Wrong CUDA backend results (while interpreter ones are good)

[wdanilo]

Hi!
I've got an example, where I faced a strange problem. I'm reading an BMP image (Word32) and I'm decomposing it to 4 channels (r,g,b and a, each Word8). Then I'm composing it back (to Word32) and the result is different than original image when using CUDA backend and identical when using interpreter.

I want to ask you, If you could look at this example and tell me if it is a bug in Accelerate or there are some not allowed operations in CUDA backend, which are working in Interpreter by accident?

Example: https://github.com/wdanilo/AccelerateHS-tests/tree/master/channels

It bases on Canny accelerate example (to parse cmd arguments) and there is not much code - all example specific code is placed inside the 'src' directory.
There is also a sample image attached.
To run the example write: ./dist/build/test/test small.bmp small2.bmp

The program prints its values to the screen.
When running on CUDA, we get:
Image {_channels = fromList [("rgba",Raw Array (Z :. 3 :. 3) [0,0,0,33153,8454273,8487168,0,0,0])]})
and with interpreter:
Right (Image {_channels = fromList [("rgba",Raw Array (Z :. 3 :. 3) [4278255615,4294902015,4294967040,4278222976,4286578816,4286611456,4278190335,4278255360,4294901760])]}) (which is the correct value)

[wdanilo]

I've simplified the example. Now all code is in Main.hs (between line 49 and 65). The error is still the same.

[tmcdonell]

It doesn't print anything for me... does that mean it is correct?

What GPU are you using?

[tmcdonell]

Actually, if I add print $ ParseArgs.run backend myimg2 as the last thing in main, then I get

Array (Z :. 3 :. 3) [4278255615,4294902015,4294967040,4278222976,4286578816,4286611456,4278190335,4278255360,4294901760]

for both the CUDA and interpreter backends.

This might be #58 ?

[wdanilo]

@tmcdonell: I'm sorry - of course the line print $ ParseArgs.run backend myimg2 was missing. Running on CUDA I get wrong results there (you can look at the small2.bmp file - this is my results from CUDA computations)

The issue #58 migh be related to this one. I'm suprised, that you are getting good results, while I'm getting wrong. My card has compute capability of 1.2 - maybe this is related?

[tmcdonell]

I guess that that is it, as I have a compute capability 3.0 card.

I never tracked down the root of the problem with #58. Perhaps texture references for 8- and 16-bit types need to be done specially in CUDA? Maybe it is a CUDA bug? Probably testing to see if 8-bit texture reference work when using plain CUDA is the best first step.

[wdanilo]

@tmcdonell: I do not know plain CUDA well. I'll check what I can do in this topic :)

[tmcdonell]

I think this was just a bug with the (now dead) accelerate-cuda backend; closing.