1 /*****************************************************************************
2 * checkasm.c: assembly check tool
3 *****************************************************************************
4 * Copyright (C) 2003-2012 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 30 // 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[12] = { "16x16", "16x8", "8x16", "8x8", "8x4", "4x8", "4x4", "4x16", "4x2", "2x8", "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;
85 static const char **intra_predict_8x16c_names = intra_predict_8x8c_names;
87 #define set_func_name(...) snprintf( func_name, sizeof(func_name), __VA_ARGS__ )
89 static inline uint32_t read_time(void)
92 #if HAVE_X86_INLINE_ASM
93 asm volatile( "rdtsc" :"=a"(a) ::"edx" );
95 asm volatile( "mftb %0" : "=r" (a) );
96 #elif ARCH_ARM // ARMv7 only
97 asm volatile( "mrc p15, 0, %0, c9, c13, 0" : "=r"(a) );
102 static bench_t* get_bench( const char *name, int cpu )
105 for( i = 0; benchs[i].name && strcmp(name, benchs[i].name); i++ )
106 assert( i < MAX_FUNCS );
107 if( !benchs[i].name )
108 benchs[i].name = strdup( name );
110 return &benchs[i].vers[0];
111 for( j = 1; benchs[i].vers[j].cpu && benchs[i].vers[j].cpu != cpu; j++ )
112 assert( j < MAX_CPUS );
113 benchs[i].vers[j].cpu = cpu;
114 return &benchs[i].vers[j];
117 static int cmp_nop( const void *a, const void *b )
119 return *(uint16_t*)a - *(uint16_t*)b;
122 static int cmp_bench( const void *a, const void *b )
124 // asciibetical sort except preserving numbers
125 const char *sa = ((bench_func_t*)a)->name;
126 const char *sb = ((bench_func_t*)b)->name;
131 if( isdigit( *sa ) && isdigit( *sb ) && isdigit( sa[1] ) != isdigit( sb[1] ) )
132 return isdigit( sa[1] ) - isdigit( sb[1] );
138 static void print_bench(void)
140 uint16_t nops[10000] = {0};
141 int nfuncs, nop_time=0;
143 for( int i = 0; i < 10000; i++ )
146 nops[i] = read_time() - t;
148 qsort( nops, 10000, sizeof(uint16_t), cmp_nop );
149 for( int i = 500; i < 9500; i++ )
152 printf( "nop: %d\n", nop_time );
154 for( nfuncs = 0; nfuncs < MAX_FUNCS && benchs[nfuncs].name; nfuncs++ );
155 qsort( benchs, nfuncs, sizeof(bench_func_t), cmp_bench );
156 for( int i = 0; i < nfuncs; i++ )
157 for( int j = 0; j < MAX_CPUS && (!j || benchs[i].vers[j].cpu); j++ )
160 bench_t *b = &benchs[i].vers[j];
163 for( k = 0; k < j && benchs[i].vers[k].pointer != b->pointer; k++ );
166 printf( "%s_%s%s: %"PRId64"\n", benchs[i].name,
167 b->cpu&X264_CPU_AVX2 ? "avx2" :
168 b->cpu&X264_CPU_FMA3 ? "fma3" :
169 b->cpu&X264_CPU_FMA4 ? "fma4" :
170 b->cpu&X264_CPU_XOP ? "xop" :
171 b->cpu&X264_CPU_AVX ? "avx" :
172 b->cpu&X264_CPU_SSE4 ? "sse4" :
173 b->cpu&X264_CPU_SSSE3 ? "ssse3" :
174 b->cpu&X264_CPU_SSE3 ? "sse3" :
175 /* print sse2slow only if there's also a sse2fast version of the same func */
176 b->cpu&X264_CPU_SSE2_IS_SLOW && j<MAX_CPUS-1 && b[1].cpu&X264_CPU_SSE2_IS_FAST && !(b[1].cpu&X264_CPU_SSE3) ? "sse2slow" :
177 b->cpu&X264_CPU_SSE2 ? "sse2" :
178 b->cpu&X264_CPU_MMX ? "mmx" :
179 b->cpu&X264_CPU_ALTIVEC ? "altivec" :
180 b->cpu&X264_CPU_NEON ? "neon" :
181 b->cpu&X264_CPU_ARMV6 ? "armv6" : "c",
182 b->cpu&X264_CPU_CACHELINE_32 ? "_c32" :
183 b->cpu&X264_CPU_CACHELINE_64 ? "_c64" :
184 b->cpu&X264_CPU_SHUFFLE_IS_FAST && !(b->cpu&X264_CPU_SSE4) ? "_fastshuffle" :
185 b->cpu&X264_CPU_SSE_MISALIGN ? "_misalign" :
186 b->cpu&X264_CPU_LZCNT ? "_lzcnt" :
187 b->cpu&X264_CPU_BMI2 ? "_bmi2" :
188 b->cpu&X264_CPU_TBM ? "_tbm" :
189 b->cpu&X264_CPU_BMI1 ? "_bmi1" :
190 b->cpu&X264_CPU_FAST_NEON_MRC ? "_fast_mrc" :
191 b->cpu&X264_CPU_SLOW_CTZ ? "_slow_ctz" :
192 b->cpu&X264_CPU_SLOW_ATOM ? "_slow_atom" : "",
193 ((int64_t)10*b->cycles/b->den - nop_time)/4 );
197 #if ARCH_X86 || ARCH_X86_64
198 int x264_stack_pagealign( int (*func)(), int align );
200 /* detect when callee-saved regs aren't saved
201 * needs an explicit asm check because it only sometimes crashes in normal use. */
202 intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
204 #define x264_stack_pagealign( func, align ) func()
207 #define call_c1(func,...) func(__VA_ARGS__)
210 /* Evil hack: detect incorrect assumptions that 32-bit ints are zero-extended to 64-bit.
211 * This is done by clobbering the stack with junk around the stack pointer and calling the
212 * assembly function through x264_checkasm_call with added dummy arguments which forces all
213 * real arguments to be passed on the stack and not in registers. For 32-bit argument the
214 * upper half of the 64-bit register location on the stack will now contain junk. Note that
215 * this is dependant on compiler behaviour and that interrupts etc. at the wrong time may
216 * overwrite the junk written to the stack so there's no guarantee that it will always
217 * detect all functions that assumes zero-extension.
219 void x264_checkasm_stack_clobber( uint64_t clobber, ... );
220 #define call_a1(func,...) ({ \
221 uint64_t r = (rand() & 0xffff) * 0x0001000100010001ULL; \
222 x264_checkasm_stack_clobber( r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r ); /* max_args+6 */ \
223 x264_checkasm_call(( intptr_t(*)())func, &ok, 0, 0, 0, 0, __VA_ARGS__ ); })
225 #define call_a1(func,...) x264_checkasm_call( (intptr_t(*)())func, &ok, __VA_ARGS__ )
227 #define call_a1 call_c1
230 #define call_bench(func,cpu,...)\
231 if( do_bench && !strncmp(func_name, bench_pattern, bench_pattern_len) )\
235 call_a1(func, __VA_ARGS__);\
236 for( int ti = 0; ti < (cpu?BENCH_RUNS:BENCH_RUNS/4); ti++ )\
238 uint32_t t = read_time();\
243 t = read_time() - t;\
244 if( t*tcount <= tsum*4 && ti > 0 )\
250 bench_t *b = get_bench( func_name, cpu );\
256 /* for most functions, run benchmark and correctness test at the same time.
257 * for those that modify their inputs, run the above macros separately */
258 #define call_a(func,...) ({ call_a2(func,__VA_ARGS__); call_a1(func,__VA_ARGS__); })
259 #define call_c(func,...) ({ call_c2(func,__VA_ARGS__); call_c1(func,__VA_ARGS__); })
260 #define call_a2(func,...) ({ call_bench(func,cpu_new,__VA_ARGS__); })
261 #define call_c2(func,...) ({ call_bench(func,0,__VA_ARGS__); })
264 static int check_pixel( int cpu_ref, int cpu_new )
266 x264_pixel_function_t pixel_c;
267 x264_pixel_function_t pixel_ref;
268 x264_pixel_function_t pixel_asm;
269 x264_predict_t predict_4x4[12];
270 x264_predict8x8_t predict_8x8[12];
271 x264_predict_8x8_filter_t predict_8x8_filter;
272 ALIGNED_16( pixel edge[36] );
273 uint16_t cost_mv[32];
274 int ret = 0, ok, used_asm;
276 x264_pixel_init( 0, &pixel_c );
277 x264_pixel_init( cpu_ref, &pixel_ref );
278 x264_pixel_init( cpu_new, &pixel_asm );
279 x264_predict_4x4_init( 0, predict_4x4 );
280 x264_predict_8x8_init( 0, predict_8x8, &predict_8x8_filter );
281 predict_8x8_filter( pbuf2+40, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
284 for( int i = 0; i < 256; i++ )
289 pbuf4[i] = -(z&1) & PIXEL_MAX;
290 pbuf3[i] = ~pbuf4[i] & PIXEL_MAX;
292 // random pattern made of maxed pixel differences, in case an intermediate value overflows
293 for( int i = 256; i < 0x1000; i++ )
295 pbuf4[i] = -(pbuf1[i&~0x88]&1) & PIXEL_MAX;
296 pbuf3[i] = ~(pbuf4[i]) & PIXEL_MAX;
299 #define TEST_PIXEL( name, align ) \
300 ok = 1, used_asm = 0; \
301 for( int i = 0; i < 8; i++ ) \
303 int res_c, res_asm; \
304 if( pixel_asm.name[i] != pixel_ref.name[i] ) \
306 set_func_name( "%s_%s", #name, pixel_names[i] ); \
308 for( int j = 0; j < 64; j++ ) \
310 res_c = call_c( pixel_c.name[i], pbuf1, (intptr_t)16, pbuf2+j*!align, (intptr_t)64 ); \
311 res_asm = call_a( pixel_asm.name[i], pbuf1, (intptr_t)16, pbuf2+j*!align, (intptr_t)64 ); \
312 if( res_c != res_asm ) \
315 fprintf( stderr, #name "[%d]: %d != %d [FAILED]\n", i, res_c, res_asm ); \
319 for( int j = 0; j < 0x1000 && ok; j += 256 ) \
321 res_c = pixel_c .name[i]( pbuf3+j, 16, pbuf4+j, 16 ); \
322 res_asm = pixel_asm.name[i]( pbuf3+j, 16, pbuf4+j, 16 ); \
323 if( res_c != res_asm ) \
326 fprintf( stderr, #name "[%d]: overflow %d != %d\n", i, res_c, res_asm ); \
331 report( "pixel " #name " :" );
333 TEST_PIXEL( sad, 0 );
334 TEST_PIXEL( sad_aligned, 1 );
335 TEST_PIXEL( ssd, 1 );
336 TEST_PIXEL( satd, 0 );
337 TEST_PIXEL( sa8d, 1 );
339 #define TEST_PIXEL_X( N ) \
340 ok = 1; used_asm = 0; \
341 for( int i = 0; i < 7; i++ ) \
343 int res_c[4]={0}, res_asm[4]={0}; \
344 if( pixel_asm.sad_x##N[i] && pixel_asm.sad_x##N[i] != pixel_ref.sad_x##N[i] ) \
346 set_func_name( "sad_x%d_%s", N, pixel_names[i] ); \
348 for( int j = 0; j < 64; j++ ) \
350 pixel *pix2 = pbuf2+j; \
351 res_c[0] = pixel_c.sad[i]( pbuf1, 16, pix2, 64 ); \
352 res_c[1] = pixel_c.sad[i]( pbuf1, 16, pix2+6, 64 ); \
353 res_c[2] = pixel_c.sad[i]( pbuf1, 16, pix2+1, 64 ); \
356 res_c[3] = pixel_c.sad[i]( pbuf1, 16, pix2+10, 64 ); \
357 call_a( pixel_asm.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, (intptr_t)64, res_asm ); \
360 call_a( pixel_asm.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, (intptr_t)64, res_asm ); \
361 if( memcmp(res_c, res_asm, sizeof(res_c)) ) \
364 fprintf( stderr, "sad_x"#N"[%d]: %d,%d,%d,%d != %d,%d,%d,%d [FAILED]\n", \
365 i, res_c[0], res_c[1], res_c[2], res_c[3], \
366 res_asm[0], res_asm[1], res_asm[2], res_asm[3] ); \
369 call_c2( pixel_c.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, (intptr_t)64, res_asm ); \
371 call_c2( pixel_c.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, (intptr_t)64, res_asm ); \
375 report( "pixel sad_x"#N" :" );
380 #define TEST_PIXEL_VAR( i ) \
381 if( pixel_asm.var[i] != pixel_ref.var[i] ) \
383 set_func_name( "%s_%s", "var", pixel_names[i] ); \
385 /* abi-check wrapper can't return uint64_t, so separate it from return value check */ \
386 call_c1( pixel_c.var[i], pbuf1, 16 ); \
387 call_a1( pixel_asm.var[i], pbuf1, (intptr_t)16 ); \
388 uint64_t res_c = pixel_c.var[i]( pbuf1, 16 ); \
389 uint64_t res_asm = pixel_asm.var[i]( pbuf1, 16 ); \
390 if( res_c != res_asm ) \
393 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) ); \
395 call_c2( pixel_c.var[i], pbuf1, (intptr_t)16 ); \
396 call_a2( pixel_asm.var[i], pbuf1, (intptr_t)16 ); \
399 ok = 1; used_asm = 0;
400 TEST_PIXEL_VAR( PIXEL_16x16 );
401 TEST_PIXEL_VAR( PIXEL_8x16 );
402 TEST_PIXEL_VAR( PIXEL_8x8 );
403 report( "pixel var :" );
405 #define TEST_PIXEL_VAR2( i ) \
406 if( pixel_asm.var2[i] != pixel_ref.var2[i] ) \
408 int res_c, res_asm, ssd_c, ssd_asm; \
409 set_func_name( "%s_%s", "var2", pixel_names[i] ); \
411 res_c = call_c( pixel_c.var2[i], pbuf1, (intptr_t)16, pbuf2, (intptr_t)16, &ssd_c ); \
412 res_asm = call_a( pixel_asm.var2[i], pbuf1, (intptr_t)16, pbuf2, (intptr_t)16, &ssd_asm ); \
413 if( res_c != res_asm || ssd_c != ssd_asm ) \
416 fprintf( stderr, "var2[%d]: %d != %d or %d != %d [FAILED]\n", i, res_c, res_asm, ssd_c, ssd_asm ); \
420 ok = 1; used_asm = 0;
421 TEST_PIXEL_VAR2( PIXEL_8x16 );
422 TEST_PIXEL_VAR2( PIXEL_8x8 );
423 report( "pixel var2 :" );
425 ok = 1; used_asm = 0;
426 for( int i = 0; i < 4; i++ )
427 if( pixel_asm.hadamard_ac[i] != pixel_ref.hadamard_ac[i] )
429 set_func_name( "hadamard_ac_%s", pixel_names[i] );
431 for( int j = 0; j < 32; j++ )
433 pixel *pix = (j&16 ? pbuf1 : pbuf3) + (j&15)*256;
434 call_c1( pixel_c.hadamard_ac[i], pbuf1, (intptr_t)16 );
435 call_a1( pixel_asm.hadamard_ac[i], pbuf1, (intptr_t)16 );
436 uint64_t rc = pixel_c.hadamard_ac[i]( pix, 16 );
437 uint64_t ra = pixel_asm.hadamard_ac[i]( pix, 16 );
441 fprintf( stderr, "hadamard_ac[%d]: %d,%d != %d,%d\n", i, (int)rc, (int)(rc>>32), (int)ra, (int)(ra>>32) );
445 call_c2( pixel_c.hadamard_ac[i], pbuf1, (intptr_t)16 );
446 call_a2( pixel_asm.hadamard_ac[i], pbuf1, (intptr_t)16 );
448 report( "pixel hadamard_ac :" );
451 for( int i = 0; i < 32; i++ )
452 for( int j = 0; j < 16; j++ )
453 pbuf4[16*i+j] = -((i+j)&1) & PIXEL_MAX;
454 ok = 1; used_asm = 0;
455 if( pixel_asm.vsad != pixel_ref.vsad )
457 for( int h = 2; h <= 32; h += 2 )
460 set_func_name( "vsad" );
462 for( int j = 0; j < 2 && ok; j++ )
464 pixel *p = j ? pbuf4 : pbuf1;
465 res_c = call_c( pixel_c.vsad, p, (intptr_t)16, h );
466 res_asm = call_a( pixel_asm.vsad, p, (intptr_t)16, h );
467 if( res_c != res_asm )
470 fprintf( stderr, "vsad: height=%d, %d != %d\n", h, res_c, res_asm );
476 report( "pixel vsad :" );
478 ok = 1; used_asm = 0;
479 if( pixel_asm.asd8 != pixel_ref.asd8 )
481 set_func_name( "asd8" );
483 int res_c = call_c( pixel_c.asd8, pbuf1, (intptr_t)8, pbuf2, (intptr_t)8, 16 );
484 int res_a = call_a( pixel_asm.asd8, pbuf1, (intptr_t)8, pbuf2, (intptr_t)8, 16 );
488 fprintf( stderr, "asd: %d != %d\n", res_c, res_a );
491 report( "pixel asd :" );
493 #define TEST_INTRA_X3( name, i8x8, ... ) \
494 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
496 int res_c[3], res_asm[3]; \
497 set_func_name( #name ); \
499 call_c( pixel_c.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_c ); \
500 call_a( pixel_asm.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_asm ); \
501 if( memcmp(res_c, res_asm, sizeof(res_c)) ) \
504 fprintf( stderr, #name": %d,%d,%d != %d,%d,%d [FAILED]\n", \
505 res_c[0], res_c[1], res_c[2], \
506 res_asm[0], res_asm[1], res_asm[2] ); \
510 #define TEST_INTRA_X9( name, cmp ) \
511 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
513 set_func_name( #name ); \
515 ALIGNED_ARRAY_64( uint16_t, bitcosts,[17] ); \
516 for( int i=0; i<17; i++ ) \
517 bitcosts[i] = 9*(i!=8); \
518 memcpy( pbuf3, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) ); \
519 memcpy( pbuf4, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) ); \
520 for( int i=0; i<32; i++ ) \
522 pixel *fenc = pbuf1+48+i*12; \
523 pixel *fdec1 = pbuf3+48+i*12; \
524 pixel *fdec2 = pbuf4+48+i*12; \
525 int pred_mode = i%9; \
526 int res_c = INT_MAX; \
527 for( int j=0; j<9; j++ ) \
529 predict_4x4[j]( fdec1 ); \
530 int cost = pixel_c.cmp[PIXEL_4x4]( fenc, FENC_STRIDE, fdec1, FDEC_STRIDE ) + 9*(j!=pred_mode); \
531 if( cost < (uint16_t)res_c ) \
532 res_c = cost + (j<<16); \
534 predict_4x4[res_c>>16]( fdec1 ); \
535 int res_a = call_a( pixel_asm.name, fenc, fdec2, bitcosts+8-pred_mode ); \
536 if( res_c != res_a ) \
539 fprintf( stderr, #name": %d,%d != %d,%d [FAILED]\n", res_c>>16, res_c&0xffff, res_a>>16, res_a&0xffff ); \
542 if( memcmp(fdec1, fdec2, 4*FDEC_STRIDE*sizeof(pixel)) ) \
545 fprintf( stderr, #name" [FAILED]\n" ); \
546 for( int j=0; j<16; j++ ) \
547 fprintf( stderr, "%02x ", fdec1[(j&3)+(j>>2)*FDEC_STRIDE] ); \
548 fprintf( stderr, "\n" ); \
549 for( int j=0; j<16; j++ ) \
550 fprintf( stderr, "%02x ", fdec2[(j&3)+(j>>2)*FDEC_STRIDE] ); \
551 fprintf( stderr, "\n" ); \
557 #define TEST_INTRA8_X9( name, cmp ) \
558 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
560 set_func_name( #name ); \
562 ALIGNED_ARRAY_64( uint16_t, bitcosts,[17] ); \
563 ALIGNED_ARRAY_16( uint16_t, satds_c,[16] ); \
564 ALIGNED_ARRAY_16( uint16_t, satds_a,[16] ); \
565 memset( satds_c, 0, 16 * sizeof(*satds_c) ); \
566 memset( satds_a, 0, 16 * sizeof(*satds_a) ); \
567 for( int i=0; i<17; i++ ) \
568 bitcosts[i] = 9*(i!=8); \
569 for( int i=0; i<32; i++ ) \
571 pixel *fenc = pbuf1+48+i*12; \
572 pixel *fdec1 = pbuf3+48+i*12; \
573 pixel *fdec2 = pbuf4+48+i*12; \
574 int pred_mode = i%9; \
575 int res_c = INT_MAX; \
576 predict_8x8_filter( fdec1, edge, ALL_NEIGHBORS, ALL_NEIGHBORS ); \
577 for( int j=0; j<9; j++ ) \
579 predict_8x8[j]( fdec1, edge ); \
580 satds_c[j] = pixel_c.cmp[PIXEL_8x8]( fenc, FENC_STRIDE, fdec1, FDEC_STRIDE ) + 9*(j!=pred_mode); \
581 if( satds_c[j] < (uint16_t)res_c ) \
582 res_c = satds_c[j] + (j<<16); \
584 predict_8x8[res_c>>16]( fdec1, edge ); \
585 int res_a = call_a( pixel_asm.name, fenc, fdec2, edge, bitcosts+8-pred_mode, satds_a ); \
586 if( res_c != res_a || memcmp(satds_c, satds_a, sizeof(satds_c)) ) \
589 fprintf( stderr, #name": %d,%d != %d,%d [FAILED]\n", res_c>>16, res_c&0xffff, res_a>>16, res_a&0xffff ); \
590 for( int j = 0; j < 9; j++ ) \
591 fprintf( stderr, "%5d ", satds_c[j]); \
592 fprintf( stderr, "\n" ); \
593 for( int j = 0; j < 9; j++ ) \
594 fprintf( stderr, "%5d ", satds_a[j]); \
595 fprintf( stderr, "\n" ); \
598 for( int j=0; j<8; j++ ) \
599 if( memcmp(fdec1+j*FDEC_STRIDE, fdec2+j*FDEC_STRIDE, 8*sizeof(pixel)) ) \
603 fprintf( stderr, #name" [FAILED]\n" ); \
604 for( int j=0; j<8; j++ ) \
606 for( int k=0; k<8; k++ ) \
607 fprintf( stderr, "%02x ", fdec1[k+j*FDEC_STRIDE] ); \
608 fprintf( stderr, "\n" ); \
610 fprintf( stderr, "\n" ); \
611 for( int j=0; j<8; j++ ) \
613 for( int k=0; k<8; k++ ) \
614 fprintf( stderr, "%02x ", fdec2[k+j*FDEC_STRIDE] ); \
615 fprintf( stderr, "\n" ); \
617 fprintf( stderr, "\n" ); \
623 memcpy( pbuf3, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) );
624 ok = 1; used_asm = 0;
625 TEST_INTRA_X3( intra_satd_x3_16x16, 0 );
626 TEST_INTRA_X3( intra_satd_x3_8x16c, 0 );
627 TEST_INTRA_X3( intra_satd_x3_8x8c, 0 );
628 TEST_INTRA_X3( intra_sa8d_x3_8x8, 1, edge );
629 TEST_INTRA_X3( intra_satd_x3_4x4, 0 );
630 report( "intra satd_x3 :" );
631 ok = 1; used_asm = 0;
632 TEST_INTRA_X3( intra_sad_x3_16x16, 0 );
633 TEST_INTRA_X3( intra_sad_x3_8x16c, 0 );
634 TEST_INTRA_X3( intra_sad_x3_8x8c, 0 );
635 TEST_INTRA_X3( intra_sad_x3_8x8, 1, edge );
636 TEST_INTRA_X3( intra_sad_x3_4x4, 0 );
637 report( "intra sad_x3 :" );
638 ok = 1; used_asm = 0;
639 TEST_INTRA_X9( intra_satd_x9_4x4, satd );
640 TEST_INTRA8_X9( intra_sa8d_x9_8x8, sa8d );
641 report( "intra satd_x9 :" );
642 ok = 1; used_asm = 0;
643 TEST_INTRA_X9( intra_sad_x9_4x4, sad );
644 TEST_INTRA8_X9( intra_sad_x9_8x8, sad );
645 report( "intra sad_x9 :" );
647 ok = 1; used_asm = 0;
648 if( pixel_asm.ssd_nv12_core != pixel_ref.ssd_nv12_core )
651 set_func_name( "ssd_nv12" );
652 uint64_t res_u_c, res_v_c, res_u_a, res_v_a;
653 pixel_c.ssd_nv12_core( pbuf1, 368, pbuf2, 368, 360, 8, &res_u_c, &res_v_c );
654 pixel_asm.ssd_nv12_core( pbuf1, 368, pbuf2, 368, 360, 8, &res_u_a, &res_v_a );
655 if( res_u_c != res_u_a || res_v_c != res_v_a )
658 fprintf( stderr, "ssd_nv12: %"PRIu64",%"PRIu64" != %"PRIu64",%"PRIu64"\n",
659 res_u_c, res_v_c, res_u_a, res_v_a );
661 call_c( pixel_c.ssd_nv12_core, pbuf1, (intptr_t)368, pbuf2, (intptr_t)368, 360, 8, &res_u_c, &res_v_c );
662 call_a( pixel_asm.ssd_nv12_core, pbuf1, (intptr_t)368, pbuf2, (intptr_t)368, 360, 8, &res_u_a, &res_v_a );
664 report( "ssd_nv12 :" );
666 if( pixel_asm.ssim_4x4x2_core != pixel_ref.ssim_4x4x2_core ||
667 pixel_asm.ssim_end4 != pixel_ref.ssim_end4 )
671 ALIGNED_16( int sums[5][4] ) = {{0}};
674 res_c = x264_pixel_ssim_wxh( &pixel_c, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
675 res_a = x264_pixel_ssim_wxh( &pixel_asm, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
676 if( fabs( res_c - res_a ) > 1e-6 )
679 fprintf( stderr, "ssim: %.7f != %.7f [FAILED]\n", res_c, res_a );
681 set_func_name( "ssim_core" );
682 call_c2( pixel_c.ssim_4x4x2_core, pbuf1+2, (intptr_t)32, pbuf2+2, (intptr_t)32, sums );
683 call_a2( pixel_asm.ssim_4x4x2_core, pbuf1+2, (intptr_t)32, pbuf2+2, (intptr_t)32, sums );
684 set_func_name( "ssim_end" );
685 call_c2( pixel_c.ssim_end4, sums, sums, 4 );
686 call_a2( pixel_asm.ssim_end4, sums, sums, 4 );
690 ok = 1; used_asm = 0;
691 for( int i = 0; i < 32; i++ )
693 for( int i = 0; i < 100 && ok; i++ )
694 if( pixel_asm.ads[i&3] != pixel_ref.ads[i&3] )
696 ALIGNED_16( uint16_t sums[72] );
697 ALIGNED_16( int dc[4] );
698 ALIGNED_16( int16_t mvs_a[32] );
699 ALIGNED_16( int16_t mvs_c[32] );
701 int thresh = rand() & 0x3fff;
702 set_func_name( "esa_ads" );
703 for( int j = 0; j < 72; j++ )
704 sums[j] = rand() & 0x3fff;
705 for( int j = 0; j < 4; j++ )
706 dc[j] = rand() & 0x3fff;
708 mvn_c = call_c( pixel_c.ads[i&3], dc, sums, 32, cost_mv, mvs_c, 28, thresh );
709 mvn_a = call_a( pixel_asm.ads[i&3], dc, sums, 32, cost_mv, mvs_a, 28, thresh );
710 if( mvn_c != mvn_a || memcmp( mvs_c, mvs_a, mvn_c*sizeof(*mvs_c) ) )
713 printf( "c%d: ", i&3 );
714 for( int j = 0; j < mvn_c; j++ )
715 printf( "%d ", mvs_c[j] );
716 printf( "\na%d: ", i&3 );
717 for( int j = 0; j < mvn_a; j++ )
718 printf( "%d ", mvs_a[j] );
722 report( "esa ads:" );
727 static int check_dct( int cpu_ref, int cpu_new )
729 x264_dct_function_t dct_c;
730 x264_dct_function_t dct_ref;
731 x264_dct_function_t dct_asm;
732 x264_quant_function_t qf;
733 int ret = 0, ok, used_asm, interlace = 0;
734 ALIGNED_16( dctcoef dct1[16][16] );
735 ALIGNED_16( dctcoef dct2[16][16] );
736 ALIGNED_16( dctcoef dct4[16][16] );
737 ALIGNED_16( dctcoef dct8[4][64] );
738 ALIGNED_16( dctcoef dctdc[2][8] );
742 x264_dct_init( 0, &dct_c );
743 x264_dct_init( cpu_ref, &dct_ref);
744 x264_dct_init( cpu_new, &dct_asm );
746 memset( h, 0, sizeof(*h) );
747 x264_param_default( &h->param );
748 h->sps->i_chroma_format_idc = 1;
749 h->chroma_qp_table = i_chroma_qp_table + 12;
750 h->param.analyse.i_luma_deadzone[0] = 0;
751 h->param.analyse.i_luma_deadzone[1] = 0;
752 h->param.analyse.b_transform_8x8 = 1;
753 for( int i = 0; i < 6; i++ )
754 h->pps->scaling_list[i] = x264_cqm_flat16;
756 x264_quant_init( h, 0, &qf );
758 /* overflow test cases */
759 for( int i = 0; i < 5; i++ )
761 pixel *enc = &pbuf3[16*i*FENC_STRIDE];
762 pixel *dec = &pbuf4[16*i*FDEC_STRIDE];
764 for( int j = 0; j < 16; j++ )
766 int cond_a = (i < 2) ? 1 : ((j&3) == 0 || (j&3) == (i-1));
767 int cond_b = (i == 0) ? 1 : !cond_a;
768 enc[0] = enc[1] = enc[4] = enc[5] = enc[8] = enc[9] = enc[12] = enc[13] = cond_a ? PIXEL_MAX : 0;
769 enc[2] = enc[3] = enc[6] = enc[7] = enc[10] = enc[11] = enc[14] = enc[15] = cond_b ? PIXEL_MAX : 0;
771 for( int k = 0; k < 4; k++ )
772 dec[k] = PIXEL_MAX - enc[k];
779 #define TEST_DCT( name, t1, t2, size ) \
780 if( dct_asm.name != dct_ref.name ) \
782 set_func_name( #name ); \
784 pixel *enc = pbuf3; \
785 pixel *dec = pbuf4; \
786 for( int j = 0; j < 5; j++) \
788 call_c( dct_c.name, t1, &pbuf1[j*64], &pbuf2[j*64] ); \
789 call_a( dct_asm.name, t2, &pbuf1[j*64], &pbuf2[j*64] ); \
790 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
793 fprintf( stderr, #name " [FAILED]\n" ); \
794 for( int k = 0; k < size; k++ )\
795 printf( "%d ", ((dctcoef*)t1)[k] );\
797 for( int k = 0; k < size; k++ )\
798 printf( "%d ", ((dctcoef*)t2)[k] );\
802 call_c( dct_c.name, t1, enc, dec ); \
803 call_a( dct_asm.name, t2, enc, dec ); \
804 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
807 fprintf( stderr, #name " [FAILED] (overflow)\n" ); \
810 enc += 16*FENC_STRIDE; \
811 dec += 16*FDEC_STRIDE; \
814 ok = 1; used_asm = 0;
815 TEST_DCT( sub4x4_dct, dct1[0], dct2[0], 16 );
816 TEST_DCT( sub8x8_dct, dct1, dct2, 16*4 );
817 TEST_DCT( sub8x8_dct_dc, dctdc[0], dctdc[1], 4 );
818 TEST_DCT( sub8x16_dct_dc, dctdc[0], dctdc[1], 8 );
819 TEST_DCT( sub16x16_dct, dct1, dct2, 16*16 );
820 report( "sub_dct4 :" );
822 ok = 1; used_asm = 0;
823 TEST_DCT( sub8x8_dct8, (void*)dct1[0], (void*)dct2[0], 64 );
824 TEST_DCT( sub16x16_dct8, (void*)dct1, (void*)dct2, 64*4 );
825 report( "sub_dct8 :" );
828 // fdct and idct are denormalized by different factors, so quant/dequant
829 // is needed to force the coefs into the right range.
830 dct_c.sub16x16_dct( dct4, pbuf1, pbuf2 );
831 dct_c.sub16x16_dct8( dct8, pbuf1, pbuf2 );
832 for( int i = 0; i < 16; i++ )
834 qf.quant_4x4( dct4[i], h->quant4_mf[CQM_4IY][20], h->quant4_bias[CQM_4IY][20] );
835 qf.dequant_4x4( dct4[i], h->dequant4_mf[CQM_4IY], 20 );
837 for( int i = 0; i < 4; i++ )
839 qf.quant_8x8( dct8[i], h->quant8_mf[CQM_8IY][20], h->quant8_bias[CQM_8IY][20] );
840 qf.dequant_8x8( dct8[i], h->dequant8_mf[CQM_8IY], 20 );
842 x264_cqm_delete( h );
844 #define TEST_IDCT( name, src ) \
845 if( dct_asm.name != dct_ref.name ) \
847 set_func_name( #name ); \
849 memcpy( pbuf3, pbuf1, 32*32 * sizeof(pixel) ); \
850 memcpy( pbuf4, pbuf1, 32*32 * sizeof(pixel) ); \
851 memcpy( dct1, src, 256 * sizeof(dctcoef) ); \
852 memcpy( dct2, src, 256 * sizeof(dctcoef) ); \
853 call_c1( dct_c.name, pbuf3, (void*)dct1 ); \
854 call_a1( dct_asm.name, pbuf4, (void*)dct2 ); \
855 if( memcmp( pbuf3, pbuf4, 32*32 * sizeof(pixel) ) ) \
858 fprintf( stderr, #name " [FAILED]\n" ); \
860 call_c2( dct_c.name, pbuf3, (void*)dct1 ); \
861 call_a2( dct_asm.name, pbuf4, (void*)dct2 ); \
863 ok = 1; used_asm = 0;
864 TEST_IDCT( add4x4_idct, dct4 );
865 TEST_IDCT( add8x8_idct, dct4 );
866 TEST_IDCT( add8x8_idct_dc, dct4 );
867 TEST_IDCT( add16x16_idct, dct4 );
868 TEST_IDCT( add16x16_idct_dc, dct4 );
869 report( "add_idct4 :" );
871 ok = 1; used_asm = 0;
872 TEST_IDCT( add8x8_idct8, dct8 );
873 TEST_IDCT( add16x16_idct8, dct8 );
874 report( "add_idct8 :" );
877 #define TEST_DCTDC( name )\
878 ok = 1; used_asm = 0;\
879 if( dct_asm.name != dct_ref.name )\
881 set_func_name( #name );\
883 uint16_t *p = (uint16_t*)buf1;\
884 for( int i = 0; i < 16 && ok; i++ )\
886 for( int j = 0; j < 16; j++ )\
887 dct1[0][j] = !i ? (j^j>>1^j>>2^j>>3)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
888 : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
889 : ((*p++)&0x1fff)-0x1000; /* general case */\
890 memcpy( dct2, dct1, 16 * sizeof(dctcoef) );\
891 call_c1( dct_c.name, dct1[0] );\
892 call_a1( dct_asm.name, dct2[0] );\
893 if( memcmp( dct1, dct2, 16 * sizeof(dctcoef) ) )\
896 call_c2( dct_c.name, dct1[0] );\
897 call_a2( dct_asm.name, dct2[0] );\
899 report( #name " :" );
901 TEST_DCTDC( dct4x4dc );
902 TEST_DCTDC( idct4x4dc );
905 #define TEST_DCTDC_CHROMA( name )\
906 ok = 1; used_asm = 0;\
907 if( dct_asm.name != dct_ref.name )\
909 set_func_name( #name );\
911 uint16_t *p = (uint16_t*)buf1;\
912 for( int i = 0; i < 16 && ok; i++ )\
914 for( int j = 0; j < 8; j++ )\
915 dct1[j][0] = !i ? (j^j>>1^j>>2)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
916 : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
917 : ((*p++)&0x1fff)-0x1000; /* general case */\
918 memcpy( dct2, dct1, 8*16 * sizeof(dctcoef) );\
919 call_c1( dct_c.name, dctdc[0], dct1 );\
920 call_a1( dct_asm.name, dctdc[1], dct2 );\
921 if( memcmp( dctdc[0], dctdc[1], 8 * sizeof(dctcoef) ) || memcmp( dct1, dct2, 8*16 * sizeof(dctcoef) ) )\
924 fprintf( stderr, #name " [FAILED]\n" ); \
927 call_c2( dct_c.name, dctdc[0], dct1 );\
928 call_a2( dct_asm.name, dctdc[1], dct2 );\
930 report( #name " :" );
932 TEST_DCTDC_CHROMA( dct2x4dc );
933 #undef TEST_DCTDC_CHROMA
935 x264_zigzag_function_t zigzag_c[2];
936 x264_zigzag_function_t zigzag_ref[2];
937 x264_zigzag_function_t zigzag_asm[2];
939 ALIGNED_16( dctcoef level1[64] );
940 ALIGNED_16( dctcoef level2[64] );
942 #define TEST_ZIGZAG_SCAN( name, t1, t2, dct, size ) \
943 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
945 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
947 for( int i = 0; i < size*size; i++ ) \
949 call_c( zigzag_c[interlace].name, t1, dct ); \
950 call_a( zigzag_asm[interlace].name, t2, dct ); \
951 if( memcmp( t1, t2, size*size*sizeof(dctcoef) ) ) \
954 for( int i = 0; i < 2; i++ ) \
956 dctcoef *d = (dctcoef*)(i ? t2 : t1); \
957 for( int j = 0; j < size; j++ ) \
959 for( int k = 0; k < size; k++ ) \
960 fprintf( stderr, "%2d ", d[k+j*8] ); \
961 fprintf( stderr, "\n" ); \
963 fprintf( stderr, "\n" ); \
965 fprintf( stderr, #name " [FAILED]\n" ); \
969 #define TEST_ZIGZAG_SUB( name, t1, t2, size ) \
970 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
973 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
975 memcpy( pbuf3, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
976 memcpy( pbuf4, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
977 nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
978 nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
979 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( pbuf3, pbuf4, 16*FDEC_STRIDE*sizeof(pixel) ) || nz_c != nz_a ) \
982 fprintf( stderr, #name " [FAILED]\n" ); \
984 call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
985 call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
988 #define TEST_ZIGZAG_SUBAC( name, t1, t2 ) \
989 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
992 dctcoef dc_a, dc_c; \
993 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
995 for( int i = 0; i < 2; i++ ) \
997 memcpy( pbuf3, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
998 memcpy( pbuf4, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
999 for( int j = 0; j < 4; j++ ) \
1001 memcpy( pbuf3 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
1002 memcpy( pbuf4 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
1004 nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
1005 nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
1006 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 ) \
1009 fprintf( stderr, #name " [FAILED]\n" ); \
1013 call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
1014 call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
1017 #define TEST_INTERLEAVE( name, t1, t2, dct, size ) \
1018 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1020 for( int j = 0; j < 100; j++ ) \
1022 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1024 memcpy(dct, buf1, size*sizeof(dctcoef)); \
1025 for( int i = 0; i < size; i++ ) \
1026 dct[i] = rand()&0x1F ? 0 : dct[i]; \
1027 memcpy(buf3, buf4, 10); \
1028 call_c( zigzag_c[interlace].name, t1, dct, buf3 ); \
1029 call_a( zigzag_asm[interlace].name, t2, dct, buf4 ); \
1030 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( buf3, buf4, 10 ) ) \
1037 x264_zigzag_init( 0, &zigzag_c[0], &zigzag_c[1] );
1038 x264_zigzag_init( cpu_ref, &zigzag_ref[0], &zigzag_ref[1] );
1039 x264_zigzag_init( cpu_new, &zigzag_asm[0], &zigzag_asm[1] );
1041 ok = 1; used_asm = 0;
1042 TEST_INTERLEAVE( interleave_8x8_cavlc, level1, level2, dct1[0], 64 );
1043 report( "zigzag_interleave :" );
1045 for( interlace = 0; interlace <= 1; interlace++ )
1047 ok = 1; used_asm = 0;
1048 TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, dct1[0], 8 );
1049 TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 4 );
1050 TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
1051 TEST_ZIGZAG_SUBAC( sub_4x4ac, level1, level2 );
1052 report( interlace ? "zigzag_field :" : "zigzag_frame :" );
1054 #undef TEST_ZIGZAG_SCAN
1055 #undef TEST_ZIGZAG_SUB
1060 static int check_mc( int cpu_ref, int cpu_new )
1062 x264_mc_functions_t mc_c;
1063 x264_mc_functions_t mc_ref;
1064 x264_mc_functions_t mc_a;
1065 x264_pixel_function_t pixf;
1067 pixel *src = &(pbuf1)[2*64+2];
1068 pixel *src2[4] = { &(pbuf1)[3*64+2], &(pbuf1)[5*64+2],
1069 &(pbuf1)[7*64+2], &(pbuf1)[9*64+2] };
1070 pixel *dst1 = pbuf3;
1071 pixel *dst2 = pbuf4;
1073 int ret = 0, ok, used_asm;
1075 x264_mc_init( 0, &mc_c );
1076 x264_mc_init( cpu_ref, &mc_ref );
1077 x264_mc_init( cpu_new, &mc_a );
1078 x264_pixel_init( 0, &pixf );
1080 #define MC_TEST_LUMA( w, h ) \
1081 if( mc_a.mc_luma != mc_ref.mc_luma && !(w&(w-1)) && h<=16 ) \
1083 const x264_weight_t *weight = x264_weight_none; \
1084 set_func_name( "mc_luma_%dx%d", w, h ); \
1086 for( int i = 0; i < 1024; i++ ) \
1087 pbuf3[i] = pbuf4[i] = 0xCD; \
1088 call_c( mc_c.mc_luma, dst1, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1089 call_a( mc_a.mc_luma, dst2, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1090 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1092 fprintf( stderr, "mc_luma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
1096 if( mc_a.get_ref != mc_ref.get_ref ) \
1098 pixel *ref = dst2; \
1099 intptr_t ref_stride = 32; \
1100 int w_checked = ( ( sizeof(pixel) == 2 && (w == 12 || w == 20)) ? w-2 : w ); \
1101 const x264_weight_t *weight = x264_weight_none; \
1102 set_func_name( "get_ref_%dx%d", w_checked, h ); \
1104 for( int i = 0; i < 1024; i++ ) \
1105 pbuf3[i] = pbuf4[i] = 0xCD; \
1106 call_c( mc_c.mc_luma, dst1, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1107 ref = (pixel*)call_a( mc_a.get_ref, ref, &ref_stride, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1108 for( int i = 0; i < h; i++ ) \
1109 if( memcmp( dst1+i*32, ref+i*ref_stride, w_checked * sizeof(pixel) ) ) \
1111 fprintf( stderr, "get_ref[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w_checked, h ); \
1117 #define MC_TEST_CHROMA( w, h ) \
1118 if( mc_a.mc_chroma != mc_ref.mc_chroma ) \
1120 set_func_name( "mc_chroma_%dx%d", w, h ); \
1122 for( int i = 0; i < 1024; i++ ) \
1123 pbuf3[i] = pbuf4[i] = 0xCD; \
1124 call_c( mc_c.mc_chroma, dst1, dst1+8, (intptr_t)16, src, (intptr_t)64, dx, dy, w, h ); \
1125 call_a( mc_a.mc_chroma, dst2, dst2+8, (intptr_t)16, src, (intptr_t)64, dx, dy, w, h ); \
1126 /* mc_chroma width=2 may write garbage to the right of dst. ignore that. */ \
1127 for( int j = 0; j < h; j++ ) \
1128 for( int i = w; i < 8; i++ ) \
1130 dst2[i+j*16+8] = dst1[i+j*16+8]; \
1131 dst2[i+j*16 ] = dst1[i+j*16 ]; \
1133 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1135 fprintf( stderr, "mc_chroma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
1139 ok = 1; used_asm = 0;
1140 for( int dy = -8; dy < 8; dy++ )
1141 for( int dx = -128; dx < 128; dx++ )
1143 if( rand()&15 ) continue; // running all of them is too slow
1144 MC_TEST_LUMA( 20, 18 );
1145 MC_TEST_LUMA( 16, 16 );
1146 MC_TEST_LUMA( 16, 8 );
1147 MC_TEST_LUMA( 12, 10 );
1148 MC_TEST_LUMA( 8, 16 );
1149 MC_TEST_LUMA( 8, 8 );
1150 MC_TEST_LUMA( 8, 4 );
1151 MC_TEST_LUMA( 4, 8 );
1152 MC_TEST_LUMA( 4, 4 );
1154 report( "mc luma :" );
1156 ok = 1; used_asm = 0;
1157 for( int dy = -1; dy < 9; dy++ )
1158 for( int dx = -128; dx < 128; dx++ )
1160 if( rand()&15 ) continue;
1161 MC_TEST_CHROMA( 8, 8 );
1162 MC_TEST_CHROMA( 8, 4 );
1163 MC_TEST_CHROMA( 4, 8 );
1164 MC_TEST_CHROMA( 4, 4 );
1165 MC_TEST_CHROMA( 4, 2 );
1166 MC_TEST_CHROMA( 2, 4 );
1167 MC_TEST_CHROMA( 2, 2 );
1169 report( "mc chroma :" );
1171 #undef MC_TEST_CHROMA
1173 #define MC_TEST_AVG( name, weight ) \
1175 for( int i = 0; i < 12; i++ ) \
1177 memcpy( pbuf3, pbuf1+320, 320 * sizeof(pixel) ); \
1178 memcpy( pbuf4, pbuf1+320, 320 * sizeof(pixel) ); \
1179 if( mc_a.name[i] != mc_ref.name[i] ) \
1181 set_func_name( "%s_%s", #name, pixel_names[i] ); \
1183 call_c1( mc_c.name[i], pbuf3, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1184 call_a1( mc_a.name[i], pbuf4, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1185 if( memcmp( pbuf3, pbuf4, 320 * sizeof(pixel) ) ) \
1188 fprintf( stderr, #name "[%d]: [FAILED]\n", i ); \
1190 call_c2( mc_c.name[i], pbuf3, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1191 call_a2( mc_a.name[i], pbuf4, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1196 ok = 1, used_asm = 0;
1197 for( int w = -63; w <= 127 && ok; w++ )
1198 MC_TEST_AVG( avg, w );
1199 report( "mc wpredb :" );
1201 #define MC_TEST_WEIGHT( name, weight, aligned ) \
1202 int align_off = (aligned ? 0 : rand()%16); \
1203 for( int i = 1; i <= 5; i++ ) \
1205 ALIGNED_16( pixel buffC[640] ); \
1206 ALIGNED_16( pixel buffA[640] ); \
1207 int j = X264_MAX( i*4, 2 ); \
1208 memset( buffC, 0, 640 * sizeof(pixel) ); \
1209 memset( buffA, 0, 640 * sizeof(pixel) ); \
1212 /* w12 is the same as w16 in some cases */ \
1213 if( i == 3 && mc_a.name[i] == mc_a.name[i+1] ) \
1215 if( mc_a.name[i] != mc_ref.name[i] ) \
1217 set_func_name( "%s_w%d", #name, j ); \
1219 call_c1( mc_c.weight[i], buffC, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1220 mc_a.weight_cache(&ha, &weight); \
1221 call_a1( weight.weightfn[i], buffA, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1222 for( int k = 0; k < 16; k++ ) \
1223 if( memcmp( &buffC[k*32], &buffA[k*32], j * sizeof(pixel) ) ) \
1226 fprintf( stderr, #name "[%d]: [FAILED] s:%d o:%d d%d\n", i, s, o, d ); \
1229 call_c2( mc_c.weight[i], buffC, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1230 call_a2( weight.weightfn[i], buffA, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1234 ok = 1; used_asm = 0;
1237 for( int s = 0; s <= 127 && ok; s++ )
1239 for( int o = -128; o <= 127 && ok; o++ )
1241 if( rand() & 2047 ) continue;
1242 for( int d = 0; d <= 7 && ok; d++ )
1246 x264_weight_t weight = { .i_scale = s, .i_denom = d, .i_offset = o };
1247 MC_TEST_WEIGHT( weight, weight, (align_cnt++ % 4) );
1252 report( "mc weight :" );
1254 ok = 1; used_asm = 0;
1255 for( int o = 0; o <= 127 && ok; o++ )
1258 if( rand() & 15 ) continue;
1259 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1260 MC_TEST_WEIGHT( offsetadd, weight, (align_cnt++ % 4) );
1262 report( "mc offsetadd :" );
1263 ok = 1; used_asm = 0;
1264 for( int o = -128; o < 0 && ok; o++ )
1267 if( rand() & 15 ) continue;
1268 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1269 MC_TEST_WEIGHT( offsetsub, weight, (align_cnt++ % 4) );
1271 report( "mc offsetsub :" );
1273 ok = 1; used_asm = 0;
1274 for( int height = 8; height <= 16; height += 8 )
1276 if( mc_a.store_interleave_chroma != mc_ref.store_interleave_chroma )
1278 set_func_name( "store_interleave_chroma" );
1280 memset( pbuf3, 0, 64*height );
1281 memset( pbuf4, 0, 64*height );
1282 call_c( mc_c.store_interleave_chroma, pbuf3, (intptr_t)64, pbuf1, pbuf1+16, height );
1283 call_a( mc_a.store_interleave_chroma, pbuf4, (intptr_t)64, pbuf1, pbuf1+16, height );
1284 if( memcmp( pbuf3, pbuf4, 64*height ) )
1287 fprintf( stderr, "store_interleave_chroma FAILED: h=%d\n", height );
1291 if( mc_a.load_deinterleave_chroma_fenc != mc_ref.load_deinterleave_chroma_fenc )
1293 set_func_name( "load_deinterleave_chroma_fenc" );
1295 call_c( mc_c.load_deinterleave_chroma_fenc, pbuf3, pbuf1, (intptr_t)64, height );
1296 call_a( mc_a.load_deinterleave_chroma_fenc, pbuf4, pbuf1, (intptr_t)64, height );
1297 if( memcmp( pbuf3, pbuf4, FENC_STRIDE*height ) )
1300 fprintf( stderr, "load_deinterleave_chroma_fenc FAILED: h=%d\n", height );
1304 if( mc_a.load_deinterleave_chroma_fdec != mc_ref.load_deinterleave_chroma_fdec )
1306 set_func_name( "load_deinterleave_chroma_fdec" );
1308 call_c( mc_c.load_deinterleave_chroma_fdec, pbuf3, pbuf1, (intptr_t)64, height );
1309 call_a( mc_a.load_deinterleave_chroma_fdec, pbuf4, pbuf1, (intptr_t)64, height );
1310 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*height ) )
1313 fprintf( stderr, "load_deinterleave_chroma_fdec FAILED: h=%d\n", height );
1318 report( "store_interleave :" );
1321 int w, h, src_stride;
1322 } 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} };
1323 ok = 1; used_asm = 0;
1324 if( mc_a.plane_copy != mc_ref.plane_copy )
1326 set_func_name( "plane_copy" );
1328 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1330 int w = plane_specs[i].w;
1331 int h = plane_specs[i].h;
1332 intptr_t src_stride = plane_specs[i].src_stride;
1333 intptr_t dst_stride = (w + 127) & ~63;
1334 assert( dst_stride * h <= 0x1000 );
1335 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1336 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1337 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1338 call_c( mc_c.plane_copy, pbuf3, dst_stride, src1, src_stride, w, h );
1339 call_a( mc_a.plane_copy, pbuf4, dst_stride, src1, src_stride, w, h );
1340 for( int y = 0; y < h; y++ )
1341 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(pixel) ) )
1344 fprintf( stderr, "plane_copy FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1350 if( mc_a.plane_copy_interleave != mc_ref.plane_copy_interleave )
1352 set_func_name( "plane_copy_interleave" );
1354 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1356 int w = (plane_specs[i].w + 1) >> 1;
1357 int h = plane_specs[i].h;
1358 intptr_t src_stride = (plane_specs[i].src_stride + 1) >> 1;
1359 intptr_t dst_stride = (2*w + 127) & ~63;
1360 assert( dst_stride * h <= 0x1000 );
1361 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1362 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1363 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1364 call_c( mc_c.plane_copy_interleave, pbuf3, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1365 call_a( mc_a.plane_copy_interleave, pbuf4, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1366 for( int y = 0; y < h; y++ )
1367 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
1370 fprintf( stderr, "plane_copy_interleave FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1376 if( mc_a.plane_copy_deinterleave != mc_ref.plane_copy_deinterleave )
1378 set_func_name( "plane_copy_deinterleave" );
1380 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1382 int w = (plane_specs[i].w + 1) >> 1;
1383 int h = plane_specs[i].h;
1384 intptr_t dst_stride = w;
1385 intptr_t src_stride = (2*w + 127) & ~63;
1386 intptr_t offv = (dst_stride*h + 31) & ~15;
1387 memset( pbuf3, 0, 0x1000 );
1388 memset( pbuf4, 0, 0x1000 );
1389 call_c( mc_c.plane_copy_deinterleave, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf1, src_stride, w, h );
1390 call_a( mc_a.plane_copy_deinterleave, pbuf4, dst_stride, pbuf4+offv, dst_stride, pbuf1, src_stride, w, h );
1391 for( int y = 0; y < h; y++ )
1392 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w ) ||
1393 memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w ) )
1396 fprintf( stderr, "plane_copy_deinterleave FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1401 report( "plane_copy :" );
1403 if( mc_a.hpel_filter != mc_ref.hpel_filter )
1405 pixel *srchpel = pbuf1+8+2*64;
1406 pixel *dstc[3] = { pbuf3+8, pbuf3+8+16*64, pbuf3+8+32*64 };
1407 pixel *dsta[3] = { pbuf4+8, pbuf4+8+16*64, pbuf4+8+32*64 };
1408 void *tmp = pbuf3+49*64;
1409 set_func_name( "hpel_filter" );
1410 ok = 1; used_asm = 1;
1411 memset( pbuf3, 0, 4096 * sizeof(pixel) );
1412 memset( pbuf4, 0, 4096 * sizeof(pixel) );
1413 call_c( mc_c.hpel_filter, dstc[0], dstc[1], dstc[2], srchpel, (intptr_t)64, 48, 10, tmp );
1414 call_a( mc_a.hpel_filter, dsta[0], dsta[1], dsta[2], srchpel, (intptr_t)64, 48, 10, tmp );
1415 for( int i = 0; i < 3; i++ )
1416 for( int j = 0; j < 10; j++ )
1417 //FIXME ideally the first pixels would match too, but they aren't actually used
1418 if( memcmp( dstc[i]+j*64+2, dsta[i]+j*64+2, 43 * sizeof(pixel) ) )
1421 fprintf( stderr, "hpel filter differs at plane %c line %d\n", "hvc"[i], j );
1422 for( int k = 0; k < 48; k++ )
1423 printf( "%02x%s", dstc[i][j*64+k], (k+1)&3 ? "" : " " );
1425 for( int k = 0; k < 48; k++ )
1426 printf( "%02x%s", dsta[i][j*64+k], (k+1)&3 ? "" : " " );
1430 report( "hpel filter :" );
1433 if( mc_a.frame_init_lowres_core != mc_ref.frame_init_lowres_core )
1435 pixel *dstc[4] = { pbuf3, pbuf3+1024, pbuf3+2048, pbuf3+3072 };
1436 pixel *dsta[4] = { pbuf4, pbuf4+1024, pbuf4+2048, pbuf4+3072 };
1437 set_func_name( "lowres_init" );
1438 ok = 1; used_asm = 1;
1439 for( int w = 40; w <= 48; w += 8 )
1441 intptr_t stride = (w+8)&~15;
1442 call_c( mc_c.frame_init_lowres_core, pbuf1, dstc[0], dstc[1], dstc[2], dstc[3], (intptr_t)w*2, stride, w, 16 );
1443 call_a( mc_a.frame_init_lowres_core, pbuf1, dsta[0], dsta[1], dsta[2], dsta[3], (intptr_t)w*2, stride, w, 16 );
1444 for( int i = 0; i < 16; i++ )
1446 for( int j = 0; j < 4; j++ )
1447 if( memcmp( dstc[j]+i*stride, dsta[j]+i*stride, w * sizeof(pixel) ) )
1450 fprintf( stderr, "frame_init_lowres differs at plane %d line %d\n", j, i );
1451 for( int k = 0; k < w; k++ )
1452 printf( "%d ", dstc[j][k+i*stride] );
1454 for( int k = 0; k < w; k++ )
1455 printf( "%d ", dsta[j][k+i*stride] );
1461 report( "lowres init :" );
1464 #define INTEGRAL_INIT( name, size, ... )\
1465 if( mc_a.name != mc_ref.name )\
1467 intptr_t stride = 80;\
1468 set_func_name( #name );\
1470 memcpy( buf3, buf1, size*2*stride );\
1471 memcpy( buf4, buf1, size*2*stride );\
1472 uint16_t *sum = (uint16_t*)buf3;\
1473 call_c1( mc_c.name, __VA_ARGS__ );\
1474 sum = (uint16_t*)buf4;\
1475 call_a1( mc_a.name, __VA_ARGS__ );\
1476 if( memcmp( buf3, buf4, (stride-8)*2 ) \
1477 || (size>9 && memcmp( buf3+18*stride, buf4+18*stride, (stride-8)*2 )))\
1479 call_c2( mc_c.name, __VA_ARGS__ );\
1480 call_a2( mc_a.name, __VA_ARGS__ );\
1482 ok = 1; used_asm = 0;
1483 INTEGRAL_INIT( integral_init4h, 2, sum+stride, pbuf2, stride );
1484 INTEGRAL_INIT( integral_init8h, 2, sum+stride, pbuf2, stride );
1485 INTEGRAL_INIT( integral_init4v, 14, sum, sum+9*stride, stride );
1486 INTEGRAL_INIT( integral_init8v, 9, sum, stride );
1487 report( "integral init :" );
1489 if( mc_a.mbtree_propagate_cost != mc_ref.mbtree_propagate_cost )
1491 ok = 1; used_asm = 1;
1493 for( int i = 0; i < 10; i++ )
1495 float fps_factor = (rand()&65535) / 256.;
1496 set_func_name( "mbtree_propagate" );
1497 int *dsta = (int*)buf3;
1498 int *dstc = dsta+400;
1499 uint16_t *prop = (uint16_t*)buf1;
1500 uint16_t *intra = (uint16_t*)buf4;
1501 uint16_t *inter = intra+128;
1502 uint16_t *qscale = inter+128;
1503 uint16_t *rnd = (uint16_t*)buf2;
1505 for( int j = 0; j < 100; j++ )
1507 intra[j] = *rnd++ & 0x7fff;
1508 intra[j] += !intra[j];
1509 inter[j] = *rnd++ & 0x7fff;
1510 qscale[j] = *rnd++ & 0x7fff;
1512 call_c( mc_c.mbtree_propagate_cost, dstc, prop, intra, inter, qscale, &fps_factor, 100 );
1513 call_a( mc_a.mbtree_propagate_cost, dsta, prop, intra, inter, qscale, &fps_factor, 100 );
1514 // I don't care about exact rounding, this is just how close the floating-point implementation happens to be
1516 for( int j = 0; j < 100 && ok; j++ )
1518 ok &= abs( dstc[j]-dsta[j] ) <= 1 || fabs( (double)dstc[j]/dsta[j]-1 ) < 1e-4;
1520 fprintf( stderr, "mbtree_propagate FAILED: %f !~= %f\n", (double)dstc[j], (double)dsta[j] );
1523 report( "mbtree propagate :" );
1526 if( mc_a.memcpy_aligned != mc_ref.memcpy_aligned )
1528 set_func_name( "memcpy_aligned" );
1529 ok = 1; used_asm = 1;
1530 for( size_t size = 16; size < 256; size += 16 )
1532 memset( buf4, 0xAA, size + 1 );
1533 call_c( mc_c.memcpy_aligned, buf3, buf1, size );
1534 call_a( mc_a.memcpy_aligned, buf4, buf1, size );
1535 if( memcmp( buf3, buf4, size ) || buf4[size] != 0xAA )
1538 fprintf( stderr, "memcpy_aligned FAILED: size=%d\n", (int)size );
1542 report( "memcpy aligned :" );
1545 if( mc_a.memzero_aligned != mc_ref.memzero_aligned )
1547 set_func_name( "memzero_aligned" );
1548 ok = 1; used_asm = 1;
1549 for( size_t size = 128; size < 1024; size += 128 )
1551 memset( buf4, 0xAA, size + 1 );
1552 call_c( mc_c.memzero_aligned, buf3, size );
1553 call_a( mc_a.memzero_aligned, buf4, size );
1554 if( memcmp( buf3, buf4, size ) || buf4[size] != 0xAA )
1557 fprintf( stderr, "memzero_aligned FAILED: size=%d\n", (int)size );
1561 report( "memzero aligned :" );
1567 static int check_deblock( int cpu_ref, int cpu_new )
1569 x264_deblock_function_t db_c;
1570 x264_deblock_function_t db_ref;
1571 x264_deblock_function_t db_a;
1572 int ret = 0, ok = 1, used_asm = 0;
1573 int alphas[36], betas[36];
1576 x264_deblock_init( 0, &db_c, 0 );
1577 x264_deblock_init( cpu_ref, &db_ref, 0 );
1578 x264_deblock_init( cpu_new, &db_a, 0 );
1580 /* not exactly the real values of a,b,tc but close enough */
1581 for( int i = 35, a = 255, c = 250; i >= 0; i-- )
1583 alphas[i] = a << (BIT_DEPTH-8);
1584 betas[i] = (i+1)/2 << (BIT_DEPTH-8);
1585 tcs[i][0] = tcs[i][3] = (c+6)/10 << (BIT_DEPTH-8);
1586 tcs[i][1] = (c+7)/15 << (BIT_DEPTH-8);
1587 tcs[i][2] = (c+9)/20 << (BIT_DEPTH-8);
1592 #define TEST_DEBLOCK( name, align, ... ) \
1593 for( int i = 0; i < 36; i++ ) \
1595 intptr_t off = 8*32 + (i&15)*4*!align; /* benchmark various alignments of h filter */ \
1596 for( int j = 0; j < 1024; j++ ) \
1597 /* two distributions of random to excersize different failure modes */ \
1598 pbuf3[j] = rand() & (i&1 ? 0xf : PIXEL_MAX ); \
1599 memcpy( pbuf4, pbuf3, 1024 * sizeof(pixel) ); \
1600 if( db_a.name != db_ref.name ) \
1602 set_func_name( #name ); \
1604 call_c1( db_c.name, pbuf3+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1605 call_a1( db_a.name, pbuf4+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1606 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1609 fprintf( stderr, #name "(a=%d, b=%d): [FAILED]\n", alphas[i], betas[i] ); \
1612 call_c2( db_c.name, pbuf3+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1613 call_a2( db_a.name, pbuf4+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1617 TEST_DEBLOCK( deblock_luma[0], 0, tcs[i] );
1618 TEST_DEBLOCK( deblock_luma[1], 1, tcs[i] );
1619 TEST_DEBLOCK( deblock_h_chroma_420, 0, tcs[i] );
1620 TEST_DEBLOCK( deblock_h_chroma_422, 0, tcs[i] );
1621 TEST_DEBLOCK( deblock_chroma_420_mbaff, 0, tcs[i] );
1622 TEST_DEBLOCK( deblock_chroma_422_mbaff, 0, tcs[i] );
1623 TEST_DEBLOCK( deblock_chroma[1], 1, tcs[i] );
1624 TEST_DEBLOCK( deblock_luma_intra[0], 0 );
1625 TEST_DEBLOCK( deblock_luma_intra[1], 1 );
1626 TEST_DEBLOCK( deblock_h_chroma_420_intra, 0 );
1627 TEST_DEBLOCK( deblock_h_chroma_422_intra, 0 );
1628 TEST_DEBLOCK( deblock_chroma_420_intra_mbaff, 0 );
1629 TEST_DEBLOCK( deblock_chroma_422_intra_mbaff, 0 );
1630 TEST_DEBLOCK( deblock_chroma_intra[1], 1 );
1632 if( db_a.deblock_strength != db_ref.deblock_strength )
1634 for( int i = 0; i < 100; i++ )
1636 ALIGNED_ARRAY_16( uint8_t, nnz, [X264_SCAN8_SIZE] );
1637 ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );
1638 ALIGNED_ARRAY_16( int16_t, mv, [2],[X264_SCAN8_LUMA_SIZE][2] );
1639 ALIGNED_ARRAY_16( uint8_t, bs, [2],[2][8][4] );
1640 memset( bs, 99, sizeof(bs) );
1641 for( int j = 0; j < X264_SCAN8_SIZE; j++ )
1642 nnz[j] = ((rand()&7) == 7) * rand() & 0xf;
1643 for( int j = 0; j < 2; j++ )
1644 for( int k = 0; k < X264_SCAN8_LUMA_SIZE; k++ )
1646 ref[j][k] = ((rand()&3) != 3) ? 0 : (rand() & 31) - 2;
1647 for( int l = 0; l < 2; l++ )
1648 mv[j][k][l] = ((rand()&7) != 7) ? (rand()&7) - 3 : (rand()&1023) - 512;
1650 set_func_name( "deblock_strength" );
1651 call_c( db_c.deblock_strength, nnz, ref, mv, bs[0], 2<<(i&1), ((i>>1)&1) );
1652 call_a( db_a.deblock_strength, nnz, ref, mv, bs[1], 2<<(i&1), ((i>>1)&1) );
1653 if( memcmp( bs[0], bs[1], sizeof(bs[0]) ) )
1656 fprintf( stderr, "deblock_strength: [FAILED]\n" );
1657 for( int j = 0; j < 2; j++ )
1659 for( int k = 0; k < 2; k++ )
1660 for( int l = 0; l < 4; l++ )
1662 for( int m = 0; m < 4; m++ )
1663 printf("%d ",bs[j][k][l][m]);
1673 report( "deblock :" );
1678 static int check_quant( int cpu_ref, int cpu_new )
1680 x264_quant_function_t qf_c;
1681 x264_quant_function_t qf_ref;
1682 x264_quant_function_t qf_a;
1683 ALIGNED_16( dctcoef dct1[64] );
1684 ALIGNED_16( dctcoef dct2[64] );
1685 ALIGNED_16( dctcoef dct3[8][16] );
1686 ALIGNED_16( dctcoef dct4[8][16] );
1687 ALIGNED_16( uint8_t cqm_buf[64] );
1688 int ret = 0, ok, used_asm;
1689 int oks[3] = {1,1,1}, used_asms[3] = {0,0,0};
1692 memset( h, 0, sizeof(*h) );
1693 h->sps->i_chroma_format_idc = 1;
1694 x264_param_default( &h->param );
1695 h->chroma_qp_table = i_chroma_qp_table + 12;
1696 h->param.analyse.b_transform_8x8 = 1;
1698 for( int i_cqm = 0; i_cqm < 4; i_cqm++ )
1702 for( int i = 0; i < 6; i++ )
1703 h->pps->scaling_list[i] = x264_cqm_flat16;
1704 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_FLAT;
1706 else if( i_cqm == 1 )
1708 for( int i = 0; i < 6; i++ )
1709 h->pps->scaling_list[i] = x264_cqm_jvt[i];
1710 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_JVT;
1714 int max_scale = BIT_DEPTH < 10 ? 255 : 228;
1716 for( int i = 0; i < 64; i++ )
1717 cqm_buf[i] = 10 + rand() % (max_scale - 9);
1719 for( int i = 0; i < 64; i++ )
1721 for( int i = 0; i < 6; i++ )
1722 h->pps->scaling_list[i] = cqm_buf;
1723 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_CUSTOM;
1726 h->param.rc.i_qp_min = 0;
1727 h->param.rc.i_qp_max = QP_MAX;
1729 x264_quant_init( h, 0, &qf_c );
1730 x264_quant_init( h, cpu_ref, &qf_ref );
1731 x264_quant_init( h, cpu_new, &qf_a );
1733 #define INIT_QUANT8(j) \
1735 static const int scale1d[8] = {32,31,24,31,32,31,24,31}; \
1736 for( int i = 0; i < 64; i++ ) \
1738 unsigned int scale = (255*scale1d[i>>3]*scale1d[i&7])/16; \
1739 dct1[i] = dct2[i] = j ? (rand()%(2*scale+1))-scale : 0; \
1743 #define INIT_QUANT4(j) \
1745 static const int scale1d[4] = {4,6,4,6}; \
1746 for( int i = 0; i < 16; i++ ) \
1748 unsigned int scale = 255*scale1d[i>>2]*scale1d[i&3]; \
1749 dct1[i] = dct2[i] = j ? (rand()%(2*scale+1))-scale : 0; \
1753 #define TEST_QUANT_DC( name, cqm ) \
1754 if( qf_a.name != qf_ref.name ) \
1756 set_func_name( #name ); \
1758 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1760 for( int j = 0; j < 2; j++ ) \
1762 int result_c, result_a; \
1763 for( int i = 0; i < 16; i++ ) \
1764 dct1[i] = dct2[i] = j ? (rand() & 0x1fff) - 0xfff : 0; \
1765 result_c = call_c1( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1766 result_a = call_a1( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1767 if( memcmp( dct1, dct2, 16*sizeof(dctcoef) ) || result_c != result_a ) \
1770 fprintf( stderr, #name "(cqm=%d): [FAILED]\n", i_cqm ); \
1773 call_c2( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1774 call_a2( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1779 #define TEST_QUANT( qname, block, w ) \
1780 if( qf_a.qname != qf_ref.qname ) \
1782 set_func_name( #qname ); \
1784 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1786 for( int j = 0; j < 2; j++ ) \
1789 int result_c = call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1790 int result_a = call_a1( qf_a.qname, dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1791 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) || result_c != result_a ) \
1794 fprintf( stderr, #qname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1797 call_c2( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1798 call_a2( qf_a.qname, dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1803 TEST_QUANT( quant_8x8, CQM_8IY, 8 );
1804 TEST_QUANT( quant_8x8, CQM_8PY, 8 );
1805 TEST_QUANT( quant_4x4, CQM_4IY, 4 );
1806 TEST_QUANT( quant_4x4, CQM_4PY, 4 );
1807 TEST_QUANT_DC( quant_4x4_dc, **h->quant4_mf[CQM_4IY] );
1808 TEST_QUANT_DC( quant_2x2_dc, **h->quant4_mf[CQM_4IC] );
1810 #define TEST_DEQUANT( qname, dqname, block, w ) \
1811 if( qf_a.dqname != qf_ref.dqname ) \
1813 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
1815 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1818 qf_c.qname( dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
1819 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1820 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1821 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1822 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1825 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1828 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1829 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1833 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8IY, 8 );
1834 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8PY, 8 );
1835 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4IY, 4 );
1836 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4PY, 4 );
1838 #define TEST_DEQUANT_DC( qname, dqname, block, w ) \
1839 if( qf_a.dqname != qf_ref.dqname ) \
1841 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
1843 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1845 for( int i = 0; i < 16; i++ ) \
1846 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16; \
1847 qf_c.qname( dct1, h->quant##w##_mf[block][qp][0]>>1, h->quant##w##_bias[block][qp][0]>>1 ); \
1848 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
1849 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1850 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1851 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
1854 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
1856 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
1857 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
1861 TEST_DEQUANT_DC( quant_4x4_dc, dequant_4x4_dc, CQM_4IY, 4 );
1863 if( qf_a.idct_dequant_2x4_dc != qf_ref.idct_dequant_2x4_dc )
1865 set_func_name( "idct_dequant_2x4_dc_%s", i_cqm?"cqm":"flat" );
1867 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- )
1869 for( int i = 0; i < 8; i++ )
1870 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16;
1871 qf_c.quant_2x2_dc( &dct1[0], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
1872 qf_c.quant_2x2_dc( &dct1[4], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
1873 call_c( qf_c.idct_dequant_2x4_dc, dct1, dct3, h->dequant4_mf[CQM_4IC], qp+3 );
1874 call_a( qf_a.idct_dequant_2x4_dc, dct1, dct4, h->dequant4_mf[CQM_4IC], qp+3 );
1875 for( int i = 0; i < 8; i++ )
1876 if( dct3[i][0] != dct4[i][0] )
1879 fprintf( stderr, "idct_dequant_2x4_dc (qp=%d, cqm=%d): [FAILED]\n", qp, i_cqm );
1885 if( qf_a.idct_dequant_2x4_dconly != qf_ref.idct_dequant_2x4_dconly )
1887 set_func_name( "idct_dequant_2x4_dc_%s", i_cqm?"cqm":"flat" );
1889 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- )
1891 for( int i = 0; i < 8; i++ )
1892 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16;
1893 qf_c.quant_2x2_dc( &dct1[0], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
1894 qf_c.quant_2x2_dc( &dct1[4], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
1895 memcpy( dct2, dct1, 8*sizeof(dctcoef) );
1896 call_c1( qf_c.idct_dequant_2x4_dconly, dct1, h->dequant4_mf[CQM_4IC], qp+3 );
1897 call_a1( qf_a.idct_dequant_2x4_dconly, dct2, h->dequant4_mf[CQM_4IC], qp+3 );
1898 if( memcmp( dct1, dct2, 8*sizeof(dctcoef) ) )
1901 fprintf( stderr, "idct_dequant_2x4_dconly (qp=%d, cqm=%d): [FAILED]\n", qp, i_cqm );
1904 call_c2( qf_c.idct_dequant_2x4_dconly, dct1, h->dequant4_mf[CQM_4IC], qp+3 );
1905 call_a2( qf_a.idct_dequant_2x4_dconly, dct2, h->dequant4_mf[CQM_4IC], qp+3 );
1909 #define TEST_OPTIMIZE_CHROMA_DC( optname, size ) \
1910 if( qf_a.optname != qf_ref.optname ) \
1912 set_func_name( #optname ); \
1914 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1916 int qpdc = qp + (size == 8 ? 3 : 0); \
1917 int dmf = h->dequant4_mf[CQM_4IC][qpdc%6][0] << qpdc/6; \
1920 for( int i = 16; ; i <<= 1 ) \
1922 int res_c, res_asm; \
1923 int max = X264_MIN( i, PIXEL_MAX*16 ); \
1924 for( int j = 0; j < size; j++ ) \
1925 dct1[j] = rand()%(max*2+1) - max; \
1926 for( int j = 0; i <= size; j += 4 ) \
1927 qf_c.quant_2x2_dc( &dct1[j], h->quant4_mf[CQM_4IC][qpdc][0]>>1, h->quant4_bias[CQM_4IC][qpdc][0]>>1 ); \
1928 memcpy( dct2, dct1, size*sizeof(dctcoef) ); \
1929 res_c = call_c1( qf_c.optname, dct1, dmf ); \
1930 res_asm = call_a1( qf_a.optname, dct2, dmf ); \
1931 if( res_c != res_asm || memcmp( dct1, dct2, size*sizeof(dctcoef) ) ) \
1934 fprintf( stderr, #optname "(qp=%d, res_c=%d, res_asm=%d): [FAILED]\n", qp, res_c, res_asm ); \
1936 call_c2( qf_c.optname, dct1, dmf ); \
1937 call_a2( qf_a.optname, dct2, dmf ); \
1938 if( i >= PIXEL_MAX*16 ) \
1944 TEST_OPTIMIZE_CHROMA_DC( optimize_chroma_2x2_dc, 4 );
1945 TEST_OPTIMIZE_CHROMA_DC( optimize_chroma_2x4_dc, 8 );
1947 x264_cqm_delete( h );
1950 ok = oks[0]; used_asm = used_asms[0];
1951 report( "quant :" );
1953 ok = oks[1]; used_asm = used_asms[1];
1954 report( "dequant :" );
1956 ok = oks[2]; used_asm = used_asms[2];
1957 report( "optimize chroma dc :" );
1959 ok = 1; used_asm = 0;
1960 if( qf_a.denoise_dct != qf_ref.denoise_dct )
1963 for( int size = 16; size <= 64; size += 48 )
1965 set_func_name( "denoise_dct" );
1966 memcpy( dct1, buf1, size*sizeof(dctcoef) );
1967 memcpy( dct2, buf1, size*sizeof(dctcoef) );
1968 memcpy( buf3+256, buf3, 256 );
1969 call_c1( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
1970 call_a1( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
1971 if( memcmp( dct1, dct2, size*sizeof(dctcoef) ) || memcmp( buf3+4, buf3+256+4, (size-1)*sizeof(uint32_t) ) )
1973 call_c2( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
1974 call_a2( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
1977 report( "denoise dct :" );
1979 #define TEST_DECIMATE( decname, w, ac, thresh ) \
1980 if( qf_a.decname != qf_ref.decname ) \
1982 set_func_name( #decname ); \
1984 for( int i = 0; i < 100; i++ ) \
1986 static const int distrib[16] = {1,1,1,1,1,1,1,1,1,1,1,1,2,3,4};\
1987 static const int zerorate_lut[4] = {3,7,15,31};\
1988 int zero_rate = zerorate_lut[i&3];\
1989 for( int idx = 0; idx < w*w; idx++ ) \
1991 int sign = (rand()&1) ? -1 : 1; \
1992 int abs_level = distrib[rand()&15]; \
1993 if( abs_level == 4 ) abs_level = rand()&0x3fff; \
1994 int zero = !(rand()&zero_rate); \
1995 dct1[idx] = zero * abs_level * sign; \
1999 int result_c = call_c( qf_c.decname, dct1 ); \
2000 int result_a = call_a( qf_a.decname, dct1 ); \
2001 if( X264_MIN(result_c,thresh) != X264_MIN(result_a,thresh) ) \
2004 fprintf( stderr, #decname ": [FAILED]\n" ); \
2010 ok = 1; used_asm = 0;
2011 TEST_DECIMATE( decimate_score64, 8, 0, 6 );
2012 TEST_DECIMATE( decimate_score16, 4, 0, 6 );
2013 TEST_DECIMATE( decimate_score15, 4, 1, 7 );
2014 report( "decimate_score :" );
2016 #define TEST_LAST( last, lastname, size, ac ) \
2017 if( qf_a.last != qf_ref.last ) \
2019 set_func_name( #lastname ); \
2021 for( int i = 0; i < 100; i++ ) \
2024 int max = rand() & (size-1); \
2025 memset( dct1, 0, size*sizeof(dctcoef) ); \
2026 for( int idx = ac; idx < max; idx++ ) \
2027 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
2030 int result_c = call_c( qf_c.last, dct1+ac ); \
2031 int result_a = call_a( qf_a.last, dct1+ac ); \
2032 if( result_c != result_a ) \
2035 fprintf( stderr, #lastname ": [FAILED]\n" ); \
2041 ok = 1; used_asm = 0;
2042 TEST_LAST( coeff_last4 , coeff_last4, 4, 0 );
2043 TEST_LAST( coeff_last8 , coeff_last8, 8, 0 );
2044 TEST_LAST( coeff_last[ DCT_LUMA_AC], coeff_last15, 16, 1 );
2045 TEST_LAST( coeff_last[ DCT_LUMA_4x4], coeff_last16, 16, 0 );
2046 TEST_LAST( coeff_last[ DCT_LUMA_8x8], coeff_last64, 64, 0 );
2047 report( "coeff_last :" );
2049 #define TEST_LEVELRUN( lastname, name, size, ac ) \
2050 if( qf_a.lastname != qf_ref.lastname ) \
2052 set_func_name( #name ); \
2054 for( int i = 0; i < 100; i++ ) \
2056 x264_run_level_t runlevel_c, runlevel_a; \
2058 int max = rand() & (size-1); \
2059 memset( dct1, 0, size*sizeof(dctcoef) ); \
2060 memcpy( &runlevel_a, buf1+i, sizeof(x264_run_level_t) ); \
2061 memcpy( &runlevel_c, buf1+i, sizeof(x264_run_level_t) ); \
2062 for( int idx = ac; idx < max; idx++ ) \
2063 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
2066 int result_c = call_c( qf_c.lastname, dct1+ac, &runlevel_c ); \
2067 int result_a = call_a( qf_a.lastname, dct1+ac, &runlevel_a ); \
2068 if( result_c != result_a || runlevel_c.last != runlevel_a.last || \
2069 runlevel_c.mask != runlevel_a.mask || \
2070 memcmp(runlevel_c.level, runlevel_a.level, sizeof(dctcoef)*result_c)) \
2073 fprintf( stderr, #name ": [FAILED]\n" ); \
2079 ok = 1; used_asm = 0;
2080 TEST_LEVELRUN( coeff_level_run4 , coeff_level_run4, 4, 0 );
2081 TEST_LEVELRUN( coeff_level_run8 , coeff_level_run8, 8, 0 );
2082 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_AC], coeff_level_run15, 16, 1 );
2083 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_4x4], coeff_level_run16, 16, 0 );
2084 report( "coeff_level_run :" );
2089 static int check_intra( int cpu_ref, int cpu_new )
2091 int ret = 0, ok = 1, used_asm = 0;
2092 ALIGNED_ARRAY_32( pixel, edge,[36] );
2093 ALIGNED_ARRAY_32( pixel, edge2,[36] );
2094 ALIGNED_16( pixel fdec[FDEC_STRIDE*20] );
2097 x264_predict_t predict_16x16[4+3];
2098 x264_predict_t predict_8x8c[4+3];
2099 x264_predict_t predict_8x16c[4+3];
2100 x264_predict8x8_t predict_8x8[9+3];
2101 x264_predict_t predict_4x4[9+3];
2102 x264_predict_8x8_filter_t predict_8x8_filter;
2103 } ip_c, ip_ref, ip_a;
2105 x264_predict_16x16_init( 0, ip_c.predict_16x16 );
2106 x264_predict_8x8c_init( 0, ip_c.predict_8x8c );
2107 x264_predict_8x16c_init( 0, ip_c.predict_8x16c );
2108 x264_predict_8x8_init( 0, ip_c.predict_8x8, &ip_c.predict_8x8_filter );
2109 x264_predict_4x4_init( 0, ip_c.predict_4x4 );
2111 x264_predict_16x16_init( cpu_ref, ip_ref.predict_16x16 );
2112 x264_predict_8x8c_init( cpu_ref, ip_ref.predict_8x8c );
2113 x264_predict_8x16c_init( cpu_ref, ip_ref.predict_8x16c );
2114 x264_predict_8x8_init( cpu_ref, ip_ref.predict_8x8, &ip_ref.predict_8x8_filter );
2115 x264_predict_4x4_init( cpu_ref, ip_ref.predict_4x4 );
2117 x264_predict_16x16_init( cpu_new, ip_a.predict_16x16 );
2118 x264_predict_8x8c_init( cpu_new, ip_a.predict_8x8c );
2119 x264_predict_8x16c_init( cpu_new, ip_a.predict_8x16c );
2120 x264_predict_8x8_init( cpu_new, ip_a.predict_8x8, &ip_a.predict_8x8_filter );
2121 x264_predict_4x4_init( cpu_new, ip_a.predict_4x4 );
2123 memcpy( fdec, pbuf1, 32*20 * sizeof(pixel) );\
2125 ip_c.predict_8x8_filter( fdec+48, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
2127 #define INTRA_TEST( name, dir, w, h, align, bench, ... )\
2128 if( ip_a.name[dir] != ip_ref.name[dir] )\
2130 set_func_name( "intra_%s_%s", #name, intra_##name##_names[dir] );\
2132 memcpy( pbuf3, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
2133 memcpy( pbuf4, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
2134 for( int a = 0; a < (do_bench ? 64/sizeof(pixel) : 1); a += align )\
2136 call_c##bench( ip_c.name[dir], pbuf3+48+a, ##__VA_ARGS__ );\
2137 call_a##bench( ip_a.name[dir], pbuf4+48+a, ##__VA_ARGS__ );\
2138 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*20 * sizeof(pixel) ) )\
2140 fprintf( stderr, #name "[%d] : [FAILED]\n", dir );\
2142 for( int k = -1; k < 16; k++ )\
2143 printf( "%2x ", edge[16+k] );\
2145 for( int j = 0; j < h; j++ )\
2147 printf( "%2x ", edge[14-j] );\
2148 for( int k = 0; k < w; k++ )\
2149 printf( "%2x ", pbuf4[48+k+j*FDEC_STRIDE] );\
2153 for( int j = 0; j < h; j++ )\
2156 for( int k = 0; k < w; k++ )\
2157 printf( "%2x ", pbuf3[48+k+j*FDEC_STRIDE] );\
2165 for( int i = 0; i < 12; i++ )
2166 INTRA_TEST( predict_4x4, i, 4, 4, 4, );
2167 for( int i = 0; i < 7; i++ )
2168 INTRA_TEST( predict_8x8c, i, 8, 8, 16, );
2169 for( int i = 0; i < 7; i++ )
2170 INTRA_TEST( predict_8x16c, i, 8, 16, 16, );
2171 for( int i = 0; i < 7; i++ )
2172 INTRA_TEST( predict_16x16, i, 16, 16, 16, );
2173 for( int i = 0; i < 12; i++ )
2174 INTRA_TEST( predict_8x8, i, 8, 8, 8, , edge );
2176 set_func_name("intra_predict_8x8_filter");
2177 if( ip_a.predict_8x8_filter != ip_ref.predict_8x8_filter )
2180 for( int i = 0; i < 32; i++ )
2182 if( !(i&7) || ((i&MB_TOPRIGHT) && !(i&MB_TOP)) )
2184 int neighbor = (i&24)>>1;
2185 memset( edge, 0, 36*sizeof(pixel) );
2186 memset( edge2, 0, 36*sizeof(pixel) );
2187 call_c( ip_c.predict_8x8_filter, pbuf1+48, edge, neighbor, i&7 );
2188 call_a( ip_a.predict_8x8_filter, pbuf1+48, edge2, neighbor, i&7 );
2189 if( !(neighbor&MB_TOPLEFT) )
2190 edge[15] = edge2[15] = 0;
2191 if( memcmp( edge+7, edge2+7, (i&MB_TOPRIGHT ? 26 : i&MB_TOP ? 17 : 8) * sizeof(pixel) ) )
2193 fprintf( stderr, "predict_8x8_filter : [FAILED] %d %d\n", (i&24)>>1, i&7);
2199 #define EXTREMAL_PLANE( w, h ) \
2202 for( int j = 0; j < 7; j++ ) \
2203 max[j] = test ? rand()&PIXEL_MAX : PIXEL_MAX; \
2204 fdec[48-1-FDEC_STRIDE] = (i&1)*max[0]; \
2205 for( int j = 0; j < w/2; j++ ) \
2206 fdec[48+j-FDEC_STRIDE] = (!!(i&2))*max[1]; \
2207 for( int j = w/2; j < w-1; j++ ) \
2208 fdec[48+j-FDEC_STRIDE] = (!!(i&4))*max[2]; \
2209 fdec[48+(w-1)-FDEC_STRIDE] = (!!(i&8))*max[3]; \
2210 for( int j = 0; j < h/2; j++ ) \
2211 fdec[48+j*FDEC_STRIDE-1] = (!!(i&16))*max[4]; \
2212 for( int j = h/2; j < h-1; j++ ) \
2213 fdec[48+j*FDEC_STRIDE-1] = (!!(i&32))*max[5]; \
2214 fdec[48+(h-1)*FDEC_STRIDE-1] = (!!(i&64))*max[6]; \
2216 /* Extremal test case for planar prediction. */
2217 for( int test = 0; test < 100 && ok; test++ )
2218 for( int i = 0; i < 128 && ok; i++ )
2220 EXTREMAL_PLANE( 8, 8 );
2221 INTRA_TEST( predict_8x8c, I_PRED_CHROMA_P, 8, 8, 64, 1 );
2222 EXTREMAL_PLANE( 8, 16 );
2223 INTRA_TEST( predict_8x16c, I_PRED_CHROMA_P, 8, 16, 64, 1 );
2224 EXTREMAL_PLANE( 16, 16 );
2225 INTRA_TEST( predict_16x16, I_PRED_16x16_P, 16, 16, 64, 1 );
2227 report( "intra pred :" );
2231 #define DECL_CABAC(cpu) \
2232 static void run_cabac_decision_##cpu( x264_t *h, uint8_t *dst )\
2235 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2236 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2237 for( int i = 0; i < 0x1000; i++ )\
2238 x264_cabac_encode_decision_##cpu( &cb, buf1[i]>>1, buf1[i]&1 );\
2240 static void run_cabac_bypass_##cpu( x264_t *h, uint8_t *dst )\
2243 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2244 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2245 for( int i = 0; i < 0x1000; i++ )\
2246 x264_cabac_encode_bypass_##cpu( &cb, buf1[i]&1 );\
2248 static void run_cabac_terminal_##cpu( x264_t *h, uint8_t *dst )\
2251 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2252 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2253 for( int i = 0; i < 0x1000; i++ )\
2254 x264_cabac_encode_terminal_##cpu( &cb );\
2260 #define run_cabac_decision_asm run_cabac_decision_c
2261 #define run_cabac_bypass_asm run_cabac_bypass_c
2262 #define run_cabac_terminal_asm run_cabac_terminal_c
2265 static int check_cabac( int cpu_ref, int cpu_new )
2267 int ret = 0, ok, used_asm = 1;
2269 h.sps->i_chroma_format_idc = 3;
2270 if( cpu_ref || run_cabac_decision_c == run_cabac_decision_asm )
2272 x264_cabac_init( &h );
2274 set_func_name( "cabac_encode_decision" );
2275 memcpy( buf4, buf3, 0x1000 );
2276 call_c( run_cabac_decision_c, &h, buf3 );
2277 call_a( run_cabac_decision_asm, &h, buf4 );
2278 ok = !memcmp( buf3, buf4, 0x1000 );
2279 report( "cabac decision:" );
2281 set_func_name( "cabac_encode_bypass" );
2282 memcpy( buf4, buf3, 0x1000 );
2283 call_c( run_cabac_bypass_c, &h, buf3 );
2284 call_a( run_cabac_bypass_asm, &h, buf4 );
2285 ok = !memcmp( buf3, buf4, 0x1000 );
2286 report( "cabac bypass:" );
2288 set_func_name( "cabac_encode_terminal" );
2289 memcpy( buf4, buf3, 0x1000 );
2290 call_c( run_cabac_terminal_c, &h, buf3 );
2291 call_a( run_cabac_terminal_asm, &h, buf4 );
2292 ok = !memcmp( buf3, buf4, 0x1000 );
2293 report( "cabac terminal:" );
2298 static int check_bitstream( int cpu_ref, int cpu_new )
2300 x264_bitstream_function_t bs_c;
2301 x264_bitstream_function_t bs_ref;
2302 x264_bitstream_function_t bs_a;
2304 int ret = 0, ok = 1, used_asm = 0;
2306 x264_bitstream_init( 0, &bs_c );
2307 x264_bitstream_init( cpu_ref, &bs_ref );
2308 x264_bitstream_init( cpu_new, &bs_a );
2309 if( bs_a.nal_escape != bs_ref.nal_escape )
2312 uint8_t *input = malloc(size+100);
2313 uint8_t *output1 = malloc(size*2);
2314 uint8_t *output2 = malloc(size*2);
2316 set_func_name( "nal_escape" );
2317 for( int i = 0; i < 100; i++ )
2319 /* Test corner-case sizes */
2320 int test_size = i < 10 ? i+1 : rand() & 0x3fff;
2321 /* Test 8 different probability distributions of zeros */
2322 for( int j = 0; j < test_size+32; j++ )
2323 input[j] = (rand()&((1 << ((i&7)+1)) - 1)) * rand();
2324 uint8_t *end_c = (uint8_t*)call_c1( bs_c.nal_escape, output1, input, input+test_size );
2325 uint8_t *end_a = (uint8_t*)call_a1( bs_a.nal_escape, output2, input, input+test_size );
2326 int size_c = end_c-output1;
2327 int size_a = end_a-output2;
2328 if( size_c != size_a || memcmp( output1, output2, size_c ) )
2330 fprintf( stderr, "nal_escape : [FAILED] %d %d\n", size_c, size_a );
2335 for( int j = 0; j < size+32; j++ )
2337 call_c2( bs_c.nal_escape, output1, input, input+size );
2338 call_a2( bs_a.nal_escape, output2, input, input+size );
2343 report( "nal escape:" );
2348 static int check_all_funcs( int cpu_ref, int cpu_new )
2350 return check_pixel( cpu_ref, cpu_new )
2351 + check_dct( cpu_ref, cpu_new )
2352 + check_mc( cpu_ref, cpu_new )
2353 + check_intra( cpu_ref, cpu_new )
2354 + check_deblock( cpu_ref, cpu_new )
2355 + check_quant( cpu_ref, cpu_new )
2356 + check_cabac( cpu_ref, cpu_new )
2357 + check_bitstream( cpu_ref, cpu_new );
2360 static int add_flags( int *cpu_ref, int *cpu_new, int flags, const char *name )
2362 *cpu_ref = *cpu_new;
2364 #if BROKEN_STACK_ALIGNMENT
2365 *cpu_new |= X264_CPU_STACK_MOD4;
2367 if( *cpu_new & X264_CPU_SSE2_IS_FAST )
2368 *cpu_new &= ~X264_CPU_SSE2_IS_SLOW;
2370 fprintf( stderr, "x264: %s\n", name );
2371 return check_all_funcs( *cpu_ref, *cpu_new );
2374 static int check_all_flags( void )
2377 int cpu0 = 0, cpu1 = 0;
2379 if( x264_cpu_detect() & X264_CPU_MMX2 )
2381 ret |= add_flags( &cpu0, &cpu1, X264_CPU_MMX | X264_CPU_MMX2, "MMX" );
2382 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "MMX Cache64" );
2383 cpu1 &= ~X264_CPU_CACHELINE_64;
2385 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_32, "MMX Cache32" );
2386 cpu1 &= ~X264_CPU_CACHELINE_32;
2388 if( x264_cpu_detect() & X264_CPU_LZCNT )
2390 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "MMX_LZCNT" );
2391 cpu1 &= ~X264_CPU_LZCNT;
2393 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "MMX SlowCTZ" );
2394 cpu1 &= ~X264_CPU_SLOW_CTZ;
2396 if( x264_cpu_detect() & X264_CPU_SSE2 )
2398 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE | X264_CPU_SSE2 | X264_CPU_SSE2_IS_SLOW, "SSE2Slow" );
2399 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2_IS_FAST, "SSE2Fast" );
2400 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSE2Fast Cache64" );
2401 cpu1 &= ~X264_CPU_CACHELINE_64;
2402 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SHUFFLE_IS_FAST, "SSE2 FastShuffle" );
2403 cpu1 &= ~X264_CPU_SHUFFLE_IS_FAST;
2404 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSE2 SlowCTZ" );
2405 cpu1 &= ~X264_CPU_SLOW_CTZ;
2406 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSE2 SlowAtom" );
2407 cpu1 &= ~X264_CPU_SLOW_ATOM;
2409 if( x264_cpu_detect() & X264_CPU_SSE_MISALIGN )
2411 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE_MISALIGN, "SSE_Misalign" );
2412 cpu1 &= ~X264_CPU_SSE_MISALIGN;
2414 if( x264_cpu_detect() & X264_CPU_LZCNT )
2416 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSE_LZCNT" );
2417 cpu1 &= ~X264_CPU_LZCNT;
2419 if( x264_cpu_detect() & X264_CPU_SSE3 )
2421 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE3 | X264_CPU_CACHELINE_64, "SSE3" );
2422 cpu1 &= ~X264_CPU_CACHELINE_64;
2424 if( x264_cpu_detect() & X264_CPU_SSSE3 )
2426 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSSE3, "SSSE3" );
2427 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64" );
2428 cpu1 &= ~X264_CPU_CACHELINE_64;
2429 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SHUFFLE_IS_FAST, "SSSE3 FastShuffle" );
2430 cpu1 &= ~X264_CPU_SHUFFLE_IS_FAST;
2431 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSSE3 SlowCTZ" );
2432 cpu1 &= ~X264_CPU_SLOW_CTZ;
2433 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSSE3 SlowAtom" );
2434 cpu1 &= ~X264_CPU_SLOW_ATOM;
2436 if( x264_cpu_detect() & X264_CPU_SSE4 )
2437 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE4 | X264_CPU_SHUFFLE_IS_FAST, "SSE4" );
2438 if( x264_cpu_detect() & X264_CPU_AVX )
2439 ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX, "AVX" );
2440 if( x264_cpu_detect() & X264_CPU_XOP )
2441 ret |= add_flags( &cpu0, &cpu1, X264_CPU_XOP, "XOP" );
2442 if( x264_cpu_detect() & X264_CPU_FMA4 )
2444 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA4, "FMA4" );
2445 cpu1 &= ~X264_CPU_FMA4;
2447 if( x264_cpu_detect() & X264_CPU_FMA3 )
2449 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA3, "FMA3" );
2450 cpu1 &= ~X264_CPU_FMA3;
2452 if( x264_cpu_detect() & X264_CPU_BMI1 )
2454 ret |= add_flags( &cpu0, &cpu1, X264_CPU_BMI1, "BMI1" );
2455 if( x264_cpu_detect() & X264_CPU_TBM )
2457 ret |= add_flags( &cpu0, &cpu1, X264_CPU_TBM, "TBM" );
2458 cpu1 &= ~X264_CPU_TBM;
2460 if( x264_cpu_detect() & X264_CPU_BMI2 )
2462 ret |= add_flags( &cpu0, &cpu1, X264_CPU_BMI2, "BMI2" );
2463 cpu1 &= ~X264_CPU_BMI2;
2465 cpu1 &= ~X264_CPU_BMI1;
2467 if( x264_cpu_detect() & X264_CPU_AVX2 )
2468 ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX2, "AVX2" );
2470 if( x264_cpu_detect() & X264_CPU_ALTIVEC )
2472 fprintf( stderr, "x264: ALTIVEC against C\n" );
2473 ret = check_all_funcs( 0, X264_CPU_ALTIVEC );
2476 if( x264_cpu_detect() & X264_CPU_ARMV6 )
2477 ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV6, "ARMv6" );
2478 if( x264_cpu_detect() & X264_CPU_NEON )
2479 ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
2480 if( x264_cpu_detect() & X264_CPU_FAST_NEON_MRC )
2481 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FAST_NEON_MRC, "Fast NEON MRC" );
2486 int main(int argc, char *argv[])
2490 if( argc > 1 && !strncmp( argv[1], "--bench", 7 ) )
2492 #if !ARCH_X86 && !ARCH_X86_64 && !ARCH_PPC && !ARCH_ARM
2493 fprintf( stderr, "no --bench for your cpu until you port rdtsc\n" );
2497 if( argv[1][7] == '=' )
2499 bench_pattern = argv[1]+8;
2500 bench_pattern_len = strlen(bench_pattern);
2506 int seed = ( argc > 1 ) ? atoi(argv[1]) : x264_mdate();
2507 fprintf( stderr, "x264: using random seed %u\n", seed );
2510 buf1 = x264_malloc( 0x1e00 + 0x2000*sizeof(pixel) + 16*BENCH_ALIGNS );
2511 pbuf1 = x264_malloc( 0x1e00*sizeof(pixel) + 16*BENCH_ALIGNS );
2512 if( !buf1 || !pbuf1 )
2514 fprintf( stderr, "malloc failed, unable to initiate tests!\n" );
2517 #define INIT_POINTER_OFFSETS\
2518 buf2 = buf1 + 0xf00;\
2519 buf3 = buf2 + 0xf00;\
2520 buf4 = buf3 + 0x1000*sizeof(pixel);\
2521 pbuf2 = pbuf1 + 0xf00;\
2522 pbuf3 = (pixel*)buf3;\
2523 pbuf4 = (pixel*)buf4;
2524 INIT_POINTER_OFFSETS;
2525 for( int i = 0; i < 0x1e00; i++ )
2527 buf1[i] = rand() & 0xFF;
2528 pbuf1[i] = rand() & PIXEL_MAX;
2530 memset( buf1+0x1e00, 0, 0x2000*sizeof(pixel) );
2532 /* 16-byte alignment is guaranteed whenever it's useful, but some functions also vary in speed depending on %64 */
2534 for( int i = 0; i < BENCH_ALIGNS && !ret; i++ )
2536 INIT_POINTER_OFFSETS;
2537 ret |= x264_stack_pagealign( check_all_flags, i*16 );
2541 fprintf( stderr, "%d/%d\r", i+1, BENCH_ALIGNS );
2544 ret = check_all_flags();
2548 fprintf( stderr, "x264: at least one test has failed. Go and fix that Right Now!\n" );
2551 fprintf( stderr, "x264: All tests passed Yeah :)\n" );