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.
27 # define _GNU_SOURCE // for syscall (performance monitoring API)
36 #include "libavutil/common.h"
37 #include "libavutil/cpu.h"
38 #include "libavutil/intfloat.h"
39 #include "libavutil/random_seed.h"
45 #if HAVE_SETCONSOLETEXTATTRIBUTE
47 #define COLOR_RED FOREGROUND_RED
48 #define COLOR_GREEN FOREGROUND_GREEN
49 #define COLOR_YELLOW (FOREGROUND_RED|FOREGROUND_GREEN)
53 #define COLOR_YELLOW 3
64 #if ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
65 #include "libavutil/arm/cpu.h"
67 void (*checkasm_checked_call)(void *func, int dummy, ...) = checkasm_checked_call_novfp;
70 /* List of tests to invoke */
76 #if CONFIG_AAC_DECODER
77 { "aacpsdsp", checkasm_check_aacpsdsp },
78 { "sbrdsp", checkasm_check_sbrdsp },
80 #if CONFIG_ALAC_DECODER
81 { "alacdsp", checkasm_check_alacdsp },
84 { "audiodsp", checkasm_check_audiodsp },
87 { "blockdsp", checkasm_check_blockdsp },
90 { "bswapdsp", checkasm_check_bswapdsp },
92 #if CONFIG_DCA_DECODER
93 { "synth_filter", checkasm_check_synth_filter },
95 #if CONFIG_EXR_DECODER
96 { "exrdsp", checkasm_check_exrdsp },
99 { "flacdsp", checkasm_check_flacdsp },
101 #if CONFIG_FMTCONVERT
102 { "fmtconvert", checkasm_check_fmtconvert },
105 { "g722dsp", checkasm_check_g722dsp },
108 { "h264dsp", checkasm_check_h264dsp },
111 { "h264pred", checkasm_check_h264pred },
114 { "h264qpel", checkasm_check_h264qpel },
116 #if CONFIG_HEVC_DECODER
117 { "hevc_add_res", checkasm_check_hevc_add_res },
118 { "hevc_idct", checkasm_check_hevc_idct },
120 #if CONFIG_HUFFYUV_DECODER
121 { "huffyuvdsp", checkasm_check_huffyuvdsp },
123 #if CONFIG_JPEG2000_DECODER
124 { "jpeg2000dsp", checkasm_check_jpeg2000dsp },
126 #if CONFIG_HUFFYUVDSP
127 { "llviddsp", checkasm_check_llviddsp },
129 #if CONFIG_LLVIDENCDSP
130 { "llviddspenc", checkasm_check_llviddspenc },
132 #if CONFIG_PIXBLOCKDSP
133 { "pixblockdsp", checkasm_check_pixblockdsp },
135 #if CONFIG_UTVIDEO_DECODER
136 { "utvideodsp", checkasm_check_utvideodsp },
138 #if CONFIG_V210_ENCODER
139 { "v210enc", checkasm_check_v210enc },
142 { "vp8dsp", checkasm_check_vp8dsp },
144 #if CONFIG_VP9_DECODER
145 { "vp9dsp", checkasm_check_vp9dsp },
148 { "videodsp", checkasm_check_videodsp },
152 #if CONFIG_BLEND_FILTER
153 { "vf_blend", checkasm_check_blend },
155 #if CONFIG_COLORSPACE_FILTER
156 { "vf_colorspace", checkasm_check_colorspace },
158 #if CONFIG_HFLIP_FILTER
159 { "vf_hflip", checkasm_check_vf_hflip },
161 #if CONFIG_THRESHOLD_FILTER
162 { "vf_threshold", checkasm_check_vf_threshold },
166 { "fixed_dsp", checkasm_check_fixed_dsp },
167 { "float_dsp", checkasm_check_float_dsp },
172 /* List of cpu flags to check */
173 static const struct {
179 { "ARMV8", "armv8", AV_CPU_FLAG_ARMV8 },
180 { "NEON", "neon", AV_CPU_FLAG_NEON },
182 { "ARMV5TE", "armv5te", AV_CPU_FLAG_ARMV5TE },
183 { "ARMV6", "armv6", AV_CPU_FLAG_ARMV6 },
184 { "ARMV6T2", "armv6t2", AV_CPU_FLAG_ARMV6T2 },
185 { "VFP", "vfp", AV_CPU_FLAG_VFP },
186 { "VFP_VM", "vfp_vm", AV_CPU_FLAG_VFP_VM },
187 { "VFPV3", "vfp3", AV_CPU_FLAG_VFPV3 },
188 { "NEON", "neon", AV_CPU_FLAG_NEON },
190 { "ALTIVEC", "altivec", AV_CPU_FLAG_ALTIVEC },
191 { "VSX", "vsx", AV_CPU_FLAG_VSX },
192 { "POWER8", "power8", AV_CPU_FLAG_POWER8 },
194 { "MMX", "mmx", AV_CPU_FLAG_MMX|AV_CPU_FLAG_CMOV },
195 { "MMXEXT", "mmxext", AV_CPU_FLAG_MMXEXT },
196 { "3DNOW", "3dnow", AV_CPU_FLAG_3DNOW },
197 { "3DNOWEXT", "3dnowext", AV_CPU_FLAG_3DNOWEXT },
198 { "SSE", "sse", AV_CPU_FLAG_SSE },
199 { "SSE2", "sse2", AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW },
200 { "SSE3", "sse3", AV_CPU_FLAG_SSE3|AV_CPU_FLAG_SSE3SLOW },
201 { "SSSE3", "ssse3", AV_CPU_FLAG_SSSE3|AV_CPU_FLAG_ATOM },
202 { "SSE4.1", "sse4", AV_CPU_FLAG_SSE4 },
203 { "SSE4.2", "sse42", AV_CPU_FLAG_SSE42 },
204 { "AES-NI", "aesni", AV_CPU_FLAG_AESNI },
205 { "AVX", "avx", AV_CPU_FLAG_AVX },
206 { "XOP", "xop", AV_CPU_FLAG_XOP },
207 { "FMA3", "fma3", AV_CPU_FLAG_FMA3 },
208 { "FMA4", "fma4", AV_CPU_FLAG_FMA4 },
209 { "AVX2", "avx2", AV_CPU_FLAG_AVX2 },
210 { "AVX-512", "avx512", AV_CPU_FLAG_AVX512 },
215 typedef struct CheckasmFuncVersion {
216 struct CheckasmFuncVersion *next;
221 } CheckasmFuncVersion;
223 /* Binary search tree node */
224 typedef struct CheckasmFunc {
225 struct CheckasmFunc *child[2];
226 CheckasmFuncVersion versions;
227 uint8_t color; /* 0 = red, 1 = black */
234 CheckasmFunc *current_func;
235 CheckasmFuncVersion *current_func_ver;
236 const char *current_test_name;
237 const char *bench_pattern;
238 int bench_pattern_len;
247 const char *cpu_flag_name;
248 const char *test_name;
254 /* float compare support code */
255 static int is_negative(union av_intfloat32 u)
260 int float_near_ulp(float a, float b, unsigned max_ulp)
262 union av_intfloat32 x, y;
267 if (is_negative(x) != is_negative(y)) {
268 // handle -0.0 == +0.0
272 if (llabs((int64_t)x.i - y.i) <= max_ulp)
278 int float_near_ulp_array(const float *a, const float *b, unsigned max_ulp,
283 for (i = 0; i < len; i++) {
284 if (!float_near_ulp(a[i], b[i], max_ulp))
290 int float_near_abs_eps(float a, float b, float eps)
292 float abs_diff = fabsf(a - b);
294 return abs_diff < eps;
297 int float_near_abs_eps_array(const float *a, const float *b, float eps,
302 for (i = 0; i < len; i++) {
303 if (!float_near_abs_eps(a[i], b[i], eps))
309 int float_near_abs_eps_ulp(float a, float b, float eps, unsigned max_ulp)
311 return float_near_ulp(a, b, max_ulp) || float_near_abs_eps(a, b, eps);
314 int float_near_abs_eps_array_ulp(const float *a, const float *b, float eps,
315 unsigned max_ulp, unsigned len)
319 for (i = 0; i < len; i++) {
320 if (!float_near_abs_eps_ulp(a[i], b[i], eps, max_ulp))
326 int double_near_abs_eps(double a, double b, double eps)
328 double abs_diff = fabs(a - b);
330 return abs_diff < eps;
333 int double_near_abs_eps_array(const double *a, const double *b, double eps,
338 for (i = 0; i < len; i++) {
339 if (!double_near_abs_eps(a[i], b[i], eps))
345 /* Print colored text to stderr if the terminal supports it */
346 static void color_printf(int color, const char *fmt, ...)
348 static int use_color = -1;
351 #if HAVE_SETCONSOLETEXTATTRIBUTE
353 static WORD org_attributes;
356 CONSOLE_SCREEN_BUFFER_INFO con_info;
357 con = GetStdHandle(STD_ERROR_HANDLE);
358 if (con && con != INVALID_HANDLE_VALUE && GetConsoleScreenBufferInfo(con, &con_info)) {
359 org_attributes = con_info.wAttributes;
365 SetConsoleTextAttribute(con, (org_attributes & 0xfff0) | (color & 0x0f));
368 const char *term = getenv("TERM");
369 use_color = term && strcmp(term, "dumb") && isatty(2);
372 fprintf(stderr, "\x1b[%d;3%dm", (color & 0x08) >> 3, color & 0x07);
376 vfprintf(stderr, fmt, arg);
380 #if HAVE_SETCONSOLETEXTATTRIBUTE
381 SetConsoleTextAttribute(con, org_attributes);
383 fprintf(stderr, "\x1b[0m");
388 /* Deallocate a tree */
389 static void destroy_func_tree(CheckasmFunc *f)
392 CheckasmFuncVersion *v = f->versions.next;
394 CheckasmFuncVersion *next = v->next;
399 destroy_func_tree(f->child[0]);
400 destroy_func_tree(f->child[1]);
405 /* Allocate a zero-initialized block, clean up and exit on failure */
406 static void *checkasm_malloc(size_t size)
408 void *ptr = calloc(1, size);
410 fprintf(stderr, "checkasm: malloc failed\n");
411 destroy_func_tree(state.funcs);
417 /* Get the suffix of the specified cpu flag */
418 static const char *cpu_suffix(int cpu)
420 int i = FF_ARRAY_ELEMS(cpus);
423 if (cpu & cpus[i].flag)
424 return cpus[i].suffix;
429 static int cmp_nop(const void *a, const void *b)
431 return *(const uint16_t*)a - *(const uint16_t*)b;
434 /* Measure the overhead of the timing code (in decicycles) */
435 static int measure_nop_time(void)
437 uint16_t nops[10000];
439 av_unused const int sysfd = state.sysfd;
442 for (i = 0; i < 10000; i++) {
448 qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
449 for (i = 2500; i < 7500; i++)
452 return nop_sum / 500;
455 /* Print benchmark results */
456 static void print_benchs(CheckasmFunc *f)
459 print_benchs(f->child[0]);
461 /* Only print functions with at least one assembly version */
462 if (f->versions.cpu || f->versions.next) {
463 CheckasmFuncVersion *v = &f->versions;
465 CheckasmPerf *p = &v->perf;
467 int decicycles = (10*p->cycles/p->iterations - state.nop_time) / 4;
468 printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
470 } while ((v = v->next));
473 print_benchs(f->child[1]);
477 /* ASCIIbetical sort except preserving natural order for numbers */
478 static int cmp_func_names(const char *a, const char *b)
480 const char *start = a;
481 int ascii_diff, digit_diff;
483 for (; !(ascii_diff = *(const unsigned char*)a - *(const unsigned char*)b) && *a; a++, b++);
484 for (; av_isdigit(*a) && av_isdigit(*b); a++, b++);
486 if (a > start && av_isdigit(a[-1]) && (digit_diff = av_isdigit(*a) - av_isdigit(*b)))
492 /* Perform a tree rotation in the specified direction and return the new root */
493 static CheckasmFunc *rotate_tree(CheckasmFunc *f, int dir)
495 CheckasmFunc *r = f->child[dir^1];
496 f->child[dir^1] = r->child[dir];
503 #define is_red(f) ((f) && !(f)->color)
505 /* Balance a left-leaning red-black tree at the specified node */
506 static void balance_tree(CheckasmFunc **root)
508 CheckasmFunc *f = *root;
510 if (is_red(f->child[0]) && is_red(f->child[1])) {
512 f->child[0]->color = f->child[1]->color = 1;
515 if (!is_red(f->child[0]) && is_red(f->child[1]))
516 *root = rotate_tree(f, 0); /* Rotate left */
517 else if (is_red(f->child[0]) && is_red(f->child[0]->child[0]))
518 *root = rotate_tree(f, 1); /* Rotate right */
521 /* Get a node with the specified name, creating it if it doesn't exist */
522 static CheckasmFunc *get_func(CheckasmFunc **root, const char *name)
524 CheckasmFunc *f = *root;
527 /* Search the tree for a matching node */
528 int cmp = cmp_func_names(name, f->name);
530 f = get_func(&f->child[cmp > 0], name);
532 /* Rebalance the tree on the way up if a new node was inserted */
533 if (!f->versions.func)
537 /* Allocate and insert a new node into the tree */
538 int name_length = strlen(name);
539 f = *root = checkasm_malloc(sizeof(CheckasmFunc) + name_length);
540 memcpy(f->name, name, name_length + 1);
546 /* Perform tests and benchmarks for the specified cpu flag if supported by the host */
547 static void check_cpu_flag(const char *name, int flag)
549 int old_cpu_flag = state.cpu_flag;
551 flag |= old_cpu_flag;
552 av_force_cpu_flags(-1);
553 state.cpu_flag = flag & av_get_cpu_flags();
554 av_force_cpu_flags(state.cpu_flag);
556 if (!flag || state.cpu_flag != old_cpu_flag) {
559 state.cpu_flag_name = name;
560 for (i = 0; tests[i].func; i++) {
561 if (state.test_name && strcmp(tests[i].name, state.test_name))
563 state.current_test_name = tests[i].name;
569 /* Print the name of the current CPU flag, but only do it once */
570 static void print_cpu_name(void)
572 if (state.cpu_flag_name) {
573 color_printf(COLOR_YELLOW, "%s:\n", state.cpu_flag_name);
574 state.cpu_flag_name = NULL;
578 #if CONFIG_LINUX_PERF
579 static int bench_init_linux(void)
581 struct perf_event_attr attr = {
582 .type = PERF_TYPE_HARDWARE,
583 .size = sizeof(struct perf_event_attr),
584 .config = PERF_COUNT_HW_CPU_CYCLES,
585 .disabled = 1, // start counting only on demand
590 printf("benchmarking with Linux Perf Monitoring API\n");
592 state.sysfd = syscall(__NR_perf_event_open, &attr, 0, -1, -1, 0);
593 if (state.sysfd == -1) {
601 static int bench_init_ffmpeg(void)
604 printf("benchmarking with native FFmpeg timers\n");
607 fprintf(stderr, "checkasm: --bench is not supported on your system\n");
612 static int bench_init(void)
614 #if CONFIG_LINUX_PERF
615 int ret = bench_init_linux();
617 int ret = bench_init_ffmpeg();
622 state.nop_time = measure_nop_time();
623 printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
627 static void bench_uninit(void)
629 #if CONFIG_LINUX_PERF
635 int main(int argc, char *argv[])
637 unsigned int seed = av_get_random_seed();
640 #if ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
641 if (have_vfp(av_get_cpu_flags()) || have_neon(av_get_cpu_flags()))
642 checkasm_checked_call = checkasm_checked_call_vfp;
645 if (!tests[0].func || !cpus[0].flag) {
646 fprintf(stderr, "checkasm: no tests to perform\n");
651 if (!strncmp(argv[1], "--bench", 7)) {
652 if (bench_init() < 0)
654 if (argv[1][7] == '=') {
655 state.bench_pattern = argv[1] + 8;
656 state.bench_pattern_len = strlen(state.bench_pattern);
658 state.bench_pattern = "";
659 } else if (!strncmp(argv[1], "--test=", 7)) {
660 state.test_name = argv[1] + 7;
662 seed = strtoul(argv[1], NULL, 10);
669 fprintf(stderr, "checkasm: using random seed %u\n", seed);
670 av_lfg_init(&checkasm_lfg, seed);
672 check_cpu_flag(NULL, 0);
673 for (i = 0; cpus[i].flag; i++)
674 check_cpu_flag(cpus[i].name, cpus[i].flag);
676 if (state.num_failed) {
677 fprintf(stderr, "checkasm: %d of %d tests have failed\n", state.num_failed, state.num_checked);
680 fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
681 if (state.bench_pattern) {
682 print_benchs(state.funcs);
686 destroy_func_tree(state.funcs);
691 /* Decide whether or not the specified function needs to be tested and
692 * allocate/initialize data structures if needed. Returns a pointer to a
693 * reference function if the function should be tested, otherwise NULL */
694 void *checkasm_check_func(void *func, const char *name, ...)
698 CheckasmFuncVersion *v;
703 name_length = vsnprintf(name_buf, sizeof(name_buf), name, arg);
706 if (!func || name_length <= 0 || name_length >= sizeof(name_buf))
709 state.current_func = get_func(&state.funcs, name_buf);
710 state.funcs->color = 1;
711 v = &state.current_func->versions;
714 CheckasmFuncVersion *prev;
716 /* Only test functions that haven't already been tested */
724 } while ((v = v->next));
726 v = prev->next = checkasm_malloc(sizeof(CheckasmFuncVersion));
731 v->cpu = state.cpu_flag;
732 state.current_func_ver = v;
740 /* Decide whether or not the current function needs to be benchmarked */
741 int checkasm_bench_func(void)
743 return !state.num_failed && state.bench_pattern &&
744 !strncmp(state.current_func->name, state.bench_pattern, state.bench_pattern_len);
747 /* Indicate that the current test has failed */
748 void checkasm_fail_func(const char *msg, ...)
750 if (state.current_func_ver->cpu && state.current_func_ver->ok) {
754 fprintf(stderr, " %s_%s (", state.current_func->name, cpu_suffix(state.current_func_ver->cpu));
756 vfprintf(stderr, msg, arg);
758 fprintf(stderr, ")\n");
760 state.current_func_ver->ok = 0;
765 /* Get the benchmark context of the current function */
766 CheckasmPerf *checkasm_get_perf_context(void)
768 CheckasmPerf *perf = &state.current_func_ver->perf;
769 memset(perf, 0, sizeof(*perf));
770 perf->sysfd = state.sysfd;
774 /* Print the outcome of all tests performed since the last time this function was called */
775 void checkasm_report(const char *name, ...)
777 static int prev_checked, prev_failed, max_length;
779 if (state.num_checked > prev_checked) {
780 int pad_length = max_length + 4;
784 pad_length -= fprintf(stderr, " - %s.", state.current_test_name);
786 pad_length -= vfprintf(stderr, name, arg);
788 fprintf(stderr, "%*c", FFMAX(pad_length, 0) + 2, '[');
790 if (state.num_failed == prev_failed)
791 color_printf(COLOR_GREEN, "OK");
793 color_printf(COLOR_RED, "FAILED");
794 fprintf(stderr, "]\n");
796 prev_checked = state.num_checked;
797 prev_failed = state.num_failed;
798 } else if (!state.cpu_flag) {
799 /* Calculate the amount of padding required to make the output vertically aligned */
800 int length = strlen(state.current_test_name);
804 length += vsnprintf(NULL, 0, name, arg);
807 if (length > max_length)