CPU性能监控

armv7 performance monitor接口 更多...

宏定义
#define	ARMV7_PMU_EVENT_PMNC_SW_INCR   0x00
	Software increment.
#define	ARMV7_PMU_EVENT_L1_ICACHE_REFILL   0x01
	Instruction fetch that causes a refill.
#define	ARMV7_PMU_EVENT_ITLB_REFILL   0x02
	Instruction fetch that causes a TLB refill.
#define	ARMV7_PMU_EVENT_L1_DCACHE_REFILL   0x03
	Data access causes a refill.
#define	ARMV7_PMU_EVENT_L1_DCACHE_ACCESS   0x04
	Data cache hit.
#define	ARMV7_PMU_EVENT_DTLB_REFILL   0x05
	Data access causes a TLB refill.
#define	ARMV7_PMU_EVENT_MEM_READ   0x06
	Data read architecturally executed.
#define	ARMV7_PMU_EVENT_MEM_WRITE   0x07
	Data write architecturally executed.
#define	ARMV7_PMU_EVENT_INSTR_EXECUTED   0x08
	Instruction architecturally executed.
#define	ARMV7_PMU_EVENT_EXC_TAKEN   0x09
	Exception taken.
#define	ARMV7_PMU_EVENT_EXC_EXECUTED   0x0A
	Exception return architecturally executed.
#define	ARMV7_PMU_EVENT_CID_WRITE   0x0B
	Change to ContexID retired.
#define	ARMV7_PMU_EVENT_PC_WRITE   0x0C
	Software change of PC.
#define	ARMV7_PMU_EVENT_PC_IMM_BRANCH   0x0D
	Immediate branch architecturally executed.
#define	ARMV7_PMU_EVENT_PC_RETURN   0x0E
	Procedure return architecturally executed.
#define	ARMV7_PMU_EVENT_MEM_UNALIGNED_ACCESS   0x0F
	Unaligned load-store.
#define	ARMV7_PMU_EVENT_PC_BRANCH_MIS_PRED   0x10
	Branch mispredicted/not predicted.
#define	ARMV7_PMU_EVENT_CLOCK_CYCLES   0x11
	Cycle counter.
#define	ARMV7_PMU_EVENT_PC_BRANCH_PRED   0x12
	Branch predicted.
#define	ARMV7_PMU_EVENT_MEM_ACCESS   0x13
	Data memory access.
#define	ARMV7_PMU_EVENT_L1_ICACHE_ACCESS   0x14
	Instruction Cache access.
#define	ARMV7_PMU_EVENT_L1_DCACHE_WB   0x15
	Data cache eviction.
#define	ARMV7_PMU_EVENT_L2_CACHE_ACCESS   0x16
	L2 data cache access.
#define	ARMV7_PMU_EVENT_L2_CACHE_REFILL   0x17
	L2 data cache refill.
#define	ARMV7_PMU_EVENT_L2_CACHE_WB   0x18
	L2 data cache write-back.
#define	ARMV7_PMU_EVENT_BUS_ACCESS   0x19
	Bus access.
#define	ARMV7_PMU_EVENT_BUS_CYCLES   0x1D
	Bus cycle.
#define	ARMV7_A7_PMU_EVENT_BUS_ACCESS_READ   0x60
	Bus access, read.
#define	ARMV7_A7_PMU_EVENT_BUS_ACCESS_WRITE   0x61
	Bus access, write.
#define	ARMV7_A7_PMU_EVENT_IRQ_EXCEPTION   0x86
	IRQ exception taken.
#define	ARMV7_A7_PMU_EVENT_FIQ_EXCEPTION   0x87
	FIQ exception taken.
#define	ARMV7_A7_PMU_EVENT_EXT_MEM   0xC0
	External memory request.
#define	ARMV7_A7_PMU_EVENT_NO_CACHE_EXT_MEM   0xC1
	Non-cacheable external memory request.
#define	ARMV7_A7_PMU_EVENT_PREF_LINFILL   0xC2
	Linefill because of prefetch.
#define	ARMV7_A7_PMU_EVENT_PREF_LINFILL_DROP   0xC3
	Prefetch linefill dropped.
#define	ARMV7_A7_PMU_EVENT_ENTER_READ_ALLOC_MODE   0xC4
	Entering read allocate mode.
#define	ARMV7_A7_PMU_EVENT_READ_ALLOC_MODE   0xC5
	Read allocate mode.
#define	ARMV7_A7_PMU_EVENT_ETM_EXT_OUT0   0xC7
	ETM Ext Out[0].
#define	ARMV7_A7_PMU_EVENT_ETM_EXT_OUT1   0xC8
	ETM Ext Out[1].
#define	ARMV7_A7_PMU_EVENT_PIPELINE_STALL   0xC9
	Pipeline stall.
#define	ARMV7_A7_PMU_EVENT_DATA_SNOOP   0xCA
	Data snooped from other processor.

枚举

函数
void	rt_hw_pmu_enable_event (rt_int32_t idx, rt_uint32_t event, rt_bool_t int_en)
	配置计数器idx监控事件event,使能计数器，并根据需要使能中断 更多...
void	rt_hw_pmu_disable_event (rt_int32_t idx, rt_bool_t int_dis)
	关闭计数器监控，并根据需要关闭中断 更多...
void	rt_hw_pmu_start (void)
	打开所有计数器
void	rt_hw_pmu_stop (void)
	关闭所有计数器
rt_uint32_t	rt_hw_pmu_get_counter_val (rt_int32_t idx)
	获取计数器的值 更多...
void	rt_hw_pmu_set_counter_val (rt_int32_t idx, rt_uint32_t val)
	配置计数器的值 更多...
rt_uint32_t	rt_hw_pmu_get_overflow_flag (void)
	获取计数器overflow标志 更多...
rt_uint32_t	rt_hw_pmu_cnt_overflowed (rt_int32_t idx)
	获取指定计数器overflow标志 更多...
void	rt_hw_pmu_write_swinc (rt_int32_t idx)
	写PMSWINC寄存器 更多...
void	rt_hw_pmu_reset (void)
	重置所有计数器，包括关闭计数器，关闭中断，reset计数器值为0

详细描述

armv7 performance monitor接口

PMU即性能监视器，它运行时可以收集关于处理器和内存的各种统计信息。这些统计信息中的事件非常有用， 可以利用它们来调试或者剖析代码，对软件进行性能的调优。

PMU寄存器可以分成两类：一部分寄存器是用来做计数（主要是cycle counter），比如测量一段代码的运行时 间，从寄存器里面读取cpu cycles数，并根据cpu的主频计算代码运行时间；另一部分寄存器是用来统计CPU和 内存等的信息，即event事件，如cache的miss和hit，可以帮助调整代码优化性能。

需要注意的是，对于多核CPU，每个core都有一套独立的PMU寄存器。

PMU驱动提供couter监控事件设置，中断控制，计数器溢出查询，设置计数器值，获取计数器值，计数器reset 功能等。

以监控Icache miss事件为例，如下是sample code：

rt_hw_pmu_set_counter_val(IDX_EVENT_COUNTER1, 0);
rt_hw_pmu_enable_event(IDX_EVENT_COUNTER1, ARMV7_PMU_EVENT_L1_ICACHE_REFILL, 0);
rt_hw_pmu_start();

1. 设置event counter1计数开始值为0，计数器从0开始计数;
2. 设置event counter1监控事件ARMV7_PMU_EVENT_L1_ICACHE_REFILL，使能event counter1;
3. 打开pmu counter总开关。

value = rt_hw_pmu_get_counter_val(IDX_EVENT_COUNTER1);

获取Icache miss的次数。

枚举类型说明

enum rt_hw_counter_idx

枚举值
IDX_CYCLE_COUNTER	CPU周期计数器
IDX_EVENT_COUNTER0	Event0计数器
IDX_EVENT_COUNTER1	Event1计数器
IDX_EVENT_COUNTER2	Event2计数器
IDX_EVENT_COUNTER3	Event3计数器
IDX_EVENT_COUNTER4	Event4计数器
IDX_EVENT_COUNTER5	...

函数说明

void rt_hw_pmu_enable_event	(	rt_int32_t	idx,
		rt_uint32_t	event,
		rt_bool_t	int_en
	)

配置计数器idx监控事件event,使能计数器，并根据需要使能中断

参数

idx	计数器id
event	监控事件编码
int_en	计数器中断使能标志, 1:使能中断; 0:忽略

void rt_hw_pmu_disable_event	(	rt_int32_t	idx,
		rt_bool_t	int_dis
	)

关闭计数器监控，并根据需要关闭中断

参数

idx	计数器id
int_en	计数器中断使能标志, 1:关闭中断; 0:忽略

rt_uint32_t rt_hw_pmu_get_counter_val

(

rt_int32_t

idx

)

获取计数器的值

参数

idx

计数器id

返回

计数器的值

void rt_hw_pmu_set_counter_val	(	rt_int32_t	idx,
		rt_uint32_t	val
	)

配置计数器的值

参数

idx	计数器id
val	计数值

注解

当计数器的值配置为val，并使能计数器后，计数器将从该值开始递增计数

rt_uint32_t rt_hw_pmu_get_overflow_flag

(

void

)

获取计数器overflow标志

返回

返回overflow标志，一个bit对应一个counter，1:counter溢出; 0:counter未溢出

rt_uint32_t rt_hw_pmu_cnt_overflowed

(

rt_int32_t

idx

)

获取指定计数器overflow标志

参数

idx

计数器id

返回

返回指定计数器overflow标志，0:未溢出; 非0:发生溢出

void rt_hw_pmu_write_swinc

(

rt_int32_t

idx

)

写PMSWINC寄存器

参数

idx

计数器id

注解

当配置计数器idx监控SWINC事件，调用本接口写PMSWINC寄存器，计数器idx值就会加1