Attempting single precision GPU computation throws double precision error about clojurecl HOT 3 CLOSED

Hi. You are indeed doing SP yourself, but you're trying to invoke built-in Neanderthal code written with OpenCL that tries to generate double and single precision variants of certain code.

We need to identify why that feature gets invoked in your program. If you look here (https://github.com/uncomplicate/neanderthal/blob/d7f683cab829a401420e03b52869d953341377f0/src/clojure/uncomplicate/neanderthal/internal/device/clblast.clj#L2316) you'll see that both double and single precision engines are created lazily, only when you explicitly create them. (That's dealt in https://github.com/uncomplicate/neanderthal/blob/master/src/clojure/uncomplicate/neanderthal/opencl.clj). That all seems to give warnings (it would be useful to test why), so it might not be crashing your program.

However, the file https://github.com/uncomplicate/neanderthal/blob/5c76fb83eb43b7aa3307e6e90d5025e41750a00f/src/device/uncomplicate/neanderthal/internal/device/opencl/random.cl contains code for both double and single precision numbers, and this produces errors. I must have overlook that the GPU might be not supporting doubles at all, so I didn't split that into two files.

Long term solution is (I hope) that I split this file into double and single specific parts, and load them in https://github.com/uncomplicate/neanderthal/blob/d7f683cab829a401420e03b52869d953341377f0/src/clojure/uncomplicate/neanderthal/internal/device/clblast.clj#L2312 appropriately.

Short term solution is that you test this approach on your machine:

1. clone neanderthal
2. split random.cl into random-float.cl and random-double.cl
3. in clblast.clj provide separate source files for random instead of the unified random.cl
4. build neanderthal (skip tests because these obviously can't all pass on your machine)
5. see whether your program works, so I can apply this in upstream Neanderthal.
6. (if possible, open new issue for this discussion in Neanderthal, since the problem is not in ClojureCL)

from clojurecl.

That makes a lot of sense, thank you! I was also confused why the warnings seemed to trigger build failures, but there are also errors mixed in that huge text dump.

I'll give that workaround a try and open up a new issue in neanderthal with the results.

from clojurecl.

I've opened uncomplicate/neanderthal#114, and will close this issue out. Let me know in that PR if you'd like a companion issue opened in Neanderthal.

from clojurecl.