* - create a gpu metrics (internal) registry
- the objective is to separate the cpu and gpu implementations such that they evolve
indepedently. to that end, this approach will:
- preserve the same metrics configuration (from the end user perspective)
- internally delegate the responsibility to the gpu metrics builder when there is a
valid device present
- decouple the gpu metrics builder from the cpu ones to prevent misuse
- move away from including the cuda file from within the cc file and segregate the code
via ifdef's
- move segment sorter to common
- this is the first of a handful of pr's that splits the larger pr #5326
- it moves this facility to common (from ranking objective class), so that it can be
used for metric computation
- it also wraps all the bald device pointers into span.
* - implementation of map ranking algorithm
- also effected necessary suggestions mentioned in the earlier ranking pr's
- made some performance improvements to the ndcg algo as well
* - pairwise ranking objective implementation on gpu
- there are couple of more algorithms (ndcg and map) for which support will be added
as follow-up pr's
- with no label groups defined, get gradient is 90x faster on gpu (120m instance
mortgage dataset)
- it can perform by an order of magnitude faster with ~ 10 groups (and adequate cores
for the cpu implementation)
* Add JSON config to rank obj.
* - do not create device vectors for the entire sparse page while computing histograms...
- while creating the compressed histogram indices, the row vector is created for the entire
sparse page batch. this is needless as we only process chunks at a time based on a slice
of the total gpu memory
- this pr will allocate only as much as required to store the ppropriate row indices and the entries
* - do not dereference row_ptrs once the device_vector has been created to elide host copies of those counts
- instead, grab the entry counts directly from the sparsepage
* - set the appropriate device before freeing device memory...
- pr #4532 added a global memory tracker/logger to keep track of number of (de)allocations
and peak memory usage on a per device basis.
- this pr adds the appropriate check to make sure that the (de)allocation counts and memory usages
makes sense for the device. since verbosity is typically increased on debug/non-retail builds.
* - pre-create cub allocators and reuse them
- create them once and not resize them dynamically. we need to ensure that these allocators
are created and destroyed exactly once so that the appropriate device id's are set
* - training with external memory - part 2 of 2
- when external memory support is enabled, building of histogram indices are
done incrementally for every sparse page
- the entire set of input data is divided across multiple gpu's and the relative
row positions within each device is tracked when building the compressed histogram buffer
- this was tested using a mortgage dataset containing ~ 670m rows before 4xt4's could be
saturated
* - training with external memory part 1 of 2
- this pr focuses on computing the quantiles using multiple gpus on a
dataset that uses the external cache capabilities
- there will a follow-up pr soon after this that will support creation
of histogram indices on large dataset as well
- both of these changes are required to support training with external memory
- the sparse pages in dmatrix are taken in batches and the the cut matrices
are incrementally built
- also snuck in some (perf) changes related to sketches aggregation amongst multiple
features across multiple sparse page batches. instead of aggregating the summary
inside each device and merged later, it is aggregated in-place when the device
is working on different rows but the same feature
* - fix issues with training with external memory on cpu
- use the batch size to determine the correct number of rows in a batch
- use the right number of threads in omp parallalization if the batch size
is less than the default omp max threads (applicable for the last batch)
* - handle scenarios where last batch size is < available number of threads
- augment tests such that we can test all scenarios (batch size <, >, = number of threads)
* make the assignments of HostDeviceVector exception safe.
* storing a dummy GPUDistribution instance in HDV for CPU based code.
* change testxgboost binary location to build directory.
sriramch