[doc] Document list generation #990
Conversation
docs/internal.rst
Outdated
| - (Tasks ``$0`` and ``$1``) from the list of ``root`` node (``S0``) to list of the ``dense`` nodes (``S1``); | ||
| - (Tasks ``$2`` and ``$3``) from the list of ``dense`` nodes (``S1``) to the list of ``bitmasked`` nodes (``S2``). |
There was a problem hiding this comment.
What does from ... to mean? Can we clarify this?
| @ti.kernel | ||
| def func(): | ||
| for i in x: | ||
| print(i) |
There was a problem hiding this comment.
I guess this will always get not printed since bitmasked is inactivated by default.
There was a problem hiding this comment.
That's true. We are just illustrating list generation here.
docs/internal.rst
Outdated
| Our strategy is to generate lists of SNode elements layer by layer. | ||
| This flattens the data structure leaf elements into a 1D dense array, circumventing the irregularity of | ||
| incomplete trees. Then we can simply invoke a regular **parallel for** over the list. |
There was a problem hiding this comment.
| Our strategy is to generate lists of SNode elements layer by layer. | |
| This flattens the data structure leaf elements into a 1D dense array, circumventing the irregularity of | |
| incomplete trees. Then we can simply invoke a regular **parallel for** over the list. | |
| Our strategy is to generate a list of SNode element indices for each layer. | |
| This flattens the data structure leaf elements into a 1D dense array, circumventing the irregularity of | |
| incomplete trees. Then we can simply invoke a regular **parallel for** over the list. |
Also, is the generated list stored in CPU or GPU memory? If you find the process hard to explain, please use LLVM implementation for example.
There was a problem hiding this comment.
It's on GPU, and these lists are updated in a kernel. There are multiple lists so I think plural form is the correct one. I'd suggest to say lists of active SNode elements layer by layer.
There was a problem hiding this comment.
| Our strategy is to generate lists of SNode elements layer by layer. | |
| This flattens the data structure leaf elements into a 1D dense array, circumventing the irregularity of | |
| incomplete trees. Then we can simply invoke a regular **parallel for** over the list. | |
| Our strategy is to generate several lists of active SNode elements on GPU, layer by layer. | |
| This flattens the data structure leaf elements into a 1D dense array, circumventing the irregularity of | |
| incomplete trees. Then we can simply invoke a regular **parallel for** over the list. |
There was a problem hiding this comment.
I'll be more specific and say generation will happen on the device.
docs/internal.rst
Outdated
| Our strategy is to generate lists of SNode elements layer by layer. | ||
| This flattens the data structure leaf elements into a 1D dense array, circumventing the irregularity of | ||
| incomplete trees. Then we can simply invoke a regular **parallel for** over the list. |
There was a problem hiding this comment.
It's on GPU, and these lists are updated in a kernel. There are multiple lists so I think plural form is the correct one. I'd suggest to say lists of active SNode elements layer by layer.
|
Thanks for the reviews! I'm merging this, so that I can add more details on the internal designs (for #986) following this PR. |
Related issue = #709 #941 #986
[Click here for the format server]