Godot 4.4 doesn't allocate shared GPU memory correctly

[NovaXP]

Tested versions

• Reproducible in Godot 4.4 Dev 1, Godot 4.4 Beta 1 (haven't tested versions in between)
• Not reproducible in Godot 4.3 Stable

System information

Godot v4.4.beta1 - Windows 10 (build 19045) - Multi-window, 1 monitor - Vulkan (Forward+) - dedicated AMD Radeon R9 M200X Series (Advanced Micro Devices, Inc.; 27.20.20913.2000) - Intel(R) Core(TM) i7-4720HQ CPU @ 2.60GHz (8 threads)

Issue description

While testing the Godot Benchmarks project using Godot 4.4 Beta 1 on some older computer setups, I found that the benchmark would always crash at the "Viewport > Activation > Activate 1024 Viewports" part on a system with only 1 GB of VRAM. When the console is open, the output reads:

Running benchmark 1 of 1: viewport/activation/activate_1024_viewports
ERROR: Can't allocate memory for image, error: -2.
   at: (drivers/vulkan/rendering_device_driver_vulkan.cpp:1812)
ERROR: Condition "!texture.driver_id" is true. Returning: RID()
   at: texture_create (servers/rendering/rendering_device.cpp:1020)
ERROR: Condition "texture.rd_texture.is_null()" is true.
   at: texture_2d_initialize (servers/rendering/renderer_rd/storage_rd/texture_storage.cpp:820)
ERROR: Attempting to use an uninitialized RID
   at: (./core/templates/rid_owner.h:259)

This same issue occurs on both Windows 10 and Ubuntu 22.04 LTS, and none of the other tests fail, only Activate 1024 Viewports. When converting the benchmark project to 4.3 Stable, the test is able to run successfully.

I tried reproducing the issue on a separate computer with 2 GB of VRAM, and while the issue didn't occur, I found in Task Manager that the Activate 1024 Viewports test would use around 0.75 GB of VRAM in 4.3 Stable, but around 1.1 GB in 4.4 Dev 1 and 4.4 Beta 1 (I haven't tested any of the versions in between yet). In the images below, the green line represents the test being run in 4.3 Stable, and the red line represents the same test being run in 4.4 Beta 1.

(I understand that old hardware can't be supported forever, so if this issue is a side effect of a new feature that will benefit most users then that's okay. I just wanted to bring attention to it in case it's not a known or intended change.)

Steps to reproduce

Run the Godot Benchmarks project in Godot 4.4 Beta 1, select and run only the following tests and note how much VRAM was used:

• Activate 1024 Viewports
• Activate 256 Viewports
• Activate 64 Viewports

Then, repeat the same process using Godot 4.3 Stable and compare the VRAM usage to 4.4 Beta 1.

Minimal reproduction project (MRP)

https://github.com/godotengine/godot-benchmarks

[JekSun97]

I noticed this too, maybe there are big memory leaks, because after the last tests in 4.4 beta1, my computer started to slow down a lot.

[Calinou]

I can't reproduce this on 4.4.beta1.

PC specifications

• CPU: Intel Core i9-13900K
• GPU: NVIDIA GeForce RTX 4090
• RAM: 64 GB (2×32 GB DDR5-5800 C30)
• SSD: Solidigm P44 Pro 2 TB
• OS: Linux (Fedora 41)

The benchmark completes very fast, so even watch -n0.1 "free -m && nvidia-smi" has trouble keeping up (and watch doesn't allow values below 0.1). I did try running 10 instances of the benchmarks simultaneously to increase VRAM usage, but regardless of this, 4.3.stable and 4.4.beta1 will peak at similar memory utilizations (within margin of error).

[NovaXP]

I'm currently testing this on as many other devices as I can and will report back with my findings.

[NovaXP]

I did some tests using Libre Hardware Monitor to get actual numbers instead of just Task Manager screenshots, and initially I found that the amount of VRAM being used between 4.3 and 4.4 were very similar, enough that'd I'd consider it margin of error.

However, something that clearly varied between the results of 4.3 and 4.4 was the amount of shared GPU memory being used, where 4.3 was using much more than 4.4.

Shared GPU Memory Used (averages from multiple tests):

Radeon R9 M265X (2 GB):
4.3 - 286.96 MB
4.4 - 51.26 MB

GTX 1060 (6 GB):
4.3 - 374.43 MB
4.4 - 122.13 MB

RTX 4070 (12 GB):
4.3 - 439.68 MB
4.4 - 179.77 MB

Ryzen 7 4700G (512 MB):
4.3 - 1598.55 MB
4.4 - 1242.95 MB

GTX 650 Ti (1 GB):
4.3 - 1083.75 MB
4.4 - All tests crashed with "ERROR: Can't allocate memory for image, error: -2"

The GTX 1060 and 4700G were on Windows 11, the rest were on Windows 10. I did all the tests one after another and made sure there weren't any unnecessary background programs running for any of them.

I'm afraid I'm not familiar enough with the inner workings of the engine to know what might be causing this, but I can provide screenshots of the data I got from any of the tests, and can try to scrounge up from other systems if it'd help.

[clayjohn]

I wonder if this makes a difference #101972

We were requesting the wrong memory type for certain buffers, I wonder if that triggered GPU drivers to allocate memory inefficiently.

[darksylinc]

I wonder if this makes a difference #101972

Fix #101972 could strongly be related; since we were wrongly allocating almost all memory into VRAM, whereas some of that memory was supposed to be allocated in normal RAM (which is normally a lot more abundant) and be accessed over PCIe bus.

[NovaXP]

I downloaded the Windows Editor artifact from that pull request on the 650 Ti and M265X computers from my previous test and the results were the same as with 4.4 Beta 1. I'm admittedly a bit new to using GitHub, so if that wasn't the right way to get it then let me know and I'll try again.

[clayjohn]

Since this is coming from dev1, I still suspect #90993

@NovaXP Can you try using D3D12 in both 4.3 and 4.4 dev1 and tell me if there is a difference in memory usage?