GitLab

[Performance Improvement] Make GPU sampling and to_block use pinned memory to decrease required synchronization (#5685)

* fix empty tensor is treated as pinned

* avoid calling cudaHostGetDevicePointer on nullptr

* update empty array

* add a comment

* replace

* blabla

* balbla

* blabla
Co-authored-by: Steve <ubuntu@ip-172-31-34-29.ap-northeast-1.compute.internal>

* param

* brief

* note

* return

* tparam

* brief2

* file

* return2

* return

* blabla

* all
Co-authored-by: Steve <ubuntu@ip-172-31-34-29.ap-northeast-1.compute.internal>

* [Misc] clang-format auto fix.

* manual

* manual

* manual

* manual

* todo

* fix
Co-authored-by: Steve <ubuntu@ip-172-31-34-29.ap-northeast-1.compute.internal>

* get device pointers

* change if condition to IsPinned

* sample neighbors with masks

* oops

* refactor again

* remove

* remove debug code

* rename macro

* address comments

* address comment

* address comments

* rename a lot of stuff

* oops

* rename `DLContext` to `DGLContext`

* rename `kDLGPU` to `kDLCUDA`

* replace DLTensor with DGLArray

* fix linting

* Unify DGLType and DLDataType to DGLDataType

* Fix FFI

* rename DLDeviceType to DGLDeviceType

* decouple dlpack from the core library

* fix bug

* fix lint

* fix merge

* fix build

* address comments

* rename dl_converter to dlpack_convert

* remove redundant comments

* add set_stream

* add .record_stream for NDArray and HeteroGraph

* refactor dgl stream Python APIs

* test record_stream

* add unit test for record stream

* use pytorch's stream

* fix lint

* fix cpu build

* address comments

* address comments

* add record stream tests for dgl.graph

* record frames and update dataloder

* add docstring

* update frame

* add backend check for record_stream

* remove CUDAThreadEntry::stream

* record stream for newly created formats

* fix bug

* fix cpp test

* fix None c_void_p to c_handle

* Use an internal cuda stream for CopyDataFromTo

* small fix white space

* Fix to compile

* Make stream optional in copydata for compile

* fix lint issue

* Update cub functions to use internal stream

* Lint check

* Update CopyTo/CopyFrom/CopyFromTo to use internal stream

* Address comments

* Fix backward CUDA stream

* Avoid overloading CopyFromTo()

* Minor comment update

* Overload copydatafromto in cuda device api
Co-authored-by: xiny <xiny@nvidia.com>

* add weighted sampling without replacement (A-Chao)

* improve Algorithm A-Chao with block-wise prefix sum

* correctly fill out_idxs

* implement weighted sampling with replacement

* small fix

* merge host-side code of weighted/uniform sampling

* enable unit tests for cuda weighted sampling

* move thrust/cub wrapper to the cmake file

* update docs accordingly

* fix linting

* fix linting

* fix unit test

* Bump external CUB/Thrust versions

* Fix code style and update description of algorithm design

* [Feature] GPU support weighted graph neighbor sampling
commit by pengqirong(OPPO)

* merge pengqirong's implementation

* revert the change to cub and thrust

* fix linting

* use DeviceSegmentedSort for better performance

* add more comments

* add necessary notes

* add necessary notes

* resolve some comments

* define THRUST_CUB_WRAPPED_NAMESPACE

* fix doc
Co-authored-by: 彭齐荣 <657017034@qq.com>

* Change the curand_init parameter

* Change the curand_init parameter

* commit

* commit

* change the curandState and launch dim of CSRRowwiseSample kernel

* commit

* keep  _CSRRowWiseSampleReplaceKernel in sync
Co-authored-by: nv-dlasalle <63612878+nv-dlasalle@users.noreply.github.com>

* Change the curand_init parameter

* Change the curand_init parameter

* commit

* commit
Co-authored-by: nv-dlasalle <63612878+nv-dlasalle@users.noreply.github.com>

* implement pin_memory/unpin_memory/is_pinned for dgl.graph

* update python docstring

* update c++ docstring

* add test

* fix the broken UnifiedTensor

* XPU_SWITCH for kDLCPUPinned

* a rough version ready for testing

* eliminate extra context parameter for pin/unpin

* update train_sampling

* fix linting

* fix typo

* multi-gpu uva sampling case

* disable new format materialization for pinned graphs

* update python doc for pin_memory_

* fix unit test

* UVA sampling for link prediction

* dispatch most csr ops

* update graphsage example to combine uva sampling and UnifiedTensor

* update graphsage example to combine uva sampling and UnifiedTensor

* update graphsage example to combine uva sampling and UnifiedTensor

* update doc

* update examples

* change unitgraph and heterograph's PinMemory to in-place

* update examples for multi-gpu uva sampling

* update doc

* fix linting

* fix cpu build

* fix is_pinned for DistGraph

* fix is_pinned for DistGraph

* update graphsage unsupervised example

* update doc for gpu sampling

* update some check for sampling device switching

* fix linting

* adapt for new dataloader

* fix linting

* fix

* fix some name issue

* adjust device check

* add unit test for uva sampling & fix some zero_copy bug

* fix linting

* update num_threads in graphsage examples
Co-authored-by: Quan (Andy) Gan <coin2028@hotmail.com>
Co-authored-by: Jinjing Zhou <VoVAllen@users.noreply.github.com>

* remove

* remove

* fix

* remove

* remove

* Split out separate generators for each thread

* Amortize cost of curand_init

* Improve readability
Co-authored-by: Jinjing Zhou <VoVAllen@users.noreply.github.com>
Co-authored-by: Quan (Andy) Gan <coin2028@hotmail.com>

* Start on uniform GPU sampling

* Save more work

* Get cu file compiling

* Update sampling

* More changes

* Get GPU sampling for uniform probabilities solved

* Fix batch tensor migration

* Fix

* update kernels

* expand blocking

* Undo testing change

* Cut down on sampling overhead

* Fix replacement

* Update unit tests

* Add option to gpu sample in graphsage

* Copy only csc to gpu

* Add ogbn support

* Fix linting

* Remove nvtx from sample

* Improve documentation and error checking

* Expand documentation

* Update assert checking

* delete extra space

* Use standard dataloader when dataset is a dictionary

* ogb -> ogbn

* Fix edge selection determinism

* Fix typos

* Remove nvtx

* Add comment for self.fanout_arrays and assert

* Fix linting

* Migrate to scalarbatcher

* Fix indentation

* Fix batcher

* Fix indexing

* Only use databatcher for GPU

* Convert to DGL NDArray to PyTorch Tensor

* Add optimization for PyTorch's F.tensor() for list of GPU tensors
Co-authored-by: Da Zheng <zhengda1936@gmail.com>