Fix dart inplace prediction with GPU input. #6777
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Member
trivialfis
commented
Mar 24, 2021
•
trivialfis
commented
- On C API there's a check for output prediction being device-readable. With dart this doesn't hold.
- Added optimization to avoid memory copying whenever possible. Will post the result later.
* Fix dart inplace predict with data on GPU, which might trigger a fatal check for device access right. * Avoid copying data whenever possible.
trivialfis
force-pushed
the
fix-dart-inplace
branch
from
fb0f2f6 to
29a78c5
Compare
trivialfis
commented
Mar 24, 2021
| const auto& host_data = page.data.ConstHostVector(); | ||
| dh::device_vector<Entry> sorted_entries(host_data.begin() + begin, | ||
| host_data.begin() + end); | ||
| dh::device_vector<Entry> sorted_entries; |
There was a problem hiding this comment.
This is to avoid copying data when input is already on GPU.
| @@ -92,7 +92,10 @@ class HostDeviceVectorImpl { | |||
| } else { | |||
| gpu_access_ = GPUAccess::kWrite; | |||
| SetDevice(); | |||
| thrust::fill(data_d_->begin(), data_d_->end(), v); | |||
| auto s_data = dh::ToSpan(*data_d_); | |||
There was a problem hiding this comment.
Avoid synchronization.
| dh::safe_cuda(cudaMemcpyAsync(entries_d.data().get(), | ||
| data_vec.data() + ent_cnt_begin, | ||
| n_entries * sizeof(Entry), cudaMemcpyDefault)); | ||
| if (row_batch.data.DeviceCanRead()) { |
There was a problem hiding this comment.
Avoid copying data when it's already on GPU.
|
100 rounds on higgs: Before: After: |
hcho3
reviewed
Mar 25, 2021
There was a problem hiding this comment.
Trying to understand what this PR does. Is it correct to summarize it as thus:
- Current behavior: Copy the predicted scores
predts.predictionsto the host before summing them intoh_out_predts - New behavior: Sum the predicted scores
predts.predictionsintopredts.predictions.DeviceSpan(), keeping all the data on the device.
| thrust::fill(data_d_->begin(), data_d_->end(), v); | ||
| auto s_data = dh::ToSpan(*data_d_); | ||
| dh::LaunchN(device_, data_d_->size(), [=]XGBOOST_DEVICE(size_t i) { | ||
| s_data[i] = v; |
There was a problem hiding this comment.
Should we use the bound-checked interface here, given that the size of data_d_ is already clear?
Your summary is correct. |
RAMitchell
approved these changes
Mar 25, 2021
hcho3
approved these changes
Mar 25, 2021
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.