Internal error with radix sort example

[acfoltzer]

A gist of the code I'm running, and the result is here.

I can duplicate this with GHCs 7.2.1 and 7.4.1. My device info is as follows:

Prelude Foreign.CUDA.Driver> initialise []
Prelude Foreign.CUDA.Driver> props =<< device 0
DeviceProperties {deviceName = "Quadro 5000", computeCapability = 2.0, totalGlobalMem = 2683502592, totalConstMem = 65536, sharedMemPerBlock = 49152, regsPerBlock = 32768, warpSize = 32, maxThreadsPerBlock = 1024, maxThreadsPerMultiProcessor = 1536, maxBlockSize = (1024,1024,64), maxGridSize = (65535,65535,65535), maxTextureDim1D = 65536, maxTextureDim2D = (65536,65535), maxTextureDim3D = (2048,2048,2048), clockRate = 1026000, multiProcessorCount = 11, memPitch = 2147483647, memBusWidth = 320, memClockRate = 1500000, textureAlignment = 512, computeMode = Default, deviceOverlap = True, concurrentKernels = True, eccEnabled = False, asyncEngineCount = 2, cacheMemL2 = 655360, tccDriverEnabled = False, pciInfo = PCI {busID = 2, deviceID = 0, domainID = 0}, kernelExecTimeoutEnabled = False, integrated = False, canMapHostMemory = True, unifiedAddressing = True}

[tmcdonell]

sadly, "lost device memory reference" was a common problem with the old reference counting system, which has been completely replaced now with finalisers. This is available on my branch, but give me chance to track down any other bugs it may hit in the new code generator...

[mchakravarty]

@tmcdonell Did you try the code in the Gist on your branch? Does it work?

[tmcdonell]

it does, but for large vectors hits a bug in (I think) scan, which causes the program to segfault.

[mchakravarty]

Weren't there some scan fixes recently? Is the problem reported here still happening in 0.12?

[mchakravarty]

The radix sort example in accelerate-examples fails for me with:

radixsort: Failed: Prelude.(!!): index too large

in the recent version of Accelerate.

[tmcdonell]

radix sort in the examples package now works correctly. thanks for reporting!