import numpy as np
import conf
class PE:
PEBuffer = conf.PEBuffer
PEState = conf.ClockGate
def __init__(self):
self.SetFilterWeight((0, 0))
self.SetImageRow((0, 0))
def SetPEState(self, State):
self.PEState = State
def SetFilterWeight(self, FilterWeight):
self.FilterWeight = FilterWeight
def SetImageRow(self, ImageRow):
self.ImageRow = ImageRow
def SetPEImgAndFlt(self, ImageNum, FilterNum):
self.ImageNum = ImageNum
self.FilterNum = FilterNum
def __SetPsum__(self, Psum):
self.Psum = Psum
def __Conv1d__(self, ImageRow, FilterWeight):
result = list()
for x in range(0, len(ImageRow) - 1 + len(FilterWeight)):
y = x + len(FilterWeight)
if y > len(ImageRow):
break
r = ImageRow[x:y] * FilterWeight
result.append(r.sum())
return np.array(result)
def __Conv__(self):
ImageRow = self.ImageRow
FilterWeight = self.FilterWeight
ImageNum = self.ImageNum
FilterNum = self.FilterNum
l = list()
if FilterNum == 1 and ImageNum == 1:
# 图和核都为1 直接运行卷积
return self.__Conv1d__(ImageRow, FilterWeight)
else:
# 核为1 ， filter重用
if FilterNum == 1:
# 水平分割为原始图的每一行
pics = np.hsplit(ImageRow, ImageNum)
# 遍历，卷积
for x in pics:
# 卷积后的结果加入l中临时保存
l.append(self.__Conv1d__(x, FilterWeight))
# 将l中的结果组合成一个新的矩阵
# 横向组合
result = np.hstack(np.array(l))
# 返回结果
return result
# 图为1 ，img重用
if ImageNum == 1:
# 将FilterWeight变为矩阵
FilterWeight = np.reshape(FilterWeight, (int(FilterWeight.size / FilterNum), FilterNum))
flts = np.array(FilterWeight.T)
for x in flts:
l.append(self.__Conv1d__(ImageRow, x))
result = np.array(l)
result = result.T
result = np.reshape(result, (1, result.size))
return result
def CountPsum(self):
if self.PEState == conf.ClockGate:
self.__SetPsum__(conf.EmptyPsum)
elif self.PEState == conf.Running:
self.__SetPsum__(self.__Conv__())
