Tensorflow and TVM gives different results with the same model

#1

I had found something strange…
When I tried to convert a .pb model(containing reshape op and conv2d op) into TVM, as seen in the following pictrue, i obatin the different results by Tensorflow and TVM.

%7B2C5C6CA0-5931-419A-9282-79077F1346D1%7D_20191106113052
{2C5C6CA0-5931-419A-9282-79077F1346D1}_20191106113052.jpg1240×915 66.5 KB

I test the result by numpy.testing.allclose(tvm_out, tf_out, atol=1e-5, rtol=1e-5), then the result is

 AssertionError: 
Not equal to tolerance rtol=1e-05, atol=1e-05
(mismatch 0.2635099449936291%)
 x: array([ 11.083764, -14.04364 ,  -1.184902, ..., -32.929897, -35.814198,
        76.43094 ], dtype=float32)
 y: array([ 11.083763, -14.043638,  -1.184902, ..., -32.929893, -35.814198,
        76.43094 ], dtype=float32)

Can anyone help me?
Some guesses are the tensorflow implements the conv2d by im2col+gemm(blas) which is different from tvm. and loss of accuracy when performing multiply add.

#2

Latest progress:
when setting target = 'cuda -libs=cublas' , correct results are obtained. So blas is the key point that results in the accuracy lost.
Finally, can anybody provide some tutorials that show how to implement conv2d by cblas in cpu?