-
Notifications
You must be signed in to change notification settings - Fork 286
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Another attempt at supporting non-contiguous arrays #172
base: main
Are you sure you want to change the base?
Conversation
…avior of users of get_elwise_module or get_elwise_range_module, to test backwards compatibility and performance. This squashes multiple commits to put all major changes in one place.
…ctual GPUArrays and using ARRAY_i indices (rather than just i). Squashed multiple commits into one to make changes clearer.
Hi SyamGadde: Is it possible to use constant memory with the new Deferred SourceModule? I think any memcpy to a constant memory symbol would happen before the module is called. Is that right? Would this make it impossible to have both support for non-contiguous arrays and support for constant memory in ElementwiseKernel at the same time? I could try to implement a deferred memcpy to support that usecase, but I'm not sure if that's a good idea. The more 'magic' happens behind the scenes, the more unpredictable the performance becomes. On the other hand, maybe people who want low level control should compile their own source modules anyway, and ElementwiseKernel is for convenience already. Curious to know your thoughts. |
By the way, get_texrefs and get_module are not documented yet, so we could decide to remove those altogether if it makes adding noncontiguous support easier. |
Thanks for the comments! To tell you the truth, I don't really understand texrefs and globals yet, I merely understood that there was a need internally for get_texref, and deferring the binding of texrefs was the easiest way to maintain the current "flow" (pre-generated callable objects that took (now) arbitrary array arguments). But I agree about introducing too much magic. Are there use cases for get_global that I could look at to get a better idea how it might be used? If it truly is constant and wouldn't change for any version of the kernel, perhaps a set_global() call could be added, and the value would be stored and applied after compilation? |
I'm a beginner with CUDA. So far as I understand, globals work like this: First you declare some data __constant__ in your kernel source code If I understand correctly, on modern devices the constant cache is not strictly necessary, because the normal caches will already be sufficient for most purposes. You can get a small benefit from the constant cache if your constants are constantly being pushed out of the other caches by lots of memory traffic. (Additionally, the constant cache is only useful if every thread in the warp accesses the same value) If we get texture objects into PyCUDA, I don't think it would be really necessary to have get_texref and get_global on an ElementwiseKernel, since you could use texture objects for textures and normal global memory instead of constant memory. People who need lower level control could use normal SourceModules. If removing those two functions makes implementing noncontiguous support simpler, I'd say we should go with that. Does it? Was get_texref the only thing requiring deferred sourcemodule? |
No -- the main reason for DeferredSourceModule was to defer compilation until the time when you call the module/kernel with actual array arguments, and then choose between a generic kernel (contiguous) or a custom kernel (non-contiguous). Supporting the deferred binding of texrefs added some complexity but is not the majority of deferred.py. I don't know enough about the costs/benefits of Textures in some of the elementwise-kernels to advise on how to remove them. My goal was to disturb Andreas' code as little as possible and to remain backwards-compatible, at least as a first stab. |
Just wanted to report in to say that this hasn't fallen off my radar, but due to tenure-related crunch time at work, I'll have to push this out to mid-May (by when I have to submit my materials). |
…iable name change from "slicer" to "flipper".
- move as much computation as possible out of create_key(), either to __init__() if it can be pre-computed without knowing shape/strides of args, or to create_source() otherwise. - add support for kernel to see the N-D indices via the do_indices keyword argument and the INDEX kernel variable - change all internal kernel variable names to all-caps and precede by underscore, aside from those variables likely to be used by custom kernels, such as indexing variables (i, A_i, etc.), NDIM, INDEX, total_threads (used by curandom) - allow user to explicitly specify the arguments (by index into arguments) that should be traversed elementwise as arrays - allow user to specify order (fortran or C) to traverse elementwise arrays - remove some accidentally committed @Profile decorators
…comes _eval) and documentation. Get rid of evalcont parameter in favor of explicit _set_mod() call.
… 0. This seems to allow some forms of broadcasting, though technically callers should be sending all elementwise arrays with the same shape (and valid strides) anyway.
…args specified by elwise_arg_inds be the same shape, eliminating the need for shape_arg_index.
This supercedes #171 after I polluted my original 'noncontig' branch with some commits. This squashes all the related updates into reasonably partitioned changes.
Issues that may be impacted/fixed by this include:
#15, #66, #121, #145, #151, #154, #162
Here is the original comment from the original commit:
Inspired by:
https://lists.tiker.net/pipermail/pycuda/2016-December/004986.html
https://gitlab.tiker.net/inducer/pycuda/merge_requests/1
I tried a new approach to supporting non-contiguous arrays in PyCUDA (could be ported to PyOpenCL somewhat easily I think). The goals (some elicited by the above discussion and comments in the WIP) were:
get_elwise_module
andget_elwise_range_module
The only way I could think to support all those goals was to delay compilation (and source generation) to call-time, to take advantage of knowledge of input array strides. Contiguous arrays get the kernels that PyCUDA has always given them, non-contiguous arrays get specialized kernels. The nice thing about doing this is that the actual shape and strides can be sent as '#define's to aid compiler optimization (could even help with the contiguous kernels). The tricky thing about doing this is that some functions in the current implementation require the Module/Function before call-time, to get texrefs etc. So I basically implemented a Proxy class for SourceModule, called
DeferredSourceModule
which also defers the generation of the values created byget_function()
,get_texref()
, etc. until call-time.To make this all work, indexing (for non-contiguous arrays at least) for an array
A
needs to beA[A_i]
, rather thanA[i]
. If it detects matching contiguous arrays as inputs, thenX_i
is '#define'd to bei
, so kernels using the old method will still work (as long as the input arrays are contiguous and match in strides). No regexes needed to transform the user-specified kernel fragments, it's all directed by the user. Also, if you want to support non-contiguous arrays, you need to send the actualGPUArray
objects, rather than theirgpudata
members to the call or prepare_ functions.All existing tests succeed. More would probably need to be added if it made sense to integrate this into PyCUDA.
Positive side-effects:
GPUArray.get()
andGPUArray.copy()
now work for arbitrarily sliced/strided arrays.The performance hit for contiguous arrays is around 15% for modest-sized arrays (i.e. the 1000x100 array tested by Keegan in the above discussion) and, looking at profiling output, I think the hit is due to generating the key (in
ElementwiseSourceModule.create_key()
) for hashing cached kernels. This could probably be fixed. Performance for non-contiguous arrays is infinitely better, given that they weren't supported before, but I've seen a 40% slowdown over the contiguous version for theb1[::2,:-1]**2
test Keegan tried, due to the need to calculate indexes at each iteration of the loop. It tries to do this in a smart way, by pre-calculating the per-thread step for each dimension, and only using division/modulo to calculate the starting indices for each thread before the loop.Independently of whether these changes are merged in, I will continue to use and develop them to support some local needs, so comments are welcome. I hope this is useful!