[iOS Metal] Warning / panic / crash can happen when allocating a lot of glb asset instances in default engine setup #7015
Answered
by
bejado
azhong-git
asked this question in
Q&A
-
|
Hi Filament community, In my usecase, I need to load a simple glb asset like a unit cube (with textures) with 5000 instances on iOS w/ Metal. I am seeing warnings / panics / crashes. These can include: HandleAllocator arena full, Backend CommandStream overflow, random EXC_BAD_ACCESS on FEngine::loop (more details below). I expect 5000 not quite heavy and hope these issues are unexpected from engine side. Are these expected? Any solution towards this? (e.g. increase HandleAllocator arena size, metal command buffer size, any way of tracking their use dynamically?) I can repro using hello-gltf and my own app on the latest release 1.40.3. The glb asset (~70KB) is a cube with two JPG textures. The code snippet I am using: size_t numInstances = 5000;
std::vector<filament::gltfio::FilamentInstance *> tempInstances(numInstances,
nullptr);
app.asset = app.assetLoader->createInstancedAsset(
buffer.data(),
static_cast<uint32_t>(size),
tempInstances.data(),
tempInstances.size());
resourceLoader->loadResources(app.asset);
delete resourceLoader;
delete stbDecoder;
for (size_t i = 0; i < numInstances; i++) {
auto * instance = app.asset->getAssetInstances()[i];
scene->addEntities(instance->getEntities(), instance->getEntityCount());
}
|
Beta Was this translation helpful? Give feedback.
Answered by
bejado
Aug 1, 2023
Replies: 1 comment 4 replies
-
|
Are you compiling Filament yourself? If so, what happens if you increase |
Beta Was this translation helpful? Give feedback.
4 replies
-
|
I usually used the Cocoapods for production and unfortunately cannot increase buffer sizes there. In order to repro, I did compile Filament v1.40.3 in debug build and used Here is the glb I am using: I did increase both sizes four times(see diff at the bottom). Although no crash it seems, UI become super slow with warnings on blocking CommandStream which used too much space. There are also some rendering artifacts (not sure if they are related).
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -45,15 +45,15 @@ set(FILAMENT_PER_FRAME_COMMANDS_SIZE_IN_MB "2" CACHE STRING
-set(FILAMENT_MIN_COMMAND_BUFFERS_SIZE_IN_MB "2" CACHE STRING
+set(FILAMENT_MIN_COMMAND_BUFFERS_SIZE_IN_MB "8" CACHE STRING
"Size of the command-stream buffer. As a rule of thumb use the same value as FILAMENT_PER_FRRAME_COMMANDS_SIZE_IN_MB, default 2."
)
-set(FILAMENT_METAL_HANDLE_ARENA_SIZE_IN_MB "8" CACHE STRING
+set(FILAMENT_METAL_HANDLE_ARENA_SIZE_IN_MB "32" CACHE STRING
"Size of the Metal handle arena, default 8." |
Beta Was this translation helpful? Give feedback.
-
|
The solution is to use GPU hardware instancing instead of loading 5000 separate assets. Our gltfio library isn't set up to do this, but you can view a discussion here on hacking it yourself. Alternatively, create the cube geometry yourself instead of using gltfio. |
Beta Was this translation helpful? Give feedback.
Answer selected by
azhong-git
-
|
Thank you for the answer and pointer to the discussion. Hacking gltfio is probably not ideal for me since I relied on asset instance for loading other complex glb assets. I understand creating the cube from buffers + hardware instancing could work. (More context: I did went from creating cubes with textures (w/o hardware instancing) -> loading glb instances, and found glb asset instances much faster than creating individual cubes (w/o hardware instancing)). Would there be future plans on supporting hardware instancing in gltfio library? (I don't see examples of |
Beta Was this translation helpful? Give feedback.
-
|
At this moment, it's something we would like to add but don't have any concrete timeline. |
Beta Was this translation helpful? Give feedback.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The solution is to use GPU hardware instancing instead of loading 5000 separate assets. Our gltfio library isn't set up to do this, but you can view a discussion here on hacking it yourself. Alternatively, create the cube geometry yourself instead of using gltfio.