Backend Configuration¶
Usually no direct interaction is required
Keep each system's defaults unless application-level profiling demonstrates that a different Warp block size is beneficial on the target hardware.
User-facing configuration for optional system execution backends.
dataclass
¶
Tune execution of the optional GVS Warp kernels.
Attributes:
| Name | Type | Description |
|---|---|---|
warp_block_dim |
int
|
CUDA threads per persistent dynamics and body-state recurrence block. It must be a multiple of 32 between 32 and 1024. Models with at most 64 active coordinates default to 128 threads; larger models default to 192. |
Validate the immutable launch configuration.
Returns:
| Type | Description |
|---|---|
None
|
None. |
Raises:
| Type | Description |
|---|---|
TypeError
|
If :attr: |
ValueError
|
If :attr: |
dataclass
¶
Tune execution of the optional PlanarPCS and PCS Warp kernels.
Backend parameters are intentionally separate from the backend choice:
backend selects "auto", "jax", or "warp", whereas this
object configures the selected PCS implementation. Keeping the options in
one typed object allows future tuning parameters to be added without
expanding the system constructors with backend-specific keywords.
Attributes:
| Name | Type | Description |
|---|---|---|
warp_block_dim |
int
|
CUDA threads per persistent dynamics block. It must be
a multiple of 32 between 32 and 1024. The recommended defaults are
128 for :class: |
Validate the immutable launch configuration.
Returns:
| Type | Description |
|---|---|
None
|
None. |
Raises:
| Type | Description |
|---|---|
TypeError
|
If :attr: |
ValueError
|
If :attr: |