lp:~florian-rathgeber/ffc/gpu-wrappers

Branch merges

Propose for merging

No branches dependent on this one.

Related source package recipes

No recipes using this branch. (?)

Create packaging recipe

Related snap packages

Related bugs

Related blueprints

1513. By Florian Rathgeber on 2011-01-21

updated DOLFIN GPU includes

1512. By Florian Rathgeber on 2011-01-13

merge (trunk revno 1556)

1511. By Florian Rathgeber <frat@xeon1> on 2010-12-08

merged up to trunk rev 1537

1510. By Florian Rathgeber <florian@thehoff> on 2010-09-30

updated block and grid sizes (optimization)

1509. By Florian Rathgeber <florian@thehoff> on 2010-08-23

bugfix: wrapper code generation return immediately if no (or invalid) -l parameter is given; gpu_utils: set GRID_DIM to 128 for better occupancy

1508. By Florian Rathgeber <florian@thehoff> on 2010-08-10

Fixed bug in indexing GPU data structures (forgot parentheses)

1507. By Florian Rathgeber <florian@thehoff> on 2010-08-07

Updated wrapper code to reflect changes in GPUForm: now initialized only using GPUFunctionSpace(s), GPUMesh extracted from function space 0

1506. By Florian Rathgeber <florian@thehoff> on 2010-08-06

Generate GPU dof tabulation kernels, restructured such that both full CPU and GPU code can be generated at the same time

codegeneration:
  * tabulate_dofs_gpu
  * return list of mesh entities need for dof tabulation (needed to generate calling code)
  * update for changes format

codesnippets: added jacobian_gpu to dict

cpp:
  * added reset_gpu() to reset formatting to intial state
  * override "dofs", "entity index", "offset entities"

formatting:
  * extract function space and form class names from wrapper code
  * format GPU dofmap code: generate kernel for each dofmap, calling code for each function space
  * format GPU integral code: generate kernel and calling code for each form
  * generate CPU dofmap and integral code as usual

quadraturegenerator: override_gpu(), generate tabulate_tensor_gpu code, reset_gpu()

quadraturerepresentation: split compute_integral_ir into additional function that is called a second time with overriden format in case of dolfin-gpu target

representation: removed GPU override

tensorgenerator: override_gpu(), generate tabulate_tensor_gpu code, reset_gpu()

wrappers:
  * return wrapper code only for dolfin case
  * wrapper code + function space + form class names for dolfin-gpu case (function space class names obtained from gpu_utils.wrappers)

1505. By Florian Rathgeber <florian@thehoff> on 2010-08-06

Updated gpu_utils to latest FFC version:
templates
  * moved all templates to gpu_utils/__init__.py
  * added templates for dofmap tabulation kernel + calling code
  * split templates into kernel and calling code
wrappers
  * added function space wrapper, function space template
  * generate_dolfin_classes: collect class names of all generated function space classes and the corresponding dofmaps (needed for filling the templates in the code formatting stage)

1504. By Florian Rathgeber <florian@thehoff> on 2010-08-02

* moved call to override_gpu from codegeneration.py:generate_code to representation.py:compute_ir
* added code snippets for ip_coordinates_gpu