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/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_V210_ENCODER
70 { "v210enc", checkasm_check_v210enc },
75 /* List of cpu flags to check */
82 { "ARMV8", "armv8", AV_CPU_FLAG_ARMV8 },
83 { "NEON", "neon", AV_CPU_FLAG_NEON },
85 { "ARMV5TE", "armv5te", AV_CPU_FLAG_ARMV5TE },
86 { "ARMV6", "armv6", AV_CPU_FLAG_ARMV6 },
87 { "ARMV6T2", "armv6t2", AV_CPU_FLAG_ARMV6T2 },
88 { "VFP", "vfp", AV_CPU_FLAG_VFP },
89 { "VFPV3", "vfp3", AV_CPU_FLAG_VFPV3 },
90 { "NEON", "neon", AV_CPU_FLAG_NEON },
92 { "ALTIVEC", "altivec", AV_CPU_FLAG_ALTIVEC },
93 { "VSX", "vsx", AV_CPU_FLAG_VSX },
94 { "POWER8", "power8", AV_CPU_FLAG_POWER8 },
96 { "MMX", "mmx", AV_CPU_FLAG_MMX|AV_CPU_FLAG_CMOV },
97 { "MMXEXT", "mmxext", AV_CPU_FLAG_MMXEXT },
98 { "3DNOW", "3dnow", AV_CPU_FLAG_3DNOW },
99 { "3DNOWEXT", "3dnowext", AV_CPU_FLAG_3DNOWEXT },
100 { "SSE", "sse", AV_CPU_FLAG_SSE },
101 { "SSE2", "sse2", AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW },
102 { "SSE3", "sse3", AV_CPU_FLAG_SSE3|AV_CPU_FLAG_SSE3SLOW },
103 { "SSSE3", "ssse3", AV_CPU_FLAG_SSSE3|AV_CPU_FLAG_ATOM },
104 { "SSE4.1", "sse4", AV_CPU_FLAG_SSE4 },
105 { "SSE4.2", "sse42", AV_CPU_FLAG_SSE42 },
106 { "AVX", "avx", AV_CPU_FLAG_AVX },
107 { "XOP", "xop", AV_CPU_FLAG_XOP },
108 { "FMA3", "fma3", AV_CPU_FLAG_FMA3 },
109 { "FMA4", "fma4", AV_CPU_FLAG_FMA4 },
110 { "AVX2", "avx2", AV_CPU_FLAG_AVX2 },
115 typedef struct CheckasmFuncVersion {
116 struct CheckasmFuncVersion *next;
122 } CheckasmFuncVersion;
124 /* Binary search tree node */
125 typedef struct CheckasmFunc {
126 struct CheckasmFunc *child[2];
127 CheckasmFuncVersion versions;
134 CheckasmFunc *current_func;
135 CheckasmFuncVersion *current_func_ver;
136 const char *current_test_name;
137 const char *bench_pattern;
138 int bench_pattern_len;
143 const char *cpu_flag_name;
149 /* Print colored text to stderr if the terminal supports it */
150 static void color_printf(int color, const char *fmt, ...)
152 static int use_color = -1;
155 #if HAVE_SETCONSOLETEXTATTRIBUTE
157 static WORD org_attributes;
160 CONSOLE_SCREEN_BUFFER_INFO con_info;
161 con = GetStdHandle(STD_ERROR_HANDLE);
162 if (con && con != INVALID_HANDLE_VALUE && GetConsoleScreenBufferInfo(con, &con_info)) {
163 org_attributes = con_info.wAttributes;
169 SetConsoleTextAttribute(con, (org_attributes & 0xfff0) | (color & 0x0f));
172 const char *term = getenv("TERM");
173 use_color = term && strcmp(term, "dumb") && isatty(2);
176 fprintf(stderr, "\x1b[%d;3%dm", (color & 0x08) >> 3, color & 0x07);
180 vfprintf(stderr, fmt, arg);
184 #if HAVE_SETCONSOLETEXTATTRIBUTE
185 SetConsoleTextAttribute(con, org_attributes);
187 fprintf(stderr, "\x1b[0m");
192 /* Deallocate a tree */
193 static void destroy_func_tree(CheckasmFunc *f)
196 CheckasmFuncVersion *v = f->versions.next;
198 CheckasmFuncVersion *next = v->next;
203 destroy_func_tree(f->child[0]);
204 destroy_func_tree(f->child[1]);
209 /* Allocate a zero-initialized block, clean up and exit on failure */
210 static void *checkasm_malloc(size_t size)
212 void *ptr = calloc(1, size);
214 fprintf(stderr, "checkasm: malloc failed\n");
215 destroy_func_tree(state.funcs);
221 /* Get the suffix of the specified cpu flag */
222 static const char *cpu_suffix(int cpu)
224 int i = FF_ARRAY_ELEMS(cpus);
227 if (cpu & cpus[i].flag)
228 return cpus[i].suffix;
234 static int cmp_nop(const void *a, const void *b)
236 return *(const uint16_t*)a - *(const uint16_t*)b;
239 /* Measure the overhead of the timing code (in decicycles) */
240 static int measure_nop_time(void)
242 uint16_t nops[10000];
245 for (i = 0; i < 10000; i++) {
246 uint64_t t = AV_READ_TIME();
247 nops[i] = AV_READ_TIME() - t;
250 qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
251 for (i = 2500; i < 7500; i++)
254 return nop_sum / 500;
257 /* Print benchmark results */
258 static void print_benchs(CheckasmFunc *f)
261 print_benchs(f->child[0]);
263 /* Only print functions with at least one assembly version */
264 if (f->versions.cpu || f->versions.next) {
265 CheckasmFuncVersion *v = &f->versions;
268 int decicycles = (10*v->cycles/v->iterations - state.nop_time) / 4;
269 printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
271 } while ((v = v->next));
274 print_benchs(f->child[1]);
279 /* ASCIIbetical sort except preserving natural order for numbers */
280 static int cmp_func_names(const char *a, const char *b)
282 int ascii_diff, digit_diff;
284 for (; !(ascii_diff = *a - *b) && *a; a++, b++);
285 for (; av_isdigit(*a) && av_isdigit(*b); a++, b++);
287 return (digit_diff = av_isdigit(*a) - av_isdigit(*b)) ? digit_diff : ascii_diff;
290 /* Get a node with the specified name, creating it if it doesn't exist */
291 static CheckasmFunc *get_func(const char *name, int length)
293 CheckasmFunc *f, **f_ptr = &state.funcs;
295 /* Search the tree for a matching node */
296 while ((f = *f_ptr)) {
297 int cmp = cmp_func_names(name, f->name);
301 f_ptr = &f->child[(cmp > 0)];
304 /* Allocate and insert a new node into the tree */
305 f = *f_ptr = checkasm_malloc(sizeof(CheckasmFunc) + length);
306 memcpy(f->name, name, length+1);
311 /* Perform tests and benchmarks for the specified cpu flag if supported by the host */
312 static void check_cpu_flag(const char *name, int flag)
314 int old_cpu_flag = state.cpu_flag;
316 flag |= old_cpu_flag;
317 av_set_cpu_flags_mask(flag);
318 state.cpu_flag = av_get_cpu_flags();
320 if (!flag || state.cpu_flag != old_cpu_flag) {
323 state.cpu_flag_name = name;
324 for (i = 0; tests[i].func; i++) {
325 state.current_test_name = tests[i].name;
331 /* Print the name of the current CPU flag, but only do it once */
332 static void print_cpu_name(void)
334 if (state.cpu_flag_name) {
335 color_printf(COLOR_YELLOW, "%s:\n", state.cpu_flag_name);
336 state.cpu_flag_name = NULL;
340 int main(int argc, char *argv[])
342 int i, seed, ret = 0;
344 if (!tests[0].func || !cpus[0].flag) {
345 fprintf(stderr, "checkasm: no tests to perform\n");
349 if (argc > 1 && !strncmp(argv[1], "--bench", 7)) {
351 fprintf(stderr, "checkasm: --bench is not supported on your system\n");
354 if (argv[1][7] == '=') {
355 state.bench_pattern = argv[1] + 8;
356 state.bench_pattern_len = strlen(state.bench_pattern);
358 state.bench_pattern = "";
364 seed = (argc > 1) ? atoi(argv[1]) : av_get_random_seed();
365 fprintf(stderr, "checkasm: using random seed %u\n", seed);
366 av_lfg_init(&checkasm_lfg, seed);
368 check_cpu_flag(NULL, 0);
369 for (i = 0; cpus[i].flag; i++)
370 check_cpu_flag(cpus[i].name, cpus[i].flag);
372 if (state.num_failed) {
373 fprintf(stderr, "checkasm: %d of %d tests have failed\n", state.num_failed, state.num_checked);
376 fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
378 if (state.bench_pattern) {
379 state.nop_time = measure_nop_time();
380 printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
381 print_benchs(state.funcs);
386 destroy_func_tree(state.funcs);
390 /* Decide whether or not the specified function needs to be tested and
391 * allocate/initialize data structures if needed. Returns a pointer to a
392 * reference function if the function should be tested, otherwise NULL */
393 void *checkasm_check_func(void *func, const char *name, ...)
397 CheckasmFuncVersion *v;
402 name_length = vsnprintf(name_buf, sizeof(name_buf), name, arg);
405 if (!func || name_length <= 0 || name_length >= sizeof(name_buf))
408 state.current_func = get_func(name_buf, name_length);
409 v = &state.current_func->versions;
412 CheckasmFuncVersion *prev;
414 /* Only test functions that haven't already been tested */
422 } while ((v = v->next));
424 v = prev->next = checkasm_malloc(sizeof(CheckasmFuncVersion));
429 v->cpu = state.cpu_flag;
430 state.current_func_ver = v;
438 /* Decide whether or not the current function needs to be benchmarked */
439 int checkasm_bench_func(void)
441 return !state.num_failed && state.bench_pattern &&
442 !strncmp(state.current_func->name, state.bench_pattern, state.bench_pattern_len);
445 /* Indicate that the current test has failed */
446 void checkasm_fail_func(const char *msg, ...)
448 if (state.current_func_ver->cpu && state.current_func_ver->ok) {
452 fprintf(stderr, " %s_%s (", state.current_func->name, cpu_suffix(state.current_func_ver->cpu));
454 vfprintf(stderr, msg, arg);
456 fprintf(stderr, ")\n");
458 state.current_func_ver->ok = 0;
463 /* Update benchmark results of the current function */
464 void checkasm_update_bench(int iterations, uint64_t cycles)
466 state.current_func_ver->iterations += iterations;
467 state.current_func_ver->cycles += cycles;
470 /* Print the outcome of all tests performed since the last time this function was called */
471 void checkasm_report(const char *name, ...)
473 static int prev_checked, prev_failed, max_length;
475 if (state.num_checked > prev_checked) {
476 int pad_length = max_length + 4;
480 pad_length -= fprintf(stderr, " - %s.", state.current_test_name);
482 pad_length -= vfprintf(stderr, name, arg);
484 fprintf(stderr, "%*c", FFMAX(pad_length, 0) + 2, '[');
486 if (state.num_failed == prev_failed)
487 color_printf(COLOR_GREEN, "OK");
489 color_printf(COLOR_RED, "FAILED");
490 fprintf(stderr, "]\n");
492 prev_checked = state.num_checked;
493 prev_failed = state.num_failed;
494 } else if (!state.cpu_flag) {
495 /* Calculate the amount of padding required to make the output vertically aligned */
496 int length = strlen(state.current_test_name);
500 length += vsnprintf(NULL, 0, name, arg);
503 if (length > max_length)