How to profile the cycles and timing of a inference in tflite-micro

PlatformIO IDE
1

I am trying to get information on the number of cycles and the time taken to run a inference in a teensy board with M7 cortex running at 600 Mhz.

The result from serial monitor :
Total cycles used: 410
Prediction Time: 411 us
Output Tensor:
0.08 0.25

Looks weird , shouldnt it be 410/600000000 =0.6667 microseconds ?

imgg
imgg583Ă—1495 158 KB

2

Both of these user micros(). Your “Prediction Time” directly, the “Total Cycles used” as returned by the difference between two calls to tflite::GetCurrentTimeTicks() indirectly.

So your code is not measuring clock cycles at all.

You can measure clock cycles of the Cortex-M7 with the ARM DWT (debug, watchpoints, traces) facility. An example for that is seen here.

#define CPU_RESET_CYCLECOUNTER    do { ARM_DEMCR |= ARM_DEMCR_TRCENA;          \
                                       ARM_DWT_CTRL |= ARM_DWT_CTRL_CYCCNTENA; \
                                       ARM_DWT_CYCCNT = 0; } while(0)
volatile int cycles;

void some_function_that_does_work() {
  CPU_RESET_CYCLECOUNTER; // reset to 0
  do_heavy_work();
  cycles = ARM_DWT_CYCCNT; // capture cycle count
  // print number of cycles etc
}
1 Like
3

Ah I see ! Is there a universal cycle counter that can be used for all chips ?

This code works for any arduino platform chips but now the ARM specific code might not work with esp32 chips right?

4

The CPU cycle count is not abstracted away by the Arduino API or the TFLite API. They only care about microsecond or millisecond delays / timing. So you will have to create some abstraction layer or case work yourself.

#ifdef __arm__ 
/* some ARM CPU, assume we can use the ARM DWT */
#define CPU_RESET_CYCLECOUNTER    do { ARM_DEMCR |= ARM_DEMCR_TRCENA;          \
                                       ARM_DWT_CTRL |= ARM_DWT_CTRL_CYCCNTENA; \
                                       ARM_DWT_CYCCNT = 0; } while(0)
#define CPU_GET_CYCLECOUNTER()  ARM_DWT_CYCCNT
#elif defined(ESP32) // or ARDUINO_ARCH_ESP32 or __xtensa__ I guess
/* some ESP32 stuff, use their SDK */
#define CPU_RESET_CYCLECOUNTER    do { esp_cpu_set_cycle_count(0); } while(0)
#define CPU_GET_CYCLECOUNTER()  ((uint32_t) esp_cpu_get_cycle_count())
#else
#error "tf is this architecture"
#endif

// use CPU_RESET_CYCLECOUNTER and CPU_GET_CYCLECOUNTER in code

Similiarly, the ESP32 SDK has esp_cpu_get_cycle_count, providing the same functionality.

1 Like