如何测量 ARM Cortex-A53 处理器中的程序执行时间？

Question

我正在使用以下方法读取 cortex-a15 中的时钟：

static void readticks(unsigned int *result)
{
    struct timeval t;
    unsigned int cc;
    if (!enabled) {
        // program the performance-counter control-register:
        asm volatile("mcr p15, 0, %0, c9, c12, 0" :: "r"(17));
        //enable all counters
        asm volatile("mcr p15, 0, %0, c9, c12, 1" :: "r"(0x8000000f));
        //Clear overflow.
        asm volatile("mcr p15, 0, %0, c9, c12, 3" :: "r"(0x8000000f));
        enabled = 1;
    }
    asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(cc));
    gettimeofday(&t,(struct timezone *) 0);
    result[0] = cc;
    result[1] = t.tv_usec;
    result[2] = t.tv_sec;
}

最终的性能分析看起来像：

before = readticks();
  foo();
after = readticks();
clock_cycles = after - before. 

我想在 cortex-A53、ARM64（不是 aarch32）中使用相同的逻辑。

我在关注在线门户网站后尝试过这个：

    /* All counters, including PMCCNTR_EL0, are disabled/enabled */

        #define QUADD_ARMV8_PMCR_E      (1 << 0)
        /* Reset all event counters, not including PMCCNTR_EL0, to 0

 */
    #define QUADD_ARMV8_PMCR_P      (1 << 1)
    /* Reset PMCCNTR_EL0 to 0 */
    #define QUADD_ARMV8_PMCR_C      (1 << 2)
    /* Clock divider: PMCCNTR_EL0 counts every clock cycle/every 64 clock cycles */
    #define QUADD_ARMV8_PMCR_D      (1 << 3)
    /* Export of events is disabled/enabled */
    #define QUADD_ARMV8_PMCR_X      (1 << 4)
    /* Disable cycle counter, PMCCNTR_EL0 when event counting is prohibited */
    #define QUADD_ARMV8_PMCR_DP     (1 << 5)
    /* Long cycle count enable */
    #define QUADD_ARMV8_PMCR_LC     (1 << 6)

static inline unsigned int armv8_pmu_pmcr_read(void)
{

        unsigned int val;
        /* Read Performance Monitors Control Register */
        asm volatile("mrs %0, pmcr_el0" : "=r" (val));
        return val;
}
static inline void armv8_pmu_pmcr_write(unsigned int val)
{
    asm volatile("msr pmcr_el0, %0" : :"r" (val & QUADD_ARMV8_PMCR_WR_MASK));
}

static void enable_all_counters(void)
{
    unsigned int val;
    /* Enable all counters */
    val = armv8_pmu_pmcr_read();
    val |= QUADD_ARMV8_PMCR_E | QUADD_ARMV8_PMCR_X;
    armv8_pmu_pmcr_write(val);
}

static void reset_all_counters(void)
{

    unsigned int val;
    val = armv8_pmu_pmcr_read();
    val |= QUADD_ARMV8_PMCR_P | QUADD_ARMV8_PMCR_C;
    armv8_pmu_pmcr_write(val);
}

static void readticks(unsigned int *result)
{
    struct timeval t;
    unsigned int cc;
    unsigned int val;
    if (!enabled) {
        reset_all_counters();
        enable_all_counters();
        enabled = 1;
    }
    cc = armv8_pmu_pmcr_read();
    gettimeofday(&t,(struct timezone *) 0);
    result[0] = cc;
    result[1] = t.tv_usec;
    result[2] = t.tv_sec;
}

但是当我尝试分析时，它给出了“非法指令”作为错误。谁能帮我改一下上面的cortex-a53代码？

Answer 1

您需要为用户模式启用 PMU。这是我为它编写的内核模块（适用于 Raspberry Pi 2 中的 ARM V7）：

/* Module source file 'module.c'. */
#include <linux/module.h>
#include <linux/init.h>
#include <linux/kernel.h>

arm_write(unsigned long val)
{
        //Enabling both read and write - note difference between mcr and mrc
        asm volatile("mrc p15, 0, %0, c9, c14, 0" :: "r"(1));
        asm volatile("mcr p15, 0, %0, c9, c14, 0" :: "r"(1));

}

static int enabler(void)
{
     unsigned long value = 1;
     printk(KERN_INFO "Enabling PMU usermode.\n");
     arm_write(value);
     return 0;
}

static void end(void)
{
     printk(KERN_INFO "module unloaded.\n");
}

module_init(enabler);
module_exit(end);

MODULE_AUTHOR("Sama");
MODULE_LICENSE("GPL");
MODULE_DESCRIPTION("Blahblah");

这将启用对 PMU 的用户模式访问。编译完成后，您需要启用 PMU 计数器，如下所示：

int main(int argc, char **argv){
    int enable_divider =1;
    int do_reset=1;
        int value = 1;

        // peform reset:
        if (do_reset) {
                value |= 2;     // reset all counters to zero.
                value |= 4;     // reset cycle counter to zero.
        }

        if (enable_divider)
                value |= 8;     // enable "by 64" divider for CCNT. You really do not want to get all cycle count. This will increment the counter by 1 for every 64 cpu cycle.

        value |= 16;
        // program the performance-counter control-register with mask constructed above
        asm volatile ("MCR p15, 0, %0, c9, c12, 0\t\n" :: "r"(value));
        // enable all counters:
        asm volatile ("MCR p15, 0, %0, c9, c12, 1\t\n" :: "r"(0x8000000f));

        // clear overflows:
        asm volatile ("MCR p15, 0, %0, c9, c12, 3\t\n" :: "r"(0x80000001));

        // Select individual counter (0)
        asm volatile ("MCR p15,   0,    %0,  c9  ,   c12 ,   5\t\n":: "r"(0x00));

        // Write event (0x11 = Cycle count)
        asm volatile ("MCR p15,   0,    %0,  c9  ,   c13 ,   1\t\n":: "r"(0xD));

    printf("Hi");
        unsigned int output;

        // Read current event counter
        asm volatile ("MRC p15,   0,    %0,  c9  ,   c13 ,   2\t\n": "=r"(output));
        printf("Event count 0: %ul\n", output);
    printf("Normal Execution, No Buffer Overflow Occurred.\n");
   return 0;
}

但是不幸的是，您得到的不仅仅是您的程序 cpu 周期，而是整个系统 cpu 周期！所以我推荐的是使用 perf。

用 C 语言将你的 asm 代码写成内联汇编代码，然后这样写：

int dummya(int z, int b){
//This is my function you need to change it for yourself
struct perf_event_attr pe;
long long count;
int fd;
        memset(&pe, 0, sizeof(struct perf_event_attr));
        pe.type = PERF_TYPE_HARDWARE;
        pe.size = sizeof(struct perf_event_attr);
        pe.config = PERF_COUNT_HW_CPU_CYCLES;
        pe.disabled = 1;
        pe.exclude_kernel = 1;
        pe.exclude_hv = 1;
        fd = perf_event_open(&pe, 0, -1, -1, 0);
        if (fd == -1) {
                fprintf(stderr, "Error opening leader %llx\n", pe.config);
                exit(EXIT_FAILURE);
               }

        ioctl(fd, PERF_EVENT_IOC_RESET, 0);
        ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);
//From here the counter starts.
        asm("Your ASM Codes");
        asm("Your ASM Codes");
        asm("Your ASM Codes");
        asm("Your ASM Codes");
        asm("Your ASM Codes");
        asm("Your ASM Codes");
        asm("Your ASM Codes");
        asm("Your ASM Codes");
//Disabling Counter
        ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);

        read(fd, &count, sizeof(long long));
        printf("%lld\n", count);
        close(fd);
return 5;
}

请注意，您需要新的内核才能访问 Perf 驱动程序。