name, t1, pbuf2, pbuf3, &dc_c ); \
772 nz_a = call_a1( zigzag_asm.name, t2, pbuf2, pbuf4, &dc_a ); \
773 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 ) \
776 fprintf( stderr, #name " [FAILED]\n" ); \
780 call_c2( zigzag_c.name, t1, pbuf2, pbuf3, &dc_c ); \
781 call_a2( zigzag_asm.name, t2, pbuf2, pbuf4, &dc_a ); \
784 #define TEST_INTERLEAVE( name, t1, t2, dct, size ) \
785 if( zigzag_asm.name != zigzag_ref.name ) \
787 for( int j = 0; j < 100; j++ ) \
789 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
791 memcpy(dct, buf1, size*sizeof(dctcoef)); \
792 for( int i = 0; i < size; i++ ) \
793 dct[i] = rand()&0x1F ? 0 : dct[i]; \
794 memcpy(buf3, buf4, 10); \
795 call_c( zigzag_c.name, t1, dct, buf3 ); \
796 call_a( zigzag_asm.name, t2, dct, buf4 ); \
797 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( buf3, buf4, 10 ) ) \
805 x264_zigzag_init( 0, &zigzag_c, 0 );
806 x264_zigzag_init( cpu_ref, &zigzag_ref, 0 );
807 x264_zigzag_init( cpu_new, &zigzag_asm, 0 );
809 ok = 1; used_asm = 0;
810 TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, (void*)dct1, 64 );
811 TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 16 );
812 TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
813 TEST_ZIGZAG_SUBAC( sub_4x4ac, level1, level2 );
814 report( "zigzag_frame :" );
817 x264_zigzag_init( 0, &zigzag_c, 1 );
818 x264_zigzag_init( cpu_ref, &zigzag_ref, 1 );
819 x264_zigzag_init( cpu_new, &zigzag_asm, 1 );
821 ok = 1; used_asm = 0;
822 TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, (void*)dct1, 64 );
823 TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 16 );
824 TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
825 TEST_ZIGZAG_SUBAC( sub_4x4ac, level1, level2 );
826 report( "zigzag_field :" );
828 ok = 1; used_asm = 0;
829 TEST_INTERLEAVE( interleave_8x8_cavlc, level1, level2, dct1[0], 64 );
830 report( "zigzag_interleave :" );
831 #undef TEST_ZIGZAG_SCAN
832 #undef TEST_ZIGZAG_SUB
837 static int check_mc( int cpu_ref, int cpu_new )
839 x264_mc_functions_t mc_c;
840 x264_mc_functions_t mc_ref;
841 x264_mc_functions_t mc_a;
842 x264_pixel_function_t pixf;
844 pixel *src = &(pbuf1)[2*64+2];
845 pixel *src2[4] = { &(pbuf1)[3*64+2], &(pbuf1)[5*64+2],
846 &(pbuf1)[7*64+2], &(pbuf1)[9*64+2] };
850 int ret = 0, ok, used_asm;
852 x264_mc_init( 0, &mc_c );
853 x264_mc_init( cpu_ref, &mc_ref );
854 x264_mc_init( cpu_new, &mc_a );
855 x264_pixel_init( 0, &pixf );
857 #define MC_TEST_LUMA( w, h ) \
858 if( mc_a.mc_luma != mc_ref.mc_luma && !(w&(w-1)) && h<=16 ) \
860 const x264_weight_t *weight = weight_none; \
861 set_func_name( "mc_luma_%dx%d", w, h ); \
863 for( int i = 0; i < 1024; i++ ) \
864 pbuf3[i] = pbuf4[i] = 0xCD; \
865 call_c( mc_c.mc_luma, dst1, 32, src2, 64, dx, dy, w, h, weight ); \
866 call_a( mc_a.mc_luma, dst2, 32, src2, 64, dx, dy, w, h, weight ); \
867 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
869 fprintf( stderr, "mc_luma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
873 if( mc_a.get_ref != mc_ref.get_ref ) \
876 int ref_stride = 32; \
877 int w_checked = ( ( sizeof(pixel) == 2 && (w == 12 || w == 20)) ? w-2 : w ); \
878 const x264_weight_t *weight = weight_none; \
879 set_func_name( "get_ref_%dx%d", w_checked, h ); \
881 for( int i = 0; i < 1024; i++ ) \
882 pbuf3[i] = pbuf4[i] = 0xCD; \
883 call_c( mc_c.mc_luma, dst1, 32, src2, 64, dx, dy, w, h, weight ); \
884 ref = (pixel*)call_a( mc_a.get_ref, ref, &ref_stride, src2, 64, dx, dy, w, h, weight ); \
885 for( int i = 0; i < h; i++ ) \
886 if( memcmp( dst1+i*32, ref+i*ref_stride, w_checked * sizeof(pixel) ) ) \
888 fprintf( stderr, "get_ref[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w_checked, h ); \
894 #define MC_TEST_CHROMA( w, h ) \
895 if( mc_a.mc_chroma != mc_ref.mc_chroma ) \
897 set_func_name( "mc_chroma_%dx%d", w, h ); \
899 for( int i = 0; i < 1024; i++ ) \
900 pbuf3[i] = pbuf4[i] = 0xCD; \
901 call_c( mc_c.mc_chroma, dst1, dst1+8, 16, src, 64, dx, dy, w, h ); \
902 call_a( mc_a.mc_chroma, dst2, dst2+8, 16, src, 64, dx, dy, w, h ); \
903 /* mc_chroma width=2 may write garbage to the right of dst. ignore that. */ \
904 for( int j = 0; j < h; j++ ) \
905 for( int i = w; i < 8; i++ ) \
907 dst2[i+j*16+8] = dst1[i+j*16+8]; \
908 dst2[i+j*16] = dst1[i+j*16]; \
910 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
912 fprintf( stderr, "mc_chroma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
916 ok = 1; used_asm = 0;
917 for( int dy = -8; dy < 8; dy++ )
918 for( int dx = -128; dx < 128; dx++ )
920 if( rand()&15 ) continue; // running all of them is too slow
921 MC_TEST_LUMA( 20, 18 );
922 MC_TEST_LUMA( 16, 16 );
923 MC_TEST_LUMA( 16, 8 );
924 MC_TEST_LUMA( 12, 10 );
925 MC_TEST_LUMA( 8, 16 );
926 MC_TEST_LUMA( 8, 8 );
927 MC_TEST_LUMA( 8, 4 );
928 MC_TEST_LUMA( 4, 8 );
929 MC_TEST_LUMA( 4, 4 );
931 report( "mc luma :" );
933 ok = 1; used_asm = 0;
934 for( int dy = -1; dy < 9; dy++ )
935 for( int dx = -128; dx < 128; dx++ )
937 if( rand()&15 ) continue;
938 MC_TEST_CHROMA( 8, 8 );
939 MC_TEST_CHROMA( 8, 4 );
940 MC_TEST_CHROMA( 4, 8 );
941 MC_TEST_CHROMA( 4, 4 );
942 MC_TEST_CHROMA( 4, 2 );
943 MC_TEST_CHROMA( 2, 4 );
944 MC_TEST_CHROMA( 2, 2 );
946 report( "mc chroma :" );
948 #undef MC_TEST_CHROMA
950 #define MC_TEST_AVG( name, weight ) \
952 ok = 1, used_asm = 0; \
953 for( int i = 0; i < 10; i++ ) \
955 memcpy( pbuf3, pbuf1+320, 320 * sizeof(pixel) ); \
956 memcpy( pbuf4, pbuf1+320, 320 * sizeof(pixel) ); \
957 if( mc_a.name[i] != mc_ref.name[i] ) \
959 set_func_name( "%s_%s", #name, pixel_names[i] ); \
961 call_c1( mc_c.name[i], pbuf3, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
962 call_a1( mc_a.name[i], pbuf4, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
963 if( memcmp( pbuf3, pbuf4, 320 * sizeof(pixel) ) ) \
966 fprintf( stderr, #name "[%d]: [FAILED]\n", i ); \
968 call_c2( mc_c.name[i], pbuf3, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
969 call_a2( mc_a.name[i], pbuf4, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
974 for( int w = -63; w <= 127 && ok; w++ )
975 MC_TEST_AVG( avg, w );
976 report( "mc wpredb :" );
978 #define MC_TEST_WEIGHT( name, weight, aligned ) \
979 int align_off = (aligned ? 0 : rand()%16); \
980 ok = 1, used_asm = 0; \
981 for( int i = 1; i <= 5; i++ ) \
983 ALIGNED_16( pixel buffC[640] ); \
984 ALIGNED_16( pixel buffA[640] ); \
985 int j = X264_MAX( i*4, 2 ); \
986 memset( buffC, 0, 640 * sizeof(pixel) ); \
987 memset( buffA, 0, 640 * sizeof(pixel) ); \
990 /* w12 is the same as w16 in some cases */ \
991 if( i == 3 && mc_a.name[i] == mc_a.name[i+1] ) \
993 if( mc_a.name[i] != mc_ref.name[i] ) \
995 set_func_name( "%s_w%d", #name, j ); \
997 call_c1( mc_c.weight[i], buffC, 32, pbuf2+align_off, 32, &weight, 16 ); \
998 mc_a.weight_cache(&ha, &weight); \
999 call_a1( weight.weightfn[i], buffA, 32, pbuf2+align_off, 32, &weight, 16 ); \
1000 for( int k = 0; k < 16; k++ ) \
1001 if( memcmp( &buffC[k*32], &buffA[k*32], j * sizeof(pixel) ) ) \
1004 fprintf( stderr, #name "[%d]: [FAILED] s:%d o:%d d%d\n", i, s, o, d ); \
1007 call_c2( mc_c.weight[i], buffC, 32, pbuf2+align_off, 32, &weight, 16 ); \
1008 call_a2( weight.weightfn[i], buffA, 32, pbuf2+align_off, 32, &weight, 16 ); \
1012 ok = 1; used_asm = 0;
1015 for( int s = 0; s <= 127 && ok; s++ )
1017 for( int o = -128; o <= 127 && ok; o++ )
1019 if( rand() & 2047 ) continue;
1020 for( int d = 0; d <= 7 && ok; d++ )
1024 x264_weight_t weight = { .i_scale = s, .i_denom = d, .i_offset = o };
1025 MC_TEST_WEIGHT( weight, weight, (align_cnt++ % 4) );
1030 report( "mc weight :" );
1032 ok = 1; used_asm = 0;
1033 for( int o = 0; o <= 127 && ok; o++ )
1036 if( rand() & 15 ) continue;
1037 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1038 MC_TEST_WEIGHT( offsetadd, weight, (align_cnt++ % 4) );
1040 report( "mc offsetadd :" );
1041 ok = 1; used_asm = 0;
1042 for( int o = -128; o < 0 && ok; o++ )
1045 if( rand() & 15 ) continue;
1046 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1047 MC_TEST_WEIGHT( offsetsub, weight, (align_cnt++ % 4) );
1049 report( "mc offsetsub :" );
1051 ok = 1; used_asm = 0;
1052 if( mc_a.store_interleave_8x8x2 != mc_ref.store_interleave_8x8x2 )
1054 set_func_name( "store_interleave_8x8x2" );
1056 memset( pbuf3, 0, 64*8 );
1057 memset( pbuf4, 0, 64*8 );
1058 call_c( mc_c.store_interleave_8x8x2, pbuf3, 64, pbuf1, pbuf1+16 );
1059 call_a( mc_a.store_interleave_8x8x2, pbuf4, 64, pbuf1, pbuf1+16 );
1060 if( memcmp( pbuf3, pbuf4, 64*8 ) )
1063 if( mc_a.load_deinterleave_8x8x2_fenc != mc_ref.load_deinterleave_8x8x2_fenc )
1065 set_func_name( "load_deinterleave_8x8x2_fenc" );
1067 call_c( mc_c.load_deinterleave_8x8x2_fenc, pbuf3, pbuf1, 64 );
1068 call_a( mc_a.load_deinterleave_8x8x2_fenc, pbuf4, pbuf1, 64 );
1069 if( memcmp( pbuf3, pbuf4, FENC_STRIDE*8 ) )
1072 if( mc_a.load_deinterleave_8x8x2_fdec != mc_ref.load_deinterleave_8x8x2_fdec )
1074 set_func_name( "load_deinterleave_8x8x2_fdec" );
1076 call_c( mc_c.load_deinterleave_8x8x2_fdec, pbuf3, pbuf1, 64 );
1077 call_a( mc_a.load_deinterleave_8x8x2_fdec, pbuf4, pbuf1, 64 );
1078 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*8 ) )
1081 report( "store_interleave :" );
1084 int w, h, src_stride;
1085 } 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} };
1086 ok = 1; used_asm = 0;
1087 if( mc_a.plane_copy != mc_ref.plane_copy )
1089 set_func_name( "plane_copy" );
1091 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1093 int w = plane_specs[i].w;
1094 int h = plane_specs[i].h;
1095 int src_stride = plane_specs[i].src_stride;
1096 int dst_stride = (w + 127) & ~63;
1097 assert( dst_stride * h <= 0x1000 );
1098 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1099 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1100 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1101 call_c( mc_c.plane_copy, pbuf3, dst_stride, src1, src_stride, w, h );
1102 call_a( mc_a.plane_copy, pbuf4, dst_stride, src1, src_stride, w, h );
1103 for( int y = 0; y < h; y++ )
1104 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(pixel) ) )
1107 fprintf( stderr, "plane_copy FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
1113 if( mc_a.plane_copy_interleave != mc_ref.plane_copy_interleave )
1115 set_func_name( "plane_copy_interleave" );
1117 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1119 int w = (plane_specs[i].w + 1) >> 1;
1120 int h = plane_specs[i].h;
1121 int src_stride = (plane_specs[i].src_stride + 1) >> 1;
1122 int dst_stride = (2*w + 127) & ~63;
1123 assert( dst_stride * h <= 0x1000 );
1124 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1125 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1126 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1127 call_c( mc_c.plane_copy_interleave, pbuf3, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1128 call_a( mc_a.plane_copy_interleave, pbuf4, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1129 for( int y = 0; y < h; y++ )
1130 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
1133 fprintf( stderr, "plane_copy_interleave FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
1139 if( mc_a.plane_copy_deinterleave != mc_ref.plane_copy_deinterleave )
1141 set_func_name( "plane_copy_deinterleave" );
1143 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1145 int w = (plane_specs[i].w + 1) >> 1;
1146 int h = plane_specs[i].h;
1148 int src_stride = (2*w + 127) & ~63;
1149 int offv = (dst_stride*h + 31) & ~15;
1150 memset( pbuf3, 0, 0x1000 );
1151 memset( pbuf4, 0, 0x1000 );
1152 call_c( mc_c.plane_copy_deinterleave, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf1, src_stride, w, h );
1153 call_a( mc_a.plane_copy_deinterleave, pbuf4, dst_stride, pbuf4+offv, dst_stride, pbuf1, src_stride, w, h );
1154 for( int y = 0; y < h; y++ )
1155 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w ) ||
1156 memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w ) )
1159 fprintf( stderr, "plane_copy_deinterleave FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
1164 report( "plane_copy :" );
1166 if( mc_a.hpel_filter != mc_ref.hpel_filter )
1168 pixel *srchpel = pbuf1+8+2*64;
1169 pixel *dstc[3] = { pbuf3+8, pbuf3+8+16*64, pbuf3+8+32*64 };
1170 pixel *dsta[3] = { pbuf4+8, pbuf4+8+16*64, pbuf4+8+32*64 };
1171 void *tmp = pbuf3+49*64;
1172 set_func_name( "hpel_filter" );
1173 ok = 1; used_asm = 1;
1174 memset( pbuf3, 0, 4096 * sizeof(pixel) );
1175 memset( pbuf4, 0, 4096 * sizeof(pixel) );
1176 call_c( mc_c.hpel_filter, dstc[0], dstc[1], dstc[2], srchpel, 64, 48, 10, tmp );
1177 call_a( mc_a.hpel_filter, dsta[0], dsta[1], dsta[2], srchpel, 64, 48, 10, tmp );
1178 for( int i = 0; i < 3; i++ )
1179 for( int j = 0; j < 10; j++ )
1180 //FIXME ideally the first pixels would match too, but they aren't actually used
1181 if( memcmp( dstc[i]+j*64+2, dsta[i]+j*64+2, 43 * sizeof(pixel) ) )
1184 fprintf( stderr, "hpel filter differs at plane %c line %d\n", "hvc"[i], j );
1185 for( int k = 0; k < 48; k++ )
1186 printf( "%02x%s", dstc[i][j*64+k], (k+1)&3 ? "" : " " );
1188 for( int k = 0; k < 48; k++ )
1189 printf( "%02x%s", dsta[i][j*64+k], (k+1)&3 ? "" : " " );
1193 report( "hpel filter :" );
1196 if( mc_a.frame_init_lowres_core != mc_ref.frame_init_lowres_core )
1198 pixel *dstc[4] = { pbuf3, pbuf3+1024, pbuf3+2048, pbuf3+3072 };
1199 pixel *dsta[4] = { pbuf4, pbuf4+1024, pbuf4+2048, pbuf4+3072 };
1200 set_func_name( "lowres_init" );
1201 ok = 1; used_asm = 1;
1202 for( int w = 40; w <= 48; w += 8 )
1204 int stride = (w+8)&~15;
1205 call_c( mc_c.frame_init_lowres_core, pbuf1, dstc[0], dstc[1], dstc[2], dstc[3], w*2, stride, w, 16 );
1206 call_a( mc_a.frame_init_lowres_core, pbuf1, dsta[0], dsta[1], dsta[2], dsta[3], w*2, stride, w, 16 );
1207 for( int i = 0; i < 16; i++ )
1209 for( int j = 0; j < 4; j++ )
1210 if( memcmp( dstc[j]+i*stride, dsta[j]+i*stride, w * sizeof(pixel) ) )
1213 fprintf( stderr, "frame_init_lowres differs at plane %d line %d\n", j, i );
1214 for( int k = 0; k < w; k++ )
1215 printf( "%d ", dstc[j][k+i*stride] );
1217 for( int k = 0; k < w; k++ )
1218 printf( "%d ", dsta[j][k+i*stride] );
1224 report( "lowres init :" );
1227 #define INTEGRAL_INIT( name, size, ... )\
1228 if( mc_a.name != mc_ref.name )\
1231 set_func_name( #name );\
1233 memcpy( buf3, buf1, size*2*stride );\
1234 memcpy( buf4, buf1, size*2*stride );\
1235 uint16_t *sum = (uint16_t*)buf3;\
1236 call_c1( mc_c.name, __VA_ARGS__ );\
1237 sum = (uint16_t*)buf4;\
1238 call_a1( mc_a.name, __VA_ARGS__ );\
1239 if( memcmp( buf3, buf4, (stride-8)*2 ) \
1240 || (size>9 && memcmp( buf3+18*stride, buf4+18*stride, (stride-8)*2 )))\
1242 call_c2( mc_c.name, __VA_ARGS__ );\
1243 call_a2( mc_a.name, __VA_ARGS__ );\
1245 ok = 1; used_asm = 0;
1246 INTEGRAL_INIT( integral_init4h, 2, sum+stride, pbuf2, stride );
1247 INTEGRAL_INIT( integral_init8h, 2, sum+stride, pbuf2, stride );
1248 INTEGRAL_INIT( integral_init4v, 14, sum, sum+9*stride, stride );
1249 INTEGRAL_INIT( integral_init8v, 9, sum, stride );
1250 report( "integral init :" );
1252 if( mc_a.mbtree_propagate_cost != mc_ref.mbtree_propagate_cost )
1255 for( int i = 0; i < 10; i++ )
1257 float fps_factor = (rand()&65535) / 256.;
1258 ok = 1; used_asm = 1;
1259 set_func_name( "mbtree_propagate" );
1260 int *dsta = (int*)buf3;
1261 int *dstc = dsta+400;
1262 uint16_t *prop = (uint16_t*)buf1;
1263 uint16_t *intra = (uint16_t*)buf4;
1264 uint16_t *inter = intra+100;
1265 uint16_t *qscale = inter+100;
1266 uint16_t *rnd = (uint16_t*)buf2;
1268 for( int j = 0; j < 100; j++ )
1270 intra[j] = *rnd++ & 0x7fff;
1271 intra[j] += !intra[j];
1272 inter[j] = *rnd++ & 0x7fff;
1273 qscale[j] = *rnd++ & 0x7fff;
1275 call_c( mc_c.mbtree_propagate_cost, dstc, prop, intra, inter, qscale, &fps_factor, 100 );
1276 call_a( mc_a.mbtree_propagate_cost, dsta, prop, intra, inter, qscale, &fps_factor, 100 );
1277 // I don't care about exact rounding, this is just how close the floating-point implementation happens to be
1279 for( int j = 0; j < 100; j++ )
1280 ok &= abs( dstc[j]-dsta[j] ) <= 1 || fabs( (double)dstc[j]/dsta[j]-1 ) < 1e-4;
1282 report( "mbtree propagate :" );
1288 static int check_deblock( int cpu_ref, int cpu_new )
1290 x264_deblock_function_t db_c;
1291 x264_deblock_function_t db_ref;
1292 x264_deblock_function_t db_a;
1293 int ret = 0, ok = 1, used_asm = 0;
1294 int alphas[36], betas[36];
1297 x264_deblock_init( 0, &db_c, 0 );
1298 x264_deblock_init( cpu_ref, &db_ref, 0 );
1299 x264_deblock_init( cpu_new, &db_a, 0 );
1301 /* not exactly the real values of a,b,tc but close enough */
1302 for( int i = 35, a = 255, c = 250; i >= 0; i-- )
1304 alphas[i] = a << (BIT_DEPTH-8);
1305 betas[i] = (i+1)/2 << (BIT_DEPTH-8);
1306 tcs[i][0] = tcs[i][3] = (c+6)/10 << (BIT_DEPTH-8);
1307 tcs[i][1] = (c+7)/15 << (BIT_DEPTH-8);
1308 tcs[i][2] = (c+9)/20 << (BIT_DEPTH-8);
1313 #define TEST_DEBLOCK( name, align, ... ) \
1314 for( int i = 0; i < 36; i++ ) \
1316 int off = 8*32 + (i&15)*4*!align; /* benchmark various alignments of h filter */ \
1317 for( int j = 0; j < 1024; j++ ) \
1318 /* two distributions of random to excersize different failure modes */ \
1319 pbuf3[j] = rand() & (i&1 ? 0xf : PIXEL_MAX ); \
1320 memcpy( pbuf4, pbuf3, 1024 * sizeof(pixel) ); \
1321 if( db_a.name != db_ref.name ) \
1323 set_func_name( #name ); \
1325 call_c1( db_c.name, pbuf3+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1326 call_a1( db_a.name, pbuf4+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1327 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1330 fprintf( stderr, #name "(a=%d, b=%d): [FAILED]\n", alphas[i], betas[i] ); \
1333 call_c2( db_c.name, pbuf3+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1334 call_a2( db_a.name, pbuf4+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1338 TEST_DEBLOCK( deblock_luma[0], 0, tcs[i] );
1339 TEST_DEBLOCK( deblock_luma[1], 1, tcs[i] );
1340 TEST_DEBLOCK( deblock_chroma[0], 0, tcs[i] );
1341 TEST_DEBLOCK( deblock_chroma[1], 1, tcs[i] );
1342 TEST_DEBLOCK( deblock_luma_intra[0], 0 );
1343 TEST_DEBLOCK( deblock_luma_intra[1], 1 );
1344 TEST_DEBLOCK( deblock_chroma_intra[0], 0 );
1345 TEST_DEBLOCK( deblock_chroma_intra[1], 1 );
1347 if( db_a.deblock_strength != db_ref.deblock_strength )
1349 for( int i = 0; i < 100; i++ )
1351 ALIGNED_ARRAY_16( uint8_t, nnz, [X264_SCAN8_SIZE] );
1352 ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );
1353 ALIGNED_ARRAY_16( int16_t, mv, [2],[X264_SCAN8_LUMA_SIZE][2] );
1354 ALIGNED_ARRAY_16( uint8_t, bs, [2],[2][8][4] );
1355 memset( bs, 99, sizeof(bs) );
1356 for( int j = 0; j < X264_SCAN8_SIZE; j++ )
1357 nnz[j] = ((rand()&7) == 7) * rand() & 0xf;
1358 for( int j = 0; j < 2; j++ )
1359 for( int k = 0; k < X264_SCAN8_LUMA_SIZE; k++ )
1361 ref[j][k] = ((rand()&3) != 3) ? 0 : (rand() & 31) - 2;
1362 for( int l = 0; l < 2; l++ )
1363 mv[j][k][l] = ((rand()&7) != 7) ? (rand()&7) - 3 : (rand()&1023) - 512;
1365 set_func_name( "deblock_strength" );
1366 call_c( db_c.deblock_strength, nnz, ref, mv, bs[0], 2<<(i&1), ((i>>1)&1), NULL );
1367 call_a( db_a.deblock_strength, nnz, ref, mv, bs[1], 2<<(i&1), ((i>>1)&1), NULL );
1368 if( memcmp( bs[0], bs[1], sizeof(bs[0]) ) )
1371 fprintf( stderr, "deblock_strength: [FAILED]\n" );
1372 for( int j = 0; j < 2; j++ )
1374 for( int k = 0; k < 2; k++ )
1375 for( int l = 0; l < 4; l++ )
1377 for( int m = 0; m < 4; m++ )
1378 printf("%d ",bs[j][k][l][m]);
1388 report( "deblock :" );
1393 static int check_quant( int cpu_ref, int cpu_new )
1395 x264_quant_function_t qf_c;
1396 x264_quant_function_t qf_ref;
1397 x264_quant_function_t qf_a;
1398 ALIGNED_16( dctcoef dct1[64] );
1399 ALIGNED_16( dctcoef dct2[64] );
1400 ALIGNED_16( uint8_t cqm_buf[64] );
1401 int ret = 0, ok, used_asm;
1402 int oks[3] = {1,1,1}, used_asms[3] = {0,0,0};
1405 memset( h, 0, sizeof(*h) );
1406 h->pps = h->pps_array;
1407 x264_param_default( &h->param );
1408 h->chroma_qp_table = i_chroma_qp_table + 12;
1409 h->param.analyse.b_transform_8x8 = 1;
1411 for( int i_cqm = 0; i_cqm < 4; i_cqm++ )
1415 for( int i = 0; i < 6; i++ )
1416 h->pps->scaling_list[i] = x264_cqm_flat16;
1417 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_FLAT;
1419 else if( i_cqm == 1 )
1421 for( int i = 0; i < 6; i++ )
1422 h->pps->scaling_list[i] = x264_cqm_jvt[i];
1423 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_JVT;
1427 int max_scale = BIT_DEPTH < 10 ? 255 : 228;
1429 for( int i = 0; i < 64; i++ )
1430 cqm_buf[i] = 10 + rand() % (max_scale - 9);
1432 for( int i = 0; i < 64; i++ )
1434 for( int i = 0; i < 6; i++ )
1435 h->pps->scaling_list[i] = cqm_buf;
1436 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_CUSTOM;
1439 h->param.rc.i_qp_min = 0;
1440 h->param.rc.i_qp_max = QP_MAX;
1442 x264_quant_init( h, 0, &qf_c );
1443 x264_quant_init( h, cpu_ref, &qf_ref );
1444 x264_quant_init( h, cpu_new, &qf_a );
1446 #define INIT_QUANT8(j) \
1448 static const int scale1d[8] = {32,31,24,31,32,31,24,31}; \
1449 for( int i = 0; i < 64; i++ ) \
1451 unsigned int scale = (255*scale1d[i>>3]*scale1d[i&7])/16; \
1452 dct1[i] = dct2[i] = j ? (rand()%(2*scale+1))-scale : 0; \
1456 #define INIT_QUANT4(j) \
1458 static const int scale1d[4] = {4,6,4,6}; \
1459 for( int i = 0; i < 16; i++ ) \
1461 unsigned int scale = 255*scale1d[i>>2]*scale1d[i&3]; \
1462 dct1[i] = dct2[i] = j ? (rand()%(2*scale+1))-scale : 0; \
1466 #define TEST_QUANT_DC( name, cqm ) \
1467 if( qf_a.name != qf_ref.name ) \
1469 set_func_name( #name ); \
1471 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1473 for( int j = 0; j < 2; j++ ) \
1475 int result_c, result_a; \
1476 for( int i = 0; i < 16; i++ ) \
1477 dct1[i] = dct2[i] = j ? (rand() & 0x1fff) - 0xfff : 0; \
1478 result_c = call_c1( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1479 result_a = call_a1( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1480 if( memcmp( dct1, dct2, 16*sizeof(dctcoef) ) || result_c != result_a ) \
1483 fprintf( stderr, #name "(cqm=%d): [FAILED]\n", i_cqm ); \
1486 call_c2( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1487 call_a2( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1492 #define TEST_QUANT( qname, block, w ) \
1493 if( qf_a.qname != qf_ref.qname ) \
1495 set_func_name( #qname ); \
1497 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1499 for( int j = 0; j < 2; j++ ) \
1502 int result_c = call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1503 int result_a = call_a1( qf_a.qname, dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1504 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) || result_c != result_a ) \
1507 fprintf( stderr, #qname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1510 call_c2( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1511 call_a2( qf_a.qname, dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1516 TEST_QUANT( quant_8x8, CQM_8IY, 8 );
1517 TEST_QUANT( quant_8x8, CQM_8PY, 8 );
1518 TEST_QUANT( quant_4x4, CQM_4IY, 4 );
1519 TEST_QUANT( quant_4x4, CQM_4PY, 4 );
1520 TEST_QUANT_DC( quant_4x4_dc, **h->quant4_mf[CQM_4IY] );
1521 TEST_QUANT_DC( quant_2x2_dc, **h->quant4_mf[CQM_4IC] );
1523 #define TEST_DEQUANT( qname, dqname, block, w ) \
1524 if( qf_a.dqname != qf_ref.dqname ) \
1526 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
1528 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1531 call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1532 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1533 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1534 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1535 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1538 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1541 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1542 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1546 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8IY, 8 );
1547 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8PY, 8 );
1548 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4IY, 4 );
1549 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4PY, 4 );
1551 #define TEST_DEQUANT_DC( qname, dqname, block, w ) \
1552 if( qf_a.dqname != qf_ref.dqname ) \
1554 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
1556 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1558 for( int i = 0; i < 16; i++ ) \
1559 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16; \
1560 call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp][0]>>1, h->quant##w##_bias[block][qp][0]>>1 ); \
1561 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1562 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1563 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1564 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1567 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1569 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1570 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1574 TEST_DEQUANT_DC( quant_4x4_dc, dequant_4x4_dc, CQM_4IY, 4 );
1576 #define TEST_OPTIMIZE_CHROMA_DC( qname, optname, w ) \
1577 if( qf_a.optname != qf_ref.optname ) \
1579 set_func_name( #optname ); \
1581 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1583 int dmf = h->dequant4_mf[CQM_4IC][qp%6][0] << qp/6; \
1586 for( int i = 16; ; i <<= 1 )\
1588 int res_c, res_asm; \
1589 int max = X264_MIN( i, PIXEL_MAX*16 ); \
1590 for( int j = 0; j < w*w; j++ ) \
1591 dct1[j] = rand()%(max*2+1) - max; \
1592 call_c1( qf_c.qname, dct1, h->quant4_mf[CQM_4IC][qp][0]>>1, h->quant4_bias[CQM_4IC][qp][0]>>1 ); \
1593 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1594 res_c = call_c1( qf_c.optname, dct1, dmf ); \
1595 res_asm = call_a1( qf_a.optname, dct2, dmf ); \
1596 if( res_c != res_asm || memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1599 fprintf( stderr, #optname "(qp=%d, res_c=%d, res_asm=%d): [FAILED]\n", qp, res_c, res_asm ); \
1601 call_c2( qf_c.optname, dct1, dmf ); \
1602 call_a2( qf_a.optname, dct2, dmf ); \
1603 if( i >= PIXEL_MAX*16 ) \
1609 TEST_OPTIMIZE_CHROMA_DC( quant_2x2_dc, optimize_chroma_dc, 2 );
1611 x264_cqm_delete( h );
1614 ok = oks[0]; used_asm = used_asms[0];
1615 report( "quant :" );
1617 ok = oks[1]; used_asm = used_asms[1];
1618 report( "dequant :" );
1620 ok = oks[2]; used_asm = used_asms[2];
1621 report( "optimize chroma dc :" );
1623 ok = 1; used_asm = 0;
1624 if( qf_a.denoise_dct != qf_ref.denoise_dct )
1627 for( int size = 16; size <= 64; size += 48 )
1629 set_func_name( "denoise_dct" );
1630 memcpy( dct1, buf1, size*sizeof(dctcoef) );
1631 memcpy( dct2, buf1, size*sizeof(dctcoef) );
1632 memcpy( buf3+256, buf3, 256 );
1633 call_c1( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
1634 call_a1( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
1635 if( memcmp( dct1, dct2, size*sizeof(dctcoef) ) || memcmp( buf3+4, buf3+256+4, (size-1)*sizeof(uint32_t) ) )
1637 call_c2( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
1638 call_a2( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
1641 report( "denoise dct :" );
1643 #define TEST_DECIMATE( decname, w, ac, thresh ) \
1644 if( qf_a.decname != qf_ref.decname ) \
1646 set_func_name( #decname ); \
1648 for( int i = 0; i < 100; i++ ) \
1650 static const int distrib[16] = {1,1,1,1,1,1,1,1,1,1,1,1,2,3,4};\
1651 static const int zerorate_lut[4] = {3,7,15,31};\
1652 int zero_rate = zerorate_lut[i&3];\
1653 for( int idx = 0; idx < w*w; idx++ ) \
1655 int sign = (rand()&1) ? -1 : 1; \
1656 int abs_level = distrib[rand()&15]; \
1657 if( abs_level == 4 ) abs_level = rand()&0x3fff; \
1658 int zero = !(rand()&zero_rate); \
1659 dct1[idx] = zero * abs_level * sign; \
1663 int result_c = call_c( qf_c.decname, dct1 ); \
1664 int result_a = call_a( qf_a.decname, dct1 ); \
1665 if( X264_MIN(result_c,thresh) != X264_MIN(result_a,thresh) ) \
1668 fprintf( stderr, #decname ": [FAILED]\n" ); \
1674 ok = 1; used_asm = 0;
1675 TEST_DECIMATE( decimate_score64, 8, 0, 6 );
1676 TEST_DECIMATE( decimate_score16, 4, 0, 6 );
1677 TEST_DECIMATE( decimate_score15, 4, 1, 7 );
1678 report( "decimate_score :" );
1680 #define TEST_LAST( last, lastname, w, ac ) \
1681 if( qf_a.last != qf_ref.last ) \
1683 set_func_name( #lastname ); \
1685 for( int i = 0; i < 100; i++ ) \
1688 int max = rand() & (w*w-1); \
1689 memset( dct1, 0, w*w*sizeof(dctcoef) ); \
1690 for( int idx = ac; idx < max; idx++ ) \
1691 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
1694 int result_c = call_c( qf_c.last, dct1+ac ); \
1695 int result_a = call_a( qf_a.last, dct1+ac ); \
1696 if( result_c != result_a ) \
1699 fprintf( stderr, #lastname ": [FAILED]\n" ); \
1705 ok = 1; used_asm = 0;
1706 TEST_LAST( coeff_last[DCT_CHROMA_DC], coeff_last4, 2, 0 );
1707 TEST_LAST( coeff_last[ DCT_LUMA_AC], coeff_last15, 4, 1 );
1708 TEST_LAST( coeff_last[ DCT_LUMA_4x4], coeff_last16, 4, 0 );
1709 TEST_LAST( coeff_last[ DCT_LUMA_8x8], coeff_last64, 8, 0 );
1710 report( "coeff_last :" );
1712 #define TEST_LEVELRUN( lastname, name, w, ac ) \
1713 if( qf_a.lastname != qf_ref.lastname ) \
1715 set_func_name( #name ); \
1717 for( int i = 0; i < 100; i++ ) \
1719 x264_run_level_t runlevel_c, runlevel_a; \
1721 int max = rand() & (w*w-1); \
1722 memset( dct1, 0, w*w*sizeof(dctcoef) ); \
1723 memcpy( &runlevel_a, buf1+i, sizeof(x264_run_level_t) ); \
1724 memcpy( &runlevel_c, buf1+i, sizeof(x264_run_level_t) ); \
1725 for( int idx = ac; idx < max; idx++ ) \
1726 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
1729 int result_c = call_c( qf_c.lastname, dct1+ac, &runlevel_c ); \
1730 int result_a = call_a( qf_a.lastname, dct1+ac, &runlevel_a ); \
1731 if( result_c != result_a || runlevel_c.last != runlevel_a.last || \
1732 memcmp(runlevel_c.level, runlevel_a.level, sizeof(dctcoef)*result_c) || \
1733 memcmp(runlevel_c.run, runlevel_a.run, sizeof(uint8_t)*(result_c-1)) ) \
1736 fprintf( stderr, #name ": [FAILED]\n" ); \
1742 ok = 1; used_asm = 0;
1743 TEST_LEVELRUN( coeff_level_run[DCT_CHROMA_DC], coeff_level_run4, 2, 0 );
1744 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_AC], coeff_level_run15, 4, 1 );
1745 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_4x4], coeff_level_run16, 4, 0 );
1746 report( "coeff_level_run :" );
1751 static int check_intra( int cpu_ref, int cpu_new )
1753 int ret = 0, ok = 1, used_asm = 0;
1754 ALIGNED_16( pixel edge[33] );
1755 ALIGNED_16( pixel edge2[33] );
1756 ALIGNED_16( pixel fdec[FDEC_STRIDE*20] );
1759 x264_predict_t predict_16x16[4+3];
1760 x264_predict_t predict_8x8c[4+3];
1761 x264_predict8x8_t predict_8x8[9+3];
1762 x264_predict_t predict_4x4[9+3];
1763 x264_predict_8x8_filter_t predict_8x8_filter;
1764 } ip_c, ip_ref, ip_a;
1766 x264_predict_16x16_init( 0, ip_c.predict_16x16 );
1767 x264_predict_8x8c_init( 0, ip_c.predict_8x8c );
1768 x264_predict_8x8_init( 0, ip_c.predict_8x8, &ip_c.predict_8x8_filter );
1769 x264_predict_4x4_init( 0, ip_c.predict_4x4 );
1771 x264_predict_16x16_init( cpu_ref, ip_ref.predict_16x16 );
1772 x264_predict_8x8c_init( cpu_ref, ip_ref.predict_8x8c );
1773 x264_predict_8x8_init( cpu_ref, ip_ref.predict_8x8, &ip_ref.predict_8x8_filter );
1774 x264_predict_4x4_init( cpu_ref, ip_ref.predict_4x4 );
1776 x264_predict_16x16_init( cpu_new, ip_a.predict_16x16 );
1777 x264_predict_8x8c_init( cpu_new, ip_a.predict_8x8c );
1778 x264_predict_8x8_init( cpu_new, ip_a.predict_8x8, &ip_a.predict_8x8_filter );
1779 x264_predict_4x4_init( cpu_new, ip_a.predict_4x4 );
1781 memcpy( fdec, pbuf1, 32*20 * sizeof(pixel) );\
1783 ip_c.predict_8x8_filter( fdec+48, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
1785 #define INTRA_TEST( name, dir, w, bench, ... )\
1786 if( ip_a.name[dir] != ip_ref.name[dir] )\
1788 set_func_name( "intra_%s_%s", #name, intra_##name##_names[dir] );\
1790 memcpy( pbuf3, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
1791 memcpy( pbuf4, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
1792 call_c##bench( ip_c.name[dir], pbuf3+48, ##__VA_ARGS__ );\
1793 call_a##bench( ip_a.name[dir], pbuf4+48, ##__VA_ARGS__ );\
1794 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*20 * sizeof(pixel) ) )\
1796 fprintf( stderr, #name "[%d] : [FAILED]\n", dir );\
1798 for( int k = -1; k < 16; k++ )\
1799 printf( "%2x ", edge[16+k] );\
1801 for( int j = 0; j < w; j++ )\
1803 printf( "%2x ", edge[14-j] );\
1804 for( int k = 0; k < w; k++ )\
1805 printf( "%2x ", pbuf4[48+k+j*FDEC_STRIDE] );\
1809 for( int j = 0; j < w; j++ )\
1812 for( int k = 0; k < w; k++ )\
1813 printf( "%2x ", pbuf3[48+k+j*FDEC_STRIDE] );\
1819 for( int i = 0; i < 12; i++ )
1820 INTRA_TEST( predict_4x4, i, 4, );
1821 for( int i = 0; i < 7; i++ )
1822 INTRA_TEST( predict_8x8c, i, 8, );
1823 for( int i = 0; i < 7; i++ )
1824 INTRA_TEST( predict_16x16, i, 16, );
1825 for( int i = 0; i < 12; i++ )
1826 INTRA_TEST( predict_8x8, i, 8, , edge );
1828 set_func_name("intra_predict_8x8_filter");
1829 if( ip_a.predict_8x8_filter != ip_ref.predict_8x8_filter )
1832 for( int i = 0; i < 32; i++ )
1834 memcpy( edge2, edge, 33 * sizeof(pixel) );
1835 call_c(ip_c.predict_8x8_filter, pbuf1+48, edge, (i&24)>>1, i&7);
1836 call_a(ip_a.predict_8x8_filter, pbuf1+48, edge2, (i&24)>>1, i&7);
1837 if( memcmp( edge, edge2, 33 * sizeof(pixel) ) )
1839 fprintf( stderr, "predict_8x8_filter : [FAILED] %d %d\n", (i&24)>>1, i&7);
1845 #define EXTREMAL_PLANE(size) \
1848 for( int j = 0; j < 7; j++ ) \
1849 max[j] = test ? rand()&PIXEL_MAX : PIXEL_MAX; \
1850 fdec[48-1-FDEC_STRIDE] = (i&1)*max[0]; \
1851 for( int j = 0; j < size/2; j++ ) \
1852 fdec[48+j-FDEC_STRIDE] = (!!(i&2))*max[1]; \
1853 for( int j = size/2; j < size-1; j++ ) \
1854 fdec[48+j-FDEC_STRIDE] = (!!(i&4))*max[2]; \
1855 fdec[48+(size-1)-FDEC_STRIDE] = (!!(i&8))*max[3]; \
1856 for( int j = 0; j < size/2; j++ ) \
1857 fdec[48+j*FDEC_STRIDE-1] = (!!(i&16))*max[4]; \
1858 for( int j = size/2; j < size-1; j++ ) \
1859 fdec[48+j*FDEC_STRIDE-1] = (!!(i&32))*max[5]; \
1860 fdec[48+(size-1)*FDEC_STRIDE-1] = (!!(i&64))*max[6]; \
1862 /* Extremal test case for planar prediction. */
1863 for( int test = 0; test < 100 && ok; test++ )
1864 for( int i = 0; i < 128 && ok; i++ )
1866 EXTREMAL_PLANE( 8 );
1867 INTRA_TEST( predict_8x8c, I_PRED_CHROMA_P, 8, 1 );
1868 EXTREMAL_PLANE( 16 );
1869 INTRA_TEST( predict_16x16, I_PRED_16x16_P, 16, 1 );
1871 report( "intra pred :" );
1875 #define DECL_CABAC(cpu) \
1876 static void run_cabac_decision_##cpu( uint8_t *dst )\
1879 x264_cabac_context_init( &cb, SLICE_TYPE_P, 26, 0 );\
1880 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
1881 for( int i = 0; i < 0x1000; i++ )\
1882 x264_cabac_encode_decision_##cpu( &cb, buf1[i]>>1, buf1[i]&1 );\
1884 static void run_cabac_bypass_##cpu( uint8_t *dst )\
1887 x264_cabac_context_init( &cb, SLICE_TYPE_P, 26, 0 );\
1888 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
1889 for( int i = 0; i < 0x1000; i++ )\
1890 x264_cabac_encode_bypass_##cpu( &cb, buf1[i]&1 );\
1892 static void run_cabac_terminal_##cpu( uint8_t *dst )\
1895 x264_cabac_context_init( &cb, SLICE_TYPE_P, 26, 0 );\
1896 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
1897 for( int i = 0; i < 0x1000; i++ )\
1898 x264_cabac_encode_terminal_##cpu( &cb );\
1904 #define run_cabac_decision_asm run_cabac_decision_c
1905 #define run_cabac_bypass_asm run_cabac_bypass_c
1906 #define run_cabac_terminal_asm run_cabac_terminal_c
1909 static int check_cabac( int cpu_ref, int cpu_new )
1911 int ret = 0, ok, used_asm = 1;
1912 if( cpu_ref || run_cabac_decision_c == run_cabac_decision_asm )
1916 set_func_name( "cabac_encode_decision" );
1917 memcpy( buf4, buf3, 0x1000 );
1918 call_c( run_cabac_decision_c, buf3 );
1919 call_a( run_cabac_decision_asm, buf4 );
1920 ok = !memcmp( buf3, buf4, 0x1000 );
1921 report( "cabac decision:" );
1923 set_func_name( "cabac_encode_bypass" );
1924 memcpy( buf4, buf3, 0x1000 );
1925 call_c( run_cabac_bypass_c, buf3 );
1926 call_a( run_cabac_bypass_asm, buf4 );
1927 ok = !memcmp( buf3, buf4, 0x1000 );
1928 report( "cabac bypass:" );
1930 set_func_name( "cabac_encode_terminal" );
1931 memcpy( buf4, buf3, 0x1000 );
1932 call_c( run_cabac_terminal_c, buf3 );
1933 call_a( run_cabac_terminal_asm, buf4 );
1934 ok = !memcmp( buf3, buf4, 0x1000 );
1935 report( "cabac terminal:" );
1940 static int check_bitstream( int cpu_ref, int cpu_new )
1942 x264_bitstream_function_t bs_c;
1943 x264_bitstream_function_t bs_ref;
1944 x264_bitstream_function_t bs_a;
1946 int ret = 0, ok = 1, used_asm = 0;
1948 x264_bitstream_init( 0, &bs_c );
1949 x264_bitstream_init( cpu_ref, &bs_ref );
1950 x264_bitstream_init( cpu_new, &bs_a );
1951 if( bs_a.nal_escape != bs_ref.nal_escape )
1954 uint8_t *input = malloc(size+100);
1955 uint8_t *output1 = malloc(size*2);
1956 uint8_t *output2 = malloc(size*2);
1958 set_func_name( "nal_escape" );
1959 for( int i = 0; i < 100; i++ )
1961 /* Test corner-case sizes */
1962 int test_size = i < 10 ? i+1 : rand() & 0x3fff;
1963 /* Test 8 different probability distributions of zeros */
1964 for( int j = 0; j < test_size; j++ )
1965 input[j] = (rand()&((1 << ((i&7)+1)) - 1)) * rand();
1966 uint8_t *end_c = (uint8_t*)call_c1( bs_c.nal_escape, output1, input, input+test_size );
1967 uint8_t *end_a = (uint8_t*)call_a1( bs_a.nal_escape, output2, input, input+test_size );
1968 int size_c = end_c-output1;
1969 int size_a = end_a-output2;
1970 if( size_c != size_a || memcmp( output1, output2, size_c ) )
1972 fprintf( stderr, "nal_escape : [FAILED] %d %d\n", size_c, size_a );
1977 for( int j = 0; j < size; j++ )
1979 call_c2( bs_c.nal_escape, output1, input, input+size );
1980 call_a2( bs_a.nal_escape, output2, input, input+size );
1985 report( "nal escape:" );
1990 static int check_all_funcs( int cpu_ref, int cpu_new )
1992 return check_pixel( cpu_ref, cpu_new )
1993 + check_dct( cpu_ref, cpu_new )
1994 + check_mc( cpu_ref, cpu_new )
1995 + check_intra( cpu_ref, cpu_new )
1996 + check_deblock( cpu_ref, cpu_new )
1997 + check_quant( cpu_ref, cpu_new )
1998 + check_cabac( cpu_ref, cpu_new )
1999 + check_bitstream( cpu_ref, cpu_new );
2002 static int add_flags( int *cpu_ref, int *cpu_new, int flags, const char *name )
2004 *cpu_ref = *cpu_new;
2006 if( *cpu_new & X264_CPU_SSE2_IS_FAST )
2007 *cpu_new &= ~X264_CPU_SSE2_IS_SLOW;
2009 fprintf( stderr, "x264: %s\n", name );
2010 return check_all_funcs( *cpu_ref, *cpu_new );
2013 static int check_all_flags( void )
2016 int cpu0 = 0, cpu1 = 0;
2018 if( x264_cpu_detect() & X264_CPU_MMXEXT )
2020 ret |= add_flags( &cpu0, &cpu1, X264_CPU_MMX | X264_CPU_MMXEXT, "MMX" );
2021 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "MMX Cache64" );
2022 cpu1 &= ~X264_CPU_CACHELINE_64;
2024 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_32, "MMX Cache32" );
2025 cpu1 &= ~X264_CPU_CACHELINE_32;
2027 if( x264_cpu_detect() & X264_CPU_LZCNT )
2029 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "MMX_LZCNT" );
2030 cpu1 &= ~X264_CPU_LZCNT;
2032 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "MMX SlowCTZ" );
2033 cpu1 &= ~X264_CPU_SLOW_CTZ;
2035 if( x264_cpu_detect() & X264_CPU_SSE2 )
2037 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE | X264_CPU_SSE2 | X264_CPU_SSE2_IS_SLOW, "SSE2Slow" );
2038 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2_IS_FAST, "SSE2Fast" );
2039 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSE2Fast Cache64" );
2040 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SHUFFLE_IS_FAST, "SSE2 FastShuffle" );
2041 cpu1 &= ~X264_CPU_SHUFFLE_IS_FAST;
2042 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSE2 SlowCTZ" );
2043 cpu1 &= ~X264_CPU_SLOW_CTZ;
2044 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSE2 SlowAtom" );
2045 cpu1 &= ~X264_CPU_SLOW_ATOM;
2047 if( x264_cpu_detect() & X264_CPU_SSE_MISALIGN )
2049 cpu1 &= ~X264_CPU_CACHELINE_64;
2050 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE_MISALIGN, "SSE_Misalign" );
2051 cpu1 &= ~X264_CPU_SSE_MISALIGN;
2053 if( x264_cpu_detect() & X264_CPU_LZCNT )
2055 cpu1 &= ~X264_CPU_CACHELINE_64;
2056 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSE_LZCNT" );
2057 cpu1 &= ~X264_CPU_LZCNT;
2059 if( x264_cpu_detect() & X264_CPU_SSE3 )
2060 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE3 | X264_CPU_CACHELINE_64, "SSE3" );
2061 if( x264_cpu_detect() & X264_CPU_SSSE3 )
2063 cpu1 &= ~X264_CPU_CACHELINE_64;
2064 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSSE3, "SSSE3" );
2065 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64" );
2066 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SHUFFLE_IS_FAST, "SSSE3 FastShuffle" );
2067 cpu1 &= ~X264_CPU_SHUFFLE_IS_FAST;
2068 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSSE3 SlowCTZ" );
2069 cpu1 &= ~X264_CPU_SLOW_CTZ;
2070 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSSE3 SlowAtom" );
2071 cpu1 &= ~X264_CPU_SLOW_ATOM;
2073 if( x264_cpu_detect() & X264_CPU_SSE4 )
2075 cpu1 &= ~X264_CPU_CACHELINE_64;
2076 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE4, "SSE4" );
2078 if( x264_cpu_detect() & X264_CPU_AVX )
2079 ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX, "AVX" );
2081 if( x264_cpu_detect() & X264_CPU_ALTIVEC )
2083 fprintf( stderr, "x264: ALTIVEC against C\n" );
2084 ret = check_all_funcs( 0, X264_CPU_ALTIVEC );
2087 if( x264_cpu_detect() & X264_CPU_ARMV6 )
2088 ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV6, "ARMv6" );
2089 if( x264_cpu_detect() & X264_CPU_NEON )
2090 ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
2091 if( x264_cpu_detect() & X264_CPU_FAST_NEON_MRC )
2092 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FAST_NEON_MRC, "Fast NEON MRC" );
2097 int main(int argc, char *argv[])
2101 if( argc > 1 && !strncmp( argv[1], "--bench", 7 ) )
2103 #if !ARCH_X86 && !ARCH_X86_64 && !ARCH_PPC && !ARCH_ARM
2104 fprintf( stderr, "no --bench for your cpu until you port rdtsc\n" );
2108 if( argv[1][7] == '=' )
2110 bench_pattern = argv[1]+8;
2111 bench_pattern_len = strlen(bench_pattern);
2117 int seed = ( argc > 1 ) ? atoi(argv[1]) : x264_mdate();
2118 fprintf( stderr, "x264: using random seed %u\n", seed );
2121 buf1 = x264_malloc( 0x1e00 + 0x2000*sizeof(pixel) + 16*BENCH_ALIGNS );
2122 pbuf1 = x264_malloc( 0x1e00*sizeof(pixel) + 16*BENCH_ALIGNS );
2123 if( !buf1 || !pbuf1 )
2125 fprintf( stderr, "malloc failed, unable to initiate tests!\n" );
2128 #define INIT_POINTER_OFFSETS\
2129 buf2 = buf1 + 0xf00;\
2130 buf3 = buf2 + 0xf00;\
2131 buf4 = buf3 + 0x1000*sizeof(pixel);\
2132 pbuf2 = pbuf1 + 0xf00;\
2133 pbuf3 = (pixel*)buf3;\
2134 pbuf4 = (pixel*)buf4;
2135 INIT_POINTER_OFFSETS;
2136 for( int i = 0; i < 0x1e00; i++ )
2138 buf1[i] = rand() & 0xFF;
2139 pbuf1[i] = rand() & PIXEL_MAX;
2141 memset( buf1+0x1e00, 0, 0x2000*sizeof(pixel) );
2143 /* 16-byte alignment is guaranteed whenever it's useful, but some functions also vary in speed depending on %64 */
2145 for( int i = 0; i < BENCH_ALIGNS && !ret; i++ )
2147 INIT_POINTER_OFFSETS;
2148 ret |= x264_stack_pagealign( check_all_flags, i*16 );
2152 fprintf( stderr, "%d/%d\r", i+1, BENCH_ALIGNS );
2155 ret = check_all_flags();
2159 fprintf( stderr, "x264: at least one test has failed. Go and fix that Right Now!\n" );
2162 fprintf( stderr, "x264: All tests passed Yeah :)\n" );