checkasm: use perf API on Linux ARM*

On ARM platforms, accessing the PMU registers requires special user
access permissions. Since there is no other way to get accurate timers,
the current implementation of timers in FFmpeg rely on these registers.
Unfortunately, enabling user access to these registers on Linux is not
trivial, and generally involve compiling a random and unreliable github
kernel module, or patching somehow your kernel.

Such module is very unlikely to reach the upstream anytime soon. Quoting
Robin Murphin from ARM:

> Say you do give userspace direct access to the PMU; now run two or more
> programs at once that believe they can use the counters for their own
> "minimal-overhead" profiling. Have fun interpreting those results...
>
> And that's not even getting into the implications of scheduling across
> different CPUs, CPUidle, etc. where the PMU state is completely beyond
> userspace's control. In general, the plan to provide userspace with
> something which might happen to just about work in a few corner cases,
> but is meaningless, misleading or downright broken in all others, is to
> never do so.

As a result, the alternative is to use the Performance Monitoring Linux
API which makes use of these registers internally (assuming the PMU of
your ARM board is supported in the kernel, which is definitely not a
given...).

While the Linux API is obviously cross platform, it does have a
significant overhead which needs to be taken into account. As a result,
that mode is only weakly enabled on ARM platforms exclusively.

Note on the non flexibility of the implementation: the timers (native
FFmpeg vs Linux API) are selected at compilation time to prevent the
need of function calls, which would result in a negative impact on the
cycle counters.

configure

@@ -448,6 +448,7 @@ Developer options (useful when working on FFmpeg itself):
   --libfuzzer=PATH         path to libfuzzer
   --ignore-tests=TESTS     comma-separated list (without "fate-" prefix
                            in the name) of tests whose result is ignored
+  --enable-linux-perf      enable Linux Performance Monitor API
 
 NOTE: Object files are built at the place where configure is launched.
 EOF
@@ -1699,6 +1700,7 @@ CONFIG_LIST="
     $SUBSYSTEM_LIST
     autodetect
     fontconfig
+    linux_perf
     memory_poisoning
     neon_clobber_test
     ossfuzz
@@ -5015,6 +5017,7 @@ case $target_os in
     linux)
         enable dv1394
         enable section_data_rel_ro
+        enabled_any arm aarch64 && enable_weak linux_perf
         ;;
     irix*)
         target_os=irix

tests/checkasm/checkasm.c

@@ -20,6 +20,14 @@
  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  */
 
+#include "config.h"
+
+#if CONFIG_LINUX_PERF
+# ifndef _GNU_SOURCE
+#  define _GNU_SOURCE // for syscall (performance monitoring API)
+# endif
+#endif
+
 #include <stdarg.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -190,8 +198,7 @@ typedef struct CheckasmFuncVersion {
     void *func;
     int ok;
     int cpu;
-    int iterations;
-    uint64_t cycles;
+    CheckasmPerf perf;
 } CheckasmFuncVersion;
 
 /* Binary search tree node */
@@ -212,7 +219,11 @@ static struct {
     int bench_pattern_len;
     int num_checked;
     int num_failed;
+
+    /* perf */
     int nop_time;
+    int sysfd;
+
     int cpu_flag;
     const char *cpu_flag_name;
     const char *test_name;
@@ -396,7 +407,6 @@ static const char *cpu_suffix(int cpu)
     return "c";
 }
 
-#ifdef AV_READ_TIME
 static int cmp_nop(const void *a, const void *b)
 {
     return *(const uint16_t*)a - *(const uint16_t*)b;
@@ -407,10 +417,13 @@ static int measure_nop_time(void)
 {
     uint16_t nops[10000];
     int i, nop_sum = 0;
+    av_unused const int sysfd = state.sysfd;
 
+    uint64_t t = 0;
     for (i = 0; i < 10000; i++) {
-        uint64_t t = AV_READ_TIME();
-        nops[i] = AV_READ_TIME() - t;
+        PERF_START(t);
+        PERF_STOP(t);
+        nops[i] = t;
     }
 
     qsort(nops, 10000, sizeof(uint16_t), cmp_nop);
@@ -430,8 +443,9 @@ static void print_benchs(CheckasmFunc *f)
         if (f->versions.cpu || f->versions.next) {
             CheckasmFuncVersion *v = &f->versions;
             do {
-                if (v->iterations) {
-                    int decicycles = (10*v->cycles/v->iterations - state.nop_time) / 4;
+                CheckasmPerf *p = &v->perf;
+                if (p->iterations) {
+                    int decicycles = (10*p->cycles/p->iterations - state.nop_time) / 4;
                     printf("%s_%s: %d.%d\n", f->name, cpu_suffix(v->cpu), decicycles/10, decicycles%10);
                 }
             } while ((v = v->next));
@@ -440,7 +454,6 @@ static void print_benchs(CheckasmFunc *f)
         print_benchs(f->child[1]);
     }
 }
-#endif
 
 /* ASCIIbetical sort except preserving natural order for numbers */
 static int cmp_func_names(const char *a, const char *b)
@@ -543,6 +556,63 @@ static void print_cpu_name(void)
     }
 }
 
+#if CONFIG_LINUX_PERF
+static int bench_init_linux(void)
+{
+    struct perf_event_attr attr = {
+        .type           = PERF_TYPE_HARDWARE,
+        .size           = sizeof(struct perf_event_attr),
+        .config         = PERF_COUNT_HW_CPU_CYCLES,
+        .disabled       = 1, // start counting only on demand
+        .exclude_kernel = 1,
+        .exclude_hv     = 1,
+    };
+
+    printf("benchmarking with Linux Perf Monitoring API\n");
+
+    state.sysfd = syscall(__NR_perf_event_open, &attr, 0, -1, -1, 0);
+    if (state.sysfd == -1) {
+        perror("syscall");
+        return -1;
+    }
+    return 0;
+}
+#endif
+
+static int bench_init_ffmpeg(void)
+{
+#ifdef AV_READ_TIME
+    printf("benchmarking with native FFmpeg timers\n");
+    return 0;
+#else
+    fprintf(stderr, "checkasm: --bench is not supported on your system\n");
+    return -1;
+#endif
+}
+
+static int bench_init(void)
+{
+#if CONFIG_LINUX_PERF
+    int ret = bench_init_linux();
+#else
+    int ret = bench_init_ffmpeg();
+#endif
+    if (ret < 0)
+        return ret;
+
+    state.nop_time = measure_nop_time();
+    printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
+    return 0;
+}
+
+static void bench_uninit(void)
+{
+#if CONFIG_LINUX_PERF
+    if (state.sysfd > 0)
+        close(state.sysfd);
+#endif
+}
+
 int main(int argc, char *argv[])
 {
     unsigned int seed = av_get_random_seed();
@@ -560,10 +630,8 @@ int main(int argc, char *argv[])
 
     while (argc > 1) {
         if (!strncmp(argv[1], "--bench", 7)) {
-#ifndef AV_READ_TIME
-            fprintf(stderr, "checkasm: --bench is not supported on your system\n");
-            return 1;
-#endif
+            if (bench_init() < 0)
+                return 1;
             if (argv[1][7] == '=') {
                 state.bench_pattern = argv[1] + 8;
                 state.bench_pattern_len = strlen(state.bench_pattern);
@@ -591,16 +659,13 @@ int main(int argc, char *argv[])
         ret = 1;
     } else {
         fprintf(stderr, "checkasm: all %d tests passed\n", state.num_checked);
-#ifdef AV_READ_TIME
         if (state.bench_pattern) {
-            state.nop_time = measure_nop_time();
-            printf("nop: %d.%d\n", state.nop_time/10, state.nop_time%10);
             print_benchs(state.funcs);
         }
-#endif
     }
 
     destroy_func_tree(state.funcs);
+    bench_uninit();
     return ret;
 }
 
@@ -678,11 +743,13 @@ void checkasm_fail_func(const char *msg, ...)
     }
 }
 
-/* Update benchmark results of the current function */
-void checkasm_update_bench(int iterations, uint64_t cycles)
+/* Get the benchmark context of the current function */
+CheckasmPerf *checkasm_get_perf_context(void)
 {
-    state.current_func_ver->iterations += iterations;
-    state.current_func_ver->cycles += cycles;
+    CheckasmPerf *perf = &state.current_func_ver->perf;
+    memset(perf, 0, sizeof(*perf));
+    perf->sysfd = state.sysfd;
+    return perf;
 }
 
 /* Print the outcome of all tests performed since the last time this function was called */

tests/checkasm/checkasm.h

@@ -25,6 +25,14 @@
 
 #include <stdint.h>
 #include "config.h"
+
+#if CONFIG_LINUX_PERF
+#include <unistd.h> // read(3)
+#include <sys/ioctl.h>
+#include <asm/unistd.h>
+#include <linux/perf_event.h>
+#endif
+
 #include "libavutil/avstring.h"
 #include "libavutil/cpu.h"
 #include "libavutil/internal.h"
@@ -58,10 +66,12 @@ void checkasm_check_vp8dsp(void);
 void checkasm_check_vp9dsp(void);
 void checkasm_check_videodsp(void);
 
+struct CheckasmPerf;
+
 void *checkasm_check_func(void *func, const char *name, ...) av_printf_format(2, 3);
 int checkasm_bench_func(void);
 void checkasm_fail_func(const char *msg, ...) av_printf_format(1, 2);
-void checkasm_update_bench(int iterations, uint64_t cycles);
+struct CheckasmPerf *checkasm_get_perf_context(void);
 void checkasm_report(const char *name, ...) av_printf_format(1, 2);
 
 /* float compare utilities */
@@ -178,32 +188,59 @@ void checkasm_checked_call(void *func, ...);
 #define declare_new_float(ret, ...) declare_new(ret, __VA_ARGS__)
 #endif
 
+typedef struct CheckasmPerf {
+    int sysfd;
+    uint64_t cycles;
+    int iterations;
+} CheckasmPerf;
+
+#if defined(AV_READ_TIME) || CONFIG_LINUX_PERF
+
+#if CONFIG_LINUX_PERF
+#define PERF_START(t) do {                              \
+    ioctl(sysfd, PERF_EVENT_IOC_RESET, 0);              \
+    ioctl(sysfd, PERF_EVENT_IOC_ENABLE, 0);             \
+} while (0)
+#define PERF_STOP(t) do {                               \
+    ioctl(sysfd, PERF_EVENT_IOC_DISABLE, 0);            \
+    read(sysfd, &t, sizeof(t));                         \
+} while (0)
+#else
+#define PERF_START(t) t = AV_READ_TIME()
+#define PERF_STOP(t)  t = AV_READ_TIME() - t
+#endif
+
 /* Benchmark the function */
-#ifdef AV_READ_TIME
 #define bench_new(...)\
     do {\
         if (checkasm_bench_func()) {\
+            struct CheckasmPerf *perf = checkasm_get_perf_context();\
+            av_unused const int sysfd = perf->sysfd;\
             func_type *tfunc = func_new;\
             uint64_t tsum = 0;\
             int ti, tcount = 0;\
+            uint64_t t = 0; \
             for (ti = 0; ti < BENCH_RUNS; ti++) {\
-                uint64_t t = AV_READ_TIME();\
+                PERF_START(t);\
                 tfunc(__VA_ARGS__);\
                 tfunc(__VA_ARGS__);\
                 tfunc(__VA_ARGS__);\
                 tfunc(__VA_ARGS__);\
-                t = AV_READ_TIME() - t;\
+                PERF_STOP(t);\
                 if (t*tcount <= tsum*4 && ti > 0) {\
                     tsum += t;\
                     tcount++;\
                 }\
             }\
             emms_c();\
-            checkasm_update_bench(tcount, tsum);\
+            perf->cycles += t;\
+            perf->iterations++;\
         }\
     } while (0)
 #else
 #define bench_new(...) while(0)
+#define PERF_START(t)  while(0)
+#define PERF_STOP(t)   while(0)
 #endif
 
 #endif /* TESTS_CHECKASM_CHECKASM_H */