[feat] add Hashtable On GPU #74
Conversation
rhdong
commented
May 10, 2021
rhdong
commented
- add Hashtable On GPU
- switch cuda tool chain to cuda11
- update STYLE GUIDE for clang format on MacOS.
- update bazel version to 3.7.2
| size_t default_value_num = | ||
| is_full_default ? default_value.shape().dim_size(0) : 1; | ||
| CUDA_CHECK(cudaStreamCreate(&_stream)); | ||
| CUDA_CHECK(cudaMalloc((void**)&d_status, sizeof(bool) * len)); |
There was a problem hiding this comment.
Is it to use cudaMallocManaged?
Allocates memory that will be automatically managed by the Unified Memory system.
There was a problem hiding this comment.
I have tried but fail, and I will consider your advice next version, for this is a stable implement.
| 3.7.2 |
There was a problem hiding this comment.
Is the bazel version 3.7.2 necessary? If it is, please also mention it in README.
| ~CuckooHashTableOfTensorsGpu() { delete table_; } | ||
|
|
||
| void CreateTable(size_t max_size, gpu::TableWrapperBase<K, V>** pptable) { | ||
| if (runtime_dim_ <= 50) { |
There was a problem hiding this comment.
I think it is possible to use mod 50 and switch statement to achieve better performance.
There was a problem hiding this comment.
Just run one time, and mod50 & switch will have worse readability
| CUDA_CHECK(cudaStreamCreate(&_stream)); | ||
| CUDA_CHECK(cudaMalloc((void**)&d_keys, sizeof(K) * len)); | ||
| CUDA_CHECK(cudaMalloc((void**)&d_values, sizeof(V) * runtime_dim_ * len)); | ||
| CUDA_CHECK(cudaMemcpy((void*)d_keys, (void*)keys.tensor_data().data(), |
There was a problem hiding this comment.
If the operation is registered on Device_GPU, the input Tensor is already on GPU device memory. So it is not strictly needed to do the copy.
There was a problem hiding this comment.
I tested it, sometimes inputs were in CPU memory that caused crash, so I used memcpy..
rhdong
force-pushed
the
rhdong/support-gpu-ops
branch
from
ded0900
to
32eb233
Compare
rhdong
force-pushed
the
rhdong/support-gpu-ops
branch
from
9ebd069
to
9d71562
Compare
