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 #if CONFIG_ALAC_DECODER
62 { "alacdsp", checkasm_check_alacdsp },
65 { "bswapdsp", checkasm_check_bswapdsp },
68 { "flacdsp", checkasm_check_flacdsp },
71 { "h264pred", checkasm_check_h264pred },
74 { "h264qpel", checkasm_check_h264qpel },
76 #if CONFIG_JPEG2000_DECODER
77 { "jpeg2000dsp", checkasm_check_jpeg2000dsp },
79 #if CONFIG_PIXBLOCKDSP
80 { "pixblockdsp", checkasm_check_pixblockdsp },
82 #if CONFIG_V210_ENCODER
83 { "v210enc", checkasm_check_v210enc },
85 #if CONFIG_VP9_DECODER
86 { "vp9dsp", checkasm_check_vp9dsp },
92 /* List of cpu flags to check */
99 { "ARMV8", "armv8", AV_CPU_FLAG_ARMV8 },
100 { "NEON", "neon", AV_CPU_FLAG_NEON },
102 { "ARMV5TE", "armv5te", AV_CPU_FLAG_ARMV5TE },
103 { "ARMV6", "armv6", AV_CPU_FLAG_ARMV6 },
104 { "ARMV6T2", "armv6t2", AV_CPU_FLAG_ARMV6T2 },
105 { "VFP", "vfp", AV_CPU_FLAG_VFP },
106 { "VFP_VM", "vfp_vm", AV_CPU_FLAG_VFP_VM },
107 { "VFPV3", "vfp3", AV_CPU_FLAG_VFPV3 },
108 { "NEON", "neon", AV_CPU_FLAG_NEON },
110 { "ALTIVEC", "altivec", AV_CPU_FLAG_ALTIVEC },
111 { "VSX", "vsx", AV_CPU_FLAG_VSX },
112 { "POWER8", "power8", AV_CPU_FLAG_POWER8 },
114 { "MMX", "mmx", AV_CPU_FLAG_MMX|AV_CPU_FLAG_CMOV },
115 { "MMXEXT", "mmxext", AV_CPU_FLAG_MMXEXT },
116 { "3DNOW", "3dnow", AV_CPU_FLAG_3DNOW },
117 { "3DNOWEXT", "3dnowext", AV_CPU_FLAG_3DNOWEXT },
118 { "SSE", "sse", AV_CPU_FLAG_SSE },
119 { "SSE2", "sse2", AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW },
120 { "SSE3", "sse3", AV_CPU_FLAG_SSE3|AV_CPU_FLAG_SSE3SLOW },
121 { "SSSE3", "ssse3", AV_CPU_FLAG_SSSE3|AV_CPU_FLAG_ATOM },
122 { "SSE4.1", "sse4", AV_CPU_FLAG_SSE4 },
123 { "SSE4.2", "sse42", AV_CPU_FLAG_SSE42 },
124 { "AES-NI", "aesni", AV_CPU_FLAG_AESNI },
125 { "AVX", "avx", AV_CPU_FLAG_AVX },
126 { "XOP", "xop", AV_CPU_FLAG_XOP },
127 { "FMA3", "fma3", AV_CPU_FLAG_FMA3 },
128 { "FMA4", "fma4", AV_CPU_FLAG_FMA4 },
129 { "AVX2", "avx2", AV_CPU_FLAG_AVX2 },
134 typedef struct CheckasmFuncVersion {
135 struct CheckasmFuncVersion *next;
141 } CheckasmFuncVersion;
143 /* Binary search tree node */
144 typedef struct CheckasmFunc {
145 struct CheckasmFunc *child[2];
146 CheckasmFuncVersion versions;
147 uint8_t color; /* 0 = red, 1 = black */
154 CheckasmFunc *current_func;
155 CheckasmFuncVersion *current_func_ver;
156 const char *current_test_name;
157 const char *bench_pattern;
158 int bench_pattern_len;
163 const char *cpu_flag_name;
169 /* Print colored text to stderr if the terminal supports it */
170 static void color_printf(int color, const char *fmt, ...)
172 static int use_color = -1;
175 #if HAVE_SETCONSOLETEXTATTRIBUTE
177 static WORD org_attributes;
180 CONSOLE_SCREEN_BUFFER_INFO con_info;
181 con = GetStdHandle(STD_ERROR_HANDLE);
182 if (con && con != INVALID_HANDLE_VALUE && GetConsoleScreenBufferInfo(con, &con_info)) {
183 org_attributes = con_info.wAttributes;
189 SetConsoleTextAttribute(con, (org_attributes & 0xfff0) | (color & 0x0f));
192 const char *term = getenv("TERM");
193 use_color = term && strcmp(term, "dumb") && isatty(2);
196 fprintf(stderr, "\x1b[%d;3%dm", (color & 0x08) >> 3, color & 0x07);
200 vfprintf(stderr, fmt, arg);
204 #if HAVE_SETCONSOLETEXTATTRIBUTE
205 SetConsoleTextAttribute(con, org_attributes);
207 fprintf(stderr, "\x1b[0m");
212 /* Deallocate a tree */
213 static void destroy_func_tree(CheckasmFunc *f)
216 CheckasmFuncVersion *v = f->versions.next;
218 CheckasmFuncVersion *next = v->next;
223 destroy_func_tree(f->child[0]);
224 destroy_func_tree(f->child[1]);
229 /* Allocate a zero-initialized block, clean up and exit on failure */
230 static void *checkasm_malloc(size_t size)
232 void *ptr = calloc(1, size);
234 fprintf(stderr, "checkasm: malloc failed\n");
235 destroy_func_tree(state.funcs);
241 /* Get the suffix of the specified cpu flag */
242 static const char *cpu_suffix(int cpu)
244 int i = FF_ARRAY_ELEMS(cpus);
247 if (cpu & cpus[i].flag)
248 return cpus[i].suffix;
254 static int cmp_nop(const void *a, const void *b)
256 return *(const uint16_t*)a - *(const uint16_t*)b;
259 /* Measure the overhead of the timing code (in decicycles) */
260 static int measure_nop_time(void)
262 uint16_t nops[10000];
265 for (i = 0; i < 10000; i++) {
266 uint64_t t = AV_READ_TIME();
267 nops[i] = AV_READ_TIME() - t;
270 qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
271 for (i = 2500; i < 7500; i++)
274 return nop_sum / 500;
277 /* Print benchmark results */
278 static void print_benchs(CheckasmFunc *f)
281 print_benchs(f->child[0]);
283 /* Only print functions with at least one assembly version */
284 if (f->versions.cpu || f->versions.next) {
285 CheckasmFuncVersion *v = &f->versions;
288 int decicycles = (10*v->cycles/v->iterations - state.nop_time) / 4;
289 printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
291 } while ((v = v->next));
294 print_benchs(f->child[1]);
299 /* ASCIIbetical sort except preserving natural order for numbers */
300 static int cmp_func_names(const char *a, const char *b)
302 const char *start = a;
303 int ascii_diff, digit_diff;
305 for (; !(ascii_diff = *(const unsigned char*)a - *(const unsigned char*)b) && *a; a++, b++);
306 for (; av_isdigit(*a) && av_isdigit(*b); a++, b++);
308 if (a > start && av_isdigit(a[-1]) && (digit_diff = av_isdigit(*a) - av_isdigit(*b)))
314 /* Perform a tree rotation in the specified direction and return the new root */
315 static CheckasmFunc *rotate_tree(CheckasmFunc *f, int dir)
317 CheckasmFunc *r = f->child[dir^1];
318 f->child[dir^1] = r->child[dir];
325 #define is_red(f) ((f) && !(f)->color)
327 /* Balance a left-leaning red-black tree at the specified node */
328 static void balance_tree(CheckasmFunc **root)
330 CheckasmFunc *f = *root;
332 if (is_red(f->child[0]) && is_red(f->child[1])) {
334 f->child[0]->color = f->child[1]->color = 1;
337 if (!is_red(f->child[0]) && is_red(f->child[1]))
338 *root = rotate_tree(f, 0); /* Rotate left */
339 else if (is_red(f->child[0]) && is_red(f->child[0]->child[0]))
340 *root = rotate_tree(f, 1); /* Rotate right */
343 /* Get a node with the specified name, creating it if it doesn't exist */
344 static CheckasmFunc *get_func(CheckasmFunc **root, const char *name)
346 CheckasmFunc *f = *root;
349 /* Search the tree for a matching node */
350 int cmp = cmp_func_names(name, f->name);
352 f = get_func(&f->child[cmp > 0], name);
354 /* Rebalance the tree on the way up if a new node was inserted */
355 if (!f->versions.func)
359 /* Allocate and insert a new node into the tree */
360 int name_length = strlen(name);
361 f = *root = checkasm_malloc(sizeof(CheckasmFunc) + name_length);
362 memcpy(f->name, name, name_length + 1);
368 /* Perform tests and benchmarks for the specified cpu flag if supported by the host */
369 static void check_cpu_flag(const char *name, int flag)
371 int old_cpu_flag = state.cpu_flag;
373 flag |= old_cpu_flag;
374 av_force_cpu_flags(-1);
375 state.cpu_flag = flag & av_get_cpu_flags();
376 av_force_cpu_flags(state.cpu_flag);
378 if (!flag || state.cpu_flag != old_cpu_flag) {
381 state.cpu_flag_name = name;
382 for (i = 0; tests[i].func; i++) {
383 state.current_test_name = tests[i].name;
389 /* Print the name of the current CPU flag, but only do it once */
390 static void print_cpu_name(void)
392 if (state.cpu_flag_name) {
393 color_printf(COLOR_YELLOW, "%s:\n", state.cpu_flag_name);
394 state.cpu_flag_name = NULL;
398 int main(int argc, char *argv[])
400 int i, seed, ret = 0;
402 if (!tests[0].func || !cpus[0].flag) {
403 fprintf(stderr, "checkasm: no tests to perform\n");
407 if (argc > 1 && !strncmp(argv[1], "--bench", 7)) {
409 fprintf(stderr, "checkasm: --bench is not supported on your system\n");
412 if (argv[1][7] == '=') {
413 state.bench_pattern = argv[1] + 8;
414 state.bench_pattern_len = strlen(state.bench_pattern);
416 state.bench_pattern = "";
422 seed = (argc > 1) ? atoi(argv[1]) : av_get_random_seed();
423 fprintf(stderr, "checkasm: using random seed %u\n", seed);
424 av_lfg_init(&checkasm_lfg, seed);
426 check_cpu_flag(NULL, 0);
427 for (i = 0; cpus[i].flag; i++)
428 check_cpu_flag(cpus[i].name, cpus[i].flag);
430 if (state.num_failed) {
431 fprintf(stderr, "checkasm: %d of %d tests have failed\n", state.num_failed, state.num_checked);
434 fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
436 if (state.bench_pattern) {
437 state.nop_time = measure_nop_time();
438 printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
439 print_benchs(state.funcs);
444 destroy_func_tree(state.funcs);
448 /* Decide whether or not the specified function needs to be tested and
449 * allocate/initialize data structures if needed. Returns a pointer to a
450 * reference function if the function should be tested, otherwise NULL */
451 void *checkasm_check_func(void *func, const char *name, ...)
455 CheckasmFuncVersion *v;
460 name_length = vsnprintf(name_buf, sizeof(name_buf), name, arg);
463 if (!func || name_length <= 0 || name_length >= sizeof(name_buf))
466 state.current_func = get_func(&state.funcs, name_buf);
467 state.funcs->color = 1;
468 v = &state.current_func->versions;
471 CheckasmFuncVersion *prev;
473 /* Only test functions that haven't already been tested */
481 } while ((v = v->next));
483 v = prev->next = checkasm_malloc(sizeof(CheckasmFuncVersion));
488 v->cpu = state.cpu_flag;
489 state.current_func_ver = v;
497 /* Decide whether or not the current function needs to be benchmarked */
498 int checkasm_bench_func(void)
500 return !state.num_failed && state.bench_pattern &&
501 !strncmp(state.current_func->name, state.bench_pattern, state.bench_pattern_len);
504 /* Indicate that the current test has failed */
505 void checkasm_fail_func(const char *msg, ...)
507 if (state.current_func_ver->cpu && state.current_func_ver->ok) {
511 fprintf(stderr, " %s_%s (", state.current_func->name, cpu_suffix(state.current_func_ver->cpu));
513 vfprintf(stderr, msg, arg);
515 fprintf(stderr, ")\n");
517 state.current_func_ver->ok = 0;
522 /* Update benchmark results of the current function */
523 void checkasm_update_bench(int iterations, uint64_t cycles)
525 state.current_func_ver->iterations += iterations;
526 state.current_func_ver->cycles += cycles;
529 /* Print the outcome of all tests performed since the last time this function was called */
530 void checkasm_report(const char *name, ...)
532 static int prev_checked, prev_failed, max_length;
534 if (state.num_checked > prev_checked) {
535 int pad_length = max_length + 4;
539 pad_length -= fprintf(stderr, " - %s.", state.current_test_name);
541 pad_length -= vfprintf(stderr, name, arg);
543 fprintf(stderr, "%*c", FFMAX(pad_length, 0) + 2, '[');
545 if (state.num_failed == prev_failed)
546 color_printf(COLOR_GREEN, "OK");
548 color_printf(COLOR_RED, "FAILED");
549 fprintf(stderr, "]\n");
551 prev_checked = state.num_checked;
552 prev_failed = state.num_failed;
553 } else if (!state.cpu_flag) {
554 /* Calculate the amount of padding required to make the output vertically aligned */
555 int length = strlen(state.current_test_name);
559 length += vsnprintf(NULL, 0, name, arg);
562 if (length > max_length)