Add WS3 detailed technique using the "6" slot. Port the ALS haze, lighting, and math parts of the detailed and ultra shaders. The texture lookups are left out.
Shader selection based on terrain quality setting:
- Ultra - ws30-ALS-ultra frag/vert.
- High and Medium - ws30-ALS-detailed frag/vert. The shaders switch code paths based on quality level uniform.
- Low - ws30-ALS frag/vert.
Changelog:
ws30-ALS vertex shaders:
- Ambient colour material colour doesn't seem to be used in the vertex shader, and isn't sent to fragment shaders currently. Regional materials only define non-default ambient colour of vec4(0.0) for water, ocean etc. Otherwise the default value of vec4(0.2, 0.2, 0.2, 1.0) is used.
ws30-ALS.frag:
- Set alpha of color to diffuse_term.a, to be consistent with WS2 implementation.
ws30-ALS-ultra.frag and ws30-ALS-detailed.frag
- World pos is assigned a value to allow noise functions to compile.
- Swatch_size is temporarily set to 2000m instead of the xsize texture dimension to allow noise math to run while landclass search and texture arrays are being looked up. Swatch_size is used to adjust the wavelength of multiple overlay mixing noise wavelengths based on how far the textures are stretched (WiP WS2 feature). There are some noise calculations that could run while the first landclass lookup happens. If this is not enough, the selection of calculated noise wavelengths to add could change based on how far the textures are stretched, instead of changing the wavelengths at calculation time.
- Move photoscenery technique no "4" after technique no "7" without changing the index (in case the low index is needed for photoscenery). This makes viewing a diff with the WS2 effect easier.
Performance: Currently there's only 1 ground texture lookup and landclass transitions for that texture. The ultra shader looks up 5 more textures. Probably transitions for 1 more texture need to be supported, as often a base and overlay texture are mixed contributing heavily to visible colour. The math overhead is mostly present, except for noise math being better hidden than in the eventual version. Some of the texture array lookups in the full version may be hidden by the math - depending on GPU memory handling compared to calculation speed.
Squashed commit of the following:
commit e7c89ffb600d1bf5cee2936b7dbff31089452745
Author: vs <vs2009@mail.com>
Date: Fri Nov 5 23:44:35 2021 +1000
WS30 shaders:
- WS30-ALS and WS30-ALS-ultra fargment and vertex shaders: Fix NDotL being used before it is initialised. Rename varying diffuse_term from WS2 terrain shaders to light_diffuse_comp as the full diffuse term is not calculated until the fragment shader.
- WS30-ALS-ultra.frag: Enable haze and lighting by default.
commit 9b55ad051a8d7f3568dfdcd6890655942999e8d0
Author: vs <vs2009@mail.com>
Date: Wed Nov 3 22:50:29 2021 +1000
WS30 terrain shaders: ws30-ALS-ultra.frag
- For power users looking to profile transitions on different GPUs: The test phase toggles and settings for different transition options are available at the start of ws30-ALS-ultra.frag, along with explanations, and how to get accurate profile results on different GPUs. Just need to change a few numbers, save, and debug menu > configure dev extensions > reload shaders to activate.
- Note: at a minimum, small scale transitions to remove the 'squareness' due to the landclass texture need to be used even on old GPUs.
- All transition options are off by default. ws30-ALS-ultra.frag is active when the terrain quality slider is set to Ultra. Start by turning one of small or large scale transitions.
- Texture mixing for small and large scale transitions don't work together, to reduce texture lookups. Turn one off when using the other.
- Landclasses with contrasting colours make transition issues more visble. The driver control panel texture filtering settings reduces issues with seams at landclass borders.
- ws30-ALS.frag is left untouched in this commit, for comparison and reviewing. ws30-ALS.frag is active when terrain quality slider is set High to Low.
Changes:
- Implement large scale transitions and small scale de-pixelisation of landclasses by searching the landclass texture. Several options and quality levels are available for profiling on different GPUs.
- There are fixes for multiple issues dealing with texture rendering.
Changelog:
1.0 - Small-scale transitions:
-- a) Remove squareness due to landclass texture by growing neighbour landclasses onto others depending on a growth priority. This feature uses 2 landclass texture lookups, 1 noise lookup, and only 1 ground texture lookup per point. For now the growth priority is simply the landclass id number. If this approach is used, growth priority would be a materials parameter. A way to set default growth priority by landclass via an xml file would be needed. There is flickering noise at long ranges due to small scale detail in the noise function used for growing landclasses. Partial derivative are used to turn off the feature, but turning off the feature too quickly still shows some square ness in distant landclasses.
-- b) Reduce squareness due to landclass texture by mixing base textures of neighbouring landclasses. This is not perfect, as perfectly square shapes turn into perfectly square shapes with perfectly smudged edges. This uses 2 landclass texture lookups, and 2-3 ground texture lookups.
-- c) - a) and b) can be run at once. This option will also fade the growths gradually with distance. This uses 2 landclass texture lookups, 1 noise lookup, and 2-3 ground texture lookups.
2.0 - Large scale transitions:
-- a) Implemented by searching the landclass texture. The search pattern is current landlcass for the fragment at the center, and n search points in four directions along the s and t axes forming a cross. The search directions are configurable in the code, adn a minimum of 3 directions are needed. The step size is configurable. The fewer the steps, the larger the bands formed in the transitions. There are 1+4*n landclass lookups per n search points. e.g. 1 search point: 5 lookups, 4 search points:17 lookups, 10 search points: 41 lookups.
-- b) An option to dither the transition bands by adding mixing noise is availble. This breaks up the visual impact of bands.
-- c) There is some functionality to grow neighbours on a large scale. No growth priority is used - bothe neighbours will lose definition. A more advanced implementation can use several material parameters to define the nature of the transition: some transitions are very sharp in nature like with agriculture, other transitions can be very gradual, some transitions are patchy. Materials can have competing parameters, to determine which neighbour grows or has a shorter transition on one side. These parameters will need a way of specifying defaults by landclass.
3.0 - Fixes
-- a) - Fix seams at landclass borders. These are caused by different ground textures being stretched by different amounts. At the border, an incorrect mip-map level is looked up, causing a colour disontinuity in addition to the difference between landclasses. The reason the mip-map LoD is incorrect is because GPUs use 4 neighbouring pixels to figure out how fast texture coordinates change with respect to screenspace x and y (i.e. partial derivaitves), and use that information to pick a mip-map level. At a landlcass border, this calculation is incorrect.
-- b) Fix seams at borders caused by the current detiling function. This is due to the detiling function changing the amount textures are stretched, as well as messing with coordinates.
-- c) The solution to a) and b) is to use textureGrad() to lookup textures. It allows specifying partial derivatives. The partial derivatives for the normal texture coordinates are obtained by built-in funcions, and these need to be multiplied by each stretching factor, including different stretching inside conditionals. All future texture lookups that use custom coordinate scaling/manipulation need to use textureGrad. This fixes various signs of incorrect mip-map lod with distance and view angle (textures look stable/solid).
-- d) Fix ground textures being stretched out of proportion in the detiling function. This is caused by tile dimensions not being unequal and texture stretching.
commit aca259acaccffbaf1cb264be423dda72c106150e
Author: vs <vs2009@mail.com>
Date: Sat Oct 23 22:55:56 2021 +1000
WS30 shaders: varying fix
commit 91b194f6766d432206cb013d99e0a96cd34ff33b
Author: vs <vs2009@mail.com>
Date: Sat Oct 23 22:17:03 2021 +1000
WS30 ultra shader: port the ambient lighting fix.
commit e131ffaca7b2c934474f0e447b5852cb4c7c8d4a
Author: vs <vs2009@mail.com>
Date: Fri Oct 15 22:19:42 2021 +1000
WS30: landclass and texture lookups in fragment shader:
Changelog:
- Testing: Procedural random landclass function, get_random_landclass(). 1). Avoids a texture lookup for profiling. 2). Creates squares of configurable size with controllable random landclass ranges - to test landclass transition algorithms and quickly profile lots of landclasses in view. It's also possible to put pressure on GPU memory even with 1 lookup of ground textures per fragment, by reducing size to <20m and viewing from altitude.
- Add ALS ultra technique as no 5, move existing ALS technique to no. 6. Add WS30-ALS-ultra frag & vert in preparation for eventual ALS ultra port. Change comments/variables to match using a texture array instead of atlas. Sampler name: atlas -> textureArray.
- Testing/optimisation of de-tiling: WS30-ALS.frag: reorder perlin detiling to make noise source easier to switch. WS30-ALS-ultra.frag: Create alternative to perlin noise used for reducing tiling that avoids a texture lookup - to profile and optimise. This is commented out. The performance difference in the final version depends on memory pressure versus arithmetic load of the full shader. It varies by GPU and what's in view. In future, once a full ALS port is done, maybe it'll be more efficient to run a higher quality de-tiling algorithm all the time to avoid branching overhead due to effects like agriculture which need more de-tiling. De-tiling code in shaders should /eventually/ match CPU-side de-tiling of texture masks used for object and vegetation placement. De-tiling code in GPUs might be subject to precision and numerical artifacts, and if so maybe the texture lookup version would work better. The same applies to noise used for rendering WS3 water waves and collision queries. One option is to use exactly the same glsl code, with a C++ header to convert glsl syntax e.g. https://github.com/g-truc/glm .
Stuart Buchanan