U
    9%e3                     @   s  d dl Z d dlZd dlmZ d dlmZ d dlmZmZ d dl	m
Z
 d dlmZ d dlmZ dd	 d
fedddZe jedd	 ddZe jedd	 d
dZeG dd dZeG dd dZG dd dZG dd dZd#ddZdd	 d dfddZdd  Zd!d" ZdS )$    N)deque)	dataclass)DictList)_KinetoEventprofile)
DeviceTypec                 C   s   | j S N)childrenx r   T/var/www/html/Darija-Ai-API/env/lib/python3.8/site-packages/torch/profiler/_utils.py<lambda>       r   F)reversec                 c   sP   |rt ndd }t|| }|rL||}|V  |||D ]}|| q:qd S )Nc                 S   s   | S r
   r   r   r   r   r   r      r   z_traverse.<locals>.<lambda>)reversedr   append)treenext_fnZchildren_fnr   order	remaining
curr_eventchild_eventr   r   r   	_traverse   s    r   c                 C   s   |   S r
   )popr   r   r   r   r      r   T)r   r   c                 C   s   |   S r
   )popleftr   r   r   r   r      r   c                   @   sJ   e Zd ZU dZeed< dZeed< dZeed< dZeed< e	dd Z
dS )	EventMetricsr   duration_time_nsself_time_nsidle_time_nsqueue_depthc                 C   s   | j dkrdS | j| j  S )Nr   g        )r   r!   selfr   r   r   fraction_idle_time$   s    
zEventMetrics.fraction_idle_timeN)__name__
__module____qualname__r   int__annotations__r    r!   r"   propertyr%   r   r   r   r   r      s   
r   c                   @   s*   e Zd ZU eed< eed< dZeed< dS )Intervalstartendr   r"   N)r&   r'   r(   r)   r*   r"   r   r   r   r   r,   +   s   
r,   c                   @   s>   e Zd Zdd Zdd Zdd Zdd Zee d	d
dZ	dS )EventKeyc                 C   s
   || _ d S r
   event)r$   r1   r   r   r   __init__3   s    zEventKey.__init__c                 C   s   t | jjS r
   )hashr1   idr#   r   r   r   __hash__6   s    zEventKey.__hash__c                 C   s   | j j|j jkS r
   )r1   r4   )r$   otherr   r   r   __eq__9   s    zEventKey.__eq__c                 C   s
   | j j S r
   )r1   namer#   r   r   r   __repr__<   s    zEventKey.__repr__)	intervalsc           	      C   s   d}t |dd d}|rTt| jj|d j}t| jj|d j}||k rT||| 7 }d\}}|t|k r|| }|| }|d7 }|j|jkr|j|jkr|d7 }q\n|j|_|}t| jj|j}t| jj|j}||k r\||| 7 }q\|S )Nr   c                 S   s   | j S r
   r-   r   r   r   r   r   A   r   z,EventKey.intervals_overlap.<locals>.<lambda>key)r      r>   )	sortedmaxr1   start_time_nsr-   minend_time_nsr.   len)	r$   r:   Zoverlap_timeZoverlap_startZoverlap_endijZprev_intervalZcurr_intervalr   r   r   intervals_overlap?   s.    zEventKey.intervals_overlapN)
r&   r'   r(   r2   r5   r7   r9   r   r,   rG   r   r   r   r   r/   2   s
   r/   c                   @   sL   e Zd ZedddZdd Zdd Zdd	 Zd
d Zde	e
dddZdS )BasicEvaluation)profc                 C   sd   || _ i | _|   tdd | j D dd d| _dd | jD | _g | _|  | _	| 
  d S )Nc                 s   s   | ]
}|V  qd S r
   r   .0er   r   r   	<genexpr>f   s     z+BasicEvaluation.__init__.<locals>.<genexpr>c                 S   s   | j jS r
   )r1   rA   r   r   r   r   r   f   r   z*BasicEvaluation.__init__.<locals>.<lambda>r<   c                 S   s   g | ]
}|j qS r   r0   rJ   r   r   r   
<listcomp>h   s     z,BasicEvaluation.__init__.<locals>.<listcomp>)r   metricscompute_self_timer?   keysZ
event_keyseventscuda_eventscompute_queue_depthqueue_depth_listcompute_idle_time)r$   rI   r   r   r   r2   a   s     
zBasicEvaluation.__init__c                 C   s   | j jdk	stt| j j }|r| }|j}|jD ]}||j8 }|| q8t	|| j
ksxtd|j d|j t|d| j
t	|< |j| j
t	| _q dS )zM
        Computes event's self time(total time - time in child ops).
        NzDuplicate id: z, )r    )r   kineto_resultsAssertionErrorr   Zexperimental_event_treer   r   r   r   r/   rO   r4   r8   r   )r$   stackr   	self_timer   r   r   r   rP   m   s"    

z!BasicEvaluation.compute_self_timec                    s  | j jdk	st| j j }dd dd tfdd|D dd	 d
}tfdd|D dd	 d
}t|| dd	 d
| _i }d}|D ]2 t| fdd	|d}|| < |dk	r|n|}qd}d}|| | j }	dd }
g }|	j|
d
 |	D ]}t|drB|	 d }|	 |
  d }||krZ|| dk	rZ|| }nt|drZ|j}|j}|t|k r|| 	 d |kr|d7 }qZ|| d }t|d}t|dr|t||| qt|dr|| jt| _q|S )z
        Computes queue_depth at each event. This will calculate the queue depth data for
        All the events in the tree.
        This will return a list of Interval of queue depth data of cuda launch and kernels.
        Nc                 S   s
   | j dkS )NZcudaLaunchKernel)r8   rL   r   r   r   is_cuda_launch_kernel   s    zBBasicEvaluation.compute_queue_depth.<locals>.is_cuda_launch_kernelc                 S   s   |   tjkod| j kS )NZmem)Zdevice_typer	   CUDAr8   lowerr[   r   r   r   is_cuda_kernel   s    z;BasicEvaluation.compute_queue_depth.<locals>.is_cuda_kernelc                 3   s   | ]} |r|V  qd S r
   r   rJ   )r\   r   r   rM      s      z6BasicEvaluation.compute_queue_depth.<locals>.<genexpr>c                 S   s   |   S r
   start_usr   r   r   r   r      r   z5BasicEvaluation.compute_queue_depth.<locals>.<lambda>r<   c                 3   s   | ]} |r|V  qd S r
   r   rJ   )r_   r   r   rM      s      c                 S   s   |   S r
   r`   r   r   r   r   r      r   c                 S   s   |   S r
   r`   r   r   r   r   r      r   r   c                    s   |      kS r
   )Zlinked_correlation_idr   )cuda_launch_eventr   r   r      s   r;   c                 S   s2   t | dr|  d S t | dr&| jS tdd S )Nra     rA   zUnknown Event Type)hasattrra   rA   	Exceptionr0   r   r   r   new_old_event_comparator   s
    

zEBasicEvaluation.compute_queue_depth.<locals>.new_old_event_comparatorra   rd   rA   r>   )r   rW   rX   rR   r?   rS   index_of_first_matchsortre   ra   Zduration_usrA   rC   rD   r@   r   r,   rO   r/   r"   )r$   Zcuda_event_listZcuda_launch_eventsZcuda_kernel_eventsZkernel_mappingZlast_mapped_kernelindexZcurrent_kernel_indexZspawned_kernel_indexZ
all_eventsrg   rU   r1   
start_timeend_timeZcurrent_queue_depthr   )rb   r_   r\   r   rT      sr     





z#BasicEvaluation.compute_queue_depthc                 C   s   d}d}g }| j rP| jrP|t| jd j| j d jt| j d j| jd jg7 }| j D ]@}|jdkrr|sr|j}d}|jdkrV|rV|t||j d}qVdd | j	
 D }|D ]}t||| j	t| _qdS )z4
        Computes idle time of the profile.
        Fr   rc   Tc                 S   s   g | ]
}|j qS r   r0   rJ   r   r   r   rN      s     z5BasicEvaluation.compute_idle_time.<locals>.<listcomp>N)rU   rR   r,   rA   r-   r.   rC   r"   r   rO   rQ   r/   rG   r!   )r$   idleZ
idle_startZidle_intervalsZ
data_point
event_listr1   r   r   r   rV      s,    
z!BasicEvaluation.compute_idle_timec                    s  ddl }ttj}dd |D }d d}g d}|t|k r||  krV|d7 }q4t|d t|D ]l}t| fdd|d	}t|||d
}	|	dk	rh||	 |krht	||	 j
|| j
 |dk	r|n|} qqh|d7 }q4fddj D }
|
r|jfdd|
D |jd}|jfdd|
D |jd}||| || }||| || }|d|  }dd tt||
dd ddD }
|
d| }
|
S )a  
        Filter and Rank the events based on some heuristics:
        1) Events that are in the falling phase of the queue depth.
        2) Events that have a high idle_time, self_time difference.

        Parameters:
            length: The number of events to return.
        r   Nc                 S   s   g | ]
}|j qS r   )r"   rJ   r   r   r   rN     s     z/BasicEvaluation.rank_events.<locals>.<listcomp>   r>   c                    s   |  kS r
   r   r   )bottom_threasholdr   r   r     r   z-BasicEvaluation.rank_events.<locals>.<lambda>r;   )r-   r.   c                    s   g | ]}|  r|qS r   )rG   rK   r1   )decrease_intervalr   r   rN      s   
c                    s   g | ]} j | jqS r   )rO   r    rq   r#   r   r   rN   '  s     )Zdtypec                    s   g | ]} j | jqS r   )rO   r%   rq   r#   r   r   rN   +  s     g333333?c                 S   s   g | ]\}}|qS r   r   )rK   _r1   r   r   r   rN   3  s   c                 S   s   | d S )Nr   r   r   r   r   r   r   7  r   T)r=   r   )torchlistr   rU   rD   rangerh   argmaxr   r,   r-   rO   rQ   ZtensorZfloat32ZmeanZstdr?   zip)r$   lengthrt   rU   Z	qd_valuesZtop_threasholdrE   rF   Znext_minimum_idxZpeak_idxrn   rZ   Z	idle_timeZnormalized_gainZnormalized_selfZheuristic_score_listr   )rp   rr   r$   r   rank_events   sh     
  

zBasicEvaluation.rank_eventsr>   T)ry   print_enablec                    sJ     |}|s|S |rdnd}|d fdd|D 7 }|rFt| |S )NzOptimizable events:
zNo events to optimize

c                    s@   g | ]8}d  d| dt |j d j| jd ddd  	qS )zP--------------------------------------------------------------------------------z
Event:                z
Source code location: z
Percentage idle time: d   z.2fz%
)source_code_locationr1   rO   r%   rq   r#   r   r   rN   E  s   z:BasicEvaluation.get_optimizable_events.<locals>.<listcomp>)rz   joinprint)r$   ry   r{   rn   outputr   r#   r   get_optimizable_events>  s    


z&BasicEvaluation.get_optimizable_eventsN)r>   T)r&   r'   r(   r   r2   rP   rT   rV   rz   r)   boolr   r   r   r   r   rH   `   s   VIrH   c                 C   sD   |d ks|t | krt | }t||D ]}|| | r&|  S q&d S r
   )rD   rv   )seq	predicater-   r.   rE   r   r   r   rh   S  s    
rh   c                 C   s   | S r
   r   r   r   r   r   r   \  r   c                 C   s2   | || } t | dkrd S | t| |d| S )Nr   r<   )rD   rj   r@   )r   r=   r-   r.   r   r   r   rw   \  s    rw   c                 C   s0   | d k	r,t d| j}|d kr&| j} q | jS dS )Nz
\.py\(.*\)zNo source code location found)researchr8   parent)r1   matchr   r   r   r~   c  s    r~   c               	   C   s"   ddl m}  |   W 5 Q R X d S )Nr   r   )torch.autograd.profilerr   r   r   r   r   _init_for_cuda_graphsq  s    r   )r   N)	functoolsr   collectionsr   dataclassesr   typingr   r   Ztorch.autogradr   r   r   Ztorch.profilerr	   r   r   partialZtraverse_dfsZtraverse_bfsr   r,   r/   rH   rh   rw   r~   r   r   r   r   r   <module>   s2   
  . t
	