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 { "audiodsp", checkasm_check_audiodsp },
75 { "blockdsp", checkasm_check_blockdsp },
78 { "bswapdsp", checkasm_check_bswapdsp },
80 #if CONFIG_DCA_DECODER
81 { "synth_filter", checkasm_check_synth_filter },
84 { "flacdsp", checkasm_check_flacdsp },
87 { "fmtconvert", checkasm_check_fmtconvert },
90 { "h264dsp", checkasm_check_h264dsp },
93 { "h264pred", checkasm_check_h264pred },
96 { "h264qpel", checkasm_check_h264qpel },
98 #if CONFIG_HEVC_DECODER
99 { "hevc_add_res", checkasm_check_hevc_add_res },
100 { "hevc_idct", checkasm_check_hevc_idct },
102 #if CONFIG_JPEG2000_DECODER
103 { "jpeg2000dsp", checkasm_check_jpeg2000dsp },
105 #if CONFIG_HUFFYUVDSP
106 { "llviddsp", checkasm_check_llviddsp },
108 #if CONFIG_PIXBLOCKDSP
109 { "pixblockdsp", checkasm_check_pixblockdsp },
111 #if CONFIG_V210_ENCODER
112 { "v210enc", checkasm_check_v210enc },
115 { "vp8dsp", checkasm_check_vp8dsp },
117 #if CONFIG_VP9_DECODER
118 { "vp9dsp", checkasm_check_vp9dsp },
121 { "videodsp", checkasm_check_videodsp },
125 #if CONFIG_BLEND_FILTER
126 { "vf_blend", checkasm_check_blend },
128 #if CONFIG_COLORSPACE_FILTER
129 { "vf_colorspace", checkasm_check_colorspace },
135 /* List of cpu flags to check */
136 static const struct {
142 { "ARMV8", "armv8", AV_CPU_FLAG_ARMV8 },
143 { "NEON", "neon", AV_CPU_FLAG_NEON },
145 { "ARMV5TE", "armv5te", AV_CPU_FLAG_ARMV5TE },
146 { "ARMV6", "armv6", AV_CPU_FLAG_ARMV6 },
147 { "ARMV6T2", "armv6t2", AV_CPU_FLAG_ARMV6T2 },
148 { "VFP", "vfp", AV_CPU_FLAG_VFP },
149 { "VFP_VM", "vfp_vm", AV_CPU_FLAG_VFP_VM },
150 { "VFPV3", "vfp3", AV_CPU_FLAG_VFPV3 },
151 { "NEON", "neon", AV_CPU_FLAG_NEON },
153 { "ALTIVEC", "altivec", AV_CPU_FLAG_ALTIVEC },
154 { "VSX", "vsx", AV_CPU_FLAG_VSX },
155 { "POWER8", "power8", AV_CPU_FLAG_POWER8 },
157 { "MMX", "mmx", AV_CPU_FLAG_MMX|AV_CPU_FLAG_CMOV },
158 { "MMXEXT", "mmxext", AV_CPU_FLAG_MMXEXT },
159 { "3DNOW", "3dnow", AV_CPU_FLAG_3DNOW },
160 { "3DNOWEXT", "3dnowext", AV_CPU_FLAG_3DNOWEXT },
161 { "SSE", "sse", AV_CPU_FLAG_SSE },
162 { "SSE2", "sse2", AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW },
163 { "SSE3", "sse3", AV_CPU_FLAG_SSE3|AV_CPU_FLAG_SSE3SLOW },
164 { "SSSE3", "ssse3", AV_CPU_FLAG_SSSE3|AV_CPU_FLAG_ATOM },
165 { "SSE4.1", "sse4", AV_CPU_FLAG_SSE4 },
166 { "SSE4.2", "sse42", AV_CPU_FLAG_SSE42 },
167 { "AES-NI", "aesni", AV_CPU_FLAG_AESNI },
168 { "AVX", "avx", AV_CPU_FLAG_AVX },
169 { "XOP", "xop", AV_CPU_FLAG_XOP },
170 { "FMA3", "fma3", AV_CPU_FLAG_FMA3 },
171 { "FMA4", "fma4", AV_CPU_FLAG_FMA4 },
172 { "AVX2", "avx2", AV_CPU_FLAG_AVX2 },
177 typedef struct CheckasmFuncVersion {
178 struct CheckasmFuncVersion *next;
184 } CheckasmFuncVersion;
186 /* Binary search tree node */
187 typedef struct CheckasmFunc {
188 struct CheckasmFunc *child[2];
189 CheckasmFuncVersion versions;
190 uint8_t color; /* 0 = red, 1 = black */
197 CheckasmFunc *current_func;
198 CheckasmFuncVersion *current_func_ver;
199 const char *current_test_name;
200 const char *bench_pattern;
201 int bench_pattern_len;
206 const char *cpu_flag_name;
212 /* float compare support code */
213 static int is_negative(union av_intfloat32 u)
218 int float_near_ulp(float a, float b, unsigned max_ulp)
220 union av_intfloat32 x, y;
225 if (is_negative(x) != is_negative(y)) {
226 // handle -0.0 == +0.0
230 if (llabs((int64_t)x.i - y.i) <= max_ulp)
236 int float_near_ulp_array(const float *a, const float *b, unsigned max_ulp,
241 for (i = 0; i < len; i++) {
242 if (!float_near_ulp(a[i], b[i], max_ulp))
248 int float_near_abs_eps(float a, float b, float eps)
250 float abs_diff = fabsf(a - b);
252 return abs_diff < eps;
255 int float_near_abs_eps_array(const float *a, const float *b, float eps,
260 for (i = 0; i < len; i++) {
261 if (!float_near_abs_eps(a[i], b[i], eps))
267 int float_near_abs_eps_ulp(float a, float b, float eps, unsigned max_ulp)
269 return float_near_ulp(a, b, max_ulp) || float_near_abs_eps(a, b, eps);
272 int float_near_abs_eps_array_ulp(const float *a, const float *b, float eps,
273 unsigned max_ulp, unsigned len)
277 for (i = 0; i < len; i++) {
278 if (!float_near_abs_eps_ulp(a[i], b[i], eps, max_ulp))
284 /* Print colored text to stderr if the terminal supports it */
285 static void color_printf(int color, const char *fmt, ...)
287 static int use_color = -1;
290 #if HAVE_SETCONSOLETEXTATTRIBUTE
292 static WORD org_attributes;
295 CONSOLE_SCREEN_BUFFER_INFO con_info;
296 con = GetStdHandle(STD_ERROR_HANDLE);
297 if (con && con != INVALID_HANDLE_VALUE && GetConsoleScreenBufferInfo(con, &con_info)) {
298 org_attributes = con_info.wAttributes;
304 SetConsoleTextAttribute(con, (org_attributes & 0xfff0) | (color & 0x0f));
307 const char *term = getenv("TERM");
308 use_color = term && strcmp(term, "dumb") && isatty(2);
311 fprintf(stderr, "\x1b[%d;3%dm", (color & 0x08) >> 3, color & 0x07);
315 vfprintf(stderr, fmt, arg);
319 #if HAVE_SETCONSOLETEXTATTRIBUTE
320 SetConsoleTextAttribute(con, org_attributes);
322 fprintf(stderr, "\x1b[0m");
327 /* Deallocate a tree */
328 static void destroy_func_tree(CheckasmFunc *f)
331 CheckasmFuncVersion *v = f->versions.next;
333 CheckasmFuncVersion *next = v->next;
338 destroy_func_tree(f->child[0]);
339 destroy_func_tree(f->child[1]);
344 /* Allocate a zero-initialized block, clean up and exit on failure */
345 static void *checkasm_malloc(size_t size)
347 void *ptr = calloc(1, size);
349 fprintf(stderr, "checkasm: malloc failed\n");
350 destroy_func_tree(state.funcs);
356 /* Get the suffix of the specified cpu flag */
357 static const char *cpu_suffix(int cpu)
359 int i = FF_ARRAY_ELEMS(cpus);
362 if (cpu & cpus[i].flag)
363 return cpus[i].suffix;
369 static int cmp_nop(const void *a, const void *b)
371 return *(const uint16_t*)a - *(const uint16_t*)b;
374 /* Measure the overhead of the timing code (in decicycles) */
375 static int measure_nop_time(void)
377 uint16_t nops[10000];
380 for (i = 0; i < 10000; i++) {
381 uint64_t t = AV_READ_TIME();
382 nops[i] = AV_READ_TIME() - t;
385 qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
386 for (i = 2500; i < 7500; i++)
389 return nop_sum / 500;
392 /* Print benchmark results */
393 static void print_benchs(CheckasmFunc *f)
396 print_benchs(f->child[0]);
398 /* Only print functions with at least one assembly version */
399 if (f->versions.cpu || f->versions.next) {
400 CheckasmFuncVersion *v = &f->versions;
403 int decicycles = (10*v->cycles/v->iterations - state.nop_time) / 4;
404 printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
406 } while ((v = v->next));
409 print_benchs(f->child[1]);
414 /* ASCIIbetical sort except preserving natural order for numbers */
415 static int cmp_func_names(const char *a, const char *b)
417 const char *start = a;
418 int ascii_diff, digit_diff;
420 for (; !(ascii_diff = *(const unsigned char*)a - *(const unsigned char*)b) && *a; a++, b++);
421 for (; av_isdigit(*a) && av_isdigit(*b); a++, b++);
423 if (a > start && av_isdigit(a[-1]) && (digit_diff = av_isdigit(*a) - av_isdigit(*b)))
429 /* Perform a tree rotation in the specified direction and return the new root */
430 static CheckasmFunc *rotate_tree(CheckasmFunc *f, int dir)
432 CheckasmFunc *r = f->child[dir^1];
433 f->child[dir^1] = r->child[dir];
440 #define is_red(f) ((f) && !(f)->color)
442 /* Balance a left-leaning red-black tree at the specified node */
443 static void balance_tree(CheckasmFunc **root)
445 CheckasmFunc *f = *root;
447 if (is_red(f->child[0]) && is_red(f->child[1])) {
449 f->child[0]->color = f->child[1]->color = 1;
452 if (!is_red(f->child[0]) && is_red(f->child[1]))
453 *root = rotate_tree(f, 0); /* Rotate left */
454 else if (is_red(f->child[0]) && is_red(f->child[0]->child[0]))
455 *root = rotate_tree(f, 1); /* Rotate right */
458 /* Get a node with the specified name, creating it if it doesn't exist */
459 static CheckasmFunc *get_func(CheckasmFunc **root, const char *name)
461 CheckasmFunc *f = *root;
464 /* Search the tree for a matching node */
465 int cmp = cmp_func_names(name, f->name);
467 f = get_func(&f->child[cmp > 0], name);
469 /* Rebalance the tree on the way up if a new node was inserted */
470 if (!f->versions.func)
474 /* Allocate and insert a new node into the tree */
475 int name_length = strlen(name);
476 f = *root = checkasm_malloc(sizeof(CheckasmFunc) + name_length);
477 memcpy(f->name, name, name_length + 1);
483 /* Perform tests and benchmarks for the specified cpu flag if supported by the host */
484 static void check_cpu_flag(const char *name, int flag)
486 int old_cpu_flag = state.cpu_flag;
488 flag |= old_cpu_flag;
489 av_force_cpu_flags(-1);
490 state.cpu_flag = flag & av_get_cpu_flags();
491 av_force_cpu_flags(state.cpu_flag);
493 if (!flag || state.cpu_flag != old_cpu_flag) {
496 state.cpu_flag_name = name;
497 for (i = 0; tests[i].func; i++) {
498 state.current_test_name = tests[i].name;
504 /* Print the name of the current CPU flag, but only do it once */
505 static void print_cpu_name(void)
507 if (state.cpu_flag_name) {
508 color_printf(COLOR_YELLOW, "%s:\n", state.cpu_flag_name);
509 state.cpu_flag_name = NULL;
513 int main(int argc, char *argv[])
518 #if ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
519 if (have_vfp(av_get_cpu_flags()) || have_neon(av_get_cpu_flags()))
520 checkasm_checked_call = checkasm_checked_call_vfp;
523 if (!tests[0].func || !cpus[0].flag) {
524 fprintf(stderr, "checkasm: no tests to perform\n");
528 if (argc > 1 && !strncmp(argv[1], "--bench", 7)) {
530 fprintf(stderr, "checkasm: --bench is not supported on your system\n");
533 if (argv[1][7] == '=') {
534 state.bench_pattern = argv[1] + 8;
535 state.bench_pattern_len = strlen(state.bench_pattern);
537 state.bench_pattern = "";
543 seed = (argc > 1) ? strtoul(argv[1], NULL, 10) : av_get_random_seed();
544 fprintf(stderr, "checkasm: using random seed %u\n", seed);
545 av_lfg_init(&checkasm_lfg, seed);
547 check_cpu_flag(NULL, 0);
548 for (i = 0; cpus[i].flag; i++)
549 check_cpu_flag(cpus[i].name, cpus[i].flag);
551 if (state.num_failed) {
552 fprintf(stderr, "checkasm: %d of %d tests have failed\n", state.num_failed, state.num_checked);
555 fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
557 if (state.bench_pattern) {
558 state.nop_time = measure_nop_time();
559 printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
560 print_benchs(state.funcs);
565 destroy_func_tree(state.funcs);
569 /* Decide whether or not the specified function needs to be tested and
570 * allocate/initialize data structures if needed. Returns a pointer to a
571 * reference function if the function should be tested, otherwise NULL */
572 void *checkasm_check_func(void *func, const char *name, ...)
576 CheckasmFuncVersion *v;
581 name_length = vsnprintf(name_buf, sizeof(name_buf), name, arg);
584 if (!func || name_length <= 0 || name_length >= sizeof(name_buf))
587 state.current_func = get_func(&state.funcs, name_buf);
588 state.funcs->color = 1;
589 v = &state.current_func->versions;
592 CheckasmFuncVersion *prev;
594 /* Only test functions that haven't already been tested */
602 } while ((v = v->next));
604 v = prev->next = checkasm_malloc(sizeof(CheckasmFuncVersion));
609 v->cpu = state.cpu_flag;
610 state.current_func_ver = v;
618 /* Decide whether or not the current function needs to be benchmarked */
619 int checkasm_bench_func(void)
621 return !state.num_failed && state.bench_pattern &&
622 !strncmp(state.current_func->name, state.bench_pattern, state.bench_pattern_len);
625 /* Indicate that the current test has failed */
626 void checkasm_fail_func(const char *msg, ...)
628 if (state.current_func_ver->cpu && state.current_func_ver->ok) {
632 fprintf(stderr, " %s_%s (", state.current_func->name, cpu_suffix(state.current_func_ver->cpu));
634 vfprintf(stderr, msg, arg);
636 fprintf(stderr, ")\n");
638 state.current_func_ver->ok = 0;
643 /* Update benchmark results of the current function */
644 void checkasm_update_bench(int iterations, uint64_t cycles)
646 state.current_func_ver->iterations += iterations;
647 state.current_func_ver->cycles += cycles;
650 /* Print the outcome of all tests performed since the last time this function was called */
651 void checkasm_report(const char *name, ...)
653 static int prev_checked, prev_failed, max_length;
655 if (state.num_checked > prev_checked) {
656 int pad_length = max_length + 4;
660 pad_length -= fprintf(stderr, " - %s.", state.current_test_name);
662 pad_length -= vfprintf(stderr, name, arg);
664 fprintf(stderr, "%*c", FFMAX(pad_length, 0) + 2, '[');
666 if (state.num_failed == prev_failed)
667 color_printf(COLOR_GREEN, "OK");
669 color_printf(COLOR_RED, "FAILED");
670 fprintf(stderr, "]\n");
672 prev_checked = state.num_checked;
673 prev_failed = state.num_failed;
674 } else if (!state.cpu_flag) {
675 /* Calculate the amount of padding required to make the output vertically aligned */
676 int length = strlen(state.current_test_name);
680 length += vsnprintf(NULL, 0, name, arg);
683 if (length > max_length)