Frequently Asked Questions (FAQ)¶
Does it automatically support lazy evaluation (also called: late evaluation, expression templates)?
Yes, Bohrium will lazy evaluate all Python/NumPy operations until it encounters a “Python Read”, such a printing an array or having an if-statement testing the value of an array.
Does it support “views” in the sense that a sub-slice is simply a view on the same array?
Yes, Bohrium supports NumPy views fully thus operating on array slices does not involve data copying.
Does it support generator functions (which only start calculating once the evaluation is forced)? Which ones are supported? Which conditions force evaluations? Presumably reduce operations?
Yes, Bohrium uses a fusion algorithm that fuses (or merges) array operations into the same computation kernel that are then JIT-compiled and executed. However, Bohrium can only fuse operations that have some common sized dimension and no horizontal data conflicts. Typically, reducing a vector to a scalar will force evaluate (but reducing a matrix to a vector will not force an evaluate on it own).
On GPUs, will Bohrium automatically keep all data (i.e. all Bohrium arrays) on the card?
Yes, we only move data back to the host when the data is accessed directly by Python or a Python C-extension.
Does it fully support operations on the complex datatype in Bohrium arrays?
Will it lazily operate even over for-loops effectively unrolling them?
Yes, a for-loop in Python does not force evaluation. However, loops in Python with many iterations will hurt performance, just like it does in regular NumPy or Matlab
Is Bohrium using CUDA on Nvidia Cards or generic OpenCL for any GPU?
Bohrium can use both CUDA and OpenCL.
What is the disadvantage of Bohrium? I wonder why it exists as a separate project. From my point of view it looks like Bohrium is “just reimplementing” NumPy. That’s probably extremely oversimplified, but is there a plan to feed the results of Bohrium into the NumPy project?
The only disadvantage of Bohrium is the extra dependencies e.g. Bohrium need a C99 compiler for JIT-complication. Thus, the idea of incorporating Bohrium into NumPy as an alternative “backend” is very appealing and we hope it could be realized some day.
I get the error: “Failed to open map segment shared object”
This is because
TMPDIRis mounted using the
noexecflag. Bohrium uses
TMPDIRto write JIT-compiled kernels, which must be execuable. Please set
TMPDIRto a location not mounted using
noexec(thanks to Jonas Große Sundrup).