ICO uses a depth of field effect for the fog. Depth is extracted
into the alpha channel of a texture. And then used as blending factor.
You need a 1:1 texture/pixel mapping otherwise you will line at boundaries.
In order to extract the DoF, ICO moves the depth buffer around the GS
memory. Memory moves are implemented in the not-scaled world. It means
that we can't have the above 1:1 ratio. And we don't know anymore that
data are coming from the current depth buffer.
The solution: I reused an HLE channel shader to read the depth buffer directly.
This way I have the guarantee that pixel/depth are aligned.
Close#1816
Otherwise you have a write before read typical race condition. It works
most of the time because textures are stored in temporary buffers (aka
texture cache). So the race condition requires texture invalidation in the mix.
I hope the perf impact will be small enough.
Fix#1691
Blood Will Tell: gray scale effect description
Frame is renderer in 0x700
Sync 0x700 (RT will be used as input)
Foreach page of frame
// The missing Sync was this one. You can't copy new data to 0x2800
// until you finish the rendering that use 0x2800 as input texture
// (AKA end of this foreach loop)
Sync 0x2800 (not the first iteration, texture will be used as a RT)
Copy page from 0x700+offset to 0x2800
Sync 0x2800 (RT will be used as input)
Render Effect line1 from 0x2800 to 0x700
Alpha test should only be disabled when writes to all of the alpha bits in the Framebuffer are masked. Fixes a regression in Dragon Ball Z: Budokai 3 scouter image rendering.
It allow to do the division before the size multiplication
It avoid a float overflow if T is too big.
Old behavior: (T * size) / Q
New behavior: (T / Q) * size
Performance Note:
* Rcp was replaced by a slow division (more accurate)
* At least we avoid a 2nd loop on the vertex buffer
It helps on Pro Soccer Club and Galerians Ash rendering
Tric Note:
SPRITE must be handled differently because the 'q' of first vertex could
be invalid
Bilinear applies to all renderer
* Common code done in GSVertexTrace
* Extend it with forced but sprite (trade-off between linear/upscale glitches)
* Linux GUI option was moved at the top with the renderer selection
Trilinear is moved to OGL hack
close#1837
Thanks to Flatout for the review and feedback.
It will take care to update the Window GUI :)
Typical bug, missing/wrong texture on the SW renderer but working fine on the HW renderer
Debugged on ATV Quad Power Racing 2 but I suspect couple of game are impacted
Bug description:
GSdx flatten the Q value of sprite. So m_vt.m_eq.q is true when Q(2N+1) are the same.
Q(2N) values could be random. The fix replaces Q0 by Q1 for the uniform Q value.
Previously, the NTSC saturation was also applied for double scan mode (Interlaced and Frame) where the developers send double the height to the DISP registers, saturation shouldn't be performed at such cases as the developers could send a value of 780 while the real size of the output would be 390 due to double scan mode. Doing the saturation later after identifying the real size also seems a bit counter-intuitive as we haven't discovered any cases where double scan games require the NTSC saturation hack. So let's just apply the saturation only for Interlaced (Field) Mode and omit the saturation step for other modes.
Isolate all the hacks into a separate subroutine and properly document about them, should make it easier for people to understand the display rectangle setup code, the hacks were totally messing up the readability of the function earlier.
The buffer contains extra room to avoid a segmentation fault due to an overflow.
Unfortunately the end of the buffer wasn't initialized which can lead to unexpected behavior.
Based on issue #1806 it could impact Guilty Gear X2
Mesa AMD was updated :)
all drivers[1] that support GL_ARB_shader_image_load_store got GL_ARB_clear_texture
[1] Intel driver misses others extensions to run GSdx
Rendering will be corrupted (for advance effects) if the driver doesn't support it.
However it allow to run with Mesa software emulation (or inside a virtual machine)
Note: mesa still requires an override of the buffer storage extension
MESA_EXTENSION_OVERRIDE=GL_ARB_buffer_storage
Previously, the combobox will reach an indeterminate state whenever it's passed with a value out of range via ComboBoxInit(). To avoid such cases, let's initialize the current selection of the combobox with the front element of the settings vector whenever we detect an out of range value which is not declared in the vector.
To reproduce the issue, set "Renderer" to some sort of crazy value like 50 in the GSdx.ini file and it'll mess up the whole GSdx plugin dialog really bad. This patch prevents such undesirable behavior by simply selecting the front element in the vector when we read an unsupported value.
Previously, the auto output circuit selection of the GSdx wasn't good, it simply defaulted to the second output circuit even when the first output circuit is also enabled. The new algorithm for auto selecting returns the merged rectangle dimensions when both of the output circuits are enabled and if the condition for merge is not satisfied then it returns the bigger output circuit.
I didn't fix all the warnings (purpose was to realign code with "recent" update)
Linux note: only miss 2 major items
* res/tfx.cl loading
* device descriptor
* And various bug fixes ;)
Previously, when F8 was triggered multiple times in a single second, the latest captured image would replace the previous captured one as it has the same name as the previous image.
The following patch detects such cases and adds a number along with the filename when new image capture is requested under the same time as the previous capture.
This reverts commit
d6383e6c21
It created a regression in Everybody's Golf 4/Hot Shots Golf 4, breaking the renderering when depth emulation is disabled/when using a Direct3D Hardware renderer.
The code is now a mirror of the ::add. So 1 insert == 1 erase
This way it won't crash on future update. And it will support future GS
memory wrapping improvement.
ZoE2:
RemoveAt overhead plummet to 0.5%. It was 17% !
However insertion is a bit slower. Due to the begin() after the push_front
v2: use std:: for lists and arrays
Removes Alpine Racer 3 hack. Issue has been resolved.
Moves NanoBreaker hack. Issue has been resolved for OpenGL and hack has
been moved to DX only.
Moves Tri-Ace games hacks. Hacks are also necessary for OpenGL with "Partial" CRC Hack Level to prevent massive slowdown.
Move Tales Of Legendia hack back as it's also necessary for OpenGL with "Partial" CRC Hack Level to prevent graphical issues.
Close: https://github.com/PCSX2/pcsx2/issues/1698
Added PAL and NTSC-U CRC's for Ar tonelico II.
Unfortunately it requires at least GCC6. If a nice guy can check the generated code on GCC6.
I don't know clang status.
Here the only example, I have found on the web
https://developers.redhat.com/blog/2016/02/25/new-asm-flags-feature-for-x86-in-gcc-6/
Current generated code in GSTextureCache::SourceMap::Add
38b3: bsf eax,esi
38b6: add esp,0x10
38b9: test esi,esi
38bb: jne 387e <GSTextureCache::SourceMap::Add(GSTextureCache::Source*, GIFRegTEX0 const&, GSOffset*)+0x6e>
BSF already set the Z flag when input (esi) is 0. So it would be better
to not put a silly add before the jump and to skip the test operation.
OMG, Zone of Ender got a speed boost from 11 fps to 45 fps
Seriously, the goal is to allow benchmarking GSdx without too much overhead of the main renderer draw call
Note: unlike the null renderer, texture/vertex uploading, 2D draw, texture conversions are still done.
* move the post-processing frame into the OSD tab
* Rename Global Settings to Renderer Settings
* put monitor and indicator check box on the same line
At least we have a similar number of options by tab
This reverts commit f77c1900fa.
Conflicts:
plugins/GSdx/GSTextureCache.cpp
Another fix was done later for Jak cut scene (or FMV). One game got a regression (don't remember which)
It's only ever updated after the queue is updated, so its state will
always lag slightly behind it. It's sufficient to just use empty().
This seems to fix some caching issues that were noticeable on Skylake
CPUs (#998).
In the previous code, the worker thread would notify the MTGS thread
while the mutex is still locked, which could cause the MTGS thread to
wake up and immediately go back to sleep again since it can't lock the
mutex.
Use a separate mutex for waiting, which avoids the issue.
Some PSX games seem to store image data of the drawing results in an undeterminate area out of range from the current context buffer. At such cases, calculate the height of both the frame memory rectangles combined.
What happens on "Crash bash" -
* At first draw, scissoring is limited to SCAY0- 0 & SCAY1- 255
* At second draw, scissoring is limited to SCAY0- 255 & SCAY0-511
Previously, we limited the height to the value of one single output texture, so instead of that let's calculate the total height of both the two buffers combined to prevent such issues.
Previously, we only calculated the width of a single output circuit which lead to missing a single pixel from the other output circuit which in turn causes offset issues in Persona games, I have customized GetDisplayRect() to now also calculate the dimensions of the merged rectangle when both the output circuits are enabled through the PMODE register, so this hack is no longer needed. :)
TL;DR - The above commit of mine accurately handles the offset issues by calculating union of the rects, removing this stupid hack. (not insulting any other developers, this stupid hack was mine :)
Passes the merged output circuit as the base size for texture cache scaling code. Helps fixing scaling issues where games use both of the output circuits for rendering.
Future Note: Alter the behavior of IsEnabled() check always preferring the second output circuit for some weird reason. I plan on changing it to a better auto-output circuit selection mechanism but that could probably be done some time in the future.
* Upgrade the counter to signed 32 bits. 16 bits is too small to contains the 64K value.
* Read ThreadProc/m_count when the mutex is locked
* Use old value of the fetch instead to read back the new value
Use a reinterpret_cast instead of casting the function pointer address
to a void** and dereferencing it.
Also remove an unnecessary (void) and avoid including stdafx.h.
Don't adjust 'image' and just use an additional offset.
'success' was kinda unnecessary when true or false could just be
directly returned.
Move 'compression' clamping out to GSPng::Save instead.
And throw in a whole bunch of const for good measure.
JMMT uses a bigger display height on NTSC progressive scan mode, which is not really unusual hence adjust the saturation hack to only take effect on interlaced NTSC mode.
However, the whole double screen issue on FMV still exists. As a bit of information, this game has the second output disabled but seems to have some valid data inside of it, maybe the second output data is leaked into the first one? most likely a bug in the frambuffer data management rather than a CRTC issue (needs to be investigated)
Previously, the height of the frame offset was also considered for the total height of the texture which was obviously wrong as the portion before the offset value isn't part of the frame memory.
In the previous code, the threads were created and destroyed in the base
class constructor and destructor, so the threads could potentially be
active while the object is in a partially constructed or destroyed state.
The thread however, relies on a virtual function to process the queue
items, and the vtable might not be in the desired state when the object
is partially constructed or destroyed.
This probably only matters during object destruction - no items are in
the queue during object construction so the virtual function won't be
called, but items may still be queued up when the destructor is called,
so the virtual function can be called. It wasn't an issue because all
uses of the thread explicitly waited for the queues to be empty before
invoking the destructor.
Adjust the constructor to take a std::function parameter, which the
thread will use instead to process queue items, and avoid inheriting
from the GSJobQueue class. This will also eliminate the need to
explicitly wait for all jobs to finish (unless there are other external
factors, of course), which would probably make future code safer.
The previous implementation of HPO adds an offset on vertex position. It
doesn't always work beside it moves the rendering window.
The new implementation will add a texture offset so that instead to sample
the middle of the GS texel, we will sample the middle of the real texture texel.
It must be manually enabled with
* UserHacks_HalfPixelOffset_New = 1 (keep a small offset as intended by GS effect)
* UserHacks_HalfPixelOffset_New = 2 (no offset)
v2: always apply a 0.5 offset in case of float coordinates (Tales of Abyss)
Might break other games but few of them uses float coordinates to read
back the target
Doesn't fully work yet
* Unknown stack frame
* Outside any known module
Potential root cause:
* Nvidia driver
* VU code as ebp is required for emulation so likely no frame
* Use POD type to avoid SSE/AVX compilation dependency
* global object to reduce cache miss
* dynamically object so give a chance to allocate below 2GB (allow x64
optimization)
tex address is a3
vm address is a1
Could help to avoid REX prefix
Reduce prologue/epilogue register copy
Byte code size 41893 => 38912 (on my testcase)
* bin2hex.h is removed
* vptest/vpblendvb YMM support integrated upsteam
* better support of rip for 64 bits
* AVX512 support (only miss the CPU now)
Local change: add BSD3 clause
It will requires a generic (register naming) linear interpolation to use it properly
Gather instruction requires an extra mask register therefore all registers name will be shuffled
Perf wise, initial haswell implementation seems to be microcode emulated.
Based on Gabest's work.
* Miss mipmap
Note: dithering info
It is a bit tricky as a2 on linux was rdx register which overlap with fzm (dh/dl)
It might require dedicated windows code
The main FindMinMax methods is perf critical so instead I created a separate function
to ensure the constness of the depth
Fix letter regression on Xenosaga3
* Explicitly cast w_pages and h_pages into uint32.
* Prevent signed/unsigned comparison by converting lod into unsigned integer, honestly how coud a mipmapping level be negative?
Previously the dedicated custom resolution scaling equation was ignored for the second SetScale() call, generalizing the equations will also fix the DMC scaling issue on custom resolution. Also remove unnecessary checks for null on scale factors. The possibility for having a null scale factor value only exists on custom resolution and it will only happen on cases where the output circuit isn't ready yet. So the ideal way would be to handle all the required conditions of output circuit on "m_renderer->CanUpscale()" itself.
Cost ought to remain small. Worst case is 2 extra "and" operation by group of pixels in scanline renderer
I think PixelAddressN functions are mostly call in the init.
CloseThread is called in the GSJobQueue destructor, so don't call it
again in the GSThread destructor.
Fixes#392, which was caused by a use after free.
Also prevents pthread_join() from being called twice for each thread
on non-Windows operating systems, which is undefined behaviour.
Code can be enabled with "wrap_gs_mem = 1". Code only allow a single shared memory but
I don't think we need more anyway.
Linux only, Kernel panic expected with the HW renderer.
Fix FMV on Silent Hill 3 with the SW renderer
Hardcode location of interface to the location 0. If I understand the
spec correctly (unlikely), variable in interface will get successive
location.
Goal is to reduce driver work. Instead to compute some location based on
name matching approach (and silly validation), the driver can now use
static allocation.
Tests on future Mesa 13 are welcome
Just clear the buffer. The generic solution will be a copy from buffer A
to buffer B But it requires
1/ a big buffer A (otherwise it would overflow)
2/ a line width rescaling (+ the upscaling mess support)
* Add full PMODE register to replace slbg/mmod
* Add full EXTBUF register (will allow to emulate write feedback)
* Add a third source (which will actually be the destination of the
write feedback)
mipmap option 3. Actually maybe a separate tri-linear option will be better
m_mipmap == 2 => use manual PS2 trilinear/mipmap
Otherwise
m_filter == 3 => always use full automatic trilinear interpolation
m_filter == 4 => use automatic trilinear interpolation when PS2 uses mipmap
m_filter == 5 => like 4 but force bilinear interpolation inside layer
Fix Berserk #1526 Well done guys but we're more clever than you ;)
So instead to mask the color channels as any guy that RTFM, they decided to use the illegal 8H frame format
WMS/WMT 2 is the region clamping mode.
Hw unit can't emulate it right so it can give you bad filtering (Fix#1025)
Note: I only did the fix because I wanted to remove the TEXA hack. Otherwise
it is still recommended to use openGL
It must work fine without it now.
From the google code comments:
It would be nice to test those games
* Ar Tonelico 2 (line in sprite regression?)
* breath of fire dragon quarter (overlayed user interface in the game)
v2: update Dx code to use the good format
* As sw renderer, don't bother to bypass it when it is ATST_ALWAYS
* Don't update the ATE register value
=> It is a really bad idea. Next draw call will be wrong if TEST register isn't written.
The TryAlphaTest context could have been updated
GS really uses an invalid texture located at 0.
Improve the rounding for R&C. The idea is to avoid the corner case were only
the corner of the triangle touch the 7/16 edge.
* Always do +1 before the draw call
* Prefix texture name with i (as input) to keep them before the FB
Goal is to ensure that all renderers share the same draw call value.
Game: harley davidson
* write tex0 ctx0
* write tex0 ctx1
* draw ctx 0
Previous GSdx behavior will load the clut every write of TEX0. In the
above case the draw will take the wrong clut.
To be honest, it could be a wrong emulation on the EE core emulation.
The hardware likely got a single clut (1KB cache is quite expensive)
So clut loading must be skipped if the context is wrong.
Next draw will use the ctx1 clut so I apply TEX0 when the context is switched
Please test harley davidson :)
v2: detect context switch from UpdateContext function
V3: always set m_env.CTXT[i].offset.tex, avoid crash (Thanks to FlatOutPS2 that spot the issue)
V4: move bad psm correction code (rebase put it in the wrong place)
Free bt
3 0xe676d194 in ~Source ../plugins/GSdx/GSTextureCache.cpp:1526
4 0xe676d194 in GSTextureCache::SourceMap::RemoveAt(GSTextureCache::Source*) ../plugins/GSdx/GSTextureCache.cpp:1990
5 0xe676f0fe in GSTextureCache::IncAge() ../plugins/GSdx/GSTextureCache.cpp:1022
Use bt
0 0xe6772a83 in GSTextureCache::LookupSource(GIFRegTEX0 const&, GIFRegTEXA const&, GSVector4i const&) ../plugins/GSdx/GSTextureCache.cpp:204
1 0xe66b0c9f in GSRendererHW::Draw() ../plugins/GSdx/GSRendererHW.cpp:579
2 0xe66fb43e in GSState::FlushPrim() ../plugins/GSdx/GSState.cpp:1509
Hypothesis the m_map array of list contains an invalid pointer
It is populated GSTextureCache::SourceMap::Add based on the coverage. The coverage is based on the offset.
So offset is potentially wrong. As mipmap code hack the offset value. It would be a nice culprit.
This commit avoids a potential bad transition between MIPMAP (which
overwrite the "offset") and the base layer (which wrongly keep an old "offset")
Conclusion, pray for my soul as it is very hard to reproduce
Ratchet & Clank (the third) uses an address of 0 for invalid mipmap.
It would be very awkward to put the middle layer of texture in start of
memory. So let's use this information to correct the lod.
It make the game more robust on the lod rounding
* Use texraw for the unconverted texture (keep index fmt)
=> avoid bad filename order with the multiple texture layers
* add the real mipmap address
* Use a nice string format
gsdx ogl: only use geometry shader to convert big enough draw call
The purpose of geometry shader is to reduce bandwidth (72 bytes by sprite)
and CPU load.
Unfortunately it increases CPU load due to extra shader validations.
So geometry shader will only be enabled for draw call with more than
16 sprites (arbitrarily, smallest number before shadow hearts plummet)
v2: don't disable geometry shader in replayer.
It is easier to spot sprite rendering and to manually read vertex info.
Strangely the game uses large texture to handle texture buffer.
I think it plays with WMS/WMT. I'm not sure texture shuffling is 100%
correct here. But without it, it's completely broken.
* 0x9C712FF0, Jak1, EU
* 0x472E7699, Jak1, US
* 0x2479F4A9, Jak2, EU
* 0x12804727, Jak3, EU
* 0xDF659E77, JakX, EU
Please report me the CRC of the US version too so I can add them.
Please test the shadows rendering (openGL HW + accurate blending at least basic)
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
Jonathan Li