Opencl datahandling
Closes #15 (closed)
OpenCL kernels are now integrated in the normal create_kernel workflow. Also there exists a created a opencljit.init_globally function that just creates some CL queue/contex if you do not want to give it as a parameter to every kernel.
SerialDatahandling is extended to work with alternative GPU array libraries to PyCuda.
There is now some overlapping code with the _custom_transfer_functions but I suppose they are for certain quantities that have a separate transfer function as oppose to using a whole different backend.
@kuron can you have a look on it? I think the solution is not as elegant as I thought it would be.
pycuda.gpuarray.GPUArrays are not wrapped. So if you use dh.gpuarrays['foo'] you get either a pycuda array or a opencl array. I thought this step would be to drastic for one PR. Using OpenCL should still be a lot easier now.
Merge request reports
Activity
added 1 commit
- 491f84ce - Fix test_kernel to work with quicktest again
added 1 commit
- 0a35ee9b - Fix test_kernel to work with quicktest again
added 1 commit
- 2adf1984 - Extend SerialDataHandling to handle alternative GPU libraries
added 1 commit
- 45d9bf5d - Extend SerialDataHandling to handle alternative GPU libraries
-
dh.synchronization_functionalso still calls intopycuda:./pystencils/pystencils/datahandling/serial_datahandling.py in synchronization_function(self, names, stencil, target, **_) 318 index_dim_shape=values_per_cell, 319 dtype=self.fields[name].dtype.numpy_dtype, --> 320 ghost_layers=gls)) 321 322 if target == 'cpu': ./pystencils/pystencils/gpucuda/periodicity.py in get_periodic_boundary_functor(stencil, domain_size, index_dimensions, index_dim_shape, ghost_layers, thickness, dtype) 35 index_dimensions = index_dimensions 36 for src_slice, dst_slice in src_dst_slice_tuples: ---> 37 kernels.append(create_copy_kernel(domain_size, src_slice, dst_slice, index_dimensions, index_dim_shape, dtype)) 38 39 def functor(pdfs, **_): ./pystencils/pystencils/gpucuda/periodicity.py in create_copy_kernel(domain_size, from_slice, to_slice, index_dimensions, index_dim_shape, dtype) 26 27 ast = create_cuda_kernel(update_eqs, iteration_slice=to_slice, skip_independence_check=True) ---> 28 return make_python_function(ast) 29 30 ./pystencils/pystencils/gpucuda/cudajit.py in make_python_function(kernel_function_node, argument_dict, custom_backend) 27 compiled kernel as Python function 28 """ ---> 29 import pycuda.autoinit # NOQA 30 from pycuda.compiler import SourceModule 31 ModuleNotFoundError: No module named 'pycuda' - Resolved by Michael Kuron
Should I change
create_copy_kernelto return an AST. Might correlate better to the normalcreate_kernelfunctions? Though a breaking change.Edited by Stephan Seitz - Last reply by Stephan Seitz
added 5 commits
- 5ea958c6 - Extend SerialDataHandling to handle alternative GPU libraries
- 3d68eeb5 - Allow creation of indexed kernels with opencl
- b7861977 - Let SerialDataHandling.synchronization_function work with target OpenCL
- da9a4852 - Teach SympyPrinters how to print Type
- 0940124f - Make SerialDataHandling.serialization_function work with OpenCL
Toggle commit listadded 7 commits
- d50a59f6 - Extend SerialDataHandling to handle alternative GPU libraries
- 305ef514 - Allow creation of indexed kernels with opencl
- 8346e686 - Let SerialDataHandling.synchronization_function work with target OpenCL
- 8ac8c1ef - Teach SympyPrinters how to print Type
- 4beb4151 - Make SerialDataHandling.serialization_function work with OpenCL
- d712410d - Add two helper functions to manage global cl ctx
- 79e33121 - Add pystencils.opencl.autoinit
Toggle commit listadded 1 commit
- 2bd8fced - Protect pystencils.opencl.autoinit with try..except
added 1 commit
- b1c4889f - Protect pystencils.opencl.autoinit with try..except
added 5 commits
Toggle commit list- Resolved by Stephan Seitz
The current state of this merge request works well for me, except for the
dh.default_targetthing. I think the target names should be consistent betweencreate_kernelandDataHandling. This could be achieved with--- a/pystencils/datahandling/serial_datahandling.py +++ b/pystencils/datahandling/serial_datahandling.py @@ -57,5 +57,4 @@ class SerialDataHandling(DataHandling): - self.array_handler = None - - if default_target == 'opencl' or opencl_queue: - default_target = 'gpu' - self.array_handler = PyOpenClArrayHandler(opencl_queue) + try: + self.array_handler = PyOpenClArrayHandler(opencl_queue) + except Exception: + self.array_handler = None @@ -102 +101 @@ class SerialDataHandling(DataHandling): - gpu = self.default_target == 'gpu' + gpu = self.default_target == 'gpu' or self.default_target == 'opencl' @@ -227 +226 @@ class SerialDataHandling(DataHandling): - gpu = self.default_target == "gpu" + gpu = self.default_target == "gpu" or self.default_target == 'opencl'
added 1 commit
- 16451a74 - Allow default_target=='opencl' in SerialDataHandling
added 1 commit
- e8523f07 - Allow default_target=='opencl' in SerialDataHandling
@kuron I extended the PR to regard boundary handling. An example that a lot of code checks whether the target of the data handling is 'gpu'. Basic LBM scenarios can run now on OpenCL. lbmpy!13 (merged)
Edited by Stephan Seitz
Resolved by Stephan SeitzI started the full integration test, and there are test failures when using these changes in lbmpy
added 19 commits
-
f9225965...974febd7 - 6 commits from branch
pycodegen:master - 65c8db0a - Improve OpenCL integration via create_kernel
- ff03108b - Extend SerialDataHandling to handle alternative GPU libraries
- 38d88ddf - Allow creation of indexed kernels with opencl
- f1537570 - Let SerialDataHandling.synchronization_function work with target OpenCL
- c5f59fed - Teach SympyPrinters how to print Type
- 45922ae5 - Make SerialDataHandling.serialization_function work with OpenCL
- 3f142a0e - Add two helper functions to manage global cl ctx
- ae004cfe - Add pystencils.opencl.autoinit
- c045b3ad - Protect pystencils.opencl.autoinit with try..except
- 207381b1 - Add pycuda.autoinit to GPU test
- 4418f31d - Allow default_target=='opencl' in SerialDataHandling
- 65d4c6d1 - Add friendly hint when calling OpenCL kernels with PyCUDA/Numpy arrays
- 85484500 - Add weird construction to handle OpenCL boundary handling (don't show
Toggle commit list-
f9225965...974febd7 - 6 commits from branch
-
Is this merge request still under consideration or has it been superseded by https://i10git.cs.fau.de/pycodegen/pystencils/tree/target_dh_refactoring?
-
Yes, the data handling and the kernel targets need a major refactoring. The target_dh_refactoring branch introduces some severe API changes, which will not all be necessary. I plan to something like in this branch, but with less API changes. It's on my list, but will take some time :)
The plan is to
- remove 'backend' attribute of kernel, only have 'target' that can either be 'cpu', 'cuda', 'opencl', 'llvm_cpu', 'llvm_gpu'
- the data handling receives a 'target' upon construction (instead of 'default_target')
- in the branch I renamed all 'gpu' arguments in the data handling to 'accelerator' which is not necessary and would break a lot
- we can also leave a deprecated target called 'gpu' that is a synonym for 'cuda'
Edited by Martin Bauer