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_PIXBLOCKDSP
130 { "pixblockdsp", checkasm_check_pixblockdsp },
132 #if CONFIG_UTVIDEO_DECODER
133 { "utvideodsp", checkasm_check_utvideodsp },
135 #if CONFIG_V210_ENCODER
136 { "v210enc", checkasm_check_v210enc },
139 { "vp8dsp", checkasm_check_vp8dsp },
141 #if CONFIG_VP9_DECODER
142 { "vp9dsp", checkasm_check_vp9dsp },
145 { "videodsp", checkasm_check_videodsp },
149 #if CONFIG_BLEND_FILTER
150 { "vf_blend", checkasm_check_blend },
152 #if CONFIG_COLORSPACE_FILTER
153 { "vf_colorspace", checkasm_check_colorspace },
155 #if CONFIG_THRESHOLD_FILTER
156 { "vf_threshold", checkasm_check_vf_threshold },
160 { "fixed_dsp", checkasm_check_fixed_dsp },
161 { "float_dsp", checkasm_check_float_dsp },
166 /* List of cpu flags to check */
167 static const struct {
173 { "ARMV8", "armv8", AV_CPU_FLAG_ARMV8 },
174 { "NEON", "neon", AV_CPU_FLAG_NEON },
176 { "ARMV5TE", "armv5te", AV_CPU_FLAG_ARMV5TE },
177 { "ARMV6", "armv6", AV_CPU_FLAG_ARMV6 },
178 { "ARMV6T2", "armv6t2", AV_CPU_FLAG_ARMV6T2 },
179 { "VFP", "vfp", AV_CPU_FLAG_VFP },
180 { "VFP_VM", "vfp_vm", AV_CPU_FLAG_VFP_VM },
181 { "VFPV3", "vfp3", AV_CPU_FLAG_VFPV3 },
182 { "NEON", "neon", AV_CPU_FLAG_NEON },
184 { "ALTIVEC", "altivec", AV_CPU_FLAG_ALTIVEC },
185 { "VSX", "vsx", AV_CPU_FLAG_VSX },
186 { "POWER8", "power8", AV_CPU_FLAG_POWER8 },
188 { "MMX", "mmx", AV_CPU_FLAG_MMX|AV_CPU_FLAG_CMOV },
189 { "MMXEXT", "mmxext", AV_CPU_FLAG_MMXEXT },
190 { "3DNOW", "3dnow", AV_CPU_FLAG_3DNOW },
191 { "3DNOWEXT", "3dnowext", AV_CPU_FLAG_3DNOWEXT },
192 { "SSE", "sse", AV_CPU_FLAG_SSE },
193 { "SSE2", "sse2", AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW },
194 { "SSE3", "sse3", AV_CPU_FLAG_SSE3|AV_CPU_FLAG_SSE3SLOW },
195 { "SSSE3", "ssse3", AV_CPU_FLAG_SSSE3|AV_CPU_FLAG_ATOM },
196 { "SSE4.1", "sse4", AV_CPU_FLAG_SSE4 },
197 { "SSE4.2", "sse42", AV_CPU_FLAG_SSE42 },
198 { "AES-NI", "aesni", AV_CPU_FLAG_AESNI },
199 { "AVX", "avx", AV_CPU_FLAG_AVX },
200 { "XOP", "xop", AV_CPU_FLAG_XOP },
201 { "FMA3", "fma3", AV_CPU_FLAG_FMA3 },
202 { "FMA4", "fma4", AV_CPU_FLAG_FMA4 },
203 { "AVX2", "avx2", AV_CPU_FLAG_AVX2 },
208 typedef struct CheckasmFuncVersion {
209 struct CheckasmFuncVersion *next;
214 } CheckasmFuncVersion;
216 /* Binary search tree node */
217 typedef struct CheckasmFunc {
218 struct CheckasmFunc *child[2];
219 CheckasmFuncVersion versions;
220 uint8_t color; /* 0 = red, 1 = black */
227 CheckasmFunc *current_func;
228 CheckasmFuncVersion *current_func_ver;
229 const char *current_test_name;
230 const char *bench_pattern;
231 int bench_pattern_len;
240 const char *cpu_flag_name;
241 const char *test_name;
247 /* float compare support code */
248 static int is_negative(union av_intfloat32 u)
253 int float_near_ulp(float a, float b, unsigned max_ulp)
255 union av_intfloat32 x, y;
260 if (is_negative(x) != is_negative(y)) {
261 // handle -0.0 == +0.0
265 if (llabs((int64_t)x.i - y.i) <= max_ulp)
271 int float_near_ulp_array(const float *a, const float *b, unsigned max_ulp,
276 for (i = 0; i < len; i++) {
277 if (!float_near_ulp(a[i], b[i], max_ulp))
283 int float_near_abs_eps(float a, float b, float eps)
285 float abs_diff = fabsf(a - b);
287 return abs_diff < eps;
290 int float_near_abs_eps_array(const float *a, const float *b, float eps,
295 for (i = 0; i < len; i++) {
296 if (!float_near_abs_eps(a[i], b[i], eps))
302 int float_near_abs_eps_ulp(float a, float b, float eps, unsigned max_ulp)
304 return float_near_ulp(a, b, max_ulp) || float_near_abs_eps(a, b, eps);
307 int float_near_abs_eps_array_ulp(const float *a, const float *b, float eps,
308 unsigned max_ulp, unsigned len)
312 for (i = 0; i < len; i++) {
313 if (!float_near_abs_eps_ulp(a[i], b[i], eps, max_ulp))
319 int double_near_abs_eps(double a, double b, double eps)
321 double abs_diff = fabs(a - b);
323 return abs_diff < eps;
326 int double_near_abs_eps_array(const double *a, const double *b, double eps,
331 for (i = 0; i < len; i++) {
332 if (!double_near_abs_eps(a[i], b[i], eps))
338 /* Print colored text to stderr if the terminal supports it */
339 static void color_printf(int color, const char *fmt, ...)
341 static int use_color = -1;
344 #if HAVE_SETCONSOLETEXTATTRIBUTE
346 static WORD org_attributes;
349 CONSOLE_SCREEN_BUFFER_INFO con_info;
350 con = GetStdHandle(STD_ERROR_HANDLE);
351 if (con && con != INVALID_HANDLE_VALUE && GetConsoleScreenBufferInfo(con, &con_info)) {
352 org_attributes = con_info.wAttributes;
358 SetConsoleTextAttribute(con, (org_attributes & 0xfff0) | (color & 0x0f));
361 const char *term = getenv("TERM");
362 use_color = term && strcmp(term, "dumb") && isatty(2);
365 fprintf(stderr, "\x1b[%d;3%dm", (color & 0x08) >> 3, color & 0x07);
369 vfprintf(stderr, fmt, arg);
373 #if HAVE_SETCONSOLETEXTATTRIBUTE
374 SetConsoleTextAttribute(con, org_attributes);
376 fprintf(stderr, "\x1b[0m");
381 /* Deallocate a tree */
382 static void destroy_func_tree(CheckasmFunc *f)
385 CheckasmFuncVersion *v = f->versions.next;
387 CheckasmFuncVersion *next = v->next;
392 destroy_func_tree(f->child[0]);
393 destroy_func_tree(f->child[1]);
398 /* Allocate a zero-initialized block, clean up and exit on failure */
399 static void *checkasm_malloc(size_t size)
401 void *ptr = calloc(1, size);
403 fprintf(stderr, "checkasm: malloc failed\n");
404 destroy_func_tree(state.funcs);
410 /* Get the suffix of the specified cpu flag */
411 static const char *cpu_suffix(int cpu)
413 int i = FF_ARRAY_ELEMS(cpus);
416 if (cpu & cpus[i].flag)
417 return cpus[i].suffix;
422 static int cmp_nop(const void *a, const void *b)
424 return *(const uint16_t*)a - *(const uint16_t*)b;
427 /* Measure the overhead of the timing code (in decicycles) */
428 static int measure_nop_time(void)
430 uint16_t nops[10000];
432 av_unused const int sysfd = state.sysfd;
435 for (i = 0; i < 10000; i++) {
441 qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
442 for (i = 2500; i < 7500; i++)
445 return nop_sum / 500;
448 /* Print benchmark results */
449 static void print_benchs(CheckasmFunc *f)
452 print_benchs(f->child[0]);
454 /* Only print functions with at least one assembly version */
455 if (f->versions.cpu || f->versions.next) {
456 CheckasmFuncVersion *v = &f->versions;
458 CheckasmPerf *p = &v->perf;
460 int decicycles = (10*p->cycles/p->iterations - state.nop_time) / 4;
461 printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
463 } while ((v = v->next));
466 print_benchs(f->child[1]);
470 /* ASCIIbetical sort except preserving natural order for numbers */
471 static int cmp_func_names(const char *a, const char *b)
473 const char *start = a;
474 int ascii_diff, digit_diff;
476 for (; !(ascii_diff = *(const unsigned char*)a - *(const unsigned char*)b) && *a; a++, b++);
477 for (; av_isdigit(*a) && av_isdigit(*b); a++, b++);
479 if (a > start && av_isdigit(a[-1]) && (digit_diff = av_isdigit(*a) - av_isdigit(*b)))
485 /* Perform a tree rotation in the specified direction and return the new root */
486 static CheckasmFunc *rotate_tree(CheckasmFunc *f, int dir)
488 CheckasmFunc *r = f->child[dir^1];
489 f->child[dir^1] = r->child[dir];
496 #define is_red(f) ((f) && !(f)->color)
498 /* Balance a left-leaning red-black tree at the specified node */
499 static void balance_tree(CheckasmFunc **root)
501 CheckasmFunc *f = *root;
503 if (is_red(f->child[0]) && is_red(f->child[1])) {
505 f->child[0]->color = f->child[1]->color = 1;
508 if (!is_red(f->child[0]) && is_red(f->child[1]))
509 *root = rotate_tree(f, 0); /* Rotate left */
510 else if (is_red(f->child[0]) && is_red(f->child[0]->child[0]))
511 *root = rotate_tree(f, 1); /* Rotate right */
514 /* Get a node with the specified name, creating it if it doesn't exist */
515 static CheckasmFunc *get_func(CheckasmFunc **root, const char *name)
517 CheckasmFunc *f = *root;
520 /* Search the tree for a matching node */
521 int cmp = cmp_func_names(name, f->name);
523 f = get_func(&f->child[cmp > 0], name);
525 /* Rebalance the tree on the way up if a new node was inserted */
526 if (!f->versions.func)
530 /* Allocate and insert a new node into the tree */
531 int name_length = strlen(name);
532 f = *root = checkasm_malloc(sizeof(CheckasmFunc) + name_length);
533 memcpy(f->name, name, name_length + 1);
539 /* Perform tests and benchmarks for the specified cpu flag if supported by the host */
540 static void check_cpu_flag(const char *name, int flag)
542 int old_cpu_flag = state.cpu_flag;
544 flag |= old_cpu_flag;
545 av_force_cpu_flags(-1);
546 state.cpu_flag = flag & av_get_cpu_flags();
547 av_force_cpu_flags(state.cpu_flag);
549 if (!flag || state.cpu_flag != old_cpu_flag) {
552 state.cpu_flag_name = name;
553 for (i = 0; tests[i].func; i++) {
554 if (state.test_name && strcmp(tests[i].name, state.test_name))
556 state.current_test_name = tests[i].name;
562 /* Print the name of the current CPU flag, but only do it once */
563 static void print_cpu_name(void)
565 if (state.cpu_flag_name) {
566 color_printf(COLOR_YELLOW, "%s:\n", state.cpu_flag_name);
567 state.cpu_flag_name = NULL;
571 #if CONFIG_LINUX_PERF
572 static int bench_init_linux(void)
574 struct perf_event_attr attr = {
575 .type = PERF_TYPE_HARDWARE,
576 .size = sizeof(struct perf_event_attr),
577 .config = PERF_COUNT_HW_CPU_CYCLES,
578 .disabled = 1, // start counting only on demand
583 printf("benchmarking with Linux Perf Monitoring API\n");
585 state.sysfd = syscall(__NR_perf_event_open, &attr, 0, -1, -1, 0);
586 if (state.sysfd == -1) {
594 static int bench_init_ffmpeg(void)
597 printf("benchmarking with native FFmpeg timers\n");
600 fprintf(stderr, "checkasm: --bench is not supported on your system\n");
605 static int bench_init(void)
607 #if CONFIG_LINUX_PERF
608 int ret = bench_init_linux();
610 int ret = bench_init_ffmpeg();
615 state.nop_time = measure_nop_time();
616 printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
620 static void bench_uninit(void)
622 #if CONFIG_LINUX_PERF
628 int main(int argc, char *argv[])
630 unsigned int seed = av_get_random_seed();
633 #if ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
634 if (have_vfp(av_get_cpu_flags()) || have_neon(av_get_cpu_flags()))
635 checkasm_checked_call = checkasm_checked_call_vfp;
638 if (!tests[0].func || !cpus[0].flag) {
639 fprintf(stderr, "checkasm: no tests to perform\n");
644 if (!strncmp(argv[1], "--bench", 7)) {
645 if (bench_init() < 0)
647 if (argv[1][7] == '=') {
648 state.bench_pattern = argv[1] + 8;
649 state.bench_pattern_len = strlen(state.bench_pattern);
651 state.bench_pattern = "";
652 } else if (!strncmp(argv[1], "--test=", 7)) {
653 state.test_name = argv[1] + 7;
655 seed = strtoul(argv[1], NULL, 10);
662 fprintf(stderr, "checkasm: using random seed %u\n", seed);
663 av_lfg_init(&checkasm_lfg, seed);
665 check_cpu_flag(NULL, 0);
666 for (i = 0; cpus[i].flag; i++)
667 check_cpu_flag(cpus[i].name, cpus[i].flag);
669 if (state.num_failed) {
670 fprintf(stderr, "checkasm: %d of %d tests have failed\n", state.num_failed, state.num_checked);
673 fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
674 if (state.bench_pattern) {
675 print_benchs(state.funcs);
679 destroy_func_tree(state.funcs);
684 /* Decide whether or not the specified function needs to be tested and
685 * allocate/initialize data structures if needed. Returns a pointer to a
686 * reference function if the function should be tested, otherwise NULL */
687 void *checkasm_check_func(void *func, const char *name, ...)
691 CheckasmFuncVersion *v;
696 name_length = vsnprintf(name_buf, sizeof(name_buf), name, arg);
699 if (!func || name_length <= 0 || name_length >= sizeof(name_buf))
702 state.current_func = get_func(&state.funcs, name_buf);
703 state.funcs->color = 1;
704 v = &state.current_func->versions;
707 CheckasmFuncVersion *prev;
709 /* Only test functions that haven't already been tested */
717 } while ((v = v->next));
719 v = prev->next = checkasm_malloc(sizeof(CheckasmFuncVersion));
724 v->cpu = state.cpu_flag;
725 state.current_func_ver = v;
733 /* Decide whether or not the current function needs to be benchmarked */
734 int checkasm_bench_func(void)
736 return !state.num_failed && state.bench_pattern &&
737 !strncmp(state.current_func->name, state.bench_pattern, state.bench_pattern_len);
740 /* Indicate that the current test has failed */
741 void checkasm_fail_func(const char *msg, ...)
743 if (state.current_func_ver->cpu && state.current_func_ver->ok) {
747 fprintf(stderr, " %s_%s (", state.current_func->name, cpu_suffix(state.current_func_ver->cpu));
749 vfprintf(stderr, msg, arg);
751 fprintf(stderr, ")\n");
753 state.current_func_ver->ok = 0;
758 /* Get the benchmark context of the current function */
759 CheckasmPerf *checkasm_get_perf_context(void)
761 CheckasmPerf *perf = &state.current_func_ver->perf;
762 memset(perf, 0, sizeof(*perf));
763 perf->sysfd = state.sysfd;
767 /* Print the outcome of all tests performed since the last time this function was called */
768 void checkasm_report(const char *name, ...)
770 static int prev_checked, prev_failed, max_length;
772 if (state.num_checked > prev_checked) {
773 int pad_length = max_length + 4;
777 pad_length -= fprintf(stderr, " - %s.", state.current_test_name);
779 pad_length -= vfprintf(stderr, name, arg);
781 fprintf(stderr, "%*c", FFMAX(pad_length, 0) + 2, '[');
783 if (state.num_failed == prev_failed)
784 color_printf(COLOR_GREEN, "OK");
786 color_printf(COLOR_RED, "FAILED");
787 fprintf(stderr, "]\n");
789 prev_checked = state.num_checked;
790 prev_failed = state.num_failed;
791 } else if (!state.cpu_flag) {
792 /* Calculate the amount of padding required to make the output vertically aligned */
793 int length = strlen(state.current_test_name);
797 length += vsnprintf(NULL, 0, name, arg);
800 if (length > max_length)