2 * Assembly testing and benchmarking tool
3 * Copyright (c) 2015 Henrik Gramner
4 * Copyright (c) 2008 Loren Merritt
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License along
19 * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
20 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
28 #include "libavutil/common.h"
29 #include "libavutil/cpu.h"
30 #include "libavutil/intfloat.h"
31 #include "libavutil/random_seed.h"
37 #if HAVE_SETCONSOLETEXTATTRIBUTE
39 #define COLOR_RED FOREGROUND_RED
40 #define COLOR_GREEN FOREGROUND_GREEN
41 #define COLOR_YELLOW (FOREGROUND_RED|FOREGROUND_GREEN)
45 #define COLOR_YELLOW 3
56 #if ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
57 #include "libavutil/arm/cpu.h"
59 void (*checkasm_checked_call)(void *func, int dummy, ...) = checkasm_checked_call_novfp;
62 /* List of tests to invoke */
68 #if CONFIG_ALAC_DECODER
69 { "alacdsp", checkasm_check_alacdsp },
72 { "bswapdsp", checkasm_check_bswapdsp },
74 #if CONFIG_DCA_DECODER
75 { "synth_filter", checkasm_check_synth_filter },
78 { "flacdsp", checkasm_check_flacdsp },
81 { "fmtconvert", checkasm_check_fmtconvert },
84 { "h264pred", checkasm_check_h264pred },
87 { "h264qpel", checkasm_check_h264qpel },
89 #if CONFIG_JPEG2000_DECODER
90 { "jpeg2000dsp", checkasm_check_jpeg2000dsp },
92 #if CONFIG_PIXBLOCKDSP
93 { "pixblockdsp", checkasm_check_pixblockdsp },
95 #if CONFIG_V210_ENCODER
96 { "v210enc", checkasm_check_v210enc },
98 #if CONFIG_VP9_DECODER
99 { "vp9dsp", checkasm_check_vp9dsp },
102 { "videodsp", checkasm_check_videodsp },
106 #if CONFIG_BLEND_FILTER
107 { "vf_blend", checkasm_check_blend },
113 /* List of cpu flags to check */
114 static const struct {
120 { "ARMV8", "armv8", AV_CPU_FLAG_ARMV8 },
121 { "NEON", "neon", AV_CPU_FLAG_NEON },
123 { "ARMV5TE", "armv5te", AV_CPU_FLAG_ARMV5TE },
124 { "ARMV6", "armv6", AV_CPU_FLAG_ARMV6 },
125 { "ARMV6T2", "armv6t2", AV_CPU_FLAG_ARMV6T2 },
126 { "VFP", "vfp", AV_CPU_FLAG_VFP },
127 { "VFP_VM", "vfp_vm", AV_CPU_FLAG_VFP_VM },
128 { "VFPV3", "vfp3", AV_CPU_FLAG_VFPV3 },
129 { "NEON", "neon", AV_CPU_FLAG_NEON },
131 { "ALTIVEC", "altivec", AV_CPU_FLAG_ALTIVEC },
132 { "VSX", "vsx", AV_CPU_FLAG_VSX },
133 { "POWER8", "power8", AV_CPU_FLAG_POWER8 },
135 { "MMX", "mmx", AV_CPU_FLAG_MMX|AV_CPU_FLAG_CMOV },
136 { "MMXEXT", "mmxext", AV_CPU_FLAG_MMXEXT },
137 { "3DNOW", "3dnow", AV_CPU_FLAG_3DNOW },
138 { "3DNOWEXT", "3dnowext", AV_CPU_FLAG_3DNOWEXT },
139 { "SSE", "sse", AV_CPU_FLAG_SSE },
140 { "SSE2", "sse2", AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW },
141 { "SSE3", "sse3", AV_CPU_FLAG_SSE3|AV_CPU_FLAG_SSE3SLOW },
142 { "SSSE3", "ssse3", AV_CPU_FLAG_SSSE3|AV_CPU_FLAG_ATOM },
143 { "SSE4.1", "sse4", AV_CPU_FLAG_SSE4 },
144 { "SSE4.2", "sse42", AV_CPU_FLAG_SSE42 },
145 { "AES-NI", "aesni", AV_CPU_FLAG_AESNI },
146 { "AVX", "avx", AV_CPU_FLAG_AVX },
147 { "XOP", "xop", AV_CPU_FLAG_XOP },
148 { "FMA3", "fma3", AV_CPU_FLAG_FMA3 },
149 { "FMA4", "fma4", AV_CPU_FLAG_FMA4 },
150 { "AVX2", "avx2", AV_CPU_FLAG_AVX2 },
155 typedef struct CheckasmFuncVersion {
156 struct CheckasmFuncVersion *next;
162 } CheckasmFuncVersion;
164 /* Binary search tree node */
165 typedef struct CheckasmFunc {
166 struct CheckasmFunc *child[2];
167 CheckasmFuncVersion versions;
168 uint8_t color; /* 0 = red, 1 = black */
175 CheckasmFunc *current_func;
176 CheckasmFuncVersion *current_func_ver;
177 const char *current_test_name;
178 const char *bench_pattern;
179 int bench_pattern_len;
184 const char *cpu_flag_name;
190 /* float compare support code */
191 static int is_negative(union av_intfloat32 u)
196 int float_near_ulp(float a, float b, unsigned max_ulp)
198 union av_intfloat32 x, y;
203 if (is_negative(x) != is_negative(y)) {
204 // handle -0.0 == +0.0
208 if (abs(x.i - y.i) <= max_ulp)
214 int float_near_ulp_array(const float *a, const float *b, unsigned max_ulp,
219 for (i = 0; i < len; i++) {
220 if (!float_near_ulp(a[i], b[i], max_ulp))
226 int float_near_abs_eps(float a, float b, float eps)
228 float abs_diff = fabsf(a - b);
230 return abs_diff < eps;
233 int float_near_abs_eps_array(const float *a, const float *b, float eps,
238 for (i = 0; i < len; i++) {
239 if (!float_near_abs_eps(a[i], b[i], eps))
245 int float_near_abs_eps_ulp(float a, float b, float eps, unsigned max_ulp)
247 return float_near_ulp(a, b, max_ulp) || float_near_abs_eps(a, b, eps);
250 int float_near_abs_eps_array_ulp(const float *a, const float *b, float eps,
251 unsigned max_ulp, unsigned len)
255 for (i = 0; i < len; i++) {
256 if (!float_near_abs_eps_ulp(a[i], b[i], eps, max_ulp))
262 /* Print colored text to stderr if the terminal supports it */
263 static void color_printf(int color, const char *fmt, ...)
265 static int use_color = -1;
268 #if HAVE_SETCONSOLETEXTATTRIBUTE
270 static WORD org_attributes;
273 CONSOLE_SCREEN_BUFFER_INFO con_info;
274 con = GetStdHandle(STD_ERROR_HANDLE);
275 if (con && con != INVALID_HANDLE_VALUE && GetConsoleScreenBufferInfo(con, &con_info)) {
276 org_attributes = con_info.wAttributes;
282 SetConsoleTextAttribute(con, (org_attributes & 0xfff0) | (color & 0x0f));
285 const char *term = getenv("TERM");
286 use_color = term && strcmp(term, "dumb") && isatty(2);
289 fprintf(stderr, "\x1b[%d;3%dm", (color & 0x08) >> 3, color & 0x07);
293 vfprintf(stderr, fmt, arg);
297 #if HAVE_SETCONSOLETEXTATTRIBUTE
298 SetConsoleTextAttribute(con, org_attributes);
300 fprintf(stderr, "\x1b[0m");
305 /* Deallocate a tree */
306 static void destroy_func_tree(CheckasmFunc *f)
309 CheckasmFuncVersion *v = f->versions.next;
311 CheckasmFuncVersion *next = v->next;
316 destroy_func_tree(f->child[0]);
317 destroy_func_tree(f->child[1]);
322 /* Allocate a zero-initialized block, clean up and exit on failure */
323 static void *checkasm_malloc(size_t size)
325 void *ptr = calloc(1, size);
327 fprintf(stderr, "checkasm: malloc failed\n");
328 destroy_func_tree(state.funcs);
334 /* Get the suffix of the specified cpu flag */
335 static const char *cpu_suffix(int cpu)
337 int i = FF_ARRAY_ELEMS(cpus);
340 if (cpu & cpus[i].flag)
341 return cpus[i].suffix;
347 static int cmp_nop(const void *a, const void *b)
349 return *(const uint16_t*)a - *(const uint16_t*)b;
352 /* Measure the overhead of the timing code (in decicycles) */
353 static int measure_nop_time(void)
355 uint16_t nops[10000];
358 for (i = 0; i < 10000; i++) {
359 uint64_t t = AV_READ_TIME();
360 nops[i] = AV_READ_TIME() - t;
363 qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
364 for (i = 2500; i < 7500; i++)
367 return nop_sum / 500;
370 /* Print benchmark results */
371 static void print_benchs(CheckasmFunc *f)
374 print_benchs(f->child[0]);
376 /* Only print functions with at least one assembly version */
377 if (f->versions.cpu || f->versions.next) {
378 CheckasmFuncVersion *v = &f->versions;
381 int decicycles = (10*v->cycles/v->iterations - state.nop_time) / 4;
382 printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
384 } while ((v = v->next));
387 print_benchs(f->child[1]);
392 /* ASCIIbetical sort except preserving natural order for numbers */
393 static int cmp_func_names(const char *a, const char *b)
395 const char *start = a;
396 int ascii_diff, digit_diff;
398 for (; !(ascii_diff = *(const unsigned char*)a - *(const unsigned char*)b) && *a; a++, b++);
399 for (; av_isdigit(*a) && av_isdigit(*b); a++, b++);
401 if (a > start && av_isdigit(a[-1]) && (digit_diff = av_isdigit(*a) - av_isdigit(*b)))
407 /* Perform a tree rotation in the specified direction and return the new root */
408 static CheckasmFunc *rotate_tree(CheckasmFunc *f, int dir)
410 CheckasmFunc *r = f->child[dir^1];
411 f->child[dir^1] = r->child[dir];
418 #define is_red(f) ((f) && !(f)->color)
420 /* Balance a left-leaning red-black tree at the specified node */
421 static void balance_tree(CheckasmFunc **root)
423 CheckasmFunc *f = *root;
425 if (is_red(f->child[0]) && is_red(f->child[1])) {
427 f->child[0]->color = f->child[1]->color = 1;
430 if (!is_red(f->child[0]) && is_red(f->child[1]))
431 *root = rotate_tree(f, 0); /* Rotate left */
432 else if (is_red(f->child[0]) && is_red(f->child[0]->child[0]))
433 *root = rotate_tree(f, 1); /* Rotate right */
436 /* Get a node with the specified name, creating it if it doesn't exist */
437 static CheckasmFunc *get_func(CheckasmFunc **root, const char *name)
439 CheckasmFunc *f = *root;
442 /* Search the tree for a matching node */
443 int cmp = cmp_func_names(name, f->name);
445 f = get_func(&f->child[cmp > 0], name);
447 /* Rebalance the tree on the way up if a new node was inserted */
448 if (!f->versions.func)
452 /* Allocate and insert a new node into the tree */
453 int name_length = strlen(name);
454 f = *root = checkasm_malloc(sizeof(CheckasmFunc) + name_length);
455 memcpy(f->name, name, name_length + 1);
461 /* Perform tests and benchmarks for the specified cpu flag if supported by the host */
462 static void check_cpu_flag(const char *name, int flag)
464 int old_cpu_flag = state.cpu_flag;
466 flag |= old_cpu_flag;
467 av_force_cpu_flags(-1);
468 state.cpu_flag = flag & av_get_cpu_flags();
469 av_force_cpu_flags(state.cpu_flag);
471 if (!flag || state.cpu_flag != old_cpu_flag) {
474 state.cpu_flag_name = name;
475 for (i = 0; tests[i].func; i++) {
476 state.current_test_name = tests[i].name;
482 /* Print the name of the current CPU flag, but only do it once */
483 static void print_cpu_name(void)
485 if (state.cpu_flag_name) {
486 color_printf(COLOR_YELLOW, "%s:\n", state.cpu_flag_name);
487 state.cpu_flag_name = NULL;
491 int main(int argc, char *argv[])
493 int i, seed, ret = 0;
495 #if ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
496 if (have_vfp(av_get_cpu_flags()) || have_neon(av_get_cpu_flags()))
497 checkasm_checked_call = checkasm_checked_call_vfp;
500 if (!tests[0].func || !cpus[0].flag) {
501 fprintf(stderr, "checkasm: no tests to perform\n");
505 if (argc > 1 && !strncmp(argv[1], "--bench", 7)) {
507 fprintf(stderr, "checkasm: --bench is not supported on your system\n");
510 if (argv[1][7] == '=') {
511 state.bench_pattern = argv[1] + 8;
512 state.bench_pattern_len = strlen(state.bench_pattern);
514 state.bench_pattern = "";
520 seed = (argc > 1) ? atoi(argv[1]) : av_get_random_seed();
521 fprintf(stderr, "checkasm: using random seed %u\n", seed);
522 av_lfg_init(&checkasm_lfg, seed);
524 check_cpu_flag(NULL, 0);
525 for (i = 0; cpus[i].flag; i++)
526 check_cpu_flag(cpus[i].name, cpus[i].flag);
528 if (state.num_failed) {
529 fprintf(stderr, "checkasm: %d of %d tests have failed\n", state.num_failed, state.num_checked);
532 fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
534 if (state.bench_pattern) {
535 state.nop_time = measure_nop_time();
536 printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
537 print_benchs(state.funcs);
542 destroy_func_tree(state.funcs);
546 /* Decide whether or not the specified function needs to be tested and
547 * allocate/initialize data structures if needed. Returns a pointer to a
548 * reference function if the function should be tested, otherwise NULL */
549 void *checkasm_check_func(void *func, const char *name, ...)
553 CheckasmFuncVersion *v;
558 name_length = vsnprintf(name_buf, sizeof(name_buf), name, arg);
561 if (!func || name_length <= 0 || name_length >= sizeof(name_buf))
564 state.current_func = get_func(&state.funcs, name_buf);
565 state.funcs->color = 1;
566 v = &state.current_func->versions;
569 CheckasmFuncVersion *prev;
571 /* Only test functions that haven't already been tested */
579 } while ((v = v->next));
581 v = prev->next = checkasm_malloc(sizeof(CheckasmFuncVersion));
586 v->cpu = state.cpu_flag;
587 state.current_func_ver = v;
595 /* Decide whether or not the current function needs to be benchmarked */
596 int checkasm_bench_func(void)
598 return !state.num_failed && state.bench_pattern &&
599 !strncmp(state.current_func->name, state.bench_pattern, state.bench_pattern_len);
602 /* Indicate that the current test has failed */
603 void checkasm_fail_func(const char *msg, ...)
605 if (state.current_func_ver->cpu && state.current_func_ver->ok) {
609 fprintf(stderr, " %s_%s (", state.current_func->name, cpu_suffix(state.current_func_ver->cpu));
611 vfprintf(stderr, msg, arg);
613 fprintf(stderr, ")\n");
615 state.current_func_ver->ok = 0;
620 /* Update benchmark results of the current function */
621 void checkasm_update_bench(int iterations, uint64_t cycles)
623 state.current_func_ver->iterations += iterations;
624 state.current_func_ver->cycles += cycles;
627 /* Print the outcome of all tests performed since the last time this function was called */
628 void checkasm_report(const char *name, ...)
630 static int prev_checked, prev_failed, max_length;
632 if (state.num_checked > prev_checked) {
633 int pad_length = max_length + 4;
637 pad_length -= fprintf(stderr, " - %s.", state.current_test_name);
639 pad_length -= vfprintf(stderr, name, arg);
641 fprintf(stderr, "%*c", FFMAX(pad_length, 0) + 2, '[');
643 if (state.num_failed == prev_failed)
644 color_printf(COLOR_GREEN, "OK");
646 color_printf(COLOR_RED, "FAILED");
647 fprintf(stderr, "]\n");
649 prev_checked = state.num_checked;
650 prev_failed = state.num_failed;
651 } else if (!state.cpu_flag) {
652 /* Calculate the amount of padding required to make the output vertically aligned */
653 int length = strlen(state.current_test_name);
657 length += vsnprintf(NULL, 0, name, arg);
660 if (length > max_length)