Implement cuDNN `conv2D_batch`. - Githubissues

feiwang3311 / Lantern

BSD 3-Clause "New" or "Revised" License

168 stars 15 forks source link

Implement cuDNN `conv2D_batch`. #31

Closed dan-zheng closed 6 years ago

dan-zheng commented 6 years ago

Move CPU implementation of conv2D_batch and related helpers to BackendCPU.
Implement cuDNN conv2D_batch using cudnnConvolutionForward.
Make fillWithBias backend-defined, implement cuDNN conv2D_batch with bias.

TODO:

Implement cudNN conv2D gradient using cudnnConvolutionBackward.
Consider merging conv2D and conv2D_batch functions. Also consider merging conv, convBP and convBBP defined on TensorR.