v0.9.18

NNlib v0.9.18

Diff since v0.9.17

Merged pull requests:

• Transition to LinearAlgebra.BLAS.libblastrampoline instead of Base.libblas_name (#589) (@staticfloat)
• fix typos in the doc for batched_mul! (#590) (@lanceXwq)
• Use GPU-friendly rounding (#591) (@pxl-th)
• Support conv_direct! on custom datatypes (#592) (@adrhill)
• Fix padding causing device to host copies (#593) (@pxl-th)

Closed issues:

• Error when using Float64: ERROR: UndefRefError: access to undefined reference (#490)
• Warning on using Upsample on AMDGPU (#573)