Support CPU input for device QuantileDMatrix.
#8136
Conversation
trivialfis
force-pushed
the
ghist-ellpack
branch
from
7794124 to
3bf9af7
Compare
src/data/ellpack_page.cu
Outdated
| auto r_end = d_row_ptr[ridx + 1]; | ||
| size_t rsize = r_end - r_begin; | ||
|
|
||
| if (ifeature >= rsize) { |
There was a problem hiding this comment.
Can't you set the null values here?
There was a problem hiding this comment.
Thanks for the suggestion. Let me try to merge the kernels.
src/data/iterative_dmatrix.h
Outdated
| * - The CPU format and the GPU format are different, the former uses a CSR + CSC for | ||
| * histogram index while the latter uses only Ellpack. This results into a design that | ||
| * we can obtain the GPU format from CPU but not the other way around since we can't | ||
| * recover the CSC from Ellpack. More concretely, if users want to construct a CPU |
There was a problem hiding this comment.
What is the problem with obtaining CSC from ellpack?
There was a problem hiding this comment.
How to get the feature index for each element from ellpack?
There was a problem hiding this comment.
It's encoded in the bin number, which are the values in Ellpack. If the dataset is sparse you would have to look this up with binary search I guess.
There was a problem hiding this comment.
Didn't think of that. Yup, we can recover the feature index by binary searching the cut values. Will update the comment and leave it for a different PR for copying data from GPU to CPU.
There was a problem hiding this comment.
Updated the note. Will work on the other direction of conversion.
- Copy `GHistIndexMatrix` to `Ellpack` when needed.
trivialfis
force-pushed
the
ghist-ellpack
branch
from
afd281f to
43ae4ad
Compare
GHistIndexMatrixtoEllpackwhen needed.