Official Progressive GPU Lightmapper preview

@KEngelstoft I'm not a lighting guy but I'm super impressed with the work Unity have done on the Progressive GPU lightmapper. I have a demo scene that was taking 12+ hours to bake on 1 CPU i7-8700 (U201830f2) with Prog CPU lightmapper. Now with U201910b8 and Prog GPU on same workstation (GTX 1070 GPU), it take around 15 minutes. That's around 50X performance improvement. Nice.

I've trying to get a softer AO effect for some models in my scene using the Progressive GPU lightmapper. However, I'm not sure what I should be doing with Max Distance and Indirect Contribution to achieve that. The two images are with and without any AO as can be seen in the settings. I'd like to get somewhere in between (I think).

With AO


Without AO


EDIT: The images seems a little dark after uploading. It is more obvious in the Unity Editor.

Also, I'm sometimes seeing the following error (although not always).

Code (CSharp):

1. [PathTracer] InitializeLightmapData job with hash: 922d943a0bc935b9f9076e697f013e56 failed with exit code 2.

Any ideas why this is occurring and how to "fix" or avoid it? There seems to be a drop in GPU memory usage from 7.6 out of 8GB to 6.4GB when this happens. I'm assuming it is a kernel method in a Compute Shader failing or similar (but that's just a guess on my part).

 

Indirect contribution is the "intensity" of the effect (so if you want something in between, take the value from 0.5 to 0.25).

Distance is how much it "spreads" (well not exactly but that's the end result roughly)

 

Yes, I was going for a very localised effect - large bricks or blocks in walls of a stone building. I achieved the result I was looking for with the smaller values. Thanks for your response.

 

I believe there was still around 3-4 GB of available system memory (16GB total on Win10Pro), however GPU memory was pushing the limits (around 7.5 of 8GB dedicated on a GTX 1070). Is the method allocating system memory or GPU memory when it fails? Another possibility is that there is not a big enough contiguous block of system memory. Maybe after running GPU lightmapper many times (with some failures), memory becomes fragmented and doesn't get cleaned up correctly.

 

You could have told me that 3 days ago. Hey, I was impressed with the "old" performance in 2019.1.0b8. It literally saved me days of work on a demo video.

 

Great work!
That version is not available in the hub yet, is it?
Any idea when it will be available?

 

Hello, which gpu will be great to speed up baking and higher sizes of light map atlas? Radeon VII 16 GB or 2080ti 11 GB? Could you give advice?

 

I hope you do it fast as if you are using other tools like octane a cuda based GPU is far more useful.

 

@KEngelstoft Not sure if this is related to HDRP or not but Terrain lightmap baked using GPU Lightmapper are broken, while using CPU lightmapper it worked just fine
PGLM CPU


PGLM GPU

 

I'll try to create a repro project

Edit:
@KEngelstoft Umm after updating my driver today i cannot reproduce the issue anymore, in legacy pipeline and HDRP

 

This is great news!
Tried to get gpu lightmapping going with my Radeon Vega 56 egpu on macOS but no luck. The lightmapper always selects the internal Radeon 560 and I can't get it to use the egpu through the command line parameter.

This is how I start Unity 2019.1.0f1 (didn't work on 2019.2 alpha either):
./Unity -OpenCL-PlatformAndDeviceIndices 0 3 -projectPath [...]
also tried this:
./Unity -force-device-index 2 -OpenCL-PlatformAndDeviceIndices 0 3 -projectPath [...]

This is the OpenCL part from the Editor Log:

-- Listing OpenCL platforms(s) --
* OpenCL platform 0
PROFILE = FULL_PROFILE
VERSION = OpenCL 1.2 (Feb 22 2019 20:16:07)
NAME = Apple
VENDOR = Apple
EXTENSIONS = cl_APPLE_SetMemObjectDestructor cl_APPLE_ContextLoggingFunctions cl_APPLE_clut cl_APPLE_query_kernel_names cl_APPLE_gl_sharing cl_khr_gl_event
-- Listing OpenCL device(s) --
* OpenCL platform 0, device 0
DEVICE_TYPE = 2
DEVICE_NAME = Intel(R) Core(TM) i7-7820HQ CPU @ 2.90GHz
DEVICE_VENDOR = Intel
DEVICE_VERSION = OpenCL 1.2
DRIVER_VERSION = 1.1
DEVICE_MAX_COMPUTE_UNITS = 8
DEVICE_MAX_CLOCK_FREQUENCY = 2900
CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE = 65536
CL_DEVICE_HOST_UNIFIED_MEMORY = true
CL_DEVICE_MAX_MEM_ALLOC_SIZE = 4294967296
DEVICE_GLOBAL_MEM_SIZE = 17179869184
DEVICE_EXTENSIONS = cl_APPLE_SetMemObjectDestructor cl_APPLE_ContextLoggingFunctions cl_APPLE_clut cl_APPLE_query_kernel_names cl_APPLE_gl_sharing cl_khr_gl_event cl_khr_fp64 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_byte_addressable_store cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_image2d_from_buffer cl_APPLE_fp64_basic_ops cl_APPLE_fixed_alpha_channel_orders cl_APPLE_biased_fixed_point_image_formats cl_APPLE_command_queue_priority
* OpenCL platform 0, device 1
DEVICE_TYPE = 4
DEVICE_NAME = Intel(R) HD Graphics 630
DEVICE_VENDOR = Intel Inc.
DEVICE_VERSION = OpenCL 1.2
DRIVER_VERSION = 1.2(Mar 11 2019 21:25:40)
DEVICE_MAX_COMPUTE_UNITS = 24
DEVICE_MAX_CLOCK_FREQUENCY = 1100
CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE = 65536
CL_DEVICE_HOST_UNIFIED_MEMORY = true
CL_DEVICE_MAX_MEM_ALLOC_SIZE = 402653184
DEVICE_GLOBAL_MEM_SIZE = 1610612736
DEVICE_EXTENSIONS = cl_APPLE_SetMemObjectDestructor cl_APPLE_ContextLoggingFunctions cl_APPLE_clut cl_APPLE_query_kernel_names cl_APPLE_gl_sharing cl_khr_gl_event 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_byte_addressable_store cl_khr_image2d_from_buffer cl_khr_gl_depth_images cl_khr_depth_images cl_khr_3d_image_writes
OpenCL GPU device Intel(R) HD Graphics 630 from Intel Inc. has less than 2 GB of global memory, ignoring device for lightmapping.


(Filename: ./Editor/Src/GI/Progressive/RadeonRays/RadeonRaysContextManager.cpp Line: 196)

* OpenCL platform 0, device 2
DEVICE_TYPE = 4
DEVICE_NAME = AMD Radeon Pro 560 Compute Engine
DEVICE_VENDOR = AMD
DEVICE_VERSION = OpenCL 1.2
DRIVER_VERSION = 1.2 (Mar 11 2019 21:14:43)
DEVICE_MAX_COMPUTE_UNITS = 16
DEVICE_MAX_CLOCK_FREQUENCY = 300
CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE = 65536
CL_DEVICE_HOST_UNIFIED_MEMORY = false
CL_DEVICE_MAX_MEM_ALLOC_SIZE = 1073741824
DEVICE_GLOBAL_MEM_SIZE = 4294967296
DEVICE_EXTENSIONS = cl_APPLE_SetMemObjectDestructor cl_APPLE_ContextLoggingFunctions cl_APPLE_clut cl_APPLE_query_kernel_names cl_APPLE_gl_sharing cl_khr_gl_event 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_byte_addressable_store cl_khr_image2d_from_buffer cl_khr_depth_images cl_APPLE_command_queue_priority cl_APPLE_command_queue_select_compute_units cl_khr_fp64
* OpenCL platform 0, device 3
DEVICE_TYPE = 4
DEVICE_NAME = AMD Radeon RX Vega 56 Compute Engine
DEVICE_VENDOR = AMD
DEVICE_VERSION = OpenCL 1.2
DRIVER_VERSION = 1.2 (Mar 11 2019 21:14:50)
DEVICE_MAX_COMPUTE_UNITS = 56
DEVICE_MAX_CLOCK_FREQUENCY = 800
CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE = 65536
CL_DEVICE_HOST_UNIFIED_MEMORY = false
CL_DEVICE_MAX_MEM_ALLOC_SIZE = 2143289344
DEVICE_GLOBAL_MEM_SIZE = 8573157376
DEVICE_EXTENSIONS = cl_APPLE_SetMemObjectDestructor cl_APPLE_ContextLoggingFunctions cl_APPLE_clut cl_APPLE_query_kernel_names cl_APPLE_gl_sharing cl_khr_gl_event 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_byte_addressable_store cl_khr_image2d_from_buffer cl_khr_depth_images cl_APPLE_command_queue_priority cl_APPLE_command_queue_select_compute_units cl_khr_fp64
-- GPU Progressive lightmapper will use OpenCL device 'AMD Radeon Pro 560 Compute Engine' from 'AMD'--
use -OpenCL-PlatformAndDeviceIndices <platformIdx> <deviceIdx> or -OpenCL-ForceCPU as command line arguments if you want to select a specific adapter for OpenCL

 

Dear Unity Team, let me say thank you for this great tool!
I'm coming from the traditional offline 3d archviz rendering world.
Currently I'm working on a floorplan image. GI render times are totally awsome! But in some corner areas there are not enough light.
I'd like to ask, that is it possible to set more bounces than 4?

 

I don't believe it is. What is the use case for more than 4?

I also have a background in archviz and I generally consider 4 to be "enough".

 

Green areas are well lit, but red ones receives few indirect light:

 

Another example:


The back corner of the room is far from the window and it's covered by a wall. Needs more indirect light.

 

Can you show the Baked lightmap scene view for that last image?

 

Shaded:


Baked:


Photo reference:

 

Just curious are you using legacy renderpipeline or HDRP? because in HDRP you can increase the indirect light intensity. This setting help me a lot tweaking the global lighting

 

You can increase indirect intensity in any pipeline.

 

after it was baked?

 

Ah, no, I meant from lighting panel that you can adjust before or during baking.

(also, I can do and after baking in the legacy pipeline, but I specifically wrote a shader for that )

 

Increasing Indirect Intensity gives unnatural result, previously well lit areas become overexposed.

Indirect intensity 2:


The most ideal would be to use much weaker area light at the window with more (~8-10) indirect bounce.

 

I don't know exactly, but I think progressive lightmapper is a simple path tracer, so it will have these issues with darkness, no matter how many bounces you use. I think Enlighten has a more advanced light transport so should give better results, but obviously not GPU enabled. Maybe progressive has been updated since then, but I don't think so.

 

Is this ready for big scenes yet? I was testing it on the ArcViz 6 Scene. and got this warning:

OpenCL Error. Falling back to CPU lightmapper. Error callback from context: CL_MEM_OBJECT_ALLOCATION_FAILURE error executing CL_COMMAND_WRITE_BUFFER on GeForce GTX TITAN X (Device 0).

Testing it on a simple scene (with even higher settings) worked fine, within the same project.

 

Looking forward to H2 2019, in terms of GPU selection, is it advisable to have a 8GB or 11GB GPU, or is 6GB enough?

Specifically, I'm looking to buy three new computers for a small team later this year, target spec of a 1660ti allows us to do everything, but should I have one dedicated GPU lightmapper machine with eg. 2080ti?

 

Thanks for the responses. I'm working on strategy games where 512 resolution CPU lightmapper has been sufficient to date - I'll get one PC with 8GB of VRAM (probably AMD Navi which helps test game compatibility those devices anyway).

 

Currently running 2019.1 but I’ve experience this on 2018.3. I assume it’s all indirect lighting causing it, but I’m getting a lot of “blotching” on surfaces and lighting looking worse than it does when unbaked. I assume this is something to do with the lack of one of the filtering stages...

We also have some coloured lights in one scene and indirect lighting from them seems to be putting spots on very, very far away walls that you generally wouldn’t expect them to have, and it’s all very spotty and not smoothed out at all.

Do I just need to use much higher sample rates to compensate for some of the filtering or something?

I’ll grab some screenshots later

 

You need to enable one of the Denoisers. In 2019.1 you can only use the Optix Denoiser and only for the CPU Lightmapper. 2019.2 however you get the Open Image Denoiser and the Optix Denoiser that works with the GPU lightmapper.

 

I'm unfortunately not brave enough to use 2019.2 in production just yet. No way to use the Optix denoiser CPU stage with a GPU bake I assume?

Though, I never have this issue with the CPU baker, and Optix was only just introduced?

 

Is there a video about Jesper Mortensen’s Unity GDC 2019 presentation?
A here it is:

https://software.intel.com/en-us/videos/open-source-denoising-for-lightmaps

​

 

Just to add to my previous post, here's a shot of the kind of errors I'm getting. I've never seen these issues on the CPU baker before, so honestly have no idea why I'm getting them here. Unfortunately these scenes sometimes command upwards of 40 hours for a CPU bake when using 16/100/100 direct/indirect/enviro samples which is pretty low in itself. So I really want to get the GPU baker working.

 

Not that I'm aware of. It's actually fine if I kill all indirect lighting as much as possible (i.e. turn Bounces to 0 and Indirect to 8). Then it bakes a fairly decent looking scene. But without that, it looks pretty awful. Of course, then I'm lacking direct lighting.

 

Sure, I'll give it a go when I've got a spare minute. Extracting that part of the scene would prove difficult though, this is a rather big project. I'll see what I can do.