Re: Question about OpenCL VLOOKUP implementation

[Miklos Vajna <vmiklos@xxxxxxxxxxxxx>]

Hi Dhiraj,

On Tue, Jan 23, 2024 at 05:59:06PM -0600, Dhiraj Holden <dhiraj.holden@xxxxxxxxx> wrote:
> I have a question about the OpenCL implementation of VLOOKUP as given in
> sc/source/core/opencl/op_spreadsheet.cxx. Right now, both the unsorted and
> sorted vlookup both do a linear search to find the right value. I am
> wondering if it would be better to do a binary search for sorted vlookup. I
> could take care of that now that I've wrapped my head around this
> implementation.

Just a general suggestion: make sure that you profile your calculation
cost with callgrind or perf and only optimize this if you see it as a
hotspot.

Also, you might want to consider what the normal (on CPU) calculation
for VLOOKUP does: that is the default, so it typically makes sense to
improve the opencl VLOOKUP if the CPU one already does something
similar, to avoid mismatches and additional complexity.

Regards,

Miklos