I am timing the speed of a 2d convolution VS an element wise multiplication in pytorch. From my timings the conv2d operation is significantly faster than a simple element wise multiplication.
Conv Time: 0.009611368179321289
Mult Time: 0.5085432529449463
This doesn’t make much sense to me as convolution is O(N^2 k^2) where k is the filter size VS Element wise multiplication of O(N^2)
Is there some reason for this?
Edit: It would appear I need torch.cuda.synchronize() before each call.
After this the timings make sense.
Conv Time: 0.010584831237792969
Mult Time: 0.008290290832519531
import torch
import numpy as np
import time
from torch.nn import functional as F
torchType = torch.FloatTensor
inC = 32
imgH = 64
imgW = 64
outC = 32
filtSize = 3
imgs = np.random.normal(size=(100,inC,imgH,imgW))
imgs = torch.from_numpy(imgs).type(torch.float32).cuda()
filts = torch.randn((inC,outC,filtSize,filtSize)).cuda()
numIters = 500
# Warm up for GPU
for i in range(10):
cLayer = F.conv2d(imgs, filts,bias=None,padding=1)
for i in range(10):
c = torch.mul(imgs,imgs)
# Actual Timing
torch.cuda.synchronize()
st = time.time()
for i in range(numIters):
cLayer = F.conv2d(imgs, filts, bias=None, padding=1)
et = time.time()
print("Conv Time: {}".format(et-st))
torch.cuda.synchronize()
st = time.time()
for i in range(numIters):
c = torch.mul(imgs, imgs)
et = time.time()
print("Mult Time: {}".format(et-st))