On Thu, 17 Jun 2021 00:38:56 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: > > Fixed method call in glsl shaders I did a full test run on 5 different system, including manual tests on 4 of them: Windows 10 w/ Intel graphics Linux w/ NVIDIA graphics (no manual testing) Linux VM guest running on Windows 10 host Mac w/ discrete graphics Mac w/ integrated graphics No problems detected. All looks good. I think there could be some additional tuning done for point lights, but that could be looked at in a follow-on fix. I reviewed the CSR and it is ready to be Finalized. I finished reviewing the shader code, and left a couple comments on the HLSL shaders. modules/javafx.graphics/src/main/native-prism-d3d/hlsl/psConstants.h line 28: > 26: // see D3DPhongShader.h > 27: > 28: static const int numMaxLights = 3; I know we only support 3 lights, but I note that the vertex shader structures in `vsConstants.h` also have room for 5, and that number was unchanged by this PR. Have you ensured that the mismatch won't cause any problems? Also, with this change to `numMaxLights` from 5 to 3, the register assignments for the light arrays are now off (in a way that won't cause functional problems, just wasted space). I wonder if it's better to revert this change for now and deal with it in a follow-up issue? If not, I recommend to make the change from 5 to 3 more broadly. modules/javafx.graphics/src/main/native-prism-d3d/hlsl/vsConstants.h line 58: > 56: float4x3 mBones[MAX_BONES] : register(c35); > 57: > 58: float4 gReserved240[16] : register(c240); `gReserved240` is now at the wrong location (it should be 245), so if it were ever used it would be a problem. It should be updated to avoid confusion at least. ------------- PR: https://git.openjdk.java.net/jfx/pull/334
