CoCalc -- checkasm.c

05. Matplotlib / ffmpeg-3.0 / libx264 / tools / checkasm.c
⁵²⁸⁶⁷ views
1
/*****************************************************************************
2
 * checkasm.c: assembly check tool
3
 *****************************************************************************
4
 * Copyright (C) 2003-2016 x264 project
5
 *
6
 * Authors: Loren Merritt <[email protected]>
7
 *          Laurent Aimar <[email protected]>
8
 *          Fiona Glaser <[email protected]>
9
 *
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.
14
 *
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.
19
 *
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.
23
 *
24
 * This program is also available under a commercial proprietary license.
25
 * For more information, contact us at [email protected].
26
 *****************************************************************************/
27

28
#include <ctype.h>
29
#include "common/common.h"
30
#include "common/cpu.h"
31

32
// GCC doesn't align stack variables on ARM, so use .bss
33
#if ARCH_ARM
34
#undef ALIGNED_16
35
#define ALIGNED_16( var ) DECLARE_ALIGNED( static var, 16 )
36
#endif
37

38
/* buf1, buf2: initialised to random data and shouldn't write into them */
39
uint8_t *buf1, *buf2;
40
/* buf3, buf4: used to store output */
41
uint8_t *buf3, *buf4;
42
/* pbuf1, pbuf2: initialised to random pixel data and shouldn't write into them. */
43
pixel *pbuf1, *pbuf2;
44
/* pbuf3, pbuf4: point to buf3, buf4, just for type convenience */
45
pixel *pbuf3, *pbuf4;
46

47
int quiet = 0;
48

49
#define report( name ) { \
50
    if( used_asm && !quiet ) \
51
        fprintf( stderr, " - %-21s [%s]\n", name, ok ? "OK" : "FAILED" ); \
52
    if( !ok ) ret = -1; \
53
}
54

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
59

60
typedef struct
61
{
62
    void *pointer; // just for detecting duplicates
63
    uint32_t cpu;
64
    uint64_t cycles;
65
    uint32_t den;
66
} bench_t;
67

68
typedef struct
69
{
70
    char *name;
71
    bench_t vers[MAX_CPUS];
72
} bench_func_t;
73

74
int do_bench = 0;
75
int bench_pattern_len = 0;
76
const char *bench_pattern = "";
77
char func_name[100];
78
static bench_func_t benchs[MAX_FUNCS];
79

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;
86

87
#define set_func_name(...) snprintf( func_name, sizeof(func_name), __VA_ARGS__ )
88

89
static inline uint32_t read_time(void)
90
{
91
    uint32_t a = 0;
92
#if HAVE_X86_INLINE_ASM
93
    asm volatile( "lfence \n"
94
                  "rdtsc  \n"
95
                  : "=a"(a) :: "edx", "memory" );
96
#elif ARCH_PPC
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" );
100
#elif ARCH_AARCH64
101
    uint64_t b = 0;
102
    asm volatile( "mrs %0, pmccntr_el0" : "=r"(b) :: "memory" );
103
    a = b;
104
#elif ARCH_MIPS
105
    asm volatile( "rdhwr %0, $2" : "=r"(a) :: "memory" );
106
#endif
107
    return a;
108
}
109

110
static bench_t* get_bench( const char *name, int cpu )
111
{
112
    int i, j;
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 );
117
    if( !cpu )
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];
123
}
124

125
static int cmp_nop( const void *a, const void *b )
126
{
127
    return *(uint16_t*)a - *(uint16_t*)b;
128
}
129

130
static int cmp_bench( const void *a, const void *b )
131
{
132
    // asciibetical sort except preserving numbers
133
    const char *sa = ((bench_func_t*)a)->name;
134
    const char *sb = ((bench_func_t*)b)->name;
135
    for( ;; sa++, sb++ )
136
    {
137
        if( !*sa && !*sb )
138
            return 0;
139
        if( isdigit( *sa ) && isdigit( *sb ) && isdigit( sa[1] ) != isdigit( sb[1] ) )
140
            return isdigit( sa[1] ) - isdigit( sb[1] );
141
        if( *sa != *sb )
142
            return *sa - *sb;
143
    }
144
}
145

146
static void print_bench(void)
147
{
148
    uint16_t nops[10000];
149
    int nfuncs, nop_time=0;
150

151
    for( int i = 0; i < 10000; i++ )
152
    {
153
        uint32_t t = read_time();
154
        nops[i] = read_time() - t;
155
    }
156
    qsort( nops, 10000, sizeof(uint16_t), cmp_nop );
157
    for( int i = 500; i < 9500; i++ )
158
        nop_time += nops[i];
159
    nop_time /= 900;
160
    printf( "nop: %d\n", nop_time );
161

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++ )
166
        {
167
            int k;
168
            bench_t *b = &benchs[i].vers[j];
169
            if( !b->den )
170
                continue;
171
            for( k = 0; k < j && benchs[i].vers[k].pointer != b->pointer; k++ );
172
            if( k < j )
173
                continue;
174
            printf( "%s_%s%s: %"PRId64"\n", benchs[i].name,
175
#if HAVE_MMX
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" :
190
#elif ARCH_PPC
191
                    b->cpu&X264_CPU_ALTIVEC ? "altivec" :
192
#elif ARCH_ARM
193
                    b->cpu&X264_CPU_NEON ? "neon" :
194
                    b->cpu&X264_CPU_ARMV6 ? "armv6" :
195
#elif ARCH_AARCH64
196
                    b->cpu&X264_CPU_NEON ? "neon" :
197
                    b->cpu&X264_CPU_ARMV8 ? "armv8" :
198
#elif ARCH_MIPS
199
                    b->cpu&X264_CPU_MSA ? "msa" :
200
#endif
201
                    "c",
202
#if HAVE_MMX
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" :
212
#elif ARCH_ARM
213
                    b->cpu&X264_CPU_FAST_NEON_MRC ? "_fast_mrc" :
214
#endif
215
                    "",
216
                    (int64_t)(10*b->cycles/b->den - nop_time)/4 );
217
        }
218
}
219

220
#if ARCH_X86 || ARCH_X86_64
221
int x264_stack_pagealign( int (*func)(), int align );
222

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, ... );
226
#else
227
#define x264_stack_pagealign( func, align ) func()
228
#endif
229

230
#if ARCH_AARCH64
231
intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
232
#endif
233

234
#if ARCH_ARM
235
intptr_t x264_checkasm_call_neon( intptr_t (*func)(), int *ok, ... );
236
intptr_t x264_checkasm_call_noneon( intptr_t (*func)(), int *ok, ... );
237
intptr_t (*x264_checkasm_call)( intptr_t (*func)(), int *ok, ... ) = x264_checkasm_call_noneon;
238
#endif
239

240
#define call_c1(func,...) func(__VA_ARGS__)
241

242
#if ARCH_X86_64
243
/* Evil hack: detect incorrect assumptions that 32-bit ints are zero-extended to 64-bit.
244
 * This is done by clobbering the stack with junk around the stack pointer and calling the
245
 * assembly function through x264_checkasm_call with added dummy arguments which forces all
246
 * real arguments to be passed on the stack and not in registers. For 32-bit argument the
247
 * upper half of the 64-bit register location on the stack will now contain junk. Note that
248
 * this is dependant on compiler behaviour and that interrupts etc. at the wrong time may
249
 * overwrite the junk written to the stack so there's no guarantee that it will always
250
 * detect all functions that assumes zero-extension.
251
 */
252
void x264_checkasm_stack_clobber( uint64_t clobber, ... );
253
#define call_a1(func,...) ({ \
254
    uint64_t r = (rand() & 0xffff) * 0x0001000100010001ULL; \
255
    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 */ \
256
    x264_checkasm_call(( intptr_t(*)())func, &ok, 0, 0, 0, 0, __VA_ARGS__ ); })
257
#elif ARCH_X86 || (ARCH_AARCH64 && !defined(__APPLE__)) || ARCH_ARM
258
#define call_a1(func,...) x264_checkasm_call( (intptr_t(*)())func, &ok, __VA_ARGS__ )
259
#else
260
#define call_a1 call_c1
261
#endif
262

263
#if ARCH_ARM
264
#define call_a1_64(func,...) ((uint64_t (*)(intptr_t(*)(), int*, ...))x264_checkasm_call)( (intptr_t(*)())func, &ok, __VA_ARGS__ )
265
#else
266
#define call_a1_64 call_a1
267
#endif
268

269
#define call_bench(func,cpu,...)\
270
    if( do_bench && !strncmp(func_name, bench_pattern, bench_pattern_len) )\
271
    {\
272
        uint64_t tsum = 0;\
273
        int tcount = 0;\
274
        call_a1(func, __VA_ARGS__);\
275
        for( int ti = 0; ti < (cpu?BENCH_RUNS:BENCH_RUNS/4); ti++ )\
276
        {\
277
            uint32_t t = read_time();\
278
            func(__VA_ARGS__);\
279
            func(__VA_ARGS__);\
280
            func(__VA_ARGS__);\
281
            func(__VA_ARGS__);\
282
            t = read_time() - t;\
283
            if( (uint64_t)t*tcount <= tsum*4 && ti > 0 )\
284
            {\
285
                tsum += t;\
286
                tcount++;\
287
            }\
288
        }\
289
        bench_t *b = get_bench( func_name, cpu );\
290
        b->cycles += tsum;\
291
        b->den += tcount;\
292
        b->pointer = func;\
293
    }
294

295
/* for most functions, run benchmark and correctness test at the same time.
296
 * for those that modify their inputs, run the above macros separately */
297
#define call_a(func,...) ({ call_a2(func,__VA_ARGS__); call_a1(func,__VA_ARGS__); })
298
#define call_c(func,...) ({ call_c2(func,__VA_ARGS__); call_c1(func,__VA_ARGS__); })
299
#define call_a2(func,...) ({ call_bench(func,cpu_new,__VA_ARGS__); })
300
#define call_c2(func,...) ({ call_bench(func,0,__VA_ARGS__); })
301
#define call_a64(func,...) ({ call_a2(func,__VA_ARGS__); call_a1_64(func,__VA_ARGS__); })
302

303

304
static int check_pixel( int cpu_ref, int cpu_new )
305
{
306
    x264_pixel_function_t pixel_c;
307
    x264_pixel_function_t pixel_ref;
308
    x264_pixel_function_t pixel_asm;
309
    x264_predict_t predict_4x4[12];
310
    x264_predict8x8_t predict_8x8[12];
311
    x264_predict_8x8_filter_t predict_8x8_filter;
312
    ALIGNED_16( pixel edge[36] );
313
    uint16_t cost_mv[32];
314
    int ret = 0, ok, used_asm;
315

316
    x264_pixel_init( 0, &pixel_c );
317
    x264_pixel_init( cpu_ref, &pixel_ref );
318
    x264_pixel_init( cpu_new, &pixel_asm );
319
    x264_predict_4x4_init( 0, predict_4x4 );
320
    x264_predict_8x8_init( 0, predict_8x8, &predict_8x8_filter );
321
    predict_8x8_filter( pbuf2+40, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
322

323
    // maximize sum
324
    for( int i = 0; i < 256; i++ )
325
    {
326
        int z = i|(i>>4);
327
        z ^= z>>2;
328
        z ^= z>>1;
329
        pbuf4[i] = -(z&1) & PIXEL_MAX;
330
        pbuf3[i] = ~pbuf4[i] & PIXEL_MAX;
331
    }
332
    // random pattern made of maxed pixel differences, in case an intermediate value overflows
333
    for( int i = 256; i < 0x1000; i++ )
334
    {
335
        pbuf4[i] = -(pbuf1[i&~0x88]&1) & PIXEL_MAX;
336
        pbuf3[i] = ~(pbuf4[i]) & PIXEL_MAX;
337
    }
338

339
#define TEST_PIXEL( name, align ) \
340
    ok = 1, used_asm = 0; \
341
    for( int i = 0; i < ARRAY_ELEMS(pixel_c.name); i++ ) \
342
    { \
343
        int res_c, res_asm; \
344
        if( pixel_asm.name[i] != pixel_ref.name[i] ) \
345
        { \
346
            set_func_name( "%s_%s", #name, pixel_names[i] ); \
347
            used_asm = 1; \
348
            for( int j = 0; j < 64; j++ ) \
349
            { \
350
                res_c   = call_c( pixel_c.name[i],   pbuf1, (intptr_t)16, pbuf2+j*!align, (intptr_t)64 ); \
351
                res_asm = call_a( pixel_asm.name[i], pbuf1, (intptr_t)16, pbuf2+j*!align, (intptr_t)64 ); \
352
                if( res_c != res_asm ) \
353
                { \
354
                    ok = 0; \
355
                    fprintf( stderr, #name "[%d]: %d != %d [FAILED]\n", i, res_c, res_asm ); \
356
                    break; \
357
                } \
358
            } \
359
            for( int j = 0; j < 0x1000 && ok; j += 256 ) \
360
            { \
361
                res_c   = pixel_c  .name[i]( pbuf3+j, 16, pbuf4+j, 16 ); \
362
                res_asm = pixel_asm.name[i]( pbuf3+j, 16, pbuf4+j, 16 ); \
363
                if( res_c != res_asm ) \
364
                { \
365
                    ok = 0; \
366
                    fprintf( stderr, #name "[%d]: overflow %d != %d\n", i, res_c, res_asm ); \
367
                } \
368
            } \
369
        } \
370
    } \
371
    report( "pixel " #name " :" );
372

373
    TEST_PIXEL( sad, 0 );
374
    TEST_PIXEL( sad_aligned, 1 );
375
    TEST_PIXEL( ssd, 1 );
376
    TEST_PIXEL( satd, 0 );
377
    TEST_PIXEL( sa8d, 1 );
378

379
    ok = 1, used_asm = 0;
380
    if( pixel_asm.sa8d_satd[PIXEL_16x16] != pixel_ref.sa8d_satd[PIXEL_16x16] )
381
    {
382
        set_func_name( "sa8d_satd_%s", pixel_names[PIXEL_16x16] );
383
        used_asm = 1;
384
        for( int j = 0; j < 64; j++ )
385
        {
386
            uint32_t cost8_c = pixel_c.sa8d[PIXEL_16x16]( pbuf1, 16, pbuf2, 64 );
387
            uint32_t cost4_c = pixel_c.satd[PIXEL_16x16]( pbuf1, 16, pbuf2, 64 );
388
            uint64_t res_a = call_a64( pixel_asm.sa8d_satd[PIXEL_16x16], pbuf1, (intptr_t)16, pbuf2, (intptr_t)64 );
389
            uint32_t cost8_a = res_a;
390
            uint32_t cost4_a = res_a >> 32;
391
            if( cost8_a != cost8_c || cost4_a != cost4_c )
392
            {
393
                ok = 0;
394
                fprintf( stderr, "sa8d_satd [%d]: (%d,%d) != (%d,%d) [FAILED]\n", PIXEL_16x16,
395
                         cost8_c, cost4_c, cost8_a, cost4_a );
396
                break;
397
            }
398
        }
399
        for( int j = 0; j < 0x1000 && ok; j += 256 ) \
400
        {
401
            uint32_t cost8_c = pixel_c.sa8d[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
402
            uint32_t cost4_c = pixel_c.satd[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
403
            uint64_t res_a = pixel_asm.sa8d_satd[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
404
            uint32_t cost8_a = res_a;
405
            uint32_t cost4_a = res_a >> 32;
406
            if( cost8_a != cost8_c || cost4_a != cost4_c )
407
            {
408
                ok = 0;
409
                fprintf( stderr, "sa8d_satd [%d]: overflow (%d,%d) != (%d,%d) [FAILED]\n", PIXEL_16x16,
410
                         cost8_c, cost4_c, cost8_a, cost4_a );
411
            }
412
        }
413
    }
414
    report( "pixel sa8d_satd :" );
415

416
#define TEST_PIXEL_X( N ) \
417
    ok = 1; used_asm = 0; \
418
    for( int i = 0; i < 7; i++ ) \
419
    { \
420
        ALIGNED_16( int res_c[4] ) = {0}; \
421
        ALIGNED_16( int res_asm[4] ) = {0}; \
422
        if( pixel_asm.sad_x##N[i] && pixel_asm.sad_x##N[i] != pixel_ref.sad_x##N[i] ) \
423
        { \
424
            set_func_name( "sad_x%d_%s", N, pixel_names[i] ); \
425
            used_asm = 1; \
426
            for( int j = 0; j < 64; j++ ) \
427
            { \
428
                pixel *pix2 = pbuf2+j; \
429
                res_c[0] = pixel_c.sad[i]( pbuf1, 16, pix2,   64 ); \
430
                res_c[1] = pixel_c.sad[i]( pbuf1, 16, pix2+6, 64 ); \
431
                res_c[2] = pixel_c.sad[i]( pbuf1, 16, pix2+1, 64 ); \
432
                if( N == 4 ) \
433
                { \
434
                    res_c[3] = pixel_c.sad[i]( pbuf1, 16, pix2+10, 64 ); \
435
                    call_a( pixel_asm.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, (intptr_t)64, res_asm ); \
436
                } \
437
                else \
438
                    call_a( pixel_asm.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, (intptr_t)64, res_asm ); \
439
                if( memcmp(res_c, res_asm, N*sizeof(int)) ) \
440
                { \
441
                    ok = 0; \
442
                    fprintf( stderr, "sad_x"#N"[%d]: %d,%d,%d,%d != %d,%d,%d,%d [FAILED]\n", \
443
                             i, res_c[0], res_c[1], res_c[2], res_c[3], \
444
                             res_asm[0], res_asm[1], res_asm[2], res_asm[3] ); \
445
                } \
446
                if( N == 4 ) \
447
                    call_c2( pixel_c.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, (intptr_t)64, res_asm ); \
448
                else \
449
                    call_c2( pixel_c.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, (intptr_t)64, res_asm ); \
450
            } \
451
        } \
452
    } \
453
    report( "pixel sad_x"#N" :" );
454

455
    TEST_PIXEL_X(3);
456
    TEST_PIXEL_X(4);
457

458
#define TEST_PIXEL_VAR( i ) \
459
    if( pixel_asm.var[i] != pixel_ref.var[i] ) \
460
    { \
461
        set_func_name( "%s_%s", "var", pixel_names[i] ); \
462
        used_asm = 1; \
463
        /* abi-check wrapper can't return uint64_t, so separate it from return value check */ \
464
        call_c1( pixel_c.var[i],   pbuf1,           16 ); \
465
        call_a1( pixel_asm.var[i], pbuf1, (intptr_t)16 ); \
466
        uint64_t res_c   = pixel_c.var[i]( pbuf1, 16 ); \
467
        uint64_t res_asm = pixel_asm.var[i]( pbuf1, 16 ); \
468
        if( res_c != res_asm ) \
469
        { \
470
            ok = 0; \
471
            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) ); \
472
        } \
473
        call_c2( pixel_c.var[i],   pbuf1, (intptr_t)16 ); \
474
        call_a2( pixel_asm.var[i], pbuf1, (intptr_t)16 ); \
475
    }
476

477
    ok = 1; used_asm = 0;
478
    TEST_PIXEL_VAR( PIXEL_16x16 );
479
    TEST_PIXEL_VAR( PIXEL_8x16 );
480
    TEST_PIXEL_VAR( PIXEL_8x8 );
481
    report( "pixel var :" );
482

483
#define TEST_PIXEL_VAR2( i ) \
484
    if( pixel_asm.var2[i] != pixel_ref.var2[i] ) \
485
    { \
486
        int res_c, res_asm, ssd_c, ssd_asm; \
487
        set_func_name( "%s_%s", "var2", pixel_names[i] ); \
488
        used_asm = 1; \
489
        res_c   = call_c( pixel_c.var2[i],   pbuf1, (intptr_t)16, pbuf2, (intptr_t)16, &ssd_c   ); \
490
        res_asm = call_a( pixel_asm.var2[i], pbuf1, (intptr_t)16, pbuf2, (intptr_t)16, &ssd_asm ); \
491
        if( res_c != res_asm || ssd_c != ssd_asm ) \
492
        { \
493
            ok = 0; \
494
            fprintf( stderr, "var2[%d]: %d != %d or %d != %d [FAILED]\n", i, res_c, res_asm, ssd_c, ssd_asm ); \
495
        } \
496
    }
497

498
    ok = 1; used_asm = 0;
499
    TEST_PIXEL_VAR2( PIXEL_8x16 );
500
    TEST_PIXEL_VAR2( PIXEL_8x8 );
501
    report( "pixel var2 :" );
502

503
    ok = 1; used_asm = 0;
504
    for( int i = 0; i < 4; i++ )
505
        if( pixel_asm.hadamard_ac[i] != pixel_ref.hadamard_ac[i] )
506
        {
507
            set_func_name( "hadamard_ac_%s", pixel_names[i] );
508
            used_asm = 1;
509
            for( int j = 0; j < 32; j++ )
510
            {
511
                pixel *pix = (j&16 ? pbuf1 : pbuf3) + (j&15)*256;
512
                call_c1( pixel_c.hadamard_ac[i],   pbuf1, (intptr_t)16 );
513
                call_a1( pixel_asm.hadamard_ac[i], pbuf1, (intptr_t)16 );
514
                uint64_t rc = pixel_c.hadamard_ac[i]( pix, 16 );
515
                uint64_t ra = pixel_asm.hadamard_ac[i]( pix, 16 );
516
                if( rc != ra )
517
                {
518
                    ok = 0;
519
                    fprintf( stderr, "hadamard_ac[%d]: %d,%d != %d,%d\n", i, (int)rc, (int)(rc>>32), (int)ra, (int)(ra>>32) );
520
                    break;
521
                }
522
            }
523
            call_c2( pixel_c.hadamard_ac[i],   pbuf1, (intptr_t)16 );
524
            call_a2( pixel_asm.hadamard_ac[i], pbuf1, (intptr_t)16 );
525
        }
526
    report( "pixel hadamard_ac :" );
527

528
    // maximize sum
529
    for( int i = 0; i < 32; i++ )
530
        for( int j = 0; j < 16; j++ )
531
            pbuf4[16*i+j] = -((i+j)&1) & PIXEL_MAX;
532
    ok = 1; used_asm = 0;
533
    if( pixel_asm.vsad != pixel_ref.vsad )
534
    {
535
        for( int h = 2; h <= 32; h += 2 )
536
        {
537
            int res_c, res_asm;
538
            set_func_name( "vsad" );
539
            used_asm = 1;
540
            for( int j = 0; j < 2 && ok; j++ )
541
            {
542
                pixel *p = j ? pbuf4 : pbuf1;
543
                res_c   = call_c( pixel_c.vsad,   p, (intptr_t)16, h );
544
                res_asm = call_a( pixel_asm.vsad, p, (intptr_t)16, h );
545
                if( res_c != res_asm )
546
                {
547
                    ok = 0;
548
                    fprintf( stderr, "vsad: height=%d, %d != %d\n", h, res_c, res_asm );
549
                    break;
550
                }
551
            }
552
        }
553
    }
554
    report( "pixel vsad :" );
555

556
    ok = 1; used_asm = 0;
557
    if( pixel_asm.asd8 != pixel_ref.asd8 )
558
    {
559
        set_func_name( "asd8" );
560
        used_asm = 1;
561
        int res_c = call_c( pixel_c.asd8,   pbuf1, (intptr_t)8, pbuf2, (intptr_t)8, 16 );
562
        int res_a = call_a( pixel_asm.asd8, pbuf1, (intptr_t)8, pbuf2, (intptr_t)8, 16 );
563
        if( res_c != res_a )
564
        {
565
            ok = 0;
566
            fprintf( stderr, "asd: %d != %d\n", res_c, res_a );
567
        }
568
    }
569
    report( "pixel asd :" );
570

571
#define TEST_INTRA_X3( name, i8x8, ... ) \
572
    if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
573
    { \
574
        ALIGNED_16( int res_c[3] ); \
575
        ALIGNED_16( int res_asm[3] ); \
576
        set_func_name( #name ); \
577
        used_asm = 1; \
578
        call_c( pixel_c.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_c ); \
579
        call_a( pixel_asm.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_asm ); \
580
        if( memcmp(res_c, res_asm, sizeof(res_c)) ) \
581
        { \
582
            ok = 0; \
583
            fprintf( stderr, #name": %d,%d,%d != %d,%d,%d [FAILED]\n", \
584
                     res_c[0], res_c[1], res_c[2], \
585
                     res_asm[0], res_asm[1], res_asm[2] ); \
586
        } \
587
    }
588

589
#define TEST_INTRA_X9( name, cmp ) \
590
    if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
591
    { \
592
        set_func_name( #name ); \
593
        used_asm = 1; \
594
        ALIGNED_ARRAY_64( uint16_t, bitcosts,[17] ); \
595
        for( int i=0; i<17; i++ ) \
596
            bitcosts[i] = 9*(i!=8); \
597
        memcpy( pbuf3, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) ); \
598
        memcpy( pbuf4, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) ); \
599
        for( int i=0; i<32; i++ ) \
600
        { \
601
            pixel *fenc = pbuf1+48+i*12; \
602
            pixel *fdec1 = pbuf3+48+i*12; \
603
            pixel *fdec2 = pbuf4+48+i*12; \
604
            int pred_mode = i%9; \
605
            int res_c = INT_MAX; \
606
            for( int j=0; j<9; j++ ) \
607
            { \
608
                predict_4x4[j]( fdec1 ); \
609
                int cost = pixel_c.cmp[PIXEL_4x4]( fenc, FENC_STRIDE, fdec1, FDEC_STRIDE ) + 9*(j!=pred_mode); \
610
                if( cost < (uint16_t)res_c ) \
611
                    res_c = cost + (j<<16); \
612
            } \
613
            predict_4x4[res_c>>16]( fdec1 ); \
614
            int res_a = call_a( pixel_asm.name, fenc, fdec2, bitcosts+8-pred_mode ); \
615
            if( res_c != res_a ) \
616
            { \
617
                ok = 0; \
618
                fprintf( stderr, #name": %d,%d != %d,%d [FAILED]\n", res_c>>16, res_c&0xffff, res_a>>16, res_a&0xffff ); \
619
                break; \
620
            } \
621
            if( memcmp(fdec1, fdec2, 4*FDEC_STRIDE*sizeof(pixel)) ) \
622
            { \
623
                ok = 0; \
624
                fprintf( stderr, #name" [FAILED]\n" ); \
625
                for( int j=0; j<16; j++ ) \
626
                    fprintf( stderr, "%02x ", fdec1[(j&3)+(j>>2)*FDEC_STRIDE] ); \
627
                fprintf( stderr, "\n" ); \
628
                for( int j=0; j<16; j++ ) \
629
                    fprintf( stderr, "%02x ", fdec2[(j&3)+(j>>2)*FDEC_STRIDE] ); \
630
                fprintf( stderr, "\n" ); \
631
                break; \
632
            } \
633
        } \
634
    }
635

636
#define TEST_INTRA8_X9( name, cmp ) \
637
    if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
638
    { \
639
        set_func_name( #name ); \
640
        used_asm = 1; \
641
        ALIGNED_ARRAY_64( uint16_t, bitcosts,[17] ); \
642
        ALIGNED_ARRAY_16( uint16_t, satds_c,[16] ); \
643
        ALIGNED_ARRAY_16( uint16_t, satds_a,[16] ); \
644
        memset( satds_c, 0, 16 * sizeof(*satds_c) ); \
645
        memset( satds_a, 0, 16 * sizeof(*satds_a) ); \
646
        for( int i=0; i<17; i++ ) \
647
            bitcosts[i] = 9*(i!=8); \
648
        for( int i=0; i<32; i++ ) \
649
        { \
650
            pixel *fenc = pbuf1+48+i*12; \
651
            pixel *fdec1 = pbuf3+48+i*12; \
652
            pixel *fdec2 = pbuf4+48+i*12; \
653
            int pred_mode = i%9; \
654
            int res_c = INT_MAX; \
655
            predict_8x8_filter( fdec1, edge, ALL_NEIGHBORS, ALL_NEIGHBORS ); \
656
            for( int j=0; j<9; j++ ) \
657
            { \
658
                predict_8x8[j]( fdec1, edge ); \
659
                satds_c[j] = pixel_c.cmp[PIXEL_8x8]( fenc, FENC_STRIDE, fdec1, FDEC_STRIDE ) + 9*(j!=pred_mode); \
660
                if( satds_c[j] < (uint16_t)res_c ) \
661
                    res_c = satds_c[j] + (j<<16); \
662
            } \
663
            predict_8x8[res_c>>16]( fdec1, edge ); \
664
            int res_a = call_a( pixel_asm.name, fenc, fdec2, edge, bitcosts+8-pred_mode, satds_a ); \
665
            if( res_c != res_a || memcmp(satds_c, satds_a, 16 * sizeof(*satds_c)) ) \
666
            { \
667
                ok = 0; \
668
                fprintf( stderr, #name": %d,%d != %d,%d [FAILED]\n", res_c>>16, res_c&0xffff, res_a>>16, res_a&0xffff ); \
669
                for( int j = 0; j < 9; j++ ) \
670
                    fprintf( stderr, "%5d ", satds_c[j]); \
671
                fprintf( stderr, "\n" ); \
672
                for( int j = 0; j < 9; j++ ) \
673
                    fprintf( stderr, "%5d ", satds_a[j]); \
674
                fprintf( stderr, "\n" ); \
675
                break; \
676
            } \
677
            for( int j=0; j<8; j++ ) \
678
                if( memcmp(fdec1+j*FDEC_STRIDE, fdec2+j*FDEC_STRIDE, 8*sizeof(pixel)) ) \
679
                    ok = 0; \
680
            if( !ok ) \
681
            { \
682
                fprintf( stderr, #name" [FAILED]\n" ); \
683
                for( int j=0; j<8; j++ ) \
684
                { \
685
                    for( int k=0; k<8; k++ ) \
686
                        fprintf( stderr, "%02x ", fdec1[k+j*FDEC_STRIDE] ); \
687
                    fprintf( stderr, "\n" ); \
688
                } \
689
                fprintf( stderr, "\n" ); \
690
                for( int j=0; j<8; j++ ) \
691
                { \
692
                    for( int k=0; k<8; k++ ) \
693
                        fprintf( stderr, "%02x ", fdec2[k+j*FDEC_STRIDE] ); \
694
                    fprintf( stderr, "\n" ); \
695
                } \
696
                fprintf( stderr, "\n" ); \
697
                break; \
698
            } \
699
        } \
700
    }
701

702
    memcpy( pbuf3, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) );
703
    ok = 1; used_asm = 0;
704
    TEST_INTRA_X3( intra_satd_x3_16x16, 0 );
705
    TEST_INTRA_X3( intra_satd_x3_8x16c, 0 );
706
    TEST_INTRA_X3( intra_satd_x3_8x8c, 0 );
707
    TEST_INTRA_X3( intra_sa8d_x3_8x8, 1, edge );
708
    TEST_INTRA_X3( intra_satd_x3_4x4, 0 );
709
    report( "intra satd_x3 :" );
710
    ok = 1; used_asm = 0;
711
    TEST_INTRA_X3( intra_sad_x3_16x16, 0 );
712
    TEST_INTRA_X3( intra_sad_x3_8x16c, 0 );
713
    TEST_INTRA_X3( intra_sad_x3_8x8c, 0 );
714
    TEST_INTRA_X3( intra_sad_x3_8x8, 1, edge );
715
    TEST_INTRA_X3( intra_sad_x3_4x4, 0 );
716
    report( "intra sad_x3 :" );
717
    ok = 1; used_asm = 0;
718
    TEST_INTRA_X9( intra_satd_x9_4x4, satd );
719
    TEST_INTRA8_X9( intra_sa8d_x9_8x8, sa8d );
720
    report( "intra satd_x9 :" );
721
    ok = 1; used_asm = 0;
722
    TEST_INTRA_X9( intra_sad_x9_4x4, sad );
723
    TEST_INTRA8_X9( intra_sad_x9_8x8, sad );
724
    report( "intra sad_x9 :" );
725

726
    ok = 1; used_asm = 0;
727
    if( pixel_asm.ssd_nv12_core != pixel_ref.ssd_nv12_core )
728
    {
729
        used_asm = 1;
730
        set_func_name( "ssd_nv12" );
731
        uint64_t res_u_c, res_v_c, res_u_a, res_v_a;
732
        for( int w = 8; w <= 360; w += 8 )
733
        {
734
            pixel_c.ssd_nv12_core(   pbuf1, 368, pbuf2, 368, w, 8, &res_u_c, &res_v_c );
735
            pixel_asm.ssd_nv12_core( pbuf1, 368, pbuf2, 368, w, 8, &res_u_a, &res_v_a );
736
            if( res_u_c != res_u_a || res_v_c != res_v_a )
737
            {
738
                ok = 0;
739
                fprintf( stderr, "ssd_nv12: %"PRIu64",%"PRIu64" != %"PRIu64",%"PRIu64"\n",
740
                         res_u_c, res_v_c, res_u_a, res_v_a );
741
            }
742
        }
743
        call_c( pixel_c.ssd_nv12_core,   pbuf1, (intptr_t)368, pbuf2, (intptr_t)368, 360, 8, &res_u_c, &res_v_c );
744
        call_a( pixel_asm.ssd_nv12_core, pbuf1, (intptr_t)368, pbuf2, (intptr_t)368, 360, 8, &res_u_a, &res_v_a );
745
    }
746
    report( "ssd_nv12 :" );
747

748
    if( pixel_asm.ssim_4x4x2_core != pixel_ref.ssim_4x4x2_core ||
749
        pixel_asm.ssim_end4 != pixel_ref.ssim_end4 )
750
    {
751
        int cnt;
752
        float res_c, res_a;
753
        ALIGNED_16( int sums[5][4] ) = {{0}};
754
        used_asm = ok = 1;
755
        x264_emms();
756
        res_c = x264_pixel_ssim_wxh( &pixel_c,   pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
757
        res_a = x264_pixel_ssim_wxh( &pixel_asm, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
758
        if( fabs( res_c - res_a ) > 1e-6 )
759
        {
760
            ok = 0;
761
            fprintf( stderr, "ssim: %.7f != %.7f [FAILED]\n", res_c, res_a );
762
        }
763
        set_func_name( "ssim_core" );
764
        call_c( pixel_c.ssim_4x4x2_core,   pbuf1+2, (intptr_t)32, pbuf2+2, (intptr_t)32, sums );
765
        call_a( pixel_asm.ssim_4x4x2_core, pbuf1+2, (intptr_t)32, pbuf2+2, (intptr_t)32, sums );
766
        set_func_name( "ssim_end" );
767
        call_c2( pixel_c.ssim_end4,   sums, sums, 4 );
768
        call_a2( pixel_asm.ssim_end4, sums, sums, 4 );
769
        /* check incorrect assumptions that 32-bit ints are zero-extended to 64-bit */
770
        call_c1( pixel_c.ssim_end4,   sums, sums, 3 );
771
        call_a1( pixel_asm.ssim_end4, sums, sums, 3 );
772
        report( "ssim :" );
773
    }
774

775
    ok = 1; used_asm = 0;
776
    for( int i = 0; i < 32; i++ )
777
        cost_mv[i] = i*10;
778
    for( int i = 0; i < 100 && ok; i++ )
779
        if( pixel_asm.ads[i&3] != pixel_ref.ads[i&3] )
780
        {
781
            ALIGNED_16( uint16_t sums[72] );
782
            ALIGNED_16( int dc[4] );
783
            ALIGNED_16( int16_t mvs_a[48] );
784
            ALIGNED_16( int16_t mvs_c[48] );
785
            int mvn_a, mvn_c;
786
            int thresh = rand() & 0x3fff;
787
            set_func_name( "esa_ads" );
788
            for( int j = 0; j < 72; j++ )
789
                sums[j] = rand() & 0x3fff;
790
            for( int j = 0; j < 4; j++ )
791
                dc[j] = rand() & 0x3fff;
792
            used_asm = 1;
793
            mvn_c = call_c( pixel_c.ads[i&3], dc, sums, 32, cost_mv, mvs_c, 28, thresh );
794
            mvn_a = call_a( pixel_asm.ads[i&3], dc, sums, 32, cost_mv, mvs_a, 28, thresh );
795
            if( mvn_c != mvn_a || memcmp( mvs_c, mvs_a, mvn_c*sizeof(*mvs_c) ) )
796
            {
797
                ok = 0;
798
                printf( "c%d: ", i&3 );
799
                for( int j = 0; j < mvn_c; j++ )
800
                    printf( "%d ", mvs_c[j] );
801
                printf( "\na%d: ", i&3 );
802
                for( int j = 0; j < mvn_a; j++ )
803
                    printf( "%d ", mvs_a[j] );
804
                printf( "\n\n" );
805
            }
806
        }
807
    report( "esa ads:" );
808

809
    return ret;
810
}
811

812
static int check_dct( int cpu_ref, int cpu_new )
813
{
814
    x264_dct_function_t dct_c;
815
    x264_dct_function_t dct_ref;
816
    x264_dct_function_t dct_asm;
817
    x264_quant_function_t qf;
818
    int ret = 0, ok, used_asm, interlace = 0;
819
    ALIGNED_ARRAY_N( dctcoef, dct1, [16],[16] );
820
    ALIGNED_ARRAY_N( dctcoef, dct2, [16],[16] );
821
    ALIGNED_ARRAY_N( dctcoef, dct4, [16],[16] );
822
    ALIGNED_ARRAY_N( dctcoef, dct8, [4],[64] );
823
    ALIGNED_16( dctcoef dctdc[2][8] );
824
    x264_t h_buf;
825
    x264_t *h = &h_buf;
826

827
    x264_dct_init( 0, &dct_c );
828
    x264_dct_init( cpu_ref, &dct_ref);
829
    x264_dct_init( cpu_new, &dct_asm );
830

831
    memset( h, 0, sizeof(*h) );
832
    x264_param_default( &h->param );
833
    h->sps->i_chroma_format_idc = 1;
834
    h->chroma_qp_table = i_chroma_qp_table + 12;
835
    h->param.analyse.i_luma_deadzone[0] = 0;
836
    h->param.analyse.i_luma_deadzone[1] = 0;
837
    h->param.analyse.b_transform_8x8 = 1;
838
    for( int i = 0; i < 6; i++ )
839
        h->pps->scaling_list[i] = x264_cqm_flat16;
840
    x264_cqm_init( h );
841
    x264_quant_init( h, 0, &qf );
842

843
    /* overflow test cases */
844
    for( int i = 0; i < 5; i++ )
845
    {
846
        pixel *enc = &pbuf3[16*i*FENC_STRIDE];
847
        pixel *dec = &pbuf4[16*i*FDEC_STRIDE];
848

849
        for( int j = 0; j < 16; j++ )
850
        {
851
            int cond_a = (i < 2) ? 1 : ((j&3) == 0 || (j&3) == (i-1));
852
            int cond_b = (i == 0) ? 1 : !cond_a;
853
            enc[0] = enc[1] = enc[4] = enc[5] = enc[8] = enc[9] = enc[12] = enc[13] = cond_a ? PIXEL_MAX : 0;
854
            enc[2] = enc[3] = enc[6] = enc[7] = enc[10] = enc[11] = enc[14] = enc[15] = cond_b ? PIXEL_MAX : 0;
855

856
            for( int k = 0; k < 4; k++ )
857
                dec[k] = PIXEL_MAX - enc[k];
858

859
            enc += FENC_STRIDE;
860
            dec += FDEC_STRIDE;
861
        }
862
    }
863

864
#define TEST_DCT( name, t1, t2, size ) \
865
    if( dct_asm.name != dct_ref.name ) \
866
    { \
867
        set_func_name( #name ); \
868
        used_asm = 1; \
869
        pixel *enc = pbuf3; \
870
        pixel *dec = pbuf4; \
871
        for( int j = 0; j < 5; j++) \
872
        { \
873
            call_c( dct_c.name, t1, &pbuf1[j*64], &pbuf2[j*64] ); \
874
            call_a( dct_asm.name, t2, &pbuf1[j*64], &pbuf2[j*64] ); \
875
            if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
876
            { \
877
                ok = 0; \
878
                fprintf( stderr, #name " [FAILED]\n" ); \
879
                for( int k = 0; k < size; k++ )\
880
                    printf( "%d ", ((dctcoef*)t1)[k] );\
881
                printf("\n");\
882
                for( int k = 0; k < size; k++ )\
883
                    printf( "%d ", ((dctcoef*)t2)[k] );\
884
                printf("\n");\
885
                break; \
886
            } \
887
            call_c( dct_c.name, t1, enc, dec ); \
888
            call_a( dct_asm.name, t2, enc, dec ); \
889
            if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
890
            { \
891
                ok = 0; \
892
                fprintf( stderr, #name " [FAILED] (overflow)\n" ); \
893
                break; \
894
            } \
895
            enc += 16*FENC_STRIDE; \
896
            dec += 16*FDEC_STRIDE; \
897
        } \
898
    }
899
    ok = 1; used_asm = 0;
900
    TEST_DCT( sub4x4_dct, dct1[0], dct2[0], 16 );
901
    TEST_DCT( sub8x8_dct, dct1, dct2, 16*4 );
902
    TEST_DCT( sub8x8_dct_dc, dctdc[0], dctdc[1], 4 );
903
    TEST_DCT( sub8x16_dct_dc, dctdc[0], dctdc[1], 8 );
904
    TEST_DCT( sub16x16_dct, dct1, dct2, 16*16 );
905
    report( "sub_dct4 :" );
906

907
    ok = 1; used_asm = 0;
908
    TEST_DCT( sub8x8_dct8, (void*)dct1[0], (void*)dct2[0], 64 );
909
    TEST_DCT( sub16x16_dct8, (void*)dct1, (void*)dct2, 64*4 );
910
    report( "sub_dct8 :" );
911
#undef TEST_DCT
912

913
    // fdct and idct are denormalized by different factors, so quant/dequant
914
    // is needed to force the coefs into the right range.
915
    dct_c.sub16x16_dct( dct4, pbuf1, pbuf2 );
916
    dct_c.sub16x16_dct8( dct8, pbuf1, pbuf2 );
917
    for( int i = 0; i < 16; i++ )
918
    {
919
        qf.quant_4x4( dct4[i], h->quant4_mf[CQM_4IY][20], h->quant4_bias[CQM_4IY][20] );
920
        qf.dequant_4x4( dct4[i], h->dequant4_mf[CQM_4IY], 20 );
921
    }
922
    for( int i = 0; i < 4; i++ )
923
    {
924
        qf.quant_8x8( dct8[i], h->quant8_mf[CQM_8IY][20], h->quant8_bias[CQM_8IY][20] );
925
        qf.dequant_8x8( dct8[i], h->dequant8_mf[CQM_8IY], 20 );
926
    }
927
    x264_cqm_delete( h );
928

929
#define TEST_IDCT( name, src ) \
930
    if( dct_asm.name != dct_ref.name ) \
931
    { \
932
        set_func_name( #name ); \
933
        used_asm = 1; \
934
        memcpy( pbuf3, pbuf1, 32*32 * sizeof(pixel) ); \
935
        memcpy( pbuf4, pbuf1, 32*32 * sizeof(pixel) ); \
936
        memcpy( dct1, src, 256 * sizeof(dctcoef) ); \
937
        memcpy( dct2, src, 256 * sizeof(dctcoef) ); \
938
        call_c1( dct_c.name, pbuf3, (void*)dct1 ); \
939
        call_a1( dct_asm.name, pbuf4, (void*)dct2 ); \
940
        if( memcmp( pbuf3, pbuf4, 32*32 * sizeof(pixel) ) ) \
941
        { \
942
            ok = 0; \
943
            fprintf( stderr, #name " [FAILED]\n" ); \
944
        } \
945
        call_c2( dct_c.name, pbuf3, (void*)dct1 ); \
946
        call_a2( dct_asm.name, pbuf4, (void*)dct2 ); \
947
    }
948
    ok = 1; used_asm = 0;
949
    TEST_IDCT( add4x4_idct, dct4 );
950
    TEST_IDCT( add8x8_idct, dct4 );
951
    TEST_IDCT( add8x8_idct_dc, dct4 );
952
    TEST_IDCT( add16x16_idct, dct4 );
953
    TEST_IDCT( add16x16_idct_dc, dct4 );
954
    report( "add_idct4 :" );
955

956
    ok = 1; used_asm = 0;
957
    TEST_IDCT( add8x8_idct8, dct8 );
958
    TEST_IDCT( add16x16_idct8, dct8 );
959
    report( "add_idct8 :" );
960
#undef TEST_IDCT
961

962
#define TEST_DCTDC( name )\
963
    ok = 1; used_asm = 0;\
964
    if( dct_asm.name != dct_ref.name )\
965
    {\
966
        set_func_name( #name );\
967
        used_asm = 1;\
968
        uint16_t *p = (uint16_t*)buf1;\
969
        for( int i = 0; i < 16 && ok; i++ )\
970
        {\
971
            for( int j = 0; j < 16; j++ )\
972
                dct1[0][j] = !i ? (j^j>>1^j>>2^j>>3)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
973
                           : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
974
                           : ((*p++)&0x1fff)-0x1000; /* general case */\
975
            memcpy( dct2, dct1, 16 * sizeof(dctcoef) );\
976
            call_c1( dct_c.name, dct1[0] );\
977
            call_a1( dct_asm.name, dct2[0] );\
978
            if( memcmp( dct1, dct2, 16 * sizeof(dctcoef) ) )\
979
                ok = 0;\
980
        }\
981
        call_c2( dct_c.name, dct1[0] );\
982
        call_a2( dct_asm.name, dct2[0] );\
983
    }\
984
    report( #name " :" );
985

986
    TEST_DCTDC(  dct4x4dc );
987
    TEST_DCTDC( idct4x4dc );
988
#undef TEST_DCTDC
989

990
#define TEST_DCTDC_CHROMA( name )\
991
    ok = 1; used_asm = 0;\
992
    if( dct_asm.name != dct_ref.name )\
993
    {\
994
        set_func_name( #name );\
995
        used_asm = 1;\
996
        uint16_t *p = (uint16_t*)buf1;\
997
        for( int i = 0; i < 16 && ok; i++ )\
998
        {\
999
            for( int j = 0; j < 8; j++ )\
1000
                dct1[j][0] = !i ? (j^j>>1^j>>2)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
1001
                           : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
1002
                           : ((*p++)&0x1fff)-0x1000; /* general case */\
1003
            memcpy( dct2, dct1, 8*16 * sizeof(dctcoef) );\
1004
            call_c1( dct_c.name, dctdc[0], dct1 );\
1005
            call_a1( dct_asm.name, dctdc[1], dct2 );\
1006
            if( memcmp( dctdc[0], dctdc[1], 8 * sizeof(dctcoef) ) || memcmp( dct1, dct2, 8*16 * sizeof(dctcoef) ) )\
1007
            {\
1008
                ok = 0;\
1009
                fprintf( stderr, #name " [FAILED]\n" ); \
1010
            }\
1011
        }\
1012
        call_c2( dct_c.name, dctdc[0], dct1 );\
1013
        call_a2( dct_asm.name, dctdc[1], dct2 );\
1014
    }\
1015
    report( #name " :" );
1016

1017
    TEST_DCTDC_CHROMA( dct2x4dc );
1018
#undef TEST_DCTDC_CHROMA
1019

1020
    x264_zigzag_function_t zigzag_c[2];
1021
    x264_zigzag_function_t zigzag_ref[2];
1022
    x264_zigzag_function_t zigzag_asm[2];
1023

1024
    ALIGNED_16( dctcoef level1[64] );
1025
    ALIGNED_16( dctcoef level2[64] );
1026

1027
#define TEST_ZIGZAG_SCAN( name, t1, t2, dct, size ) \
1028
    if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1029
    { \
1030
        set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1031
        used_asm = 1; \
1032
        for( int i = 0; i < size*size; i++ ) \
1033
            dct[i] = i; \
1034
        call_c( zigzag_c[interlace].name, t1, dct ); \
1035
        call_a( zigzag_asm[interlace].name, t2, dct ); \
1036
        if( memcmp( t1, t2, size*size*sizeof(dctcoef) ) ) \
1037
        { \
1038
            ok = 0; \
1039
            for( int i = 0; i < 2; i++ ) \
1040
            { \
1041
                dctcoef *d = (dctcoef*)(i ? t2 : t1); \
1042
                for( int j = 0; j < size; j++ ) \
1043
                { \
1044
                    for( int k = 0; k < size; k++ ) \
1045
                        fprintf( stderr, "%2d ", d[k+j*8] ); \
1046
                    fprintf( stderr, "\n" ); \
1047
                } \
1048
                fprintf( stderr, "\n" ); \
1049
            } \
1050
            fprintf( stderr, #name " [FAILED]\n" ); \
1051
        } \
1052
    }
1053

1054
#define TEST_ZIGZAG_SUB( name, t1, t2, size ) \
1055
    if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1056
    { \
1057
        int nz_a, nz_c; \
1058
        set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1059
        used_asm = 1; \
1060
        memcpy( pbuf3, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
1061
        memcpy( pbuf4, pbuf1, 16*FDEC_STRIDE * sizeof(pixel) ); \
1062
        nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
1063
        nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
1064
        if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( pbuf3, pbuf4, 16*FDEC_STRIDE*sizeof(pixel) ) || nz_c != nz_a ) \
1065
        { \
1066
            ok = 0; \
1067
            fprintf( stderr, #name " [FAILED]\n" ); \
1068
        } \
1069
        call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3 ); \
1070
        call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4 ); \
1071
    }
1072

1073
#define TEST_ZIGZAG_SUBAC( name, t1, t2 ) \
1074
    if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1075
    { \
1076
        int nz_a, nz_c; \
1077
        dctcoef dc_a, dc_c; \
1078
        set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1079
        used_asm = 1; \
1080
        for( int i = 0; i < 2; i++ ) \
1081
        { \
1082
            memcpy( pbuf3, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
1083
            memcpy( pbuf4, pbuf2, 16*FDEC_STRIDE * sizeof(pixel) ); \
1084
            for( int j = 0; j < 4; j++ ) \
1085
            { \
1086
                memcpy( pbuf3 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
1087
                memcpy( pbuf4 + j*FDEC_STRIDE, (i?pbuf1:pbuf2) + j*FENC_STRIDE, 4 * sizeof(pixel) ); \
1088
            } \
1089
            nz_c = call_c1( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
1090
            nz_a = call_a1( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
1091
            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 ) \
1092
            { \
1093
                ok = 0; \
1094
                fprintf( stderr, #name " [FAILED]\n" ); \
1095
                break; \
1096
            } \
1097
        } \
1098
        call_c2( zigzag_c[interlace].name, t1, pbuf2, pbuf3, &dc_c ); \
1099
        call_a2( zigzag_asm[interlace].name, t2, pbuf2, pbuf4, &dc_a ); \
1100
    }
1101

1102
#define TEST_INTERLEAVE( name, t1, t2, dct, size ) \
1103
    if( zigzag_asm[interlace].name != zigzag_ref[interlace].name ) \
1104
    { \
1105
        for( int j = 0; j < 100; j++ ) \
1106
        { \
1107
            set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
1108
            used_asm = 1; \
1109
            memcpy(dct, buf1, size*sizeof(dctcoef)); \
1110
            for( int i = 0; i < size; i++ ) \
1111
                dct[i] = rand()&0x1F ? 0 : dct[i]; \
1112
            memcpy(buf3, buf4, 10); \
1113
            call_c( zigzag_c[interlace].name, t1, dct, buf3 ); \
1114
            call_a( zigzag_asm[interlace].name, t2, dct, buf4 ); \
1115
            if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( buf3, buf4, 10 ) ) \
1116
            { \
1117
                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;\
1118
            } \
1119
        } \
1120
    }
1121

1122
    x264_zigzag_init( 0, &zigzag_c[0], &zigzag_c[1] );
1123
    x264_zigzag_init( cpu_ref, &zigzag_ref[0], &zigzag_ref[1] );
1124
    x264_zigzag_init( cpu_new, &zigzag_asm[0], &zigzag_asm[1] );
1125

1126
    ok = 1; used_asm = 0;
1127
    TEST_INTERLEAVE( interleave_8x8_cavlc, level1, level2, dct8[0], 64 );
1128
    report( "zigzag_interleave :" );
1129

1130
    for( interlace = 0; interlace <= 1; interlace++ )
1131
    {
1132
        ok = 1; used_asm = 0;
1133
        TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, dct8[0], 8 );
1134
        TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 4 );
1135
        TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
1136
        TEST_ZIGZAG_SUB( sub_8x8, level1, level2, 64 );
1137
        TEST_ZIGZAG_SUBAC( sub_4x4ac, level1, level2 );
1138
        report( interlace ? "zigzag_field :" : "zigzag_frame :" );
1139
    }
1140
#undef TEST_ZIGZAG_SCAN
1141
#undef TEST_ZIGZAG_SUB
1142

1143
    return ret;
1144
}
1145

1146
static int check_mc( int cpu_ref, int cpu_new )
1147
{
1148
    x264_mc_functions_t mc_c;
1149
    x264_mc_functions_t mc_ref;
1150
    x264_mc_functions_t mc_a;
1151
    x264_pixel_function_t pixf;
1152

1153
    pixel *src     = &(pbuf1)[2*64+2];
1154
    pixel *src2[4] = { &(pbuf1)[3*64+2], &(pbuf1)[5*64+2],
1155
                       &(pbuf1)[7*64+2], &(pbuf1)[9*64+2] };
1156
    pixel *dst1    = pbuf3;
1157
    pixel *dst2    = pbuf4;
1158

1159
    int ret = 0, ok, used_asm;
1160

1161
    x264_mc_init( 0, &mc_c, 0 );
1162
    x264_mc_init( cpu_ref, &mc_ref, 0 );
1163
    x264_mc_init( cpu_new, &mc_a, 0 );
1164
    x264_pixel_init( 0, &pixf );
1165

1166
#define MC_TEST_LUMA( w, h ) \
1167
        if( mc_a.mc_luma != mc_ref.mc_luma && !(w&(w-1)) && h<=16 ) \
1168
        { \
1169
            const x264_weight_t *weight = x264_weight_none; \
1170
            set_func_name( "mc_luma_%dx%d", w, h ); \
1171
            used_asm = 1; \
1172
            for( int i = 0; i < 1024; i++ ) \
1173
                pbuf3[i] = pbuf4[i] = 0xCD; \
1174
            call_c( mc_c.mc_luma, dst1, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1175
            call_a( mc_a.mc_luma, dst2, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1176
            if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1177
            { \
1178
                fprintf( stderr, "mc_luma[mv(%d,%d) %2dx%-2d]     [FAILED]\n", dx, dy, w, h ); \
1179
                ok = 0; \
1180
            } \
1181
        } \
1182
        if( mc_a.get_ref != mc_ref.get_ref ) \
1183
        { \
1184
            pixel *ref = dst2; \
1185
            intptr_t ref_stride = 32; \
1186
            int w_checked = ( ( sizeof(pixel) == 2 && (w == 12 || w == 20)) ? w-2 : w ); \
1187
            const x264_weight_t *weight = x264_weight_none; \
1188
            set_func_name( "get_ref_%dx%d", w_checked, h ); \
1189
            used_asm = 1; \
1190
            for( int i = 0; i < 1024; i++ ) \
1191
                pbuf3[i] = pbuf4[i] = 0xCD; \
1192
            call_c( mc_c.mc_luma, dst1, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1193
            ref = (pixel*)call_a( mc_a.get_ref, ref, &ref_stride, src2, (intptr_t)64, dx, dy, w, h, weight ); \
1194
            for( int i = 0; i < h; i++ ) \
1195
                if( memcmp( dst1+i*32, ref+i*ref_stride, w_checked * sizeof(pixel) ) ) \
1196
                { \
1197
                    fprintf( stderr, "get_ref[mv(%d,%d) %2dx%-2d]     [FAILED]\n", dx, dy, w_checked, h ); \
1198
                    ok = 0; \
1199
                    break; \
1200
                } \
1201
        }
1202

1203
#define MC_TEST_CHROMA( w, h ) \
1204
        if( mc_a.mc_chroma != mc_ref.mc_chroma ) \
1205
        { \
1206
            set_func_name( "mc_chroma_%dx%d", w, h ); \
1207
            used_asm = 1; \
1208
            for( int i = 0; i < 1024; i++ ) \
1209
                pbuf3[i] = pbuf4[i] = 0xCD; \
1210
            call_c( mc_c.mc_chroma, dst1, dst1+8, (intptr_t)16, src, (intptr_t)64, dx, dy, w, h ); \
1211
            call_a( mc_a.mc_chroma, dst2, dst2+8, (intptr_t)16, src, (intptr_t)64, dx, dy, w, h ); \
1212
            /* mc_chroma width=2 may write garbage to the right of dst. ignore that. */ \
1213
            for( int j = 0; j < h; j++ ) \
1214
                for( int i = w; i < 8; i++ ) \
1215
                { \
1216
                    dst2[i+j*16+8] = dst1[i+j*16+8]; \
1217
                    dst2[i+j*16  ] = dst1[i+j*16  ]; \
1218
                } \
1219
            if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1220
            { \
1221
                fprintf( stderr, "mc_chroma[mv(%d,%d) %2dx%-2d]     [FAILED]\n", dx, dy, w, h ); \
1222
                ok = 0; \
1223
            } \
1224
        }
1225
    ok = 1; used_asm = 0;
1226
    for( int dy = -8; dy < 8; dy++ )
1227
        for( int dx = -128; dx < 128; dx++ )
1228
        {
1229
            if( rand()&15 ) continue; // running all of them is too slow
1230
            MC_TEST_LUMA( 20, 18 );
1231
            MC_TEST_LUMA( 16, 16 );
1232
            MC_TEST_LUMA( 16, 8 );
1233
            MC_TEST_LUMA( 12, 10 );
1234
            MC_TEST_LUMA( 8, 16 );
1235
            MC_TEST_LUMA( 8, 8 );
1236
            MC_TEST_LUMA( 8, 4 );
1237
            MC_TEST_LUMA( 4, 8 );
1238
            MC_TEST_LUMA( 4, 4 );
1239
        }
1240
    report( "mc luma :" );
1241

1242
    ok = 1; used_asm = 0;
1243
    for( int dy = -1; dy < 9; dy++ )
1244
        for( int dx = -128; dx < 128; dx++ )
1245
        {
1246
            if( rand()&15 ) continue;
1247
            MC_TEST_CHROMA( 8, 8 );
1248
            MC_TEST_CHROMA( 8, 4 );
1249
            MC_TEST_CHROMA( 4, 8 );
1250
            MC_TEST_CHROMA( 4, 4 );
1251
            MC_TEST_CHROMA( 4, 2 );
1252
            MC_TEST_CHROMA( 2, 4 );
1253
            MC_TEST_CHROMA( 2, 2 );
1254
        }
1255
    report( "mc chroma :" );
1256
#undef MC_TEST_LUMA
1257
#undef MC_TEST_CHROMA
1258

1259
#define MC_TEST_AVG( name, weight ) \
1260
{ \
1261
    for( int i = 0; i < 12; i++ ) \
1262
    { \
1263
        memcpy( pbuf3, pbuf1+320, 320 * sizeof(pixel) ); \
1264
        memcpy( pbuf4, pbuf1+320, 320 * sizeof(pixel) ); \
1265
        if( mc_a.name[i] != mc_ref.name[i] ) \
1266
        { \
1267
            set_func_name( "%s_%s", #name, pixel_names[i] ); \
1268
            used_asm = 1; \
1269
            call_c1( mc_c.name[i], pbuf3, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1270
            call_a1( mc_a.name[i], pbuf4, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1271
            if( memcmp( pbuf3, pbuf4, 320 * sizeof(pixel) ) ) \
1272
            { \
1273
                ok = 0; \
1274
                fprintf( stderr, #name "[%d]: [FAILED]\n", i ); \
1275
            } \
1276
            call_c2( mc_c.name[i], pbuf3, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1277
            call_a2( mc_a.name[i], pbuf4, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
1278
        } \
1279
    } \
1280
}
1281

1282
    ok = 1, used_asm = 0;
1283
    for( int w = -63; w <= 127 && ok; w++ )
1284
        MC_TEST_AVG( avg, w );
1285
    report( "mc wpredb :" );
1286

1287
#define MC_TEST_WEIGHT( name, weight, aligned ) \
1288
    int align_off = (aligned ? 0 : rand()%16); \
1289
    for( int i = 1; i <= 5; i++ ) \
1290
    { \
1291
        ALIGNED_16( pixel buffC[640] ); \
1292
        ALIGNED_16( pixel buffA[640] ); \
1293
        int j = X264_MAX( i*4, 2 ); \
1294
        memset( buffC, 0, 640 * sizeof(pixel) ); \
1295
        memset( buffA, 0, 640 * sizeof(pixel) ); \
1296
        x264_t ha; \
1297
        ha.mc = mc_a; \
1298
        /* w12 is the same as w16 in some cases */ \
1299
        if( i == 3 && mc_a.name[i] == mc_a.name[i+1] ) \
1300
            continue; \
1301
        if( mc_a.name[i] != mc_ref.name[i] ) \
1302
        { \
1303
            set_func_name( "%s_w%d", #name, j ); \
1304
            used_asm = 1; \
1305
            call_c1( mc_c.weight[i],     buffC, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1306
            mc_a.weight_cache(&ha, &weight); \
1307
            call_a1( weight.weightfn[i], buffA, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1308
            for( int k = 0; k < 16; k++ ) \
1309
                if( memcmp( &buffC[k*32], &buffA[k*32], j * sizeof(pixel) ) ) \
1310
                { \
1311
                    ok = 0; \
1312
                    fprintf( stderr, #name "[%d]: [FAILED] s:%d o:%d d%d\n", i, s, o, d ); \
1313
                    break; \
1314
                } \
1315
            /* omit unlikely high scales for benchmarking */ \
1316
            if( (s << (8-d)) < 512 ) \
1317
            { \
1318
                call_c2( mc_c.weight[i],     buffC, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1319
                call_a2( weight.weightfn[i], buffA, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
1320
            } \
1321
        } \
1322
    }
1323

1324
    ok = 1; used_asm = 0;
1325

1326
    int align_cnt = 0;
1327
    for( int s = 0; s <= 127 && ok; s++ )
1328
    {
1329
        for( int o = -128; o <= 127 && ok; o++ )
1330
        {
1331
            if( rand() & 2047 ) continue;
1332
            for( int d = 0; d <= 7 && ok; d++ )
1333
            {
1334
                if( s == 1<<d )
1335
                    continue;
1336
                x264_weight_t weight = { .i_scale = s, .i_denom = d, .i_offset = o };
1337
                MC_TEST_WEIGHT( weight, weight, (align_cnt++ % 4) );
1338
            }
1339
        }
1340

1341
    }
1342
    report( "mc weight :" );
1343

1344
    ok = 1; used_asm = 0;
1345
    for( int o = 0; o <= 127 && ok; o++ )
1346
    {
1347
        int s = 1, d = 0;
1348
        if( rand() & 15 ) continue;
1349
        x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1350
        MC_TEST_WEIGHT( offsetadd, weight, (align_cnt++ % 4) );
1351
    }
1352
    report( "mc offsetadd :" );
1353
    ok = 1; used_asm = 0;
1354
    for( int o = -128; o < 0 && ok; o++ )
1355
    {
1356
        int s = 1, d = 0;
1357
        if( rand() & 15 ) continue;
1358
        x264_weight_t weight = { .i_scale = 1, .i_denom = 0, .i_offset = o };
1359
        MC_TEST_WEIGHT( offsetsub, weight, (align_cnt++ % 4) );
1360
    }
1361
    report( "mc offsetsub :" );
1362

1363
    ok = 1; used_asm = 0;
1364
    for( int height = 8; height <= 16; height += 8 )
1365
    {
1366
        if( mc_a.store_interleave_chroma != mc_ref.store_interleave_chroma )
1367
        {
1368
            set_func_name( "store_interleave_chroma" );
1369
            used_asm = 1;
1370
            memset( pbuf3, 0, 64*height );
1371
            memset( pbuf4, 0, 64*height );
1372
            call_c( mc_c.store_interleave_chroma, pbuf3, (intptr_t)64, pbuf1, pbuf1+16, height );
1373
            call_a( mc_a.store_interleave_chroma, pbuf4, (intptr_t)64, pbuf1, pbuf1+16, height );
1374
            if( memcmp( pbuf3, pbuf4, 64*height ) )
1375
            {
1376
                ok = 0;
1377
                fprintf( stderr, "store_interleave_chroma FAILED: h=%d\n", height );
1378
                break;
1379
            }
1380
        }
1381
        if( mc_a.load_deinterleave_chroma_fenc != mc_ref.load_deinterleave_chroma_fenc )
1382
        {
1383
            set_func_name( "load_deinterleave_chroma_fenc" );
1384
            used_asm = 1;
1385
            call_c( mc_c.load_deinterleave_chroma_fenc, pbuf3, pbuf1, (intptr_t)64, height );
1386
            call_a( mc_a.load_deinterleave_chroma_fenc, pbuf4, pbuf1, (intptr_t)64, height );
1387
            if( memcmp( pbuf3, pbuf4, FENC_STRIDE*height ) )
1388
            {
1389
                ok = 0;
1390
                fprintf( stderr, "load_deinterleave_chroma_fenc FAILED: h=%d\n", height );
1391
                break;
1392
            }
1393
        }
1394
        if( mc_a.load_deinterleave_chroma_fdec != mc_ref.load_deinterleave_chroma_fdec )
1395
        {
1396
            set_func_name( "load_deinterleave_chroma_fdec" );
1397
            used_asm = 1;
1398
            call_c( mc_c.load_deinterleave_chroma_fdec, pbuf3, pbuf1, (intptr_t)64, height );
1399
            call_a( mc_a.load_deinterleave_chroma_fdec, pbuf4, pbuf1, (intptr_t)64, height );
1400
            if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*height ) )
1401
            {
1402
                ok = 0;
1403
                fprintf( stderr, "load_deinterleave_chroma_fdec FAILED: h=%d\n", height );
1404
                break;
1405
            }
1406
        }
1407
    }
1408
    report( "store_interleave :" );
1409

1410
    struct plane_spec {
1411
        int w, h, src_stride;
1412
    } 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} };
1413
    ok = 1; used_asm = 0;
1414
    if( mc_a.plane_copy != mc_ref.plane_copy )
1415
    {
1416
        set_func_name( "plane_copy" );
1417
        used_asm = 1;
1418
        for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1419
        {
1420
            int w = plane_specs[i].w;
1421
            int h = plane_specs[i].h;
1422
            intptr_t src_stride = plane_specs[i].src_stride;
1423
            intptr_t dst_stride = (w + 127) & ~63;
1424
            assert( dst_stride * h <= 0x1000 );
1425
            pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1426
            memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1427
            memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1428
            call_c( mc_c.plane_copy, pbuf3, dst_stride, src1, src_stride, w, h );
1429
            call_a( mc_a.plane_copy, pbuf4, dst_stride, src1, src_stride, w, h );
1430
            for( int y = 0; y < h; y++ )
1431
                if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(pixel) ) )
1432
                {
1433
                    ok = 0;
1434
                    fprintf( stderr, "plane_copy FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1435
                    break;
1436
                }
1437
        }
1438
    }
1439

1440
    if( mc_a.plane_copy_swap != mc_ref.plane_copy_swap )
1441
    {
1442
        set_func_name( "plane_copy_swap" );
1443
        used_asm = 1;
1444
        for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1445
        {
1446
            int w = (plane_specs[i].w + 1) >> 1;
1447
            int h = plane_specs[i].h;
1448
            intptr_t src_stride = plane_specs[i].src_stride;
1449
            intptr_t dst_stride = (2*w + 127) & ~63;
1450
            assert( dst_stride * h <= 0x1000 );
1451
            pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1452
            memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1453
            memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1454
            call_c( mc_c.plane_copy_swap, pbuf3, dst_stride, src1, src_stride, w, h );
1455
            call_a( mc_a.plane_copy_swap, pbuf4, dst_stride, src1, src_stride, w, h );
1456
            for( int y = 0; y < h; y++ )
1457
                if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
1458
                {
1459
                    ok = 0;
1460
                    fprintf( stderr, "plane_copy_swap FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1461
                    break;
1462
                }
1463
        }
1464
    }
1465

1466
    if( mc_a.plane_copy_interleave != mc_ref.plane_copy_interleave )
1467
    {
1468
        set_func_name( "plane_copy_interleave" );
1469
        used_asm = 1;
1470
        for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1471
        {
1472
            int w = (plane_specs[i].w + 1) >> 1;
1473
            int h = plane_specs[i].h;
1474
            intptr_t src_stride = (plane_specs[i].src_stride + 1) >> 1;
1475
            intptr_t dst_stride = (2*w + 127) & ~63;
1476
            assert( dst_stride * h <= 0x1000 );
1477
            pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
1478
            memset( pbuf3, 0, 0x1000*sizeof(pixel) );
1479
            memset( pbuf4, 0, 0x1000*sizeof(pixel) );
1480
            call_c( mc_c.plane_copy_interleave, pbuf3, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1481
            call_a( mc_a.plane_copy_interleave, pbuf4, dst_stride, src1, src_stride, src1+1024, src_stride+16, w, h );
1482
            for( int y = 0; y < h; y++ )
1483
                if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
1484
                {
1485
                    ok = 0;
1486
                    fprintf( stderr, "plane_copy_interleave FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1487
                    break;
1488
                }
1489
        }
1490
    }
1491

1492
    if( mc_a.plane_copy_deinterleave != mc_ref.plane_copy_deinterleave )
1493
    {
1494
        set_func_name( "plane_copy_deinterleave" );
1495
        used_asm = 1;
1496
        for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1497
        {
1498
            int w = (plane_specs[i].w + 1) >> 1;
1499
            int h = plane_specs[i].h;
1500
            intptr_t dst_stride = w;
1501
            intptr_t src_stride = (2*w + 127) & ~63;
1502
            intptr_t offv = (dst_stride*h + 31) & ~15;
1503
            memset( pbuf3, 0, 0x1000 );
1504
            memset( pbuf4, 0, 0x1000 );
1505
            call_c( mc_c.plane_copy_deinterleave, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf1, src_stride, w, h );
1506
            call_a( mc_a.plane_copy_deinterleave, pbuf4, dst_stride, pbuf4+offv, dst_stride, pbuf1, src_stride, w, h );
1507
            for( int y = 0; y < h; y++ )
1508
                if( memcmp( pbuf3+y*dst_stride,      pbuf4+y*dst_stride, w ) ||
1509
                    memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w ) )
1510
                {
1511
                    ok = 0;
1512
                    fprintf( stderr, "plane_copy_deinterleave FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1513
                    break;
1514
                }
1515
        }
1516
    }
1517

1518
    if( mc_a.plane_copy_deinterleave_rgb != mc_ref.plane_copy_deinterleave_rgb )
1519
    {
1520
        set_func_name( "plane_copy_deinterleave_rgb" );
1521
        used_asm = 1;
1522
        for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1523
        {
1524
            int w = (plane_specs[i].w + 2) >> 2;
1525
            int h = plane_specs[i].h;
1526
            intptr_t src_stride = plane_specs[i].src_stride;
1527
            intptr_t dst_stride = ALIGN( w, 16 );
1528
            intptr_t offv = dst_stride*h + 16;
1529

1530
            for( int pw = 3; pw <= 4; pw++ )
1531
            {
1532
                memset( pbuf3, 0, 0x1000 );
1533
                memset( pbuf4, 0, 0x1000 );
1534
                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 );
1535
                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 );
1536
                for( int y = 0; y < h; y++ )
1537
                    if( memcmp( pbuf3+y*dst_stride+0*offv, pbuf4+y*dst_stride+0*offv, w ) ||
1538
                        memcmp( pbuf3+y*dst_stride+1*offv, pbuf4+y*dst_stride+1*offv, w ) ||
1539
                        memcmp( pbuf3+y*dst_stride+2*offv, pbuf4+y*dst_stride+2*offv, w ) )
1540
                    {
1541
                        ok = 0;
1542
                        fprintf( stderr, "plane_copy_deinterleave_rgb FAILED: w=%d h=%d stride=%d pw=%d\n", w, h, (int)src_stride, pw );
1543
                        break;
1544
                    }
1545
            }
1546
        }
1547
    }
1548
    report( "plane_copy :" );
1549

1550
    if( mc_a.plane_copy_deinterleave_v210 != mc_ref.plane_copy_deinterleave_v210 )
1551
    {
1552
        set_func_name( "plane_copy_deinterleave_v210" );
1553
        ok = 1; used_asm = 1;
1554
        for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
1555
        {
1556
            int w = (plane_specs[i].w + 1) >> 1;
1557
            int h = plane_specs[i].h;
1558
            intptr_t dst_stride = ALIGN( w, 16 );
1559
            intptr_t src_stride = (w + 47) / 48 * 128 / sizeof(uint32_t);
1560
            intptr_t offv = dst_stride*h + 32;
1561
            memset( pbuf3, 0, 0x1000 );
1562
            memset( pbuf4, 0, 0x1000 );
1563
            call_c( mc_c.plane_copy_deinterleave_v210, pbuf3, dst_stride, pbuf3+offv, dst_stride, (uint32_t *)buf1, src_stride, w, h );
1564
            call_a( mc_a.plane_copy_deinterleave_v210, pbuf4, dst_stride, pbuf4+offv, dst_stride, (uint32_t *)buf1, src_stride, w, h );
1565
            for( int y = 0; y < h; y++ )
1566
                if( memcmp( pbuf3+y*dst_stride,      pbuf4+y*dst_stride,      w*sizeof(uint16_t) ) ||
1567
                    memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w*sizeof(uint16_t) ) )
1568
                {
1569
                    ok = 0;
1570
                    fprintf( stderr, "plane_copy_deinterleave_v210 FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
1571
                    break;
1572
                }
1573
        }
1574
        report( "v210 :" );
1575
    }
1576

1577
    if( mc_a.hpel_filter != mc_ref.hpel_filter )
1578
    {
1579
        pixel *srchpel = pbuf1+8+2*64;
1580
        pixel *dstc[3] = { pbuf3+8, pbuf3+8+16*64, pbuf3+8+32*64 };
1581
        pixel *dsta[3] = { pbuf4+8, pbuf4+8+16*64, pbuf4+8+32*64 };
1582
        void *tmp = pbuf3+49*64;
1583
        set_func_name( "hpel_filter" );
1584
        ok = 1; used_asm = 1;
1585
        memset( pbuf3, 0, 4096 * sizeof(pixel) );
1586
        memset( pbuf4, 0, 4096 * sizeof(pixel) );
1587
        call_c( mc_c.hpel_filter, dstc[0], dstc[1], dstc[2], srchpel, (intptr_t)64, 48, 10, tmp );
1588
        call_a( mc_a.hpel_filter, dsta[0], dsta[1], dsta[2], srchpel, (intptr_t)64, 48, 10, tmp );
1589
        for( int i = 0; i < 3; i++ )
1590
            for( int j = 0; j < 10; j++ )
1591
                //FIXME ideally the first pixels would match too, but they aren't actually used
1592
                if( memcmp( dstc[i]+j*64+2, dsta[i]+j*64+2, 43 * sizeof(pixel) ) )
1593
                {
1594
                    ok = 0;
1595
                    fprintf( stderr, "hpel filter differs at plane %c line %d\n", "hvc"[i], j );
1596
                    for( int k = 0; k < 48; k++ )
1597
                        printf( "%02x%s", dstc[i][j*64+k], (k+1)&3 ? "" : " " );
1598
                    printf( "\n" );
1599
                    for( int k = 0; k < 48; k++ )
1600
                        printf( "%02x%s", dsta[i][j*64+k], (k+1)&3 ? "" : " " );
1601
                    printf( "\n" );
1602
                    break;
1603
                }
1604
        report( "hpel filter :" );
1605
    }
1606

1607
    if( mc_a.frame_init_lowres_core != mc_ref.frame_init_lowres_core )
1608
    {
1609
        pixel *dstc[4] = { pbuf3, pbuf3+1024, pbuf3+2048, pbuf3+3072 };
1610
        pixel *dsta[4] = { pbuf4, pbuf4+1024, pbuf4+2048, pbuf4+3072 };
1611
        set_func_name( "lowres_init" );
1612
        ok = 1; used_asm = 1;
1613
        for( int w = 96; w <= 96+24; w += 8 )
1614
        {
1615
            intptr_t stride = (w*2+31)&~31;
1616
            intptr_t stride_lowres = (w+31)&~31;
1617
            call_c( mc_c.frame_init_lowres_core, pbuf1, dstc[0], dstc[1], dstc[2], dstc[3], stride, stride_lowres, w, 8 );
1618
            call_a( mc_a.frame_init_lowres_core, pbuf1, dsta[0], dsta[1], dsta[2], dsta[3], stride, stride_lowres, w, 8 );
1619
            for( int i = 0; i < 8; i++ )
1620
            {
1621
                for( int j = 0; j < 4; j++ )
1622
                    if( memcmp( dstc[j]+i*stride_lowres, dsta[j]+i*stride_lowres, w * sizeof(pixel) ) )
1623
                    {
1624
                        ok = 0;
1625
                        fprintf( stderr, "frame_init_lowres differs at plane %d line %d\n", j, i );
1626
                        for( int k = 0; k < w; k++ )
1627
                            printf( "%d ", dstc[j][k+i*stride_lowres] );
1628
                        printf( "\n" );
1629
                        for( int k = 0; k < w; k++ )
1630
                            printf( "%d ", dsta[j][k+i*stride_lowres] );
1631
                        printf( "\n" );
1632
                        break;
1633
                    }
1634
            }
1635
        }
1636
        report( "lowres init :" );
1637
    }
1638

1639
#define INTEGRAL_INIT( name, size, offset, cmp_len, ... )\
1640
    if( mc_a.name != mc_ref.name )\
1641
    {\
1642
        intptr_t stride = 96;\
1643
        set_func_name( #name );\
1644
        used_asm = 1;\
1645
        memcpy( buf3, buf1, size*2*stride );\
1646
        memcpy( buf4, buf1, size*2*stride );\
1647
        uint16_t *sum = (uint16_t*)buf3;\
1648
        call_c1( mc_c.name, sum+offset, __VA_ARGS__ );\
1649
        sum = (uint16_t*)buf4;\
1650
        call_a1( mc_a.name, sum+offset, __VA_ARGS__ );\
1651
        if( memcmp( buf3+2*offset, buf4+2*offset, cmp_len*2 )\
1652
            || (size>9 && memcmp( buf3+18*stride, buf4+18*stride, (stride-8)*2 )))\
1653
            ok = 0;\
1654
        call_c2( mc_c.name, sum+offset, __VA_ARGS__ );\
1655
        call_a2( mc_a.name, sum+offset, __VA_ARGS__ );\
1656
    }
1657
    ok = 1; used_asm = 0;
1658
    INTEGRAL_INIT( integral_init4h, 2, stride, stride-4, pbuf2, stride );
1659
    INTEGRAL_INIT( integral_init8h, 2, stride, stride-8, pbuf2, stride );
1660
    INTEGRAL_INIT( integral_init4v, 14, 0, stride-8, sum+9*stride, stride );
1661
    INTEGRAL_INIT( integral_init8v, 9, 0, stride-8, stride );
1662
    report( "integral init :" );
1663

1664
    ok = 1; used_asm = 0;
1665
    if( mc_a.mbtree_propagate_cost != mc_ref.mbtree_propagate_cost )
1666
    {
1667
        used_asm = 1;
1668
        x264_emms();
1669
        for( int i = 0; i < 10; i++ )
1670
        {
1671
            float fps_factor = (rand()&65535) / 65535.0f;
1672
            set_func_name( "mbtree_propagate_cost" );
1673
            int16_t *dsta = (int16_t*)buf3;
1674
            int16_t *dstc = dsta+400;
1675
            uint16_t *prop = (uint16_t*)buf1;
1676
            uint16_t *intra = (uint16_t*)buf4;
1677
            uint16_t *inter = intra+128;
1678
            uint16_t *qscale = inter+128;
1679
            uint16_t *rnd = (uint16_t*)buf2;
1680
            x264_emms();
1681
            for( int j = 0; j < 100; j++ )
1682
            {
1683
                intra[j]  = *rnd++ & 0x7fff;
1684
                intra[j] += !intra[j];
1685
                inter[j]  = *rnd++ & 0x7fff;
1686
                qscale[j] = *rnd++ & 0x7fff;
1687
            }
1688
            call_c( mc_c.mbtree_propagate_cost, dstc, prop, intra, inter, qscale, &fps_factor, 100 );
1689
            call_a( mc_a.mbtree_propagate_cost, dsta, prop, intra, inter, qscale, &fps_factor, 100 );
1690
            // I don't care about exact rounding, this is just how close the floating-point implementation happens to be
1691
            x264_emms();
1692
            for( int j = 0; j < 100 && ok; j++ )
1693
            {
1694
                ok &= abs( dstc[j]-dsta[j] ) <= 1 || fabs( (double)dstc[j]/dsta[j]-1 ) < 1e-4;
1695
                if( !ok )
1696
                    fprintf( stderr, "mbtree_propagate_cost FAILED: %f !~= %f\n", (double)dstc[j], (double)dsta[j] );
1697
            }
1698
        }
1699
    }
1700

1701
    if( mc_a.mbtree_propagate_list != mc_ref.mbtree_propagate_list )
1702
    {
1703
        used_asm = 1;
1704
        for( int i = 0; i < 8; i++ )
1705
        {
1706
            set_func_name( "mbtree_propagate_list" );
1707
            x264_t h;
1708
            int height = 4;
1709
            int width = 128;
1710
            int size = width*height;
1711
            h.mb.i_mb_stride = width;
1712
            h.mb.i_mb_width = width;
1713
            h.mb.i_mb_height = height;
1714

1715
            uint16_t *ref_costsc = (uint16_t*)buf3;
1716
            uint16_t *ref_costsa = (uint16_t*)buf4;
1717
            int16_t (*mvs)[2] = (int16_t(*)[2])(ref_costsc + size);
1718
            int16_t *propagate_amount = (int16_t*)(mvs + width);
1719
            uint16_t *lowres_costs = (uint16_t*)(propagate_amount + width);
1720
            h.scratch_buffer2 = (uint8_t*)(ref_costsa + size);
1721
            int bipred_weight = (rand()%63)+1;
1722
            int list = i&1;
1723
            for( int j = 0; j < size; j++ )
1724
                ref_costsc[j] = ref_costsa[j] = rand()&32767;
1725
            for( int j = 0; j < width; j++ )
1726
            {
1727
                static const uint8_t list_dist[2][8] = {{0,1,1,1,1,1,1,1},{1,1,3,3,3,3,3,2}};
1728
                for( int k = 0; k < 2; k++ )
1729
                    mvs[j][k] = (rand()&127) - 64;
1730
                propagate_amount[j] = rand()&32767;
1731
                lowres_costs[j] = list_dist[list][rand()&7] << LOWRES_COST_SHIFT;
1732
            }
1733

1734
            call_c1( mc_c.mbtree_propagate_list, &h, ref_costsc, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
1735
            call_a1( mc_a.mbtree_propagate_list, &h, ref_costsa, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
1736

1737
            for( int j = 0; j < size && ok; j++ )
1738
            {
1739
                ok &= abs(ref_costsa[j] - ref_costsc[j]) <= 1;
1740
                if( !ok )
1741
                    fprintf( stderr, "mbtree_propagate_list FAILED at %d: %d !~= %d\n", j, ref_costsc[j], ref_costsa[j] );
1742
            }
1743

1744
            call_c2( mc_c.mbtree_propagate_list, &h, ref_costsc, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
1745
            call_a2( mc_a.mbtree_propagate_list, &h, ref_costsa, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
1746
        }
1747
    }
1748
    report( "mbtree :" );
1749

1750
    if( mc_a.memcpy_aligned != mc_ref.memcpy_aligned )
1751
    {
1752
        set_func_name( "memcpy_aligned" );
1753
        ok = 1; used_asm = 1;
1754
        for( size_t size = 16; size < 256; size += 16 )
1755
        {
1756
            memset( buf4, 0xAA, size + 1 );
1757
            call_c( mc_c.memcpy_aligned, buf3, buf1, size );
1758
            call_a( mc_a.memcpy_aligned, buf4, buf1, size );
1759
            if( memcmp( buf3, buf4, size ) || buf4[size] != 0xAA )
1760
            {
1761
                ok = 0;
1762
                fprintf( stderr, "memcpy_aligned FAILED: size=%d\n", (int)size );
1763
                break;
1764
            }
1765
        }
1766
        report( "memcpy aligned :" );
1767
    }
1768

1769
    if( mc_a.memzero_aligned != mc_ref.memzero_aligned )
1770
    {
1771
        set_func_name( "memzero_aligned" );
1772
        ok = 1; used_asm = 1;
1773
        for( size_t size = 128; size < 1024; size += 128 )
1774
        {
1775
            memset( buf4, 0xAA, size + 1 );
1776
            call_c( mc_c.memzero_aligned, buf3, size );
1777
            call_a( mc_a.memzero_aligned, buf4, size );
1778
            if( memcmp( buf3, buf4, size ) || buf4[size] != 0xAA )
1779
            {
1780
                ok = 0;
1781
                fprintf( stderr, "memzero_aligned FAILED: size=%d\n", (int)size );
1782
                break;
1783
            }
1784
        }
1785
        report( "memzero aligned :" );
1786
    }
1787

1788
    return ret;
1789
}
1790

1791
static int check_deblock( int cpu_ref, int cpu_new )
1792
{
1793
    x264_deblock_function_t db_c;
1794
    x264_deblock_function_t db_ref;
1795
    x264_deblock_function_t db_a;
1796
    int ret = 0, ok = 1, used_asm = 0;
1797
    int alphas[36], betas[36];
1798
    int8_t tcs[36][4];
1799

1800
    x264_deblock_init( 0, &db_c, 0 );
1801
    x264_deblock_init( cpu_ref, &db_ref, 0 );
1802
    x264_deblock_init( cpu_new, &db_a, 0 );
1803

1804
    /* not exactly the real values of a,b,tc but close enough */
1805
    for( int i = 35, a = 255, c = 250; i >= 0; i-- )
1806
    {
1807
        alphas[i] = a << (BIT_DEPTH-8);
1808
        betas[i] = (i+1)/2 << (BIT_DEPTH-8);
1809
        tcs[i][0] = tcs[i][3] = (c+6)/10 << (BIT_DEPTH-8);
1810
        tcs[i][1] = (c+7)/15 << (BIT_DEPTH-8);
1811
        tcs[i][2] = (c+9)/20 << (BIT_DEPTH-8);
1812
        a = a*9/10;
1813
        c = c*9/10;
1814
    }
1815

1816
#define TEST_DEBLOCK( name, align, ... ) \
1817
    for( int i = 0; i < 36; i++ ) \
1818
    { \
1819
        intptr_t off = 8*32 + (i&15)*4*!align; /* benchmark various alignments of h filter */ \
1820
        for( int j = 0; j < 1024; j++ ) \
1821
            /* two distributions of random to excersize different failure modes */ \
1822
            pbuf3[j] = rand() & (i&1 ? 0xf : PIXEL_MAX ); \
1823
        memcpy( pbuf4, pbuf3, 1024 * sizeof(pixel) ); \
1824
        if( db_a.name != db_ref.name ) \
1825
        { \
1826
            set_func_name( #name ); \
1827
            used_asm = 1; \
1828
            call_c1( db_c.name, pbuf3+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1829
            call_a1( db_a.name, pbuf4+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1830
            if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
1831
            { \
1832
                ok = 0; \
1833
                fprintf( stderr, #name "(a=%d, b=%d): [FAILED]\n", alphas[i], betas[i] ); \
1834
                break; \
1835
            } \
1836
            call_c2( db_c.name, pbuf3+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1837
            call_a2( db_a.name, pbuf4+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
1838
        } \
1839
    }
1840

1841
    TEST_DEBLOCK( deblock_luma[0], 0, tcs[i] );
1842
    TEST_DEBLOCK( deblock_luma[1], 1, tcs[i] );
1843
    TEST_DEBLOCK( deblock_h_chroma_420, 0, tcs[i] );
1844
    TEST_DEBLOCK( deblock_h_chroma_422, 0, tcs[i] );
1845
    TEST_DEBLOCK( deblock_chroma_420_mbaff, 0, tcs[i] );
1846
    TEST_DEBLOCK( deblock_chroma_422_mbaff, 0, tcs[i] );
1847
    TEST_DEBLOCK( deblock_chroma[1], 1, tcs[i] );
1848
    TEST_DEBLOCK( deblock_luma_intra[0], 0 );
1849
    TEST_DEBLOCK( deblock_luma_intra[1], 1 );
1850
    TEST_DEBLOCK( deblock_h_chroma_420_intra, 0 );
1851
    TEST_DEBLOCK( deblock_h_chroma_422_intra, 0 );
1852
    TEST_DEBLOCK( deblock_chroma_420_intra_mbaff, 0 );
1853
    TEST_DEBLOCK( deblock_chroma_422_intra_mbaff, 0 );
1854
    TEST_DEBLOCK( deblock_chroma_intra[1], 1 );
1855

1856
    if( db_a.deblock_strength != db_ref.deblock_strength )
1857
    {
1858
        for( int i = 0; i < 100; i++ )
1859
        {
1860
            ALIGNED_ARRAY_16( uint8_t, nnz, [X264_SCAN8_SIZE] );
1861
            ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );
1862
            ALIGNED_ARRAY_16( int16_t, mv, [2],[X264_SCAN8_LUMA_SIZE][2] );
1863
            ALIGNED_ARRAY_N( uint8_t, bs, [2],[2][8][4] );
1864
            memset( bs, 99, sizeof(uint8_t)*2*4*8*2 );
1865
            for( int j = 0; j < X264_SCAN8_SIZE; j++ )
1866
                nnz[j] = ((rand()&7) == 7) * rand() & 0xf;
1867
            for( int j = 0; j < 2; j++ )
1868
                for( int k = 0; k < X264_SCAN8_LUMA_SIZE; k++ )
1869
                {
1870
                    ref[j][k] = ((rand()&3) != 3) ? 0 : (rand() & 31) - 2;
1871
                    for( int l = 0; l < 2; l++ )
1872
                        mv[j][k][l] = ((rand()&7) != 7) ? (rand()&7) - 3 : (rand()&1023) - 512;
1873
                }
1874
            set_func_name( "deblock_strength" );
1875
            call_c( db_c.deblock_strength, nnz, ref, mv, bs[0], 2<<(i&1), ((i>>1)&1) );
1876
            call_a( db_a.deblock_strength, nnz, ref, mv, bs[1], 2<<(i&1), ((i>>1)&1) );
1877
            if( memcmp( bs[0], bs[1], sizeof(uint8_t)*2*4*8 ) )
1878
            {
1879
                ok = 0;
1880
                fprintf( stderr, "deblock_strength: [FAILED]\n" );
1881
                for( int j = 0; j < 2; j++ )
1882
                {
1883
                    for( int k = 0; k < 2; k++ )
1884
                        for( int l = 0; l < 4; l++ )
1885
                        {
1886
                            for( int m = 0; m < 4; m++ )
1887
                                printf("%d ",bs[j][k][l][m]);
1888
                            printf("\n");
1889
                        }
1890
                    printf("\n");
1891
                }
1892
                break;
1893
            }
1894
        }
1895
    }
1896

1897
    report( "deblock :" );
1898

1899
    return ret;
1900
}
1901

1902
static int check_quant( int cpu_ref, int cpu_new )
1903
{
1904
    x264_quant_function_t qf_c;
1905
    x264_quant_function_t qf_ref;
1906
    x264_quant_function_t qf_a;
1907
    ALIGNED_ARRAY_N( dctcoef, dct1,[64] );
1908
    ALIGNED_ARRAY_N( dctcoef, dct2,[64] );
1909
    ALIGNED_ARRAY_N( dctcoef, dct3,[8],[16] );
1910
    ALIGNED_ARRAY_N( dctcoef, dct4,[8],[16] );
1911
    ALIGNED_ARRAY_N( uint8_t, cqm_buf,[64] );
1912
    int ret = 0, ok, used_asm;
1913
    int oks[3] = {1,1,1}, used_asms[3] = {0,0,0};
1914
    x264_t h_buf;
1915
    x264_t *h = &h_buf;
1916
    memset( h, 0, sizeof(*h) );
1917
    h->sps->i_chroma_format_idc = 1;
1918
    x264_param_default( &h->param );
1919
    h->chroma_qp_table = i_chroma_qp_table + 12;
1920
    h->param.analyse.b_transform_8x8 = 1;
1921

1922
    for( int i_cqm = 0; i_cqm < 4; i_cqm++ )
1923
    {
1924
        if( i_cqm == 0 )
1925
        {
1926
            for( int i = 0; i < 6; i++ )
1927
                h->pps->scaling_list[i] = x264_cqm_flat16;
1928
            h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_FLAT;
1929
        }
1930
        else if( i_cqm == 1 )
1931
        {
1932
            for( int i = 0; i < 6; i++ )
1933
                h->pps->scaling_list[i] = x264_cqm_jvt[i];
1934
            h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_JVT;
1935
        }
1936
        else
1937
        {
1938
            int max_scale = BIT_DEPTH < 10 ? 255 : 228;
1939
            if( i_cqm == 2 )
1940
                for( int i = 0; i < 64; i++ )
1941
                    cqm_buf[i] = 10 + rand() % (max_scale - 9);
1942
            else
1943
                for( int i = 0; i < 64; i++ )
1944
                    cqm_buf[i] = 1;
1945
            for( int i = 0; i < 6; i++ )
1946
                h->pps->scaling_list[i] = cqm_buf;
1947
            h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_CUSTOM;
1948
        }
1949

1950
        h->param.rc.i_qp_min = 0;
1951
        h->param.rc.i_qp_max = QP_MAX_SPEC;
1952
        x264_cqm_init( h );
1953
        x264_quant_init( h, 0, &qf_c );
1954
        x264_quant_init( h, cpu_ref, &qf_ref );
1955
        x264_quant_init( h, cpu_new, &qf_a );
1956

1957
#define INIT_QUANT8(j,max) \
1958
        { \
1959
            static const int scale1d[8] = {32,31,24,31,32,31,24,31}; \
1960
            for( int i = 0; i < max; i++ ) \
1961
            { \
1962
                unsigned int scale = (255*scale1d[(i>>3)&7]*scale1d[i&7])/16; \
1963
                dct1[i] = dct2[i] = (j>>(i>>6))&1 ? (rand()%(2*scale+1))-scale : 0; \
1964
            } \
1965
        }
1966

1967
#define INIT_QUANT4(j,max) \
1968
        { \
1969
            static const int scale1d[4] = {4,6,4,6}; \
1970
            for( int i = 0; i < max; i++ ) \
1971
            { \
1972
                unsigned int scale = 255*scale1d[(i>>2)&3]*scale1d[i&3]; \
1973
                dct1[i] = dct2[i] = (j>>(i>>4))&1 ? (rand()%(2*scale+1))-scale : 0; \
1974
            } \
1975
        }
1976

1977
#define TEST_QUANT_DC( name, cqm ) \
1978
        if( qf_a.name != qf_ref.name ) \
1979
        { \
1980
            set_func_name( #name ); \
1981
            used_asms[0] = 1; \
1982
            for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
1983
            { \
1984
                for( int j = 0; j < 2; j++ ) \
1985
                { \
1986
                    int result_c, result_a; \
1987
                    for( int i = 0; i < 16; i++ ) \
1988
                        dct1[i] = dct2[i] = j ? (rand() & 0x1fff) - 0xfff : 0; \
1989
                    result_c = call_c1( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1990
                    result_a = call_a1( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1991
                    if( memcmp( dct1, dct2, 16*sizeof(dctcoef) ) || result_c != result_a ) \
1992
                    { \
1993
                        oks[0] = 0; \
1994
                        fprintf( stderr, #name "(cqm=%d): [FAILED]\n", i_cqm ); \
1995
                        break; \
1996
                    } \
1997
                    call_c2( qf_c.name, dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1998
                    call_a2( qf_a.name, dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
1999
                } \
2000
            } \
2001
        }
2002

2003
#define TEST_QUANT( qname, block, type, w, maxj ) \
2004
        if( qf_a.qname != qf_ref.qname ) \
2005
        { \
2006
            set_func_name( #qname ); \
2007
            used_asms[0] = 1; \
2008
            for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2009
            { \
2010
                for( int j = 0; j < maxj; j++ ) \
2011
                { \
2012
                    INIT_QUANT##type(j, w*w) \
2013
                    int result_c = call_c1( qf_c.qname, (void*)dct1, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
2014
                    int result_a = call_a1( qf_a.qname, (void*)dct2, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
2015
                    if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) || result_c != result_a ) \
2016
                    { \
2017
                        oks[0] = 0; \
2018
                        fprintf( stderr, #qname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
2019
                        break; \
2020
                    } \
2021
                    call_c2( qf_c.qname, (void*)dct1, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
2022
                    call_a2( qf_a.qname, (void*)dct2, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
2023
                } \
2024
            } \
2025
        }
2026

2027
        TEST_QUANT( quant_8x8, CQM_8IY, 8, 8, 2 );
2028
        TEST_QUANT( quant_8x8, CQM_8PY, 8, 8, 2 );
2029
        TEST_QUANT( quant_4x4, CQM_4IY, 4, 4, 2 );
2030
        TEST_QUANT( quant_4x4, CQM_4PY, 4, 4, 2 );
2031
        TEST_QUANT( quant_4x4x4, CQM_4IY, 4, 8, 16 );
2032
        TEST_QUANT( quant_4x4x4, CQM_4PY, 4, 8, 16 );
2033
        TEST_QUANT_DC( quant_4x4_dc, **h->quant4_mf[CQM_4IY] );
2034
        TEST_QUANT_DC( quant_2x2_dc, **h->quant4_mf[CQM_4IC] );
2035

2036
#define TEST_DEQUANT( qname, dqname, block, w ) \
2037
        if( qf_a.dqname != qf_ref.dqname ) \
2038
        { \
2039
            set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
2040
            used_asms[1] = 1; \
2041
            for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2042
            { \
2043
                INIT_QUANT##w(1, w*w) \
2044
                qf_c.qname( dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
2045
                memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
2046
                call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2047
                call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2048
                if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
2049
                { \
2050
                    oks[1] = 0; \
2051
                    fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
2052
                    break; \
2053
                } \
2054
                call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2055
                call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2056
            } \
2057
        }
2058

2059
        TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8IY, 8 );
2060
        TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8PY, 8 );
2061
        TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4IY, 4 );
2062
        TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4PY, 4 );
2063

2064
#define TEST_DEQUANT_DC( qname, dqname, block, w ) \
2065
        if( qf_a.dqname != qf_ref.dqname ) \
2066
        { \
2067
            set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
2068
            used_asms[1] = 1; \
2069
            for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2070
            { \
2071
                for( int i = 0; i < 16; i++ ) \
2072
                    dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16; \
2073
                qf_c.qname( dct1, h->quant##w##_mf[block][qp][0]>>1, h->quant##w##_bias[block][qp][0]>>1 ); \
2074
                memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
2075
                call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2076
                call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2077
                if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
2078
                { \
2079
                    oks[1] = 0; \
2080
                    fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
2081
                } \
2082
                call_c2( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
2083
                call_a2( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
2084
            } \
2085
        }
2086

2087
        TEST_DEQUANT_DC( quant_4x4_dc, dequant_4x4_dc, CQM_4IY, 4 );
2088

2089
        if( qf_a.idct_dequant_2x4_dc != qf_ref.idct_dequant_2x4_dc )
2090
        {
2091
            set_func_name( "idct_dequant_2x4_dc_%s", i_cqm?"cqm":"flat" );
2092
            used_asms[1] = 1;
2093
            for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- )
2094
            {
2095
                for( int i = 0; i < 8; i++ )
2096
                    dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16;
2097
                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 );
2098
                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 );
2099
                call_c( qf_c.idct_dequant_2x4_dc, dct1, dct3, h->dequant4_mf[CQM_4IC], qp+3 );
2100
                call_a( qf_a.idct_dequant_2x4_dc, dct1, dct4, h->dequant4_mf[CQM_4IC], qp+3 );
2101
                for( int i = 0; i < 8; i++ )
2102
                    if( dct3[i][0] != dct4[i][0] )
2103
                    {
2104
                        oks[1] = 0;
2105
                        fprintf( stderr, "idct_dequant_2x4_dc (qp=%d, cqm=%d): [FAILED]\n", qp, i_cqm );
2106
                        break;
2107
                    }
2108
            }
2109
        }
2110

2111
        if( qf_a.idct_dequant_2x4_dconly != qf_ref.idct_dequant_2x4_dconly )
2112
        {
2113
            set_func_name( "idct_dequant_2x4_dc_%s", i_cqm?"cqm":"flat" );
2114
            used_asms[1] = 1;
2115
            for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- )
2116
            {
2117
                for( int i = 0; i < 8; i++ )
2118
                    dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16;
2119
                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 );
2120
                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 );
2121
                memcpy( dct2, dct1, 8*sizeof(dctcoef) );
2122
                call_c1( qf_c.idct_dequant_2x4_dconly, dct1, h->dequant4_mf[CQM_4IC], qp+3 );
2123
                call_a1( qf_a.idct_dequant_2x4_dconly, dct2, h->dequant4_mf[CQM_4IC], qp+3 );
2124
                if( memcmp( dct1, dct2, 8*sizeof(dctcoef) ) )
2125
                {
2126
                    oks[1] = 0;
2127
                    fprintf( stderr, "idct_dequant_2x4_dconly (qp=%d, cqm=%d): [FAILED]\n", qp, i_cqm );
2128
                    break;
2129
                }
2130
                call_c2( qf_c.idct_dequant_2x4_dconly, dct1, h->dequant4_mf[CQM_4IC], qp+3 );
2131
                call_a2( qf_a.idct_dequant_2x4_dconly, dct2, h->dequant4_mf[CQM_4IC], qp+3 );
2132
            }
2133
        }
2134

2135
#define TEST_OPTIMIZE_CHROMA_DC( optname, size ) \
2136
        if( qf_a.optname != qf_ref.optname ) \
2137
        { \
2138
            set_func_name( #optname ); \
2139
            used_asms[2] = 1; \
2140
            for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
2141
            { \
2142
                int qpdc = qp + (size == 8 ? 3 : 0); \
2143
                int dmf = h->dequant4_mf[CQM_4IC][qpdc%6][0] << qpdc/6; \
2144
                if( dmf > 32*64 ) \
2145
                    continue; \
2146
                for( int i = 16; ; i <<= 1 ) \
2147
                { \
2148
                    int res_c, res_asm; \
2149
                    int max = X264_MIN( i, PIXEL_MAX*16 ); \
2150
                    for( int j = 0; j < size; j++ ) \
2151
                        dct1[j] = rand()%(max*2+1) - max; \
2152
                    for( int j = 0; i <= size; j += 4 ) \
2153
                        qf_c.quant_2x2_dc( &dct1[j], h->quant4_mf[CQM_4IC][qpdc][0]>>1, h->quant4_bias[CQM_4IC][qpdc][0]>>1 ); \
2154
                    memcpy( dct2, dct1, size*sizeof(dctcoef) ); \
2155
                    res_c   = call_c1( qf_c.optname, dct1, dmf ); \
2156
                    res_asm = call_a1( qf_a.optname, dct2, dmf ); \
2157
                    if( res_c != res_asm || memcmp( dct1, dct2, size*sizeof(dctcoef) ) ) \
2158
                    { \
2159
                        oks[2] = 0; \
2160
                        fprintf( stderr, #optname "(qp=%d, res_c=%d, res_asm=%d): [FAILED]\n", qp, res_c, res_asm ); \
2161
                    } \
2162
                    call_c2( qf_c.optname, dct1, dmf ); \
2163
                    call_a2( qf_a.optname, dct2, dmf ); \
2164
                    if( i >= PIXEL_MAX*16 ) \
2165
                        break; \
2166
                } \
2167
            } \
2168
        }
2169

2170
        TEST_OPTIMIZE_CHROMA_DC( optimize_chroma_2x2_dc, 4 );
2171
        TEST_OPTIMIZE_CHROMA_DC( optimize_chroma_2x4_dc, 8 );
2172

2173
        x264_cqm_delete( h );
2174
    }
2175

2176
    ok = oks[0]; used_asm = used_asms[0];
2177
    report( "quant :" );
2178

2179
    ok = oks[1]; used_asm = used_asms[1];
2180
    report( "dequant :" );
2181

2182
    ok = oks[2]; used_asm = used_asms[2];
2183
    report( "optimize chroma dc :" );
2184

2185
    ok = 1; used_asm = 0;
2186
    if( qf_a.denoise_dct != qf_ref.denoise_dct )
2187
    {
2188
        used_asm = 1;
2189
        for( int size = 16; size <= 64; size += 48 )
2190
        {
2191
            set_func_name( "denoise_dct" );
2192
            memcpy( dct1, buf1, size*sizeof(dctcoef) );
2193
            memcpy( dct2, buf1, size*sizeof(dctcoef) );
2194
            memcpy( buf3+256, buf3, 256 );
2195
            call_c1( qf_c.denoise_dct, dct1, (uint32_t*)buf3,       (udctcoef*)buf2, size );
2196
            call_a1( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
2197
            if( memcmp( dct1, dct2, size*sizeof(dctcoef) ) || memcmp( buf3+4, buf3+256+4, (size-1)*sizeof(uint32_t) ) )
2198
                ok = 0;
2199
            call_c2( qf_c.denoise_dct, dct1, (uint32_t*)buf3,       (udctcoef*)buf2, size );
2200
            call_a2( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
2201
        }
2202
    }
2203
    report( "denoise dct :" );
2204

2205
#define TEST_DECIMATE( decname, w, ac, thresh ) \
2206
    if( qf_a.decname != qf_ref.decname ) \
2207
    { \
2208
        set_func_name( #decname ); \
2209
        used_asm = 1; \
2210
        for( int i = 0; i < 100; i++ ) \
2211
        { \
2212
            static const int distrib[16] = {1,1,1,1,1,1,1,1,1,1,1,1,2,3,4};\
2213
            static const int zerorate_lut[4] = {3,7,15,31};\
2214
            int zero_rate = zerorate_lut[i&3];\
2215
            for( int idx = 0; idx < w*w; idx++ ) \
2216
            { \
2217
                int sign = (rand()&1) ? -1 : 1; \
2218
                int abs_level = distrib[rand()&15]; \
2219
                if( abs_level == 4 ) abs_level = rand()&0x3fff; \
2220
                int zero = !(rand()&zero_rate); \
2221
                dct1[idx] = zero * abs_level * sign; \
2222
            } \
2223
            if( ac ) \
2224
                dct1[0] = 0; \
2225
            int result_c = call_c( qf_c.decname, dct1 ); \
2226
            int result_a = call_a( qf_a.decname, dct1 ); \
2227
            if( X264_MIN(result_c,thresh) != X264_MIN(result_a,thresh) ) \
2228
            { \
2229
                ok = 0; \
2230
                fprintf( stderr, #decname ": [FAILED]\n" ); \
2231
                break; \
2232
            } \
2233
        } \
2234
    }
2235

2236
    ok = 1; used_asm = 0;
2237
    TEST_DECIMATE( decimate_score64, 8, 0, 6 );
2238
    TEST_DECIMATE( decimate_score16, 4, 0, 6 );
2239
    TEST_DECIMATE( decimate_score15, 4, 1, 7 );
2240
    report( "decimate_score :" );
2241

2242
#define TEST_LAST( last, lastname, size, ac ) \
2243
    if( qf_a.last != qf_ref.last ) \
2244
    { \
2245
        set_func_name( #lastname ); \
2246
        used_asm = 1; \
2247
        for( int i = 0; i < 100; i++ ) \
2248
        { \
2249
            int nnz = 0; \
2250
            int max = rand() & (size-1); \
2251
            memset( dct1, 0, size*sizeof(dctcoef) ); \
2252
            for( int idx = ac; idx < max; idx++ ) \
2253
                nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
2254
            if( !nnz ) \
2255
                dct1[ac] = 1; \
2256
            int result_c = call_c( qf_c.last, dct1+ac ); \
2257
            int result_a = call_a( qf_a.last, dct1+ac ); \
2258
            if( result_c != result_a ) \
2259
            { \
2260
                ok = 0; \
2261
                fprintf( stderr, #lastname ": [FAILED]\n" ); \
2262
                break; \
2263
            } \
2264
        } \
2265
    }
2266

2267
    ok = 1; used_asm = 0;
2268
    TEST_LAST( coeff_last4              , coeff_last4,   4, 0 );
2269
    TEST_LAST( coeff_last8              , coeff_last8,   8, 0 );
2270
    TEST_LAST( coeff_last[  DCT_LUMA_AC], coeff_last15, 16, 1 );
2271
    TEST_LAST( coeff_last[ DCT_LUMA_4x4], coeff_last16, 16, 0 );
2272
    TEST_LAST( coeff_last[ DCT_LUMA_8x8], coeff_last64, 64, 0 );
2273
    report( "coeff_last :" );
2274

2275
#define TEST_LEVELRUN( lastname, name, size, ac ) \
2276
    if( qf_a.lastname != qf_ref.lastname ) \
2277
    { \
2278
        set_func_name( #name ); \
2279
        used_asm = 1; \
2280
        for( int i = 0; i < 100; i++ ) \
2281
        { \
2282
            x264_run_level_t runlevel_c, runlevel_a; \
2283
            int nnz = 0; \
2284
            int max = rand() & (size-1); \
2285
            memset( dct1, 0, size*sizeof(dctcoef) ); \
2286
            memcpy( &runlevel_a, buf1+i, sizeof(x264_run_level_t) ); \
2287
            memcpy( &runlevel_c, buf1+i, sizeof(x264_run_level_t) ); \
2288
            for( int idx = ac; idx < max; idx++ ) \
2289
                nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
2290
            if( !nnz ) \
2291
                dct1[ac] = 1; \
2292
            int result_c = call_c( qf_c.lastname, dct1+ac, &runlevel_c ); \
2293
            int result_a = call_a( qf_a.lastname, dct1+ac, &runlevel_a ); \
2294
            if( result_c != result_a || runlevel_c.last != runlevel_a.last || \
2295
                runlevel_c.mask != runlevel_a.mask || \
2296
                memcmp(runlevel_c.level, runlevel_a.level, sizeof(dctcoef)*result_c)) \
2297
            { \
2298
                ok = 0; \
2299
                fprintf( stderr, #name ": [FAILED]\n" ); \
2300
                break; \
2301
            } \
2302
        } \
2303
    }
2304

2305
    ok = 1; used_asm = 0;
2306
    TEST_LEVELRUN( coeff_level_run4              , coeff_level_run4,   4, 0 );
2307
    TEST_LEVELRUN( coeff_level_run8              , coeff_level_run8,   8, 0 );
2308
    TEST_LEVELRUN( coeff_level_run[  DCT_LUMA_AC], coeff_level_run15, 16, 1 );
2309
    TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_4x4], coeff_level_run16, 16, 0 );
2310
    report( "coeff_level_run :" );
2311

2312
    return ret;
2313
}
2314

2315
static int check_intra( int cpu_ref, int cpu_new )
2316
{
2317
    int ret = 0, ok = 1, used_asm = 0;
2318
    ALIGNED_ARRAY_32( pixel, edge,[36] );
2319
    ALIGNED_ARRAY_32( pixel, edge2,[36] );
2320
    ALIGNED_ARRAY_32( pixel, fdec,[FDEC_STRIDE*20] );
2321
    struct
2322
    {
2323
        x264_predict_t      predict_16x16[4+3];
2324
        x264_predict_t      predict_8x8c[4+3];
2325
        x264_predict_t      predict_8x16c[4+3];
2326
        x264_predict8x8_t   predict_8x8[9+3];
2327
        x264_predict_t      predict_4x4[9+3];
2328
        x264_predict_8x8_filter_t predict_8x8_filter;
2329
    } ip_c, ip_ref, ip_a;
2330

2331
    x264_predict_16x16_init( 0, ip_c.predict_16x16 );
2332
    x264_predict_8x8c_init( 0, ip_c.predict_8x8c );
2333
    x264_predict_8x16c_init( 0, ip_c.predict_8x16c );
2334
    x264_predict_8x8_init( 0, ip_c.predict_8x8, &ip_c.predict_8x8_filter );
2335
    x264_predict_4x4_init( 0, ip_c.predict_4x4 );
2336

2337
    x264_predict_16x16_init( cpu_ref, ip_ref.predict_16x16 );
2338
    x264_predict_8x8c_init( cpu_ref, ip_ref.predict_8x8c );
2339
    x264_predict_8x16c_init( cpu_ref, ip_ref.predict_8x16c );
2340
    x264_predict_8x8_init( cpu_ref, ip_ref.predict_8x8, &ip_ref.predict_8x8_filter );
2341
    x264_predict_4x4_init( cpu_ref, ip_ref.predict_4x4 );
2342

2343
    x264_predict_16x16_init( cpu_new, ip_a.predict_16x16 );
2344
    x264_predict_8x8c_init( cpu_new, ip_a.predict_8x8c );
2345
    x264_predict_8x16c_init( cpu_new, ip_a.predict_8x16c );
2346
    x264_predict_8x8_init( cpu_new, ip_a.predict_8x8, &ip_a.predict_8x8_filter );
2347
    x264_predict_4x4_init( cpu_new, ip_a.predict_4x4 );
2348

2349
    memcpy( fdec, pbuf1, 32*20 * sizeof(pixel) );\
2350

2351
    ip_c.predict_8x8_filter( fdec+48, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
2352

2353
#define INTRA_TEST( name, dir, w, h, align, bench, ... )\
2354
    if( ip_a.name[dir] != ip_ref.name[dir] )\
2355
    {\
2356
        set_func_name( "intra_%s_%s", #name, intra_##name##_names[dir] );\
2357
        used_asm = 1;\
2358
        memcpy( pbuf3, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
2359
        memcpy( pbuf4, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
2360
        for( int a = 0; a < (do_bench ? 64/sizeof(pixel) : 1); a += align )\
2361
        {\
2362
            call_c##bench( ip_c.name[dir], pbuf3+48+a, ##__VA_ARGS__ );\
2363
            call_a##bench( ip_a.name[dir], pbuf4+48+a, ##__VA_ARGS__ );\
2364
            if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*20 * sizeof(pixel) ) )\
2365
            {\
2366
                fprintf( stderr, #name "[%d] :  [FAILED]\n", dir );\
2367
                ok = 0;\
2368
                if( ip_c.name == (void *)ip_c.predict_8x8 )\
2369
                {\
2370
                    for( int k = -1; k < 16; k++ )\
2371
                        printf( "%2x ", edge[16+k] );\
2372
                    printf( "\n" );\
2373
                }\
2374
                for( int j = 0; j < h; j++ )\
2375
                {\
2376
                    if( ip_c.name == (void *)ip_c.predict_8x8 )\
2377
                        printf( "%2x ", edge[14-j] );\
2378
                    for( int k = 0; k < w; k++ )\
2379
                        printf( "%2x ", pbuf4[48+k+j*FDEC_STRIDE] );\
2380
                    printf( "\n" );\
2381
                }\
2382
                printf( "\n" );\
2383
                for( int j = 0; j < h; j++ )\
2384
                {\
2385
                    if( ip_c.name == (void *)ip_c.predict_8x8 )\
2386
                        printf( "   " );\
2387
                    for( int k = 0; k < w; k++ )\
2388
                        printf( "%2x ", pbuf3[48+k+j*FDEC_STRIDE] );\
2389
                    printf( "\n" );\
2390
                }\
2391
                break;\
2392
            }\
2393
        }\
2394
    }
2395

2396
    for( int i = 0; i < 12; i++ )
2397
        INTRA_TEST(   predict_4x4, i,  4,  4,  4, );
2398
    for( int i = 0; i < 7; i++ )
2399
        INTRA_TEST(  predict_8x8c, i,  8,  8, 16, );
2400
    for( int i = 0; i < 7; i++ )
2401
        INTRA_TEST( predict_8x16c, i,  8, 16, 16, );
2402
    for( int i = 0; i < 7; i++ )
2403
        INTRA_TEST( predict_16x16, i, 16, 16, 16, );
2404
    for( int i = 0; i < 12; i++ )
2405
        INTRA_TEST(   predict_8x8, i,  8,  8,  8, , edge );
2406

2407
    set_func_name("intra_predict_8x8_filter");
2408
    if( ip_a.predict_8x8_filter != ip_ref.predict_8x8_filter )
2409
    {
2410
        used_asm = 1;
2411
        for( int i = 0; i < 32; i++ )
2412
        {
2413
            if( !(i&7) || ((i&MB_TOPRIGHT) && !(i&MB_TOP)) )
2414
                continue;
2415
            int neighbor = (i&24)>>1;
2416
            memset( edge,  0, 36*sizeof(pixel) );
2417
            memset( edge2, 0, 36*sizeof(pixel) );
2418
            call_c( ip_c.predict_8x8_filter, pbuf1+48, edge,  neighbor, i&7 );
2419
            call_a( ip_a.predict_8x8_filter, pbuf1+48, edge2, neighbor, i&7 );
2420
            if( !(neighbor&MB_TOPLEFT) )
2421
                edge[15] = edge2[15] = 0;
2422
            if( memcmp( edge+7, edge2+7, (i&MB_TOPRIGHT ? 26 : i&MB_TOP ? 17 : 8) * sizeof(pixel) ) )
2423
            {
2424
                fprintf( stderr, "predict_8x8_filter :  [FAILED] %d %d\n", (i&24)>>1, i&7);
2425
                ok = 0;
2426
            }
2427
        }
2428
    }
2429

2430
#define EXTREMAL_PLANE( w, h ) \
2431
    { \
2432
        int max[7]; \
2433
        for( int j = 0; j < 7; j++ ) \
2434
            max[j] = test ? rand()&PIXEL_MAX : PIXEL_MAX; \
2435
        fdec[48-1-FDEC_STRIDE] = (i&1)*max[0]; \
2436
        for( int j = 0; j < w/2; j++ ) \
2437
            fdec[48+j-FDEC_STRIDE] = (!!(i&2))*max[1]; \
2438
        for( int j = w/2; j < w-1; j++ ) \
2439
            fdec[48+j-FDEC_STRIDE] = (!!(i&4))*max[2]; \
2440
        fdec[48+(w-1)-FDEC_STRIDE] = (!!(i&8))*max[3]; \
2441
        for( int j = 0; j < h/2; j++ ) \
2442
            fdec[48+j*FDEC_STRIDE-1] = (!!(i&16))*max[4]; \
2443
        for( int j = h/2; j < h-1; j++ ) \
2444
            fdec[48+j*FDEC_STRIDE-1] = (!!(i&32))*max[5]; \
2445
        fdec[48+(h-1)*FDEC_STRIDE-1] = (!!(i&64))*max[6]; \
2446
    }
2447
    /* Extremal test case for planar prediction. */
2448
    for( int test = 0; test < 100 && ok; test++ )
2449
        for( int i = 0; i < 128 && ok; i++ )
2450
        {
2451
            EXTREMAL_PLANE(  8,  8 );
2452
            INTRA_TEST(  predict_8x8c, I_PRED_CHROMA_P,  8,  8, 64, 1 );
2453
            EXTREMAL_PLANE(  8, 16 );
2454
            INTRA_TEST( predict_8x16c, I_PRED_CHROMA_P,  8, 16, 64, 1 );
2455
            EXTREMAL_PLANE( 16, 16 );
2456
            INTRA_TEST( predict_16x16,  I_PRED_16x16_P, 16, 16, 64, 1 );
2457
        }
2458
    report( "intra pred :" );
2459
    return ret;
2460
}
2461

2462
#define DECL_CABAC(cpu) \
2463
static void run_cabac_decision_##cpu( x264_t *h, uint8_t *dst )\
2464
{\
2465
    x264_cabac_t cb;\
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_decision_##cpu( &cb, buf1[i]>>1, buf1[i]&1 );\
2470
}\
2471
static void run_cabac_bypass_##cpu( x264_t *h, uint8_t *dst )\
2472
{\
2473
    x264_cabac_t cb;\
2474
    x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2475
    x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2476
    for( int i = 0; i < 0x1000; i++ )\
2477
        x264_cabac_encode_bypass_##cpu( &cb, buf1[i]&1 );\
2478
}\
2479
static void run_cabac_terminal_##cpu( x264_t *h, uint8_t *dst )\
2480
{\
2481
    x264_cabac_t cb;\
2482
    x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
2483
    x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
2484
    for( int i = 0; i < 0x1000; i++ )\
2485
        x264_cabac_encode_terminal_##cpu( &cb );\
2486
}
2487
DECL_CABAC(c)
2488
#if HAVE_MMX
2489
DECL_CABAC(asm)
2490
#elif defined(ARCH_AARCH64)
2491
DECL_CABAC(asm)
2492
#else
2493
#define run_cabac_decision_asm run_cabac_decision_c
2494
#define run_cabac_bypass_asm run_cabac_bypass_c
2495
#define run_cabac_terminal_asm run_cabac_terminal_c
2496
#endif
2497

2498
extern const uint8_t x264_count_cat_m1[14];
2499
void x264_cabac_block_residual_c( x264_t *h, x264_cabac_t *cb, int ctx_block_cat, dctcoef *l );
2500
void x264_cabac_block_residual_8x8_rd_c( x264_t *h, x264_cabac_t *cb, int ctx_block_cat, dctcoef *l );
2501
void x264_cabac_block_residual_rd_c( x264_t *h, x264_cabac_t *cb, int ctx_block_cat, dctcoef *l );
2502

2503
static int check_cabac( int cpu_ref, int cpu_new )
2504
{
2505
    int ret = 0, ok = 1, used_asm = 0;
2506
    x264_t h;
2507
    h.sps->i_chroma_format_idc = 3;
2508

2509
    x264_bitstream_function_t bs_ref;
2510
    x264_bitstream_function_t bs_a;
2511
    x264_bitstream_init( cpu_ref, &bs_ref );
2512
    x264_bitstream_init( cpu_new, &bs_a );
2513
    x264_quant_init( &h, cpu_new, &h.quantf );
2514
    h.quantf.coeff_last[DCT_CHROMA_DC] = h.quantf.coeff_last4;
2515

2516
#define CABAC_RESIDUAL(name, start, end, rd)\
2517
{\
2518
    if( bs_a.name##_internal && (bs_a.name##_internal != bs_ref.name##_internal || (cpu_new&X264_CPU_SSE2_IS_SLOW)) )\
2519
    {\
2520
        used_asm = 1;\
2521
        set_func_name( #name );\
2522
        for( int i = 0; i < 2; i++ )\
2523
        {\
2524
            for( intptr_t ctx_block_cat = start; ctx_block_cat <= end; ctx_block_cat++ )\
2525
            {\
2526
                for( int j = 0; j < 256; j++ )\
2527
                {\
2528
                    ALIGNED_ARRAY_N( dctcoef, dct, [2],[64] );\
2529
                    uint8_t bitstream[2][1<<16];\
2530
                    static const uint8_t ctx_ac[14] = {0,1,0,0,1,0,0,1,0,0,0,1,0,0};\
2531
                    int ac = ctx_ac[ctx_block_cat];\
2532
                    int nz = 0;\
2533
                    while( !nz )\
2534
                    {\
2535
                        for( int k = 0; k <= x264_count_cat_m1[ctx_block_cat]; k++ )\
2536
                        {\
2537
                            /* Very rough distribution that covers possible inputs */\
2538
                            int rnd = rand();\
2539
                            int coef = !(rnd&3);\
2540
                            coef += !(rnd&  15) * (rand()&0x0006);\
2541
                            coef += !(rnd&  63) * (rand()&0x0008);\
2542
                            coef += !(rnd& 255) * (rand()&0x00F0);\
2543
                            coef += !(rnd&1023) * (rand()&0x7F00);\
2544
                            nz |= dct[0][ac+k] = dct[1][ac+k] = coef * ((rand()&1) ? 1 : -1);\
2545
                        }\
2546
                    }\
2547
                    h.mb.b_interlaced = i;\
2548
                    x264_cabac_t cb[2];\
2549
                    x264_cabac_context_init( &h, &cb[0], SLICE_TYPE_P, 26, 0 );\
2550
                    x264_cabac_context_init( &h, &cb[1], SLICE_TYPE_P, 26, 0 );\
2551
                    x264_cabac_encode_init( &cb[0], bitstream[0], bitstream[0]+0xfff0 );\
2552
                    x264_cabac_encode_init( &cb[1], bitstream[1], bitstream[1]+0xfff0 );\
2553
                    cb[0].f8_bits_encoded = 0;\
2554
                    cb[1].f8_bits_encoded = 0;\
2555
                    if( !rd ) memcpy( bitstream[1], bitstream[0], 0x400 );\
2556
                    call_c1( x264_##name##_c, &h, &cb[0], ctx_block_cat, dct[0]+ac );\
2557
                    call_a1( bs_a.name##_internal, dct[1]+ac, i, ctx_block_cat, &cb[1] );\
2558
                    ok = cb[0].f8_bits_encoded == cb[1].f8_bits_encoded && !memcmp(cb[0].state, cb[1].state, 1024);\
2559
                    if( !rd ) ok |= !memcmp( bitstream[1], bitstream[0], 0x400 ) && !memcmp( &cb[1], &cb[0], offsetof(x264_cabac_t, p_start) );\
2560
                    if( !ok )\
2561
                    {\
2562
                        fprintf( stderr, #name " :  [FAILED] ctx_block_cat %d", (int)ctx_block_cat );\
2563
                        if( rd && cb[0].f8_bits_encoded != cb[1].f8_bits_encoded )\
2564
                            fprintf( stderr, " (%d != %d)", cb[0].f8_bits_encoded, cb[1].f8_bits_encoded );\
2565
                        fprintf( stderr, "\n");\
2566
                        goto name##fail;\
2567
                    }\
2568
                    if( (j&15) == 0 )\
2569
                    {\
2570
                        call_c2( x264_##name##_c, &h, &cb[0], ctx_block_cat, dct[0]+ac );\
2571
                        call_a2( bs_a.name##_internal, dct[1]+ac, i, ctx_block_cat, &cb[1] );\
2572
                    }\
2573
                }\
2574
            }\
2575
        }\
2576
    }\
2577
}\
2578
name##fail:
2579

2580
    CABAC_RESIDUAL( cabac_block_residual, 0, DCT_LUMA_8x8, 0 )
2581
    report( "cabac residual:" );
2582

2583
    ok = 1; used_asm = 0;
2584
    CABAC_RESIDUAL( cabac_block_residual_rd, 0, DCT_LUMA_8x8-1, 1 )
2585
    CABAC_RESIDUAL( cabac_block_residual_8x8_rd, DCT_LUMA_8x8, DCT_LUMA_8x8, 1 )
2586
    report( "cabac residual rd:" );
2587

2588
    if( cpu_ref || run_cabac_decision_c == run_cabac_decision_asm )
2589
        return ret;
2590
    ok = 1; used_asm = 0;
2591
    x264_cabac_init( &h );
2592

2593
    set_func_name( "cabac_encode_decision" );
2594
    memcpy( buf4, buf3, 0x1000 );
2595
    call_c( run_cabac_decision_c, &h, buf3 );
2596
    call_a( run_cabac_decision_asm, &h, buf4 );
2597
    ok = !memcmp( buf3, buf4, 0x1000 );
2598
    report( "cabac decision:" );
2599

2600
    set_func_name( "cabac_encode_bypass" );
2601
    memcpy( buf4, buf3, 0x1000 );
2602
    call_c( run_cabac_bypass_c, &h, buf3 );
2603
    call_a( run_cabac_bypass_asm, &h, buf4 );
2604
    ok = !memcmp( buf3, buf4, 0x1000 );
2605
    report( "cabac bypass:" );
2606

2607
    set_func_name( "cabac_encode_terminal" );
2608
    memcpy( buf4, buf3, 0x1000 );
2609
    call_c( run_cabac_terminal_c, &h, buf3 );
2610
    call_a( run_cabac_terminal_asm, &h, buf4 );
2611
    ok = !memcmp( buf3, buf4, 0x1000 );
2612
    report( "cabac terminal:" );
2613

2614
    return ret;
2615
}
2616

2617
static int check_bitstream( int cpu_ref, int cpu_new )
2618
{
2619
    x264_bitstream_function_t bs_c;
2620
    x264_bitstream_function_t bs_ref;
2621
    x264_bitstream_function_t bs_a;
2622

2623
    int ret = 0, ok = 1, used_asm = 0;
2624

2625
    x264_bitstream_init( 0, &bs_c );
2626
    x264_bitstream_init( cpu_ref, &bs_ref );
2627
    x264_bitstream_init( cpu_new, &bs_a );
2628
    if( bs_a.nal_escape != bs_ref.nal_escape )
2629
    {
2630
        int size = 0x4000;
2631
        uint8_t *input = malloc(size+100);
2632
        uint8_t *output1 = malloc(size*2);
2633
        uint8_t *output2 = malloc(size*2);
2634
        used_asm = 1;
2635
        set_func_name( "nal_escape" );
2636
        for( int i = 0; i < 100; i++ )
2637
        {
2638
            /* Test corner-case sizes */
2639
            int test_size = i < 10 ? i+1 : rand() & 0x3fff;
2640
            /* Test 8 different probability distributions of zeros */
2641
            for( int j = 0; j < test_size+32; j++ )
2642
                input[j] = (rand()&((1 << ((i&7)+1)) - 1)) * rand();
2643
            uint8_t *end_c = (uint8_t*)call_c1( bs_c.nal_escape, output1, input, input+test_size );
2644
            uint8_t *end_a = (uint8_t*)call_a1( bs_a.nal_escape, output2, input, input+test_size );
2645
            int size_c = end_c-output1;
2646
            int size_a = end_a-output2;
2647
            if( size_c != size_a || memcmp( output1, output2, size_c ) )
2648
            {
2649
                fprintf( stderr, "nal_escape :  [FAILED] %d %d\n", size_c, size_a );
2650
                ok = 0;
2651
                break;
2652
            }
2653
        }
2654
        for( int j = 0; j < size+32; j++ )
2655
            input[j] = rand();
2656
        call_c2( bs_c.nal_escape, output1, input, input+size );
2657
        call_a2( bs_a.nal_escape, output2, input, input+size );
2658
        free(input);
2659
        free(output1);
2660
        free(output2);
2661
    }
2662
    report( "nal escape:" );
2663

2664
    return ret;
2665
}
2666

2667
static int check_all_funcs( int cpu_ref, int cpu_new )
2668
{
2669
    return check_pixel( cpu_ref, cpu_new )
2670
         + check_dct( cpu_ref, cpu_new )
2671
         + check_mc( cpu_ref, cpu_new )
2672
         + check_intra( cpu_ref, cpu_new )
2673
         + check_deblock( cpu_ref, cpu_new )
2674
         + check_quant( cpu_ref, cpu_new )
2675
         + check_cabac( cpu_ref, cpu_new )
2676
         + check_bitstream( cpu_ref, cpu_new );
2677
}
2678

2679
static int add_flags( int *cpu_ref, int *cpu_new, int flags, const char *name )
2680
{
2681
    *cpu_ref = *cpu_new;
2682
    *cpu_new |= flags;
2683
#if STACK_ALIGNMENT < 16
2684
    *cpu_new |= X264_CPU_STACK_MOD4;
2685
#endif
2686
    if( *cpu_new & X264_CPU_SSE2_IS_FAST )
2687
        *cpu_new &= ~X264_CPU_SSE2_IS_SLOW;
2688
    if( !quiet )
2689
        fprintf( stderr, "x264: %s\n", name );
2690
    return check_all_funcs( *cpu_ref, *cpu_new );
2691
}
2692

2693
static int check_all_flags( void )
2694
{
2695
    int ret = 0;
2696
    int cpu0 = 0, cpu1 = 0;
2697
    uint32_t cpu_detect = x264_cpu_detect();
2698
#if HAVE_MMX
2699
    if( cpu_detect & X264_CPU_MMX2 )
2700
    {
2701
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_MMX | X264_CPU_MMX2, "MMX" );
2702
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "MMX Cache64" );
2703
        cpu1 &= ~X264_CPU_CACHELINE_64;
2704
#if ARCH_X86
2705
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_32, "MMX Cache32" );
2706
        cpu1 &= ~X264_CPU_CACHELINE_32;
2707
#endif
2708
        if( cpu_detect & X264_CPU_LZCNT )
2709
        {
2710
            ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "MMX LZCNT" );
2711
            cpu1 &= ~X264_CPU_LZCNT;
2712
        }
2713
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "MMX SlowCTZ" );
2714
        cpu1 &= ~X264_CPU_SLOW_CTZ;
2715
    }
2716
    if( cpu_detect & X264_CPU_SSE )
2717
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE, "SSE" );
2718
    if( cpu_detect & X264_CPU_SSE2 )
2719
    {
2720
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2 | X264_CPU_SSE2_IS_SLOW, "SSE2Slow" );
2721
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2_IS_FAST, "SSE2Fast" );
2722
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSE2Fast Cache64" );
2723
        cpu1 &= ~X264_CPU_CACHELINE_64;
2724
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_SHUFFLE, "SSE2 SlowShuffle" );
2725
        cpu1 &= ~X264_CPU_SLOW_SHUFFLE;
2726
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSE2 SlowCTZ" );
2727
        cpu1 &= ~X264_CPU_SLOW_CTZ;
2728
        if( cpu_detect & X264_CPU_LZCNT )
2729
        {
2730
            ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSE2 LZCNT" );
2731
            cpu1 &= ~X264_CPU_LZCNT;
2732
        }
2733
    }
2734
    if( cpu_detect & X264_CPU_SSE3 )
2735
    {
2736
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE3 | X264_CPU_CACHELINE_64, "SSE3" );
2737
        cpu1 &= ~X264_CPU_CACHELINE_64;
2738
    }
2739
    if( cpu_detect & X264_CPU_SSSE3 )
2740
    {
2741
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSSE3, "SSSE3" );
2742
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64" );
2743
        cpu1 &= ~X264_CPU_CACHELINE_64;
2744
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_SHUFFLE, "SSSE3 SlowShuffle" );
2745
        cpu1 &= ~X264_CPU_SLOW_SHUFFLE;
2746
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSSE3 SlowCTZ" );
2747
        cpu1 &= ~X264_CPU_SLOW_CTZ;
2748
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSSE3 SlowAtom" );
2749
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64 SlowAtom" );
2750
        cpu1 &= ~X264_CPU_CACHELINE_64;
2751
        cpu1 &= ~X264_CPU_SLOW_ATOM;
2752
        if( cpu_detect & X264_CPU_LZCNT )
2753
        {
2754
            ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSSE3 LZCNT" );
2755
            cpu1 &= ~X264_CPU_LZCNT;
2756
        }
2757
    }
2758
    if( cpu_detect & X264_CPU_SSE4 )
2759
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE4, "SSE4" );
2760
    if( cpu_detect & X264_CPU_SSE42 )
2761
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE42, "SSE4.2" );
2762
    if( cpu_detect & X264_CPU_AVX )
2763
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX, "AVX" );
2764
    if( cpu_detect & X264_CPU_XOP )
2765
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_XOP, "XOP" );
2766
    if( cpu_detect & X264_CPU_FMA4 )
2767
    {
2768
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA4, "FMA4" );
2769
        cpu1 &= ~X264_CPU_FMA4;
2770
    }
2771
    if( cpu_detect & X264_CPU_FMA3 )
2772
    {
2773
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA3, "FMA3" );
2774
        cpu1 &= ~X264_CPU_FMA3;
2775
    }
2776
    if( cpu_detect & X264_CPU_AVX2 )
2777
    {
2778
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA3 | X264_CPU_AVX2, "AVX2" );
2779
        if( cpu_detect & X264_CPU_LZCNT )
2780
        {
2781
            ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "AVX2 LZCNT" );
2782
            cpu1 &= ~X264_CPU_LZCNT;
2783
        }
2784
    }
2785
    if( cpu_detect & X264_CPU_BMI1 )
2786
    {
2787
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_BMI1, "BMI1" );
2788
        cpu1 &= ~X264_CPU_BMI1;
2789
    }
2790
    if( cpu_detect & X264_CPU_BMI2 )
2791
    {
2792
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_BMI1|X264_CPU_BMI2, "BMI2" );
2793
        cpu1 &= ~(X264_CPU_BMI1|X264_CPU_BMI2);
2794
    }
2795
#elif ARCH_PPC
2796
    if( cpu_detect & X264_CPU_ALTIVEC )
2797
    {
2798
        fprintf( stderr, "x264: ALTIVEC against C\n" );
2799
        ret = check_all_funcs( 0, X264_CPU_ALTIVEC );
2800
    }
2801
#elif ARCH_ARM
2802
    if( cpu_detect & X264_CPU_NEON )
2803
        x264_checkasm_call = x264_checkasm_call_neon;
2804
    if( cpu_detect & X264_CPU_ARMV6 )
2805
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV6, "ARMv6" );
2806
    if( cpu_detect & X264_CPU_NEON )
2807
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
2808
    if( cpu_detect & X264_CPU_FAST_NEON_MRC )
2809
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_FAST_NEON_MRC, "Fast NEON MRC" );
2810
#elif ARCH_AARCH64
2811
    if( cpu_detect & X264_CPU_ARMV8 )
2812
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV8, "ARMv8" );
2813
    if( cpu_detect & X264_CPU_NEON )
2814
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
2815
#elif ARCH_MIPS
2816
    if( cpu_detect & X264_CPU_MSA )
2817
        ret |= add_flags( &cpu0, &cpu1, X264_CPU_MSA, "MSA" );
2818
#endif
2819
    return ret;
2820
}
2821

2822
int main(int argc, char *argv[])
2823
{
2824
    int ret = 0;
2825

2826
    if( argc > 1 && !strncmp( argv[1], "--bench", 7 ) )
2827
    {
2828
#if !ARCH_X86 && !ARCH_X86_64 && !ARCH_PPC && !ARCH_ARM && !ARCH_AARCH64 && !ARCH_MIPS
2829
        fprintf( stderr, "no --bench for your cpu until you port rdtsc\n" );
2830
        return 1;
2831
#endif
2832
        do_bench = 1;
2833
        if( argv[1][7] == '=' )
2834
        {
2835
            bench_pattern = argv[1]+8;
2836
            bench_pattern_len = strlen(bench_pattern);
2837
        }
2838
        argc--;
2839
        argv++;
2840
    }
2841

2842
    int seed = ( argc > 1 ) ? atoi(argv[1]) : x264_mdate();
2843
    fprintf( stderr, "x264: using random seed %u\n", seed );
2844
    srand( seed );
2845

2846
    buf1 = x264_malloc( 0x1e00 + 0x2000*sizeof(pixel) + 32*BENCH_ALIGNS );
2847
    pbuf1 = x264_malloc( 0x1e00*sizeof(pixel) + 32*BENCH_ALIGNS );
2848
    if( !buf1 || !pbuf1 )
2849
    {
2850
        fprintf( stderr, "malloc failed, unable to initiate tests!\n" );
2851
        return -1;
2852
    }
2853
#define INIT_POINTER_OFFSETS\
2854
    buf2 = buf1 + 0xf00;\
2855
    buf3 = buf2 + 0xf00;\
2856
    buf4 = buf3 + 0x1000*sizeof(pixel);\
2857
    pbuf2 = pbuf1 + 0xf00;\
2858
    pbuf3 = (pixel*)buf3;\
2859
    pbuf4 = (pixel*)buf4;
2860
    INIT_POINTER_OFFSETS;
2861
    for( int i = 0; i < 0x1e00; i++ )
2862
    {
2863
        buf1[i] = rand() & 0xFF;
2864
        pbuf1[i] = rand() & PIXEL_MAX;
2865
    }
2866
    memset( buf1+0x1e00, 0, 0x2000*sizeof(pixel) );
2867

2868
    /* 32-byte alignment is guaranteed whenever it's useful, but some functions also vary in speed depending on %64 */
2869
    if( do_bench )
2870
        for( int i = 0; i < BENCH_ALIGNS && !ret; i++ )
2871
        {
2872
            INIT_POINTER_OFFSETS;
2873
            ret |= x264_stack_pagealign( check_all_flags, i*32 );
2874
            buf1 += 32;
2875
            pbuf1 += 32;
2876
            quiet = 1;
2877
            fprintf( stderr, "%d/%d\r", i+1, BENCH_ALIGNS );
2878
        }
2879
    else
2880
        ret = x264_stack_pagealign( check_all_flags, 0 );
2881

2882
    if( ret )
2883
    {
2884
        fprintf( stderr, "x264: at least one test has failed. Go and fix that Right Now!\n" );
2885
        return -1;
2886
    }
2887
    fprintf( stderr, "x264: All tests passed Yeah :)\n" );
2888
    if( do_bench )
2889
        print_bench();
2890
    return 0;
2891
}
2892

2893

2894
Product

Resources

Company
