[add] support for half (GPU only) #6
Conversation
[fix] pytorch binding case torch::ScalarType::Half
There was a problem hiding this comment.
Thanks for taking that work!
As we talked off-the-list, the following changes are needed:
- torch bindings are WIP/borked
- remove
__host__fromHOSTDEVICEmacro and re-use it. - Subsequently,
log_sum_expfromrnnt_helpercan be re-used from that.
Also, more a general note: This PR drop support for Kepler and Maxwell archs. I'm not against it but I have to confirm it won't affect too many people (unlikely though).
|
Thank you for your feedback. |
There was a problem hiding this comment.
Sorry for the delay! Compilation is OK but import will fail with the following error:
ImportError: /home/b-flo/warp-transducer/pytorch_binding/warprnnt_pytorch/warp_rnnt.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZNK2at10TensorBase8data_ptrI6__halfEEPT_vI
I'll let you figure it out but feel free to ask off the list for any help/hints on resolving this error!
P.S.: Postponing the full review until it's in a stable state.
CMakeLists.txt
Outdated
| # Drop support for old GPU to use CUDA math library functions | ||
| IF(NOT (CUDA_VERSION GREATER 10.2)) | ||
| set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_30,code=sm_30 -O2") | ||
| set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_35,code=sm_35") | ||
| set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_50,code=sm_50") | ||
| # set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_30,code=sm_30 -O2") | ||
| # set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_35,code=sm_35") | ||
| # set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_50,code=sm_50") | ||
| ENDIF() | ||
|
|
||
| set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_52,code=sm_52") | ||
| #set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_52,code=sm_52") |
There was a problem hiding this comment.
Btw, you can remove the condition + commented lines
mangel9742
force-pushed
the
espnet_v1.1_half_support
branch
from
a1bb689 to
77ae1f6
Compare
mangel9742
force-pushed
the
espnet_v1.1_half_support
branch
from
77ae1f6 to
7f48e85
Compare
Support added only for GPU use.
The half type used is from Cuda Math library and some of the functions are only usable for GPU.
Most of the work consists to specialize templates in order to avoid ambiguous calls.