The latter was introduced last year in order to allow debugging without
turning off Fastmem; I wanted to see if the old way was faster. In
fact, it is not, but may as well keep the define around in case future
testing is necessary.
It only ever did anything on 32-bit OS X.
Anyway, it wasn't even on the right functions, and these days
ABI_PushRegistersAndAdjustStack should handle maintaining the ABI
correctly.
wxGetActiveWindow is implemented as "return NULL" on OS X, while
wxWindow::FindFocus works. On Windows, the difference is in the use of
GetActiveWindow() vs. GetForegroundWindow(). A MSDN comment says:
> A system has only one active window, which GetForegroundWindow()
> returns. GetActiveWindow() seems to return the same window as
> GetForegroundWindow() if the foreground window belongs to the current
> thread. Otherwise, it always returns null, rather than the topmost
> window of the calling thread.
Since we are on the GUI thread, it shouldn't make any difference.
This noticeably includes GL_ARB_get_program_binary, which was previously
thought unsupported on OS X. Well, actually, the OS X implementation is
trivial and reports 0 binary formats (as of 10.10; this is hardcoded in
GLEngine, by the way), but at least it'll work if it's fixed someday.
It now affects the GPU determinism mode as well as some miscellaneous
things that were calling IsNetPlayRunning. Probably incomplete.
Notably, this can change while paused, if the user starts recording a
movie. The movie code appears to have been missing locking between
setting g_playMode and doing other things, which probably had a small
chance of causing crashes or even desynced movies; fix that with
PauseAndLock.
The next commit will add a hidden config variable to override GPU
determinism mode.
It's a relatively big commit (less big with -w), but it's hard to test
any of this separately...
The basic problem is that in netplay or movies, the state of the CPU
must be deterministic, including when the game receives notification
that the GPU has processed FIFO data. Dual core mode notifies the game
whenever the GPU thread actually gets around to doing the work, so it
isn't deterministic. Single core mode is because it notifies the game
'instantly' (after processing the data synchronously), but it's too slow
for many systems and games.
My old dc-netplay branch worked as follows: everything worked as normal
except the state of the CP registers was a lie, and the CPU thread only
delivered results when idle detection triggered (waiting for the GPU if
they weren't ready at that point). Usually, a game is idle iff all the
work for the frame has been done, except for a small amount of work
depending on the GPU result, so neither the CPU or the GPU waiting on
the other affected performance much. However, it's possible that the
game could be waiting for some earlier interrupt, and any of several
games which, for whatever reason, never went into a detectable idle
(even when I tried to improve the detection) would never receive results
at all. (The current method should have better compatibility, but it
also has slightly higher overhead and breaks some other things, so I
want to reimplement this, hopefully with less impact on the code, in the
future.)
With this commit, the basic idea is that the CPU thread acts as if the
work has been done instantly, like single core mode, but actually hands
it off asynchronously to the GPU thread (after backing up some data that
the game might change in memory before it's actually done). Since the
work isn't done, any feedback from the GPU to the CPU, such as real
XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is
broken; but most games work with these options disabled, and there is no
need to try to detect what the CPU thread is doing.
Technically: when the flag g_use_deterministic_gpu_thread (currently
stuck on) is on, the CPU thread calls RunGpu like in single core mode.
This function synchronously copies the data from the FIFO to the
internal video buffer and updates the CP registers, interrupts, etc.
However, instead of the regular ReadDataFromFifo followed by running the
opcode decoder, it runs ReadDataFromFifoOnCPU ->
OpcodeDecoder_Preprocess, which relatively quickly scans through the
FIFO data, detects SetFinish calls etc., which are immediately fired,
and saves certain associated data from memory (e.g. display lists) in
AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at
the moment), before handing the data off to the GPU thread to actually
render. That makes up the bulk of this commit.
In various circumstances, including the aforementioned EFB pokes and
performance queries as well as swap requests (i.e. the end of a frame -
we don't want the CPU potentially pumping out frames too quickly and the
GPU falling behind*), SyncGPU is called to wait for actual completion.
The overhead mainly comes from OpcodeDecoder_Preprocess (which is,
again, synchronous), as well as the actual copying.
Currently, display lists and such are escrowed from main memory even
though they usually won't change over the course of a frame, and
textures are not even though they might, resulting in a small chance of
graphical glitches. When the texture locking (i.e. fault on write) code
lands, I can make this all correct and maybe a little faster.
* This suggests an alternate determinism method of just delaying results
until a short time before the end of each frame. For all I know this
might mostly work - I haven't tried it - but if any significant work
hinges on the competion of render to texture etc., the frame will be
missed.
videoBuffer -> s_video_buffer
size -> s_video_buffer_write_ptr
g_pVideoData -> g_video_buffer_read_ptr (impl moved to Fifo.cpp)
This eradicates the wonderful use of 'size' as a global name, and makes
it clear that s_video_buffer_write_ptr and g_video_buffer_read_ptr are
the two ends of the FIFO buffer s_video_buffer.
Oh, and remove a useless namespace {}.
This state will be used to calculate sizes for skipping over commands on
a separate thread. An alternative to having these state variables would
be to have the preprocessor stash "state as we go" somewhere, but I
think that would be much uglier.
GetVertexSize now takes an extra argument to determine which state to
use, as does FifoCommandRunnable, which calls it. While I'm modifying
FifoCommandRunnable, I also change it to take a buffer and size as
parameters rather than using g_pVideoData, which will also be necessary
later. I also get rid of an unused overload.
VertexLoader::VertexLoader was setting loop_counter, a *static*
variable, to 0. This was nonsensical, but harmless until I started to
run it on a separate thread, where it had a chance of interfering with a
running vertex translator.
Switch to just using a register for the loop counter.
- Lazily create the native vertex format (which involves GL calls) from
RunVertices rather than RefreshLoader itself, freeing the latter to be
run from the CPU thread (hopefully).
- In order to avoid useless allocations while doing so, store the native
format inside the VertexLoader rather than using a cache entry.
- Wrap the s_vertex_loader_map in a lock, for similar reasons.
That commit reorganized fastmem a bit; I wrote it before the patch to
support fastmem in JitIL landed, and forgot to edit it to account for
the fact. Since JitILBase now derives from Jitx86Base, the HandleFault
override can just be removed.
Rather than *MemTools.cpp checking whether the address is in the
emulated range itself (which, as of the next commit, doesn't cover every
kind of access the JIT might want to intercept) and doing PC
replacement, they just pass the access address and context to
jit->HandleFault, which does the rest itself.
Because SContext is now in JitInterface, I wanted JitBackpatch.h (which
defines it) to be lightweight, so I moved TrampolineCache and associated
x64{Analyzer,Emitter} dependencies into its own file. I hate adding new
files in three places, two of which are MSVC...
While I'm at it, edit a misleading comment.
When executing a BL-type instruction, push the new LR onto the stack,
then CALL the dispatcher or linked block rather than JMPing to it. When
executing BLR, compare [rsp+8] to LR, and RET if it's right, which it
usually will be unless the thread was switched out. If it's not right,
reset RSP to avoid overflow.
This both saves a trip through the dispatcher and improves branch
prediction.
There is a small possibility of stack overflow anyway, which should
be handled... *yawn*
These calls are made outside of JIT blocks, and thus previously did not
read any protection - register use is taken into account and the outer
dispatcher stack frame is sufficient. However, if data is to be stored
on the stack, these calls must reserve stack shadow space on Windows to
avoid clobbering it.
To avoid FPRs being pushed unnecessarily, I checked the uses: DSPEmitter
doesn't use FPRs, and VertexLoader doesn't use anything but RAX, so I
specified the register list accordingly. The regular JIT, however, does
use FPRs, and as far as I can tell, it was incorrect not to save them in
the outer routine. Since the dispatcher loop is only exited when
pausing or stopping, this should have no noticeable performance impact.
- Factor common work into a helper function.
- Replace confusingly named "noProlog" with "rsp_alignment". Now that
x86 is not supported, we can just specify it explicitly as 8 for
clarity.
- Add the option to include more frame size, which I'll need later.
- Revert a change by magumagu in March which replaced MOVAPD with MOVUPD
on account of 32-bit Windows, since it's no longer supported. True,
apparently recent processors don't execute the former any faster if the
pointer is, in fact, aligned, but there's no point using MOVUPD for
something that's guaranteed to be aligned...
(I discovered that GenFrsqrte and GenFres were incorrectly passing false
to noProlog - they were, in fact, functions without prologs, the
original meaning of the parameter - which caused the previous change to
break. This is now fixed.)
Uses are split into three categories:
- Arbitrary (except for size savings) - constants like RSCRATCH are
used.
- ABI (i.e. RAX as return value) - ABI_RETURN is used.
- Fixed by architecture (RCX shifts, RDX/RAX for some instructions) -
explicit register is kept.
In theory this allows the assignments to be modified easily. I verified
that I was able to run Melee with all the registers changed, although
there may be issues if RSCRATCH[2] and ABI_PARAM{1,2} conflict.
And switch to a register order that consistently prefers callee-save to
caller-save. phire suggested putting rdi/rsi first, even though they're
caller-save, to save code space; this is more conservative and I can do
that later.
Rather than using a variety of registers including RSI, ABI_PARAM1
(either RCX or RDI), RCX, and RDX, the rule is:
- RDI and RSI are never used. This allows them to be allocated on Unix,
bringing parity with Windows.
- RDX is a permanent temporary register along with RAX (and is thus not
FlushLocked). It's used frequently enough that allocating it would
probably be a bad idea, as it would constantly get flushed.
- RCX is allocatable, but is flushed in two situations:
- Non-immediate shifts (rlwnm), because x86 requires RCX to be used.
- Paired single loads and stores, because they require three
temporary registers: the helper functions take two integer
arguments, and another register is used as an index to get the
function address.
These should be relatively rare.
While we're at it, in stores, use the registers directly where possible
rather than always using temporaries (by making SafeWriteRegToReg
clobber less). The address doesn't need to be clobbered in the usual
case, and on CPUs with MOVBE, neither does the value.
Oh, and get rid of a useless MEMCHECK.
This commit does not actually add new registers to the allocation order;
it is intended to test for any performance or correctness issues
separately.
The special case is where the registers are actually to be swapped (i.e.
func(ABI_PARAM2, ABI_PARAM1); this was previously impossible but would
be ugly not to handle anyway.
In two cases, my old code was using a temporary register but not saving
it properly; it basically worked by accident (an otherwise useless
FlushLock was causing CallerSavedRegistersInUse to think it was in use
by the GPR cache, even though it was actually a temporary).
I'm going to modify this in the next commit to use RDX, but I didn't
want to leave a broken revision in the middle.
The register is RBP, previously in the GPR allocation order. The next
commit will investigate whether there are too few GPRs (now or before),
but for now there is no replacement.
Previously, it was accessed RIP relatively; using RBP, anything in the
first 0x100 bytes of ppcState (including all the GPRs) can be accessed
with three fewer bytes. Code to access ppcState is generated constantly
(mostly by register save/load), so in principle, this should improve
instruction cache footprint significantly. It seems that this makes a
significant performance difference in practice.
The vast majority of this commit is mechanically replacing
M(&PowerPC::ppcState.x) with a new macro PPCSTATE(x).
Version 2: gets most of the cases which were using the register access
macros.
The old method would always evict the first suitable register, i.e. the
same register every time once the cache got full. The cache doesn't get
terribly often, but the result is pathological...
This causes glDrawArrays to fail in core profile, and thus on OS X, see:
http://renderingpipeline.com/2012/03/attribute-less-rendering/
There must be something bound, even though it is not used.
Fixes#7599. I'm not sure this is actually the best way to fix it,
since AFAICT it makes a nonobvious assumption that *something* will be
bound before the first attributeless rendering in
TextureConverter::DecodeToTexture, but it's what degasus suggested and
seems to work.
Separated out from my gpu-determinism branch by request. It's not a big
commit; I just like to write long commit messages.
The main reason to kill it is hopefully a slight performance improvement
from avoiding the double switch (especially in single core mode);
however, this also improves cycle calculation, as described below.
- FifoCommandRunnable is removed; in its stead, Decode returns the
number of cycles (which only matters for "sync" GPU mode), or 0 if there
was not enough data, and is also responsible for unknown opcode alerts.
Decode and DecodeSemiNop are almost identical, so the latter is replaced
with a skipped_frame parameter to Decode. Doesn't mean we can't improve
skipped_frame mode to do less work; if, at such a point, branching on it
has too much overhead (it certainly won't now), it can always be changed
to a template parameter.
- FifoCommandRunnable used a fixed, large cycle count for display lists,
regardless of the contents. Presumably the actual hardware's processing
time is mostly the processing time of whatever commands are in the list,
and with this change InterpretDisplayList can just return the list's
cycle count to be added to the total. (Since the calculation for this
is part of Decode, it didn't seem easy to split this change up.)
To facilitate this, Decode also gains an explicit 'end' parameter in
lieu of FifoCommandRunnable's call to GetVideoBufferEndPtr, which can
point to there or to the end of a display list (or elsewhere in
gpu-determinism, but that's another story). Also, as a small
optimization, InterpretDisplayList now calls OpcodeDecoder_Run rather
than having its own Decode loop, to allow Decode to be inlined (haven't
checked whether this actually happens though).
skipped_frame mode still does not traverse display lists and uses the
old fake value of 45 cycles. degasus has suggested that this hack is
not essential for performance and can be removed, but I want to separate
any potential performance impact of that from this commit.
comex