update memory calculator #5244
update memory calculator #5244
Conversation
BrianMarre
force-pushed
the
topic-updateMemoryCalculatorWithAtomicPhysis
branch
from
83ab2f0 to
702b222
Compare
| row_division = np.full(global_gpu_extent[dim], int(global_cell_extent[dim] / global_gpu_extent[dim]), dtype=np.int_) | ||
| global_domain_division.append(row_division) | ||
|
|
||
| print(f"global domain division: {global_domain_division}") |
There was a problem hiding this comment.
For me, it is not clear what domain division is supposed to be. I think you should name it differently.
There was a problem hiding this comment.
after an offline discussion I followed Pawel's advice and renamed it to grid_distribution to match more closely the PIConGPU terminology.
|
|
||
| print(f"global domain division: {global_domain_division}") | ||
|
|
||
| # get cell extent of each GPU, <extent of the gpu domain>[simulation dimension, multi dimensional gpu index] |
There was a problem hiding this comment.
rephrased the comment to make it clearer.
| raise ValueError("unsupported precision {precision}") | ||
|
|
||
| @staticmethod | ||
| def get_predefined_attribute_dict(simulation_dimension: int, precision: int) -> dict[str, int]: |
There was a problem hiding this comment.
Why is it a static method? simulation_dimension, precision are also instance attributes. So as long as you put in the class, I would make just use these values.
There was a problem hiding this comment.
since these methods might be usable in other contexts I decided to make them independent of the current implementation.
There was a problem hiding this comment.
@BrianMarre ok, but in that case I would just take it out of the class :)
ci: no-compile
ci: no-compile
ci: no-compile
BrianMarre
force-pushed
the
topic-updateMemoryCalculatorWithAtomicPhysis
branch
from
702b222 to
8593bdb
Compare
ci: no-compile
BrianMarre
force-pushed
the
topic-updateMemoryCalculatorWithAtomicPhysis
branch
from
8593bdb to
79ee6ef
Compare
There was a problem hiding this comment.
I have strong opinions about the qualities of my bike's accommodation but there's actually one thing about the counting in bits and bytes that needs reworking. Others are optional. Haven't checked if the details of the memory usage actually match up with what the code does.
| self.value_size * num_fields * local_cells | ||
| + double_buffer_mem | ||
| + self.value_size * num_pml_fields * local_pml_cells | ||
| number_local_cells = int(np.prod(cell_extent + self.super_cell_size * 2 * self.guard_size)) |
There was a problem hiding this comment.
It's not quite obvious what cell_extent refers to. It looks like it means extent_in_number_of_cells while I first read it as extent_of_a_cell. Better name would be great.
|
Forgot: Offline discussion in the meeting: We want to keep the main example minimal and concrete (and not have the "kind of general" computation for varying parameters) and might add the "how to compute the used values" in a separate section. |
done |
This PR,