The game sets the framebuffer as an input texture. So I did the same for
openGL. Code is protected with a CRC. It is working because the game want to sample
pixels.
For the record, I tested it GTA too, it doesn't work as expected because
the game will resize the framebuffer to a smaller one. So you don't have
the guarantee that pixel will be read before a data write.
Note: it requires at least accurate blending set on basic
Note: I need CRC of all Jak games that suffers of this issue. Thanks you :)
Default copy-constructor is eight 32 bits move
GSRendererOGL::Lines2Sprites code shrinks from 510B to 398B
(loop of the function 296B => 181B). Hopefully it will reduce the cost
to convert line to sprite on the CPU (i.e. when geometry shader is disabled)
It impacts all renderers. It ought to fix issue in GTA radiosity,
Shadows in Jak series. (note shadows will suck in upscaling)
Implementation is really brutal. Expect a massive slow down, but at least we can test the effect
easily.
Normally perf impact will remain reasonable if the game doesn't use a Read-Write effect
Performances number are welcomes (my guess is really awful in HW mode, slow in SW mode).
You can enable it with "UserHacks_AutoFlush = 1"
GS memory is only 4MB but rendering is allowed to be 2048x2048
with 32 bits format (so 16MB). Technically the frame/depth buffer can start
at the end of the GS memory. Let's not waste too much memory.
Fix crash with BASARAX
(game draws a 2048x1664 32 bits area)
The hack only fix the HW renderer but not the SW renderer. However I'm not sure
the issue is from GSdx.
The hack will disable alpha test that used to generate empty draw call.
Manual gives all setup to upload a palette from the host. But nothing forbid to render
directly in the palette buffer. (GS rule nb 1, there is no rule ^^)
Fix Virtua Fighter 2 dark colors
However I'm not sure we can fix HW renderer. Rendering is done on the GPU but palette
handling is done on the CPU... So we need to read back data (ouch, and slow). A quick
test didn't get the expected results. Potentially there are others bugs (aka not gonna
happen on the HW renderer)
Fix motocross mania missing texture. Close#1319
As far as I understand, transfer is initialized in DIR. But the real
write only occured later so the blit buffer could have been overwritten
by a new value.
BLIT 0 13700
TREG 40 40
DIR 0 0
BLIT 0 13f00 <=== the bad guy
Write! ... => 0x3f00 W:1 F:C_32 (DIR 00), dPos(0 0) size(64 64)
v2: set a value in m_tr.m_blit for load state
It creates a regression on game that uses a small temporary target to
upload textures of various sizes. Inital code was done to handle direct
frame write (background, FMV) so big target
Purpose is to control the filtering when final image is displayed on the screen
Could improve the sharpness of the output in some games (ofc, it will be pixelated)
When depth primitive is constant and depth test is greater or equal, we can
execute the depth write after color (depth status will only depends on the initial
value)
New case for RGB_ONLY ate:
If the blending equation uses a fixed alpha or a source alpha. We can postpone the alpha write
in a 2nd pass.
If depth can also be postponed, we can guarantee the order of correctness of the value.
1st pass => do RGB
2nd pass => do Alpha & Depth
It fixed Stuntman letter rendering :) Remaining of the game is still broken :(
At higher resolutions it takes too much time to save a screenshot at the
maximum compression level. So let's allow the user to set the
compression level.
This re-uses the png_compression_level setting. The default compression
level is 1 for speed, but if the user wishes to increase the compression
level (without using an external tool) and doesn't mind if the
screenshot takes more time to save then they can increase the
compression level up to a maximum of 9 (which can take quite a while).
Fixes#1527.
PNG_LIBRARIES adds both libpng and zlib to the command line.
PNG_LIBRARY only adds libpng to the linker command line, and the cmake
documentation also suggests not to use it.
Value seems wrongly rounded and you can't distinguish 0xFFFF from 0xFFFE
Instead check that depth is constant for the draw call and the value from the vertex buffer
Fix recent regression on GTA (and likely various games)
In FB_ONLY mode the alpha test impacts (discard) only the depth value.
If there is no depth buffer, we don't care about depth write. So alpha
test is useless and we can do the draw with a single draw call and no program
switch
Extend GSVector to support float move
Initial code likely used integer move for performance reason. However due to
the nan correction, register is now in float domain.
CID 168626 (#1 of 1): Uninitialized scalar field
uninit_member: Non-static class member m_end_block is not initialized in this constructor nor in any functions that it calls.
Fix rendering issue on letters on Kengo/burnout 3/...
Default algo will execute the alpha test in 2 passes. However due to blending
you can't handle accurately the color.
Fortunately for us, the rendering uses an always pass depth test so you
can execute first all the color rendering (which doesn't depends on the alpha test)
And then the depth part which depends on the alpha test.
* Code was factorized a bit with the help of max_z
* Add an extra optimization if test is ZTST_GEQUAL and min z value is
the biggest value. Z test will always be pass.
Note: due to float rounding (23 bits mantissa vs 24 bits depth) the test
is done against 0xFF_FFFE and not 0xFF_FFFF. It is wrong but GPU will
also use float so impact will be null.
The solution files are unused and for ancient Visual Studio versions -
GSDumpGUI has its own solution file, and bin2cpp is included in the main
solution file.
The property sheets have either fallen out of use or were never used in
the first place.
Fixes a regression introduced by 46ba9aa117,
where the Linux GS replayer would always use the options in inis/GSdx.ini
(or use the default options if that doesn't exist) to replay the dump,
instead of using the GSdx.ini from the specified ini folder.
Update done on f712c5c6d0
Previous code use the size of the draw to compute latest block. I
don't know why I use .x/.y which are the origin offset so the start of the block.
Check the instruction set first in GPUinit, GPUconfigure and GPUtext
to prevent unsupported vector instructions from being executed.
Move the vector initialisation in GPUinit to a separate function - it
avoids a vzeroupper instruction.
vector push_back causes a SIGILL signal on a Nehalem (SSE4.2) QEMU VM
when compiled with GCC 6.1.1.
However, an empty constructor causes illegal instruction exceptions to be
generated on a Windows VM.
So here's an inbetween that looks stupid but works on what I've tested.
Fix#1457 (GTA)
The game uses a depth format for a pure color buffer (cokes do ravage
in gaming industry)
However I'm really afraid that it migth break another effect in other games.
Combine all the different configurations together so the project files
are more generic and maintainable.
Also standardise the layout so all the project files will be similar and
all have the same standard elements (even if empty).
Add 64-bit configurations.
Additional specifics:
spu2-x: FLOAT_SAMPLES preprocessor definition removed since it's unused.
Use a relaxed atomic to read the exit variable in the hot path
Wait that exit is deasserted in the destructor, so we are sure the
thread will "soon" return
Value could range from 1 to 9. Default is 4 and it is potentially the
best option. Feel free to test some values on your system, behavior
might depends on the core number and thread number
Value is exponential so 4 is 2 times more pixels than 3.
Small value increased thread overhead, big value increase wait/sync latency
memory overhead by thead is only 256KB
However it will reduce the probability to block the push thread to nearly 0
I tested a couple of dumps and only manage 4000 element with 1 extrathread.
Add a factor 2 on the VRAM to get the quantity of available memory for the textures.
The driver is allowed to put some textures in RAM. Of course it is bad for performance
but it won't crash.
Due to the 4GB by process limit, I keep a (reasonable) maximum of 3.8GB.
In order to avoid a crash when memory is too low an exception will be risen
with no guarantee on rendering and big performance impact. In this situation
you ought to reduce upscaling/disable large framebuffer.
* Does the first vsync (start counter) after the sleep
* Dump data after the rendering, avoid to count extra destructor,sleep time
* Dump data into a basic csv file (if people want nice graph)
In file included from GSRenderer.cpp:23:0:
GSRenderer.h: In constructor ‘GSRenderer::GSRenderer()’:
GSRenderer.h:58:12: warning: ‘GSRenderer::m_dev’ will be initialized after [-Wreorder]
GSDevice* m_dev;
^
GSRenderer.h:52:13: warning: ‘GSVector2i GSRenderer::m_real_size’ [-Wreorder]
GSVector2i m_real_size;
^
GSRenderer.cpp:32:1: warning: when initialized here [-Wreorder]
GLLoader: cast passed parameters to required type.
GSDeviceOGL: cast variables to required type and silence warnings.
GSRendererOGL: cast variables to required type and silence warnings.
It creates some slowdowns for unknown reason. My best hypothesis is
that stencil will be cleared too which is slow.
Let's keep the code for the future when stencil will be dropped.
Fix#1420
If baseline and display rectangle offsets differ by small values then consider the status of frame memory offsets, prevents blurring on Tenchu: Fatal Shadows, Worms 3D, ProStroke Golf, Vexx
* Improve frame buffer height management on custom resolution. Width seems to be fine with the same size as scaled image output.
* Prevent offset issues on Persona 3 based on the data from merge circuit.
Note: Fixes custom resolution upscaling on ICO 50Hz/60Hz mode when large frame buffer is enabled. previously 60Hz mode only displayed half of the screen and 50Hz mode only worked due to the scissor hack.
* Ignore Frame memory offsets for calculating dimensions value of display rectangle.
* Remove hack which limited scaling size based on the scissor value.
Note: With the following commit, SilentHill 2 now properly outputs the desired resolution by the users on custom resolution. Previously if we set 1024 x 1024 , it'll output a lower height value which was caused by the hack removed on this commit.
It ought to be the same in performance but code will be easier this way
v2: print the sync status
v3: use a performance print so it doesn't spam the console
* Support Mesa Nouveau IR (free driver for Nvidia's GPU)
=> Print intermediate representation + final shader
=> Dump GPR usage
* Move dumped shader in /tmp/GSdx_Shader/<sub_dir>
=> Avoid the landing of 3 thousands of files in $PWD ^^
* Use function instead of macro
Using __declspec(dllexport) causes duplicate export warnings to be
generated when compiling 64-bit builds. Name mangling also occurs on
functions that are exported this way, so it doesn't actually work with
the plugin system, which uses unmangled names.
The module definition file exports the functions without name mangling
and is sufficient on its own.
Increase the performance on the free driver (Nouveau)
Currently the driver validates all UBO when only 1 is updated. It
is clearly a bad idea to put all UBO in a single common headers.
find . -name *.vcxproj -exec sed -i -e '/user.props/d' {} \;
Microsoft recommends against using .user files. From
https://msdn.microsoft.com/en-us/library/669zx6zc.aspx :
"The best practice is to delete the reference to them in Property
Manager to ensure that your projects operate independently of any
per-user, per-computer settings. This is important to ensure correct
behaviour in a SCC (source code control) environment."
If you cannot compile SPU2-X after this commit (since that still relies
on the old DirectX SDK), you'll need to fix your build environment.
Legacy GPU:
Older driver will be broken.
Still supported GPU:
Please upgrade to the latest AMD driver 16.5.2 or 16.5.3 (and prey that future driver will still work)
Potential issue
* miss a default value
* Renderer is DX9 again on windows
* Case issue
* Same default for all code paths
* overload of GetConfig
v2: separate specific OS option
Enable it to ensure correct rendering (FMV)
Disable it to reduce GPU/memory requirement
Option will likely be removed when the perf impact will be reduced.
* Fast accurate DATE is always enabled, it was faster than standard DATE
* The less fast version is always enabled too. It is likely barely used
so perf impact will be small on few game that could hit this path.
Nice rendering has a higher priority
* The "slow" path will depends on the date option.
Note normally it isn't too slow (-10%) if GL_ARB_shader_image_load_store
is supported but AMD crimson is an epic fail.
Fixes games like Time Crisis 2/3, which use two half-width display
rectangles placed side by side in split screen mode by using different
DISPLAY.DX values.
Issue1: Depth buffer is wrongly invalidated only the first page is detected.
Issue2: First page seems to be partially written. Could be a GSdx transfer bug.
Anyway, invalidation only support a page granularity.
So here a quick workaround that will clear depth buffer in case of very small partial write.
Might worth to check regression on nocturne/digital saga
It would requires some texture dynamic width convert shaders.
So as a quick solution, let's add a new CRC hack.
For issue #1362 (granted the CRC is correct)
Faster :) Reduce further the cost of accurate date
The optimization will clear the stencil to 1. So all pixels will have a
single sample that pass both the depth & stencil test. No primitive
overlaps So the destination alpha test can be done directly in the
shader.
Game often uses date to allow a single pixel pass. If this
use case is detected, stencil buffer will be cleared after first pixels
that pass both depth&stencil test.
It seems to reduce the load on the GPU.
Note: with the help of texture barriere, maybe we could implement the algo
with a single pass.
If a color buffer is still attached and is smaller than depth buffer,
the latter won't be fully cleared.
As a faster alternative, use GL4.4 clear texture function. Avoid to fiddle with
framebuffer and pixel tests.
Fix #1362x Ar Tolenico 2 map clip
So just reuse GT hle shader :) Acid stage is now correct. However it might need
some tuning for others stages.
Still look awful with uspcaled resolution (note internal game framebuffer is around 160x128)
Until AMD release the driver with a fix, I can't use 2nd blending source with SSO.
So let's use the first source. Blending/Alpha will be wrong. But it is likely better
than an uninitialized alpha value.
"Enable Hardware Depth" removed from main dialog.
"Disable Depth Emulation" and "Fast Texture Invalidation" added to Hacks
dialog.
And fix lots of whitespace issues.
GSDX: Improvements to the config interface.
- GSDX: Add new logos to dialog
- GSDX: Remove all the extra null renderers
- GSDX: Changes to renderer combobox
- Sort all the renderers in ascending order. (the fact that D3D11 was
above D3D9 really annoyed me >_<)
- Properly display usage of D3D10/D3D11 on the combobox.
- Use highest available version of DX by default.
- GSDX: gray out upscaling hacks at native resolution
- GSDX-PSX: Modifications to the dialog
- Add new logos
- Remove SDL renderer from combobox since it was removed long ago.
GSOffset is already based on a lookup of PSM/BP/BW. Coverage only adds
the size parameters (so only 256 possibilities)
It replaces the hash lookup with a free array access.
Will use integral coordinate to avoid any rescaling.
Bilinear interpolation isn't supported. I don't think it is allowed to
filter a depth texture anyway.
The hypothesis is that game will use a depth (aka Z32/Z24/Z16/Z16S)
format when sampling depth texture as color. Technically one could use
a standard color format but block/pixel order won't be the same.
(otherwise I'm screwed)
=> Hypothesis invalid on GoW. They just do a scrambled rendering...
Lookup info:
* The first searched list is the depth pool as we search a depth
texture.
* 2nd one is the render target pool (if a depth was converted to a
render target already)
To avoid any CPU overhead, the source will be a pointer to the real texture
* Conversion (if float texture) will be done on the fly by the shader (GPU).
* Relative rescaling won't be supported. Texture must be fetched with
integral coordinate
Cache page coverage of texture into a hash map
Test done on Champion of Norrath (paltex + DisablePartialInvalidation)
Profiler:
Self of GSTextureCache::SourceMap::Add 5.39% => 0.23%
Self of GSTextureCache::LookupSource 15.27% => 10.82%
Hard to measure on CoN as it depends on memory transfer. Seem to be 5-10 fps faster.
This reverts commit 53690cf9d0.
Quoting user:
For aliasing, the option allow of reduce a little but always very
visible compared with DX11 even with anisotropic OFF, , furthermore
many textures bug added with option activated (predictable but not see
on DX11 with anisotropic ON).
TL;DR doesn't worth it.
Note: it seem to work on DX because DX uses HW texturing in clamp region
mode (and others invalid case). OpenGL uses SW texturing to ensure accuracy
* keep a reference of program/pipeline created to ease the deletion
* extend a bit the API to support multiple pipeline
Final goal will be to use a pre link pipeline for SW shaders. And uses
the default pipeline for HW shaders.
By default, anisotropic filtering was disabled when textures aren't countinuous.
This hack allows to force it. It can help to reduce aliasing but it would create
unexpected effect on texture boundaries.
Again, someone ought to add the option on Windows too
Someone ought to add the Windows option too (and DisablePartialInvalidation too)
It might break a couple of games but most of them run better with depth enabled.
Fixes an issue with the D3D backends crashing if the configure dialog
is accessed and ok is pressed. The D3Dcompiler dll is freed and a null
pointer is dereferenced.
It might break gsdxgui but GSshutdown really should not be called unless
GSdx is shutting down. GSDumpGUI on Windows provides the same (or
better) functionality.
* Silent Hill 2 doesn't need the CRC hack
* GSRenderer: no need to explicitly set bottom value for r.
* Texture Cache: Removed a check which couldn't possibly enter true
branch.
For some reason some Windows 7 systems (most are unaffected) cannot cope
with LoadLibraryEx and return error code 87 - "The parameter is
incorrect".
Switch to using LoadLibrary instead for any case where Windows 7 is
expected to successfully load the requested dll. Potentially Windows
Vista is also affected.
So let's increase the height. It will increase the memory requirement on some games
v2: try to do it automatically
(not sure it will useful as most game will requires it)
v3: let's back to an hardcoded 1280 size. It generates too much issue
Try to avoid random black screen frame
v2: don't force the preload hack on the frame
It creates a ghost image over FMV
v3: support offset within a frame
The long story:
Game blits FMV far aways of the RT which is actually the input of the RO texture...
Currently GSdx suffers of 2 bugs.
1/ RT is too small
2/ texture isn't properly updated with the rendered value. Texture is invalidated
but it reads back the pixels from the GS memory whereas the correct
value is located on the GPU.
This commit will replace the standard draw by a manual blit. Therefore it avoid
size issue and bad upscaling issue.
v2:
* Use various copy to be more compatible with dx api
* Move all part of the hack info the BlitFMV function
v3: add log message
It often happens the game try to upload the FMV directly which typically
gave a black screen.
Commit fix rules of roses and I hope various black screen FMV
Performance impact must be tested, and I'm afraid of strange texture cache behavior.
V2: check the size of the transfer too
V3: add support of 16 bits format
V4: avoid division by 0
Using D3DX11 requires the end user to install the DirectX redist files.
Switch to using D3DCompile, and distribute D3DCompiler_47.dll for
Windows Vista, 7 and 8 users (Windows 8.1 onwards supplies
D3DCompiler_47.dll with the OS).
It actually removes the previous hack that read the full target.
Unfortunately snowblind engine game uses big target so the read is very big too (1280x448)
which is killer for the perf. Whereas the game requires only 24x12 texels
Give a 2x speed boost on Champion of Norrath !!!
Games uses very special texture with a lots of repeating.
It is much faster to send the full texture rather than trying to partially invalidate it.
On my gs dump:
FPS: 29 => 68 !
Basically I ran
find . -name "*.vcxproj" -exec sed -i -e 's/_xp//' {} \;
This will likely break XP, but it paves the way on Windows for a PCSX2
that does not require the DirectX redistributables to be installed for
Windows 8, 8.1 and 10 users. Windows Vista and 7 users will still require
the DirectX redistributable files for XInput and XAudio, though PCSX2
should still be capable of running if a user does not actually use either
of them.
All GL4 extensions supported by DX10 class GPU will be soon mandatory
Namely:
* GL_ARB_copy_image
* GL_ARB_texture_barrier
* GL_ARB_clip_control
* GL_ARB_direct_state_access
* GL_ARB_separate_shader_objects
* GL_ARB_buffer_storage
There are likely few games (RE4) which constantly change the FBW register value causing the framebuffer width to be updated at every interval. Adding a safe limit (512) similar to frame buffer height would prevent such constant changes of the framebuffer width when FBW changes once again to an even lower value.
Valid values for png_compression_level are from 0 (no compression) to 9
(max compression). The default is 1.
v2: Use zlib Z_BEST_SPEED (1) and Z_BEST_COMPRESSION (9) defines.
The following patch uses the height value of the display rectangle rather than make an estimation of the Frame buffer height when the game uses a non-referenceable height (or) width.
Don't lookup a depth buffer if depth test is always pass without write
Boost performance on Tekken5 when depth emulation is enabled in openGL
(Tekken5 sets same address for both the RT and the depth but depth is disabled)
v2:
Keep ds if DATE is enabled (some implementation uses a stencil buffer)
Be more aggressive to avoid an useless depth lookup
Texture coordinate could be dummy/float/int integral/int normalized.
Old behavior:
* VS was in charge to select the texture coordinate
* int integral format wasn't supported
New behavior:
* Always compute all formats
* FS will be in charge to select the good format
Impact:
* VS will be slightly slower but it reduces shaders permutation from
little to 0 (won't be bad for CPU)
* FS speed isn't impacted as 2 separate code paths were already required
to support both format
* Rasterizer will be 33% slower but unlikely to be the limited factor of
the GPU
* In future we could directly use the integral format in the FS.
V2: remove useless PSin_t
It would be on by default. Unsafe & fast path.
The hack is a security if someone encounters any issue
v2: update Windows gui file
v3: fix typo in tooltip and linux gui
Visual Studio Find and Replace can only be trusted if all the files are
included in the project. I suppose it's time to add any missing files
to the relevant projects...
The old implementation saved the current value of a GSSetting as uint in
a field called 'id'. The implementation of GSSettings suggests that
GSSettings could be saved in a database with id as primary key. This
would require a translation look up from id to value but could have all
advantages of a database. However the interface to GSSetting was never
implemented like that.
In the new implementation GSSetting has a 'value' field that stores an
int representative value of the desired state. Additionally the
constructor is 'overloaded' as template to reduce casting in the
consumer code. However all consumer values need to be castable to int.
Accordingly combobox initialization was adjusted.
Initially it was free to do the SW blending because safe fbmask
will already do a sw blending.
Unsafe version uses a fast path with a limited blending. Therefore
SW blending isn't free anymore.
Improve the speed of the previous speed hack (xenosaga 1)
The hack relies on the undefined behavior of the hardware so it can
potentially generate rendering corruption.
This new hack drops the cache flusing when only the alpha channel is masked.
Alpha is a direct copy of the fragment. Normally masked bits will be constant
everywhere (RT, FS output, texture cache) so it would likely work.
Just in case, code is only enabled with the new shiny hack
The previous behaviour loaded the saved renderer config whenever the
adapter combobox was changed. The renderer will now only change if the
new adapter doesn't support the currently selected renderer (i.e
Direct3D11 might not be supported, so it'll revert to Direct3D 9).
Fixes#1080.
The Wild Arms Offset text was slightly cut off due to the label being
too small. Make the dialog slightly wider so the full text will fit.
Someone should probably make the dialog look nicer at some point.
Avoid a crash on Onimusha3 (PAL 60HZ)
In theory it will be better to find the root cause of overflow. I.e. somewhere in this
code below. Dirty rectangle is too big.
***********************************************************************
if(rowsize > 0 && offset % rowsize == 0)
{
int y = GSLocalMemory::m_psm[psm].pgs.y * offset / rowsize;
if(r.bottom > y)
{
GL_CACHE("TC: Dirty After Target(%s) %d (0x%x)", to_string(type),
t->m_texture ? t->m_texture->GetID() : 0,
t->m_TEX0.TBP0);
// TODO: do not add this rect above too
t->m_dirty.push_back(GSDirtyRect(GSVector4i(r.left, r.top - y, r.right, r.bottom - y), psm));
t->m_TEX0.TBW = bw;
continue;
}
}
***********************************************************************
So as a temporary solution (that will likely stay for a couple of
years), buffers were increased.
Height of the dirty rectangle must be the GS size of the RT. Of course
RT doesn't have any height so we compute the max safest value.
Fix issue #987
Candidate for 1.4 release
Both the Linux and Windows config dialogs now have a TV Shaders combobox,
so the F7 toggle can be made temporary. This makes the hotkey behaviour
consistent with all the other hotkeys.
The old one isn't working. I don't think there's a URL that redirects to
whatever language the user is using (unless my browser settings are
wrong), so I've just used the English US URL.
1. Add GS_Renderer Enum
Replace all instances of int/uint32 renderer identifier by a strongly
typed enum and appropriate casts.
Only instances in GS[*].cpp/h classes were touched. GPU[*].cpp/h classes
do not to follow the same convention.
2. Add default renderer according to OS
The default renderer is OS dependent (Win -> Dx9HW, others -> OGLHW).
Consequently one should always check againt the appropriate default
value on config load.
The old behaviour was only - if a at all - problematic if the respective
element in the gsdx.ini was missing and probably even then didn't create
issues. The current implementation is still more stable and does not
depend on the implementation of GS.cpp -> GetConfig()
The following patch adds Mipmap option (software mode exclusive) and Preload Data Frame (Hardware mode exclusive) to the GSDX plugin settings for debug purposes.
The goal is to check the impact on game that have wrong RT content.
It helps a bit Smash Court Tennis Pro Tournament 2 but the game suffers
another texture cache bug. (RT BW is 10 whereas texture BW is 8)
Note: Armored Core: Last Raven must be tested (only game so far
that rely on the option and I didn't want to add a new one).
Typical wrong draw:
1/ draw in 32 bits
2/ draw in 24 bits
3/ Use alpha as a texure. (Must reuse the GPU data)
4/ Write alpha from EE
5/ Use alpha as a texure. (Must upload new data)
This commit fixes the step 5.
Fix#917 (Conflict - Desert Storm)
CID 146973 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR)2. uninit_member: Non-static class member overflow is not initialized in this constructor nor in any functions that it calls.
2999}
runion/rempty/rinter requires x < z and y < w
Help issue #762 (accurate blending issue)
If you want to shine, please put better GSVector code (AVX512 is 2 instruction :p)
In the DATE42 algo, first pass must find the primitive that write the
bad alpha value. If depth test is fail, alpha value won't be written therefore
you mustn't keep the primitive id.
In theory to ensure 100% correctness, depth would need to be fully executed
(currently depth write is disabled). However it requires to copy the depth buffer.
It is likely bad for the perf.
Issue reported on DBZInfWorld
Function pointer was mangled to avoid any collision. Nowadays all symbols
are hidden so no risk of collision.
Syntax is nicer beside it would allow to put back GLES3.2. I think it
supports most of the used extension.
glActiveTexture & glBlendColor are provided without symbol query.
CID 146834 (#2-1 of 2): Division or modulo by zero (DIVIDE_BY_ZERO)9. divide_by_zero: In expression tpf * 10000ULL / ttpf, division by expression ttpf which may be zero has undefined behavior.
CID 146839 (#1 of 1): Explicit null dereferenced (FORWARD_NULL)11. var_deref_model: Passing null pointer fb_pages to UsePages, which dereferences it.
CID 146840 (#1 of 1): Explicit null dereferenced (FORWARD_NULL)11. var_deref_model: Passing null pointer zb_pages to UsePages, which dereferences it.
* Prevent a potential null pointer deference in ```void GSRendererSW::UsePages()```
All combobox text can now be seen in full without having to click on the
combobox.
The internal and custom resolution stuff has been moved into the Hardware
Mode Settings groupbox since it doesn't affect software mode.
The dialog has also been rearranged a bit.
shaders/GSdx.fx is now the default location and is no longer hardcoded.
The external shader and external shader config can now be selected. (The
OpenGL renderer already has this feature.)
Note: It is still possible to not use a config file, just use an invalid
value for shaderfx_conf.
Don't use D3DX compile from file and compile from resource functions -
use the compile from memory function instead. It does the same thing,
except you have to set things up yourself.
Benefits:
Allows external shaders to be split into a config file and a shader file
without hardcoding the config file name.
Less code.
Yes, I more or less used the same message as the dx11 one.
Don't use D3DX compile from file and compile from resource functions -
use the compile from memory function instead. It does the same thing,
except you have to set things up yourself.
Benefits:
Easier move to D3DCompile when it becomes necessary.
Allows external shaders to be split into a config file and a shader
file without hardcoding the config file name.
Less code.
OpenGL does not use the cdecl calling convention (which is the default
calling convention for GSdx on Windows). Since DebugOutputToFile is used
by OpenGL, it needs to use the same calling convention that OpenGL uses.
This fixes a debug build crash when the OpenGL renderers are used and
debug_opengl is nonzero in the ini.
A couple of useless members were removed too.
Also fix wnd initialization
Coverity:
CID 146955 (#1 of 1): Uninitialized pointer read (UNINIT)
18. uninit_use: Using uninitialized value wnd[i].
Coverity:
CID 146816 (#1 of 1): Calling risky function (DC.STREAM_BUFFER)
dont_call: fscanf(FILE *, char const *, ...) assumes an arbitrarily large string, so callers must use correct precision specifiers or never use fscanf(FILE *, char const *, ...)
Without this patch, if a user initiates a recording and then cancels at the GSdx
dialog, the audio was recording anyway, which is probably highly unexpected.
However, while probably highly unexpected, it could still be useful to record
only audio, but with this patch it's now impossible.
We can reconsider if it turns out that people are actually using this "feature",
though one might as well set the video setting to be very unobtrusive (very low
resolution/bitrate) such that it uses very little CPU.
This is the internal resolution which GSdx uses and recording at this resolution
is optimal, i.e. without any dumb scaling, with all relevant pixels and without
redundant pixels.
The resulting clip still doesn't have the correct aspect ratio set, but that's
just a property which can be set to the clip afterwards, which is where the DAR
becomes useful. Since it's usually anamorphic, when muxing later with the audio
use the DAR to set the playback aspect ratio.
gsdx changes:
Remove native resolution checkbox from GUI and rework associated code
Small changes to Windows and Linux GUI
Support 8x native resolution
Fix custom resolution width less than native width use case
Previously it was saving the display name to the config but trying to restore
according to the friendly name.
Now store and restore according to the "displayName" which is more unique than
"friendlyName" since it includes GUID[s], and handle it consistently as _bstr_t.
This fixes the following issues when custom resolution is selected.
- When the width is smaller than the native resolution width, the
texture cache targets are removed on every Vsync signal, causing a
black screen issue.
- The texture cache code needs a 1 returned for the custom resolution
upscale multiplier or there'll be some really funny graphical issues.
It also removes unnecessary GetConfig (which I think unconditionally
does a a file read on Windows) calls if the width was increased - the
upscale multiplier is already stored, and the custom resolution width
and height calls are now unnecessary.
Also fix some whitespace issues.
Before this patch, when recording Progressive (frame) mode, it recorded all
the frames correctly but set the clip's fps property to 25/29.97, so when
played back it played at half the speed (but was fine when played at double speed).
This patch does not affect the number of frames recorded per second, but rather
only sets the resulting clip fps property to the correct one (double than before).
Also fixes a bug that in a non-managed window in progressive mode, the title
displayed "200%" speed when it should have displayed 100% speed.
Fixes#832
upscale_multiplier function values have been changed to allocate native resolution and also move custom resolution to 9.
Remove the old native checkbox value and include Native in the combo box.
Internal GSDX functions have also been updated with this new update to the upscale_multiplier variable.
* Greatly reduce the number of clut read (factor 10x)
* Avoid to get wrong TEXA texture in the cache.
* Fix "jump depends on uninitialized variable" Valgrind warning.
Fix#748
I try my best to avoid any breakage of DX but please test it too.
* add lengthly comment to explain the format
* Likely reduce the number of shader permutation
* Avoid slow AEM (on GPU)
Expect regressions because TC needs some fixes
v2: fix palette mode
CID 146833 (#2-1 of 2): Division or modulo by zero (DIVIDE_BY_ZERO)
divide_by_zero: In expression this->m_width / this->m_upscale_multiplier, division by expression this->m_upscale_multiplier which may be zero has undefined behavior.
CID 146835 (#1 of 1): Division or modulo by float zero (DIVIDE_BY_ZERO)
50. divide_by_zero: In expression (float)(end - start) / (float)frame_number, division by expression frame_number which may be zero has undefined behavior
Add 2 new shaders:
* ps_main12: cast a 16 bit depth to a RGB5A1 color
* ps_main16: cast a a RGB5A1 color to a 16 bit depth
Shader might be used in future commit as it seems Silent Hill uses this
kind of format.
Fix tab/indentation too
Add tooltips to Windows GUI. Fixes#578.
Rearrange Window GSdx GUI a bit.
Reformat tooltip descriptions to be both Linux and Windows (un)friendly.
Add new tooltip descriptions.
For Linux and Windows, add tooltip descriptions for AF, AA1, extra
rendering threads, Enable Shade Boost, FX Shader/External Shader, FXAA.
For Windows, also add Alpha Correction and Logarithmic Z tooltips.
For Linux, add software mipmap tooltip.
Thanks to tsunami2311, ssakash, gregory38 and whoever did the wiki for
helping with descriptions.
The hack description box messages were replaced by tooltips on Windows,
so there is no need for the first line in the description to refer to
the hack anymore.
3-state checkboxes were also converted to comboboxes, so remove any
references to checkbox states.
Reformat the descriptions so it looks okay on both Windows and Linux.
Could be better, but I'm out of ideas.
Also fix a few typos and missing punctuation.
Changes the three-state checkboxes to comboboxes.
Removes the hack description box.
Hides Alpha and Alpha Stencil when OpenGL is selected.
Rearranges the hacks dialog.
This moves the hardware, non-renderer dependent options to the top and
makes the renderer options visible only when it affects the current
renderer.
The texture filtering checkbox has been changed to a combobox and now
uses the same description as Linux does.
Also changes the Title to GSdx Settings.
The purpose is to avoid issue on MS-Intel driver without
a dedicated hack in the compilation.
Code doesn't use it so I suspect others implement to discard the
statement.
The CRC hack option is always used regardless of whether the HW Hack
checkbox is ticked. Move it to the main gsdx dialog and make it use the
same descriptions that are used in Linux.
And fix the accurate_blend_unit whitespace formatting.
according to some of the internal functions in GSDevice files, Anisotropic Filtering is only done when paltex is disabled. do the same on the GUI for user awareness.
Removes the checkbox of Anisotropic filtering from the GSDX plugin settings, the checkbox was usually used to enable & disable the AF which is not necessary since there is an option in the drop down list for disabling AF.
the internal function value of "AnisotropicFiltering" has been replaced with "MaxAnisotropy" for detection.
the detection uses the function getconfig("MaxAnisotropy", value) where value 0 means disabled and value is the default value when no value is set in the INI file.
I: pcsx2: spelling-error-in-binary usr/lib/i386-linux-gnu/pcsx2/libGSdx-1.0.0.so allows to allows one to
I: pcsx2: spelling-error-in-binary usr/lib/i386-linux-gnu/pcsx2/libGSdx-1.0.0.so Allow to Allow one to
.
Apparently lintian checks grammar too (most common ones).
The Wild Arms Offset hack can only be on or off. There is no third
state. Convert the three-state checkbox into a two-state checkbox, fix
the description and remove a leftover.
Intially GSBlendStateOGL was an alias of the m_blendMapD3D9 array
The object was replaced by an index in the array. Save 2k of memory duplication.
And too much useless code.
v2: push/pop blending state in DATE stuff
v3: remove m_state which is useless now
I used to disable the RT but it doesn't work well with 50cents. It seems
texture writes weren't propagated correctly to the depth buffer. Besides
the game write an alpha value whereas depth is 24 bits only...
The purpose is to emulate correctly destination alpha factor
An alpha channel of 128 is 1.0 in the GS but only ~0.5 in the GPU
I think few draw call use destination alpha so impact on perf must remains small.
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)
Previous CopyRect function does a memcopy without conversion.
This function will allow to use different format for input/output. Just a
possibility for the future
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.
Sourceforge was dead for more than a week therefore add the license
information. I could not find the original TGM source (dead link) so I'm not
even sure if this still applies or if the glsl was totally rewritten. None
of the glsl files have a copyright header so it's hard to tell.
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)
It fixes the bad light on Silent Hill with the SW renderer.
Full story
if Q is NaN, m_vt.m_eq.q becomes wrongly true
/Q will wrongly be optimized in the "Vertex Shader" of the SW
Note: Add an assert for the STQ handler
Code path is quite hot so no need to add extra check for nothing
GS uses integer value and does integer operation too.
This commit trunc the sampled texture, the interpoled fragment color
and the product of the 2.
It impacts negatively the perf of about 3/4% (GPU) but it fixes rendering on
suikoden and potentially some others games too.
Code was completey bitrotten
Code was a partial test (and yet 500 lines already)
Shader is more and more complex and multithreading support greatly
reduce the cost of shader switch
Nvidia allows to get the ASM of the shader of the compiled shader. It is useful
to check the performance.
It also allow me to compile most of shader code path for QA
Dump is enabled in linux replayer + debug_glsl_shader = 2
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.
Partially invalidate RT when there is a write in the middle of it (actually 2 pages below)
Code is not yet enabled because
1/ I want to stabilize latest update
2/ not sure of the impact of the code
3/ maybe it need a more generic version
Frame is always 32 bits but game can reuse it later as a 16 bits RT.
Fix half screen issue with Ricky Ponting Cricket
Unfortunately it triggers texture shuffle wrongly. I hope there is no
regression.
Fix depth issue.
First issue: RT and Depth are located at the same address.
=> solution, just lookup the depth (done for any games)
Second issue: Instead to draw a X,Y rectangle, they draw a 32 pix, 4096 pix
rectangle with a FBW of 1.
=> solution, don't bother just write the depth directly
Please someone add some CRC to enable the code
// 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 is mostly the texture shuffle effect which is now emulated correctly on openGL
They remain in the Dx level as speedhack in particular for
* Castlevania;
* GodOfWar;
* StarOcean3;
* ValkyrieProfile2;
* RadiataStories;
Remove old commented hack: Drakengard2, ArctheLad, ReZ
Separate hacks that are only enabled in aggressive mode
It might save a couple of fps
Add a define to test the perf if we keep only the blue channel. It brokes
the code in Prince Of Persia that use the Red/Green channel... Maybe the
speed hack :( Or find a way to replace all if with a lookup table
Note: it is only supported on OpenGL currently
Code unscale the texture to ease the conversion. Quality is awful (same as before)
but I'm not sure we can support an upscaled texture
Maybe the quality loss is due to the reduction without mipmap
Maybe the best solution will be to add an hack to extract the blue channel
(with texture swizzle), and uses a "full page/screen" spirte instead.
(it would be faster too)
Note: won't be compatible with MSAA (but gl doesn't support it anyway)
// In theory new textures contain invalidated data. Still in theory a new target
// must contains the content of the GS memory.
// In practice, TC will wrongly invalidate some RT. For example due to write on the alpha
// channel but colors is still valid. Unfortunately TC doesn't support the upload of data
// in target.
//
// Cleaning the code here will likely break several games. However it might reduce
// the noise in draw call debugging. It is the main reason to enable it on debug build.
//
// From a performance point of view, it might cost a little on big upscaling
// but normally few RT are miss so it must remain reasonable.
Game can directly uploads a background or the full image in
"CTRC" buffer. Previous code was a full black screen.
It will also avoid various black screen issue in gs dump.
hidden option: preload_frame_with_gs_data
Note: impact on upscaling was not tested and it's likely broken
Improve the rendering in MGS3 (even if the game is still broken
due to others TC issues)
// Typical bug (MGS3 blue cloud):
// 1/ RT used as 32 bits => alpha channel written
// 2/ RT used as 24 bits => no update of alpha channel
// 3/ Lookup of texture that used alpha channel as index, HasSharedBits will return false
// because of the previous draw call format
//
// Solution: consider the RT as 32 bits if the alpha was used in the past
"Regression" introduced in 8 bits RT support
The case appears when
1/ address of RT is the same as input texture
2/ input format is 8 bits
3/ previous lookup of RT was miss
It avoid various upscaling glitches on GS post-processing effect
// 1/ Palette is used to interpret the alpha channel of the RT as an index.
// Star Ocean 3 uses it to emulate a stencil buffer.
// 2/ Z formats are a bad idea to interpolate (discontinuties).
// 3/ 16 bits buffer is used to move data from a channel to another.
//
// I keep linear filtering for standard color even if I'm not sure that it is
// working correctly.
// Indeed, texture is reduced so you need to read all covered pixels (9 in 3x)
// to correctly interpolate the value. Linear interpolation is likely acceptable
// only in 2x scaling
//
// Src texture will still be bilinear interpolated so I'm really not sure
// that we need to do it here too.
//
// Future note: instead to do
// RT 2048x2048 -> T 1024x1024 -> RT 2048x2048
// We can maybe sample directly a bigger texture
// RT 2048x2048 -> T 2048x2048 -> RT 2048x2048
// Pro: better quality. Copy instead of StretchRect (must be faster)
// Cons: consume more memory
//
// In distant future: investigate to reuse the RT directly without any
// copy. Likely a speed boost and memory usage reduction.
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
It seems to impacts lots of games that still have issue (VP2, MTG3, PoP)
The PSMT32 format is read a PSMT8. I think we need to convert it as PSMT8H (i.e.
unpack it to have only an alpha channel)
To clear a (W, H) RT/Depth, GS can only send a (W , H/2) primitive with
a Z (or RT) pointer in the middle of the buffer
Top-half will be cleared by the color (or the depth).
Bottom-half will be cleared by the depth (or the color).
Code isn't enabled yet but it was tested with success on GoW2.
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
Rework the renderer selection logic to reload the saved renderer after
it's been changed in the config panel and to properly detect whether a
renderer toggle has been requested.
Changed the name of the Direct3D renderers so they would match the way
they are named in GSdx's Plugin Settings window. Also changed a printf
to say OpenCL instead of opencl.
Just use 1 in skip draw if you want to use it
openGL code will soon be fixed to support depth correctly (and I hope dx in the future)
Conflicts:
plugins/GSdx/GSState.cpp
plugins/GSdx/GSState.h
Old way various check boxes to configure the plugin
New way an unified drop down list
* Level 0 (dev only) => disable (mostly) all hacks and auto skip depth.
* Level 1 (dev only ) => enable oi/oo/cu hacks, others remains disabled
* Level 2 (GL) => enable most hacks except a couple of one that were fixed on openGL (speed impact)
* Level 3 (DX) => enable all hacks
* Level 4 => enable also aggresive CRC
Note: windows gui must be updated, and it will be nice to create a global tooltip
Gregory: Get all changes but keep C11 code path which will be the future
(if someone can find info on Visual Studio support)
* Keep the old posix api to use address sanitizer on gcc 4.9
Conflicts:
common/build/Utilities/utilities.vcxproj.filters
common/build/Utilities/utilities_vs2012.vcxproj.filters
common/build/Utilities/utilities_vs2013.vcxproj.filters
common/include/Utilities/MemcpyFast.h
common/include/Utilities/StringHelpers.h
common/src/Utilities/AlignedMalloc.cpp
common/src/Utilities/vssprintf.cpp
plugins/GSdx/stdafx.cpp
-Last commit broke Snowblind games again
-Width now based on output circuit
-Try to use the smallest possible while being at least the size of the screen
The idea is to easily extract bad frame from gs dump and to compare them
with the sw renderer.
Here the summary of the options:
savet => dump all input textures
save => dump the render target
savez => dump the depth buffer
savef => dump the frame
The idea will be to replace StretchRect for standard case with a framebuffer
blit. Potentially it toggles less gl state.
Worth a test on Star Ocean 3 that uses a lots this function for stencil emulation
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
When the RT is used as an input texture, we need to rescale it.
Previous behavior was to always uses a linear filtering (more smooth).
Unfortunately it broke some games that expected an exact value like Star Ocean 3
This commit will disable the linear filtering in normal filtering mode (filter = 0
or filter = 2)
This way, shadow of Star Ocean 3 will appear correctly in upscaling (not
100% perfect but can't do better)
Note: SO3 only requires a nearest sampling of the alpha channel but
I don't know the behavior for others games.
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.
Still not yet enabled by default
Potentially it can be optimized with the dot product but special care
need to be taken to ensure float accuracy.
Bonus: it could work on old GPU (aka DX9)
* 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 we're trying to infer the conversion shader based on the output format
It only works if the input data is RGBA8. It might not be true in the future
The idea was to merge them later with ffmpeg or other external tool
It kinds of work but current fps is around 2 !!! So not really usable
The speed issue is related to PNG.
So either the code to store pixels data must be optimized. Or maybe it misses some MT loves :)
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.
Persistent is slower (at least on my gs dump) because data is put
in host instead of the video memory
I don't understand why upload the data directly to the video memory
is faster
Initially we copy pitch by line in the PBO and tell the dma to only
use the first valid byte.
Now, we only copy useful data to the PBO. It reduce the copy and PBO memory requirement.
It seems a bit faster on native resolution
Initially I was afraid of the size increase but it can be
higly compressed.
Sometimes image was computed inside a RT but never outputed to
the screen which remains black. It is annoying to see the glitches
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 :)
initialization of IDC_ROUND_SPRITE
set&get config for round offset.
three state checkbox for the hack.
define IDC for round offset.
Linux (GUI) refresh.
typo
0: don't dump input texture
1: dump input texture
Now, you can do a first pass with only RT. When you find the wrong call, you can redump the input texture of the bad draw.
Initial code use a PBO to do asynchronous transfer. It is silly because
GSdx doesn't use this free time. So let's use a sync read. Same speed but
no PBO to manage.
-- slower (but that a debug feature)
++ smaller (40x-50x)
++ native support of alpha
Require libpng++ and the define ENABLE_OGL_PNG
Note: depth is not supported yet.
This way ogl_vertex_storage must be safer to activate
And it brings a nice performance boost (game with lots of primitives and
reasonable upscaling)
SotC testcase 4x: 61fps => 78fps
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
ogl_texture_storage 1 creates texture corruption.
Advance date is too slow, code need to be updated (properly) to uses 2 passes only not 3
Maybe one could be enough (sometimes)
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
Note yet enabled because I'm afraid of data corruption but feel free to test it
The option:
ogl_vertex_storage = 1
Performance note (warm cache+gs replay on colin3)
60 fps -> 76 fps
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).
* separate VS/GS and FS
* separate subroutine part of the FS
It already complex enough without subroutine stuff. Besides I'm not sure
we will keep subroutine on the future.
If someone has a more elegant solution, feel free to share it
spin_thread = 0
spin_thread = 1 // the faster but GS thread will never stop, very bad for laptop
It is faster on linux, it requires less code, and it is "portable"
It requires boost (only hpp files) + MSVC 2013 (for atomic) (seem doable by 2012 too)
Actually there are several queues that either use spinlock or full sleep
* Use tooltip for hack
* Update string of previous hack
* Remove unused hack
Note: hack_sprite_check requires 3 states (and potentially others hack too) but
I don't know how to do it. It likely requires a "scale button"
It works as bad as a "clever" implementation.
It seems to be enough for games such as venus/taisho-monoke/FFX
Note: it might creates glitches. Code will never be nice, so it is just
a trade-off
-1 is annoying because minimum value is 0. Instead to add more logic,
let's try to use 0 which seems to be good enough (fix regressions on DQ8/AT)
Unfortunately it causes a mini regression on taisho-mononoke. Rotation of sprite is done with 2 triangles.
Potentially previous value wrongly recover this section.
Tekken is broken with setting UserHacks_round_sprite_offset = 2 use 1 instead.
=> game use upscaling internaly so my rounding of coordinate break everything
Yakuza uses float coordinate but this hack only correct the integer coordinate
=> the solution will be to add a dedicated hack for this game.
Colin3 got a regression but I don't know when...
PS2 uses a -0.5 offset before sampling so texels must be rounded to half-pixel boundary
If fixes glitches on Venus and taisho-mononoke
Note: I didn't fixed yet texture render in reverse because I don't have any test for it.
UserHacks_round_sprite_offset = 1 <= enable correction of flat sprites
UserHacks_round_sprite_offset = 2 <= enable correction of all sprites (better on a couple of dump but not sure of the consequence)
I completely redo the algorithm. This time I do the projection and
interpolation of the 2 extrem vertex. This way I can compute the min/max
valid texture coordinate.
It gives stronger guarantee that texture sampling will be done inside the texture.
However it might have a performance impact, likely reasonable because it
is limited to sprite vertex.
A big thanks you to all people that provide me GS dump and test reports.
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.
Gregory Hainaut