[P0] Questions about FCFS

[tancheng]

Hi @MeowMJ, let's discuss these tmr, thanks~

1. CGRA-Mapper/tools/expandable/NeuraDemo.py

   Line 364 in 6c19f2d

   execution_time = self.kernel.total_iterations * self.ii

Shouldn't we manipulate number of iterations (instead of II) to calculate execution time?

2. CGRA-Mapper/tools/expandable/NeuraDemo.py

   Line 338 in 6c19f2d

   def calculate_execution_time(self):

We should also have migratable/expanded II?

3. CGRA-Mapper/tools/expandable/NeuraDemo.py

   Line 368 in 6c19f2d

   def copy_with_valid(self):

Why we need this func?

4. CGRA-Mapper/tools/expandable/NeuraDemo.py

   Lines 413 to 414 in 6c19f2d

   	execution_time = instance.calculate_execution_time()
   	instance.end_time = current_time + execution_time

If it is current time + instance.calculate_execution_time(), then instance.calculate_execution_time() should be updated for every func call?

5. CGRA-Mapper/tools/expandable/NeuraDemo.py

   Line 485 in 6c19f2d

   instance.end_time = new_end_time

   
   instance.end_time is updated by both release() and reallocate(), how to distinguish, they are by default exclude to each other?

6. I don't see a CGRA class, how do we calculate per-CGRA utilization?

7. https://github.com/tancheng/CGRA-Mapper/blob/6c19f2d233714a3135a06295cf5453dcda9a5025/tools/expandable/NeuraDemo.py#L543What is none here?

8. CGRA-Mapper/tools/expandable/NeuraDemo.py

   Lines 588 to 591 in 6c19f2d

   	# Check running instances for possible re-allocation
   	if priority_bosting:
   	for running in running_instances:
   	available_cgras, total_cgra_runtime = re_allocate(running, current_time, available_cgras, events, total_cgra_runtime)

How could this affect the end time to deque an event from the heapq? This would cause one kernel has two end events?

9. CGRA-Mapper/tools/expandable/NeuraDemo.py

   Lines 593 to 601 in 6c19f2d

   	# Calculate ratio for each kernel
   	for kernel in kernels:
   	total_execution_time = sum(
   	[inst.pure_execution_time for inst in completed_instances if inst.kernel.kernel_name == kernel.kernel_name])
   	total_waiting_time = sum(
   	[inst.pure_waiting_time for inst in completed_instances if inst.kernel.kernel_name == kernel.kernel_name])
   	total_time = total_execution_time + total_waiting_time
   	kernel_execution_ratio[kernel.kernel_name] = total_execution_time / total_time if total_time > 0 else 0
   	kernel_waiting_ratio[kernel.kernel_name] = total_waiting_time / total_time if total_time > 0 else 0

Let's discuss these stats.

[tancheng]

All resolved/answered offline except sth wrong with:

I don't see a CGRA class, how do we calculate overall CGRA utilization?

@MeowMJ will fix this in the updated script, and close this issue accordingly.