Impossible to test API/ABI version from kernel code

[dhdaines]

Hi,

I have some kernel code that I'd like to maintain compatibility with libgpuarray 0.6 and 0.7. You would think that it would work to just put this:

#include <gpuarray/config.h>
#if GPUARRAY_API_VERSION == 2
#include <cluda.h>
#endif 

But that doesn't work! I'm just speculating here but it appears that either NVCC doesn't see the libgpuarray header files, or its preprocessor is broken and can't do #if tests.

Instead I have to test for the new-style DTYPE_INPUT_N / DTYPE_OUTPUT_N macros, which is less than ideal.

Is there any way to fix this? Is it a problem with libgpuarray or CUDA?

[nouiz]

I'll let @abergeron answer your question. But I think using the test of the new style macro is fine. Why do you want to support both? We don't support 0.6 anymore ourself. Out of curiosity, which kernel are you building?

…

On Wed, Jan 17, 2018 at 11:10 AM David Huggins Daines < ***@***.***> wrote: Hi, I have some kernel code that I'd like to maintain compatibility with libgpuarray 0.6 and 0.7. You would *think* that it would work to just put this: #include <gpuarray/config.h> #if GPUARRAY_API_VERSION == 2 #include <cluda.h> #endif But that doesn't work! I'm just speculating here but it appears that either NVCC doesn't see the libgpuarray header files, or its preprocessor is broken and can't do #if tests. Instead I have to test for the new-style DTYPE_INPUT_N / DTYPE_OUTPUT_N macros, which is less than ideal. Is there any way to fix this? Is it a problem with libgpuarray or CUDA? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#565>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AALC-2dnfUze2GiPUUtZWaBi2Xvz2ouFks5tLhuEgaJpZM4RhlfL> .

[abergeron]

The problem is that we use NVRTC for compilation and it doesn't see any headers that we haven't previously registered. As it's not able to find <gpuarray/config.h> that #include will not work.

If you are doing this from Theano, you can check for the theano version since Theano 0.9 == libgpuarray 0.6 and Theano 1.0 == libgpuarray 0.7.

[dhdaines]

Hi, We want to support 0.6.9 because Theano 1.0.1 is slower than 0.9.0 and doesn't support the old backend, which is also faster and uses less GPU memory than the new backend. But ... our custom code (which does a specific kind of max-pooling) was written for the new backend. Oddly enough the new backend also seems to be faster in 0.9.0 and also doesn't throw out of memory errors. Still trying to figure out what is going on here, maybe I will file another bug report once I do. On Wed, Jan 17, 2018 at 11:22 AM, Frédéric Bastien <notifications@github.com

…

wrote: I'll let @abergeron answer your question. But I think using the test of the new style macro is fine. Why do you want to support both? We don't support 0.6 anymore ourself. Out of curiosity, which kernel are you building? On Wed, Jan 17, 2018 at 11:10 AM David Huggins Daines < ***@***.***> wrote: > Hi, > > I have some kernel code that I'd like to maintain compatibility with > libgpuarray 0.6 and 0.7. You would *think* that it would work to just put > this: > > #include <gpuarray/config.h> > #if GPUARRAY_API_VERSION == 2 > #include <cluda.h> > #endif > > > But that doesn't work! I'm just speculating here but it appears that > either NVCC doesn't see the libgpuarray header files, or its preprocessor > is broken and can't do #if tests. > > Instead I have to test for the new-style DTYPE_INPUT_N / DTYPE_OUTPUT_N > macros, which is less than ideal. > > Is there any way to fix this? Is it a problem with libgpuarray or CUDA? > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#565>, or mute the thread > <https://github.com/notifications/unsubscribe-auth/AALC- 2dnfUze2GiPUUtZWaBi2Xvz2ouFks5tLhuEgaJpZM4RhlfL> > . > — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#565 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADK8UK0yAEKaeyYvc0ROz9zJmx4EdS_4ks5tLh5GgaJpZM4RhlfL> .

[nouiz]

It is not normal that 0.9 is faster then 1.0.1. Can you run both version with Theano profiler(use Theano flag profile=True) and give me the output? Fred On Tue, Jan 23, 2018 at 6:05 PM David Huggins Daines < notifications@github.com> wrote:

…

Hi, We want to support 0.6.9 because Theano 1.0.1 is slower than 0.9.0 and doesn't support the old backend, which is also faster and uses less GPU memory than the new backend. But ... our custom code (which does a specific kind of max-pooling) was written for the new backend. Oddly enough the new backend also seems to be faster in 0.9.0 and also doesn't throw out of memory errors. Still trying to figure out what is going on here, maybe I will file another bug report once I do. On Wed, Jan 17, 2018 at 11:22 AM, Frédéric Bastien < ***@***.*** > wrote: > I'll let @abergeron answer your question. But I think using the test of the > new style macro is fine. > > Why do you want to support both? We don't support 0.6 anymore ourself. > > Out of curiosity, which kernel are you building? > > On Wed, Jan 17, 2018 at 11:10 AM David Huggins Daines < > ***@***.***> wrote: > > > Hi, > > > > I have some kernel code that I'd like to maintain compatibility with > > libgpuarray 0.6 and 0.7. You would *think* that it would work to just put > > this: > > > > #include <gpuarray/config.h> > > #if GPUARRAY_API_VERSION == 2 > > #include <cluda.h> > > #endif > > > > > > But that doesn't work! I'm just speculating here but it appears that > > either NVCC doesn't see the libgpuarray header files, or its preprocessor > > is broken and can't do #if tests. > > > > Instead I have to test for the new-style DTYPE_INPUT_N / DTYPE_OUTPUT_N > > macros, which is less than ideal. > > > > Is there any way to fix this? Is it a problem with libgpuarray or CUDA? > > > > — > > You are receiving this because you are subscribed to this thread. > > Reply to this email directly, view it on GitHub > > <#565>, or mute the thread > > <https://github.com/notifications/unsubscribe-auth/AALC- > 2dnfUze2GiPUUtZWaBi2Xvz2ouFks5tLhuEgaJpZM4RhlfL> > > . > > > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#565 (comment) >, > or mute the thread > < https://github.com/notifications/unsubscribe-auth/ADK8UK0yAEKaeyYvc0ROz9zJmx4EdS_4ks5tLh5GgaJpZM4RhlfL > > . > — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#565 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AALC-6d5XBGFs1hbW-Q6BXm5BrCeA_-Dks5tNmUJgaJpZM4RhlfL> .

[nouiz]

no news in a long time, so closing.