Further optimisations for gpu_hist. #4283
Conversation
- Fuse final update position functions into a single more efficient kernel - Refactor gpu_hist with a more explicit ellpack matrix representation
|
A quick review of my optimisation work over the last month. My improvements have been:
On a 10M*100 dense input matrix, boosting for 500 iterations, the performance improvement is approximately: In particular multi-GPU scalability seems to have improved considerably. |
| @@ -485,10 +485,10 @@ class LearnerImpl : public Learner { | |||
| this->PerformTreeMethodHeuristic(train); | |||
|
|
|||
| monitor_.Start("PredictRaw"); | |||
| this->PredictRaw(train, &preds_); | |||
| this->PredictRaw(train, &preds_[train]); | |||
There was a problem hiding this comment.
Given that DMatrix pointers are increasingly used as cache indices, the parameter should probably be changed from DMatrix* to shared_ptr<DMatrix> in all those places. We can then use weak_ptr<DMatrix> as the index into the cache.
This can be done in another pull request, however.
src/tree/updater_gpu_hist.cu
Outdated
| GradientSumT* d_node_hist, | ||
| const GradientPair* d_gpair, | ||
| size_t segment_begin, size_t n_elements) { | ||
| extern __shared__ char smem[]; | ||
| GradientSumT* smem_arr = reinterpret_cast<GradientSumT*>(smem); // NOLINT | ||
| for (auto i : dh::BlockStrideRange(0, null_gidx_value)) { | ||
| for (auto i : dh::BlockStrideRange(0, matrix.null_gidx_value)) { |
There was a problem hiding this comment.
Could you add a function like matrix.BinCount(), just to make the code more readable? null_gidx_value can then be used in cases where it means 'no value'
RAMitchell
force-pushed
the
perf
branch
2 times, most recently
from
51269ae
to
8751221
Compare
Fuse final update position functions into a single more efficient kernel
Refactor gpu_hist with a more explicit ellpack matrix representation