mapping.py

from util import *
from visualize import *
import math
def auto_mapping_average_tile(model,hardware,Parallism=[1,1,1,1,1],auto_split=False):
    pp_num=Parallism[0]
    assert pp_num<=len(model)
    total_cp=0
    for op in model:
        total_cp+=op.fd_macs
    total_cp=total_cp
    cp_per_stage=total_cp/pp_num
    stage_ops = [[] for _ in range(pp_num)]
    cur_cp_status=0
    cur_id=0
    #mapping
    for op in model:
        cur_cp_status+=(op.fd_macs)
        #print(cur_cp_status,cp_per_stage)
        if cur_cp_status<=cp_per_stage or cur_id==pp_num-1:
            stage_ops[cur_id].append(op)
        else:
            stage_ops[cur_id+1].append(op)
            cur_cp_status=(op.fd_macs)
            cur_id+=1 
    pp_tiles_num=[]
    for i in range(cur_id+1):
        cur_cp_status=0
        for op in stage_ops[i]:
            cur_cp_status+=(op.fd_macs)
        #print(math.ceil(cur_cp_status/total_cp*hardware.tile_num))
        pp_tiles_num.append(math.ceil(cur_cp_status/total_cp*hardware.tile_num))
    stage_devices=hardware.tile_split_by_pp(pp_tiles_num)
    #parallelism
    for i in range(cur_id+1):
        for op in stage_ops[i]:
            if auto_split:
                op.dpmap(devices=stage_devices[i],p_sgy=[1,len(stage_devices[i]),1,1])
            else:
                #print(len(stage_devices[i]),Parallism[1:])
                op.dpmap(devices=stage_devices[i],p_sgy=Parallism[1:])
    if hardware.topo_tpye!='gpu_like':
        draw_mapping(hardware,'test',tiles=stage_devices,path='sim_visualize',ori=False)
    return stage_devices,stage_ops