Official Progressive GPU Lightmapper preview

As written above, I don´t have the time to prepare and document an example. My problem.
Did you ever try Bakery on HDRP?
Without being precise, there are some more issues with Bakery and HDRP, Frank indirectly confirmed this.

Please don´t get me wrong, I am the biggest friend of Bakery.
Hopefully I will be able to prepare an example with my problem soon.

 

Yeah I've used bakery for a long time. I have a pretty good idea why you're getting that, but I don't want to guess. It seems to me you can't send the bakery author a few boxes as in your screenshot.

You may be having problems with PLM as well - PLM is far from finished in my view.

 

No problem, you´re welcome to guess here .
(The geometry is a spline with extrusion modifier in 3dsmax, nothing more.)
This particular problem does not appear in PLM, same geometry bakes fine.

 

PLM is too much fast on my RTX 3060ti 8GB... You can't bake high lightmap size like 4096x4096 on the 3060ti 8gb in most cases... SO use 2048x2048 lightmap size with a very high lightmap resolution (100~200)...
About 15x faster than my Ryzen 5 3600XT (1.5 min vs 27 min bake time performance)

 

Sorry, what do you compare, GPU vs CPU or 4096 vs 2048?

 

Both
2048 is limitation for 8GB gpu
GPU is faster than cpu for baking

 

Wrong, I bake 4096x4096 lightmaps on my RTX2080super (8GB).
Just ensure your lightmap resolution in the Scene-Lighting is not too high, 75 in my case doing archviz interieur.
Open taskmanager and keep an eye on the gpu during baking.
When I get >7.5GB Progressive Lightmapper falls back to CPU.

Right!

 

4096 is not useful for small objects... It's useful for large objects baking like terrains... You can use a very high lightmap resolution for small objects instead

 

I typically do not bake higher than 1024, and it used to work fine but recently, 2020.3 onward I started having GPU memory issues again like when GPU lightmapper had first been introduced. and I had only 3GB of vram. It could be there are some GPU memory management issues or whatever changes they have made after 2020.3 they are hogging a lot of GPU memory. The editor has become slower too and shaders take longer to compile.

My project used to bake fine and superfast even at 2K until Unity 2020.2.
I have tried in 2020.3, 2021.1, and 2021.2 because I need the cloud feature. But no luck.

My very complex scene used to bake in 20-25 minutes even at 60+ resolution.
I am running on a 2070 8GB card.

 

Is the lightmapper totally broken on AMD GPUs? I can't bake 4096x4096 with my 6900 XT, and Unity never even tries to use more than 5 of the 16 GB during baking.

 

By broken I mean I can not use it at all, it falls back to the CPU lightmapper every time as if I didn't have enough VRAM. Will submit a bug report later today.

 

I have a simple question
Does the lightmap parameters affect only on the Enlightens Lightmapper or does it also affect GPU PLM?

 

Most of the stuff in the Baked GI section do affect PLM (Antialiasing, pushoff, Baked tag, Limit lightmap count and Backface tolerance).

The rest doesn't do anything for PLM.

 

So lightmap parameters doesn't do anything?

 

For PLM here are the ones that do something:

 

Is there any way to force the GPU Lightmapper to work even if I have less than 4GB VRAM?
I'm aware that in that case, I won't be able to ask for support if I have issues, but I'm still willing to try it.
I have a 2GB GPU at the moment. I just wanted to make sure before spending on a new GPU.

 

Not very well, no. Expect it to crash constantly. Don't bother. 2GB ram barely leaves any left over after the editor, the os and so on. Just wait a bit and invest in a cheap GPU with bags of ram.

Even 8GB will be unstable in some larger scenes. More is better!

 

Thanks for the quick answer!
I asked because, when I used the GPU Lightmapper in Unity 2019 (before the 4GB min requirement increase), it worked flawlessly in my scenes. I currently have an AMD Radeon R7 350 2GB, and I didn't have any issues with the GPU Lightmapper at the time.
But I understand the need for the extra VRAM to avoid issues.

And thanks for the advice about 8GB and larger scenes!
Guess I'll have to go for a 12GB GPU next. That won't be cheap in my country (Brazil)...

 

Didn't the GPU Lightmapper Tilling should help with the VRAM usage? i mean that system are created to reduce the memory usage and requirement by splitting the lightmap (when baking) into a small chunks of texture based on current available VRAM?

 

Yes, but AFAIK currently it still has too many points of failure and stuff it doesn't quite support for it to be really useful.

 

Is it just not possible to bake with HDR skyboxes? I've been trying forever to get rid of the artifacts produced by the sun and backlit clouds in the HDRI:



I read the manual page about the lightmapper which says that you need to increase environment samples beyond 500 for HDR skyboxes, but that seems to do nothing at all. The exposure of the skybox is also as low as it can go without looking too dark as well. The improvement is barely noticeable with increased environment samples, even at 30.000 environment samples which is the maximum that is realistically possible within working hours. 40.000 would already take a whole day on my 6900 XT, and even if that yielded a 30% improvement over 30.000 samples that would still be nowhere near acceptable.

Maybe I could try increasing samples further if the lightmapper wasn't so broken that it only uses a third of the available VRAM on the 6900 XT, but that problem has gone ignored for so long that I guess I can give up now. Might as well just sell it and go back to my GTX 970 since that barely bakes any slower.

 

Sadly, coming back after years and trying out the GPU lightmapper, it is still worse.
It often crashes Unity. I'm using the lastest LTS.

1. Bug: Viewing the Baked Lightmap Scene View during baking caused a crash.
2. Bug: Sometimes, after baking and entering game mode, I get error messages that the main camera was "destroyed" resulting just in a blue game view. To get rid off this bug, I have to restart Unity. I have a static scene with no scripts destroying the main camera ...

Clearly, I'm wondering when GPU lightmapper will finally gets out of "Preview".

 

I'm not sure if this will help you but that seems like crazy amount of samples. Try your regular 1024 value, and switch to Non-Directional if you are in Directional lightmaps.

I've had these white artifacts with GPU PLM absolutely everywhere in my scene with Directional as soon as there is at least 1 Area Light.

Could be something else for you but worth a try (took me many days to figure it out).

PS: Re-created my most as I had originally replied to the wrong post

Cheers

 

Those scattered bright spots you see are called "fireflies" and will happen in every lightmapper and even in offline ray tracers aswell. When using a HDRI sky, make sure to remove any super bright areas in it - like the sun. For the sun, simply use a directional light.
.

 

Any estimations when Progressive GPU will work on Mac? For now I got the error using the commit with lightmap generated on Windows using Progressive GPU. Same I can't generate lightmap using Progressive GPU on mac

 

Does anyone have an educated guess how a mobile RTX 3050ti 4GB (GPU Lightmapper) would do against a Ryzen 6800 (CPU lightmapper)?

Edit: Is a 4GB card even viable at all? I get the impression that this is the bare minimum, but may not work for larger scenes anyways.

 

I tried it with 2021.3.6 and a GTX 680 2GB and it still bails out immediately because it does not have 4GB. Still looks like a hard coded limit? I though I might give it a try with smaller light map sizes, but there really seems to be no way.

 

I think 4GB is still minimum regardless.

 

Sorry for the many questions, but here's another one:
Anyone have experience with using eGPUs for baking? Is the performance loss acceptable or do the bandwidth limitations hit too hard?

 

@KEngelstoft any comment?

 

Here what I see trying to use Progressive GPU on my MacbookPro 2015 (Radeon VGA)


I asked it because Unity manual says:

 

So, I got my hands on a Ryzen 6900HS laptop with the Radeon 680m iGPU and a RTX 3050ti 4GB dGPU.
To my surprise I can indeed use the 680m for baking! That's unexpected since it only reports 512MB of VRAM (or course it does not actually have "VRAM" at all). Performance is about half of the 3050, which is in line with performance in games. But still better than using the CPU.
I used the Backyard asset and got about those times for lightbaking:
CPU: 22:40 at 12 MRays/s
Radeon: 13:28 at 50 MRays/s
RTX: 7:20 at 100 MRays/s

I guess CPU and GPU MRays are somehow different?

 

Just to report about "Failed to find a suitable OpenCL device" problem.

My original project is under Unity 2019.1.10, then updated Unity to 2021.3.6.
Still have the same problem "Failed to find a suitable OpenCL device" when selecting Progressive GPU as Lightmapper.
After that error, I can't select Progressive GPU, and keep showing "Falling back to CPU lightmapper."
CPU: Intel i9-10980XE
GPU: 2x NVIDIA Geforce RTX2080 Ti
OS: Windows

Not the VRAM size problem, it turn out to be that OpenCL really can't get my GPU device.
But I've already install nvidia driver 516.93 by quick installation mode.
By checking Device manager, GPU card property->driver details, and find my OpenCL.dll(in my case it's under C:\Windows\SysWOW64\).
It shows that the OpenCL.dll is not signed, meaning that nvidia driver might skip installing OpenCL.dll if I already have one.
So delete that OpenCL.dll and reinstall nvidia driver 516.93, select custom installation, and check the complete installation box.
After the second time nvidia driver installation, I got a new OpenCL.dll under the same path.
And can generate lighting by Progressive GPU.

 

Hi, I'm currently using Unity 2020.3.37f1. At time of this post is the latest TLS for 2020. I read in the documentation that if there are two GPU in the computer Unity will use the one not used in the Editor for the GPU lightmapper. In my case I've got two 3080 RTX Nvidia cards. The first issue is, in the GPU PLM, the dropdown list box only shows one card.
In the editor log, I've got two.

[00:00:14] Enlighten: Precompute started.

-- Listing OpenCL platforms(s) --

 * OpenCL platform 0

    PROFILE = FULL_PROFILE

    VERSION = OpenCL 3.0 CUDA 11.7.99

    NAME = NVIDIA CUDA

    VENDOR = NVIDIA Corporation

    EXTENSIONS = cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_d3d10_sharing cl_khr_d3d10_sharing cl_nv_d3d11_sharing cl_nv_copy_opts cl_nv_create_buffer cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_device_uuid cl_khr_pci_bus_info cl_khr_external_semaphore cl_khr_external_memory cl_khr_external_semaphore_win32 cl_khr_external_memory_win32

-- Listing OpenCL device(s) --

 * OpenCL platform 0, device 0 

    DEVICE_TYPE = 4

    DEVICE_NAME = NVIDIA GeForce RTX 3080

    DEVICE_VENDOR = NVIDIA Corporation

    DEVICE_VERSION = OpenCL 3.0 CUDA

    DRIVER_VERSION = 516.59

    DEVICE_MAX_COMPUTE_UNITS = 68

    DEVICE_MAX_CLOCK_FREQUENCY = 1710

    CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE = 65536

    CL_DEVICE_HOST_UNIFIED_MEMORY = false

    CL_DEVICE_MAX_MEM_ALLOC_SIZE = 2684223488

    DEVICE_GLOBAL_MEM_SIZE = 10736893952

    DEVICE_EXTENSIONS = cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_d3d10_sharing cl_khr_d3d10_sharing cl_nv_d3d11_sharing cl_nv_copy_opts cl_nv_create_buffer cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_device_uuid cl_khr_pci_bus_info cl_khr_external_semaphore cl_khr_external_memory cl_khr_external_semaphore_win32 cl_khr_external_memory_win32

 * OpenCL platform 0, device 1 

    DEVICE_TYPE = 4

    DEVICE_NAME = NVIDIA GeForce RTX 3080

    DEVICE_VENDOR = NVIDIA Corporation

    DEVICE_VERSION = OpenCL 3.0 CUDA

    DRIVER_VERSION = 516.59

    DEVICE_MAX_COMPUTE_UNITS = 68

    DEVICE_MAX_CLOCK_FREQUENCY = 1800

    CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE = 65536

    CL_DEVICE_HOST_UNIFIED_MEMORY = false

    CL_DEVICE_MAX_MEM_ALLOC_SIZE = 2684223488

    DEVICE_GLOBAL_MEM_SIZE = 10736893952

    DEVICE_EXTENSIONS = cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_d3d10_sharing cl_khr_d3d10_sharing cl_nv_d3d11_sharing cl_nv_copy_opts cl_nv_create_buffer cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_device_uuid cl_khr_pci_bus_info cl_khr_external_semaphore cl_khr_external_memory cl_khr_external_semaphore_win32 cl_khr_external_memory_win32

User defined PlatformAndDeviceIndices 0 0, name:NVIDIA GeForce RTX 3080

Found command line argument OpenCL-PlatformAndDeviceIndices 0 0

-- GPU Progressive lightmapper will use OpenCL device 'NVIDIA GeForce RTX 3080' from 'NVIDIA Corporation'--

   use -OpenCL-PlatformAndDeviceIndices <platformIdx> <deviceIdx> as command line arguments if you want to select a specific adapter for OpenCL.

Also it looks there is a user defined parameter for using device 0 0 in the command line, but I'm not launching Unity with any command line. I also reviews the Unity Hub, and there are no parameters used.

So, can you explain to me what's happening?

Thanks.

 

Can I use HDRP path tracing as a kind of PLM realtime preview? Will the result of path trace be similiar enough to the final result of baking?

 

Okay. I sent you a new road map idea .
I think it's the feature that make Bakery so good even if it has limited hardware support.
Please, please consider it. We'll be stuck with light mapping in the foreseeable future and this single feature alone will help a lot in the light baking workflow, instead of having to do an educated guess and end up wasting time anyway like current workflow.

 

Sorry for delayed answer. That command line was already added. It didn't work. I decided to change to a RTX 3090 for having more memory, but the issue is not fixed.

 

Hello! I am having some serious problems with the GPU lightmapper since around... October I think?
Using 2019.1.8f1, and a 2080Ti. (I have also tested this on an older computer with a 980 and the problem is exactly the same).

Newer NVIDIA drivers just made the GPU lightmapper stop working. I get the "OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_INVALID_KERNEL".
The newest driver that seems to work right now is 517.48. Right after that, the GPU lightmapper apparently broke in all versions of Unity I tried it on.

Yesterday I tried the latest-latest driver, and I got some hope because with it the GPU lightmapper seemed to work in 2019.4 and a 2020 build. The problem is, my game is built on 2019.1.8, and it's impossible for me to update the unity version for this project. I can stay on the older drivers (in fact, the driver Windows puts me in after a DDU is 516.94, so it's even an earlier one) but, if for some reason I had to update... would I be completely out of luck? Is there any hope that a new NVIDIA driver will fix this issue? I've read this thread and some people seemed to fix similar issues by fiddling with .dll files, but none of that has helped me so far.

Thanks in advance!