2 * Assembly testing and benchmarking tool
3 * Copyright (c) 2015 Henrik Gramner
4 * Copyright (c) 2008 Loren Merritt
6 * This file is part of Libav.
8 * Libav 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 * Libav 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 Libav; 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/random_seed.h"
36 #if HAVE_SETCONSOLETEXTATTRIBUTE
38 #define COLOR_RED FOREGROUND_RED
39 #define COLOR_GREEN FOREGROUND_GREEN
40 #define COLOR_YELLOW (FOREGROUND_RED|FOREGROUND_GREEN)
44 #define COLOR_YELLOW 3
55 /* List of tests to invoke */
61 { "bswapdsp", checkasm_check_bswapdsp },
64 { "h264pred", checkasm_check_h264pred },
67 { "h264qpel", checkasm_check_h264qpel },
69 #if CONFIG_HEVC_DECODER
70 { "hevc_mc", checkasm_check_hevc_mc },
72 #if CONFIG_V210_ENCODER
73 { "v210enc", checkasm_check_v210enc },
78 /* List of cpu flags to check */
85 { "ARMV8", "armv8", AV_CPU_FLAG_ARMV8 },
86 { "NEON", "neon", AV_CPU_FLAG_NEON },
88 { "ARMV5TE", "armv5te", AV_CPU_FLAG_ARMV5TE },
89 { "ARMV6", "armv6", AV_CPU_FLAG_ARMV6 },
90 { "ARMV6T2", "armv6t2", AV_CPU_FLAG_ARMV6T2 },
91 { "VFP", "vfp", AV_CPU_FLAG_VFP },
92 { "VFPV3", "vfp3", AV_CPU_FLAG_VFPV3 },
93 { "NEON", "neon", AV_CPU_FLAG_NEON },
95 { "ALTIVEC", "altivec", AV_CPU_FLAG_ALTIVEC },
96 { "VSX", "vsx", AV_CPU_FLAG_VSX },
97 { "POWER8", "power8", AV_CPU_FLAG_POWER8 },
99 { "MMX", "mmx", AV_CPU_FLAG_MMX|AV_CPU_FLAG_CMOV },
100 { "MMXEXT", "mmxext", AV_CPU_FLAG_MMXEXT },
101 { "3DNOW", "3dnow", AV_CPU_FLAG_3DNOW },
102 { "3DNOWEXT", "3dnowext", AV_CPU_FLAG_3DNOWEXT },
103 { "SSE", "sse", AV_CPU_FLAG_SSE },
104 { "SSE2", "sse2", AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW },
105 { "SSE3", "sse3", AV_CPU_FLAG_SSE3|AV_CPU_FLAG_SSE3SLOW },
106 { "SSSE3", "ssse3", AV_CPU_FLAG_SSSE3|AV_CPU_FLAG_ATOM },
107 { "SSE4.1", "sse4", AV_CPU_FLAG_SSE4 },
108 { "SSE4.2", "sse42", AV_CPU_FLAG_SSE42 },
109 { "AVX", "avx", AV_CPU_FLAG_AVX },
110 { "XOP", "xop", AV_CPU_FLAG_XOP },
111 { "FMA3", "fma3", AV_CPU_FLAG_FMA3 },
112 { "FMA4", "fma4", AV_CPU_FLAG_FMA4 },
113 { "AVX2", "avx2", AV_CPU_FLAG_AVX2 },
118 typedef struct CheckasmFuncVersion {
119 struct CheckasmFuncVersion *next;
125 } CheckasmFuncVersion;
127 /* Binary search tree node */
128 typedef struct CheckasmFunc {
129 struct CheckasmFunc *child[2];
130 CheckasmFuncVersion versions;
131 uint8_t color; /* 0 = red, 1 = black */
138 CheckasmFunc *current_func;
139 CheckasmFuncVersion *current_func_ver;
140 const char *current_test_name;
141 const char *bench_pattern;
142 int bench_pattern_len;
147 const char *cpu_flag_name;
153 /* Print colored text to stderr if the terminal supports it */
154 static void color_printf(int color, const char *fmt, ...)
156 static int use_color = -1;
159 #if HAVE_SETCONSOLETEXTATTRIBUTE
161 static WORD org_attributes;
164 CONSOLE_SCREEN_BUFFER_INFO con_info;
165 con = GetStdHandle(STD_ERROR_HANDLE);
166 if (con && con != INVALID_HANDLE_VALUE && GetConsoleScreenBufferInfo(con, &con_info)) {
167 org_attributes = con_info.wAttributes;
173 SetConsoleTextAttribute(con, (org_attributes & 0xfff0) | (color & 0x0f));
176 const char *term = getenv("TERM");
177 use_color = term && strcmp(term, "dumb") && isatty(2);
180 fprintf(stderr, "\x1b[%d;3%dm", (color & 0x08) >> 3, color & 0x07);
184 vfprintf(stderr, fmt, arg);
188 #if HAVE_SETCONSOLETEXTATTRIBUTE
189 SetConsoleTextAttribute(con, org_attributes);
191 fprintf(stderr, "\x1b[0m");
196 /* Deallocate a tree */
197 static void destroy_func_tree(CheckasmFunc *f)
200 CheckasmFuncVersion *v = f->versions.next;
202 CheckasmFuncVersion *next = v->next;
207 destroy_func_tree(f->child[0]);
208 destroy_func_tree(f->child[1]);
213 /* Allocate a zero-initialized block, clean up and exit on failure */
214 static void *checkasm_malloc(size_t size)
216 void *ptr = calloc(1, size);
218 fprintf(stderr, "checkasm: malloc failed\n");
219 destroy_func_tree(state.funcs);
225 /* Get the suffix of the specified cpu flag */
226 static const char *cpu_suffix(int cpu)
228 int i = FF_ARRAY_ELEMS(cpus);
231 if (cpu & cpus[i].flag)
232 return cpus[i].suffix;
238 static int cmp_nop(const void *a, const void *b)
240 return *(const uint16_t*)a - *(const uint16_t*)b;
243 /* Measure the overhead of the timing code (in decicycles) */
244 static int measure_nop_time(void)
246 uint16_t nops[10000];
249 for (i = 0; i < 10000; i++) {
250 uint64_t t = AV_READ_TIME();
251 nops[i] = AV_READ_TIME() - t;
254 qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
255 for (i = 2500; i < 7500; i++)
258 return nop_sum / 500;
261 /* Print benchmark results */
262 static void print_benchs(CheckasmFunc *f)
265 print_benchs(f->child[0]);
267 /* Only print functions with at least one assembly version */
268 if (f->versions.cpu || f->versions.next) {
269 CheckasmFuncVersion *v = &f->versions;
272 int decicycles = (10*v->cycles/v->iterations - state.nop_time) / 4;
273 printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
275 } while ((v = v->next));
278 print_benchs(f->child[1]);
283 /* ASCIIbetical sort except preserving natural order for numbers */
284 static int cmp_func_names(const char *a, const char *b)
286 const char *start = a;
287 int ascii_diff, digit_diff;
289 for (; !(ascii_diff = *(const unsigned char*)a - *(const unsigned char*)b) && *a; a++, b++);
290 for (; av_isdigit(*a) && av_isdigit(*b); a++, b++);
292 if (a > start && av_isdigit(a[-1]) && (digit_diff = av_isdigit(*a) - av_isdigit(*b)))
298 /* Perform a tree rotation in the specified direction and return the new root */
299 static CheckasmFunc *rotate_tree(CheckasmFunc *f, int dir)
301 CheckasmFunc *r = f->child[dir^1];
302 f->child[dir^1] = r->child[dir];
309 #define is_red(f) ((f) && !(f)->color)
311 /* Balance a left-leaning red-black tree at the specified node */
312 static void balance_tree(CheckasmFunc **root)
314 CheckasmFunc *f = *root;
316 if (is_red(f->child[0]) && is_red(f->child[1])) {
318 f->child[0]->color = f->child[1]->color = 1;
321 if (!is_red(f->child[0]) && is_red(f->child[1]))
322 *root = rotate_tree(f, 0); /* Rotate left */
323 else if (is_red(f->child[0]) && is_red(f->child[0]->child[0]))
324 *root = rotate_tree(f, 1); /* Rotate right */
327 /* Get a node with the specified name, creating it if it doesn't exist */
328 static CheckasmFunc *get_func(CheckasmFunc **root, const char *name)
330 CheckasmFunc *f = *root;
333 /* Search the tree for a matching node */
334 int cmp = cmp_func_names(name, f->name);
336 f = get_func(&f->child[cmp > 0], name);
338 /* Rebalance the tree on the way up if a new node was inserted */
339 if (!f->versions.func)
343 /* Allocate and insert a new node into the tree */
344 int name_length = strlen(name);
345 f = *root = checkasm_malloc(sizeof(CheckasmFunc) + name_length);
346 memcpy(f->name, name, name_length + 1);
352 /* Perform tests and benchmarks for the specified cpu flag if supported by the host */
353 static void check_cpu_flag(const char *name, int flag)
355 int old_cpu_flag = state.cpu_flag;
357 flag |= old_cpu_flag;
358 av_set_cpu_flags_mask(flag);
359 state.cpu_flag = av_get_cpu_flags();
361 if (!flag || state.cpu_flag != old_cpu_flag) {
364 state.cpu_flag_name = name;
365 for (i = 0; tests[i].func; i++) {
366 state.current_test_name = tests[i].name;
372 /* Print the name of the current CPU flag, but only do it once */
373 static void print_cpu_name(void)
375 if (state.cpu_flag_name) {
376 color_printf(COLOR_YELLOW, "%s:\n", state.cpu_flag_name);
377 state.cpu_flag_name = NULL;
381 int main(int argc, char *argv[])
383 int i, seed, ret = 0;
385 if (!tests[0].func || !cpus[0].flag) {
386 fprintf(stderr, "checkasm: no tests to perform\n");
390 if (argc > 1 && !strncmp(argv[1], "--bench", 7)) {
392 fprintf(stderr, "checkasm: --bench is not supported on your system\n");
395 if (argv[1][7] == '=') {
396 state.bench_pattern = argv[1] + 8;
397 state.bench_pattern_len = strlen(state.bench_pattern);
399 state.bench_pattern = "";
405 seed = (argc > 1) ? atoi(argv[1]) : av_get_random_seed();
406 fprintf(stderr, "checkasm: using random seed %u\n", seed);
407 av_lfg_init(&checkasm_lfg, seed);
409 check_cpu_flag(NULL, 0);
410 for (i = 0; cpus[i].flag; i++)
411 check_cpu_flag(cpus[i].name, cpus[i].flag);
413 if (state.num_failed) {
414 fprintf(stderr, "checkasm: %d of %d tests have failed\n", state.num_failed, state.num_checked);
417 fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
419 if (state.bench_pattern) {
420 state.nop_time = measure_nop_time();
421 printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
422 print_benchs(state.funcs);
427 destroy_func_tree(state.funcs);
431 /* Decide whether or not the specified function needs to be tested and
432 * allocate/initialize data structures if needed. Returns a pointer to a
433 * reference function if the function should be tested, otherwise NULL */
434 void *checkasm_check_func(void *func, const char *name, ...)
438 CheckasmFuncVersion *v;
443 name_length = vsnprintf(name_buf, sizeof(name_buf), name, arg);
446 if (!func || name_length <= 0 || name_length >= sizeof(name_buf))
449 state.current_func = get_func(&state.funcs, name_buf);
450 state.funcs->color = 1;
451 v = &state.current_func->versions;
454 CheckasmFuncVersion *prev;
456 /* Only test functions that haven't already been tested */
464 } while ((v = v->next));
466 v = prev->next = checkasm_malloc(sizeof(CheckasmFuncVersion));
471 v->cpu = state.cpu_flag;
472 state.current_func_ver = v;
480 /* Decide whether or not the current function needs to be benchmarked */
481 int checkasm_bench_func(void)
483 return !state.num_failed && state.bench_pattern &&
484 !strncmp(state.current_func->name, state.bench_pattern, state.bench_pattern_len);
487 /* Indicate that the current test has failed */
488 void checkasm_fail_func(const char *msg, ...)
490 if (state.current_func_ver->cpu && state.current_func_ver->ok) {
494 fprintf(stderr, " %s_%s (", state.current_func->name, cpu_suffix(state.current_func_ver->cpu));
496 vfprintf(stderr, msg, arg);
498 fprintf(stderr, ")\n");
500 state.current_func_ver->ok = 0;
505 /* Update benchmark results of the current function */
506 void checkasm_update_bench(int iterations, uint64_t cycles)
508 state.current_func_ver->iterations += iterations;
509 state.current_func_ver->cycles += cycles;
512 /* Print the outcome of all tests performed since the last time this function was called */
513 void checkasm_report(const char *name, ...)
515 static int prev_checked, prev_failed, max_length;
517 if (state.num_checked > prev_checked) {
518 int pad_length = max_length + 4;
522 pad_length -= fprintf(stderr, " - %s.", state.current_test_name);
524 pad_length -= vfprintf(stderr, name, arg);
526 fprintf(stderr, "%*c", FFMAX(pad_length, 0) + 2, '[');
528 if (state.num_failed == prev_failed)
529 color_printf(COLOR_GREEN, "OK");
531 color_printf(COLOR_RED, "FAILED");
532 fprintf(stderr, "]\n");
534 prev_checked = state.num_checked;
535 prev_failed = state.num_failed;
536 } else if (!state.cpu_flag) {
537 /* Calculate the amount of padding required to make the output vertically aligned */
538 int length = strlen(state.current_test_name);
542 length += vsnprintf(NULL, 0, name, arg);
545 if (length > max_length)