armv8 timer & cycle counter

armv8_tsc.h

#pragma once

#if defined(__linux__)
#	define _GNU_SOURCE
#	include <sched.h>
#	include <unistd.h>
#	include <sys/syscall.h>
#	include <linux/perf_event.h>
#elif defined(_WIN32)
#	include <intrin.h>
#	include <windows.h>
#else
#	error not supported
#endif

#include <stdint.h>
#include <stdbool.h>

//
// fixed frequency counter, always available
//

static inline uint64_t armv8_cntvct(void)
{
	uint64_t value;
#if defined(__linux__)
	__asm__ __volatile__("mrs %0, cntvct_el0" : "=r"(value));
#elif defined(_WIN32)
	// "Accessing CNTVCT_EL0" in https://developer.arm.com/documentation/ddi0601/latest/AArch64-Registers/CNTVCT-EL0--Counter-timer-Virtual-Count-Register
	value = _ReadStatusReg(ARM64_SYSREG(3, 3, 14, 0, 2));
#endif
	return value;
}

static inline uint64_t armv8_cntfrq(void)
{
	uint64_t value;
#if defined(__linux__)
	__asm__ __volatile__("mrs %0, cntfrq_el0" : "=r"(value));
#elif defined(_WIN32)
	// "Accessing CNTFRQ_EL0" in https://developer.arm.com/documentation/ddi0601/latest/AArch64-Registers/CNTFRQ-EL0--Counter-timer-Frequency-Register
	value = _ReadStatusReg(ARM64_SYSREG(3, 3, 14, 0, 0));
#endif
	return value;
}

//
// performance monitor cycle counter, pinned to one core
// on Linux requires extra setup - read comments below
//

#if defined(__linux__)
static int armv8_perf_fd;
static cpu_set_t armv8_perf_mask;
#elif defined(_WIN32)
static DWORD_PTR armv8_thread_mask;
#endif

static inline bool armv8_tsc_init(void)
{
#if defined(__linux__)
	int core = sched_getcpu();
	sched_getaffinity(0, sizeof(cpu_set_t), &armv8_perf_mask);

	cpu_set_t set;
	CPU_ZERO(&set);
	CPU_SET(core, &set);
	sched_setaffinity(0, sizeof(cpu_set_t), &set);

	struct perf_event_attr attr =
	{
		.size = sizeof(attr),
		.type = PERF_TYPE_HARDWARE,
		.config = PERF_COUNT_HW_CPU_CYCLES,
		.config1 = 1 | 2, // 1=64-bit counters, 2=allow user access
		.pinned = 1,
	};

	int fd = syscall(__NR_perf_event_open, &attr, 0, core, -1, 0);
	if (fd < 0)
	{
		// perf not enabled in kernel, or perf requires root
		// to allow non-root access, run the following:
		// echo 1 | sudo tee /proc/sys/kernel/perf_event_paranoid
		return false;
	}

	uint64_t value;
	__asm__ __volatile__("mrs %0, pmuserenr_el0" : "=r"(value)); 
	if (!(value & 4))
	{
		// PMU not allowed for user-space access, to allow run this:
		// echo 1 | sudo tee /proc/sys/kernel/perf_user_access
		close(fd);
		return false;
	}

	armv8_perf_fd = fd;
	return true;
#elif defined(_WIN32)
	armv8_thread_mask = SetThreadAffinityMask(GetCurrentThread(), 1ULL << GetCurrentProcessorNumber());
	// no setup needed to access cycle counter from user-space
	// https://learn.microsoft.com/en-us/cpp/build/arm64-windows-abi-conventions?view=msvc-170#cycle-counter
	return true;
#else
	return false;
#endif
}

static inline void armv8_tsc_done(void)
{
#if defined(__linux__)
	close(armv8_perf_fd);
	sched_setaffinity(0, sizeof(cpu_set_t), &armv8_perf_mask);
#elif defined(_WIN32)
	SetThreadAffinityMask(GetCurrentThread(), armv8_thread_mask);
#endif
}

// use only when armv8_tsc_init() returned true, otherwise SIGILL will be raised
static inline uint64_t armv8_pmccntr(void)
{
	uint64_t value;
#if defined(__linux__)
	__asm__ __volatile__("mrs %0, pmccntr_el0" : "=r"(value));
#elif defined(_WIN32)
	// "Accessing PMCCNTR_EL0" in https://developer.arm.com/documentation/ddi0601/latest/AArch64-Registers/PMCCNTR-EL0--Performance-Monitors-Cycle-Count-Register
	value = _ReadStatusReg(ARM64_SYSREG(3, 3, 9, 13, 0));
#endif
	return value;
}

armv8_tsc_example.c

#include "armv8_tsc.h"

#include <stdio.h>

static void loop()
{
	for (int i=0; i<1000000; i++)
	{
#if defined(_MSC_VER)
		__nop();
#else
		__asm__ __volatile__("");
#endif
	}
}

int main()
{
	uint64_t ticks0 = armv8_cntvct();
	loop();
	uint64_t ticks1 = armv8_cntvct();
	uint64_t freq = armv8_cntfrq();

	printf("cntvct : %zu ticks @ %zu MHz = %.2f msec\n",
		(size_t)(ticks1 - ticks0),
		(size_t)(freq / 1000000),
		(ticks1 - ticks0) * 1000.0 / freq);

	if (armv8_tsc_init())
	{
		uint64_t cycles0 = armv8_pmccntr();
		loop();
		uint64_t cycles1 = armv8_pmccntr();

		printf("pmccntr: %zu cycles\n", (size_t)(cycles1 - cycles0));
		armv8_tsc_done();
	}
	else
	{
		printf("perf not available! not enough privileges?\n");
	}
}