o
    h3                     @   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def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 )$    N)deque)	dataclass)DictList)_KinetoEventprofile)
DeviceTypec                 C      | j S N)childrenx r   K/var/www/html/ai/venv/lib/python3.10/site-packages/torch/profiler/_utils.py<lambda>       r   Freversec                 c   sX    |rt ndd }t|| }|r*||}|V  |||D ]}|| q|sd S d S )Nc                 S      | S r   r   r   r   r   r   r          z_traverse.<locals>.<lambda>)reversedr   append)treenext_fnchildren_fnr   order	remaining
curr_eventchild_eventr   r   r   	_traverse   s   r   c                 C      |   S r   )popr   r   r   r   r          T)r   r   c                 C   r    r   )popleftr   r   r   r   r      r"   c                   @   sJ   e Zd ZU dZeed< dZeed< dZeed< dZeed< e	dd Z
dS )	EventMetricsr   duration_time_nsself_time_nsidle_time_nsqueue_depthc                 C   s   | j dkrdS | j| j  S )Nr   g        )r%   r'   selfr   r   r   fraction_idle_time$   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/   r0   r(   r   r   r   r   r2   +   s   
 r2   c                   @   s>   e Zd Zdd Zdd Zdd Zdd Zd	ee fd
dZ	dS )EventKeyc                 C   s
   || _ d S r   event)r*   r7   r   r   r   __init__3      
zEventKey.__init__c                 C   s   t | jjS r   )hashr7   idr)   r   r   r   __hash__6   s   zEventKey.__hash__c                 C   s   | j j|j jkS r   )r7   r;   )r*   otherr   r   r   __eq__9   s   zEventKey.__eq__c                 C   s
   | j j S r   )r7   namer)   r   r   r   __repr__<   r9   zEventKey.__repr__	intervalsc           	      C   s   d}t |dd d}|r*t| jj|d j}t| jj|d j}||k r*||| 7 }d\}}|t|k rw|| }|| }|d7 }|j|jkrW|j|jkrQ|d7 }q.|j|_|}t| jj|j}t| jj|j}||k rq||| 7 }|t|k s4|S )Nr   c                 S   r
   r   r3   r   r   r   r   r   A   r   z,EventKey.intervals_overlap.<locals>.<lambda>key)r      rE   )	sortedmaxr7   start_time_nsr3   minend_time_nsr4   len)	r*   rA   overlap_timeoverlap_startoverlap_endijprev_intervalcurr_intervalr   r   r   intervals_overlap?   s0   zEventKey.intervals_overlapN)
r,   r-   r.   r8   r<   r>   r@   r   r2   rS   r   r   r   r   r5   2   s    r5   c                   @   sN   e Zd ZdefddZdd Zdd Zdd	 Zd
d Zdde	de
f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   )r7   rH   r   r   r   r   r   f   r"   z*BasicEvaluation.__init__.<locals>.<lambda>rC   c                 S      g | ]}|j qS r   r6   rV   r   r   r   
<listcomp>h       z,BasicEvaluation.__init__.<locals>.<listcomp>)r   metricscompute_self_timerF   keys
event_keyseventscuda_eventscompute_queue_depthqueue_depth_listcompute_idle_time)r*   rU   r   r   r   r8   a   s   
zBasicEvaluation.__init__c                 C   s   | j jdusJ t| j j }|rS| }|j}|jD ]}||j8 }|| qt|| j	vs<J d|j
 d|j t|d| j	t|< |j| j	t| _|sdS dS )zM
        Computes event's self time(total time - time in child ops).
        NzDuplicate id: z, )r&   )r   kineto_resultsr   experimental_event_treer!   r%   r   r   r5   r]   r;   r?   r$   )r*   stackr   	self_timer   r   r   r   r^   m   s$   

z!BasicEvaluation.compute_self_timec                    s  | j jdusJ | 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 ] t| fdd	|d}|| < |dur\|n|}qEd}d}|| | j }	dd }
g }|	j|
d
 |	D ]|}t|dr| d }| |	  d }||v r|| dur|| }nt|dr|j
}|j}|t|k r||  d |kr|d7 }|t|k r||  d |ks|| d }t|d}t|dr|t||| qxt|dr|| jt| _qx|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 )NcudaLaunchKernel)r?   rX   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 vS )Nmem)device_typer	   CUDAr?   lowerrk   r   r   r   is_cuda_kernel   s   z;BasicEvaluation.compute_queue_depth.<locals>.is_cuda_kernelc                 3       | ]	} |r|V  qd S r   r   rV   )rl   r   r   rY          z6BasicEvaluation.compute_queue_depth.<locals>.<genexpr>c                 S   r    r   start_usr   r   r   r   r      r"   z5BasicEvaluation.compute_queue_depth.<locals>.<lambda>rC   c                 3   rr   r   r   rV   )rq   r   r   rY      rs   c                 S   r    r   rt   r   r   r   r   r      r"   c                 S   r    r   rt   r   r   r   r   r      r"   r   c                    s   |      kS r   )linked_correlation_idr   )cuda_launch_eventr   r   r      s    rB   c                 S   s.   t | dr|  d S t | dr| jS td)Nru     rH   zUnknown Event Type)hasattrru   rH   	Exceptionr6   r   r   r   new_old_event_comparator   s
   

zEBasicEvaluation.compute_queue_depth.<locals>.new_old_event_comparatorru   ry   rH   rE   )r   rf   ra   rF   rb   index_of_first_matchsortrz   ru   duration_usrH   rJ   rK   rG   r   r2   r]   r5   r(   )r*   cuda_event_listcuda_launch_eventscuda_kernel_eventskernel_mappinglast_mapped_kernelindexcurrent_kernel_indexspawned_kernel_index
all_eventsr|   rd   r7   
start_timeend_timecurrent_queue_depthr   )rw   rq   rl   r   rc      sx   






z#BasicEvaluation.compute_queue_depthc                 C   s   d}d}g }| j r(| jr(|t| jd j| j d jt| j d j| jd jg7 }| j D ] }|jdkr9|s9|j}d}|jdkrK|rK|t||j d}q+dd | j	
 D }|D ]}t||| j	t| _qXdS )z4
        Computes idle time of the profile.
        Fr   rx   Tc                 S   rZ   r   r6   rV   r   r   r   r[      r\   z5BasicEvaluation.compute_idle_time.<locals>.<listcomp>N)rd   ra   r2   rH   r3   r4   rJ   r(   r   r]   r_   r5   rS   r'   )r*   idle
idle_startidle_intervals
data_point
event_listr7   r   r   r   re      s0   
z!BasicEvaluation.compute_idle_timec                    s  ddl }ttj}dd |D }d d}g d}|t|k ru||  kr+|d7 }qt|d t|D ]6}t| fdd|d	}t|||d
}	|	durj||	 |krjt	||	 j
|| j
 |durf|n|} nq4|d7 }|t|k s 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   rZ   r   )r(   rV   r   r   r   r[     r\   z/BasicEvaluation.rank_events.<locals>.<listcomp>   rE   c                    s   |  kS r   r   r   )bottom_threasholdr   r   r     r"   z-BasicEvaluation.rank_events.<locals>.<lambda>rB   )r3   r4   c                    s   g | ]	}|  r|qS r   )rS   rW   r7   )decrease_intervalr   r   r[      s    c                       g | ]} j | jqS r   )r]   r&   r   r)   r   r   r[   '      )dtypec                    r   r   )r]   r+   r   r)   r   r   r[   +  r   g333333?c                 S   s   g | ]\}}|qS r   r   )rW   _r7   r   r   r   r[   3  s    c                 S   s   | d S )Nr   r   r   r   r   r   r   7  r"   T)rD   r   )torchlistr   rd   rK   ranger}   argmaxr   r2   r3   r]   r_   tensorfloat32meanstdrF   zip)r*   lengthr   rd   	qd_valuestop_threasholdrO   rP   next_minimum_idxpeak_idxr   ri   	idle_timenormalized_gainnormalized_selfheuristic_score_listr   )r   r   r*   r   rank_events   sf   
zBasicEvaluation.rank_eventsrE   Tr   print_enablec                    sJ     |}|s	|S |rdnd}|d fdd|D 7 }|r#t| |S )NzOptimizable events:
zNo events to optimize

c                    s@   g | ]}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_locationr7   r]   r+   r   r)   r   r   r[   E  s    z:BasicEvaluation.get_optimizable_events.<locals>.<listcomp>)r   joinprint)r*   r   r   r   outputr   r)   r   get_optimizable_events>  s   


z&BasicEvaluation.get_optimizable_eventsN)rE   T)r,   r-   r.   r   r8   r^   rc   re   r   r/   boolr   r   r   r   r   rT   `   s    VIrT   c                 C   sD   |d u s
|t | krt | }t||D ]}|| | r|  S qd S r   )rK   r   )seq	predicater3   r4   rO   r   r   r   r}   S  s   r}   c                 C   r   r   r   r   r   r   r   r   \  r   c                 C   s2   | || } t | dkrd S | t| |d| S )Nr   rC   )rK   r   rG   )r   rD   r3   r4   r   r   r   r   \  s   r   c                 C   s0   | d urt d| j}|d u r| j} q | jS dS )Nz
\.py\(.*\)zNo source code location found)researchr?   parent)r7   matchr   r   r   r   c  s   r   c                  C   s8   ddl m}  |  	 W d    d S 1 sw   Y  d S )Nr   r   )torch.autograd.profilerr   r   r   r   r   _init_for_cuda_graphsq  s   "r   )r   N)	functoolsr   collectionsr   dataclassesr   typingr   r   torch.autogradr   r   r   torch.profilerr	   r   r   partialtraverse_dfstraverse_bfsr$   r2   r5   rT   r}   r   r   r   r   r   r   r   <module>   s0    

. 
t	