Replaced std::vector with HostDeviceVector in MetaInfo and SparsePage. #3446
Conversation
canonizer
commented
Jul 5, 2018
canonizer
commented
- added distributions to HostDeviceVector
- using HostDeviceVector for labels, weights and base margings in MetaInfo
- using HostDeviceVector for offset and data in SparsePage
- other necessary refactoring
- added distributions to HostDeviceVector - using HostDeviceVector for labels, weights and base margings in MetaInfo - using HostDeviceVector for offset and data in SparsePage - other necessary refactoring
|
Yes, the vectors are currently copied between the host and device even if not modified. If you want to avoid this, I can add const-only versions of methods that get host or GPU data. E.g., |
|
This seems like a good idea, otherwise they will be constantly syncing. |
- const versions added to calls that can trigger data transfers, e.g. DevicePointer() - updated the code that uses HostDeviceVector - objective functions now accept const HostDeviceVector<bst_float>& for predictions
- this means no copies are performed if both host and devices access the HostDeviceVector read-only
|
Done. |
- updated the lz4 plugin - added ConstDeviceSpan to HostDeviceVector - using device % dh::NVisibleDevices() for the physical device number, e.g. in calls to cudaSetDevice()
|
Is it worth to store some extra information in one of the GPU related class like number of SMs? Currently all the grid stride functions just use one grid. |
- replaced HostDeviceVector<unsigned int> with HostDeviceVector<int>
Codecov Report
@@ Coverage Diff @@
## master #3446 +/- ##
============================================
+ Coverage 50.39% 50.52% +0.13%
Complexity 188 188
============================================
Files 172 172
Lines 13986 14096 +110
Branches 457 457
============================================
+ Hits 7048 7122 +74
- Misses 6713 6749 +36
Partials 225 225
Continue to review full report at Codecov.
|
RAMitchell
force-pushed
the
dmatrix-hdvec
branch
from
11f0bc6 to
7ec438e
Compare
- added distributions to HostDeviceVector - using HostDeviceVector for labels, weights and base margings in MetaInfo - using HostDeviceVector for offset and data in SparsePage - other necessary refactoring
- const versions added to calls that can trigger data transfers, e.g. DevicePointer() - updated the code that uses HostDeviceVector - objective functions now accept const HostDeviceVector<bst_float>& for predictions
- this means no copies are performed if both host and devices access the HostDeviceVector read-only
- updated the lz4 plugin - added ConstDeviceSpan to HostDeviceVector - using device % dh::NVisibleDevices() for the physical device number, e.g. in calls to cudaSetDevice()
- replaced HostDeviceVector<unsigned int> with HostDeviceVector<int>
RAMitchell
force-pushed
the
dmatrix-hdvec
branch
3 times, most recently
from
0fb1ef1 to
32cb6f4
Compare
RAMitchell
force-pushed
the
dmatrix-hdvec
branch
from
32cb6f4 to
2a04551
Compare
|
@canonizer I have rebased this. Can you please review before I merge? It was a complicated process so there are probably mistakes. I have a backup of your PR before the rebase if necessary. |
|
A couple of GPU tests are currently failing because I removed the behaviour of "wrapping around" when a number of devices is specified that is greater than the number of available devices. This was to be consistent with what @trivialfis did in his PR. We need to decide if we are going to "wrap" the device ordinal or throw some error. I'm thinking we should throw an error because if the user selected multiple GPUs they probably did not intend for it to simply run and multiple blocks on the same GPU. We could check for a sufficient available devices inside of the GPUDistribution or GPUSet. |
- added a mock set device handler; when set, it is called instead of cudaSetDevice()
Codecov Report
@@ Coverage Diff @@
## master #3446 +/- ##
===========================================
+ Coverage 50.97% 51.1% +0.12%
Complexity 188 188
===========================================
Files 176 176
Lines 14090 14200 +110
Branches 457 457
===========================================
+ Hits 7183 7257 +74
- Misses 6682 6718 +36
Partials 225 225
Continue to review full report at Codecov.
|
dmlc#3446) * Replaced std::vector with HostDeviceVector in MetaInfo and SparsePage. - added distributions to HostDeviceVector - using HostDeviceVector for labels, weights and base margings in MetaInfo - using HostDeviceVector for offset and data in SparsePage - other necessary refactoring * Added const version of HostDeviceVector API calls. - const versions added to calls that can trigger data transfers, e.g. DevicePointer() - updated the code that uses HostDeviceVector - objective functions now accept const HostDeviceVector<bst_float>& for predictions * Updated src/linear/updater_gpu_coordinate.cu. * Added read-only state for HostDeviceVector sync. - this means no copies are performed if both host and devices access the HostDeviceVector read-only * Fixed linter and test errors. - updated the lz4 plugin - added ConstDeviceSpan to HostDeviceVector - using device % dh::NVisibleDevices() for the physical device number, e.g. in calls to cudaSetDevice() * Fixed explicit template instantiation errors for HostDeviceVector. - replaced HostDeviceVector<unsigned int> with HostDeviceVector<int> * Fixed HostDeviceVector tests that require multiple GPUs. - added a mock set device handler; when set, it is called instead of cudaSetDevice()
