1 /*****************************************************************************
2 * checkasm.c: assembly check tool
3 *****************************************************************************
4 * Copyright (C) 2003-2015 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( "lfence \n"
95 : "=a"(a) :: "edx", "memory" );
97 asm volatile( "mftb %0" : "=r"(a) :: "memory" );
98 #elif ARCH_ARM // ARMv7 only
99 asm volatile( "mrc p15, 0, %0, c9, c13, 0" : "=r"(a) :: "memory" );
102 asm volatile( "mrs %0, pmccntr_el0" : "=r"(b) :: "memory" );
105 asm volatile( "rdhwr %0, $2" : "=r"(a) :: "memory" );
110 static bench_t* get_bench( const char *name, int cpu )
113 for( i = 0; benchs[i].name && strcmp(name, benchs[i].name); i++ )
114 assert( i < MAX_FUNCS );
115 if( !benchs[i].name )
116 benchs[i].name = strdup( name );
118 return &benchs[i].vers[0];
119 for( j = 1; benchs[i].vers[j].cpu && benchs[i].vers[j].cpu != cpu; j++ )
120 assert( j < MAX_CPUS );
121 benchs[i].vers[j].cpu = cpu;
122 return &benchs[i].vers[j];
125 static int cmp_nop( const void *a, const void *b )
127 return *(uint16_t*)a - *(uint16_t*)b;
130 static int cmp_bench( const void *a, const void *b )
132 // asciibetical sort except preserving numbers
133 const char *sa = ((bench_func_t*)a)->name;
134 const char *sb = ((bench_func_t*)b)->name;
139 if( isdigit( *sa ) && isdigit( *sb ) && isdigit( sa[1] ) != isdigit( sb[1] ) )
140 return isdigit( sa[1] ) - isdigit( sb[1] );
146 static void print_bench(void)
148 uint16_t nops[10000];
149 int nfuncs, nop_time=0;
151 for( int i = 0; i < 10000; i++ )
153 uint32_t t = read_time();
154 nops[i] = read_time() - t;
156 qsort( nops, 10000, sizeof(uint16_t), cmp_nop );
157 for( int i = 500; i < 9500; i++ )
160 printf( "nop: %d\n", nop_time );
162 for( nfuncs = 0; nfuncs < MAX_FUNCS && benchs[nfuncs].name; nfuncs++ );
163 qsort( benchs, nfuncs, sizeof(bench_func_t), cmp_bench );
164 for( int i = 0; i < nfuncs; i++ )
165 for( int j = 0; j < MAX_CPUS && (!j || benchs[i].vers[j].cpu); j++ )
168 bench_t *b = &benchs[i].vers[j];
171 for( k = 0; k < j && benchs[i].vers[k].pointer != b->pointer; k++ );
174 printf( "%s_%s%s: %"PRId64"\n", benchs[i].name,
176 b->cpu&X264_CPU_AVX2 ? "avx2" :
177 b->cpu&X264_CPU_FMA3 ? "fma3" :
178 b->cpu&X264_CPU_FMA4 ? "fma4" :
179 b->cpu&X264_CPU_XOP ? "xop" :
180 b->cpu&X264_CPU_AVX ? "avx" :
181 b->cpu&X264_CPU_SSE42 ? "sse42" :
182 b->cpu&X264_CPU_SSE4 ? "sse4" :
183 b->cpu&X264_CPU_SSSE3 ? "ssse3" :
184 b->cpu&X264_CPU_SSE3 ? "sse3" :
185 /* print sse2slow only if there's also a sse2fast version of the same func */
186 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" :
187 b->cpu&X264_CPU_SSE2 ? "sse2" :
188 b->cpu&X264_CPU_SSE ? "sse" :
189 b->cpu&X264_CPU_MMX ? "mmx" :
191 b->cpu&X264_CPU_ALTIVEC ? "altivec" :
193 b->cpu&X264_CPU_NEON ? "neon" :
194 b->cpu&X264_CPU_ARMV6 ? "armv6" :
196 b->cpu&X264_CPU_NEON ? "neon" :
197 b->cpu&X264_CPU_ARMV8 ? "armv8" :
199 b->cpu&X264_CPU_MSA ? "msa" :
203 b->cpu&X264_CPU_CACHELINE_32 ? "_c32" :
204 b->cpu&X264_CPU_SLOW_ATOM && b->cpu&X264_CPU_CACHELINE_64 ? "_c64_atom" :
205 b->cpu&X264_CPU_CACHELINE_64 ? "_c64" :
206 b->cpu&X264_CPU_SLOW_SHUFFLE ? "_slowshuffle" :
207 b->cpu&X264_CPU_LZCNT ? "_lzcnt" :
208 b->cpu&X264_CPU_BMI2 ? "_bmi2" :
209 b->cpu&X264_CPU_BMI1 ? "_bmi1" :
210 b->cpu&X264_CPU_SLOW_CTZ ? "_slow_ctz" :
211 b->cpu&X264_CPU_SLOW_ATOM ? "_atom" :
213 b->cpu&X264_CPU_FAST_NEON_MRC ? "_fast_mrc" :
216 (int64_t)(10*b->cycles/b->den - nop_time)/4 );
220 #if ARCH_X86 || ARCH_X86_64
221 int x264_stack_pagealign( int (*func)(), int align );
223 /* detect when callee-saved regs aren't saved
224 * needs an explicit asm check because it only sometimes crashes in normal use. */
225 intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
227 #define x264_stack_pagealign( func, align ) func()
231 intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
234 #define call_c1(func,...) func(__VA_ARGS__)
237 /* Evil hack: detect incorrect assumptions that 32-bit ints are zero-extended to 64-bit.
238 * This is done by clobbering the stack with junk around the stack pointer and calling the
239 * assembly function through x264_checkasm_call with added dummy arguments which forces all
240 * real arguments to be passed on the stack and not in registers. For 32-bit argument the
241 * upper half of the 64-bit register location on the stack will now contain junk. Note that
242 * this is dependant on compiler behaviour and that interrupts etc. at the wrong time may
243 * overwrite the junk written to the stack so there's no guarantee that it will always
244 * detect all functions that assumes zero-extension.
246 void x264_checkasm_stack_clobber( uint64_t clobber, ... );
247 #define call_a1(func,...) ({ \
248 uint64_t r = (rand() & 0xffff) * 0x0001000100010001ULL; \
249 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 */ \
250 x264_checkasm_call(( intptr_t(*)())func, &ok, 0, 0, 0, 0, __VA_ARGS__ ); })
251 #elif ARCH_X86 || (ARCH_AARCH64 && !defined(__APPLE__))
252 #define call_a1(func,...) x264_checkasm_call( (intptr_t(*)())func, &ok, __VA_ARGS__ )
254 #define call_a1 call_c1
257 #define call_bench(func,cpu,...)\
258 if( do_bench && !strncmp(func_name, bench_pattern, bench_pattern_len) )\
262 call_a1(func, __VA_ARGS__);\
263 for( int ti = 0; ti < (cpu?BENCH_RUNS:BENCH_RUNS/4); ti++ )\
265 uint32_t t = read_time();\
270 t = read_time() - t;\
271 if( (uint64_t)t*tcount <= tsum*4 && ti > 0 )\
277 bench_t *b = get_bench( func_name, cpu );\
283 /* for most functions, run benchmark and correctness test at the same time.
284 * for those that modify their inputs, run the above macros separately */
285 #define call_a(func,...) ({ call_a2(func,__VA_ARGS__); call_a1(func,__VA_ARGS__); })
286 #define call_c(func,...) ({ call_c2(func,__VA_ARGS__); call_c1(func,__VA_ARGS__); })
287 #define call_a2(func,...) ({ call_bench(func,cpu_new,__VA_ARGS__); })
288 #define call_c2(func,...) ({ call_bench(func,0,__VA_ARGS__); })
291 static int check_pixel( int cpu_ref, int cpu_new )
293 x264_pixel_function_t pixel_c;
294 x264_pixel_function_t pixel_ref;
295 x264_pixel_function_t pixel_asm;
296 x264_predict_t predict_4x4[12];
297 x264_predict8x8_t predict_8x8[12];
298 x264_predict_8x8_filter_t predict_8x8_filter;
299 ALIGNED_16( pixel edge[36] );
300 uint16_t cost_mv[32];
301 int ret = 0, ok, used_asm;
303 x264_pixel_init( 0, &pixel_c );
304 x264_pixel_init( cpu_ref, &pixel_ref );
305 x264_pixel_init( cpu_new, &pixel_asm );
306 x264_predict_4x4_init( 0, predict_4x4 );
307 x264_predict_8x8_init( 0, predict_8x8, &predict_8x8_filter );
308 predict_8x8_filter( pbuf2+40, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
311 for( int i = 0; i < 256; i++ )
316 pbuf4[i] = -(z&1) & PIXEL_MAX;
317 pbuf3[i] = ~pbuf4[i] & PIXEL_MAX;
319 // random pattern made of maxed pixel differences, in case an intermediate value overflows
320 for( int i = 256; i < 0x1000; i++ )
322 pbuf4[i] = -(pbuf1[i&~0x88]&1) & PIXEL_MAX;
323 pbuf3[i] = ~(pbuf4[i]) & PIXEL_MAX;
326 #define TEST_PIXEL( name, align ) \
327 ok = 1, used_asm = 0; \
328 for( int i = 0; i < ARRAY_ELEMS(pixel_c.name); i++ ) \
330 int res_c, res_asm; \
331 if( pixel_asm.name[i] != pixel_ref.name[i] ) \
333 set_func_name( "%s_%s", #name, pixel_names[i] ); \
335 for( int j = 0; j < 64; j++ ) \
337 res_c = call_c( pixel_c.name[i], pbuf1, (intptr_t)16, pbuf2+j*!align, (intptr_t)64 ); \
338 res_asm = call_a( pixel_asm.name[i], pbuf1, (intptr_t)16, pbuf2+j*!align, (intptr_t)64 ); \
339 if( res_c != res_asm ) \
342 fprintf( stderr, #name "[%d]: %d != %d [FAILED]\n", i, res_c, res_asm ); \
346 for( int j = 0; j < 0x1000 && ok; j += 256 ) \
348 res_c = pixel_c .name[i]( pbuf3+j, 16, pbuf4+j, 16 ); \
349 res_asm = pixel_asm.name[i]( pbuf3+j, 16, pbuf4+j, 16 ); \
350 if( res_c != res_asm ) \
353 fprintf( stderr, #name "[%d]: overflow %d != %d\n", i, res_c, res_asm ); \
358 report( "pixel " #name " :" );
360 TEST_PIXEL( sad, 0 );
361 TEST_PIXEL( sad_aligned, 1 );
362 TEST_PIXEL( ssd, 1 );
363 TEST_PIXEL( satd, 0 );
364 TEST_PIXEL( sa8d, 1 );
366 ok = 1, used_asm = 0;
367 if( pixel_asm.sa8d_satd[PIXEL_16x16] != pixel_ref.sa8d_satd[PIXEL_16x16] )
369 set_func_name( "sa8d_satd_%s", pixel_names[PIXEL_16x16] );
371 for( int j = 0; j < 64; j++ )
373 uint32_t cost8_c = pixel_c.sa8d[PIXEL_16x16]( pbuf1, 16, pbuf2, 64 );
374 uint32_t cost4_c = pixel_c.satd[PIXEL_16x16]( pbuf1, 16, pbuf2, 64 );
375 uint64_t res_a = call_a( pixel_asm.sa8d_satd[PIXEL_16x16], pbuf1, (intptr_t)16, pbuf2, (intptr_t)64 );
376 uint32_t cost8_a = res_a;
377 uint32_t cost4_a = res_a >> 32;
378 if( cost8_a != cost8_c || cost4_a != cost4_c )
381 fprintf( stderr, "sa8d_satd [%d]: (%d,%d) != (%d,%d) [FAILED]\n", PIXEL_16x16,
382 cost8_c, cost4_c, cost8_a, cost4_a );
386 for( int j = 0; j < 0x1000 && ok; j += 256 ) \
388 uint32_t cost8_c = pixel_c.sa8d[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
389 uint32_t cost4_c = pixel_c.satd[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
390 uint64_t res_a = pixel_asm.sa8d_satd[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
391 uint32_t cost8_a = res_a;
392 uint32_t cost4_a = res_a >> 32;
393 if( cost8_a != cost8_c || cost4_a != cost4_c )
396 fprintf( stderr, "sa8d_satd [%d]: overflow (%d,%d) != (%d,%d) [FAILED]\n", PIXEL_16x16,
397 cost8_c, cost4_c, cost8_a, cost4_a );
401 report( "pixel sa8d_satd :" );
403 #define TEST_PIXEL_X( N ) \
404 ok = 1; used_asm = 0; \
405 for( int i = 0; i < 7; i++ ) \
407 ALIGNED_16( int res_c[4] ) = {0}; \
408 ALIGNED_16( int res_asm[4] ) = {0}; \
409 if( pixel_asm.sad_x##N[i] && pixel_asm.sad_x##N[i] != pixel_ref.sad_x##N[i] ) \
411 set_func_name( "sad_x%d_%s", N, pixel_names[i] ); \
413 for( int j = 0; j < 64; j++ ) \
415 pixel *pix2 = pbuf2+j; \
416 res_c[0] = pixel_c.sad[i]( pbuf1, 16, pix2, 64 ); \
417 res_c[1] = pixel_c.sad[i]( pbuf1, 16, pix2+6, 64 ); \
418 res_c[2] = pixel_c.sad[i]( pbuf1, 16, pix2+1, 64 ); \
421 res_c[3] = pixel_c.sad[i]( pbuf1, 16, pix2+10, 64 ); \
422 call_a( pixel_asm.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, (intptr_t)64, res_asm ); \
425 call_a( pixel_asm.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, (intptr_t)64, res_asm ); \
426 if( memcmp(res_c, res_asm, N*sizeof(int)) ) \
429 fprintf( stderr, "sad_x"#N"[%d]: %d,%d,%d,%d != %d,%d,%d,%d [FAILED]\n", \
430 i, res_c[0], res_c[1], res_c[2], res_c[3], \
431 res_asm[0], res_asm[1], res_asm[2], res_asm[3] ); \
434 call_c2( pixel_c.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, (intptr_t)64, res_asm ); \
436 call_c2( pixel_c.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, (intptr_t)64, res_asm ); \
440 report( "pixel sad_x"#N" :" );
445 #define TEST_PIXEL_VAR( i ) \
446 if( pixel_asm.var[i] != pixel_ref.var[i] ) \
448 set_func_name( "%s_%s", "var", pixel_names[i] ); \
450 /* abi-check wrapper can't return uint64_t, so separate it from return value check */ \
451 call_c1( pixel_c.var[i], pbuf1, 16 ); \
452 call_a1( pixel_asm.var[i], pbuf1, (intptr_t)16 ); \
453 uint64_t res_c = pixel_c.var[i]( pbuf1, 16 ); \
454 uint64_t res_asm = pixel_asm.var[i]( pbuf1, 16 ); \
455 if( res_c != res_asm ) \
458 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) ); \
460 call_c2( pixel_c.var[i], pbuf1, (intptr_t)16 ); \
461 call_a2( pixel_asm.var[i], pbuf1, (intptr_t)16 ); \
464 ok = 1; used_asm = 0;
465 TEST_PIXEL_VAR( PIXEL_16x16 );
466 TEST_PIXEL_VAR( PIXEL_8x16 );
467 TEST_PIXEL_VAR( PIXEL_8x8 );
468 report( "pixel var :" );
470 #define TEST_PIXEL_VAR2( i ) \
471 if( pixel_asm.var2[i] != pixel_ref.var2[i] ) \
473 int res_c, res_asm, ssd_c, ssd_asm; \
474 set_func_name( "%s_%s", "var2", pixel_names[i] ); \
476 res_c = call_c( pixel_c.var2[i], pbuf1, (intptr_t)16, pbuf2, (intptr_t)16, &ssd_c ); \
477 res_asm = call_a( pixel_asm.var2[i], pbuf1, (intptr_t)16, pbuf2, (intptr_t)16, &ssd_asm ); \
478 if( res_c != res_asm || ssd_c != ssd_asm ) \
481 fprintf( stderr, "var2[%d]: %d != %d or %d != %d [FAILED]\n", i, res_c, res_asm, ssd_c, ssd_asm ); \
485 ok = 1; used_asm = 0;
486 TEST_PIXEL_VAR2( PIXEL_8x16 );
487 TEST_PIXEL_VAR2( PIXEL_8x8 );
488 report( "pixel var2 :" );
490 ok = 1; used_asm = 0;
491 for( int i = 0; i < 4; i++ )
492 if( pixel_asm.hadamard_ac[i] != pixel_ref.hadamard_ac[i] )
494 set_func_name( "hadamard_ac_%s", pixel_names[i] );
496 for( int j = 0; j < 32; j++ )
498 pixel *pix = (j&16 ? pbuf1 : pbuf3) + (j&15)*256;
499 call_c1( pixel_c.hadamard_ac[i], pbuf1, (intptr_t)16 );
500 call_a1( pixel_asm.hadamard_ac[i], pbuf1, (intptr_t)16 );
501 uint64_t rc = pixel_c.hadamard_ac[i]( pix, 16 );
502 uint64_t ra = pixel_asm.hadamard_ac[i]( pix, 16 );
506 fprintf( stderr, "hadamard_ac[%d]: %d,%d != %d,%d\n", i, (int)rc, (int)(rc>>32), (int)ra, (int)(ra>>32) );
510 call_c2( pixel_c.hadamard_ac[i], pbuf1, (intptr_t)16 );
511 call_a2( pixel_asm.hadamard_ac[i], pbuf1, (intptr_t)16 );
513 report( "pixel hadamard_ac :" );
516 for( int i = 0; i < 32; i++ )
517 for( int j = 0; j < 16; j++ )
518 pbuf4[16*i+j] = -((i+j)&1) & PIXEL_MAX;
519 ok = 1; used_asm = 0;
520 if( pixel_asm.vsad != pixel_ref.vsad )
522 for( int h = 2; h <= 32; h += 2 )
525 set_func_name( "vsad" );
527 for( int j = 0; j < 2 && ok; j++ )
529 pixel *p = j ? pbuf4 : pbuf1;
530 res_c = call_c( pixel_c.vsad, p, (intptr_t)16, h );
531 res_asm = call_a( pixel_asm.vsad, p, (intptr_t)16, h );
532 if( res_c != res_asm )
535 fprintf( stderr, "vsad: height=%d, %d != %d\n", h, res_c, res_asm );
541 report( "pixel vsad :" );
543 ok = 1; used_asm = 0;
544 if( pixel_asm.asd8 != pixel_ref.asd8 )
546 set_func_name( "asd8" );
548 int res_c = call_c( pixel_c.asd8, pbuf1, (intptr_t)8, pbuf2, (intptr_t)8, 16 );
549 int res_a = call_a( pixel_asm.asd8, pbuf1, (intptr_t)8, pbuf2, (intptr_t)8, 16 );
553 fprintf( stderr, "asd: %d != %d\n", res_c, res_a );
556 report( "pixel asd :" );
558 #define TEST_INTRA_X3( name, i8x8, ... ) \
559 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
561 ALIGNED_16( int res_c[3] ); \
562 ALIGNED_16( int res_asm[3] ); \
563 set_func_name( #name ); \
565 call_c( pixel_c.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_c ); \
566 call_a( pixel_asm.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_asm ); \
567 if( memcmp(res_c, res_asm, sizeof(res_c)) ) \
570 fprintf( stderr, #name": %d,%d,%d != %d,%d,%d [FAILED]\n", \
571 res_c[0], res_c[1], res_c[2], \
572 res_asm[0], res_asm[1], res_asm[2] ); \
576 #define TEST_INTRA_X9( name, cmp ) \
577 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
579 set_func_name( #name ); \
581 ALIGNED_ARRAY_64( uint16_t, bitcosts,[17] ); \
582 for( int i=0; i<17; i++ ) \
583 bitcosts[i] = 9*(i!=8); \
584 memcpy( pbuf3, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) ); \
585 memcpy( pbuf4, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) ); \
586 for( int i=0; i<32; i++ ) \
588 pixel *fenc = pbuf1+48+i*12; \
589 pixel *fdec1 = pbuf3+48+i*12; \
590 pixel *fdec2 = pbuf4+48+i*12; \
591 int pred_mode = i%9; \
592 int res_c = INT_MAX; \
593 for( int j=0; j<9; j++ ) \
595 predict_4x4[j]( fdec1 ); \
596 int cost = pixel_c.cmp[PIXEL_4x4]( fenc, FENC_STRIDE, fdec1, FDEC_STRIDE ) + 9*(j!=pred_mode); \
597 if( cost < (uint16_t)res_c ) \
598 res_c = cost + (j<<16); \
600 predict_4x4[res_c>>16]( fdec1 ); \
601 int res_a = call_a( pixel_asm.name, fenc, fdec2, bitcosts+8-pred_mode ); \
602 if( res_c != res_a ) \
605 fprintf( stderr, #name": %d,%d != %d,%d [FAILED]\n", res_c>>16, res_c&0xffff, res_a>>16, res_a&0xffff ); \
608 if( memcmp(fdec1, fdec2, 4*FDEC_STRIDE*sizeof(pixel)) ) \
611 fprintf( stderr, #name" [FAILED]\n" ); \
612 for( int j=0; j<16; j++ ) \
613 fprintf( stderr, "%02x ", fdec1[(j&3)+(j>>2)*FDEC_STRIDE] ); \
614 fprintf( stderr, "\n" ); \
615 for( int j=0; j<16; j++ ) \
616 fprintf( stderr, "%02x ", fdec2[(j&3)+(j>>2)*FDEC_STRIDE] ); \
617 fprintf( stderr, "\n" ); \
623 #define TEST_INTRA8_X9( name, cmp ) \
624 if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
626 set_func_name( #name ); \
628 ALIGNED_ARRAY_64( uint16_t, bitcosts,[17] ); \
629 ALIGNED_ARRAY_16( uint16_t, satds_c,[16] ); \
630 ALIGNED_ARRAY_16( uint16_t, satds_a,[16] ); \
631 memset( satds_c, 0, 16 * sizeof(*satds_c) ); \
632 memset( satds_a, 0, 16 * sizeof(*satds_a) ); \
633 for( int i=0; i<17; i++ ) \
634 bitcosts[i] = 9*(i!=8); \
635 for( int i=0; i<32; i++ ) \
637 pixel *fenc = pbuf1+48+i*12; \
638 pixel *fdec1 = pbuf3+48+i*12; \
639 pixel *fdec2 = pbuf4+48+i*12; \
640 int pred_mode = i%9; \
641 int res_c = INT_MAX; \
642 predict_8x8_filter( fdec1, edge, ALL_NEIGHBORS, ALL_NEIGHBORS ); \
643 for( int j=0; j<9; j++ ) \
645 predict_8x8[j]( fdec1, edge ); \
646 satds_c[j] = pixel_c.cmp[PIXEL_8x8]( fenc, FENC_STRIDE, fdec1, FDEC_STRIDE ) + 9*(j!=pred_mode); \
647 if( satds_c[j] < (uint16_t)res_c ) \
648 res_c = satds_c[j] + (j<<16); \
650 predict_8x8[res_c>>16]( fdec1, edge ); \
651 int res_a = call_a( pixel_asm.name, fenc, fdec2, edge, bitcosts+8-pred_mode, satds_a ); \
652 if( res_c != res_a || memcmp(satds_c, satds_a, 16 * sizeof(*satds_c)) ) \
655 fprintf( stderr, #name": %d,%d != %d,%d [FAILED]\n", res_c>>16, res_c&0xffff, res_a>>16, res_a&0xffff ); \
656 for( int j = 0; j < 9; j++ ) \
657 fprintf( stderr, "%5d ", satds_c[j]); \
658 fprintf( stderr, "\n" ); \
659 for( int j = 0; j < 9; j++ ) \
660 fprintf( stderr, "%5d ", satds_a[j]); \
661 fprintf( stderr, "\n" ); \
664 for( int j=0; j<8; j++ ) \
665 if( memcmp(fdec1+j*FDEC_STRIDE, fdec2+j*FDEC_STRIDE, 8*sizeof(pixel)) ) \
669 fprintf( stderr, #name" [FAILED]\n" ); \
670 for( int j=0; j<8; j++ ) \
672 for( int k=0; k<8; k++ ) \
673 fprintf( stderr, "%02x ", fdec1[k+j*FDEC_STRIDE] ); \
674 fprintf( stderr, "\n" ); \
676 fprintf( stderr, "\n" ); \
677 for( int j=0; j<8; j++ ) \
679 for( int k=0; k<8; k++ ) \
680 fprintf( stderr, "%02x ", fdec2[k+j*FDEC_STRIDE] ); \
681 fprintf( stderr, "\n" ); \
683 fprintf( stderr, "\n" ); \
689 memcpy( pbuf3, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) );
690 ok = 1; used_asm = 0;
691 TEST_INTRA_X3( intra_satd_x3_16x16, 0 );
692 TEST_INTRA_X3( intra_satd_x3_8x16c, 0 );
693 TEST_INTRA_X3( intra_satd_x3_8x8c, 0 );
694 TEST_INTRA_X3( intra_sa8d_x3_8x8, 1, edge );
695 TEST_INTRA_X3( intra_satd_x3_4x4, 0 );
696 report( "intra satd_x3 :" );
697 ok = 1; used_asm = 0;
698 TEST_INTRA_X3( intra_sad_x3_16x16, 0 );
699 TEST_INTRA_X3( intra_sad_x3_8x16c, 0 );
700 TEST_INTRA_X3( intra_sad_x3_8x8c, 0 );
701 TEST_INTRA_X3( intra_sad_x3_8x8, 1, edge );
702 TEST_INTRA_X3( intra_sad_x3_4x4, 0 );
703 report( "intra sad_x3 :" );
704 ok = 1; used_asm = 0;
705 TEST_INTRA_X9( intra_satd_x9_4x4, satd );
706 TEST_INTRA8_X9( intra_sa8d_x9_8x8, sa8d );
707 report( "intra satd_x9 :" );
708 ok = 1; used_asm = 0;
709 TEST_INTRA_X9( intra_sad_x9_4x4, sad );
710 TEST_INTRA8_X9( intra_sad_x9_8x8, sad );
711 report( "intra sad_x9 :" );
713 ok = 1; used_asm = 0;
714 if( pixel_asm.ssd_nv12_core != pixel_ref.ssd_nv12_core )
717 set_func_name( "ssd_nv12" );
718 uint64_t res_u_c, res_v_c, res_u_a, res_v_a;
719 pixel_c.ssd_nv12_core( pbuf1, 368, pbuf2, 368, 360, 8, &res_u_c, &res_v_c );
720 pixel_asm.ssd_nv12_core( pbuf1, 368, pbuf2, 368, 360, 8, &res_u_a, &res_v_a );
721 if( res_u_c != res_u_a || res_v_c != res_v_a )
724 fprintf( stderr, "ssd_nv12: %"PRIu64",%"PRIu64" != %"PRIu64",%"PRIu64"\n",
725 res_u_c, res_v_c, res_u_a, res_v_a );
727 call_c( pixel_c.ssd_nv12_core, pbuf1, (intptr_t)368, pbuf2, (intptr_t)368, 360, 8, &res_u_c, &res_v_c );
728 call_a( pixel_asm.ssd_nv12_core, pbuf1, (intptr_t)368, pbuf2, (intptr_t)368, 360, 8, &res_u_a, &res_v_a );
730 report( "ssd_nv12 :" );
732 if( pixel_asm.ssim_4x4x2_core != pixel_ref.ssim_4x4x2_core ||
733 pixel_asm.ssim_end4 != pixel_ref.ssim_end4 )
737 ALIGNED_16( int sums[5][4] ) = {{0}};
740 res_c = x264_pixel_ssim_wxh( &pixel_c, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
741 res_a = x264_pixel_ssim_wxh( &pixel_asm, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
742 if( fabs( res_c - res_a ) > 1e-6 )
745 fprintf( stderr, "ssim: %.7f != %.7f [FAILED]\n", res_c, res_a );
747 set_func_name( "ssim_core" );
748 call_c( pixel_c.ssim_4x4x2_core, pbuf1+2, (intptr_t)32, pbuf2+2, (intptr_t)32, sums );
749 call_a( pixel_asm.ssim_4x4x2_core, pbuf1+2, (intptr_t)32, pbuf2+2, (intptr_t)32, sums );
750 set_func_name( "ssim_end" );
751 call_c2( pixel_c.ssim_end4, sums, sums, 4 );
752 call_a2( pixel_asm.ssim_end4, sums, sums, 4 );
753 /* check incorrect assumptions that 32-bit ints are zero-extended to 64-bit */
754 call_c1( pixel_c.ssim_end4, sums, sums, 3 );
755 call_a1( pixel_asm.ssim_end4, sums, sums, 3 );
759 ok = 1; used_asm = 0;
760 for( int i = 0; i < 32; i++ )
762 for( int i = 0; i < 100 && ok; i++ )
763 if( pixel_asm.ads[i&3] != pixel_ref.ads[i&3] )
765 ALIGNED_16( uint16_t sums[72] );
766 ALIGNED_16( int dc[4] );
767 ALIGNED_16( int16_t mvs_a[48] );
768 ALIGNED_16( int16_t mvs_c[48] );
770 int thresh = rand() & 0x3fff;
771 set_func_name( "esa_ads" );
772 for( int j = 0; j < 72; j++ )
773 sums[j] = rand() & 0x3fff;
774 for( int j = 0; j < 4; j++ )
775 dc[j] = rand() & 0x3fff;
777 mvn_c = call_c( pixel_c.ads[i&3], dc, sums, 32, cost_mv, mvs_c, 28, thresh );
778 mvn_a = call_a( pixel_asm.ads[i&3], dc, sums, 32, cost_mv, mvs_a, 28, thresh );
779 if( mvn_c != mvn_a || memcmp( mvs_c, mvs_a, mvn_c*sizeof(*mvs_c) ) )
782 printf( "c%d: ", i&3 );
783 for( int j = 0; j < mvn_c; j++ )
784 printf( "%d ", mvs_c[j] );
785 printf( "\na%d: ", i&3 );
786 for( int j = 0; j < mvn_a; j++ )
787 printf( "%d ", mvs_a[j] );
791 report( "esa ads:" );
796 static int check_dct( int cpu_ref, int cpu_new )
798 x264_dct_function_t dct_c;
799 x264_dct_function_t dct_ref;
800 x264_dct_function_t dct_asm;
801 x264_quant_function_t qf;
802 int ret = 0, ok, used_asm, interlace = 0;
803 ALIGNED_ARRAY_N( dctcoef, dct1, [16],[16] );
804 ALIGNED_ARRAY_N( dctcoef, dct2, [16],[16] );
805 ALIGNED_ARRAY_N( dctcoef, dct4, [16],[16] );
806 ALIGNED_ARRAY_N( dctcoef, dct8, [4],[64] );
807 ALIGNED_16( dctcoef dctdc[2][8] );
811 x264_dct_init( 0, &dct_c );
812 x264_dct_init( cpu_ref, &dct_ref);
813 x264_dct_init( cpu_new, &dct_asm );
815 memset( h, 0, sizeof(*h) );
816 x264_param_default( &h->param );
817 h->sps->i_chroma_format_idc = 1;
818 h->chroma_qp_table = i_chroma_qp_table + 12;
819 h->param.analyse.i_luma_deadzone[0] = 0;
820 h->param.analyse.i_luma_deadzone[1] = 0;
821 h->param.analyse.b_transform_8x8 = 1;
822 for( int i = 0; i < 6; i++ )
823 h->pps->scaling_list[i] = x264_cqm_flat16;
825 x264_quant_init( h, 0, &qf );
827 /* overflow test cases */
828 for( int i = 0; i < 5; i++ )
830 pixel *enc = &pbuf3[16*i*FENC_STRIDE];
831 pixel *dec = &pbuf4[16*i*FDEC_STRIDE];
833 for( int j = 0; j < 16; j++ )
835 int cond_a = (i < 2) ? 1 : ((j&3) == 0 || (j&3) == (i-1));
836 int cond_b = (i == 0) ? 1 : !cond_a;
837 enc[0] = enc[1] = enc[4] = enc[5] = enc[8] = enc[9] = enc[12] = enc[13] = cond_a ? PIXEL_MAX : 0;
838 enc[2] = enc[3] = enc[6] = enc[7] = enc[10] = enc[11] = enc[14] = enc[15] = cond_b ? PIXEL_MAX : 0;
840 for( int k = 0; k < 4; k++ )
841 dec[k] = PIXEL_MAX - enc[k];
848 #define TEST_DCT( name, t1, t2, size ) \
849 if( dct_asm.name != dct_ref.name ) \
851 set_func_name( #name ); \
853 pixel *enc = pbuf3; \
854 pixel *dec = pbuf4; \
855 for( int j = 0; j < 5; j++) \
857 call_c( dct_c.name, t1, &pbuf1[j*64], &pbuf2[j*64] ); \
858 call_a( dct_asm.name, t2, &pbuf1[j*64], &pbuf2[j*64] ); \
859 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
862 fprintf( stderr, #name " [FAILED]\n" ); \
863 for( int k = 0; k < size; k++ )\
864 printf( "%d ", ((dctcoef*)t1)[k] );\
866 for( int k = 0; k < size; k++ )\
867 printf( "%d ", ((dctcoef*)t2)[k] );\
871 call_c( dct_c.name, t1, enc, dec ); \
872 call_a( dct_asm.name, t2, enc, dec ); \
873 if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
876 fprintf( stderr, #name " [FAILED] (overflow)\n" ); \
879 enc += 16*FENC_STRIDE; \
880 dec += 16*FDEC_STRIDE; \
883 ok = 1; used_asm = 0;
884 TEST_DCT( sub4x4_dct, dct1[0], dct2[0], 16 );
885 TEST_DCT( sub8x8_dct, dct1, dct2, 16*4 );
886 TEST_DCT( sub8x8_dct_dc, dctdc[0], dctdc[1], 4 );
887 TEST_DCT( sub8x16_dct_dc, dctdc[0], dctdc[1], 8 );
888 TEST_DCT( sub16x16_dct, dct1, dct2, 16*16 );
889 report( "sub_dct4 :" );
891 ok = 1; used_asm = 0;
892 TEST_DCT( sub8x8_dct8, (void*)dct1[0], (void*)dct2[0], 64 );
893 TEST_DCT( sub16x16_dct8, (void*)dct1, (void*)dct2, 64*4 );
894 report( "sub_dct8 :" );
897 // fdct and idct are denormalized by different factors, so quant/dequant
898 // is needed to force the coefs into the right range.
899 dct_c.sub16x16_dct( dct4, pbuf1, pbuf2 );
900 dct_c.sub16x16_dct8( dct8, pbuf1, pbuf2 );
901 for( int i = 0; i < 16; i++ )
903 qf.quant_4x4( dct4[i], h->quant4_mf[CQM_4IY][20], h->quant4_bias[CQM_4IY][20] );
904 qf.dequant_4x4( dct4[i], h->dequant4_mf[CQM_4IY], 20 );
906 for( int i = 0; i < 4; i++ )
908 qf.quant_8x8( dct8[i], h->quant8_mf[CQM_8IY][20], h->quant8_bias[CQM_8IY][20] );
909 qf.dequant_8x8( dct8[i], h->dequant8_mf[CQM_8IY], 20 );
911 x264_cqm_delete( h );
913 #define TEST_IDCT( name, src ) \
914 if( dct_asm.name != dct_ref.name ) \
916 set_func_name( #name ); \
918 memcpy( pbuf3, pbuf1, 32*32 * sizeof(pixel) ); \
919 memcpy( pbuf4, pbuf1, 32*32 * sizeof(pixel) ); \
920 memcpy( dct1, src, 256 * sizeof(dctcoef) ); \
921 memcpy( dct2, src, 256 * sizeof(dctcoef) ); \
922 call_c1( dct_c.name, pbuf3, (void*)dct1 ); \
923 call_a1( dct_asm.name, pbuf4, (void*)dct2 ); \
924 if( memcmp( pbuf3, pbuf4, 32*32 * sizeof(pixel) ) ) \
927 fprintf( stderr, #name " [FAILED]\n" ); \
929 call_c2( dct_c.name, pbuf3, (void*)dct1 ); \
930 call_a2( dct_asm.name, pbuf4, (void*)dct2 ); \
932 ok = 1; used_asm = 0;
933 TEST_IDCT( add4x4_idct, dct4 );
934 TEST_IDCT( add8x8_idct, dct4 );
935 TEST_IDCT( add8x8_idct_dc, dct4 );
936 TEST_IDCT( add16x16_idct, dct4 );
937 TEST_IDCT( add16x16_idct_dc, dct4 );
938 report( "add_idct4 :" );
940 ok = 1; used_asm = 0;
941 TEST_IDCT( add8x8_idct8, dct8 );
942 TEST_IDCT( add16x16_idct8, dct8 );
943 report( "add_idct8 :" );
946 #define TEST_DCTDC( name )\
947 ok = 1; used_asm = 0;\
948 if( dct_asm.name != dct_ref.name )\
950 set_func_name( #name );\
952 uint16_t *p = (uint16_t*)buf1;\
953 for( int i = 0; i < 16 && ok; i++ )\
955 for( int j = 0; j < 16; j++ )\
956 dct1[0][j] = !i ? (j^j>>1^j>>2^j>>3)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
957 : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
958 : ((*p++)&0x1fff)-0x1000; /* general case */\
959 memcpy( dct2, dct1, 16 * sizeof(dctcoef) );\
960 call_c1( dct_c.name, dct1[0] );\
961 call_a1( dct_asm.name, dct2[0] );\
962 if( memcmp( dct1, dct2, 16 * sizeof(dctcoef) ) )\
965 call_c2( dct_c.name, dct1[0] );\
966 call_a2( dct_asm.name, dct2[0] );\
968 report( #name " :" );
970 TEST_DCTDC( dct4x4dc );
971 TEST_DCTDC( idct4x4dc );
974 #define TEST_DCTDC_CHROMA( name )\
975 ok = 1; used_asm = 0;\
976 if( dct_asm.name != dct_ref.name )\
978 set_func_name( #name );\
980 uint16_t *p = (uint16_t*)buf1;\
981 for( int i = 0; i < 16 && ok; i++ )\
983 for( int j = 0; j < 8; j++ )\
984 dct1[j][0] = !i ? (j^j>>1^j>>2)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
985 : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
986 : ((*p++)&0x1fff)-0x1000; /* general case */\
987 memcpy( dct2, dct1, 8*16 * sizeof(dctcoef) );\
988 call_c1( dct_c.name, dctdc[0], dct1 );\
989 call_a1( dct_asm.name, dctdc[1], dct2 );\
990 if( memcmp( dctdc[0], dctdc[1], 8 * sizeof(dctcoef) ) || memcmp( dct1, dct2, 8*16 * sizeof(dctcoef) ) )\
993 fprintf( stderr, #name " [FAILED]\n" ); \
996 call_c2( dct_c.name, dctdc[0], dct1 );\
997 call_a2( dct_asm.name, dctdc[1], dct2 );\
999 report( #name " :" );
1001 TEST_DCTDC_CHROMA( dct2x4dc );
1002 #undef TEST_DCTDC_CHROMA
1004 x264_zigzag_function_t zigzag_c[2];
1005 x264_zigzag_function_t zigzag_ref[2];
1006 x264_zigzag_function_t zigzag_asm[2];
1008 ALIGNED_16( dctcoef level1[64] );
1009 ALIGNED_16( dctcoef level2[64] );
1011 #define TEST_ZIGZAG_SCAN( name, t1, t2, dct, size ) \
1012 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1014 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1016 for( int i = 0; i < size*size; i++ ) \
1018 call_c( zigzag_c[interlace].name, t1, dct ); \
1019 call_a( zigzag_asm[interlace].name, t2, dct ); \
1020 if( memcmp( t1, t2, size*size*sizeof(dctcoef) ) ) \
1023 for( int i = 0; i < 2; i++ ) \
1025 dctcoef *d = (dctcoef*)(i ? t2 : t1); \
1026 for( int j = 0; j < size; j++ ) \
1028 for( int k = 0; k < size; k++ ) \
1029 fprintf( stderr, "%2d ", d[k+j*8] ); \
1030 fprintf( stderr, "\n" ); \
1032 fprintf( stderr, "\n" ); \
1034 fprintf( stderr, #name " [FAILED]\n" ); \
1038 #define TEST_ZIGZAG_SUB( name, t1, t2, size ) \
1039 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1042 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1044 memcpy( pbuf3, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
1045 memcpy( pbuf4, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
1046 nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
1047 nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
1048 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( pbuf3, pbuf4, 16*FDEC_STRIDE*sizeof(pixel) ) || nz_c != nz_a ) \
1051 fprintf( stderr, #name " [FAILED]\n" ); \
1053 call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
1054 call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
1057 #define TEST_ZIGZAG_SUBAC( name, t1, t2 ) \
1058 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1061 dctcoef dc_a, dc_c; \
1062 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1064 for( int i = 0; i < 2; i++ ) \
1066 memcpy( pbuf3, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
1067 memcpy( pbuf4, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
1068 for( int j = 0; j < 4; j++ ) \
1070 memcpy( pbuf3 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
1071 memcpy( pbuf4 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
1073 nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
1074 nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
1075 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 ) \
1078 fprintf( stderr, #name " [FAILED]\n" ); \
1082 call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
1083 call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
1086 #define TEST_INTERLEAVE( name, t1, t2, dct, size ) \
1087 if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1089 for( int j = 0; j < 100; j++ ) \
1091 set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1093 memcpy(dct, buf1, size*sizeof(dctcoef)); \
1094 for( int i = 0; i < size; i++ ) \
1095 dct[i] = rand()&0x1F ? 0 : dct[i]; \
1096 memcpy(buf3, buf4, 10); \
1097 call_c( zigzag_c[interlace].name, t1, dct, buf3 ); \
1098 call_a( zigzag_asm[interlace].name, t2, dct, buf4 ); \
1099 if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( buf3, buf4, 10 ) ) \
1101 ok = 0; printf("%d: %d %d %d %d\n%d %d %d %d\n\n",memcmp( t1, t2, size*sizeof(dctcoef) ),buf3[0], buf3[1], buf3[8], buf3[9], buf4[0], buf4[1], buf4[8], buf4[9]);break;\
1106 x264_zigzag_init( 0, &zigzag_c[0], &zigzag_c[1] );
1107 x264_zigzag_init( cpu_ref, &zigzag_ref[0], &zigzag_ref[1] );
1108 x264_zigzag_init( cpu_new, &zigzag_asm[0], &zigzag_asm[1] );
1110 ok = 1; used_asm = 0;
1111 TEST_INTERLEAVE( interleave_8x8_cavlc, level1, level2, dct8[0], 64 );
1112 report( "zigzag_interleave :" );
1114 for( interlace = 0; interlace <= 1; interlace++ )
1116 ok = 1; used_asm = 0;
1117 TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, dct8[0], 8 );
1118 TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 4 );
1119 TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
1120 TEST_ZIGZAG_SUB( sub_8x8, level1, level2, 64 );
1121 TEST_ZIGZAG_SUBAC( sub_4x4ac, level1, level2 );
1122 report( interlace ? "zigzag_field :" : "zigzag_frame :" );
1124 #undef TEST_ZIGZAG_SCAN
1125 #undef TEST_ZIGZAG_SUB
1130 static int check_mc( int cpu_ref, int cpu_new )
1132 x264_mc_functions_t mc_c;
1133 x264_mc_functions_t mc_ref;
1134 x264_mc_functions_t mc_a;
1135 x264_pixel_function_t pixf;
1137 pixel *src = &(pbuf1)[2*64+2];
1138 pixel *src2[4] = { &(pbuf1)[3*64+2], &(pbuf1)[5*64+2],
1139 &(pbuf1)[7*64+2], &(pbuf1)[9*64+2] };
1140 pixel *dst1 = pbuf3;
1141 pixel *dst2 = pbuf4;
1143 int ret = 0, ok, used_asm;
1145 x264_mc_init( 0, &mc_c, 0 );
1146 x264_mc_init( cpu_ref, &mc_ref, 0 );
1147 x264_mc_init( cpu_new, &mc_a, 0 );
1148 x264_pixel_init( 0, &pixf );
1150 #define MC_TEST_LUMA( w, h ) \
1151 if( mc_a.mc_luma != mc_ref.mc_luma && !(w&(w-1)) && h<=16 ) \
1153 const x264_weight_t *weight = x264_weight_none; \
1154 set_func_name( "mc_luma_%dx%d", w, h ); \
1156 for( int i = 0; i < 1024; i++ ) \
1157 pbuf3[i] = pbuf4[i] = 0xCD; \
1158 call_c( mc_c.mc_luma, dst1, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1159 call_a( mc_a.mc_luma, dst2, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1160 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1162 fprintf( stderr, "mc_luma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
1166 if( mc_a.get_ref != mc_ref.get_ref ) \
1168 pixel *ref = dst2; \
1169 intptr_t ref_stride = 32; \
1170 int w_checked = ( ( sizeof(pixel) == 2 && (w == 12 || w == 20)) ? w-2 : w ); \
1171 const x264_weight_t *weight = x264_weight_none; \
1172 set_func_name( "get_ref_%dx%d", w_checked, h ); \
1174 for( int i = 0; i < 1024; i++ ) \
1175 pbuf3[i] = pbuf4[i] = 0xCD; \
1176 call_c( mc_c.mc_luma, dst1, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1177 ref = (pixel*)call_a( mc_a.get_ref, ref, &ref_stride, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1178 for( int i = 0; i < h; i++ ) \
1179 if( memcmp( dst1+i*32, ref+i*ref_stride, w_checked * sizeof(pixel) ) ) \
1181 fprintf( stderr, "get_ref[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w_checked, h ); \
1187 #define MC_TEST_CHROMA( w, h ) \
1188 if( mc_a.mc_chroma != mc_ref.mc_chroma ) \
1190 set_func_name( "mc_chroma_%dx%d", w, h ); \
1192 for( int i = 0; i < 1024; i++ ) \
1193 pbuf3[i] = pbuf4[i] = 0xCD; \
1194 call_c( mc_c.mc_chroma, dst1, dst1+8, (intptr_t)16, src, (intptr_t)64, dx, dy, w, h ); \
1195 call_a( mc_a.mc_chroma, dst2, dst2+8, (intptr_t)16, src, (intptr_t)64, dx, dy, w, h ); \
1196 /* mc_chroma width=2 may write garbage to the right of dst. ignore that. */ \
1197 for( int j = 0; j < h; j++ ) \
1198 for( int i = w; i < 8; i++ ) \
1200 dst2[i+j*16+8] = dst1[i+j*16+8]; \
1201 dst2[i+j*16 ] = dst1[i+j*16 ]; \
1203 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1205 fprintf( stderr, "mc_chroma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
1209 ok = 1; used_asm = 0;
1210 for( int dy = -8; dy < 8; dy++ )
1211 for( int dx = -128; dx < 128; dx++ )
1213 if( rand()&15 ) continue; // running all of them is too slow
1214 MC_TEST_LUMA( 20, 18 );
1215 MC_TEST_LUMA( 16, 16 );
1216 MC_TEST_LUMA( 16, 8 );
1217 MC_TEST_LUMA( 12, 10 );
1218 MC_TEST_LUMA( 8, 16 );
1219 MC_TEST_LUMA( 8, 8 );
1220 MC_TEST_LUMA( 8, 4 );
1221 MC_TEST_LUMA( 4, 8 );
1222 MC_TEST_LUMA( 4, 4 );
1224 report( "mc luma :" );
1226 ok = 1; used_asm = 0;
1227 for( int dy = -1; dy < 9; dy++ )
1228 for( int dx = -128; dx < 128; dx++ )
1230 if( rand()&15 ) continue;
1231 MC_TEST_CHROMA( 8, 8 );
1232 MC_TEST_CHROMA( 8, 4 );
1233 MC_TEST_CHROMA( 4, 8 );
1234 MC_TEST_CHROMA( 4, 4 );
1235 MC_TEST_CHROMA( 4, 2 );
1236 MC_TEST_CHROMA( 2, 4 );
1237 MC_TEST_CHROMA( 2, 2 );
1239 report( "mc chroma :" );
1241 #undef MC_TEST_CHROMA
1243 #define MC_TEST_AVG( name, weight ) \
1245 for( int i = 0; i < 12; i++ ) \
1247 memcpy( pbuf3, pbuf1+320, 320 * sizeof(pixel) ); \
1248 memcpy( pbuf4, pbuf1+320, 320 * sizeof(pixel) ); \
1249 if( mc_a.name[i] != mc_ref.name[i] ) \
1251 set_func_name( "%s_%s", #name, pixel_names[i] ); \
1253 call_c1( mc_c.name[i], pbuf3, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1254 call_a1( mc_a.name[i], pbuf4, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1255 if( memcmp( pbuf3, pbuf4, 320 * sizeof(pixel) ) ) \
1258 fprintf( stderr, #name "[%d]: [FAILED]\n", i ); \
1260 call_c2( mc_c.name[i], pbuf3, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1261 call_a2( mc_a.name[i], pbuf4, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1266 ok = 1, used_asm = 0;
1267 for( int w = -63; w <= 127 && ok; w++ )
1268 MC_TEST_AVG( avg, w );
1269 report( "mc wpredb :" );
1271 #define MC_TEST_WEIGHT( name, weight, aligned ) \
1272 int align_off = (aligned ? 0 : rand()%16); \
1273 for( int i = 1; i <= 5; i++ ) \
1275 ALIGNED_16( pixel buffC[640] ); \
1276 ALIGNED_16( pixel buffA[640] ); \
1277 int j = X264_MAX( i*4, 2 ); \
1278 memset( buffC, 0, 640 * sizeof(pixel) ); \
1279 memset( buffA, 0, 640 * sizeof(pixel) ); \
1282 /* w12 is the same as w16 in some cases */ \
1283 if( i == 3 && mc_a.name[i] == mc_a.name[i+1] ) \
1285 if( mc_a.name[i] != mc_ref.name[i] ) \
1287 set_func_name( "%s_w%d", #name, j ); \
1289 call_c1( mc_c.weight[i], buffC, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1290 mc_a.weight_cache(&ha, &weight); \
1291 call_a1( weight.weightfn[i], buffA, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1292 for( int k = 0; k < 16; k++ ) \
1293 if( memcmp( &buffC[k*32], &buffA[k*32], j * sizeof(pixel) ) ) \
1296 fprintf( stderr, #name "[%d]: [FAILED] s:%d o:%d d%d\n", i, s, o, d ); \
1299 /* omit unlikely high scales for benchmarking */ \
1300 if( (s << (8-d)) < 512 ) \
1302 call_c2( mc_c.weight[i], buffC, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1303 call_a2( weight.weightfn[i], buffA, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1308 ok = 1; used_asm = 0;
1311 for( int s = 0; s <= 127 && ok; s++ )
1313 for( int o = -128; o <= 127 && ok; o++ )
1315 if( rand() & 2047 ) continue;
1316 for( int d = 0; d <= 7 && ok; d++ )
1320 x264_weight_t weight = { .i_scale = s, .i_denom = d, .i_offset = o };
1321 MC_TEST_WEIGHT( weight, weight, (align_cnt++ % 4) );
1326 report( "mc weight :" );
1328 ok = 1; used_asm = 0;
1329 for( int o = 0; o <= 127 && ok; o++ )
1332 if( rand() & 15 ) continue;
1333 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1334 MC_TEST_WEIGHT( offsetadd, weight, (align_cnt++ % 4) );
1336 report( "mc offsetadd :" );
1337 ok = 1; used_asm = 0;
1338 for( int o = -128; o < 0 && ok; o++ )
1341 if( rand() & 15 ) continue;
1342 x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1343 MC_TEST_WEIGHT( offsetsub, weight, (align_cnt++ % 4) );
1345 report( "mc offsetsub :" );
1347 ok = 1; used_asm = 0;
1348 for( int height = 8; height <= 16; height += 8 )
1350 if( mc_a.store_interleave_chroma != mc_ref.store_interleave_chroma )
1352 set_func_name( "store_interleave_chroma" );
1354 memset( pbuf3, 0, 64*height );
1355 memset( pbuf4, 0, 64*height );
1356 call_c( mc_c.store_interleave_chroma, pbuf3, (intptr_t)64, pbuf1, pbuf1+16, height );
1357 call_a( mc_a.store_interleave_chroma, pbuf4, (intptr_t)64, pbuf1, pbuf1+16, height );
1358 if( memcmp( pbuf3, pbuf4, 64*height ) )
1361 fprintf( stderr, "store_interleave_chroma FAILED: h=%d\n", height );
1365 if( mc_a.load_deinterleave_chroma_fenc != mc_ref.load_deinterleave_chroma_fenc )
1367 set_func_name( "load_deinterleave_chroma_fenc" );
1369 call_c( mc_c.load_deinterleave_chroma_fenc, pbuf3, pbuf1, (intptr_t)64, height );
1370 call_a( mc_a.load_deinterleave_chroma_fenc, pbuf4, pbuf1, (intptr_t)64, height );
1371 if( memcmp( pbuf3, pbuf4, FENC_STRIDE*height ) )
1374 fprintf( stderr, "load_deinterleave_chroma_fenc FAILED: h=%d\n", height );
1378 if( mc_a.load_deinterleave_chroma_fdec != mc_ref.load_deinterleave_chroma_fdec )
1380 set_func_name( "load_deinterleave_chroma_fdec" );
1382 call_c( mc_c.load_deinterleave_chroma_fdec, pbuf3, pbuf1, (intptr_t)64, height );
1383 call_a( mc_a.load_deinterleave_chroma_fdec, pbuf4, pbuf1, (intptr_t)64, height );
1384 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*height ) )
1387 fprintf( stderr, "load_deinterleave_chroma_fdec FAILED: h=%d\n", height );
1392 report( "store_interleave :" );
1395 int w, h, src_stride;
1396 } 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} };
1397 ok = 1; used_asm = 0;
1398 if( mc_a.plane_copy != mc_ref.plane_copy )
1400 set_func_name( "plane_copy" );
1402 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1404 int w = plane_specs[i].w;
1405 int h = plane_specs[i].h;
1406 intptr_t src_stride = plane_specs[i].src_stride;
1407 intptr_t dst_stride = (w + 127) & ~63;
1408 assert( dst_stride * h <= 0x1000 );
1409 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1410 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1411 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1412 call_c( mc_c.plane_copy, pbuf3, dst_stride, src1, src_stride, w, h );
1413 call_a( mc_a.plane_copy, pbuf4, dst_stride, src1, src_stride, w, h );
1414 for( int y = 0; y < h; y++ )
1415 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(pixel) ) )
1418 fprintf( stderr, "plane_copy FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1424 if( mc_a.plane_copy_swap != mc_ref.plane_copy_swap )
1426 set_func_name( "plane_copy_swap" );
1428 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1430 int w = (plane_specs[i].w + 1) >> 1;
1431 int h = plane_specs[i].h;
1432 intptr_t src_stride = plane_specs[i].src_stride;
1433 intptr_t dst_stride = (2*w + 127) & ~63;
1434 assert( dst_stride * h <= 0x1000 );
1435 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1436 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1437 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1438 call_c( mc_c.plane_copy_swap, pbuf3, dst_stride, src1, src_stride, w, h );
1439 call_a( mc_a.plane_copy_swap, pbuf4, dst_stride, src1, src_stride, w, h );
1440 for( int y = 0; y < h; y++ )
1441 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
1444 fprintf( stderr, "plane_copy_swap FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1450 if( mc_a.plane_copy_interleave != mc_ref.plane_copy_interleave )
1452 set_func_name( "plane_copy_interleave" );
1454 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1456 int w = (plane_specs[i].w + 1) >> 1;
1457 int h = plane_specs[i].h;
1458 intptr_t src_stride = (plane_specs[i].src_stride + 1) >> 1;
1459 intptr_t dst_stride = (2*w + 127) & ~63;
1460 assert( dst_stride * h <= 0x1000 );
1461 pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1462 memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1463 memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1464 call_c( mc_c.plane_copy_interleave, pbuf3, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1465 call_a( mc_a.plane_copy_interleave, pbuf4, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1466 for( int y = 0; y < h; y++ )
1467 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
1470 fprintf( stderr, "plane_copy_interleave FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1476 if( mc_a.plane_copy_deinterleave != mc_ref.plane_copy_deinterleave )
1478 set_func_name( "plane_copy_deinterleave" );
1480 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1482 int w = (plane_specs[i].w + 1) >> 1;
1483 int h = plane_specs[i].h;
1484 intptr_t dst_stride = w;
1485 intptr_t src_stride = (2*w + 127) & ~63;
1486 intptr_t offv = (dst_stride*h + 31) & ~15;
1487 memset( pbuf3, 0, 0x1000 );
1488 memset( pbuf4, 0, 0x1000 );
1489 call_c( mc_c.plane_copy_deinterleave, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf1, src_stride, w, h );
1490 call_a( mc_a.plane_copy_deinterleave, pbuf4, dst_stride, pbuf4+offv, dst_stride, pbuf1, src_stride, w, h );
1491 for( int y = 0; y < h; y++ )
1492 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w ) ||
1493 memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w ) )
1496 fprintf( stderr, "plane_copy_deinterleave FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1502 if( mc_a.plane_copy_deinterleave_rgb != mc_ref.plane_copy_deinterleave_rgb )
1504 set_func_name( "plane_copy_deinterleave_rgb" );
1506 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1508 int w = (plane_specs[i].w + 2) >> 2;
1509 int h = plane_specs[i].h;
1510 intptr_t src_stride = plane_specs[i].src_stride;
1511 intptr_t dst_stride = ALIGN( w, 16 );
1512 intptr_t offv = dst_stride*h + 16;
1514 for( int pw = 3; pw <= 4; pw++ )
1516 memset( pbuf3, 0, 0x1000 );
1517 memset( pbuf4, 0, 0x1000 );
1518 call_c( mc_c.plane_copy_deinterleave_rgb, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf3+2*offv, dst_stride, pbuf1, src_stride, pw, w, h );
1519 call_a( mc_a.plane_copy_deinterleave_rgb, pbuf4, dst_stride, pbuf4+offv, dst_stride, pbuf4+2*offv, dst_stride, pbuf1, src_stride, pw, w, h );
1520 for( int y = 0; y < h; y++ )
1521 if( memcmp( pbuf3+y*dst_stride+0*offv, pbuf4+y*dst_stride+0*offv, w ) ||
1522 memcmp( pbuf3+y*dst_stride+1*offv, pbuf4+y*dst_stride+1*offv, w ) ||
1523 memcmp( pbuf3+y*dst_stride+2*offv, pbuf4+y*dst_stride+2*offv, w ) )
1526 fprintf( stderr, "plane_copy_deinterleave_rgb FAILED: w=%d h=%d stride=%d pw=%d\n", w, h, (int)src_stride, pw );
1532 report( "plane_copy :" );
1534 if( mc_a.plane_copy_deinterleave_v210 != mc_ref.plane_copy_deinterleave_v210 )
1536 set_func_name( "plane_copy_deinterleave_v210" );
1537 ok = 1; used_asm = 1;
1538 for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1540 int w = (plane_specs[i].w + 1) >> 1;
1541 int h = plane_specs[i].h;
1542 intptr_t dst_stride = ALIGN( w, 16 );
1543 intptr_t src_stride = (w + 47) / 48 * 128 / sizeof(uint32_t);
1544 intptr_t offv = dst_stride*h + 32;
1545 memset( pbuf3, 0, 0x1000 );
1546 memset( pbuf4, 0, 0x1000 );
1547 call_c( mc_c.plane_copy_deinterleave_v210, pbuf3, dst_stride, pbuf3+offv, dst_stride, (uint32_t *)buf1, src_stride, w, h );
1548 call_a( mc_a.plane_copy_deinterleave_v210, pbuf4, dst_stride, pbuf4+offv, dst_stride, (uint32_t *)buf1, src_stride, w, h );
1549 for( int y = 0; y < h; y++ )
1550 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(uint16_t) ) ||
1551 memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w*sizeof(uint16_t) ) )
1554 fprintf( stderr, "plane_copy_deinterleave_v210 FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1561 if( mc_a.hpel_filter != mc_ref.hpel_filter )
1563 pixel *srchpel = pbuf1+8+2*64;
1564 pixel *dstc[3] = { pbuf3+8, pbuf3+8+16*64, pbuf3+8+32*64 };
1565 pixel *dsta[3] = { pbuf4+8, pbuf4+8+16*64, pbuf4+8+32*64 };
1566 void *tmp = pbuf3+49*64;
1567 set_func_name( "hpel_filter" );
1568 ok = 1; used_asm = 1;
1569 memset( pbuf3, 0, 4096 * sizeof(pixel) );
1570 memset( pbuf4, 0, 4096 * sizeof(pixel) );
1571 call_c( mc_c.hpel_filter, dstc[0], dstc[1], dstc[2], srchpel, (intptr_t)64, 48, 10, tmp );
1572 call_a( mc_a.hpel_filter, dsta[0], dsta[1], dsta[2], srchpel, (intptr_t)64, 48, 10, tmp );
1573 for( int i = 0; i < 3; i++ )
1574 for( int j = 0; j < 10; j++ )
1575 //FIXME ideally the first pixels would match too, but they aren't actually used
1576 if( memcmp( dstc[i]+j*64+2, dsta[i]+j*64+2, 43 * sizeof(pixel) ) )
1579 fprintf( stderr, "hpel filter differs at plane %c line %d\n", "hvc"[i], j );
1580 for( int k = 0; k < 48; k++ )
1581 printf( "%02x%s", dstc[i][j*64+k], (k+1)&3 ? "" : " " );
1583 for( int k = 0; k < 48; k++ )
1584 printf( "%02x%s", dsta[i][j*64+k], (k+1)&3 ? "" : " " );
1588 report( "hpel filter :" );
1591 if( mc_a.frame_init_lowres_core != mc_ref.frame_init_lowres_core )
1593 pixel *dstc[4] = { pbuf3, pbuf3+1024, pbuf3+2048, pbuf3+3072 };
1594 pixel *dsta[4] = { pbuf4, pbuf4+1024, pbuf4+2048, pbuf4+3072 };
1595 set_func_name( "lowres_init" );
1596 ok = 1; used_asm = 1;
1597 for( int w = 96; w <= 96+24; w += 8 )
1599 intptr_t stride = (w*2+31)&~31;
1600 intptr_t stride_lowres = (w+31)&~31;
1601 call_c( mc_c.frame_init_lowres_core, pbuf1, dstc[0], dstc[1], dstc[2], dstc[3], stride, stride_lowres, w, 8 );
1602 call_a( mc_a.frame_init_lowres_core, pbuf1, dsta[0], dsta[1], dsta[2], dsta[3], stride, stride_lowres, w, 8 );
1603 for( int i = 0; i < 8; i++ )
1605 for( int j = 0; j < 4; j++ )
1606 if( memcmp( dstc[j]+i*stride_lowres, dsta[j]+i*stride_lowres, w * sizeof(pixel) ) )
1609 fprintf( stderr, "frame_init_lowres differs at plane %d line %d\n", j, i );
1610 for( int k = 0; k < w; k++ )
1611 printf( "%d ", dstc[j][k+i*stride_lowres] );
1613 for( int k = 0; k < w; k++ )
1614 printf( "%d ", dsta[j][k+i*stride_lowres] );
1620 report( "lowres init :" );
1623 #define INTEGRAL_INIT( name, size, offset, cmp_len, ... )\
1624 if( mc_a.name != mc_ref.name )\
1626 intptr_t stride = 96;\
1627 set_func_name( #name );\
1629 memcpy( buf3, buf1, size*2*stride );\
1630 memcpy( buf4, buf1, size*2*stride );\
1631 uint16_t *sum = (uint16_t*)buf3;\
1632 call_c1( mc_c.name, sum+offset, __VA_ARGS__ );\
1633 sum = (uint16_t*)buf4;\
1634 call_a1( mc_a.name, sum+offset, __VA_ARGS__ );\
1635 if( memcmp( buf3+2*offset, buf4+2*offset, cmp_len*2 )\
1636 || (size>9 && memcmp( buf3+18*stride, buf4+18*stride, (stride-8)*2 )))\
1638 call_c2( mc_c.name, sum+offset, __VA_ARGS__ );\
1639 call_a2( mc_a.name, sum+offset, __VA_ARGS__ );\
1641 ok = 1; used_asm = 0;
1642 INTEGRAL_INIT( integral_init4h, 2, stride, stride-4, pbuf2, stride );
1643 INTEGRAL_INIT( integral_init8h, 2, stride, stride-8, pbuf2, stride );
1644 INTEGRAL_INIT( integral_init4v, 14, 0, stride-8, sum+9*stride, stride );
1645 INTEGRAL_INIT( integral_init8v, 9, 0, stride-8, stride );
1646 report( "integral init :" );
1648 ok = 1; used_asm = 0;
1649 if( mc_a.mbtree_propagate_cost != mc_ref.mbtree_propagate_cost )
1653 for( int i = 0; i < 10; i++ )
1655 float fps_factor = (rand()&65535) / 65535.0f;
1656 set_func_name( "mbtree_propagate_cost" );
1657 int16_t *dsta = (int16_t*)buf3;
1658 int16_t *dstc = dsta+400;
1659 uint16_t *prop = (uint16_t*)buf1;
1660 uint16_t *intra = (uint16_t*)buf4;
1661 uint16_t *inter = intra+128;
1662 uint16_t *qscale = inter+128;
1663 uint16_t *rnd = (uint16_t*)buf2;
1665 for( int j = 0; j < 100; j++ )
1667 intra[j] = *rnd++ & 0x7fff;
1668 intra[j] += !intra[j];
1669 inter[j] = *rnd++ & 0x7fff;
1670 qscale[j] = *rnd++ & 0x7fff;
1672 call_c( mc_c.mbtree_propagate_cost, dstc, prop, intra, inter, qscale, &fps_factor, 100 );
1673 call_a( mc_a.mbtree_propagate_cost, dsta, prop, intra, inter, qscale, &fps_factor, 100 );
1674 // I don't care about exact rounding, this is just how close the floating-point implementation happens to be
1676 for( int j = 0; j < 100 && ok; j++ )
1678 ok &= abs( dstc[j]-dsta[j] ) <= 1 || fabs( (double)dstc[j]/dsta[j]-1 ) < 1e-4;
1680 fprintf( stderr, "mbtree_propagate_cost FAILED: %f !~= %f\n", (double)dstc[j], (double)dsta[j] );
1685 if( mc_a.mbtree_propagate_list != mc_ref.mbtree_propagate_list )
1688 for( int i = 0; i < 8; i++ )
1690 set_func_name( "mbtree_propagate_list" );
1694 int size = width*height;
1695 h.mb.i_mb_stride = width;
1696 h.mb.i_mb_width = width;
1697 h.mb.i_mb_height = height;
1699 uint16_t *ref_costsc = (uint16_t*)buf3;
1700 uint16_t *ref_costsa = (uint16_t*)buf4;
1701 int16_t (*mvs)[2] = (int16_t(*)[2])(ref_costsc + size);
1702 int16_t *propagate_amount = (int16_t*)(mvs + width);
1703 uint16_t *lowres_costs = (uint16_t*)(propagate_amount + width);
1704 h.scratch_buffer2 = (uint8_t*)(ref_costsa + size);
1705 int bipred_weight = (rand()%63)+1;
1707 for( int j = 0; j < size; j++ )
1708 ref_costsc[j] = ref_costsa[j] = rand()&32767;
1709 for( int j = 0; j < width; j++ )
1711 static const uint8_t list_dist[2][8] = {{0,1,1,1,1,1,1,1},{1,1,3,3,3,3,3,2}};
1712 for( int k = 0; k < 2; k++ )
1713 mvs[j][k] = (rand()&127) - 64;
1714 propagate_amount[j] = rand()&32767;
1715 lowres_costs[j] = list_dist[list][rand()&7] << LOWRES_COST_SHIFT;
1718 call_c1( mc_c.mbtree_propagate_list, &h, ref_costsc, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
1719 call_a1( mc_a.mbtree_propagate_list, &h, ref_costsa, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
1721 for( int j = 0; j < size && ok; j++ )
1723 ok &= abs(ref_costsa[j] - ref_costsc[j]) <= 1;
1725 fprintf( stderr, "mbtree_propagate_list FAILED at %d: %d !~= %d\n", j, ref_costsc[j], ref_costsa[j] );
1728 call_c2( mc_c.mbtree_propagate_list, &h, ref_costsc, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
1729 call_a2( mc_a.mbtree_propagate_list, &h, ref_costsa, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
1732 report( "mbtree :" );
1734 if( mc_a.memcpy_aligned != mc_ref.memcpy_aligned )
1736 set_func_name( "memcpy_aligned" );
1737 ok = 1; used_asm = 1;
1738 for( size_t size = 16; size < 256; size += 16 )
1740 memset( buf4, 0xAA, size + 1 );
1741 call_c( mc_c.memcpy_aligned, buf3, buf1, size );
1742 call_a( mc_a.memcpy_aligned, buf4, buf1, size );
1743 if( memcmp( buf3, buf4, size ) || buf4[size] != 0xAA )
1746 fprintf( stderr, "memcpy_aligned FAILED: size=%d\n", (int)size );
1750 report( "memcpy aligned :" );
1753 if( mc_a.memzero_aligned != mc_ref.memzero_aligned )
1755 set_func_name( "memzero_aligned" );
1756 ok = 1; used_asm = 1;
1757 for( size_t size = 128; size < 1024; size += 128 )
1759 memset( buf4, 0xAA, size + 1 );
1760 call_c( mc_c.memzero_aligned, buf3, size );
1761 call_a( mc_a.memzero_aligned, buf4, size );
1762 if( memcmp( buf3, buf4, size ) || buf4[size] != 0xAA )
1765 fprintf( stderr, "memzero_aligned FAILED: size=%d\n", (int)size );
1769 report( "memzero aligned :" );
1775 static int check_deblock( int cpu_ref, int cpu_new )
1777 x264_deblock_function_t db_c;
1778 x264_deblock_function_t db_ref;
1779 x264_deblock_function_t db_a;
1780 int ret = 0, ok = 1, used_asm = 0;
1781 int alphas[36], betas[36];
1784 x264_deblock_init( 0, &db_c, 0 );
1785 x264_deblock_init( cpu_ref, &db_ref, 0 );
1786 x264_deblock_init( cpu_new, &db_a, 0 );
1788 /* not exactly the real values of a,b,tc but close enough */
1789 for( int i = 35, a = 255, c = 250; i >= 0; i-- )
1791 alphas[i] = a << (BIT_DEPTH-8);
1792 betas[i] = (i+1)/2 << (BIT_DEPTH-8);
1793 tcs[i][0] = tcs[i][3] = (c+6)/10 << (BIT_DEPTH-8);
1794 tcs[i][1] = (c+7)/15 << (BIT_DEPTH-8);
1795 tcs[i][2] = (c+9)/20 << (BIT_DEPTH-8);
1800 #define TEST_DEBLOCK( name, align, ... ) \
1801 for( int i = 0; i < 36; i++ ) \
1803 intptr_t off = 8*32 + (i&15)*4*!align; /* benchmark various alignments of h filter */ \
1804 for( int j = 0; j < 1024; j++ ) \
1805 /* two distributions of random to excersize different failure modes */ \
1806 pbuf3[j] = rand() & (i&1 ? 0xf : PIXEL_MAX ); \
1807 memcpy( pbuf4, pbuf3, 1024 * sizeof(pixel) ); \
1808 if( db_a.name != db_ref.name ) \
1810 set_func_name( #name ); \
1812 call_c1( db_c.name, pbuf3+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1813 call_a1( db_a.name, pbuf4+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1814 if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1817 fprintf( stderr, #name "(a=%d, b=%d): [FAILED]\n", alphas[i], betas[i] ); \
1820 call_c2( db_c.name, pbuf3+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1821 call_a2( db_a.name, pbuf4+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1825 TEST_DEBLOCK( deblock_luma[0], 0, tcs[i] );
1826 TEST_DEBLOCK( deblock_luma[1], 1, tcs[i] );
1827 TEST_DEBLOCK( deblock_h_chroma_420, 0, tcs[i] );
1828 TEST_DEBLOCK( deblock_h_chroma_422, 0, tcs[i] );
1829 TEST_DEBLOCK( deblock_chroma_420_mbaff, 0, tcs[i] );
1830 TEST_DEBLOCK( deblock_chroma_422_mbaff, 0, tcs[i] );
1831 TEST_DEBLOCK( deblock_chroma[1], 1, tcs[i] );
1832 TEST_DEBLOCK( deblock_luma_intra[0], 0 );
1833 TEST_DEBLOCK( deblock_luma_intra[1], 1 );
1834 TEST_DEBLOCK( deblock_h_chroma_420_intra, 0 );
1835 TEST_DEBLOCK( deblock_h_chroma_422_intra, 0 );
1836 TEST_DEBLOCK( deblock_chroma_420_intra_mbaff, 0 );
1837 TEST_DEBLOCK( deblock_chroma_422_intra_mbaff, 0 );
1838 TEST_DEBLOCK( deblock_chroma_intra[1], 1 );
1840 if( db_a.deblock_strength != db_ref.deblock_strength )
1842 for( int i = 0; i < 100; i++ )
1844 ALIGNED_ARRAY_16( uint8_t, nnz, [X264_SCAN8_SIZE] );
1845 ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );
1846 ALIGNED_ARRAY_16( int16_t, mv, [2],[X264_SCAN8_LUMA_SIZE][2] );
1847 ALIGNED_ARRAY_N( uint8_t, bs, [2],[2][8][4] );
1848 memset( bs, 99, sizeof(uint8_t)*2*4*8*2 );
1849 for( int j = 0; j < X264_SCAN8_SIZE; j++ )
1850 nnz[j] = ((rand()&7) == 7) * rand() & 0xf;
1851 for( int j = 0; j < 2; j++ )
1852 for( int k = 0; k < X264_SCAN8_LUMA_SIZE; k++ )
1854 ref[j][k] = ((rand()&3) != 3) ? 0 : (rand() & 31) - 2;
1855 for( int l = 0; l < 2; l++ )
1856 mv[j][k][l] = ((rand()&7) != 7) ? (rand()&7) - 3 : (rand()&1023) - 512;
1858 set_func_name( "deblock_strength" );
1859 call_c( db_c.deblock_strength, nnz, ref, mv, bs[0], 2<<(i&1), ((i>>1)&1) );
1860 call_a( db_a.deblock_strength, nnz, ref, mv, bs[1], 2<<(i&1), ((i>>1)&1) );
1861 if( memcmp( bs[0], bs[1], sizeof(uint8_t)*2*4*8 ) )
1864 fprintf( stderr, "deblock_strength: [FAILED]\n" );
1865 for( int j = 0; j < 2; j++ )
1867 for( int k = 0; k < 2; k++ )
1868 for( int l = 0; l < 4; l++ )
1870 for( int m = 0; m < 4; m++ )
1871 printf("%d ",bs[j][k][l][m]);
1881 report( "deblock :" );
1886 static int check_quant( int cpu_ref, int cpu_new )
1888 x264_quant_function_t qf_c;
1889 x264_quant_function_t qf_ref;
1890 x264_quant_function_t qf_a;
1891 ALIGNED_ARRAY_N( dctcoef, dct1,[64] );
1892 ALIGNED_ARRAY_N( dctcoef, dct2,[64] );
1893 ALIGNED_ARRAY_N( dctcoef, dct3,[8],[16] );
1894 ALIGNED_ARRAY_N( dctcoef, dct4,[8],[16] );
1895 ALIGNED_ARRAY_N( uint8_t, cqm_buf,[64] );
1896 int ret = 0, ok, used_asm;
1897 int oks[3] = {1,1,1}, used_asms[3] = {0,0,0};
1900 memset( h, 0, sizeof(*h) );
1901 h->sps->i_chroma_format_idc = 1;
1902 x264_param_default( &h->param );
1903 h->chroma_qp_table = i_chroma_qp_table + 12;
1904 h->param.analyse.b_transform_8x8 = 1;
1906 for( int i_cqm = 0; i_cqm < 4; i_cqm++ )
1910 for( int i = 0; i < 6; i++ )
1911 h->pps->scaling_list[i] = x264_cqm_flat16;
1912 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_FLAT;
1914 else if( i_cqm == 1 )
1916 for( int i = 0; i < 6; i++ )
1917 h->pps->scaling_list[i] = x264_cqm_jvt[i];
1918 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_JVT;
1922 int max_scale = BIT_DEPTH < 10 ? 255 : 228;
1924 for( int i = 0; i < 64; i++ )
1925 cqm_buf[i] = 10 + rand() % (max_scale - 9);
1927 for( int i = 0; i < 64; i++ )
1929 for( int i = 0; i < 6; i++ )
1930 h->pps->scaling_list[i] = cqm_buf;
1931 h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_CUSTOM;
1934 h->param.rc.i_qp_min = 0;
1935 h->param.rc.i_qp_max = QP_MAX_SPEC;
1937 x264_quant_init( h, 0, &qf_c );
1938 x264_quant_init( h, cpu_ref, &qf_ref );
1939 x264_quant_init( h, cpu_new, &qf_a );
1941 #define INIT_QUANT8(j,max) \
1943 static const int scale1d[8] = {32,31,24,31,32,31,24,31}; \
1944 for( int i = 0; i < max; i++ ) \
1946 unsigned int scale = (255*scale1d[(i>>3)&7]*scale1d[i&7])/16; \
1947 dct1[i] = dct2[i] = (j>>(i>>6))&1 ? (rand()%(2*scale+1))-scale : 0; \
1951 #define INIT_QUANT4(j,max) \
1953 static const int scale1d[4] = {4,6,4,6}; \
1954 for( int i = 0; i < max; i++ ) \
1956 unsigned int scale = 255*scale1d[(i>>2)&3]*scale1d[i&3]; \
1957 dct1[i] = dct2[i] = (j>>(i>>4))&1 ? (rand()%(2*scale+1))-scale : 0; \
1961 #define TEST_QUANT_DC( name, cqm ) \
1962 if( qf_a.name != qf_ref.name ) \
1964 set_func_name( #name ); \
1966 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1968 for( int j = 0; j < 2; j++ ) \
1970 int result_c, result_a; \
1971 for( int i = 0; i < 16; i++ ) \
1972 dct1[i] = dct2[i] = j ? (rand() & 0x1fff) - 0xfff : 0; \
1973 result_c = call_c1( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1974 result_a = call_a1( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1975 if( memcmp( dct1, dct2, 16*sizeof(dctcoef) ) || result_c != result_a ) \
1978 fprintf( stderr, #name "(cqm=%d): [FAILED]\n", i_cqm ); \
1981 call_c2( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1982 call_a2( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1987 #define TEST_QUANT( qname, block, type, w, maxj ) \
1988 if( qf_a.qname != qf_ref.qname ) \
1990 set_func_name( #qname ); \
1992 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1994 for( int j = 0; j < maxj; j++ ) \
1996 INIT_QUANT##type(j, w*w) \
1997 int result_c = call_c1( qf_c.qname, (void*)dct1, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
1998 int result_a = call_a1( qf_a.qname, (void*)dct2, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
1999 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) || result_c != result_a ) \
2002 fprintf( stderr, #qname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
2005 call_c2( qf_c.qname, (void*)dct1, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
2006 call_a2( qf_a.qname, (void*)dct2, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
2011 TEST_QUANT( quant_8x8, CQM_8IY, 8, 8, 2 );
2012 TEST_QUANT( quant_8x8, CQM_8PY, 8, 8, 2 );
2013 TEST_QUANT( quant_4x4, CQM_4IY, 4, 4, 2 );
2014 TEST_QUANT( quant_4x4, CQM_4PY, 4, 4, 2 );
2015 TEST_QUANT( quant_4x4x4, CQM_4IY, 4, 8, 16 );
2016 TEST_QUANT( quant_4x4x4, CQM_4PY, 4, 8, 16 );
2017 TEST_QUANT_DC( quant_4x4_dc, **h->quant4_mf[CQM_4IY] );
2018 TEST_QUANT_DC( quant_2x2_dc, **h->quant4_mf[CQM_4IC] );
2020 #define TEST_DEQUANT( qname, dqname, block, w ) \
2021 if( qf_a.dqname != qf_ref.dqname ) \
2023 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
2025 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2027 INIT_QUANT##w(1, w*w) \
2028 qf_c.qname( dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
2029 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
2030 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2031 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2032 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
2035 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
2038 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2039 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2043 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8IY, 8 );
2044 TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8PY, 8 );
2045 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4IY, 4 );
2046 TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4PY, 4 );
2048 #define TEST_DEQUANT_DC( qname, dqname, block, w ) \
2049 if( qf_a.dqname != qf_ref.dqname ) \
2051 set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
2053 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2055 for( int i = 0; i < 16; i++ ) \
2056 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16; \
2057 qf_c.qname( dct1, h->quant##w##_mf[block][qp][0]>>1, h->quant##w##_bias[block][qp][0]>>1 ); \
2058 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
2059 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2060 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2061 if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
2064 fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
2066 call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2067 call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2071 TEST_DEQUANT_DC( quant_4x4_dc, dequant_4x4_dc, CQM_4IY, 4 );
2073 if( qf_a.idct_dequant_2x4_dc != qf_ref.idct_dequant_2x4_dc )
2075 set_func_name( "idct_dequant_2x4_dc_%s", i_cqm?"cqm":"flat" );
2077 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- )
2079 for( int i = 0; i < 8; i++ )
2080 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16;
2081 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 );
2082 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 );
2083 call_c( qf_c.idct_dequant_2x4_dc, dct1, dct3, h->dequant4_mf[CQM_4IC], qp+3 );
2084 call_a( qf_a.idct_dequant_2x4_dc, dct1, dct4, h->dequant4_mf[CQM_4IC], qp+3 );
2085 for( int i = 0; i < 8; i++ )
2086 if( dct3[i][0] != dct4[i][0] )
2089 fprintf( stderr, "idct_dequant_2x4_dc (qp=%d, cqm=%d): [FAILED]\n", qp, i_cqm );
2095 if( qf_a.idct_dequant_2x4_dconly != qf_ref.idct_dequant_2x4_dconly )
2097 set_func_name( "idct_dequant_2x4_dc_%s", i_cqm?"cqm":"flat" );
2099 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- )
2101 for( int i = 0; i < 8; i++ )
2102 dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16;
2103 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 );
2104 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 );
2105 memcpy( dct2, dct1, 8*sizeof(dctcoef) );
2106 call_c1( qf_c.idct_dequant_2x4_dconly, dct1, h->dequant4_mf[CQM_4IC], qp+3 );
2107 call_a1( qf_a.idct_dequant_2x4_dconly, dct2, h->dequant4_mf[CQM_4IC], qp+3 );
2108 if( memcmp( dct1, dct2, 8*sizeof(dctcoef) ) )
2111 fprintf( stderr, "idct_dequant_2x4_dconly (qp=%d, cqm=%d): [FAILED]\n", qp, i_cqm );
2114 call_c2( qf_c.idct_dequant_2x4_dconly, dct1, h->dequant4_mf[CQM_4IC], qp+3 );
2115 call_a2( qf_a.idct_dequant_2x4_dconly, dct2, h->dequant4_mf[CQM_4IC], qp+3 );
2119 #define TEST_OPTIMIZE_CHROMA_DC( optname, size ) \
2120 if( qf_a.optname != qf_ref.optname ) \
2122 set_func_name( #optname ); \
2124 for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2126 int qpdc = qp + (size == 8 ? 3 : 0); \
2127 int dmf = h->dequant4_mf[CQM_4IC][qpdc%6][0] << qpdc/6; \
2130 for( int i = 16; ; i <<= 1 ) \
2132 int res_c, res_asm; \
2133 int max = X264_MIN( i, PIXEL_MAX*16 ); \
2134 for( int j = 0; j < size; j++ ) \
2135 dct1[j] = rand()%(max*2+1) - max; \
2136 for( int j = 0; i <= size; j += 4 ) \
2137 qf_c.quant_2x2_dc( &dct1[j], h->quant4_mf[CQM_4IC][qpdc][0]>>1, h->quant4_bias[CQM_4IC][qpdc][0]>>1 ); \
2138 memcpy( dct2, dct1, size*sizeof(dctcoef) ); \
2139 res_c = call_c1( qf_c.optname, dct1, dmf ); \
2140 res_asm = call_a1( qf_a.optname, dct2, dmf ); \
2141 if( res_c != res_asm || memcmp( dct1, dct2, size*sizeof(dctcoef) ) ) \
2144 fprintf( stderr, #optname "(qp=%d, res_c=%d, res_asm=%d): [FAILED]\n", qp, res_c, res_asm ); \
2146 call_c2( qf_c.optname, dct1, dmf ); \
2147 call_a2( qf_a.optname, dct2, dmf ); \
2148 if( i >= PIXEL_MAX*16 ) \
2154 TEST_OPTIMIZE_CHROMA_DC( optimize_chroma_2x2_dc, 4 );
2155 TEST_OPTIMIZE_CHROMA_DC( optimize_chroma_2x4_dc, 8 );
2157 x264_cqm_delete( h );
2160 ok = oks[0]; used_asm = used_asms[0];
2161 report( "quant :" );
2163 ok = oks[1]; used_asm = used_asms[1];
2164 report( "dequant :" );
2166 ok = oks[2]; used_asm = used_asms[2];
2167 report( "optimize chroma dc :" );
2169 ok = 1; used_asm = 0;
2170 if( qf_a.denoise_dct != qf_ref.denoise_dct )
2173 for( int size = 16; size <= 64; size += 48 )
2175 set_func_name( "denoise_dct" );
2176 memcpy( dct1, buf1, size*sizeof(dctcoef) );
2177 memcpy( dct2, buf1, size*sizeof(dctcoef) );
2178 memcpy( buf3+256, buf3, 256 );
2179 call_c1( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
2180 call_a1( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
2181 if( memcmp( dct1, dct2, size*sizeof(dctcoef) ) || memcmp( buf3+4, buf3+256+4, (size-1)*sizeof(uint32_t) ) )
2183 call_c2( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
2184 call_a2( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
2187 report( "denoise dct :" );
2189 #define TEST_DECIMATE( decname, w, ac, thresh ) \
2190 if( qf_a.decname != qf_ref.decname ) \
2192 set_func_name( #decname ); \
2194 for( int i = 0; i < 100; i++ ) \
2196 static const int distrib[16] = {1,1,1,1,1,1,1,1,1,1,1,1,2,3,4};\
2197 static const int zerorate_lut[4] = {3,7,15,31};\
2198 int zero_rate = zerorate_lut[i&3];\
2199 for( int idx = 0; idx < w*w; idx++ ) \
2201 int sign = (rand()&1) ? -1 : 1; \
2202 int abs_level = distrib[rand()&15]; \
2203 if( abs_level == 4 ) abs_level = rand()&0x3fff; \
2204 int zero = !(rand()&zero_rate); \
2205 dct1[idx] = zero * abs_level * sign; \
2209 int result_c = call_c( qf_c.decname, dct1 ); \
2210 int result_a = call_a( qf_a.decname, dct1 ); \
2211 if( X264_MIN(result_c,thresh) != X264_MIN(result_a,thresh) ) \
2214 fprintf( stderr, #decname ": [FAILED]\n" ); \
2220 ok = 1; used_asm = 0;
2221 TEST_DECIMATE( decimate_score64, 8, 0, 6 );
2222 TEST_DECIMATE( decimate_score16, 4, 0, 6 );
2223 TEST_DECIMATE( decimate_score15, 4, 1, 7 );
2224 report( "decimate_score :" );
2226 #define TEST_LAST( last, lastname, size, ac ) \
2227 if( qf_a.last != qf_ref.last ) \
2229 set_func_name( #lastname ); \
2231 for( int i = 0; i < 100; i++ ) \
2234 int max = rand() & (size-1); \
2235 memset( dct1, 0, size*sizeof(dctcoef) ); \
2236 for( int idx = ac; idx < max; idx++ ) \
2237 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
2240 int result_c = call_c( qf_c.last, dct1+ac ); \
2241 int result_a = call_a( qf_a.last, dct1+ac ); \
2242 if( result_c != result_a ) \
2245 fprintf( stderr, #lastname ": [FAILED]\n" ); \
2251 ok = 1; used_asm = 0;
2252 TEST_LAST( coeff_last4 , coeff_last4, 4, 0 );
2253 TEST_LAST( coeff_last8 , coeff_last8, 8, 0 );
2254 TEST_LAST( coeff_last[ DCT_LUMA_AC], coeff_last15, 16, 1 );
2255 TEST_LAST( coeff_last[ DCT_LUMA_4x4], coeff_last16, 16, 0 );
2256 TEST_LAST( coeff_last[ DCT_LUMA_8x8], coeff_last64, 64, 0 );
2257 report( "coeff_last :" );
2259 #define TEST_LEVELRUN( lastname, name, size, ac ) \
2260 if( qf_a.lastname != qf_ref.lastname ) \
2262 set_func_name( #name ); \
2264 for( int i = 0; i < 100; i++ ) \
2266 x264_run_level_t runlevel_c, runlevel_a; \
2268 int max = rand() & (size-1); \
2269 memset( dct1, 0, size*sizeof(dctcoef) ); \
2270 memcpy( &runlevel_a, buf1+i, sizeof(x264_run_level_t) ); \
2271 memcpy( &runlevel_c, buf1+i, sizeof(x264_run_level_t) ); \
2272 for( int idx = ac; idx < max; idx++ ) \
2273 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
2276 int result_c = call_c( qf_c.lastname, dct1+ac, &runlevel_c ); \
2277 int result_a = call_a( qf_a.lastname, dct1+ac, &runlevel_a ); \
2278 if( result_c != result_a || runlevel_c.last != runlevel_a.last || \
2279 runlevel_c.mask != runlevel_a.mask || \
2280 memcmp(runlevel_c.level, runlevel_a.level, sizeof(dctcoef)*result_c)) \
2283 fprintf( stderr, #name ": [FAILED]\n" ); \
2289 ok = 1; used_asm = 0;
2290 TEST_LEVELRUN( coeff_level_run4 , coeff_level_run4, 4, 0 );
2291 TEST_LEVELRUN( coeff_level_run8 , coeff_level_run8, 8, 0 );
2292 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_AC], coeff_level_run15, 16, 1 );
2293 TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_4x4], coeff_level_run16, 16, 0 );
2294 report( "coeff_level_run :" );
2299 static int check_intra( int cpu_ref, int cpu_new )
2301 int ret = 0, ok = 1, used_asm = 0;
2302 ALIGNED_ARRAY_32( pixel, edge,[36] );
2303 ALIGNED_ARRAY_32( pixel, edge2,[36] );
2304 ALIGNED_ARRAY_32( pixel, fdec,[FDEC_STRIDE*20] );
2307 x264_predict_t predict_16x16[4+3];
2308 x264_predict_t predict_8x8c[4+3];
2309 x264_predict_t predict_8x16c[4+3];
2310 x264_predict8x8_t predict_8x8[9+3];
2311 x264_predict_t predict_4x4[9+3];
2312 x264_predict_8x8_filter_t predict_8x8_filter;
2313 } ip_c, ip_ref, ip_a;
2315 x264_predict_16x16_init( 0, ip_c.predict_16x16 );
2316 x264_predict_8x8c_init( 0, ip_c.predict_8x8c );
2317 x264_predict_8x16c_init( 0, ip_c.predict_8x16c );
2318 x264_predict_8x8_init( 0, ip_c.predict_8x8, &ip_c.predict_8x8_filter );
2319 x264_predict_4x4_init( 0, ip_c.predict_4x4 );
2321 x264_predict_16x16_init( cpu_ref, ip_ref.predict_16x16 );
2322 x264_predict_8x8c_init( cpu_ref, ip_ref.predict_8x8c );
2323 x264_predict_8x16c_init( cpu_ref, ip_ref.predict_8x16c );
2324 x264_predict_8x8_init( cpu_ref, ip_ref.predict_8x8, &ip_ref.predict_8x8_filter );
2325 x264_predict_4x4_init( cpu_ref, ip_ref.predict_4x4 );
2327 x264_predict_16x16_init( cpu_new, ip_a.predict_16x16 );
2328 x264_predict_8x8c_init( cpu_new, ip_a.predict_8x8c );
2329 x264_predict_8x16c_init( cpu_new, ip_a.predict_8x16c );
2330 x264_predict_8x8_init( cpu_new, ip_a.predict_8x8, &ip_a.predict_8x8_filter );
2331 x264_predict_4x4_init( cpu_new, ip_a.predict_4x4 );
2333 memcpy( fdec, pbuf1, 32*20 * sizeof(pixel) );\
2335 ip_c.predict_8x8_filter( fdec+48, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
2337 #define INTRA_TEST( name, dir, w, h, align, bench, ... )\
2338 if( ip_a.name[dir] != ip_ref.name[dir] )\
2340 set_func_name( "intra_%s_%s", #name, intra_##name##_names[dir] );\
2342 memcpy( pbuf3, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
2343 memcpy( pbuf4, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
2344 for( int a = 0; a < (do_bench ? 64/sizeof(pixel) : 1); a += align )\
2346 call_c##bench( ip_c.name[dir], pbuf3+48+a, ##__VA_ARGS__ );\
2347 call_a##bench( ip_a.name[dir], pbuf4+48+a, ##__VA_ARGS__ );\
2348 if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*20 * sizeof(pixel) ) )\
2350 fprintf( stderr, #name "[%d] : [FAILED]\n", dir );\
2352 if( ip_c.name == (void *)ip_c.predict_8x8 )\
2354 for( int k = -1; k < 16; k++ )\
2355 printf( "%2x ", edge[16+k] );\
2358 for( int j = 0; j < h; j++ )\
2360 if( ip_c.name == (void *)ip_c.predict_8x8 )\
2361 printf( "%2x ", edge[14-j] );\
2362 for( int k = 0; k < w; k++ )\
2363 printf( "%2x ", pbuf4[48+k+j*FDEC_STRIDE] );\
2367 for( int j = 0; j < h; j++ )\
2369 if( ip_c.name == (void *)ip_c.predict_8x8 )\
2371 for( int k = 0; k < w; k++ )\
2372 printf( "%2x ", pbuf3[48+k+j*FDEC_STRIDE] );\
2380 for( int i = 0; i < 12; i++ )
2381 INTRA_TEST( predict_4x4, i, 4, 4, 4, );
2382 for( int i = 0; i < 7; i++ )
2383 INTRA_TEST( predict_8x8c, i, 8, 8, 16, );
2384 for( int i = 0; i < 7; i++ )
2385 INTRA_TEST( predict_8x16c, i, 8, 16, 16, );
2386 for( int i = 0; i < 7; i++ )
2387 INTRA_TEST( predict_16x16, i, 16, 16, 16, );
2388 for( int i = 0; i < 12; i++ )
2389 INTRA_TEST( predict_8x8, i, 8, 8, 8, , edge );
2391 set_func_name("intra_predict_8x8_filter");
2392 if( ip_a.predict_8x8_filter != ip_ref.predict_8x8_filter )
2395 for( int i = 0; i < 32; i++ )
2397 if( !(i&7) || ((i&MB_TOPRIGHT) && !(i&MB_TOP)) )
2399 int neighbor = (i&24)>>1;
2400 memset( edge, 0, 36*sizeof(pixel) );
2401 memset( edge2, 0, 36*sizeof(pixel) );
2402 call_c( ip_c.predict_8x8_filter, pbuf1+48, edge, neighbor, i&7 );
2403 call_a( ip_a.predict_8x8_filter, pbuf1+48, edge2, neighbor, i&7 );
2404 if( !(neighbor&MB_TOPLEFT) )
2405 edge[15] = edge2[15] = 0;
2406 if( memcmp( edge+7, edge2+7, (i&MB_TOPRIGHT ? 26 : i&MB_TOP ? 17 : 8) * sizeof(pixel) ) )
2408 fprintf( stderr, "predict_8x8_filter : [FAILED] %d %d\n", (i&24)>>1, i&7);
2414 #define EXTREMAL_PLANE( w, h ) \
2417 for( int j = 0; j < 7; j++ ) \
2418 max[j] = test ? rand()&PIXEL_MAX : PIXEL_MAX; \
2419 fdec[48-1-FDEC_STRIDE] = (i&1)*max[0]; \
2420 for( int j = 0; j < w/2; j++ ) \
2421 fdec[48+j-FDEC_STRIDE] = (!!(i&2))*max[1]; \
2422 for( int j = w/2; j < w-1; j++ ) \
2423 fdec[48+j-FDEC_STRIDE] = (!!(i&4))*max[2]; \
2424 fdec[48+(w-1)-FDEC_STRIDE] = (!!(i&8))*max[3]; \
2425 for( int j = 0; j < h/2; j++ ) \
2426 fdec[48+j*FDEC_STRIDE-1] = (!!(i&16))*max[4]; \
2427 for( int j = h/2; j < h-1; j++ ) \
2428 fdec[48+j*FDEC_STRIDE-1] = (!!(i&32))*max[5]; \
2429 fdec[48+(h-1)*FDEC_STRIDE-1] = (!!(i&64))*max[6]; \
2431 /* Extremal test case for planar prediction. */
2432 for( int test = 0; test < 100 && ok; test++ )
2433 for( int i = 0; i < 128 && ok; i++ )
2435 EXTREMAL_PLANE( 8, 8 );
2436 INTRA_TEST( predict_8x8c, I_PRED_CHROMA_P, 8, 8, 64, 1 );
2437 EXTREMAL_PLANE( 8, 16 );
2438 INTRA_TEST( predict_8x16c, I_PRED_CHROMA_P, 8, 16, 64, 1 );
2439 EXTREMAL_PLANE( 16, 16 );
2440 INTRA_TEST( predict_16x16, I_PRED_16x16_P, 16, 16, 64, 1 );
2442 report( "intra pred :" );
2446 #define DECL_CABAC(cpu) \
2447 static void run_cabac_decision_##cpu( x264_t *h, uint8_t *dst )\
2450 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2451 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2452 for( int i = 0; i < 0x1000; i++ )\
2453 x264_cabac_encode_decision_##cpu( &cb, buf1[i]>>1, buf1[i]&1 );\
2455 static void run_cabac_bypass_##cpu( x264_t *h, uint8_t *dst )\
2458 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2459 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2460 for( int i = 0; i < 0x1000; i++ )\
2461 x264_cabac_encode_bypass_##cpu( &cb, buf1[i]&1 );\
2463 static void run_cabac_terminal_##cpu( x264_t *h, uint8_t *dst )\
2466 x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2467 x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2468 for( int i = 0; i < 0x1000; i++ )\
2469 x264_cabac_encode_terminal_##cpu( &cb );\
2474 #elif defined(ARCH_AARCH64)
2477 #define run_cabac_decision_asm run_cabac_decision_c
2478 #define run_cabac_bypass_asm run_cabac_bypass_c
2479 #define run_cabac_terminal_asm run_cabac_terminal_c
2482 extern const uint8_t x264_count_cat_m1[14];
2483 void x264_cabac_block_residual_c( x264_t *h, x264_cabac_t *cb, int ctx_block_cat, dctcoef *l );
2484 void x264_cabac_block_residual_8x8_rd_c( x264_t *h, x264_cabac_t *cb, int ctx_block_cat, dctcoef *l );
2485 void x264_cabac_block_residual_rd_c( x264_t *h, x264_cabac_t *cb, int ctx_block_cat, dctcoef *l );
2487 static int check_cabac( int cpu_ref, int cpu_new )
2489 int ret = 0, ok = 1, used_asm = 0;
2491 h.sps->i_chroma_format_idc = 3;
2493 x264_bitstream_function_t bs_ref;
2494 x264_bitstream_function_t bs_a;
2495 x264_bitstream_init( cpu_ref, &bs_ref );
2496 x264_bitstream_init( cpu_new, &bs_a );
2497 x264_quant_init( &h, cpu_new, &h.quantf );
2498 h.quantf.coeff_last[DCT_CHROMA_DC] = h.quantf.coeff_last4;
2500 #define CABAC_RESIDUAL(name, start, end, rd)\
2502 if( bs_a.name##_internal && (bs_a.name##_internal != bs_ref.name##_internal || (cpu_new&X264_CPU_SSE2_IS_SLOW)) )\
2505 set_func_name( #name );\
2506 for( int i = 0; i < 2; i++ )\
2508 for( intptr_t ctx_block_cat = start; ctx_block_cat <= end; ctx_block_cat++ )\
2510 for( int j = 0; j < 256; j++ )\
2512 ALIGNED_ARRAY_N( dctcoef, dct, [2],[64] );\
2513 uint8_t bitstream[2][1<<16];\
2514 static const uint8_t ctx_ac[14] = {0,1,0,0,1,0,0,1,0,0,0,1,0,0};\
2515 int ac = ctx_ac[ctx_block_cat];\
2519 for( int k = 0; k <= x264_count_cat_m1[ctx_block_cat]; k++ )\
2521 /* Very rough distribution that covers possible inputs */\
2523 int coef = !(rnd&3);\
2524 coef += !(rnd& 15) * (rand()&0x0006);\
2525 coef += !(rnd& 63) * (rand()&0x0008);\
2526 coef += !(rnd& 255) * (rand()&0x00F0);\
2527 coef += !(rnd&1023) * (rand()&0x7F00);\
2528 nz |= dct[0][ac+k] = dct[1][ac+k] = coef * ((rand()&1) ? 1 : -1);\
2531 h.mb.b_interlaced = i;\
2532 x264_cabac_t cb[2];\
2533 x264_cabac_context_init( &h, &cb[0], SLICE_TYPE_P, 26, 0 );\
2534 x264_cabac_context_init( &h, &cb[1], SLICE_TYPE_P, 26, 0 );\
2535 x264_cabac_encode_init( &cb[0], bitstream[0], bitstream[0]+0xfff0 );\
2536 x264_cabac_encode_init( &cb[1], bitstream[1], bitstream[1]+0xfff0 );\
2537 cb[0].f8_bits_encoded = 0;\
2538 cb[1].f8_bits_encoded = 0;\
2539 if( !rd ) memcpy( bitstream[1], bitstream[0], 0x400 );\
2540 call_c1( x264_##name##_c, &h, &cb[0], ctx_block_cat, dct[0]+ac );\
2541 call_a1( bs_a.name##_internal, dct[1]+ac, i, ctx_block_cat, &cb[1] );\
2542 ok = cb[0].f8_bits_encoded == cb[1].f8_bits_encoded && !memcmp(cb[0].state, cb[1].state, 1024);\
2543 if( !rd ) ok |= !memcmp( bitstream[1], bitstream[0], 0x400 ) && !memcmp( &cb[1], &cb[0], offsetof(x264_cabac_t, p_start) );\
2546 fprintf( stderr, #name " : [FAILED] ctx_block_cat %d", (int)ctx_block_cat );\
2547 if( rd && cb[0].f8_bits_encoded != cb[1].f8_bits_encoded )\
2548 fprintf( stderr, " (%d != %d)", cb[0].f8_bits_encoded, cb[1].f8_bits_encoded );\
2549 fprintf( stderr, "\n");\
2554 call_c2( x264_##name##_c, &h, &cb[0], ctx_block_cat, dct[0]+ac );\
2555 call_a2( bs_a.name##_internal, dct[1]+ac, i, ctx_block_cat, &cb[1] );\
2564 CABAC_RESIDUAL( cabac_block_residual, 0, DCT_LUMA_8x8, 0 )
2565 report( "cabac residual:" );
2567 ok = 1; used_asm = 0;
2568 CABAC_RESIDUAL( cabac_block_residual_rd, 0, DCT_LUMA_8x8-1, 1 )
2569 CABAC_RESIDUAL( cabac_block_residual_8x8_rd, DCT_LUMA_8x8, DCT_LUMA_8x8, 1 )
2570 report( "cabac residual rd:" );
2572 if( cpu_ref || run_cabac_decision_c == run_cabac_decision_asm )
2574 ok = 1; used_asm = 0;
2575 x264_cabac_init( &h );
2577 set_func_name( "cabac_encode_decision" );
2578 memcpy( buf4, buf3, 0x1000 );
2579 call_c( run_cabac_decision_c, &h, buf3 );
2580 call_a( run_cabac_decision_asm, &h, buf4 );
2581 ok = !memcmp( buf3, buf4, 0x1000 );
2582 report( "cabac decision:" );
2584 set_func_name( "cabac_encode_bypass" );
2585 memcpy( buf4, buf3, 0x1000 );
2586 call_c( run_cabac_bypass_c, &h, buf3 );
2587 call_a( run_cabac_bypass_asm, &h, buf4 );
2588 ok = !memcmp( buf3, buf4, 0x1000 );
2589 report( "cabac bypass:" );
2591 set_func_name( "cabac_encode_terminal" );
2592 memcpy( buf4, buf3, 0x1000 );
2593 call_c( run_cabac_terminal_c, &h, buf3 );
2594 call_a( run_cabac_terminal_asm, &h, buf4 );
2595 ok = !memcmp( buf3, buf4, 0x1000 );
2596 report( "cabac terminal:" );
2601 static int check_bitstream( int cpu_ref, int cpu_new )
2603 x264_bitstream_function_t bs_c;
2604 x264_bitstream_function_t bs_ref;
2605 x264_bitstream_function_t bs_a;
2607 int ret = 0, ok = 1, used_asm = 0;
2609 x264_bitstream_init( 0, &bs_c );
2610 x264_bitstream_init( cpu_ref, &bs_ref );
2611 x264_bitstream_init( cpu_new, &bs_a );
2612 if( bs_a.nal_escape != bs_ref.nal_escape )
2615 uint8_t *input = malloc(size+100);
2616 uint8_t *output1 = malloc(size*2);
2617 uint8_t *output2 = malloc(size*2);
2619 set_func_name( "nal_escape" );
2620 for( int i = 0; i < 100; i++ )
2622 /* Test corner-case sizes */
2623 int test_size = i < 10 ? i+1 : rand() & 0x3fff;
2624 /* Test 8 different probability distributions of zeros */
2625 for( int j = 0; j < test_size+32; j++ )
2626 input[j] = (rand()&((1 << ((i&7)+1)) - 1)) * rand();
2627 uint8_t *end_c = (uint8_t*)call_c1( bs_c.nal_escape, output1, input, input+test_size );
2628 uint8_t *end_a = (uint8_t*)call_a1( bs_a.nal_escape, output2, input, input+test_size );
2629 int size_c = end_c-output1;
2630 int size_a = end_a-output2;
2631 if( size_c != size_a || memcmp( output1, output2, size_c ) )
2633 fprintf( stderr, "nal_escape : [FAILED] %d %d\n", size_c, size_a );
2638 for( int j = 0; j < size+32; j++ )
2640 call_c2( bs_c.nal_escape, output1, input, input+size );
2641 call_a2( bs_a.nal_escape, output2, input, input+size );
2646 report( "nal escape:" );
2651 static int check_all_funcs( int cpu_ref, int cpu_new )
2653 return check_pixel( cpu_ref, cpu_new )
2654 + check_dct( cpu_ref, cpu_new )
2655 + check_mc( cpu_ref, cpu_new )
2656 + check_intra( cpu_ref, cpu_new )
2657 + check_deblock( cpu_ref, cpu_new )
2658 + check_quant( cpu_ref, cpu_new )
2659 + check_cabac( cpu_ref, cpu_new )
2660 + check_bitstream( cpu_ref, cpu_new );
2663 static int add_flags( int *cpu_ref, int *cpu_new, int flags, const char *name )
2665 *cpu_ref = *cpu_new;
2667 #if STACK_ALIGNMENT < 16
2668 *cpu_new |= X264_CPU_STACK_MOD4;
2670 if( *cpu_new & X264_CPU_SSE2_IS_FAST )
2671 *cpu_new &= ~X264_CPU_SSE2_IS_SLOW;
2673 fprintf( stderr, "x264: %s\n", name );
2674 return check_all_funcs( *cpu_ref, *cpu_new );
2677 static int check_all_flags( void )
2680 int cpu0 = 0, cpu1 = 0;
2681 uint32_t cpu_detect = x264_cpu_detect();
2683 if( cpu_detect & X264_CPU_MMX2 )
2685 ret |= add_flags( &cpu0, &cpu1, X264_CPU_MMX | X264_CPU_MMX2, "MMX" );
2686 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "MMX Cache64" );
2687 cpu1 &= ~X264_CPU_CACHELINE_64;
2689 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_32, "MMX Cache32" );
2690 cpu1 &= ~X264_CPU_CACHELINE_32;
2692 if( cpu_detect & X264_CPU_LZCNT )
2694 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "MMX LZCNT" );
2695 cpu1 &= ~X264_CPU_LZCNT;
2697 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "MMX SlowCTZ" );
2698 cpu1 &= ~X264_CPU_SLOW_CTZ;
2700 if( cpu_detect & X264_CPU_SSE )
2701 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE, "SSE" );
2702 if( cpu_detect & X264_CPU_SSE2 )
2704 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2 | X264_CPU_SSE2_IS_SLOW, "SSE2Slow" );
2705 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2_IS_FAST, "SSE2Fast" );
2706 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSE2Fast Cache64" );
2707 cpu1 &= ~X264_CPU_CACHELINE_64;
2708 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_SHUFFLE, "SSE2 SlowShuffle" );
2709 cpu1 &= ~X264_CPU_SLOW_SHUFFLE;
2710 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSE2 SlowCTZ" );
2711 cpu1 &= ~X264_CPU_SLOW_CTZ;
2712 if( cpu_detect & X264_CPU_LZCNT )
2714 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSE2 LZCNT" );
2715 cpu1 &= ~X264_CPU_LZCNT;
2718 if( cpu_detect & X264_CPU_SSE3 )
2720 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE3 | X264_CPU_CACHELINE_64, "SSE3" );
2721 cpu1 &= ~X264_CPU_CACHELINE_64;
2723 if( cpu_detect & X264_CPU_SSSE3 )
2725 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSSE3, "SSSE3" );
2726 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64" );
2727 cpu1 &= ~X264_CPU_CACHELINE_64;
2728 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_SHUFFLE, "SSSE3 SlowShuffle" );
2729 cpu1 &= ~X264_CPU_SLOW_SHUFFLE;
2730 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSSE3 SlowCTZ" );
2731 cpu1 &= ~X264_CPU_SLOW_CTZ;
2732 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSSE3 SlowAtom" );
2733 ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64 SlowAtom" );
2734 cpu1 &= ~X264_CPU_CACHELINE_64;
2735 cpu1 &= ~X264_CPU_SLOW_ATOM;
2736 if( cpu_detect & X264_CPU_LZCNT )
2738 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSSE3 LZCNT" );
2739 cpu1 &= ~X264_CPU_LZCNT;
2742 if( cpu_detect & X264_CPU_SSE4 )
2743 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE4, "SSE4" );
2744 if( cpu_detect & X264_CPU_SSE42 )
2745 ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE42, "SSE4.2" );
2746 if( cpu_detect & X264_CPU_AVX )
2747 ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX, "AVX" );
2748 if( cpu_detect & X264_CPU_XOP )
2749 ret |= add_flags( &cpu0, &cpu1, X264_CPU_XOP, "XOP" );
2750 if( cpu_detect & X264_CPU_FMA4 )
2752 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA4, "FMA4" );
2753 cpu1 &= ~X264_CPU_FMA4;
2755 if( cpu_detect & X264_CPU_FMA3 )
2757 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA3, "FMA3" );
2758 cpu1 &= ~X264_CPU_FMA3;
2760 if( cpu_detect & X264_CPU_AVX2 )
2762 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA3 | X264_CPU_AVX2, "AVX2" );
2763 if( cpu_detect & X264_CPU_LZCNT )
2765 ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "AVX2 LZCNT" );
2766 cpu1 &= ~X264_CPU_LZCNT;
2769 if( cpu_detect & X264_CPU_BMI1 )
2771 ret |= add_flags( &cpu0, &cpu1, X264_CPU_BMI1, "BMI1" );
2772 cpu1 &= ~X264_CPU_BMI1;
2774 if( cpu_detect & X264_CPU_BMI2 )
2776 ret |= add_flags( &cpu0, &cpu1, X264_CPU_BMI1|X264_CPU_BMI2, "BMI2" );
2777 cpu1 &= ~(X264_CPU_BMI1|X264_CPU_BMI2);
2780 if( cpu_detect & X264_CPU_ALTIVEC )
2782 fprintf( stderr, "x264: ALTIVEC against C\n" );
2783 ret = check_all_funcs( 0, X264_CPU_ALTIVEC );
2786 if( cpu_detect & X264_CPU_ARMV6 )
2787 ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV6, "ARMv6" );
2788 if( cpu_detect & X264_CPU_NEON )
2789 ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
2790 if( cpu_detect & X264_CPU_FAST_NEON_MRC )
2791 ret |= add_flags( &cpu0, &cpu1, X264_CPU_FAST_NEON_MRC, "Fast NEON MRC" );
2793 if( cpu_detect & X264_CPU_ARMV8 )
2794 ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV8, "ARMv8" );
2795 if( cpu_detect & X264_CPU_NEON )
2796 ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
2798 if( cpu_detect & X264_CPU_MSA )
2799 ret |= add_flags( &cpu0, &cpu1, X264_CPU_MSA, "MSA" );
2804 int main(int argc, char *argv[])
2808 if( argc > 1 && !strncmp( argv[1], "--bench", 7 ) )
2810 #if !ARCH_X86 && !ARCH_X86_64 && !ARCH_PPC && !ARCH_ARM && !ARCH_AARCH64 && !ARCH_MIPS
2811 fprintf( stderr, "no --bench for your cpu until you port rdtsc\n" );
2815 if( argv[1][7] == '=' )
2817 bench_pattern = argv[1]+8;
2818 bench_pattern_len = strlen(bench_pattern);
2824 int seed = ( argc > 1 ) ? atoi(argv[1]) : x264_mdate();
2825 fprintf( stderr, "x264: using random seed %u\n", seed );
2828 buf1 = x264_malloc( 0x1e00 + 0x2000*sizeof(pixel) + 32*BENCH_ALIGNS );
2829 pbuf1 = x264_malloc( 0x1e00*sizeof(pixel) + 32*BENCH_ALIGNS );
2830 if( !buf1 || !pbuf1 )
2832 fprintf( stderr, "malloc failed, unable to initiate tests!\n" );
2835 #define INIT_POINTER_OFFSETS\
2836 buf2 = buf1 + 0xf00;\
2837 buf3 = buf2 + 0xf00;\
2838 buf4 = buf3 + 0x1000*sizeof(pixel);\
2839 pbuf2 = pbuf1 + 0xf00;\
2840 pbuf3 = (pixel*)buf3;\
2841 pbuf4 = (pixel*)buf4;
2842 INIT_POINTER_OFFSETS;
2843 for( int i = 0; i < 0x1e00; i++ )
2845 buf1[i] = rand() & 0xFF;
2846 pbuf1[i] = rand() & PIXEL_MAX;
2848 memset( buf1+0x1e00, 0, 0x2000*sizeof(pixel) );
2850 /* 32-byte alignment is guaranteed whenever it's useful, but some functions also vary in speed depending on %64 */
2852 for( int i = 0; i < BENCH_ALIGNS && !ret; i++ )
2854 INIT_POINTER_OFFSETS;
2855 ret |= x264_stack_pagealign( check_all_flags, i*32 );
2859 fprintf( stderr, "%d/%d\r", i+1, BENCH_ALIGNS );
2862 ret = x264_stack_pagealign( check_all_flags, 0 );
2866 fprintf( stderr, "x264: at least one test has failed. Go and fix that Right Now!\n" );
2869 fprintf( stderr, "x264: All tests passed Yeah :)\n" );