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 { "h264dsp", checkasm_check_h264dsp },
87 { "h264pred", checkasm_check_h264pred },
90 { "h264qpel", checkasm_check_h264qpel },
92 #if CONFIG_JPEG2000_DECODER
93 { "jpeg2000dsp", checkasm_check_jpeg2000dsp },
95 #if CONFIG_PIXBLOCKDSP
96 { "pixblockdsp", checkasm_check_pixblockdsp },
98 #if CONFIG_V210_ENCODER
99 { "v210enc", checkasm_check_v210enc },
101 #if CONFIG_VP9_DECODER
102 { "vp9dsp", checkasm_check_vp9dsp },
105 { "videodsp", checkasm_check_videodsp },
109 #if CONFIG_BLEND_FILTER
110 { "vf_blend", checkasm_check_blend },
112 #if CONFIG_COLORSPACE_FILTER
113 { "vf_colorspace", checkasm_check_colorspace },
119 /* List of cpu flags to check */
120 static const struct {
126 { "ARMV8", "armv8", AV_CPU_FLAG_ARMV8 },
127 { "NEON", "neon", AV_CPU_FLAG_NEON },
129 { "ARMV5TE", "armv5te", AV_CPU_FLAG_ARMV5TE },
130 { "ARMV6", "armv6", AV_CPU_FLAG_ARMV6 },
131 { "ARMV6T2", "armv6t2", AV_CPU_FLAG_ARMV6T2 },
132 { "VFP", "vfp", AV_CPU_FLAG_VFP },
133 { "VFP_VM", "vfp_vm", AV_CPU_FLAG_VFP_VM },
134 { "VFPV3", "vfp3", AV_CPU_FLAG_VFPV3 },
135 { "NEON", "neon", AV_CPU_FLAG_NEON },
137 { "ALTIVEC", "altivec", AV_CPU_FLAG_ALTIVEC },
138 { "VSX", "vsx", AV_CPU_FLAG_VSX },
139 { "POWER8", "power8", AV_CPU_FLAG_POWER8 },
141 { "MMX", "mmx", AV_CPU_FLAG_MMX|AV_CPU_FLAG_CMOV },
142 { "MMXEXT", "mmxext", AV_CPU_FLAG_MMXEXT },
143 { "3DNOW", "3dnow", AV_CPU_FLAG_3DNOW },
144 { "3DNOWEXT", "3dnowext", AV_CPU_FLAG_3DNOWEXT },
145 { "SSE", "sse", AV_CPU_FLAG_SSE },
146 { "SSE2", "sse2", AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW },
147 { "SSE3", "sse3", AV_CPU_FLAG_SSE3|AV_CPU_FLAG_SSE3SLOW },
148 { "SSSE3", "ssse3", AV_CPU_FLAG_SSSE3|AV_CPU_FLAG_ATOM },
149 { "SSE4.1", "sse4", AV_CPU_FLAG_SSE4 },
150 { "SSE4.2", "sse42", AV_CPU_FLAG_SSE42 },
151 { "AES-NI", "aesni", AV_CPU_FLAG_AESNI },
152 { "AVX", "avx", AV_CPU_FLAG_AVX },
153 { "XOP", "xop", AV_CPU_FLAG_XOP },
154 { "FMA3", "fma3", AV_CPU_FLAG_FMA3 },
155 { "FMA4", "fma4", AV_CPU_FLAG_FMA4 },
156 { "AVX2", "avx2", AV_CPU_FLAG_AVX2 },
161 typedef struct CheckasmFuncVersion {
162 struct CheckasmFuncVersion *next;
168 } CheckasmFuncVersion;
170 /* Binary search tree node */
171 typedef struct CheckasmFunc {
172 struct CheckasmFunc *child[2];
173 CheckasmFuncVersion versions;
174 uint8_t color; /* 0 = red, 1 = black */
181 CheckasmFunc *current_func;
182 CheckasmFuncVersion *current_func_ver;
183 const char *current_test_name;
184 const char *bench_pattern;
185 int bench_pattern_len;
190 const char *cpu_flag_name;
196 /* float compare support code */
197 static int is_negative(union av_intfloat32 u)
202 int float_near_ulp(float a, float b, unsigned max_ulp)
204 union av_intfloat32 x, y;
209 if (is_negative(x) != is_negative(y)) {
210 // handle -0.0 == +0.0
214 if (abs(x.i - y.i) <= max_ulp)
220 int float_near_ulp_array(const float *a, const float *b, unsigned max_ulp,
225 for (i = 0; i < len; i++) {
226 if (!float_near_ulp(a[i], b[i], max_ulp))
232 int float_near_abs_eps(float a, float b, float eps)
234 float abs_diff = fabsf(a - b);
236 return abs_diff < eps;
239 int float_near_abs_eps_array(const float *a, const float *b, float eps,
244 for (i = 0; i < len; i++) {
245 if (!float_near_abs_eps(a[i], b[i], eps))
251 int float_near_abs_eps_ulp(float a, float b, float eps, unsigned max_ulp)
253 return float_near_ulp(a, b, max_ulp) || float_near_abs_eps(a, b, eps);
256 int float_near_abs_eps_array_ulp(const float *a, const float *b, float eps,
257 unsigned max_ulp, unsigned len)
261 for (i = 0; i < len; i++) {
262 if (!float_near_abs_eps_ulp(a[i], b[i], eps, max_ulp))
268 /* Print colored text to stderr if the terminal supports it */
269 static void color_printf(int color, const char *fmt, ...)
271 static int use_color = -1;
274 #if HAVE_SETCONSOLETEXTATTRIBUTE
276 static WORD org_attributes;
279 CONSOLE_SCREEN_BUFFER_INFO con_info;
280 con = GetStdHandle(STD_ERROR_HANDLE);
281 if (con && con != INVALID_HANDLE_VALUE && GetConsoleScreenBufferInfo(con, &con_info)) {
282 org_attributes = con_info.wAttributes;
288 SetConsoleTextAttribute(con, (org_attributes & 0xfff0) | (color & 0x0f));
291 const char *term = getenv("TERM");
292 use_color = term && strcmp(term, "dumb") && isatty(2);
295 fprintf(stderr, "\x1b[%d;3%dm", (color & 0x08) >> 3, color & 0x07);
299 vfprintf(stderr, fmt, arg);
303 #if HAVE_SETCONSOLETEXTATTRIBUTE
304 SetConsoleTextAttribute(con, org_attributes);
306 fprintf(stderr, "\x1b[0m");
311 /* Deallocate a tree */
312 static void destroy_func_tree(CheckasmFunc *f)
315 CheckasmFuncVersion *v = f->versions.next;
317 CheckasmFuncVersion *next = v->next;
322 destroy_func_tree(f->child[0]);
323 destroy_func_tree(f->child[1]);
328 /* Allocate a zero-initialized block, clean up and exit on failure */
329 static void *checkasm_malloc(size_t size)
331 void *ptr = calloc(1, size);
333 fprintf(stderr, "checkasm: malloc failed\n");
334 destroy_func_tree(state.funcs);
340 /* Get the suffix of the specified cpu flag */
341 static const char *cpu_suffix(int cpu)
343 int i = FF_ARRAY_ELEMS(cpus);
346 if (cpu & cpus[i].flag)
347 return cpus[i].suffix;
353 static int cmp_nop(const void *a, const void *b)
355 return *(const uint16_t*)a - *(const uint16_t*)b;
358 /* Measure the overhead of the timing code (in decicycles) */
359 static int measure_nop_time(void)
361 uint16_t nops[10000];
364 for (i = 0; i < 10000; i++) {
365 uint64_t t = AV_READ_TIME();
366 nops[i] = AV_READ_TIME() - t;
369 qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
370 for (i = 2500; i < 7500; i++)
373 return nop_sum / 500;
376 /* Print benchmark results */
377 static void print_benchs(CheckasmFunc *f)
380 print_benchs(f->child[0]);
382 /* Only print functions with at least one assembly version */
383 if (f->versions.cpu || f->versions.next) {
384 CheckasmFuncVersion *v = &f->versions;
387 int decicycles = (10*v->cycles/v->iterations - state.nop_time) / 4;
388 printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
390 } while ((v = v->next));
393 print_benchs(f->child[1]);
398 /* ASCIIbetical sort except preserving natural order for numbers */
399 static int cmp_func_names(const char *a, const char *b)
401 const char *start = a;
402 int ascii_diff, digit_diff;
404 for (; !(ascii_diff = *(const unsigned char*)a - *(const unsigned char*)b) && *a; a++, b++);
405 for (; av_isdigit(*a) && av_isdigit(*b); a++, b++);
407 if (a > start && av_isdigit(a[-1]) && (digit_diff = av_isdigit(*a) - av_isdigit(*b)))
413 /* Perform a tree rotation in the specified direction and return the new root */
414 static CheckasmFunc *rotate_tree(CheckasmFunc *f, int dir)
416 CheckasmFunc *r = f->child[dir^1];
417 f->child[dir^1] = r->child[dir];
424 #define is_red(f) ((f) && !(f)->color)
426 /* Balance a left-leaning red-black tree at the specified node */
427 static void balance_tree(CheckasmFunc **root)
429 CheckasmFunc *f = *root;
431 if (is_red(f->child[0]) && is_red(f->child[1])) {
433 f->child[0]->color = f->child[1]->color = 1;
436 if (!is_red(f->child[0]) && is_red(f->child[1]))
437 *root = rotate_tree(f, 0); /* Rotate left */
438 else if (is_red(f->child[0]) && is_red(f->child[0]->child[0]))
439 *root = rotate_tree(f, 1); /* Rotate right */
442 /* Get a node with the specified name, creating it if it doesn't exist */
443 static CheckasmFunc *get_func(CheckasmFunc **root, const char *name)
445 CheckasmFunc *f = *root;
448 /* Search the tree for a matching node */
449 int cmp = cmp_func_names(name, f->name);
451 f = get_func(&f->child[cmp > 0], name);
453 /* Rebalance the tree on the way up if a new node was inserted */
454 if (!f->versions.func)
458 /* Allocate and insert a new node into the tree */
459 int name_length = strlen(name);
460 f = *root = checkasm_malloc(sizeof(CheckasmFunc) + name_length);
461 memcpy(f->name, name, name_length + 1);
467 /* Perform tests and benchmarks for the specified cpu flag if supported by the host */
468 static void check_cpu_flag(const char *name, int flag)
470 int old_cpu_flag = state.cpu_flag;
472 flag |= old_cpu_flag;
473 av_force_cpu_flags(-1);
474 state.cpu_flag = flag & av_get_cpu_flags();
475 av_force_cpu_flags(state.cpu_flag);
477 if (!flag || state.cpu_flag != old_cpu_flag) {
480 state.cpu_flag_name = name;
481 for (i = 0; tests[i].func; i++) {
482 state.current_test_name = tests[i].name;
488 /* Print the name of the current CPU flag, but only do it once */
489 static void print_cpu_name(void)
491 if (state.cpu_flag_name) {
492 color_printf(COLOR_YELLOW, "%s:\n", state.cpu_flag_name);
493 state.cpu_flag_name = NULL;
497 int main(int argc, char *argv[])
499 int i, seed, ret = 0;
501 #if ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
502 if (have_vfp(av_get_cpu_flags()) || have_neon(av_get_cpu_flags()))
503 checkasm_checked_call = checkasm_checked_call_vfp;
506 if (!tests[0].func || !cpus[0].flag) {
507 fprintf(stderr, "checkasm: no tests to perform\n");
511 if (argc > 1 && !strncmp(argv[1], "--bench", 7)) {
513 fprintf(stderr, "checkasm: --bench is not supported on your system\n");
516 if (argv[1][7] == '=') {
517 state.bench_pattern = argv[1] + 8;
518 state.bench_pattern_len = strlen(state.bench_pattern);
520 state.bench_pattern = "";
526 seed = (argc > 1) ? atoi(argv[1]) : av_get_random_seed();
527 fprintf(stderr, "checkasm: using random seed %u\n", seed);
528 av_lfg_init(&checkasm_lfg, seed);
530 check_cpu_flag(NULL, 0);
531 for (i = 0; cpus[i].flag; i++)
532 check_cpu_flag(cpus[i].name, cpus[i].flag);
534 if (state.num_failed) {
535 fprintf(stderr, "checkasm: %d of %d tests have failed\n", state.num_failed, state.num_checked);
538 fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
540 if (state.bench_pattern) {
541 state.nop_time = measure_nop_time();
542 printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
543 print_benchs(state.funcs);
548 destroy_func_tree(state.funcs);
552 /* Decide whether or not the specified function needs to be tested and
553 * allocate/initialize data structures if needed. Returns a pointer to a
554 * reference function if the function should be tested, otherwise NULL */
555 void *checkasm_check_func(void *func, const char *name, ...)
559 CheckasmFuncVersion *v;
564 name_length = vsnprintf(name_buf, sizeof(name_buf), name, arg);
567 if (!func || name_length <= 0 || name_length >= sizeof(name_buf))
570 state.current_func = get_func(&state.funcs, name_buf);
571 state.funcs->color = 1;
572 v = &state.current_func->versions;
575 CheckasmFuncVersion *prev;
577 /* Only test functions that haven't already been tested */
585 } while ((v = v->next));
587 v = prev->next = checkasm_malloc(sizeof(CheckasmFuncVersion));
592 v->cpu = state.cpu_flag;
593 state.current_func_ver = v;
601 /* Decide whether or not the current function needs to be benchmarked */
602 int checkasm_bench_func(void)
604 return !state.num_failed && state.bench_pattern &&
605 !strncmp(state.current_func->name, state.bench_pattern, state.bench_pattern_len);
608 /* Indicate that the current test has failed */
609 void checkasm_fail_func(const char *msg, ...)
611 if (state.current_func_ver->cpu && state.current_func_ver->ok) {
615 fprintf(stderr, " %s_%s (", state.current_func->name, cpu_suffix(state.current_func_ver->cpu));
617 vfprintf(stderr, msg, arg);
619 fprintf(stderr, ")\n");
621 state.current_func_ver->ok = 0;
626 /* Update benchmark results of the current function */
627 void checkasm_update_bench(int iterations, uint64_t cycles)
629 state.current_func_ver->iterations += iterations;
630 state.current_func_ver->cycles += cycles;
633 /* Print the outcome of all tests performed since the last time this function was called */
634 void checkasm_report(const char *name, ...)
636 static int prev_checked, prev_failed, max_length;
638 if (state.num_checked > prev_checked) {
639 int pad_length = max_length + 4;
643 pad_length -= fprintf(stderr, " - %s.", state.current_test_name);
645 pad_length -= vfprintf(stderr, name, arg);
647 fprintf(stderr, "%*c", FFMAX(pad_length, 0) + 2, '[');
649 if (state.num_failed == prev_failed)
650 color_printf(COLOR_GREEN, "OK");
652 color_printf(COLOR_RED, "FAILED");
653 fprintf(stderr, "]\n");
655 prev_checked = state.num_checked;
656 prev_failed = state.num_failed;
657 } else if (!state.cpu_flag) {
658 /* Calculate the amount of padding required to make the output vertically aligned */
659 int length = strlen(state.current_test_name);
663 length += vsnprintf(NULL, 0, name, arg);
666 if (length > max_length)