Actually it can partially be done with GL_ARB_shader_image_load_store
extension. However all drivers that support shader_image have
texture barrier too.
The idea is to use a floating texture to accumulate the data and
then do a final postprocessing pass to apply the modulo
v2:
* use bounding box to
* fix vertex corruption issue
* use negative number in shader which allow to use half float (+12
fps@4x)
Basically the code does the alpha multiplication in the shader therefore
the blend unit only does a pure addition. This way the multiplication is
accurate and accurate_blending doesn't requires a costly barrier.
This code also avoid variable duplication to make the code more separated.
Hopefully blending can be done in a separated function
It is preliminary work to support fast color clipping with HDR
v2: fix assertion compilation failure
v3: fix regression in not accurate mode
v3: Cs * As/Af is not an accumulation
Those cases don't need the Cd addition and were already optimized anyway
Fix a regression on GoW2
Do DATE algo selection before blending. This way we can detect bad
interaction.
Regroup all blending/colclip in a single block. Avoid to check abe &&
rt multiple times.
v2: only enable sw blending when abe is true
The updated medium level will run for all sprites. It helps sotc blooming effect and it remains
fast enough to be enabled by default (at least on 3D games)
The new high level will run for all sprites + color clipping
The idea is that sprites are often use for post-processing effect (ofc except 2D games)
Most of the time post-processing supports SW blending with a small speed penality. SW
blending is more accurate so it is better to use it.
Gain: 1% at 4x on SotC (it partially compensates recent additions)
When the color is constant and equal to 128, the MODULATE mode is
equivalent to the DECAL mode. It saves 5 instructions on the FS.
Accurate options do a better jobs. Technically it can still
be useful for old gpu/driver that doesn't support the GL4.5 extension.
On Windows, you can still rely on Dx
On linux, free driver support it (except Intel)
Code is not yet enabled because it requires extensive test
The idea is to replace point by a 1 pixels sprite with the help of
a geometry shader. In 4x, point will be replaced by a 4x4 sprite.
// GL42 interact very badly with sw blending. GL42 uses the primitiveID to find the primitive
// that write the bad alpha value. Sw blending will force the draw to run primitive by primitive
// (therefore primitiveID will be constant to 1)
It might help to fix a bit the color on a couple of games
accurate_fbmask = 1
Code uses GL4.5 extensions. So far it seems the effect is ony used a couple
of time and often in non-overlapping primitive. Speed impact will likely remain small
GS doesn't supports texture shuffle/swizzle so it is emulated in a
complex way.
The idea is to read/write the 32 bits color format as a 16 bit format.
This way, RG (16 lsb bits) or BA (16 msb bits) can be read or written with
square texture that targets pixels 1-8 or pixels 8-16.
However shuffle is limited. For example you can copy the green channel
to either the alpha channel or another green channel.
Note: Partial masking of channel is not yet implemented
V2: improve logging
V3: better support of green channel in shader
V4: improve detection of destination (issue due to rounding)
Gow uses 24 bits buffer, so only color is updated but blending is configured as Cd
so it is a NOP
In this case, we don't lookup the target in the texture cache. It reduces the complexity
to handle depth which can be located at same address as RT
Note: please test DX renderer
Please test it!
GS supports 3 formats for the output:
32 bits: normal case
=> no change
24 bits: like 32 bits but without alpha channel
=> mask alpha channel (ie don't write it anymore)
=> Always uses 1.0f as blending coefficient
16 bits: RGB5A1, emulated by a 32 bits openGL texture. I think it will be more correct to use
a real 16 bits GL texture. Unfortunately it would cost several (slow) target conversions.
Anyway as a current solution
=> apply a mask of 0xF8 on color when SW blending is used (improve Castlevania shadow)
unfortunately normal blending mode still uses the full range of colors!
This commit also corrects a couple of blending factor. 128/255 is equivalent to 1.0f in PS2, whereas GPU uses 1.0f. So the blending factor must be 255/128 instead of 2
Note: disable CRC hack and enable accurate_colclip to see Castlevania shadow ^^
(issue #380).
Note2: SW renderer is darker on Castlevania. I don't know why maybe linked to the 16 bits format poorly emulated
The purpose of the code is to support alpha channel
of RT uses as an index for a palette texure.
I'm afraid that code will likely break pure palette texture. Only used
if paltex is enabled
It fixes missing shadow in Star Ocean 3 (issue #374) in Native resolution
with filter = 0 (no filtering) or = 2 (normal fitering)
Rendering explanation:
The game emulates a stencil buffer with the alpha channel
The alpha channel of the RT can contains a palette texture index (format 4HH)
The idea is to have a gradient of value in the palette (16/32/48/...).
This way you can implement a +16/-16 and even wrap the alpha value every time
you hit the pixel.
Bilinear filtering breaks the rendering because it interpolates between counts
so you doesn't have the exact count
Upscaling breaks the rendering because the RT is reused as an input texture. It means
that we need to scale it down which again create some interpolations.
* Dump context before the increase of s_n
=> aligned with the global call number
* Don't print colclip not supported when it is optimized away
* dettach the input texture when it is useless
=> avoid to show a wrong texture in the debugger
This way it will allow to implement all blendings operartion in FS.
Of course it will be slow, but it would be nice for debug and quickly check
game error rendering.
Currently colclip uses 2 passes to wrap the output of blending unit
However some blending mode are only a plain copy (of 0 or Cs or Cd).
So no overflow of [0:255], no need to wrap it
Note: I saw those cases in GoW.
Much faster for small batch that write the alpha value. Code can
be enabled with accurate_date option.
Here a summary of all DATE possibilities:
1/ no overlap of primitive
=> texture barrier (pro no setup of stencil and single draw)
2/ alpha written
=> small batch => texture barrier (primitive by primitive). Done in N-primitive draw calls.
(based on GL_ARB_texture_barrier)
=> bigger batch => compute the first good primitive, slow but only 2 draw calls.
(based on GL_ARB_shader_image_load_store)
=> Otherwise there is the UserHacks_AlphaStencil but it is a hack!
3/ alpha written
=> full setup of stencil ( 2 draw calls)
No barrier => draw all primitives
Barrier but without overlap => draw all primitives
Barrier with overlap => draw primitive by primitve
It will ease the implementation of accurate blending and why not date too
Group opengl calls into a nice name.
Apitrace shows them in a tree format that support folding. Previously it
was a long flat list (10K-40K of lines by frame)
I align the call number with the internal s_n variable. This way it is
easy to map GSdx dump output with the GL debugger :)
If there is no overlap, it is allowed to directly read from the render target.
On SotC testcase with 6x scaling: 30fps -> 40fps
Note: it requires GL_ARB_texture_barrier extension so be sure to have a recent driver
Note2: it requires a lots of testing too
Open question: in case of complex date (written alpha)
Will it be faster to split the draw call into multiple call with no
primitive overlap
DATE is implemented in 2 ways.
1/ with stencil
2/ purely in FS (sw)
I kept method 1 to reduce the work on method 2. It sucks for performance.
So it would be either 1 or 2.
Note: DATE has a big impact on higher upscaling
Note2: you can disable the 2nd method with this configuration parameter
override_GL_ARB_shader_image_load_store = 0
UserHacks_UnscaleSprite = 1 will unscale flat sprites
UserHacks_UnscaleSprite = 2 will unscale all sprites (don't work well so far)
The idea of the hack is to redo the interpolation of texture coordinate
based on the non-upscaled pixel position.
It avoids various glitches but sprites aren't upscaled anymore (so no
more anti-aliasing, potentially a coefficient can be added).
It is replacement of the previous hack (UserHacks_stretch_sprite). Don't enable both in the same time!
The idea of the hack is to move the sprite to the pixel boundary. It
avoids most of rounding issue. It also rescales verticaly the sprite (avoid horizontal line on ace combat).
I don't like this rescaling maybe we can limit it to only 1 pixels.
On my limited testcase, results are much better with any upscaling factor.
I still have a bad line in Kingdom heart. If you have issue with others
game please provide us a GS dump.
2x upscaling is pixel perfects. Bigger upscaling is better but not yet perfect
Feedbacks are welcomes (note it doesn't solve all upscaling issue, only wrong texture sampling)
For the history:
If you have a texture of [0;16[ texels and draws a primitive [0;16[
The formulae to sample last pixels of texture is
0.5 + (16*s-1)/(16*s) * 16
Native (s==1): 15.5 (good)
2x (s==2): 16 (bad, outside of the texture)
4x (s==4): 16.25 (bad, really outside of the texure))
* Only a single VAO
=> Format is set once
=> Only a single bind at startup
=> GSVertexBufferStateOGL is nearly useless
=> barely faster but better than nothing :)
Gregory Hainaut