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 )
492 ALIGNED_16( int sums[5][4] ) = {{0}};
495 res_c = x264_pixel_ssim_wxh( &pixel_c, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3 );
496 res_a = x264_pixel_ssim_wxh( &pixel_asm, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3 );
497 if( fabs( res_c - res_a ) > 1e-6 )
500 fprintf( stderr, "ssim: %.7f != %.7f [FAILED]\n", res_c, res_a );
502 set_func_name( "ssim_core" );
503 call_c2( pixel_c.ssim_4x4x2_core, pbuf1+2, 32, pbuf2+2, 32, sums );
504 call_a2( pixel_asm.ssim_4x4x2_core, pbuf1+2, 32, pbuf2+2, 32, sums );
505 set_func_name( "ssim_end" );
506 call_c2( pixel_c.ssim_end4, sums, sums, 4 );
507 call_a2( pixel_asm.ssim_end4, sums, sums, 4 );
511 ok = 1; used_asm = 0;
512 for( int i = 0; i < 32; i++ )
514 for( int i = 0; i < 100 && ok; i++ )
515 if( pixel_asm.ads[i&3] != pixel_ref.ads[i&3] )
517 ALIGNED_16( uint16_t sums[72] );
518 ALIGNED_16( int dc[4] );
519 int16_t mvs_a[32], mvs_c[32];
521 int thresh = rand() & 0x3fff;
522 set_func_name( "esa_ads" );
523 for( int j = 0; j < 72; j++ )
524 sums[j] = rand() & 0x3fff;
525 for( int j = 0; j < 4; j++ )
526 dc[j] = rand() & 0x3fff;
528 mvn_c = call_c( pixel_c.ads[i&3], dc, sums, 32, cost_mv, mvs_c, 28, thresh );
529 mvn_a = call_a( pixel_asm.ads[i&3], dc, sums, 32, cost_mv, mvs_a, 28, thresh );
530 if( mvn_c != mvn_a || memcmp( mvs_c, mvs_a, mvn_c*sizeof(*mvs_c) ) )
533 printf( "c%d: ", i&3 );
534 for( int j = 0; j < mvn_c; j++ )
535 printf( "%d ", mvs_c[j] );
536 printf( "\na%d: ", i&3 );
537 for( int j = 0; j < mvn_a; j++ )
538 printf( "%d ", mvs_a[j] );
542 report( "esa ads:" );
547 static int check_dct( int cpu_ref, int cpu_new )
549 x264_dct_function_t dct_c;
550 x264_dct_function_t dct_ref;
551 x264_dct_function_t dct_asm;
552 x264_quant_function_t qf;
553 int ret = 0, ok, used_asm, interlace = 0;
554 ALIGNED_16( dctcoef dct1[16][16] );
555 ALIGNED_16( dctcoef dct2[16][16] );
556 ALIGNED_16( dctcoef dct4[16][16] );
557 ALIGNED_16( dctcoef dct8[4][64] );
558 ALIGNED_16( dctcoef dctdc[2][4] );
562 x264_dct_init( 0, &dct_c );
563 x264_dct_init( cpu_ref, &dct_ref);
564 x264_dct_init( cpu_new, &dct_asm );
566 memset( h, 0, sizeof(*h) );
567 h->pps = h->pps_array;
568 x264_param_default( &h->param );
569 h->chroma_qp_table = i_chroma_qp_table + 12;
570 h->param.analyse.i_luma_deadzone[0] = 0;
571 h->param.analyse.i_luma_deadzone[1] = 0;
572 h->param.analyse.b_transform_8x8 = 1;
573 for( int i = 0; i < 6; i++ )
574 h->pps->scaling_list[i] = x264_cqm_flat16;
576 x264_quant_init( h, 0, &qf );
578 /* overflow test cases */
579 for( int i = 0; i < 5; i++ )
581 pixel *enc = &pbuf3[16*i*FENC_STRIDE];
582 pixel *dec = &pbuf4[16*i*FDEC_STRIDE];
584 for( int j = 0; j < 16; j++ )
586 int cond_a = (i < 2) ? 1 : ((j&3) == 0 || (j&3) == (i-1));
587 int cond_b = (i == 0) ? 1 : !cond_a;
588 enc[0] = enc[1] = cond_a ? PIXEL_MAX : 0;
589 enc[2] = enc[3] = cond_b ? PIXEL_MAX : 0;
591 for( int k = 0; k < 4; k++ )
592 dec[k] = PIXEL_MAX - enc[k];
599 #define TEST_DCT( name, t1, t2, size ) \
600 if( dct_asm.name != dct_ref.name ) \
602 set_func_name( #name ); \
604 pixel *enc = pbuf3; \
605 pixel *dec = pbuf4; \
606 for( int j = 0; j < 5; j++) \
608 call_c( dct_c.name, t1, &pbuf1[j*64], &pbuf2[j*64] ); \
609 call_a( dct_asm.name, t2, &pbuf1[j*64], &pbuf2[j*64] ); \
610 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
613 fprintf( stderr, #name " [FAILED]\n" ); \
616 call_c( dct_c.name, t1, enc, dec ); \
617 call_a( dct_asm.name, t2, enc, dec ); \
618 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
621 fprintf( stderr, #name " [FAILED] (overflow)\n" ); \
624 enc += 16*FENC_STRIDE; \
625 dec += 16*FDEC_STRIDE; \
628 ok = 1; used_asm = 0;
629 TEST_DCT( sub4x4_dct, dct1[0], dct2[0], 16 );
630 TEST_DCT( sub8x8_dct, dct1, dct2, 16*4 );
631 TEST_DCT( sub8x8_dct_dc, dctdc[0], dctdc[1], 4 );
632 TEST_DCT( sub16x16_dct, dct1, dct2, 16*16 );
633 report( "sub_dct4 :" );
635 ok = 1; used_asm = 0;
636 TEST_DCT( sub8x8_dct8, (void*)dct1[0], (void*)dct2[0], 64 );
637 TEST_DCT( sub16x16_dct8, (void*)dct1, (void*)dct2, 64*4 );
638 report( "sub_dct8 :" );
641 // fdct and idct are denormalized by different factors, so quant/dequant
642 // is needed to force the coefs into the right range.
643 dct_c.sub16x16_dct( dct4, pbuf1, pbuf2 );
644 dct_c.sub16x16_dct8( dct8, pbuf1, pbuf2 );
645 for( int i = 0; i < 16; i++ )
647 qf.quant_4x4( dct4[i], h->quant4_mf[CQM_4IY][20], h->quant4_bias[CQM_4IY][20] );
648 qf.dequant_4x4( dct4[i], h->dequant4_mf[CQM_4IY], 20 );
650 for( int i = 0; i < 4; i++ )
652 qf.quant_8x8( dct8[i], h->quant8_mf[CQM_8IY][20], h->quant8_bias[CQM_8IY][20] );
653 qf.dequant_8x8( dct8[i], h->dequant8_mf[CQM_8IY], 20 );
655 x264_cqm_delete( h );
657 #define TEST_IDCT( name, src ) \
658 if( dct_asm.name != dct_ref.name ) \
660 set_func_name( #name ); \
662 memcpy( pbuf3, pbuf1, 32*32 * sizeof(pixel) ); \
663 memcpy( pbuf4, pbuf1, 32*32 * sizeof(pixel) ); \
664 memcpy( dct1, src, 256 * sizeof(dctcoef) ); \
665 memcpy( dct2, src, 256 * sizeof(dctcoef) ); \
666 call_c1( dct_c.name, pbuf3, (void*)dct1 ); \
667 call_a1( dct_asm.name, pbuf4, (void*)dct2 ); \
668 if( memcmp( pbuf3, pbuf4, 32*32 * sizeof(pixel) ) ) \
671 fprintf( stderr, #name " [FAILED]\n" ); \
673 call_c2( dct_c.name, pbuf3, (void*)dct1 ); \
674 call_a2( dct_asm.name, pbuf4, (void*)dct2 ); \
676 ok = 1; used_asm = 0;
677 TEST_IDCT( add4x4_idct, dct4 );
678 TEST_IDCT( add8x8_idct, dct4 );
679 TEST_IDCT( add8x8_idct_dc, dct4 );
680 TEST_IDCT( add16x16_idct, dct4 );
681 TEST_IDCT( add16x16_idct_dc, dct4 );
682 report( "add_idct4 :" );
684 ok = 1; used_asm = 0;
685 TEST_IDCT( add8x8_idct8, dct8 );
686 TEST_IDCT( add16x16_idct8, dct8 );
687 report( "add_idct8 :" );
690 #define TEST_DCTDC( name )\
691 ok = 1; used_asm = 0;\
692 if( dct_asm.name != dct_ref.name )\
694 set_func_name( #name );\
696 uint16_t *p = (uint16_t*)buf1;\
697 for( int i = 0; i < 16 && ok; i++ )\
699 for( int j = 0; j < 16; j++ )\
700 dct1[0][j] = !i ? (j^j>>1^j>>2^j>>3)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
701 : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
702 : ((*p++)&0x1fff)-0x1000; /* general case */\
703 memcpy( dct2, dct1, 16 * sizeof(dctcoef) );\
704 call_c1( dct_c.name, dct1[0] );\
705 call_a1( dct_asm.name, dct2[0] );\
706 if( memcmp( dct1, dct2, 16 * sizeof(dctcoef) ) )\
709 call_c2( dct_c.name, dct1[0] );\
710 call_a2( dct_asm.name, dct2[0] );\
712 report( #name " :" );
714 TEST_DCTDC( dct4x4dc );
715 TEST_DCTDC( idct4x4dc );
718 x264_zigzag_function_t zigzag_c[2];
719 x264_zigzag_function_t zigzag_ref[2];
720 x264_zigzag_function_t zigzag_asm[2];
722 ALIGNED_16( dctcoef level1[64] );
723 ALIGNED_16( dctcoef level2[64] );
725 #define TEST_ZIGZAG_SCAN( name, t1, t2, dct, size ) \
726 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
728 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
730 memcpy(dct, buf1, size*sizeof(dctcoef)); \
731 call_c( zigzag_c[interlace].name, t1, dct ); \
732 call_a( zigzag_asm[interlace].name, t2, dct ); \
733 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
736 fprintf( stderr, #name " [FAILED]\n" ); \
740 #define TEST_ZIGZAG_SUB( name, t1, t2, size ) \
741 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
744 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
746 memcpy( pbuf3, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
747 memcpy( pbuf4, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
748 nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
749 nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
750 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( pbuf3, pbuf4, 16*FDEC_STRIDE*sizeof(pixel) ) || nz_c != nz_a ) \
753 fprintf( stderr, #name " [FAILED]\n" ); \
755 call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
756 call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
759 #define TEST_ZIGZAG_SUBAC( name, t1, t2 ) \
760 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
763 dctcoef dc_a, dc_c; \
764 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
766 for( int i = 0; i < 2; i++ ) \
768 memcpy( pbuf3, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
769 memcpy( pbuf4, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
770 for( int j = 0; j < 4; j++ ) \
772 memcpy( pbuf3 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
773 memcpy( pbuf4 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
775 nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
776 nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
777 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 ) \
780 fprintf( stderr, #name " [FAILED]\n" ); \
784 call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
785 call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
788 #define TEST_INTERLEAVE( name, t1, t2, dct, size ) \
789 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
791 for( int j = 0; j < 100; j++ ) \
793 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
795 memcpy(dct, buf1, size*sizeof(dctcoef)); \
796 for( int i = 0; i < size; i++ ) \
797 dct[i] = rand()&0x1F ? 0 : dct[i]; \
798 memcpy(buf3, buf4, 10); \
799 call_c( zigzag_c[interlace].name, t1, dct, buf3 ); \
800 call_a( zigzag_asm[interlace].name, t2, dct, buf4 ); \
801 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( buf3, buf4, 10 ) ) \
808 x264_zigzag_init( 0, &zigzag_c[0], &zigzag_c[1] );
809 x264_zigzag_init( cpu_ref, &zigzag_ref[0], &zigzag_ref[1] );
810 x264_zigzag_init( cpu_new, &zigzag_asm[0], &zigzag_asm[1] );
812 ok = 1; used_asm = 0;
813 TEST_INTERLEAVE( interleave_8x8_cavlc, level1, level2, dct1[0], 64 );
814 report( "zigzag_interleave :" );
816 for( interlace = 0; interlace <= 1; interlace++ )
818 ok = 1; used_asm = 0;
819 TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, (void*)dct1, 64 );
820 TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 16 );
821 TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
822 TEST_ZIGZAG_SUBAC( sub_4x4ac, level1, level2 );
823 report( interlace ? "zigzag_field :" : "zigzag_frame :" );
825 #undef TEST_ZIGZAG_SCAN
826 #undef TEST_ZIGZAG_SUB
831 static int check_mc( int cpu_ref, int cpu_new )
833 x264_mc_functions_t mc_c;
834 x264_mc_functions_t mc_ref;
835 x264_mc_functions_t mc_a;
836 x264_pixel_function_t pixf;
838 pixel *src = &(pbuf1)[2*64+2];
839 pixel *src2[4] = { &(pbuf1)[3*64+2], &(pbuf1)[5*64+2],
840 &(pbuf1)[7*64+2], &(pbuf1)[9*64+2] };
844 int ret = 0, ok, used_asm;
846 x264_mc_init( 0, &mc_c );
847 x264_mc_init( cpu_ref, &mc_ref );
848 x264_mc_init( cpu_new, &mc_a );
849 x264_pixel_init( 0, &pixf );
851 #define MC_TEST_LUMA( w, h ) \
852 if( mc_a.mc_luma != mc_ref.mc_luma && !(w&(w-1)) && h<=16 ) \
854 const x264_weight_t *weight = weight_none; \
855 set_func_name( "mc_luma_%dx%d", w, h ); \
857 for( int i = 0; i < 1024; i++ ) \
858 pbuf3[i] = pbuf4[i] = 0xCD; \
859 call_c( mc_c.mc_luma, dst1, 32, src2, 64, dx, dy, w, h, weight ); \
860 call_a( mc_a.mc_luma, dst2, 32, src2, 64, dx, dy, w, h, weight ); \
861 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
863 fprintf( stderr, "mc_luma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
867 if( mc_a.get_ref != mc_ref.get_ref ) \
870 int ref_stride = 32; \
871 int w_checked = ( ( sizeof(pixel) == 2 && (w == 12 || w == 20)) ? w-2 : w ); \
872 const x264_weight_t *weight = weight_none; \
873 set_func_name( "get_ref_%dx%d", w_checked, h ); \
875 for( int i = 0; i < 1024; i++ ) \
876 pbuf3[i] = pbuf4[i] = 0xCD; \
877 call_c( mc_c.mc_luma, dst1, 32, src2, 64, dx, dy, w, h, weight ); \
878 ref = (pixel*)call_a( mc_a.get_ref, ref, &ref_stride, src2, 64, dx, dy, w, h, weight ); \
879 for( int i = 0; i < h; i++ ) \
880 if( memcmp( dst1+i*32, ref+i*ref_stride, w_checked * sizeof(pixel) ) ) \
882 fprintf( stderr, "get_ref[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w_checked, h ); \
888 #define MC_TEST_CHROMA( w, h ) \
889 if( mc_a.mc_chroma != mc_ref.mc_chroma ) \
891 set_func_name( "mc_chroma_%dx%d", w, h ); \
893 for( int i = 0; i < 1024; i++ ) \
894 pbuf3[i] = pbuf4[i] = 0xCD; \
895 call_c( mc_c.mc_chroma, dst1, dst1+8, 16, src, 64, dx, dy, w, h ); \
896 call_a( mc_a.mc_chroma, dst2, dst2+8, 16, src, 64, dx, dy, w, h ); \
897 /* mc_chroma width=2 may write garbage to the right of dst. ignore that. */ \
898 for( int j = 0; j < h; j++ ) \
899 for( int i = w; i < 8; i++ ) \
901 dst2[i+j*16+8] = dst1[i+j*16+8]; \
902 dst2[i+j*16] = dst1[i+j*16]; \
904 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
906 fprintf( stderr, "mc_chroma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
910 ok = 1; used_asm = 0;
911 for( int dy = -8; dy < 8; dy++ )
912 for( int dx = -128; dx < 128; dx++ )
914 if( rand()&15 ) continue; // running all of them is too slow
915 MC_TEST_LUMA( 20, 18 );
916 MC_TEST_LUMA( 16, 16 );
917 MC_TEST_LUMA( 16, 8 );
918 MC_TEST_LUMA( 12, 10 );
919 MC_TEST_LUMA( 8, 16 );
920 MC_TEST_LUMA( 8, 8 );
921 MC_TEST_LUMA( 8, 4 );
922 MC_TEST_LUMA( 4, 8 );
923 MC_TEST_LUMA( 4, 4 );
925 report( "mc luma :" );
927 ok = 1; used_asm = 0;
928 for( int dy = -1; dy < 9; dy++ )
929 for( int dx = -128; dx < 128; dx++ )
931 if( rand()&15 ) continue;
932 MC_TEST_CHROMA( 8, 8 );
933 MC_TEST_CHROMA( 8, 4 );
934 MC_TEST_CHROMA( 4, 8 );
935 MC_TEST_CHROMA( 4, 4 );
936 MC_TEST_CHROMA( 4, 2 );
937 MC_TEST_CHROMA( 2, 4 );
938 MC_TEST_CHROMA( 2, 2 );
940 report( "mc chroma :" );
942 #undef MC_TEST_CHROMA
944 #define MC_TEST_AVG( name, weight ) \
946 ok = 1, used_asm = 0; \
947 for( int i = 0; i < 10; i++ ) \
949 memcpy( pbuf3, pbuf1+320, 320 * sizeof(pixel) ); \
950 memcpy( pbuf4, pbuf1+320, 320 * sizeof(pixel) ); \
951 if( mc_a.name[i] != mc_ref.name[i] ) \
953 set_func_name( "%s_%s", #name, pixel_names[i] ); \
955 call_c1( mc_c.name[i], pbuf3, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
956 call_a1( mc_a.name[i], pbuf4, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
957 if( memcmp( pbuf3, pbuf4, 320 * sizeof(pixel) ) ) \
960 fprintf( stderr, #name "[%d]: [FAILED]\n", i ); \
962 call_c2( mc_c.name[i], pbuf3, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
963 call_a2( mc_a.name[i], pbuf4, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
968 for( int w = -63; w <= 127 && ok; w++ )
969 MC_TEST_AVG( avg, w );
970 report( "mc wpredb :" );
972 #define MC_TEST_WEIGHT( name, weight, aligned ) \
973 int align_off = (aligned ? 0 : rand()%16); \
974 ok = 1, used_asm = 0; \
975 for( int i = 1; i <= 5; i++ ) \
977 ALIGNED_16( pixel buffC[640] ); \
978 ALIGNED_16( pixel buffA[640] ); \
979 int j = X264_MAX( i*4, 2 ); \
980 memset( buffC, 0, 640 * sizeof(pixel) ); \
981 memset( buffA, 0, 640 * sizeof(pixel) ); \
984 /* w12 is the same as w16 in some cases */ \
985 if( i == 3 && mc_a.name[i] == mc_a.name[i+1] ) \
987 if( mc_a.name[i] != mc_ref.name[i] ) \
989 set_func_name( "%s_w%d", #name, j ); \
991 call_c1( mc_c.weight[i], buffC, 32, pbuf2+align_off, 32, &weight, 16 ); \
992 mc_a.weight_cache(&ha, &weight); \
993 call_a1( weight.weightfn[i], buffA, 32, pbuf2+align_off, 32, &weight, 16 ); \
994 for( int k = 0; k < 16; k++ ) \
995 if( memcmp( &buffC[k*32], &buffA[k*32], j * sizeof(pixel) ) ) \
998 fprintf( stderr, #name "[%d]: [FAILED] s:%d o:%d d%d\n", i, s, o, d ); \
1001 call_c2( mc_c.weight[i], buffC, 32, pbuf2+align_off, 32, &weight, 16 ); \
1002 call_a2( weight.weightfn[i], buffA, 32, pbuf2+align_off, 32, &weight, 16 ); \
1006 ok = 1; used_asm = 0;
1009 for( int s = 0; s <= 127 && ok; s++ )
1011 for( int o = -128; o <= 127 && ok; o++ )
1013 if( rand() & 2047 ) continue;
1014 for( int d = 0; d <= 7 && ok; d++ )
1018 x264_weight_t weight = { .i_scale = s, .i_denom = d, .i_offset = o };
1019 MC_TEST_WEIGHT( weight, weight, (align_cnt++ % 4) );
1024 report( "mc weight :" );
1026 ok = 1; used_asm = 0;
1027 for( int o = 0; o <= 127 && ok; o++ )
1030 if( rand() & 15 ) continue;
1031 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1032 MC_TEST_WEIGHT( offsetadd, weight, (align_cnt++ % 4) );
1034 report( "mc offsetadd :" );
1035 ok = 1; used_asm = 0;
1036 for( int o = -128; o < 0 && ok; o++ )
1039 if( rand() & 15 ) continue;
1040 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1041 MC_TEST_WEIGHT( offsetsub, weight, (align_cnt++ % 4) );
1043 report( "mc offsetsub :" );
1045 ok = 1; used_asm = 0;
1046 if( mc_a.store_interleave_8x8x2 != mc_ref.store_interleave_8x8x2 )
1048 set_func_name( "store_interleave_8x8x2" );
1050 memset( pbuf3, 0, 64*8 );
1051 memset( pbuf4, 0, 64*8 );
1052 call_c( mc_c.store_interleave_8x8x2, pbuf3, 64, pbuf1, pbuf1+16 );
1053 call_a( mc_a.store_interleave_8x8x2, pbuf4, 64, pbuf1, pbuf1+16 );
1054 if( memcmp( pbuf3, pbuf4, 64*8 ) )
1057 if( mc_a.load_deinterleave_8x8x2_fenc != mc_ref.load_deinterleave_8x8x2_fenc )
1059 set_func_name( "load_deinterleave_8x8x2_fenc" );
1061 call_c( mc_c.load_deinterleave_8x8x2_fenc, pbuf3, pbuf1, 64 );
1062 call_a( mc_a.load_deinterleave_8x8x2_fenc, pbuf4, pbuf1, 64 );
1063 if( memcmp( pbuf3, pbuf4, FENC_STRIDE*8 ) )
1066 if( mc_a.load_deinterleave_8x8x2_fdec != mc_ref.load_deinterleave_8x8x2_fdec )
1068 set_func_name( "load_deinterleave_8x8x2_fdec" );
1070 call_c( mc_c.load_deinterleave_8x8x2_fdec, pbuf3, pbuf1, 64 );
1071 call_a( mc_a.load_deinterleave_8x8x2_fdec, pbuf4, pbuf1, 64 );
1072 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*8 ) )
1075 report( "store_interleave :" );
1078 int w, h, src_stride;
1079 } 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} };
1080 ok = 1; used_asm = 0;
1081 if( mc_a.plane_copy != mc_ref.plane_copy )
1083 set_func_name( "plane_copy" );
1085 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1087 int w = plane_specs[i].w;
1088 int h = plane_specs[i].h;
1089 int src_stride = plane_specs[i].src_stride;
1090 int dst_stride = (w + 127) & ~63;
1091 assert( dst_stride * h <= 0x1000 );
1092 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1093 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1094 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1095 call_c( mc_c.plane_copy, pbuf3, dst_stride, src1, src_stride, w, h );
1096 call_a( mc_a.plane_copy, pbuf4, dst_stride, src1, src_stride, w, h );
1097 for( int y = 0; y < h; y++ )
1098 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(pixel) ) )
1101 fprintf( stderr, "plane_copy FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
1107 if( mc_a.plane_copy_interleave != mc_ref.plane_copy_interleave )
1109 set_func_name( "plane_copy_interleave" );
1111 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1113 int w = (plane_specs[i].w + 1) >> 1;
1114 int h = plane_specs[i].h;
1115 int src_stride = (plane_specs[i].src_stride + 1) >> 1;
1116 int dst_stride = (2*w + 127) & ~63;
1117 assert( dst_stride * h <= 0x1000 );
1118 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1119 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1120 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1121 call_c( mc_c.plane_copy_interleave, pbuf3, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1122 call_a( mc_a.plane_copy_interleave, pbuf4, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1123 for( int y = 0; y < h; y++ )
1124 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
1127 fprintf( stderr, "plane_copy_interleave FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
1133 if( mc_a.plane_copy_deinterleave != mc_ref.plane_copy_deinterleave )
1135 set_func_name( "plane_copy_deinterleave" );
1137 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1139 int w = (plane_specs[i].w + 1) >> 1;
1140 int h = plane_specs[i].h;
1142 int src_stride = (2*w + 127) & ~63;
1143 int offv = (dst_stride*h + 31) & ~15;
1144 memset( pbuf3, 0, 0x1000 );
1145 memset( pbuf4, 0, 0x1000 );
1146 call_c( mc_c.plane_copy_deinterleave, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf1, src_stride, w, h );
1147 call_a( mc_a.plane_copy_deinterleave, pbuf4, dst_stride, pbuf4+offv, dst_stride, pbuf1, src_stride, w, h );
1148 for( int y = 0; y < h; y++ )
1149 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w ) ||
1150 memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w ) )
1153 fprintf( stderr, "plane_copy_deinterleave FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
1158 report( "plane_copy :" );
1160 if( mc_a.hpel_filter != mc_ref.hpel_filter )
1162 pixel *srchpel = pbuf1+8+2*64;
1163 pixel *dstc[3] = { pbuf3+8, pbuf3+8+16*64, pbuf3+8+32*64 };
1164 pixel *dsta[3] = { pbuf4+8, pbuf4+8+16*64, pbuf4+8+32*64 };
1165 void *tmp = pbuf3+49*64;
1166 set_func_name( "hpel_filter" );
1167 ok = 1; used_asm = 1;
1168 memset( pbuf3, 0, 4096 * sizeof(pixel) );
1169 memset( pbuf4, 0, 4096 * sizeof(pixel) );
1170 call_c( mc_c.hpel_filter, dstc[0], dstc[1], dstc[2], srchpel, 64, 48, 10, tmp );
1171 call_a( mc_a.hpel_filter, dsta[0], dsta[1], dsta[2], srchpel, 64, 48, 10, tmp );
1172 for( int i = 0; i < 3; i++ )
1173 for( int j = 0; j < 10; j++ )
1174 //FIXME ideally the first pixels would match too, but they aren't actually used
1175 if( memcmp( dstc[i]+j*64+2, dsta[i]+j*64+2, 43 * sizeof(pixel) ) )
1178 fprintf( stderr, "hpel filter differs at plane %c line %d\n", "hvc"[i], j );
1179 for( int k = 0; k < 48; k++ )
1180 printf( "%02x%s", dstc[i][j*64+k], (k+1)&3 ? "" : " " );
1182 for( int k = 0; k < 48; k++ )
1183 printf( "%02x%s", dsta[i][j*64+k], (k+1)&3 ? "" : " " );
1187 report( "hpel filter :" );
1190 if( mc_a.frame_init_lowres_core != mc_ref.frame_init_lowres_core )
1192 pixel *dstc[4] = { pbuf3, pbuf3+1024, pbuf3+2048, pbuf3+3072 };
1193 pixel *dsta[4] = { pbuf4, pbuf4+1024, pbuf4+2048, pbuf4+3072 };
1194 set_func_name( "lowres_init" );
1195 ok = 1; used_asm = 1;
1196 for( int w = 40; w <= 48; w += 8 )
1198 int stride = (w+8)&~15;
1199 call_c( mc_c.frame_init_lowres_core, pbuf1, dstc[0], dstc[1], dstc[2], dstc[3], w*2, stride, w, 16 );
1200 call_a( mc_a.frame_init_lowres_core, pbuf1, dsta[0], dsta[1], dsta[2], dsta[3], w*2, stride, w, 16 );
1201 for( int i = 0; i < 16; i++ )
1203 for( int j = 0; j < 4; j++ )
1204 if( memcmp( dstc[j]+i*stride, dsta[j]+i*stride, w * sizeof(pixel) ) )
1207 fprintf( stderr, "frame_init_lowres differs at plane %d line %d\n", j, i );
1208 for( int k = 0; k < w; k++ )
1209 printf( "%d ", dstc[j][k+i*stride] );
1211 for( int k = 0; k < w; k++ )
1212 printf( "%d ", dsta[j][k+i*stride] );
1218 report( "lowres init :" );
1221 #define INTEGRAL_INIT( name, size, ... )\
1222 if( mc_a.name != mc_ref.name )\
1225 set_func_name( #name );\
1227 memcpy( buf3, buf1, size*2*stride );\
1228 memcpy( buf4, buf1, size*2*stride );\
1229 uint16_t *sum = (uint16_t*)buf3;\
1230 call_c1( mc_c.name, __VA_ARGS__ );\
1231 sum = (uint16_t*)buf4;\
1232 call_a1( mc_a.name, __VA_ARGS__ );\
1233 if( memcmp( buf3, buf4, (stride-8)*2 ) \
1234 || (size>9 && memcmp( buf3+18*stride, buf4+18*stride, (stride-8)*2 )))\
1236 call_c2( mc_c.name, __VA_ARGS__ );\
1237 call_a2( mc_a.name, __VA_ARGS__ );\
1239 ok = 1; used_asm = 0;
1240 INTEGRAL_INIT( integral_init4h, 2, sum+stride, pbuf2, stride );
1241 INTEGRAL_INIT( integral_init8h, 2, sum+stride, pbuf2, stride );
1242 INTEGRAL_INIT( integral_init4v, 14, sum, sum+9*stride, stride );
1243 INTEGRAL_INIT( integral_init8v, 9, sum, stride );
1244 report( "integral init :" );
1246 if( mc_a.mbtree_propagate_cost != mc_ref.mbtree_propagate_cost )
1249 for( int i = 0; i < 10; i++ )
1251 float fps_factor = (rand()&65535) / 256.;
1252 ok = 1; used_asm = 1;
1253 set_func_name( "mbtree_propagate" );
1254 int *dsta = (int*)buf3;
1255 int *dstc = dsta+400;
1256 uint16_t *prop = (uint16_t*)buf1;
1257 uint16_t *intra = (uint16_t*)buf4;
1258 uint16_t *inter = intra+100;
1259 uint16_t *qscale = inter+100;
1260 uint16_t *rnd = (uint16_t*)buf2;
1262 for( int j = 0; j < 100; j++ )
1264 intra[j] = *rnd++ & 0x7fff;
1265 intra[j] += !intra[j];
1266 inter[j] = *rnd++ & 0x7fff;
1267 qscale[j] = *rnd++ & 0x7fff;
1269 call_c( mc_c.mbtree_propagate_cost, dstc, prop, intra, inter, qscale, &fps_factor, 100 );
1270 call_a( mc_a.mbtree_propagate_cost, dsta, prop, intra, inter, qscale, &fps_factor, 100 );
1271 // I don't care about exact rounding, this is just how close the floating-point implementation happens to be
1273 for( int j = 0; j < 100; j++ )
1274 ok &= abs( dstc[j]-dsta[j] ) <= 1 || fabs( (double)dstc[j]/dsta[j]-1 ) < 1e-4;
1276 report( "mbtree propagate :" );
1282 static int check_deblock( int cpu_ref, int cpu_new )
1284 x264_deblock_function_t db_c;
1285 x264_deblock_function_t db_ref;
1286 x264_deblock_function_t db_a;
1287 int ret = 0, ok = 1, used_asm = 0;
1288 int alphas[36], betas[36];
1291 x264_deblock_init( 0, &db_c, 0 );
1292 x264_deblock_init( cpu_ref, &db_ref, 0 );
1293 x264_deblock_init( cpu_new, &db_a, 0 );
1295 /* not exactly the real values of a,b,tc but close enough */
1296 for( int i = 35, a = 255, c = 250; i >= 0; i-- )
1298 alphas[i] = a << (BIT_DEPTH-8);
1299 betas[i] = (i+1)/2 << (BIT_DEPTH-8);
1300 tcs[i][0] = tcs[i][3] = (c+6)/10 << (BIT_DEPTH-8);
1301 tcs[i][1] = (c+7)/15 << (BIT_DEPTH-8);
1302 tcs[i][2] = (c+9)/20 << (BIT_DEPTH-8);
1307 #define TEST_DEBLOCK( name, align, ... ) \
1308 for( int i = 0; i < 36; i++ ) \
1310 int off = 8*32 + (i&15)*4*!align; /* benchmark various alignments of h filter */ \
1311 for( int j = 0; j < 1024; j++ ) \
1312 /* two distributions of random to excersize different failure modes */ \
1313 pbuf3[j] = rand() & (i&1 ? 0xf : PIXEL_MAX ); \
1314 memcpy( pbuf4, pbuf3, 1024 * sizeof(pixel) ); \
1315 if( db_a.name != db_ref.name ) \
1317 set_func_name( #name ); \
1319 call_c1( db_c.name, pbuf3+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1320 call_a1( db_a.name, pbuf4+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1321 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1324 fprintf( stderr, #name "(a=%d, b=%d): [FAILED]\n", alphas[i], betas[i] ); \
1327 call_c2( db_c.name, pbuf3+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1328 call_a2( db_a.name, pbuf4+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1332 TEST_DEBLOCK( deblock_luma[0], 0, tcs[i] );
1333 TEST_DEBLOCK( deblock_luma[1], 1, tcs[i] );
1334 TEST_DEBLOCK( deblock_chroma[0], 0, tcs[i] );
1335 TEST_DEBLOCK( deblock_chroma[1], 1, tcs[i] );
1336 TEST_DEBLOCK( deblock_luma_intra[0], 0 );
1337 TEST_DEBLOCK( deblock_luma_intra[1], 1 );
1338 TEST_DEBLOCK( deblock_chroma_intra[0], 0 );
1339 TEST_DEBLOCK( deblock_chroma_intra[1], 1 );
1341 if( db_a.deblock_strength != db_ref.deblock_strength )
1343 for( int i = 0; i < 100; i++ )
1345 ALIGNED_ARRAY_16( uint8_t, nnz, [X264_SCAN8_SIZE] );
1346 ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );
1347 ALIGNED_ARRAY_16( int16_t, mv, [2],[X264_SCAN8_LUMA_SIZE][2] );
1348 ALIGNED_ARRAY_16( uint8_t, bs, [2],[2][8][4] );
1349 memset( bs, 99, sizeof(bs) );
1350 for( int j = 0; j < X264_SCAN8_SIZE; j++ )
1351 nnz[j] = ((rand()&7) == 7) * rand() & 0xf;
1352 for( int j = 0; j < 2; j++ )
1353 for( int k = 0; k < X264_SCAN8_LUMA_SIZE; k++ )
1355 ref[j][k] = ((rand()&3) != 3) ? 0 : (rand() & 31) - 2;
1356 for( int l = 0; l < 2; l++ )
1357 mv[j][k][l] = ((rand()&7) != 7) ? (rand()&7) - 3 : (rand()&1023) - 512;
1359 set_func_name( "deblock_strength" );
1360 call_c( db_c.deblock_strength, nnz, ref, mv, bs[0], 2<<(i&1), ((i>>1)&1), NULL );
1361 call_a( db_a.deblock_strength, nnz, ref, mv, bs[1], 2<<(i&1), ((i>>1)&1), NULL );
1362 if( memcmp( bs[0], bs[1], sizeof(bs[0]) ) )
1365 fprintf( stderr, "deblock_strength: [FAILED]\n" );
1366 for( int j = 0; j < 2; j++ )
1368 for( int k = 0; k < 2; k++ )
1369 for( int l = 0; l < 4; l++ )
1371 for( int m = 0; m < 4; m++ )
1372 printf("%d ",bs[j][k][l][m]);
1382 report( "deblock :" );
1387 static int check_quant( int cpu_ref, int cpu_new )
1389 x264_quant_function_t qf_c;
1390 x264_quant_function_t qf_ref;
1391 x264_quant_function_t qf_a;
1392 ALIGNED_16( dctcoef dct1[64] );
1393 ALIGNED_16( dctcoef dct2[64] );
1394 ALIGNED_16( uint8_t cqm_buf[64] );
1395 int ret = 0, ok, used_asm;
1396 int oks[3] = {1,1,1}, used_asms[3] = {0,0,0};
1399 memset( h, 0, sizeof(*h) );
1400 h->pps = h->pps_array;
1401 x264_param_default( &h->param );
1402 h->chroma_qp_table = i_chroma_qp_table + 12;
1403 h->param.analyse.b_transform_8x8 = 1;
1405 for( int i_cqm = 0; i_cqm < 4; i_cqm++ )
1409 for( int i = 0; i < 6; i++ )
1410 h->pps->scaling_list[i] = x264_cqm_flat16;
1411 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_FLAT;
1413 else if( i_cqm == 1 )
1415 for( int i = 0; i < 6; i++ )
1416 h->pps->scaling_list[i] = x264_cqm_jvt[i];
1417 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_JVT;
1421 int max_scale = BIT_DEPTH < 10 ? 255 : 228;
1423 for( int i = 0; i < 64; i++ )
1424 cqm_buf[i] = 10 + rand() % (max_scale - 9);
1426 for( int i = 0; i < 64; i++ )
1428 for( int i = 0; i < 6; i++ )
1429 h->pps->scaling_list[i] = cqm_buf;
1430 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_CUSTOM;
1433 h->param.rc.i_qp_min = 0;
1434 h->param.rc.i_qp_max = QP_MAX;
1436 x264_quant_init( h, 0, &qf_c );
1437 x264_quant_init( h, cpu_ref, &qf_ref );
1438 x264_quant_init( h, cpu_new, &qf_a );
1440 #define INIT_QUANT8(j) \
1442 static const int scale1d[8] = {32,31,24,31,32,31,24,31}; \
1443 for( int i = 0; i < 64; i++ ) \
1445 unsigned int scale = (255*scale1d[i>>3]*scale1d[i&7])/16; \
1446 dct1[i] = dct2[i] = j ? (rand()%(2*scale+1))-scale : 0; \
1450 #define INIT_QUANT4(j) \
1452 static const int scale1d[4] = {4,6,4,6}; \
1453 for( int i = 0; i < 16; i++ ) \
1455 unsigned int scale = 255*scale1d[i>>2]*scale1d[i&3]; \
1456 dct1[i] = dct2[i] = j ? (rand()%(2*scale+1))-scale : 0; \
1460 #define TEST_QUANT_DC( name, cqm ) \
1461 if( qf_a.name != qf_ref.name ) \
1463 set_func_name( #name ); \
1465 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1467 for( int j = 0; j < 2; j++ ) \
1469 int result_c, result_a; \
1470 for( int i = 0; i < 16; i++ ) \
1471 dct1[i] = dct2[i] = j ? (rand() & 0x1fff) - 0xfff : 0; \
1472 result_c = call_c1( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1473 result_a = call_a1( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1474 if( memcmp( dct1, dct2, 16*sizeof(dctcoef) ) || result_c != result_a ) \
1477 fprintf( stderr, #name "(cqm=%d): [FAILED]\n", i_cqm ); \
1480 call_c2( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1481 call_a2( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1486 #define TEST_QUANT( qname, block, w ) \
1487 if( qf_a.qname != qf_ref.qname ) \
1489 set_func_name( #qname ); \
1491 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1493 for( int j = 0; j < 2; j++ ) \
1496 int result_c = call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1497 int result_a = call_a1( qf_a.qname, dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1498 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) || result_c != result_a ) \
1501 fprintf( stderr, #qname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1504 call_c2( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1505 call_a2( qf_a.qname, dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1510 TEST_QUANT( quant_8x8, CQM_8IY, 8 );
1511 TEST_QUANT( quant_8x8, CQM_8PY, 8 );
1512 TEST_QUANT( quant_4x4, CQM_4IY, 4 );
1513 TEST_QUANT( quant_4x4, CQM_4PY, 4 );
1514 TEST_QUANT_DC( quant_4x4_dc, **h->quant4_mf[CQM_4IY] );
1515 TEST_QUANT_DC( quant_2x2_dc, **h->quant4_mf[CQM_4IC] );
1517 #define TEST_DEQUANT( qname, dqname, block, w ) \
1518 if( qf_a.dqname != qf_ref.dqname ) \
1520 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
1522 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1525 call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1526 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1527 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1528 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1529 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1532 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1535 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1536 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1540 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8IY, 8 );
1541 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8PY, 8 );
1542 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4IY, 4 );
1543 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4PY, 4 );
1545 #define TEST_DEQUANT_DC( qname, dqname, block, w ) \
1546 if( qf_a.dqname != qf_ref.dqname ) \
1548 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
1550 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1552 for( int i = 0; i < 16; i++ ) \
1553 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16; \
1554 call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp][0]>>1, h->quant##w##_bias[block][qp][0]>>1 ); \
1555 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1556 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1557 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1558 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1561 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1563 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1564 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1568 TEST_DEQUANT_DC( quant_4x4_dc, dequant_4x4_dc, CQM_4IY, 4 );
1570 #define TEST_OPTIMIZE_CHROMA_DC( qname, optname, w ) \
1571 if( qf_a.optname != qf_ref.optname ) \
1573 set_func_name( #optname ); \
1575 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1577 int dmf = h->dequant4_mf[CQM_4IC][qp%6][0] << qp/6; \
1580 for( int i = 16; ; i <<= 1 )\
1582 int res_c, res_asm; \
1583 int max = X264_MIN( i, PIXEL_MAX*16 ); \
1584 for( int j = 0; j < w*w; j++ ) \
1585 dct1[j] = rand()%(max*2+1) - max; \
1586 call_c1( qf_c.qname, dct1, h->quant4_mf[CQM_4IC][qp][0]>>1, h->quant4_bias[CQM_4IC][qp][0]>>1 ); \
1587 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1588 res_c = call_c1( qf_c.optname, dct1, dmf ); \
1589 res_asm = call_a1( qf_a.optname, dct2, dmf ); \
1590 if( res_c != res_asm || memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1593 fprintf( stderr, #optname "(qp=%d, res_c=%d, res_asm=%d): [FAILED]\n", qp, res_c, res_asm ); \
1595 call_c2( qf_c.optname, dct1, dmf ); \
1596 call_a2( qf_a.optname, dct2, dmf ); \
1597 if( i >= PIXEL_MAX*16 ) \
1603 TEST_OPTIMIZE_CHROMA_DC( quant_2x2_dc, optimize_chroma_dc, 2 );
1605 x264_cqm_delete( h );
1608 ok = oks[0]; used_asm = used_asms[0];
1609 report( "quant :" );
1611 ok = oks[1]; used_asm = used_asms[1];
1612 report( "dequant :" );
1614 ok = oks[2]; used_asm = used_asms[2];
1615 report( "optimize chroma dc :" );
1617 ok = 1; used_asm = 0;
1618 if( qf_a.denoise_dct != qf_ref.denoise_dct )
1621 for( int size = 16; size <= 64; size += 48 )
1623 set_func_name( "denoise_dct" );
1624 memcpy( dct1, buf1, size*sizeof(dctcoef) );
1625 memcpy( dct2, buf1, size*sizeof(dctcoef) );
1626 memcpy( buf3+256, buf3, 256 );
1627 call_c1( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
1628 call_a1( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
1629 if( memcmp( dct1, dct2, size*sizeof(dctcoef) ) || memcmp( buf3+4, buf3+256+4, (size-1)*sizeof(uint32_t) ) )
1631 call_c2( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
1632 call_a2( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
1635 report( "denoise dct :" );
1637 #define TEST_DECIMATE( decname, w, ac, thresh ) \
1638 if( qf_a.decname != qf_ref.decname ) \
1640 set_func_name( #decname ); \
1642 for( int i = 0; i < 100; i++ ) \
1644 static const int distrib[16] = {1,1,1,1,1,1,1,1,1,1,1,1,2,3,4};\
1645 static const int zerorate_lut[4] = {3,7,15,31};\
1646 int zero_rate = zerorate_lut[i&3];\
1647 for( int idx = 0; idx < w*w; idx++ ) \
1649 int sign = (rand()&1) ? -1 : 1; \
1650 int abs_level = distrib[rand()&15]; \
1651 if( abs_level == 4 ) abs_level = rand()&0x3fff; \
1652 int zero = !(rand()&zero_rate); \
1653 dct1[idx] = zero * abs_level * sign; \
1657 int result_c = call_c( qf_c.decname, dct1 ); \
1658 int result_a = call_a( qf_a.decname, dct1 ); \
1659 if( X264_MIN(result_c,thresh) != X264_MIN(result_a,thresh) ) \
1662 fprintf( stderr, #decname ": [FAILED]\n" ); \
1668 ok = 1; used_asm = 0;
1669 TEST_DECIMATE( decimate_score64, 8, 0, 6 );
1670 TEST_DECIMATE( decimate_score16, 4, 0, 6 );
1671 TEST_DECIMATE( decimate_score15, 4, 1, 7 );
1672 report( "decimate_score :" );
1674 #define TEST_LAST( last, lastname, w, ac ) \
1675 if( qf_a.last != qf_ref.last ) \
1677 set_func_name( #lastname ); \
1679 for( int i = 0; i < 100; i++ ) \
1682 int max = rand() & (w*w-1); \
1683 memset( dct1, 0, w*w*sizeof(dctcoef) ); \
1684 for( int idx = ac; idx < max; idx++ ) \
1685 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
1688 int result_c = call_c( qf_c.last, dct1+ac ); \
1689 int result_a = call_a( qf_a.last, dct1+ac ); \
1690 if( result_c != result_a ) \
1693 fprintf( stderr, #lastname ": [FAILED]\n" ); \
1699 ok = 1; used_asm = 0;
1700 TEST_LAST( coeff_last[DCT_CHROMA_DC], coeff_last4, 2, 0 );
1701 TEST_LAST( coeff_last[ DCT_LUMA_AC], coeff_last15, 4, 1 );
1702 TEST_LAST( coeff_last[ DCT_LUMA_4x4], coeff_last16, 4, 0 );
1703 TEST_LAST( coeff_last[ DCT_LUMA_8x8], coeff_last64, 8, 0 );
1704 report( "coeff_last :" );
1706 #define TEST_LEVELRUN( lastname, name, w, ac ) \
1707 if( qf_a.lastname != qf_ref.lastname ) \
1709 set_func_name( #name ); \
1711 for( int i = 0; i < 100; i++ ) \
1713 x264_run_level_t runlevel_c, runlevel_a; \
1715 int max = rand() & (w*w-1); \
1716 memset( dct1, 0, w*w*sizeof(dctcoef) ); \
1717 memcpy( &runlevel_a, buf1+i, sizeof(x264_run_level_t) ); \
1718 memcpy( &runlevel_c, buf1+i, sizeof(x264_run_level_t) ); \
1719 for( int idx = ac; idx < max; idx++ ) \
1720 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
1723 int result_c = call_c( qf_c.lastname, dct1+ac, &runlevel_c ); \
1724 int result_a = call_a( qf_a.lastname, dct1+ac, &runlevel_a ); \
1725 if( result_c != result_a || runlevel_c.last != runlevel_a.last || \
1726 memcmp(runlevel_c.level, runlevel_a.level, sizeof(dctcoef)*result_c) || \
1727 memcmp(runlevel_c.run, runlevel_a.run, sizeof(uint8_t)*(result_c-1)) ) \
1730 fprintf( stderr, #name ": [FAILED]\n" ); \
1736 ok = 1; used_asm = 0;
1737 TEST_LEVELRUN( coeff_level_run[DCT_CHROMA_DC], coeff_level_run4, 2, 0 );
1738 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_AC], coeff_level_run15, 4, 1 );
1739 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_4x4], coeff_level_run16, 4, 0 );
1740 report( "coeff_level_run :" );
1745 static int check_intra( int cpu_ref, int cpu_new )
1747 int ret = 0, ok = 1, used_asm = 0;
1748 ALIGNED_16( pixel edge[33] );
1749 ALIGNED_16( pixel edge2[33] );
1750 ALIGNED_16( pixel fdec[FDEC_STRIDE*20] );
1753 x264_predict_t predict_16x16[4+3];
1754 x264_predict_t predict_8x8c[4+3];
1755 x264_predict8x8_t predict_8x8[9+3];
1756 x264_predict_t predict_4x4[9+3];
1757 x264_predict_8x8_filter_t predict_8x8_filter;
1758 } ip_c, ip_ref, ip_a;
1760 x264_predict_16x16_init( 0, ip_c.predict_16x16 );
1761 x264_predict_8x8c_init( 0, ip_c.predict_8x8c );
1762 x264_predict_8x8_init( 0, ip_c.predict_8x8, &ip_c.predict_8x8_filter );
1763 x264_predict_4x4_init( 0, ip_c.predict_4x4 );
1765 x264_predict_16x16_init( cpu_ref, ip_ref.predict_16x16 );
1766 x264_predict_8x8c_init( cpu_ref, ip_ref.predict_8x8c );
1767 x264_predict_8x8_init( cpu_ref, ip_ref.predict_8x8, &ip_ref.predict_8x8_filter );
1768 x264_predict_4x4_init( cpu_ref, ip_ref.predict_4x4 );
1770 x264_predict_16x16_init( cpu_new, ip_a.predict_16x16 );
1771 x264_predict_8x8c_init( cpu_new, ip_a.predict_8x8c );
1772 x264_predict_8x8_init( cpu_new, ip_a.predict_8x8, &ip_a.predict_8x8_filter );
1773 x264_predict_4x4_init( cpu_new, ip_a.predict_4x4 );
1775 memcpy( fdec, pbuf1, 32*20 * sizeof(pixel) );\
1777 ip_c.predict_8x8_filter( fdec+48, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
1779 #define INTRA_TEST( name, dir, w, bench, ... )\
1780 if( ip_a.name[dir] != ip_ref.name[dir] )\
1782 set_func_name( "intra_%s_%s", #name, intra_##name##_names[dir] );\
1784 memcpy( pbuf3, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
1785 memcpy( pbuf4, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
1786 call_c##bench( ip_c.name[dir], pbuf3+48, ##__VA_ARGS__ );\
1787 call_a##bench( ip_a.name[dir], pbuf4+48, ##__VA_ARGS__ );\
1788 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*20 * sizeof(pixel) ) )\
1790 fprintf( stderr, #name "[%d] : [FAILED]\n", dir );\
1792 for( int k = -1; k < 16; k++ )\
1793 printf( "%2x ", edge[16+k] );\
1795 for( int j = 0; j < w; j++ )\
1797 printf( "%2x ", edge[14-j] );\
1798 for( int k = 0; k < w; k++ )\
1799 printf( "%2x ", pbuf4[48+k+j*FDEC_STRIDE] );\
1803 for( int j = 0; j < w; j++ )\
1806 for( int k = 0; k < w; k++ )\
1807 printf( "%2x ", pbuf3[48+k+j*FDEC_STRIDE] );\
1813 for( int i = 0; i < 12; i++ )
1814 INTRA_TEST( predict_4x4, i, 4, );
1815 for( int i = 0; i < 7; i++ )
1816 INTRA_TEST( predict_8x8c, i, 8, );
1817 for( int i = 0; i < 7; i++ )
1818 INTRA_TEST( predict_16x16, i, 16, );
1819 for( int i = 0; i < 12; i++ )
1820 INTRA_TEST( predict_8x8, i, 8, , edge );
1822 set_func_name("intra_predict_8x8_filter");
1823 if( ip_a.predict_8x8_filter != ip_ref.predict_8x8_filter )
1826 for( int i = 0; i < 32; i++ )
1828 memcpy( edge2, edge, 33 * sizeof(pixel) );
1829 call_c(ip_c.predict_8x8_filter, pbuf1+48, edge, (i&24)>>1, i&7);
1830 call_a(ip_a.predict_8x8_filter, pbuf1+48, edge2, (i&24)>>1, i&7);
1831 if( memcmp( edge, edge2, 33 * sizeof(pixel) ) )
1833 fprintf( stderr, "predict_8x8_filter : [FAILED] %d %d\n", (i&24)>>1, i&7);
1839 #define EXTREMAL_PLANE(size) \
1842 for( int j = 0; j < 7; j++ ) \
1843 max[j] = test ? rand()&PIXEL_MAX : PIXEL_MAX; \
1844 fdec[48-1-FDEC_STRIDE] = (i&1)*max[0]; \
1845 for( int j = 0; j < size/2; j++ ) \
1846 fdec[48+j-FDEC_STRIDE] = (!!(i&2))*max[1]; \
1847 for( int j = size/2; j < size-1; j++ ) \
1848 fdec[48+j-FDEC_STRIDE] = (!!(i&4))*max[2]; \
1849 fdec[48+(size-1)-FDEC_STRIDE] = (!!(i&8))*max[3]; \
1850 for( int j = 0; j < size/2; j++ ) \
1851 fdec[48+j*FDEC_STRIDE-1] = (!!(i&16))*max[4]; \
1852 for( int j = size/2; j < size-1; j++ ) \
1853 fdec[48+j*FDEC_STRIDE-1] = (!!(i&32))*max[5]; \
1854 fdec[48+(size-1)*FDEC_STRIDE-1] = (!!(i&64))*max[6]; \
1856 /* Extremal test case for planar prediction. */
1857 for( int test = 0; test < 100 && ok; test++ )
1858 for( int i = 0; i < 128 && ok; i++ )
1860 EXTREMAL_PLANE( 8 );
1861 INTRA_TEST( predict_8x8c, I_PRED_CHROMA_P, 8, 1 );
1862 EXTREMAL_PLANE( 16 );
1863 INTRA_TEST( predict_16x16, I_PRED_16x16_P, 16, 1 );
1865 report( "intra pred :" );
1869 #define DECL_CABAC(cpu) \
1870 static void run_cabac_decision_##cpu( uint8_t *dst )\
1873 x264_cabac_context_init( &cb, SLICE_TYPE_P, 26, 0 );\
1874 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
1875 for( int i = 0; i < 0x1000; i++ )\
1876 x264_cabac_encode_decision_##cpu( &cb, buf1[i]>>1, buf1[i]&1 );\
1878 static void run_cabac_bypass_##cpu( uint8_t *dst )\
1881 x264_cabac_context_init( &cb, SLICE_TYPE_P, 26, 0 );\
1882 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
1883 for( int i = 0; i < 0x1000; i++ )\
1884 x264_cabac_encode_bypass_##cpu( &cb, buf1[i]&1 );\
1886 static void run_cabac_terminal_##cpu( uint8_t *dst )\
1889 x264_cabac_context_init( &cb, SLICE_TYPE_P, 26, 0 );\
1890 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
1891 for( int i = 0; i < 0x1000; i++ )\
1892 x264_cabac_encode_terminal_##cpu( &cb );\
1898 #define run_cabac_decision_asm run_cabac_decision_c
1899 #define run_cabac_bypass_asm run_cabac_bypass_c
1900 #define run_cabac_terminal_asm run_cabac_terminal_c
1903 static int check_cabac( int cpu_ref, int cpu_new )
1905 int ret = 0, ok, used_asm = 1;
1906 if( cpu_ref || run_cabac_decision_c == run_cabac_decision_asm )
1910 set_func_name( "cabac_encode_decision" );
1911 memcpy( buf4, buf3, 0x1000 );
1912 call_c( run_cabac_decision_c, buf3 );
1913 call_a( run_cabac_decision_asm, buf4 );
1914 ok = !memcmp( buf3, buf4, 0x1000 );
1915 report( "cabac decision:" );
1917 set_func_name( "cabac_encode_bypass" );
1918 memcpy( buf4, buf3, 0x1000 );
1919 call_c( run_cabac_bypass_c, buf3 );
1920 call_a( run_cabac_bypass_asm, buf4 );
1921 ok = !memcmp( buf3, buf4, 0x1000 );
1922 report( "cabac bypass:" );
1924 set_func_name( "cabac_encode_terminal" );
1925 memcpy( buf4, buf3, 0x1000 );
1926 call_c( run_cabac_terminal_c, buf3 );
1927 call_a( run_cabac_terminal_asm, buf4 );
1928 ok = !memcmp( buf3, buf4, 0x1000 );
1929 report( "cabac terminal:" );
1934 static int check_bitstream( int cpu_ref, int cpu_new )
1936 x264_bitstream_function_t bs_c;
1937 x264_bitstream_function_t bs_ref;
1938 x264_bitstream_function_t bs_a;
1940 int ret = 0, ok = 1, used_asm = 0;
1942 x264_bitstream_init( 0, &bs_c );
1943 x264_bitstream_init( cpu_ref, &bs_ref );
1944 x264_bitstream_init( cpu_new, &bs_a );
1945 if( bs_a.nal_escape != bs_ref.nal_escape )
1948 uint8_t *input = malloc(size+100);
1949 uint8_t *output1 = malloc(size*2);
1950 uint8_t *output2 = malloc(size*2);
1952 set_func_name( "nal_escape" );
1953 for( int i = 0; i < 100; i++ )
1955 /* Test corner-case sizes */
1956 int test_size = i < 10 ? i+1 : rand() & 0x3fff;
1957 /* Test 8 different probability distributions of zeros */
1958 for( int j = 0; j < test_size; j++ )
1959 input[j] = (rand()&((1 << ((i&7)+1)) - 1)) * rand();
1960 uint8_t *end_c = (uint8_t*)call_c1( bs_c.nal_escape, output1, input, input+test_size );
1961 uint8_t *end_a = (uint8_t*)call_a1( bs_a.nal_escape, output2, input, input+test_size );
1962 int size_c = end_c-output1;
1963 int size_a = end_a-output2;
1964 if( size_c != size_a || memcmp( output1, output2, size_c ) )
1966 fprintf( stderr, "nal_escape : [FAILED] %d %d\n", size_c, size_a );
1971 for( int j = 0; j < size; j++ )
1973 call_c2( bs_c.nal_escape, output1, input, input+size );
1974 call_a2( bs_a.nal_escape, output2, input, input+size );
1979 report( "nal escape:" );
1984 static int check_all_funcs( int cpu_ref, int cpu_new )
1986 return check_pixel( cpu_ref, cpu_new )
1987 + check_dct( cpu_ref, cpu_new )
1988 + check_mc( cpu_ref, cpu_new )
1989 + check_intra( cpu_ref, cpu_new )
1990 + check_deblock( cpu_ref, cpu_new )
1991 + check_quant( cpu_ref, cpu_new )
1992 + check_cabac( cpu_ref, cpu_new )
1993 + check_bitstream( cpu_ref, cpu_new );
1996 static int add_flags( int *cpu_ref, int *cpu_new, int flags, const char *name )
1998 *cpu_ref = *cpu_new;
2000 if( *cpu_new & X264_CPU_SSE2_IS_FAST )
2001 *cpu_new &= ~X264_CPU_SSE2_IS_SLOW;
2003 fprintf( stderr, "x264: %s\n", name );
2004 return check_all_funcs( *cpu_ref, *cpu_new );
2007 static int check_all_flags( void )
2010 int cpu0 = 0, cpu1 = 0;
2012 if( x264_cpu_detect() & X264_CPU_MMXEXT )
2014 ret |= add_flags( &cpu0, &cpu1, X264_CPU_MMX | X264_CPU_MMXEXT, "MMX" );
2015 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "MMX Cache64" );
2016 cpu1 &= ~X264_CPU_CACHELINE_64;
2018 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_32, "MMX Cache32" );
2019 cpu1 &= ~X264_CPU_CACHELINE_32;
2021 if( x264_cpu_detect() & X264_CPU_LZCNT )
2023 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "MMX_LZCNT" );
2024 cpu1 &= ~X264_CPU_LZCNT;
2026 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "MMX SlowCTZ" );
2027 cpu1 &= ~X264_CPU_SLOW_CTZ;
2029 if( x264_cpu_detect() & X264_CPU_SSE2 )
2031 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE | X264_CPU_SSE2 | X264_CPU_SSE2_IS_SLOW, "SSE2Slow" );
2032 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2_IS_FAST, "SSE2Fast" );
2033 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSE2Fast Cache64" );
2034 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SHUFFLE_IS_FAST, "SSE2 FastShuffle" );
2035 cpu1 &= ~X264_CPU_SHUFFLE_IS_FAST;
2036 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSE2 SlowCTZ" );
2037 cpu1 &= ~X264_CPU_SLOW_CTZ;
2038 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSE2 SlowAtom" );
2039 cpu1 &= ~X264_CPU_SLOW_ATOM;
2041 if( x264_cpu_detect() & X264_CPU_SSE_MISALIGN )
2043 cpu1 &= ~X264_CPU_CACHELINE_64;
2044 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE_MISALIGN, "SSE_Misalign" );
2045 cpu1 &= ~X264_CPU_SSE_MISALIGN;
2047 if( x264_cpu_detect() & X264_CPU_LZCNT )
2049 cpu1 &= ~X264_CPU_CACHELINE_64;
2050 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSE_LZCNT" );
2051 cpu1 &= ~X264_CPU_LZCNT;
2053 if( x264_cpu_detect() & X264_CPU_SSE3 )
2054 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE3 | X264_CPU_CACHELINE_64, "SSE3" );
2055 if( x264_cpu_detect() & X264_CPU_SSSE3 )
2057 cpu1 &= ~X264_CPU_CACHELINE_64;
2058 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSSE3, "SSSE3" );
2059 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64" );
2060 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SHUFFLE_IS_FAST, "SSSE3 FastShuffle" );
2061 cpu1 &= ~X264_CPU_SHUFFLE_IS_FAST;
2062 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSSE3 SlowCTZ" );
2063 cpu1 &= ~X264_CPU_SLOW_CTZ;
2064 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSSE3 SlowAtom" );
2065 cpu1 &= ~X264_CPU_SLOW_ATOM;
2067 if( x264_cpu_detect() & X264_CPU_SSE4 )
2069 cpu1 &= ~X264_CPU_CACHELINE_64;
2070 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE4, "SSE4" );
2072 if( x264_cpu_detect() & X264_CPU_AVX )
2073 ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX, "AVX" );
2075 if( x264_cpu_detect() & X264_CPU_ALTIVEC )
2077 fprintf( stderr, "x264: ALTIVEC against C\n" );
2078 ret = check_all_funcs( 0, X264_CPU_ALTIVEC );
2081 if( x264_cpu_detect() & X264_CPU_ARMV6 )
2082 ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV6, "ARMv6" );
2083 if( x264_cpu_detect() & X264_CPU_NEON )
2084 ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
2085 if( x264_cpu_detect() & X264_CPU_FAST_NEON_MRC )
2086 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FAST_NEON_MRC, "Fast NEON MRC" );
2091 int main(int argc, char *argv[])
2095 if( argc > 1 && !strncmp( argv[1], "--bench", 7 ) )
2097 #if !ARCH_X86 && !ARCH_X86_64 && !ARCH_PPC && !ARCH_ARM
2098 fprintf( stderr, "no --bench for your cpu until you port rdtsc\n" );
2102 if( argv[1][7] == '=' )
2104 bench_pattern = argv[1]+8;
2105 bench_pattern_len = strlen(bench_pattern);
2111 int seed = ( argc > 1 ) ? atoi(argv[1]) : x264_mdate();
2112 fprintf( stderr, "x264: using random seed %u\n", seed );
2115 buf1 = x264_malloc( 0x1e00 + 0x2000*sizeof(pixel) + 16*BENCH_ALIGNS );
2116 pbuf1 = x264_malloc( 0x1e00*sizeof(pixel) + 16*BENCH_ALIGNS );
2117 if( !buf1 || !pbuf1 )
2119 fprintf( stderr, "malloc failed, unable to initiate tests!\n" );
2122 #define INIT_POINTER_OFFSETS\
2123 buf2 = buf1 + 0xf00;\
2124 buf3 = buf2 + 0xf00;\
2125 buf4 = buf3 + 0x1000*sizeof(pixel);\
2126 pbuf2 = pbuf1 + 0xf00;\
2127 pbuf3 = (pixel*)buf3;\
2128 pbuf4 = (pixel*)buf4;
2129 INIT_POINTER_OFFSETS;
2130 for( int i = 0; i < 0x1e00; i++ )
2132 buf1[i] = rand() & 0xFF;
2133 pbuf1[i] = rand() & PIXEL_MAX;
2135 memset( buf1+0x1e00, 0, 0x2000*sizeof(pixel) );
2137 /* 16-byte alignment is guaranteed whenever it's useful, but some functions also vary in speed depending on %64 */
2139 for( int i = 0; i < BENCH_ALIGNS && !ret; i++ )
2141 INIT_POINTER_OFFSETS;
2142 ret |= x264_stack_pagealign( check_all_flags, i*16 );
2146 fprintf( stderr, "%d/%d\r", i+1, BENCH_ALIGNS );
2149 ret = check_all_flags();
2153 fprintf( stderr, "x264: at least one test has failed. Go and fix that Right Now!\n" );
2156 fprintf( stderr, "x264: All tests passed Yeah :)\n" );