On Fri, 11 Jun 2021 01:30:27 GMT, Nir Lisker <nlis...@openjdk.org> wrote:

>> Added a SpotLight only to the D3D pipeline currently.
>> 
>> ### API discussion points
>> 
>> - [X]  Added `SpotLight` as a subclass of `LightBase`. However, it could 
>> also be a subclass of `PointLight` as it's a point light with direction and 
>> extra factors. I saw that `scenario.effect.light.SpotLight` extends its 
>> respective `PointLight`, but it's not a perfect analogy. In the end, I think 
>> it's a questions of whether `PointLight` will be expanded in a way which 
>> doesn't not suit `SpotLight`, and I tend to think that the answer is no.
>> 
>> - [X] The inner and outer angles are the "diameter angles" as shown 
>> [here](https://docs.microsoft.com/en-us/windows/win32/direct3d9/light-typeshttps://docs.microsoft.com/en-us/windows/win32/direct3d9/light-types).
>>   I, personally, find it more intuitive that these are the "radius angles", 
>> so half these angles, as used in the spotlight factor formula. Do you think 
>> I can change this or do you prefer the current definition of the angles?
>> 
>> - [x] The current implementation uses an ad-hoc direction property (using a 
>> `Point3D`). It crossed my mind that we could use the rotation transforms of 
>> the node to control the direction instead, just like we use the 
>> translation/layout of the node to get the position (there is an internal 
>> Affine3D transform for lights, not sure why `AmbientLight` needs it). 
>> Wouldn't that make more sense? When I rotate the light I would expect to see 
>> a change in direction.
>> 
>> ### Implementation discussion points
>> 
>> - [ ] I've gotten advice from a graphics engineer to treat point lights as 
>> spot lights with a 360 degrees coverage, which simplifies a few places. We 
>> can still try to optimize for a point light by looking at the light 
>> parameters: `falloff = 0` and `outerAngle = 180`. These possible 
>> optimization exist in `ES2PhongShader.java` and `D3DMeshView.cc`, and in the 
>> pixel/fragment shaders in the form of 3 different ways to compute the 
>> spotlight factor (the `computeLightN` methods). We need to check which of 
>> these give the best results.
>> 
>> ## Performance
>> 
>> Testing 3 point lights and comparing this branch with `master` using a 1000 
>> division sphere, 200 meshes, and 5000 meshes.
>> Using an AMD RX 470 4GB GPU.
>> 
>> In this branch, there is a possible CPU optimization for checking the light 
>> type and using precalculated values (in `D3DMeshView.cc` for d3d and 
>> `ES2PhongShader.java` for opengl). On the GPU, I tried 3 ways of computing 
>> the spotlight factor contributions (`computeSpotlightFactor`, 
>> `computeSpotlightFactor2` and `computeSpotlightFactor3`) trying out 
>> different branching and shortcuts.
>> 
>> ### Results
>> The CPU "optimizations" made no difference, which is understandable 
>> considering it will not be the bottleneck. We can remove these if we want to 
>> simplify, though maybe if we allow a large number of lights it could make a 
>> difference (I doubt it). I don't have a strong preference either way.
>> 
>> The sphere 1000 tests always gave max fps (120 on Win and 60 on Ubuntu).
>> 
>> **Win 10**
>> Compared with the `master` branch, this patch shows 5-10 fps drop in the 
>> mesh 200 test and ~5 in the mesh 5000 test. I repeated the tests on several 
>> occasions and got different results in terms of absolute numbers, but the 
>> relative performance difference remained more or less the same. Out of the 3 
>> `computeSpotlightFactor` methods, `computeSpotlightFactor3`, which has no 
>> "optimizations", gives slightly better performance.
>> 
>> **Ubuntu 18**
>> The mesh 200 test always gave 60 fps because it is locked to this fps, so we 
>> can't measure the real GPU performance change.
>> The mesh 5000 test shows 2-6 fps drop from master, with 
>> `computeSpotlightFactor` > `computeSpotlightFactor2`  > 
>> `computeSpotlightFactor3` at in terms of performance (~2 fps difference 
>> each).
>> 
>> **Conclusion**: we can expect a 5 fps drop more or less with 3 point lights. 
>> `computeSpotlightFactor3` on d3d and `computeSpotlightFactor` on opengl gave 
>> the best performances.
>
> Nir Lisker has updated the pull request incrementally with one additional 
> commit since the last revision:
> 
>   Addressed review comments

I tested this on a somewhat older physical Linux box yesterday. It runs better 
than it does on my VirtuaBox VM, but there are still regressions.

When running the old attenuation.LightingSample sample with the SpotLight 
patch, on Linux only the last light is applied. For example, if all three 
lights are selected, only the magenta light is applied. If both red and blue 
lights are selected, only the blue light is applied. If the red light is 
selected by itself then it is applied. Without the patch, all selected lights 
are applied (up to max of 3). The same behavior happens with the new sample 
when using spot lights.

I wonder if this is indicative of using more resources than are available on 
older graphics cards? I can't spot anything obvious, though.

I do note that if I build and force the ES2 pipeline on my Windows 10 system 
(which is not supported, but is another interesting data point), that runs fine.

Here is the output of `java -Dprism.verbose=true`:


Prism pipeline init order: es2 sw
Using Double Precision Marlin Rasterizer
Using dirty region optimizations
Not using texture mask for primitives
Not forcing power of 2 sizes for textures
Using hardware CLAMP_TO_ZERO mode
Opting in for HiDPI pixel scaling
Prism pipeline name = com.sun.prism.es2.ES2Pipeline
Loading ES2 native library ... prism_es2
        succeeded.
GLFactory using com.sun.prism.es2.X11GLFactory
(X) Got class = class com.sun.prism.es2.ES2Pipeline
Initialized prism pipeline: com.sun.prism.es2.ES2Pipeline
Maximum supported texture size: 16384
Maximum texture size clamped to 4096
Non power of two texture support = true
Maximum number of vertex attributes = 16
Maximum number of uniform vertex components = 4096
Maximum number of uniform fragment components = 4096
Maximum number of varying components = 124
Maximum number of texture units usable in a vertex shader = 32
Maximum number of texture units usable in a fragment shader = 32
Graphics Vendor: NVIDIA Corporation
       Renderer: Quadro 600/PCIe/SSE2
        Version: 4.5.0 NVIDIA 384.130
 vsync: true vpipe: true

-------------

PR: https://git.openjdk.java.net/jfx/pull/334

Reply via email to