1. Add mutex when operating the memory pool.
2. Support CMake 3.19 and fix some CMake bugs. Note that CUDA_ROOT variable in CMake is modified as CUDA_TOOLKIT_ROOT. You can find this update in the README.
3. Fix some bugs on Ubuntu.

1. Add mutex when operating the memory pool.
2. Support CMake 3.19 and fix some CMake bugs. Note that CUDA_ROOT variable in CMake is modified as CUDA_TOOLKIT_ROOT. You can find this update in the README.
3. Fix some bugs on Ubuntu.

1. Support Inplace way for some functions.
2. Add new flexible constructor for XList.
3. Bug fixed.

Fix the minor bugs in IsSameShaped.

Find some fp16 bugs during decoding NMT system. We decide to roll back some codes to the last version. These codes need to be reviewed.

Find some bugs during training NMT system. We decide to roll back some codes to the last version. These codes need to be reviewed.

Fix the __half bugs in SetData.

1. Fix the bugs in backward process.
2. Support the float16 class.
3. Fix bugs.
4. Clean the codes.
5. Remove the makefile.

Merge with the branch of xuchen (NOT update the float16, this needs code review) and fix the bugs in Gather function.
1. Support Reciprocal fucntion.
2. Fix the safe delete bugs in XDevice.
3. Support new API to convert the data type of tensor.
4. Support to show the memory usage of buffer memory.
5. Fix minor errors.

1. Try to fix the bugs in destroy the stream of XDevice (Uncheck).
2. Fix the bug of memory leak in ReduceSumAll function.
2. Adjust the directory structure.
3. Fix the minor errors.

Fix the bugs in Sum and Sub when MKL or OpenBlas is used.

1. Update the codes of machine translation sample. The current version is the same with NiuTrans.NMT.
2. Update the XList class.
3. Bugs fix.

1. Offer macro to set whether use the half precision in cuda codes.
2. Update the manuals.

1. Fix the memory leak bugs in XList.
2. Replace the half with the unsigned short data type.

Fix the bugs in ReduceMax and its related functions.

Support fp16 data type for more operations and fix the minor errors.  (Don't use this! It's an incomplete version)

Support fp16 data type for more operations and fix the minor errors.  (Don't use this! It's an incomplete version)

Support fp16 data type for more operations and fix the minor errors.  (Don't use this! It's an incomplete version)

1. Fix the mistakes in manual.
2. Fix the minor errors in Stack and LogSoftmax functions.

1. Fix the mistakes in manual.
2. Fix the minor errors in Normalize and ScaleAndShift functions.

1. Fix the mistakes in manual. By the way, I have to say there are so many mistakes in the manual. I'm shocked it has been checked a lot of times, but why they are still be there. No one care about that? Really???
2. Clean the codes.

1. Call the Scale or Shift functions when the parameter shift is 0.0F or scale is 1.0F.
2. Add the unit test for ScaleAndShift for the cases which parameter shift is 0.0F or scale is 1.0F.

Optimize the MultiplyMe and DivMe functions when operate the scalar tensors.

1. Fix the bug in DivMe function which cannot handle the scalar tensor and broadcast case.
2. Clean the codes.
3. Fix minor errors.

1. Fix the bug in SubMe function which cannot handle the scalar tensor and broadcast case.
2. Clean the codes.

Fix the bug in SumMe function which cannot handle the scalar tensor and broadcast case.

1. Fix the bug in MultiplyMe function which cannot handle the scalar tensor and broadcast case.
2. Add broadcast multiply unit test.