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/avstring.h"
29 #include "libavutil/cpu.h"
30 #include "libavutil/lfg.h"
31 #include "libavutil/timer.h"
33 void checkasm_check_bswapdsp(void);
34 void checkasm_check_h264pred(void);
35 void checkasm_check_h264qpel(void);
36 void checkasm_check_hevc_mc(void);
37 void checkasm_check_v210enc(void);
39 void *checkasm_check_func(void *func, const char *name, ...) av_printf_format(2, 3);
40 int checkasm_bench_func(void);
41 void checkasm_fail_func(const char *msg, ...) av_printf_format(1, 2);
42 void checkasm_update_bench(int iterations, uint64_t cycles);
43 void checkasm_report(const char *name, ...) av_printf_format(1, 2);
45 /* float compare utilities */
46 int float_near_ulp(float a, float b, unsigned max_ulp);
47 int float_near_abs_eps(float a, float b, float eps);
48 int float_near_abs_eps_ulp(float a, float b, float eps, unsigned max_ulp);
49 int float_near_ulp_array(const float *a, const float *b, unsigned max_ulp,
51 int float_near_abs_eps_array(const float *a, const float *b, float eps,
53 int float_near_abs_eps_array_ulp(const float *a, const float *b, float eps,
54 unsigned max_ulp, unsigned len);
56 extern AVLFG checkasm_lfg;
57 #define rnd() av_lfg_get(&checkasm_lfg)
59 static av_unused void *func_ref, *func_new;
61 #define BENCH_RUNS 1000 /* Trade-off between accuracy and speed */
63 /* Decide whether or not the specified function needs to be tested */
64 #define check_func(func, ...) (func_ref = checkasm_check_func((func_new = func), __VA_ARGS__))
66 /* Declare the function prototype. The first argument is the return value, the remaining
67 * arguments are the function parameters. Naming parameters is optional. */
68 #define declare_func(ret, ...) declare_new(ret, __VA_ARGS__) typedef ret func_type(__VA_ARGS__)
69 #define declare_func_emms(cpu_flags, ret, ...) declare_new_emms(cpu_flags, ret, __VA_ARGS__) typedef ret func_type(__VA_ARGS__)
71 /* Indicate that the current test has failed */
72 #define fail() checkasm_fail_func("%s:%d", av_basename(__FILE__), __LINE__)
74 /* Print the test outcome */
75 #define report checkasm_report
77 /* Call the reference function */
78 #define call_ref(...) ((func_type *)func_ref)(__VA_ARGS__)
80 #if ARCH_X86 && HAVE_YASM
81 /* Verifies that clobbered callee-saved registers are properly saved and restored
82 * and that either no MMX registers are touched or emms is issued */
83 void checkasm_checked_call(void *func, ...);
84 /* Verifies that clobbered callee-saved registers are properly saved and restored
85 * and issues emms for asm functions which are not required to do so */
86 void checkasm_checked_call_emms(void *func, ...);
89 /* Evil hack: detect incorrect assumptions that 32-bit ints are zero-extended to 64-bit.
90 * This is done by clobbering the stack with junk around the stack pointer and calling the
91 * assembly function through checked_call() with added dummy arguments which forces all
92 * real arguments to be passed on the stack and not in registers. For 32-bit arguments the
93 * upper half of the 64-bit register locations on the stack will now contain junk which will
94 * cause misbehaving functions to either produce incorrect output or segfault. Note that
95 * even though this works extremely well in practice, it's technically not guaranteed
96 * and false negatives is theoretically possible, but there can never be any false positives.
98 void checkasm_stack_clobber(uint64_t clobber, ...);
99 #define declare_new(ret, ...) ret (*checked_call)(void *, int, int, int, int, int, __VA_ARGS__)\
100 = (void *)checkasm_checked_call;
101 #define declare_new_emms(cpu_flags, ret, ...) \
102 ret (*checked_call)(void *, int, int, int, int, int, __VA_ARGS__) = \
103 ((cpu_flags) & av_get_cpu_flags()) ? (void *)checkasm_checked_call_emms : \
104 (void *)checkasm_checked_call;
105 #define CLOB (UINT64_C(0xdeadbeefdeadbeef))
106 #define call_new(...) (checkasm_stack_clobber(CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,\
107 CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB),\
108 checked_call(func_new, 0, 0, 0, 0, 0, __VA_ARGS__))
110 #define declare_new(ret, ...) ret (*checked_call)(void *, __VA_ARGS__) = (void *)checkasm_checked_call;
111 #define declare_new_emms(cpu_flags, ret, ...) ret (*checked_call)(void *, __VA_ARGS__) = \
112 ((cpu_flags) & av_get_cpu_flags()) ? (void *)checkasm_checked_call_emms : \
113 (void *)checkasm_checked_call;
114 #define call_new(...) checked_call(func_new, __VA_ARGS__)
117 #define declare_new(ret, ...)
118 #define declare_new_emms(cpu_flags, ret, ...)
119 /* Call the function */
120 #define call_new(...) ((func_type *)func_new)(__VA_ARGS__)
123 /* Benchmark the function */
125 #define bench_new(...)\
127 if (checkasm_bench_func()) {\
128 func_type *tfunc = func_new;\
131 for (ti = 0; ti < BENCH_RUNS; ti++) {\
132 uint64_t t = AV_READ_TIME();\
137 t = AV_READ_TIME() - t;\
138 if (t*tcount <= tsum*4 && ti > 0) {\
143 checkasm_update_bench(tcount, tsum);\
147 #define bench_new(...) while(0)