如何测量python代码每行占用的CPU和GPU时间？

import time
import tensorflow as tf

# 测量CPU时间
def measure_cpu_time():
    start_time = time.process_time()
    
    # 代码行1
    # ...
    
    cpu_time_1 = time.process_time() - start_time
    
    # 代码行2
    # ...
    
    cpu_time_2 = time.process_time() - cpu_time_1
    
    # 代码行3
    # ...
    
    cpu_time_3 = time.process_time() - cpu_time_2
    
    # ...

# 测量GPU时间
@tf.function
def measure_gpu_time():
    start_time = tf.timestamp()
    
    # 代码行1
    # ...
    
    gpu_time_1 = tf.timestamp() - start_time
    
    # 代码行2
    # ...
    
    gpu_time_2 = tf.timestamp() - gpu_time_1
    
    # 代码行3
    # ...
    
    gpu_time_3 = tf.timestamp() - gpu_time_2
    
    # ...

# 调用函数进行测量
measure_cpu_time()
measure_gpu_time()