[bug] Update children_offsets & stride info to align as elem_stride #4345
Conversation
Currently we run 2 passes to generate snode tree layout in opengl codegen. 1. One pass from leaf to root so that we can calculate the real size & elem_stride for each snode and sort the children from small to large. 2. Another pass from root to leaf so that we can align as elem_stride. Note we cannot eliminate the first pass due to https://github.com/taichi-dev/taichi/blob/master/taichi/backends/opengl/struct_opengl.cpp#L50. The fix for the alignment issue was simple: we used children_offsets information in the codegen but didn't update it in the second pass. This PR fixes the issue and also reorganizes the code a bit to make it easier to read and understand. ghstack-source-id: 958b5ed2293af69ab1e2a8e7f734a651fab37f36 Pull Request resolved: #4345
|
/rebase |
taichi-gardener
force-pushed
the
gh/ailzhang/58/head
branch
from
bbef0e8
to
f35d8de
Compare
There seems to be a discrepancy between testing.yml and release.yml. For example, my opengl PR taichi-dev#4345 fails at vulkan backend on m1. Also today's nightly failed. So unifying them here to debug...
There seems to be a discrepancy between testing.yml and release.yml. For example, my opengl PR #4345 fails at vulkan backend on m1. Also today's nightly failed. So unifying them here to debug...
|
/rebase |
Currently we run 2 passes to generate snode tree layout in opengl codegen. 1. One pass from leaf to root so that we can calculate the real size & elem_stride for each snode and sort the children from small to large. 2. Another pass from root to leaf so that we can align as elem_stride. Note we cannot eliminate the first pass due to https://github.com/taichi-dev/taichi/blob/master/taichi/backends/opengl/struct_opengl.cpp#L50. The fix for the alignment issue was simple: we used children_offsets information in the codegen but didn't update it in the second pass. This PR fixes the issue and also reorganizes the code a bit to make it easier to read and understand. [ghstack-poisoned]
taichi-gardener
force-pushed
the
gh/ailzhang/58/head
branch
from
f35d8de
to
cf92af5
Compare
|
✔️ Deploy Preview for docsite-preview canceled. 🔨 Explore the source changes: cf92af5 🔍 Inspect the deploy log: https://app.netlify.com/sites/docsite-preview/deploys/621366933a524f00070f024e |
|
ohhhhh nice finally this passed CI ... btw what I did to get out of the infinite rerun loop was simply delete the |
| if (snode.type == SNodeType::root) { | ||
| snode_map_.at(snode.node_type_name).mem_offset_in_root = 0; | ||
| } | ||
| void OpenglStructCompiler::align_as_elem_stride(const SNode &snode) { |
There was a problem hiding this comment.
picky: is it possible to 1) extract this algo into a function (without the GL-dependent part), and 2) add a unit test for it...
There was a problem hiding this comment.
sure thing! Will send a followup PR tmr or Wed !
Guess it does highlight the need for a containerized Win CI env.. |
And a containerized macos CI.... ;) |
Stack from ghstack:
Currently we run 2 passes to generate snode tree layout in opengl
codegen.
elem_stride for each snode and sort the children from small to large.
Note we cannot eliminate the first pass due to
https://github.com/taichi-dev/taichi/blob/master/taichi/backends/opengl/struct_opengl.cpp#L50.
The fix for the alignment issue was simple: we used children_offsets
information in the codegen but didn't update it in the second pass.
This PR fixes the issue and also reorganizes the code a bit to make it
easier to read and understand.