[Perf] Enhance cudnn and cublas backend and enable TensorCore (#4353)
authorSiyuan Feng <Hzfengsy@vip.qq.com>
Mon, 25 Nov 2019 06:01:55 +0000 (22:01 -0800)
committerLeyuan Wang <laurawly@gmail.com>
Mon, 25 Nov 2019 06:01:55 +0000 (22:01 -0800)
commitdabde40f15423fc9632824dcd95696f36cf17a01
tree6b74eea48e21e24003093b7c40e2008ad4c8c0be
parentfbb2a356df990e8fb8b5d76d4d0124fd3b0378c8
[Perf] Enhance cudnn and cublas backend and enable TensorCore (#4353)

* add half and mix precision support to cublas backend

* add TensorCore support in CuDNN

* enhance CuDNN support

* address comments and fix lint

* fix

* add fp16 test
include/tvm/runtime/util.h
python/tvm/contrib/cublas.py
python/tvm/contrib/cudnn.py
src/runtime/contrib/cblas/gemm_common.h
src/runtime/contrib/cublas/cublas.cc
src/runtime/contrib/cublas/cublas_utils.h
src/runtime/contrib/cudnn/conv_forward.cc
tests/python/contrib/test_cublas.py
tests/python/contrib/test_cudnn.py
topi/python/topi/cuda/conv2d.py
topi/python/topi/cuda/dense.py