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/intfloat.h"
31 #include "libavutil/random_seed.h"
37 #if HAVE_SETCONSOLETEXTATTRIBUTE
39 #define COLOR_RED FOREGROUND_RED
40 #define COLOR_GREEN FOREGROUND_GREEN
41 #define COLOR_YELLOW (FOREGROUND_RED|FOREGROUND_GREEN)
45 #define COLOR_YELLOW 3
56 #if ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
57 #include "libavutil/arm/cpu.h"
59 void (*checkasm_checked_call)(void *func, int dummy, ...) = checkasm_checked_call_novfp;
62 /* List of tests to invoke */
68 #if CONFIG_ALAC_DECODER
69 { "alacdsp", checkasm_check_alacdsp },
72 { "bswapdsp", checkasm_check_bswapdsp },
74 #if CONFIG_DCA_DECODER
75 { "synth_filter", checkasm_check_synth_filter },
78 { "flacdsp", checkasm_check_flacdsp },
81 { "fmtconvert", checkasm_check_fmtconvert },
84 { "h264pred", checkasm_check_h264pred },
87 { "h264qpel", checkasm_check_h264qpel },
89 #if CONFIG_JPEG2000_DECODER
90 { "jpeg2000dsp", checkasm_check_jpeg2000dsp },
92 #if CONFIG_PIXBLOCKDSP
93 { "pixblockdsp", checkasm_check_pixblockdsp },
95 #if CONFIG_V210_ENCODER
96 { "v210enc", checkasm_check_v210enc },
98 #if CONFIG_VP9_DECODER
99 { "vp9dsp", checkasm_check_vp9dsp },
102 { "videodsp", checkasm_check_videodsp },
106 #if CONFIG_BLEND_FILTER
107 { "vf_blend", checkasm_check_blend },
109 #if CONFIG_COLORSPACE_FILTER
110 { "vf_colorspace", checkasm_check_colorspace },
116 /* List of cpu flags to check */
117 static const struct {
123 { "ARMV8", "armv8", AV_CPU_FLAG_ARMV8 },
124 { "NEON", "neon", AV_CPU_FLAG_NEON },
126 { "ARMV5TE", "armv5te", AV_CPU_FLAG_ARMV5TE },
127 { "ARMV6", "armv6", AV_CPU_FLAG_ARMV6 },
128 { "ARMV6T2", "armv6t2", AV_CPU_FLAG_ARMV6T2 },
129 { "VFP", "vfp", AV_CPU_FLAG_VFP },
130 { "VFP_VM", "vfp_vm", AV_CPU_FLAG_VFP_VM },
131 { "VFPV3", "vfp3", AV_CPU_FLAG_VFPV3 },
132 { "NEON", "neon", AV_CPU_FLAG_NEON },
134 { "ALTIVEC", "altivec", AV_CPU_FLAG_ALTIVEC },
135 { "VSX", "vsx", AV_CPU_FLAG_VSX },
136 { "POWER8", "power8", AV_CPU_FLAG_POWER8 },
138 { "MMX", "mmx", AV_CPU_FLAG_MMX|AV_CPU_FLAG_CMOV },
139 { "MMXEXT", "mmxext", AV_CPU_FLAG_MMXEXT },
140 { "3DNOW", "3dnow", AV_CPU_FLAG_3DNOW },
141 { "3DNOWEXT", "3dnowext", AV_CPU_FLAG_3DNOWEXT },
142 { "SSE", "sse", AV_CPU_FLAG_SSE },
143 { "SSE2", "sse2", AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW },
144 { "SSE3", "sse3", AV_CPU_FLAG_SSE3|AV_CPU_FLAG_SSE3SLOW },
145 { "SSSE3", "ssse3", AV_CPU_FLAG_SSSE3|AV_CPU_FLAG_ATOM },
146 { "SSE4.1", "sse4", AV_CPU_FLAG_SSE4 },
147 { "SSE4.2", "sse42", AV_CPU_FLAG_SSE42 },
148 { "AES-NI", "aesni", AV_CPU_FLAG_AESNI },
149 { "AVX", "avx", AV_CPU_FLAG_AVX },
150 { "XOP", "xop", AV_CPU_FLAG_XOP },
151 { "FMA3", "fma3", AV_CPU_FLAG_FMA3 },
152 { "FMA4", "fma4", AV_CPU_FLAG_FMA4 },
153 { "AVX2", "avx2", AV_CPU_FLAG_AVX2 },
158 typedef struct CheckasmFuncVersion {
159 struct CheckasmFuncVersion *next;
165 } CheckasmFuncVersion;
167 /* Binary search tree node */
168 typedef struct CheckasmFunc {
169 struct CheckasmFunc *child[2];
170 CheckasmFuncVersion versions;
171 uint8_t color; /* 0 = red, 1 = black */
178 CheckasmFunc *current_func;
179 CheckasmFuncVersion *current_func_ver;
180 const char *current_test_name;
181 const char *bench_pattern;
182 int bench_pattern_len;
187 const char *cpu_flag_name;
193 /* float compare support code */
194 static int is_negative(union av_intfloat32 u)
199 int float_near_ulp(float a, float b, unsigned max_ulp)
201 union av_intfloat32 x, y;
206 if (is_negative(x) != is_negative(y)) {
207 // handle -0.0 == +0.0
211 if (abs(x.i - y.i) <= max_ulp)
217 int float_near_ulp_array(const float *a, const float *b, unsigned max_ulp,
222 for (i = 0; i < len; i++) {
223 if (!float_near_ulp(a[i], b[i], max_ulp))
229 int float_near_abs_eps(float a, float b, float eps)
231 float abs_diff = fabsf(a - b);
233 return abs_diff < eps;
236 int float_near_abs_eps_array(const float *a, const float *b, float eps,
241 for (i = 0; i < len; i++) {
242 if (!float_near_abs_eps(a[i], b[i], eps))
248 int float_near_abs_eps_ulp(float a, float b, float eps, unsigned max_ulp)
250 return float_near_ulp(a, b, max_ulp) || float_near_abs_eps(a, b, eps);
253 int float_near_abs_eps_array_ulp(const float *a, const float *b, float eps,
254 unsigned max_ulp, unsigned len)
258 for (i = 0; i < len; i++) {
259 if (!float_near_abs_eps_ulp(a[i], b[i], eps, max_ulp))
265 /* Print colored text to stderr if the terminal supports it */
266 static void color_printf(int color, const char *fmt, ...)
268 static int use_color = -1;
271 #if HAVE_SETCONSOLETEXTATTRIBUTE
273 static WORD org_attributes;
276 CONSOLE_SCREEN_BUFFER_INFO con_info;
277 con = GetStdHandle(STD_ERROR_HANDLE);
278 if (con && con != INVALID_HANDLE_VALUE && GetConsoleScreenBufferInfo(con, &con_info)) {
279 org_attributes = con_info.wAttributes;
285 SetConsoleTextAttribute(con, (org_attributes & 0xfff0) | (color & 0x0f));
288 const char *term = getenv("TERM");
289 use_color = term && strcmp(term, "dumb") && isatty(2);
292 fprintf(stderr, "\x1b[%d;3%dm", (color & 0x08) >> 3, color & 0x07);
296 vfprintf(stderr, fmt, arg);
300 #if HAVE_SETCONSOLETEXTATTRIBUTE
301 SetConsoleTextAttribute(con, org_attributes);
303 fprintf(stderr, "\x1b[0m");
308 /* Deallocate a tree */
309 static void destroy_func_tree(CheckasmFunc *f)
312 CheckasmFuncVersion *v = f->versions.next;
314 CheckasmFuncVersion *next = v->next;
319 destroy_func_tree(f->child[0]);
320 destroy_func_tree(f->child[1]);
325 /* Allocate a zero-initialized block, clean up and exit on failure */
326 static void *checkasm_malloc(size_t size)
328 void *ptr = calloc(1, size);
330 fprintf(stderr, "checkasm: malloc failed\n");
331 destroy_func_tree(state.funcs);
337 /* Get the suffix of the specified cpu flag */
338 static const char *cpu_suffix(int cpu)
340 int i = FF_ARRAY_ELEMS(cpus);
343 if (cpu & cpus[i].flag)
344 return cpus[i].suffix;
350 static int cmp_nop(const void *a, const void *b)
352 return *(const uint16_t*)a - *(const uint16_t*)b;
355 /* Measure the overhead of the timing code (in decicycles) */
356 static int measure_nop_time(void)
358 uint16_t nops[10000];
361 for (i = 0; i < 10000; i++) {
362 uint64_t t = AV_READ_TIME();
363 nops[i] = AV_READ_TIME() - t;
366 qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
367 for (i = 2500; i < 7500; i++)
370 return nop_sum / 500;
373 /* Print benchmark results */
374 static void print_benchs(CheckasmFunc *f)
377 print_benchs(f->child[0]);
379 /* Only print functions with at least one assembly version */
380 if (f->versions.cpu || f->versions.next) {
381 CheckasmFuncVersion *v = &f->versions;
384 int decicycles = (10*v->cycles/v->iterations - state.nop_time) / 4;
385 printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
387 } while ((v = v->next));
390 print_benchs(f->child[1]);
395 /* ASCIIbetical sort except preserving natural order for numbers */
396 static int cmp_func_names(const char *a, const char *b)
398 const char *start = a;
399 int ascii_diff, digit_diff;
401 for (; !(ascii_diff = *(const unsigned char*)a - *(const unsigned char*)b) && *a; a++, b++);
402 for (; av_isdigit(*a) && av_isdigit(*b); a++, b++);
404 if (a > start && av_isdigit(a[-1]) && (digit_diff = av_isdigit(*a) - av_isdigit(*b)))
410 /* Perform a tree rotation in the specified direction and return the new root */
411 static CheckasmFunc *rotate_tree(CheckasmFunc *f, int dir)
413 CheckasmFunc *r = f->child[dir^1];
414 f->child[dir^1] = r->child[dir];
421 #define is_red(f) ((f) && !(f)->color)
423 /* Balance a left-leaning red-black tree at the specified node */
424 static void balance_tree(CheckasmFunc **root)
426 CheckasmFunc *f = *root;
428 if (is_red(f->child[0]) && is_red(f->child[1])) {
430 f->child[0]->color = f->child[1]->color = 1;
433 if (!is_red(f->child[0]) && is_red(f->child[1]))
434 *root = rotate_tree(f, 0); /* Rotate left */
435 else if (is_red(f->child[0]) && is_red(f->child[0]->child[0]))
436 *root = rotate_tree(f, 1); /* Rotate right */
439 /* Get a node with the specified name, creating it if it doesn't exist */
440 static CheckasmFunc *get_func(CheckasmFunc **root, const char *name)
442 CheckasmFunc *f = *root;
445 /* Search the tree for a matching node */
446 int cmp = cmp_func_names(name, f->name);
448 f = get_func(&f->child[cmp > 0], name);
450 /* Rebalance the tree on the way up if a new node was inserted */
451 if (!f->versions.func)
455 /* Allocate and insert a new node into the tree */
456 int name_length = strlen(name);
457 f = *root = checkasm_malloc(sizeof(CheckasmFunc) + name_length);
458 memcpy(f->name, name, name_length + 1);
464 /* Perform tests and benchmarks for the specified cpu flag if supported by the host */
465 static void check_cpu_flag(const char *name, int flag)
467 int old_cpu_flag = state.cpu_flag;
469 flag |= old_cpu_flag;
470 av_force_cpu_flags(-1);
471 state.cpu_flag = flag & av_get_cpu_flags();
472 av_force_cpu_flags(state.cpu_flag);
474 if (!flag || state.cpu_flag != old_cpu_flag) {
477 state.cpu_flag_name = name;
478 for (i = 0; tests[i].func; i++) {
479 state.current_test_name = tests[i].name;
485 /* Print the name of the current CPU flag, but only do it once */
486 static void print_cpu_name(void)
488 if (state.cpu_flag_name) {
489 color_printf(COLOR_YELLOW, "%s:\n", state.cpu_flag_name);
490 state.cpu_flag_name = NULL;
494 int main(int argc, char *argv[])
496 int i, seed, ret = 0;
498 #if ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
499 if (have_vfp(av_get_cpu_flags()) || have_neon(av_get_cpu_flags()))
500 checkasm_checked_call = checkasm_checked_call_vfp;
503 if (!tests[0].func || !cpus[0].flag) {
504 fprintf(stderr, "checkasm: no tests to perform\n");
508 if (argc > 1 && !strncmp(argv[1], "--bench", 7)) {
510 fprintf(stderr, "checkasm: --bench is not supported on your system\n");
513 if (argv[1][7] == '=') {
514 state.bench_pattern = argv[1] + 8;
515 state.bench_pattern_len = strlen(state.bench_pattern);
517 state.bench_pattern = "";
523 seed = (argc > 1) ? atoi(argv[1]) : av_get_random_seed();
524 fprintf(stderr, "checkasm: using random seed %u\n", seed);
525 av_lfg_init(&checkasm_lfg, seed);
527 check_cpu_flag(NULL, 0);
528 for (i = 0; cpus[i].flag; i++)
529 check_cpu_flag(cpus[i].name, cpus[i].flag);
531 if (state.num_failed) {
532 fprintf(stderr, "checkasm: %d of %d tests have failed\n", state.num_failed, state.num_checked);
535 fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
537 if (state.bench_pattern) {
538 state.nop_time = measure_nop_time();
539 printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
540 print_benchs(state.funcs);
545 destroy_func_tree(state.funcs);
549 /* Decide whether or not the specified function needs to be tested and
550 * allocate/initialize data structures if needed. Returns a pointer to a
551 * reference function if the function should be tested, otherwise NULL */
552 void *checkasm_check_func(void *func, const char *name, ...)
556 CheckasmFuncVersion *v;
561 name_length = vsnprintf(name_buf, sizeof(name_buf), name, arg);
564 if (!func || name_length <= 0 || name_length >= sizeof(name_buf))
567 state.current_func = get_func(&state.funcs, name_buf);
568 state.funcs->color = 1;
569 v = &state.current_func->versions;
572 CheckasmFuncVersion *prev;
574 /* Only test functions that haven't already been tested */
582 } while ((v = v->next));
584 v = prev->next = checkasm_malloc(sizeof(CheckasmFuncVersion));
589 v->cpu = state.cpu_flag;
590 state.current_func_ver = v;
598 /* Decide whether or not the current function needs to be benchmarked */
599 int checkasm_bench_func(void)
601 return !state.num_failed && state.bench_pattern &&
602 !strncmp(state.current_func->name, state.bench_pattern, state.bench_pattern_len);
605 /* Indicate that the current test has failed */
606 void checkasm_fail_func(const char *msg, ...)
608 if (state.current_func_ver->cpu && state.current_func_ver->ok) {
612 fprintf(stderr, " %s_%s (", state.current_func->name, cpu_suffix(state.current_func_ver->cpu));
614 vfprintf(stderr, msg, arg);
616 fprintf(stderr, ")\n");
618 state.current_func_ver->ok = 0;
623 /* Update benchmark results of the current function */
624 void checkasm_update_bench(int iterations, uint64_t cycles)
626 state.current_func_ver->iterations += iterations;
627 state.current_func_ver->cycles += cycles;
630 /* Print the outcome of all tests performed since the last time this function was called */
631 void checkasm_report(const char *name, ...)
633 static int prev_checked, prev_failed, max_length;
635 if (state.num_checked > prev_checked) {
636 int pad_length = max_length + 4;
640 pad_length -= fprintf(stderr, " - %s.", state.current_test_name);
642 pad_length -= vfprintf(stderr, name, arg);
644 fprintf(stderr, "%*c", FFMAX(pad_length, 0) + 2, '[');
646 if (state.num_failed == prev_failed)
647 color_printf(COLOR_GREEN, "OK");
649 color_printf(COLOR_RED, "FAILED");
650 fprintf(stderr, "]\n");
652 prev_checked = state.num_checked;
653 prev_failed = state.num_failed;
654 } else if (!state.cpu_flag) {
655 /* Calculate the amount of padding required to make the output vertically aligned */
656 int length = strlen(state.current_test_name);
660 length += vsnprintf(NULL, 0, name, arg);
663 if (length > max_length)