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.
23 #ifndef TESTS_CHECKASM_CHECKASM_H
24 #define TESTS_CHECKASM_CHECKASM_H
30 #include <unistd.h> // read(3)
31 #include <sys/ioctl.h>
32 #include <asm/unistd.h>
33 #include <linux/perf_event.h>
34 #elif CONFIG_MACOS_KPERF
35 #include "libavutil/macos_kperf.h"
38 #include "libavutil/avstring.h"
39 #include "libavutil/cpu.h"
40 #include "libavutil/emms.h"
41 #include "libavutil/internal.h"
42 #include "libavutil/lfg.h"
43 #include "libavutil/timer.h"
45 void checkasm_check_aacpsdsp(void);
46 void checkasm_check_afir(void);
47 void checkasm_check_alacdsp(void);
48 void checkasm_check_audiodsp(void);
49 void checkasm_check_av_tx(void);
50 void checkasm_check_blend(void);
51 void checkasm_check_blockdsp(void);
52 void checkasm_check_bswapdsp(void);
53 void checkasm_check_colorspace(void);
54 void checkasm_check_exrdsp(void);
55 void checkasm_check_fixed_dsp(void);
56 void checkasm_check_flacdsp(void);
57 void checkasm_check_float_dsp(void);
58 void checkasm_check_fmtconvert(void);
59 void checkasm_check_g722dsp(void);
60 void checkasm_check_h264chroma(void);
61 void checkasm_check_h264dsp(void);
62 void checkasm_check_h264pred(void);
63 void checkasm_check_h264qpel(void);
64 void checkasm_check_hevc_add_res(void);
65 void checkasm_check_hevc_deblock(void);
66 void checkasm_check_hevc_idct(void);
67 void checkasm_check_hevc_pel(void);
68 void checkasm_check_hevc_sao(void);
69 void checkasm_check_huffyuvdsp(void);
70 void checkasm_check_idctdsp(void);
71 void checkasm_check_jpeg2000dsp(void);
72 void checkasm_check_llviddsp(void);
73 void checkasm_check_llviddspenc(void);
74 void checkasm_check_lpc(void);
75 void checkasm_check_motion(void);
76 void checkasm_check_nlmeans(void);
77 void checkasm_check_opusdsp(void);
78 void checkasm_check_pixblockdsp(void);
79 void checkasm_check_sbrdsp(void);
80 void checkasm_check_synth_filter(void);
81 void checkasm_check_sw_gbrp(void);
82 void checkasm_check_sw_rgb(void);
83 void checkasm_check_sw_scale(void);
84 void checkasm_check_utvideodsp(void);
85 void checkasm_check_v210dec(void);
86 void checkasm_check_v210enc(void);
87 void checkasm_check_vc1dsp(void);
88 void checkasm_check_vf_bwdif(void);
89 void checkasm_check_vf_eq(void);
90 void checkasm_check_vf_gblur(void);
91 void checkasm_check_vf_hflip(void);
92 void checkasm_check_vf_threshold(void);
93 void checkasm_check_vf_sobel(void);
94 void checkasm_check_vp8dsp(void);
95 void checkasm_check_vp9dsp(void);
96 void checkasm_check_videodsp(void);
97 void checkasm_check_vorbisdsp(void);
101 void *checkasm_check_func(void *func, const char *name, ...) av_printf_format(2, 3);
102 int checkasm_bench_func(void);
103 void checkasm_fail_func(const char *msg, ...) av_printf_format(1, 2);
104 struct CheckasmPerf *checkasm_get_perf_context(void);
105 void checkasm_report(const char *name, ...) av_printf_format(1, 2);
107 /* float compare utilities */
108 int float_near_ulp(float a, float b, unsigned max_ulp);
109 int float_near_abs_eps(float a, float b, float eps);
110 int float_near_abs_eps_ulp(float a, float b, float eps, unsigned max_ulp);
111 int float_near_ulp_array(const float *a, const float *b, unsigned max_ulp,
113 int float_near_abs_eps_array(const float *a, const float *b, float eps,
115 int float_near_abs_eps_array_ulp(const float *a, const float *b, float eps,
116 unsigned max_ulp, unsigned len);
117 int double_near_abs_eps(double a, double b, double eps);
118 int double_near_abs_eps_array(const double *a, const double *b, double eps,
121 extern AVLFG checkasm_lfg;
122 #define rnd() av_lfg_get(&checkasm_lfg)
124 static av_unused void *func_ref, *func_new;
126 #define BENCH_RUNS 1000 /* Trade-off between accuracy and speed */
128 /* Decide whether or not the specified function needs to be tested */
129 #define check_func(func, ...) (func_ref = checkasm_check_func((func_new = func), __VA_ARGS__))
131 /* Declare the function prototype. The first argument is the return value, the remaining
132 * arguments are the function parameters. Naming parameters is optional. */
133 #define declare_func(ret, ...) declare_new(ret, __VA_ARGS__) typedef ret func_type(__VA_ARGS__)
134 #define declare_func_float(ret, ...) declare_new_float(ret, __VA_ARGS__) typedef ret func_type(__VA_ARGS__)
135 #define declare_func_emms(cpu_flags, ret, ...) declare_new_emms(cpu_flags, ret, __VA_ARGS__) typedef ret func_type(__VA_ARGS__)
137 /* Indicate that the current test has failed */
138 #define fail() checkasm_fail_func("%s:%d", av_basename(__FILE__), __LINE__)
140 /* Print the test outcome */
141 #define report checkasm_report
143 /* Call the reference function */
144 #define call_ref(...) ((func_type *)func_ref)(__VA_ARGS__)
146 #if ARCH_X86 && HAVE_X86ASM
147 /* Verifies that clobbered callee-saved registers are properly saved and restored
148 * and that either no MMX registers are touched or emms is issued */
149 void checkasm_checked_call(void *func, ...);
150 /* Verifies that clobbered callee-saved registers are properly saved and restored
151 * and issues emms for asm functions which are not required to do so */
152 void checkasm_checked_call_emms(void *func, ...);
153 /* Verifies that clobbered callee-saved registers are properly saved and restored
154 * but doesn't issue emms. Meant for dsp functions returning float or double */
155 void checkasm_checked_call_float(void *func, ...);
158 /* Evil hack: detect incorrect assumptions that 32-bit ints are zero-extended to 64-bit.
159 * This is done by clobbering the stack with junk around the stack pointer and calling the
160 * assembly function through checked_call() with added dummy arguments which forces all
161 * real arguments to be passed on the stack and not in registers. For 32-bit arguments the
162 * upper half of the 64-bit register locations on the stack will now contain junk which will
163 * cause misbehaving functions to either produce incorrect output or segfault. Note that
164 * even though this works extremely well in practice, it's technically not guaranteed
165 * and false negatives is theoretically possible, but there can never be any false positives.
167 void checkasm_stack_clobber(uint64_t clobber, ...);
168 #define declare_new(ret, ...) ret (*checked_call)(void *, int, int, int, int, int, __VA_ARGS__)\
169 = (void *)checkasm_checked_call;
170 #define declare_new_float(ret, ...) ret (*checked_call)(void *, int, int, int, int, int, __VA_ARGS__)\
171 = (void *)checkasm_checked_call_float;
172 #define declare_new_emms(cpu_flags, ret, ...) \
173 ret (*checked_call)(void *, int, int, int, int, int, __VA_ARGS__) = \
174 ((cpu_flags) & av_get_cpu_flags()) ? (void *)checkasm_checked_call_emms : \
175 (void *)checkasm_checked_call;
176 #define CLOB (UINT64_C(0xdeadbeefdeadbeef))
177 #define call_new(...) (checkasm_stack_clobber(CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,\
178 CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB),\
179 checked_call(func_new, 0, 0, 0, 0, 0, __VA_ARGS__))
181 #define declare_new(ret, ...) ret (*checked_call)(void *, __VA_ARGS__) = (void *)checkasm_checked_call;
182 #define declare_new_float(ret, ...) ret (*checked_call)(void *, __VA_ARGS__) = (void *)checkasm_checked_call_float;
183 #define declare_new_emms(cpu_flags, ret, ...) ret (*checked_call)(void *, __VA_ARGS__) = \
184 ((cpu_flags) & av_get_cpu_flags()) ? (void *)checkasm_checked_call_emms : \
185 (void *)checkasm_checked_call;
186 #define call_new(...) checked_call(func_new, __VA_ARGS__)
188 #elif ARCH_ARM && HAVE_ARMV5TE_EXTERNAL
189 /* Use a dummy argument, to offset the real parameters by 2, not only 1.
190 * This makes sure that potential 8-byte-alignment of parameters is kept the same
191 * even when the extra parameters have been removed. */
192 void checkasm_checked_call_vfp(void *func, int dummy, ...);
193 void checkasm_checked_call_novfp(void *func, int dummy, ...);
194 extern void (*checkasm_checked_call)(void *func, int dummy, ...);
195 #define declare_new(ret, ...) ret (*checked_call)(void *, int dummy, __VA_ARGS__, \
196 int, int, int, int, int, int, int, int, \
197 int, int, int, int, int, int, int) = (void *)checkasm_checked_call;
198 #define call_new(...) checked_call(func_new, 0, __VA_ARGS__, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 0, 0, 0)
199 #elif ARCH_AARCH64 && !defined(__APPLE__)
200 void checkasm_stack_clobber(uint64_t clobber, ...);
201 void checkasm_checked_call(void *func, ...);
202 #define declare_new(ret, ...) ret (*checked_call)(void *, int, int, int, int, int, int, int, __VA_ARGS__,\
203 int, int, int, int, int, int, int, int,\
204 int, int, int, int, int, int, int)\
205 = (void *)checkasm_checked_call;
206 #define CLOB (UINT64_C(0xdeadbeefdeadbeef))
207 #define call_new(...) (checkasm_stack_clobber(CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,\
208 CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB,CLOB),\
209 checked_call(func_new, 0, 0, 0, 0, 0, 0, 0, __VA_ARGS__,\
210 7, 6, 5, 4, 3, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0))
212 void checkasm_set_function(void *);
213 void *checkasm_get_wrapper(void);
215 #if (__riscv_xlen == 64) && defined (__riscv_d)
216 #define declare_new(ret, ...) \
217 ret (*checked_call)(__VA_ARGS__) = checkasm_get_wrapper();
218 #define call_new(...) \
219 (checkasm_set_function(func_new), checked_call(__VA_ARGS__))
222 #define declare_new(ret, ...)
223 #define declare_new_float(ret, ...)
224 #define declare_new_emms(cpu_flags, ret, ...)
225 /* Call the function */
226 #define call_new(...) ((func_type *)func_new)(__VA_ARGS__)
229 #ifndef declare_new_emms
230 #define declare_new_emms(cpu_flags, ret, ...) declare_new(ret, __VA_ARGS__)
232 #ifndef declare_new_float
233 #define declare_new_float(ret, ...) declare_new(ret, __VA_ARGS__)
236 typedef struct CheckasmPerf {
242 #if defined(AV_READ_TIME) || CONFIG_LINUX_PERF || CONFIG_MACOS_KPERF
244 #if CONFIG_LINUX_PERF
245 #define PERF_START(t) do { \
246 ioctl(sysfd, PERF_EVENT_IOC_RESET, 0); \
247 ioctl(sysfd, PERF_EVENT_IOC_ENABLE, 0); \
249 #define PERF_STOP(t) do { \
251 ioctl(sysfd, PERF_EVENT_IOC_DISABLE, 0); \
252 ret = read(sysfd, &t, sizeof(t)); \
255 #elif CONFIG_MACOS_KPERF
256 #define PERF_START(t) t = ff_kperf_cycles()
257 #define PERF_STOP(t) t = ff_kperf_cycles() - t
259 #define PERF_START(t) t = AV_READ_TIME()
260 #define PERF_STOP(t) t = AV_READ_TIME() - t
263 /* Benchmark the function */
264 #define bench_new(...)\
266 if (checkasm_bench_func()) {\
267 struct CheckasmPerf *perf = checkasm_get_perf_context();\
268 av_unused const int sysfd = perf->sysfd;\
269 func_type *tfunc = func_new;\
273 for (ti = 0; ti < BENCH_RUNS; ti++) {\
280 if (t*tcount <= tsum*4 && ti > 0) {\
291 #define bench_new(...) while(0)
292 #define PERF_START(t) while(0)
293 #define PERF_STOP(t) while(0)
296 #define DECL_CHECKASM_CHECK_FUNC(type) \
297 int checkasm_check_##type(const char *const file, const int line, \
298 const type *const buf1, const ptrdiff_t stride1, \
299 const type *const buf2, const ptrdiff_t stride2, \
300 const int w, const int h, const char *const name)
302 DECL_CHECKASM_CHECK_FUNC(uint8_t);
303 DECL_CHECKASM_CHECK_FUNC(uint16_t);
304 DECL_CHECKASM_CHECK_FUNC(uint32_t);
305 DECL_CHECKASM_CHECK_FUNC(int16_t);
306 DECL_CHECKASM_CHECK_FUNC(int32_t);
308 #define PASTE(a,b) a ## b
309 #define CONCAT(a,b) PASTE(a,b)
311 #define checkasm_check(prefix, ...) CONCAT(checkasm_check_, prefix)(__FILE__, __LINE__, __VA_ARGS__)
313 #endif /* TESTS_CHECKASM_CHECKASM_H */