Ryujinx/Ryujinx.Graphics/Graphics3d/NvGpuEngineP2mf.cs
BaronKiko 92c1726647 GPU Profiling (#570)
* Profiler initial setup

* Capture actual timing data

* Profiling data dumped to file on close

* Support for multiple sessions under the same name

* Service profiling

* Sort output for easier read

* csv output

* Split session into 2 seperate values

* Refactor name to category

* Basic profiling window dummy. Toggle with F1 or set key with config
No actual data displayed yet, just a pretty triangle

* Simple font rendering

* Display some actual timing data

* Fix font bearing being ignored

* x bearing and advance. Fixed y bearing calc

* Different coloured lines to make reading easier

* Scrolling

* Multiple columns for name

* Column titles

* display in ms rather than ticks

* Bars to display times

* Sortable columns

* Regex filtering

* Better instant timing calculation
Fixed minor regex bug

* Better filtering
Better max value calculation
Skip some rendering to reduce profiler weight

* Variable update rate

* Show/hide inactive button
Some other touchups

* Add missing project reference

* Hide inactive and pause

* Fix viewport errors

* Update initial window position

* Variable name cleanup

* Disable timing dump by default

* Internal Profile refactor and cleanup

* Timing info cleanup

* Profile config cleanup

* Settings cleanup

* Button refactor

* Profile refactor

* Profile window cleanup

* Window manager refactor

* Font service cleanup

* Fixed bug in profiling method where method was called twice without profiling enabled

* Allow update rates of less than 1hz

* Stop using window.run because it's apparently not great for performance.
Some other performance things, should only draw a new frame when something has changed

* Improved time tracking to keep history

* Profile window was getting too long so I added regions and split bar rendering out into partial class

* Dummy graph view with button to toggle

* Realtime graphing initial commit

* Display totals on new bar

* Simple zooming support with arrow keys

* Limit graph zoom and label start and stop

* Added support for timing flags

* Stop data running away when paused and frame updated

* Manual step button

* Update at when flag issued (ie every frame)

* Removed useless finish profiling call

* Enable and disable profiling at compile time.

* Better plage for frame swap flag, also kept enough flags to cover larger time spans

* No more stopwatches created, uses PerformanceCounter now

* public and internal fields to props

* Move visible update to update rather than draw as it causes a lockup if called from draw
Also added profile window disposal so closing main window closes profiler too

* Fixed optimization settings for profiled builds

* Appveyer script guess to add profiling builds

* Quotes

* 1 less quote

* Maybe escape space?

* Specify config

* Different approach

* Fix file paths

* Fix another path

* Better artifact naming

* Missing -

* test  string

* Removed for, to test

* readd for

* moved dashes around so artifacts can begin with letters

* quote env vars

* martix

* Removed configs

* Much more efficient capture, ConcurrentDictionary was causing too much overhead

* Skip repeating pixels during draw

* Stop ram usage getting too high. Compensating for cleanup doing more now

* Profile CPU, execute skipped because it's just too much work

* Fixed bug with skipping draws. Furthest needed to be reset every loop

* Less distracting colour for timing flags

* Removed profile method function. It just doesn't play nice with conditional compilation so best to remove it now before it's used a lot

* Profile vertex end

* Null check for category, group and item

* Forgot to reset instant count/time

* Increment line when blank

* Fix threading conflict
Fixed instant count and time. Now accuratly represents the total time and count in the buffer

* Fixed bug in time rendering where times were being trimmed to an int.
Also added  microsecond/millisecond formatting to reduce the number of decimal places needed

* Profiled methods used by VertexEndGl

* Texture 2d copy profiled

* Support for multiple profiling levels

* Sometimes it would have to wait a long time for lock to clear so moved it to a tryenter and skip if already locked

* Dumb bug regarding clearing of timestamps. Start is already removed so no need to add it to the start

* Optimisations in drawing routine:
Only calculate bar top and bottom once per bar rather than once per timestamp
Pre-calculate the right side of the graph as it was being calculated multiple times per bar
Skip rendering timestamps that occupy the same pixel space now uses the raw timestamp to decide. While technically not as accurate it's much easier as the right side of the bar doesn't have to be calculated for a skipped timestamp

* Couple alignment changes

* Engine 3d call method profiled

* Custom equals overload for profile config. The default implpmentation was just too slow

* Bump cleanup thread priority. It clears the timer queue so it need to be run frequently

* Fixed bug with scrolling caused by recent rendering optimisations. Simply forgot to increment the line index on a skipped line

* Stopped blocking memory disposal so much. Also parralised(?) cleanup call

* Profiled M2mf, P2mf and shader compilation

* Missing file

* Removed deprecated config

* Removed legacy npaddebug

* Using statement no longer needed

* Apply suggestions from AcK

Co-Authored-By: Ac_K <Acoustik666@gmail.com>

* Apply suggestions from code review

Co-Authored-By: Ac_K <Acoustik666@gmail.com>
2019-06-01 02:13:57 +02:00

179 lines
5.1 KiB
C#

using Ryujinx.Graphics.Memory;
using Ryujinx.Graphics.Texture;
using System.Collections.Generic;
using Ryujinx.Profiler;
namespace Ryujinx.Graphics.Graphics3d
{
class NvGpuEngineP2mf : INvGpuEngine
{
public int[] Registers { get; private set; }
private NvGpu _gpu;
private Dictionary<int, NvGpuMethod> _methods;
private int _copyStartX;
private int _copyStartY;
private int _copyWidth;
private int _copyHeight;
private int _copyGobBlockHeight;
private long _copyAddress;
private int _copyOffset;
private int _copySize;
private bool _copyLinear;
private byte[] _buffer;
public NvGpuEngineP2mf(NvGpu gpu)
{
_gpu = gpu;
Registers = new int[0x80];
_methods = new Dictionary<int, NvGpuMethod>();
void AddMethod(int meth, int count, int stride, NvGpuMethod method)
{
while (count-- > 0)
{
_methods.Add(meth, method);
meth += stride;
}
}
AddMethod(0x6c, 1, 1, Execute);
AddMethod(0x6d, 1, 1, PushData);
}
public void CallMethod(NvGpuVmm vmm, GpuMethodCall methCall)
{
if (_methods.TryGetValue(methCall.Method, out NvGpuMethod method))
{
ProfileConfig profile = Profiles.GPU.EngineP2mf.PushData;
profile.SessionItem = method.Method.Name;
Profile.Begin(profile);
method(vmm, methCall);
Profile.End(profile);
}
else
{
WriteRegister(methCall);
}
}
private void Execute(NvGpuVmm vmm, GpuMethodCall methCall)
{
Profile.Begin(Profiles.GPU.EngineP2mf.Execute);
//TODO: Some registers and copy modes are still not implemented.
int control = methCall.Argument;
long dstAddress = MakeInt64From2xInt32(NvGpuEngineP2mfReg.DstAddress);
int dstPitch = ReadRegister(NvGpuEngineP2mfReg.DstPitch);
int dstBlkDim = ReadRegister(NvGpuEngineP2mfReg.DstBlockDim);
int dstX = ReadRegister(NvGpuEngineP2mfReg.DstX);
int dstY = ReadRegister(NvGpuEngineP2mfReg.DstY);
int dstWidth = ReadRegister(NvGpuEngineP2mfReg.DstWidth);
int dstHeight = ReadRegister(NvGpuEngineP2mfReg.DstHeight);
int lineLengthIn = ReadRegister(NvGpuEngineP2mfReg.LineLengthIn);
int lineCount = ReadRegister(NvGpuEngineP2mfReg.LineCount);
_copyLinear = (control & 1) != 0;
_copyGobBlockHeight = 1 << ((dstBlkDim >> 4) & 0xf);
_copyStartX = dstX;
_copyStartY = dstY;
_copyWidth = dstWidth;
_copyHeight = dstHeight;
_copyAddress = dstAddress;
_copyOffset = 0;
_copySize = lineLengthIn * lineCount;
_buffer = new byte[_copySize];
Profile.End(Profiles.GPU.EngineP2mf.Execute);
}
private void PushData(NvGpuVmm vmm, GpuMethodCall methCall)
{
if (_buffer == null)
{
return;
}
Profile.Begin(Profiles.GPU.EngineP2mf.PushData);
for (int shift = 0; shift < 32 && _copyOffset < _copySize; shift += 8, _copyOffset++)
{
_buffer[_copyOffset] = (byte)(methCall.Argument >> shift);
}
if (methCall.IsLastCall)
{
if (_copyLinear)
{
vmm.WriteBytes(_copyAddress, _buffer);
}
else
{
BlockLinearSwizzle swizzle = new BlockLinearSwizzle(
_copyWidth,
_copyHeight, 1,
_copyGobBlockHeight, 1, 1);
int srcOffset = 0;
for (int y = _copyStartY; y < _copyHeight && srcOffset < _copySize; y++)
for (int x = _copyStartX; x < _copyWidth && srcOffset < _copySize; x++)
{
int dstOffset = swizzle.GetSwizzleOffset(x, y, 0);
vmm.WriteByte(_copyAddress + dstOffset, _buffer[srcOffset++]);
}
}
_buffer = null;
}
Profile.End(Profiles.GPU.EngineP2mf.PushData);
}
private long MakeInt64From2xInt32(NvGpuEngineP2mfReg reg)
{
return
(long)Registers[(int)reg + 0] << 32 |
(uint)Registers[(int)reg + 1];
}
private void WriteRegister(GpuMethodCall methCall)
{
Registers[methCall.Method] = methCall.Argument;
}
private int ReadRegister(NvGpuEngineP2mfReg reg)
{
return Registers[(int)reg];
}
private void WriteRegister(NvGpuEngineP2mfReg reg, int value)
{
Registers[(int)reg] = value;
}
}
}