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 },
119 { "hevc_sao", checkasm_check_hevc_sao },
121 #if CONFIG_HUFFYUV_DECODER
122 { "huffyuvdsp", checkasm_check_huffyuvdsp },
124 #if CONFIG_JPEG2000_DECODER
125 { "jpeg2000dsp", checkasm_check_jpeg2000dsp },
127 #if CONFIG_HUFFYUVDSP
128 { "llviddsp", checkasm_check_llviddsp },
130 #if CONFIG_LLVIDENCDSP
131 { "llviddspenc", checkasm_check_llviddspenc },
133 #if CONFIG_PIXBLOCKDSP
134 { "pixblockdsp", checkasm_check_pixblockdsp },
136 #if CONFIG_UTVIDEO_DECODER
137 { "utvideodsp", checkasm_check_utvideodsp },
139 #if CONFIG_V210_ENCODER
140 { "v210enc", checkasm_check_v210enc },
143 { "vp8dsp", checkasm_check_vp8dsp },
145 #if CONFIG_VP9_DECODER
146 { "vp9dsp", checkasm_check_vp9dsp },
149 { "videodsp", checkasm_check_videodsp },
153 #if CONFIG_BLEND_FILTER
154 { "vf_blend", checkasm_check_blend },
156 #if CONFIG_COLORSPACE_FILTER
157 { "vf_colorspace", checkasm_check_colorspace },
159 #if CONFIG_HFLIP_FILTER
160 { "vf_hflip", checkasm_check_vf_hflip },
162 #if CONFIG_THRESHOLD_FILTER
163 { "vf_threshold", checkasm_check_vf_threshold },
167 { "sw_rgb", checkasm_check_sw_rgb },
170 { "fixed_dsp", checkasm_check_fixed_dsp },
171 { "float_dsp", checkasm_check_float_dsp },
176 /* List of cpu flags to check */
177 static const struct {
183 { "ARMV8", "armv8", AV_CPU_FLAG_ARMV8 },
184 { "NEON", "neon", AV_CPU_FLAG_NEON },
186 { "ARMV5TE", "armv5te", AV_CPU_FLAG_ARMV5TE },
187 { "ARMV6", "armv6", AV_CPU_FLAG_ARMV6 },
188 { "ARMV6T2", "armv6t2", AV_CPU_FLAG_ARMV6T2 },
189 { "VFP", "vfp", AV_CPU_FLAG_VFP },
190 { "VFP_VM", "vfp_vm", AV_CPU_FLAG_VFP_VM },
191 { "VFPV3", "vfp3", AV_CPU_FLAG_VFPV3 },
192 { "NEON", "neon", AV_CPU_FLAG_NEON },
194 { "ALTIVEC", "altivec", AV_CPU_FLAG_ALTIVEC },
195 { "VSX", "vsx", AV_CPU_FLAG_VSX },
196 { "POWER8", "power8", AV_CPU_FLAG_POWER8 },
198 { "MMX", "mmx", AV_CPU_FLAG_MMX|AV_CPU_FLAG_CMOV },
199 { "MMXEXT", "mmxext", AV_CPU_FLAG_MMXEXT },
200 { "3DNOW", "3dnow", AV_CPU_FLAG_3DNOW },
201 { "3DNOWEXT", "3dnowext", AV_CPU_FLAG_3DNOWEXT },
202 { "SSE", "sse", AV_CPU_FLAG_SSE },
203 { "SSE2", "sse2", AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW },
204 { "SSE3", "sse3", AV_CPU_FLAG_SSE3|AV_CPU_FLAG_SSE3SLOW },
205 { "SSSE3", "ssse3", AV_CPU_FLAG_SSSE3|AV_CPU_FLAG_ATOM },
206 { "SSE4.1", "sse4", AV_CPU_FLAG_SSE4 },
207 { "SSE4.2", "sse42", AV_CPU_FLAG_SSE42 },
208 { "AES-NI", "aesni", AV_CPU_FLAG_AESNI },
209 { "AVX", "avx", AV_CPU_FLAG_AVX },
210 { "XOP", "xop", AV_CPU_FLAG_XOP },
211 { "FMA3", "fma3", AV_CPU_FLAG_FMA3 },
212 { "FMA4", "fma4", AV_CPU_FLAG_FMA4 },
213 { "AVX2", "avx2", AV_CPU_FLAG_AVX2 },
214 { "AVX-512", "avx512", AV_CPU_FLAG_AVX512 },
219 typedef struct CheckasmFuncVersion {
220 struct CheckasmFuncVersion *next;
225 } CheckasmFuncVersion;
227 /* Binary search tree node */
228 typedef struct CheckasmFunc {
229 struct CheckasmFunc *child[2];
230 CheckasmFuncVersion versions;
231 uint8_t color; /* 0 = red, 1 = black */
238 CheckasmFunc *current_func;
239 CheckasmFuncVersion *current_func_ver;
240 const char *current_test_name;
241 const char *bench_pattern;
242 int bench_pattern_len;
251 const char *cpu_flag_name;
252 const char *test_name;
258 /* float compare support code */
259 static int is_negative(union av_intfloat32 u)
264 int float_near_ulp(float a, float b, unsigned max_ulp)
266 union av_intfloat32 x, y;
271 if (is_negative(x) != is_negative(y)) {
272 // handle -0.0 == +0.0
276 if (llabs((int64_t)x.i - y.i) <= max_ulp)
282 int float_near_ulp_array(const float *a, const float *b, unsigned max_ulp,
287 for (i = 0; i < len; i++) {
288 if (!float_near_ulp(a[i], b[i], max_ulp))
294 int float_near_abs_eps(float a, float b, float eps)
296 float abs_diff = fabsf(a - b);
298 return abs_diff < eps;
301 int float_near_abs_eps_array(const float *a, const float *b, float eps,
306 for (i = 0; i < len; i++) {
307 if (!float_near_abs_eps(a[i], b[i], eps))
313 int float_near_abs_eps_ulp(float a, float b, float eps, unsigned max_ulp)
315 return float_near_ulp(a, b, max_ulp) || float_near_abs_eps(a, b, eps);
318 int float_near_abs_eps_array_ulp(const float *a, const float *b, float eps,
319 unsigned max_ulp, unsigned len)
323 for (i = 0; i < len; i++) {
324 if (!float_near_abs_eps_ulp(a[i], b[i], eps, max_ulp))
330 int double_near_abs_eps(double a, double b, double eps)
332 double abs_diff = fabs(a - b);
334 return abs_diff < eps;
337 int double_near_abs_eps_array(const double *a, const double *b, double eps,
342 for (i = 0; i < len; i++) {
343 if (!double_near_abs_eps(a[i], b[i], eps))
349 /* Print colored text to stderr if the terminal supports it */
350 static void color_printf(int color, const char *fmt, ...)
352 static int use_color = -1;
355 #if HAVE_SETCONSOLETEXTATTRIBUTE
357 static WORD org_attributes;
360 CONSOLE_SCREEN_BUFFER_INFO con_info;
361 con = GetStdHandle(STD_ERROR_HANDLE);
362 if (con && con != INVALID_HANDLE_VALUE && GetConsoleScreenBufferInfo(con, &con_info)) {
363 org_attributes = con_info.wAttributes;
369 SetConsoleTextAttribute(con, (org_attributes & 0xfff0) | (color & 0x0f));
372 const char *term = getenv("TERM");
373 use_color = term && strcmp(term, "dumb") && isatty(2);
376 fprintf(stderr, "\x1b[%d;3%dm", (color & 0x08) >> 3, color & 0x07);
380 vfprintf(stderr, fmt, arg);
384 #if HAVE_SETCONSOLETEXTATTRIBUTE
385 SetConsoleTextAttribute(con, org_attributes);
387 fprintf(stderr, "\x1b[0m");
392 /* Deallocate a tree */
393 static void destroy_func_tree(CheckasmFunc *f)
396 CheckasmFuncVersion *v = f->versions.next;
398 CheckasmFuncVersion *next = v->next;
403 destroy_func_tree(f->child[0]);
404 destroy_func_tree(f->child[1]);
409 /* Allocate a zero-initialized block, clean up and exit on failure */
410 static void *checkasm_malloc(size_t size)
412 void *ptr = calloc(1, size);
414 fprintf(stderr, "checkasm: malloc failed\n");
415 destroy_func_tree(state.funcs);
421 /* Get the suffix of the specified cpu flag */
422 static const char *cpu_suffix(int cpu)
424 int i = FF_ARRAY_ELEMS(cpus);
427 if (cpu & cpus[i].flag)
428 return cpus[i].suffix;
433 static int cmp_nop(const void *a, const void *b)
435 return *(const uint16_t*)a - *(const uint16_t*)b;
438 /* Measure the overhead of the timing code (in decicycles) */
439 static int measure_nop_time(void)
441 uint16_t nops[10000];
443 av_unused const int sysfd = state.sysfd;
446 for (i = 0; i < 10000; i++) {
452 qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
453 for (i = 2500; i < 7500; i++)
456 return nop_sum / 500;
459 /* Print benchmark results */
460 static void print_benchs(CheckasmFunc *f)
463 print_benchs(f->child[0]);
465 /* Only print functions with at least one assembly version */
466 if (f->versions.cpu || f->versions.next) {
467 CheckasmFuncVersion *v = &f->versions;
469 CheckasmPerf *p = &v->perf;
471 int decicycles = (10*p->cycles/p->iterations - state.nop_time) / 4;
472 printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
474 } while ((v = v->next));
477 print_benchs(f->child[1]);
481 /* ASCIIbetical sort except preserving natural order for numbers */
482 static int cmp_func_names(const char *a, const char *b)
484 const char *start = a;
485 int ascii_diff, digit_diff;
487 for (; !(ascii_diff = *(const unsigned char*)a - *(const unsigned char*)b) && *a; a++, b++);
488 for (; av_isdigit(*a) && av_isdigit(*b); a++, b++);
490 if (a > start && av_isdigit(a[-1]) && (digit_diff = av_isdigit(*a) - av_isdigit(*b)))
496 /* Perform a tree rotation in the specified direction and return the new root */
497 static CheckasmFunc *rotate_tree(CheckasmFunc *f, int dir)
499 CheckasmFunc *r = f->child[dir^1];
500 f->child[dir^1] = r->child[dir];
507 #define is_red(f) ((f) && !(f)->color)
509 /* Balance a left-leaning red-black tree at the specified node */
510 static void balance_tree(CheckasmFunc **root)
512 CheckasmFunc *f = *root;
514 if (is_red(f->child[0]) && is_red(f->child[1])) {
516 f->child[0]->color = f->child[1]->color = 1;
519 if (!is_red(f->child[0]) && is_red(f->child[1]))
520 *root = rotate_tree(f, 0); /* Rotate left */
521 else if (is_red(f->child[0]) && is_red(f->child[0]->child[0]))
522 *root = rotate_tree(f, 1); /* Rotate right */
525 /* Get a node with the specified name, creating it if it doesn't exist */
526 static CheckasmFunc *get_func(CheckasmFunc **root, const char *name)
528 CheckasmFunc *f = *root;
531 /* Search the tree for a matching node */
532 int cmp = cmp_func_names(name, f->name);
534 f = get_func(&f->child[cmp > 0], name);
536 /* Rebalance the tree on the way up if a new node was inserted */
537 if (!f->versions.func)
541 /* Allocate and insert a new node into the tree */
542 int name_length = strlen(name);
543 f = *root = checkasm_malloc(sizeof(CheckasmFunc) + name_length);
544 memcpy(f->name, name, name_length + 1);
550 /* Perform tests and benchmarks for the specified cpu flag if supported by the host */
551 static void check_cpu_flag(const char *name, int flag)
553 int old_cpu_flag = state.cpu_flag;
555 flag |= old_cpu_flag;
556 av_force_cpu_flags(-1);
557 state.cpu_flag = flag & av_get_cpu_flags();
558 av_force_cpu_flags(state.cpu_flag);
560 if (!flag || state.cpu_flag != old_cpu_flag) {
563 state.cpu_flag_name = name;
564 for (i = 0; tests[i].func; i++) {
565 if (state.test_name && strcmp(tests[i].name, state.test_name))
567 state.current_test_name = tests[i].name;
573 /* Print the name of the current CPU flag, but only do it once */
574 static void print_cpu_name(void)
576 if (state.cpu_flag_name) {
577 color_printf(COLOR_YELLOW, "%s:\n", state.cpu_flag_name);
578 state.cpu_flag_name = NULL;
582 #if CONFIG_LINUX_PERF
583 static int bench_init_linux(void)
585 struct perf_event_attr attr = {
586 .type = PERF_TYPE_HARDWARE,
587 .size = sizeof(struct perf_event_attr),
588 .config = PERF_COUNT_HW_CPU_CYCLES,
589 .disabled = 1, // start counting only on demand
594 printf("benchmarking with Linux Perf Monitoring API\n");
596 state.sysfd = syscall(__NR_perf_event_open, &attr, 0, -1, -1, 0);
597 if (state.sysfd == -1) {
605 static int bench_init_ffmpeg(void)
608 printf("benchmarking with native FFmpeg timers\n");
611 fprintf(stderr, "checkasm: --bench is not supported on your system\n");
616 static int bench_init(void)
618 #if CONFIG_LINUX_PERF
619 int ret = bench_init_linux();
621 int ret = bench_init_ffmpeg();
626 state.nop_time = measure_nop_time();
627 printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
631 static void bench_uninit(void)
633 #if CONFIG_LINUX_PERF
639 int main(int argc, char *argv[])
641 unsigned int seed = av_get_random_seed();
644 #if ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
645 if (have_vfp(av_get_cpu_flags()) || have_neon(av_get_cpu_flags()))
646 checkasm_checked_call = checkasm_checked_call_vfp;
649 if (!tests[0].func || !cpus[0].flag) {
650 fprintf(stderr, "checkasm: no tests to perform\n");
655 if (!strncmp(argv[1], "--bench", 7)) {
656 if (bench_init() < 0)
658 if (argv[1][7] == '=') {
659 state.bench_pattern = argv[1] + 8;
660 state.bench_pattern_len = strlen(state.bench_pattern);
662 state.bench_pattern = "";
663 } else if (!strncmp(argv[1], "--test=", 7)) {
664 state.test_name = argv[1] + 7;
666 seed = strtoul(argv[1], NULL, 10);
673 fprintf(stderr, "checkasm: using random seed %u\n", seed);
674 av_lfg_init(&checkasm_lfg, seed);
676 check_cpu_flag(NULL, 0);
677 for (i = 0; cpus[i].flag; i++)
678 check_cpu_flag(cpus[i].name, cpus[i].flag);
680 if (state.num_failed) {
681 fprintf(stderr, "checkasm: %d of %d tests have failed\n", state.num_failed, state.num_checked);
684 fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
685 if (state.bench_pattern) {
686 print_benchs(state.funcs);
690 destroy_func_tree(state.funcs);
695 /* Decide whether or not the specified function needs to be tested and
696 * allocate/initialize data structures if needed. Returns a pointer to a
697 * reference function if the function should be tested, otherwise NULL */
698 void *checkasm_check_func(void *func, const char *name, ...)
702 CheckasmFuncVersion *v;
707 name_length = vsnprintf(name_buf, sizeof(name_buf), name, arg);
710 if (!func || name_length <= 0 || name_length >= sizeof(name_buf))
713 state.current_func = get_func(&state.funcs, name_buf);
714 state.funcs->color = 1;
715 v = &state.current_func->versions;
718 CheckasmFuncVersion *prev;
720 /* Only test functions that haven't already been tested */
728 } while ((v = v->next));
730 v = prev->next = checkasm_malloc(sizeof(CheckasmFuncVersion));
735 v->cpu = state.cpu_flag;
736 state.current_func_ver = v;
744 /* Decide whether or not the current function needs to be benchmarked */
745 int checkasm_bench_func(void)
747 return !state.num_failed && state.bench_pattern &&
748 !strncmp(state.current_func->name, state.bench_pattern, state.bench_pattern_len);
751 /* Indicate that the current test has failed */
752 void checkasm_fail_func(const char *msg, ...)
754 if (state.current_func_ver->cpu && state.current_func_ver->ok) {
758 fprintf(stderr, " %s_%s (", state.current_func->name, cpu_suffix(state.current_func_ver->cpu));
760 vfprintf(stderr, msg, arg);
762 fprintf(stderr, ")\n");
764 state.current_func_ver->ok = 0;
769 /* Get the benchmark context of the current function */
770 CheckasmPerf *checkasm_get_perf_context(void)
772 CheckasmPerf *perf = &state.current_func_ver->perf;
773 memset(perf, 0, sizeof(*perf));
774 perf->sysfd = state.sysfd;
778 /* Print the outcome of all tests performed since the last time this function was called */
779 void checkasm_report(const char *name, ...)
781 static int prev_checked, prev_failed, max_length;
783 if (state.num_checked > prev_checked) {
784 int pad_length = max_length + 4;
788 pad_length -= fprintf(stderr, " - %s.", state.current_test_name);
790 pad_length -= vfprintf(stderr, name, arg);
792 fprintf(stderr, "%*c", FFMAX(pad_length, 0) + 2, '[');
794 if (state.num_failed == prev_failed)
795 color_printf(COLOR_GREEN, "OK");
797 color_printf(COLOR_RED, "FAILED");
798 fprintf(stderr, "]\n");
800 prev_checked = state.num_checked;
801 prev_failed = state.num_failed;
802 } else if (!state.cpu_flag) {
803 /* Calculate the amount of padding required to make the output vertically aligned */
804 int length = strlen(state.current_test_name);
808 length += vsnprintf(NULL, 0, name, arg);
811 if (length > max_length)