Trouble with OpenCL/WebCL Translator samples

[wolfviking0]

Hi, I just rebuild all the sample using Translator + Validator with LLVM 3.4.

I am working on OSX 10.9 with on HD 4000, I have some weird behavior with validator samples.

——————————

—> Wrong Value result only on GPU

This sample 'Convolution' is working with validator on CPU but with GPU all the value returned are 0

GPU : http://wolfviking0.github.io/webcl-translator/build_book/book.html?sample=1&gl=off&title=Convolution&gpu

GPU + VALIDATOR : http://wolfviking0.github.io/webcl-translator/build_book/book.html?sample=1&gl=off&title=Convolution&gpu&validator=on

——————————

—> Wrong Texture result

The sample must show a sinus curve, it’s a basic OpenGLInterop work perfect without Validator, but the result show incorrect curve/line with Validator.

GPU : http://wolfviking0.github.io/webcl-translator/build_book/book.html?sample=5&gl=on&title=OpenGL%20Interop&gpu

GPU + VALIDATOR : http://wolfviking0.github.io/webcl-translator/build_book/book.html?sample=5&gl=on&title=OpenGL%20Interop&gpu&validator=on

——————————

—> Black Texture result

The sample QJulia show a black screen when it use Validator. This sample doesn’t use interop, but both (with and without interop) are not working.

GPU : http://wolfviking0.github.io/webcl-translator/build_osx/osxsample.html?sample=6&gl=on&title=QJulia%20Copying&gpu

GPU + VALIDATOR : http://wolfviking0.github.io/webcl-translator/build_osx/osxsample.html?sample=6&gl=on&title=QJulia%20Copying&gpu&validator=on

——————————

—> Crash browser

The sample Transpose (and some other like Reduce) are giving a crash when it’s use Validator. If I build natively the sample and try to build the kernel generated I have a segfault.

GPU : http://wolfviking0.github.io/webcl-translator/build_osx/osxsample.html?sample=1&gl=off&title=Transpose&gpu

GPU + VALIDATOR : http://wolfviking0.github.io/webcl-translator/build_osx/osxsample.html?sample=1&gl=off&title=Transpose&gpu&validator=on

——————————

—> Build -11 Exception

Some sample like Scan can’t build the kernel generated and return -11.

GPU : http://wolfviking0.github.io/webcl-translator/build_osx/osxsample.html?sample=3&gl=off&title=Scan&gpu

GPU + VALIDATOR : http://wolfviking0.github.io/webcl-translator/build_osx/osxsample.html?sample=3&gl=off&title=Scan&gpu&validator=on

——————————

—> Wrong color texture

Two sample Mandelbulb and Qjulia are working with validator on CPU but return the result with wrong color.

CPU : http://wolfviking0.github.io/webcl-translator/build_davibu/davibu.html?sample=1&gl=on&title=QJulia&1&0&preprocessed_rendering_kernel.cl&512&512

CPU + VALIDATOR : http://wolfviking0.github.io/webcl-translator/build_davibu/davibu.html?sample=1&gl=on&title=QJulia&1&0&preprocessed_rendering_kernel.cl&512&512&validator=on

——————————

—> Froze of the Computer

The same samples are killing my Mac mini when I launch it on GPU. Take at least 5 minutes for successfully force killing webkit.

GPU : http://wolfviking0.github.io/webcl-translator/build_davibu/davibu.html?sample=1&gl=on&title=QJulia&0&1&preprocessed_rendering_kernel.cl&512&512

GPU + VALIDATOR : http://wolfviking0.github.io/webcl-translator/build_davibu/davibu.html?sample=1&gl=on&title=QJulia&0&1&preprocessed_rendering_kernel.cl&512&512&validator=on

How I say at the beginning, I am working with OSX 10.9 Maverick. Some of this trouble seems not present on Mountain Lion. The WebKit version I use it’s my own build, but it’s updated with the last change of the Samsung Webkit-WebCL repository. The WebKit version core I use it’s from the 19 December 2013 (https://github.com/wolfviking0/webcl-webkit-maverick).

Best regards

Tony

[elhigu]

Hi, HD4000 drivers are working really bad in both windows and OSX Mavericks. Mountain Lion didn't have GPU drivers for OSX so there was a bit less problems (Apples CPU drivers are not perfect either).

I would recommend not to use HD4000 neither Iris drivers on OSX, but only CPU for now. Many times HD4000 already fails on parse error with perfectly valid OpenCL.

We have actually disabled using GPU for testing on OSX from our testbench.

[wolfviking0]

@elhigu, If you have Maverick on other computer with better gpu, or want to do some test, you can launch all the samples here http://wolfviking0.github.io/webcl-translator/?debug=on with the debug parameter you can access to all the sample for the validator.

[wolfviking0]

Hi @elhigu, I try the sample on other Maverick with 'AMD Radeon HD 6970M 1024 Mo'. Many samples are now working, you have right HD 4000 have a lot of trouble. I don't have crash and froze any more.

But, I keep have some wrong result and build failed.

The sample DXT Compressor it's returning a result but the image are not good :
GPU : http://wolfviking0.github.io/webcl-translator/build_trans/translator.html?sample=3&gl=on&title=DXT%20Compressor
GPU + VAL : http://wolfviking0.github.io/webcl-translator/build_trans/translator.html?sample=3&gl=on&title=DXT%20Compressor&validator=on

The samples Transpose / Scan / Reduce (int) are all of them returning a result wrong when I use Validator.

The samples Mandelbulb and Qjulia are working with validator but return the result with wrong color.

CPU : http://wolfviking0.github.io/webcl-translator/build_davibu/davibu.html?sample=1&gl=on&title=QJulia&1&0&preprocessed_rendering_kernel.cl&512&512
CPU + VALIDATOR : http://wolfviking0.github.io/webcl-translator/build_davibu/davibu.html?sample=1&gl=on&title=QJulia&1&0&preprocessed_rendering_kernel.cl&512&512&validator=on

I have some other sample who doesn't work, but seems could be a problem during the generation of the validated kernel. One of my kernel it's using union inside struct. Same for the translator is not a funny kernel ;)

Best regards

Tony

[elhigu]

Hi, did these problems started after llvm 3.4 port? I am investigating one bug, which we seems to exist only in llvm 3.4 based version of validator.

[wolfviking0]

The trouble on DXT and QJulia/Mandelbulb are present with 3.2 too, For the other I can't say because my problem was on HD 4000.

If you need more details, ask me.

Tony

[elhigu]

Ah, ok. Probably there is somewhere increment operation used in those OpenCL codes. I just found out about the bug while ago. Currently our clamping macros are causing problems if e.g. index is incremented at the same time where array is accessed. i.e.

for(int i = 0; i < get_local_size();) {
    // manually unrolled 4 times
    out[i++] = niceeffect();
    out[i++] = niceeffect();
    out[i++] = niceeffect();
    out[i++] = niceeffect();
}

In case above memory access with i++ will be wrapped to macro, which actually uses i++ in multiple places and breaks correct functionality.

We are already fixing this, by changing address range clamping macros to be implemented as functions.

If one writes example above as

for(int i = 0; i < get_local_size();) {
    // manually unrolled 4 times
    out[i] = niceeffect(); i++;
    out[i] = niceeffect(); i++;
    out[i] = niceeffect(); i++;
    out[i] = niceeffect(); i++;
}

then it should work fine. I had to do this kind of fix, when I made instrumented version from Apple's NBody force simulation.

EDIT: I opened separate issue about this #74

[elhigu]

#74 was now fixed, maybe it fixed some of these examples too.

[wolfviking0]

Hi @elhigu, I just update the sample of the translator and a lot of sample are working now (include the wrong color).
Only the dxtcompressor and some OSX sample failed.

http://wolfviking0.github.io/webcl-translator/

Thanks

[elhigu]

Ok, great. Now we need to find out if the last non working code is because of driver or validator bug.

[elhigu]

Is dxtcompressor case failing because the instrumented code does not compile when given to driver or what is exactly the problem?

[wolfviking0]

Nope about the DXTCompressor, it just the result was wrong. Theorically must return as Lena dxt version, and when i use the validator the result show Lena but not rigtly.
I need to test with your last version, but on this link you can see the weird result.
https://twitter.com/wolfviking0/status/408870594457772032

[wolfviking0]

Seems the last version return -11 error on my mac mini on HD 4000. But you can test it here : http://wolfviking0.github.io/webcl-translator/build_trans/translator.html?sample=3&gl=on&title=DXT%20Compressor&validator=on