Merge branch 'perfqueries'.

Adds support for PE performance metrics.
Used in Super Mario Sunshine's "Scrubbing Sirena Beach" level to determine when enough goop has been cleaned up to finish the level.
Also used in TimeSplitters: Future Perfect to determine the appearance of flares around light sources (e.g. sun).

OpenGL and D3D11 only. D3D9 support unlikely to be added unless anyone bothers to do the work.
Initial work and D3D11 support by me. Kudos go to Billiard for adding the OpenGL support and reviving development of this branch that way :D

Slightly (~7%) decreases performance when performance metrics are used (and only then).

Fixes issue 1498.
Fixes issue 5368.
This commit is contained in:
NeoBrainX 2013-03-02 00:04:44 +01:00
commit 008fdc7310
38 changed files with 736 additions and 51 deletions

View File

@ -22,6 +22,7 @@
#include <vector>
#include "ChunkFile.h"
#include "../../VideoCommon/Src/PerfQueryBase.h"
typedef void (*writeFn16)(const u16,const u32);
typedef void (*writeFn32)(const u32,const u32);
@ -107,6 +108,7 @@ public:
virtual void Video_EndField() = 0;
virtual u32 Video_AccessEFB(EFBAccessType, u32, u32, u32) = 0;
virtual u32 Video_GetQueryResult(PerfQueryType type) = 0;
virtual void Video_AddMessage(const char* pstr, unsigned int milliseconds) = 0;
virtual void Video_ClearMessages() = 0;
@ -156,8 +158,10 @@ class VideoBackendHardware : public VideoBackend
void Video_ExitLoop();
void Video_BeginField(u32, FieldType, u32, u32);
void Video_EndField();
u32 Video_AccessEFB(EFBAccessType, u32, u32, u32);
u32 Video_AccessEFB(EFBAccessType, u32, u32, u32);
u32 Video_GetQueryResult(PerfQueryType type);
void Video_AddMessage(const char* pstr, unsigned int milliseconds);
void Video_ClearMessages();
bool Video_Screenshot(const char* filename);

View File

@ -16,6 +16,7 @@ set(SRCS Src/BPFunctions.cpp
Src/OpcodeDecoding.cpp
Src/OpenCL.cpp
Src/OpenCL/OCLTextureDecoder.cpp
Src/PerfQueryBase.cpp
Src/PixelEngine.cpp
Src/PixelShaderGen.cpp
Src/PixelShaderManager.cpp

View File

@ -62,7 +62,7 @@
#define BPMEM_COPYFILTER1 0x54
#define BPMEM_CLEARBBOX1 0x55
#define BPMEM_CLEARBBOX2 0x56
#define BPMEM_UNKNOWN_57 0x57
#define BPMEM_CLEAR_PIXEL_PERF 0x57
#define BPMEM_REVBITS 0x58
#define BPMEM_SCISSOROFFSET 0x59
#define BPMEM_PRELOAD_ADDR 0x60

View File

@ -31,6 +31,7 @@
#include "VertexShaderManager.h"
#include "Thread.h"
#include "HW/Memmap.h"
#include "PerfQueryBase.h"
using namespace BPFunctions;
@ -62,7 +63,6 @@ void RenderToXFB(const BPCmd &bp, const EFBRectangle &rc, float yScale, float xf
{
Renderer::RenderToXFB(xfbAddr, dstWidth, dstHeight, rc, gamma);
}
void BPWritten(const BPCmd& bp)
{
/*
@ -144,7 +144,8 @@ void BPWritten(const BPCmd& bp)
|| bp.address == BPMEM_LOADTLUT0
|| bp.address == BPMEM_LOADTLUT1
|| bp.address == BPMEM_TEXINVALIDATE
|| bp.address == BPMEM_PRELOAD_MODE))
|| bp.address == BPMEM_PRELOAD_MODE
|| bp.address == BPMEM_CLEAR_PIXEL_PERF))
{
return;
}
@ -484,9 +485,10 @@ void BPWritten(const BPCmd& bp)
case BPMEM_IND_IMASK: // Index Mask ?
case BPMEM_REVBITS: // Always set to 0x0F when GX_InitRevBits() is called.
break;
case BPMEM_UNKNOWN_57: // Sunshine alternates this register between values 0x000 and 0xAAA
DEBUG_LOG(VIDEO, "Unknown BP Reg 0x57: %08x", bp.newvalue);
case BPMEM_CLEAR_PIXEL_PERF:
// GXClearPixMetric writes 0xAAA here, Sunshine alternates this register between values 0x000 and 0xAAA
g_perf_query->ResetQuery();
break;
case BPMEM_PRELOAD_ADDR:

View File

@ -21,6 +21,10 @@ volatile u32 s_swapRequested = false;
u32 s_efbAccessRequested = false;
volatile u32 s_FifoShuttingDown = false;
std::condition_variable s_perf_query_cond;
std::mutex s_perf_query_lock;
static volatile bool s_perf_query_requested;
static volatile struct
{
u32 xfbAddr;
@ -169,6 +173,43 @@ u32 VideoBackendHardware::Video_AccessEFB(EFBAccessType type, u32 x, u32 y, u32
return 0;
}
static bool QueryResultIsReady()
{
return !s_perf_query_requested || s_FifoShuttingDown;
}
void VideoFifo_CheckPerfQueryRequest()
{
if (s_perf_query_requested)
{
g_perf_query->FlushResults();
{
std::lock_guard<std::mutex> lk(s_perf_query_lock);
s_perf_query_requested = false;
}
s_perf_query_cond.notify_one();
}
}
u32 VideoBackendHardware::Video_GetQueryResult(PerfQueryType type)
{
// TODO: Is this check sane?
if (!g_perf_query->IsFlushed())
{
if (SConfig::GetInstance().m_LocalCoreStartupParameter.bCPUThread)
{
s_perf_query_requested = true;
std::unique_lock<std::mutex> lk(s_perf_query_lock);
s_perf_query_cond.wait(lk, QueryResultIsReady);
}
else
g_perf_query->FlushResults();
}
return g_perf_query->GetQueryResult(type);
}
void VideoBackendHardware::InitializeShared()
{
@ -176,6 +217,7 @@ void VideoBackendHardware::InitializeShared()
s_swapRequested = 0;
s_efbAccessRequested = 0;
s_perf_query_requested = false;
s_FifoShuttingDown = 0;
memset((void*)&s_beginFieldArgs, 0, sizeof(s_beginFieldArgs));
memset(&s_accessEFBArgs, 0, sizeof(s_accessEFBArgs));
@ -238,6 +280,7 @@ void VideoFifo_CheckAsyncRequest()
{
VideoFifo_CheckSwapRequest();
VideoFifo_CheckEFBAccess();
VideoFifo_CheckPerfQueryRequest();
}
void VideoBackendHardware::Video_GatherPipeBursted()

View File

@ -22,7 +22,7 @@ namespace OSD
{
// On-screen message display
void AddMessage(const char* str, u32 ms);
void AddMessage(const char* str, u32 ms = 2000);
void DrawMessages(); // draw the current messages on the screen. Only call once per frame.
void ClearMessages();

View File

@ -0,0 +1,3 @@
#include "PerfQueryBase.h"
PerfQueryBase* g_perf_query = 0;

View File

@ -0,0 +1,54 @@
#ifndef _PERFQUERY_BASE_H_
#define _PERFQUERY_BASE_H_
#include "CommonTypes.h"
enum PerfQueryType
{
PQ_ZCOMP_INPUT_ZCOMPLOC = 0,
PQ_ZCOMP_OUTPUT_ZCOMPLOC,
PQ_ZCOMP_INPUT,
PQ_ZCOMP_OUTPUT,
PQ_BLEND_INPUT,
PQ_EFB_COPY_CLOCKS,
PQ_NUM_MEMBERS
};
enum PerfQueryGroup
{
PQG_ZCOMP_ZCOMPLOC,
PQG_ZCOMP,
PQG_EFB_COPY_CLOCKS,
PQG_NUM_MEMBERS,
};
class PerfQueryBase
{
public:
PerfQueryBase() {};
virtual ~PerfQueryBase() {}
// Begin querying the specified value for the following host GPU commands
virtual void EnableQuery(PerfQueryGroup type) {}
// Stop querying the specified value for the following host GPU commands
virtual void DisableQuery(PerfQueryGroup type) {}
// Reset query counters to zero and drop any pending queries
virtual void ResetQuery() {}
// Return the measured value for the specified query type
// NOTE: Called from CPU thread
virtual u32 GetQueryResult(PerfQueryType type) { return 0; }
// Request the value of any pending queries - causes a pipeline flush and thus should be used carefully!
virtual void FlushResults() {}
// True if there are no further pending query results
// NOTE: Called from CPU thread
virtual bool IsFlushed() const { return true; }
};
extern PerfQueryBase* g_perf_query;
#endif // _PERFQUERY_H_

View File

@ -28,10 +28,13 @@
#include "ConfigManager.h"
#include "PixelEngine.h"
#include "RenderBase.h"
#include "CommandProcessor.h"
#include "HW/ProcessorInterface.h"
#include "DLCache.h"
#include "State.h"
#include "PerfQueryBase.h"
namespace PixelEngine
{
@ -255,23 +258,59 @@ void Read16(u16& _uReturnValue, const u32 _iAddress)
break;
}
case PE_PERF_0L:
case PE_PERF_0H:
case PE_PERF_1L:
case PE_PERF_1H:
case PE_PERF_2L:
case PE_PERF_2H:
case PE_PERF_3L:
case PE_PERF_3H:
case PE_PERF_4L:
case PE_PERF_4H:
case PE_PERF_5L:
case PE_PERF_5H:
INFO_LOG(PIXELENGINE, "(r16) perf counter @ %08x", _iAddress);
// git r90a2096a24f4 (svn r3663) added the PE_PERF cases, without setting
// _uReturnValue to anything, this reverts to the previous behaviour which allows
// The timer in SMS:Scrubbing Serena Beach to countdown correctly
_uReturnValue = 1;
// NOTE(neobrain): only PE_PERF_ZCOMP_OUTPUT is implemented in D3D11, but the other values shouldn't be contradictionary to the value of that register (i.e. INPUT registers should always be greater or equal to their corresponding OUTPUT registers).
case PE_PERF_ZCOMP_INPUT_ZCOMPLOC_L:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_ZCOMP_INPUT_ZCOMPLOC) & 0xFFFF;
break;
case PE_PERF_ZCOMP_INPUT_ZCOMPLOC_H:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_ZCOMP_INPUT_ZCOMPLOC) >> 16;
break;
case PE_PERF_ZCOMP_OUTPUT_ZCOMPLOC_L:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_ZCOMP_OUTPUT_ZCOMPLOC) & 0xFFFF;
break;
case PE_PERF_ZCOMP_OUTPUT_ZCOMPLOC_H:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_ZCOMP_OUTPUT_ZCOMPLOC) >> 16;
break;
case PE_PERF_ZCOMP_INPUT_L:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_ZCOMP_INPUT) & 0xFFFF;
break;
case PE_PERF_ZCOMP_INPUT_H:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_ZCOMP_INPUT) >> 16;
break;
case PE_PERF_ZCOMP_OUTPUT_L:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_ZCOMP_OUTPUT) & 0xFFFF;
break;
case PE_PERF_ZCOMP_OUTPUT_H:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_ZCOMP_OUTPUT) >> 16;
break;
case PE_PERF_BLEND_INPUT_L:
// Super Mario Sunshine uses this register in episode 6 of Sirena Beach:
// The amount of remaining goop is determined by checking how many pixels reach the blending stage.
// Once this register falls below a particular value (around 0x90), the game regards the challenge finished.
// In very old builds, Dolphin only returned 0. That caused the challenge to be immediately finished without any goop being cleaned (the timer just didn't even start counting from 3:00:00).
// Later builds returned 1 for the high register. That caused the timer to actually count down, but made the challenge unbeatable because the game always thought you didn't clear any goop at all.
// Note that currently this functionality is only implemented in the D3D11 backend.
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_BLEND_INPUT) & 0xFFFF;
break;
case PE_PERF_BLEND_INPUT_H:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_BLEND_INPUT) >> 16;
break;
case PE_PERF_EFB_COPY_CLOCKS_L:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_EFB_COPY_CLOCKS) & 0xFFFF;
break;
case PE_PERF_EFB_COPY_CLOCKS_H:
_uReturnValue = g_video_backend->Video_GetQueryResult(PQ_EFB_COPY_CLOCKS) >> 16;
break;
default:

View File

@ -36,19 +36,20 @@ enum
PE_BBOX_TOP = 0x14, // Flip Top
PE_BBOX_BOTTOM = 0x16, // Flip Bottom
// These have not yet been RE:d. They are the perf counters.
PE_PERF_0L = 0x18,
PE_PERF_0H = 0x1a,
PE_PERF_1L = 0x1c,
PE_PERF_1H = 0x1e,
PE_PERF_2L = 0x20,
PE_PERF_2H = 0x22,
PE_PERF_3L = 0x24,
PE_PERF_3H = 0x26,
PE_PERF_4L = 0x28,
PE_PERF_4H = 0x2a,
PE_PERF_5L = 0x2c,
PE_PERF_5H = 0x2e,
// NOTE: Order not verified
// These indicate the number of quads that are being used as input/output for each particular stage
PE_PERF_ZCOMP_INPUT_ZCOMPLOC_L = 0x18,
PE_PERF_ZCOMP_INPUT_ZCOMPLOC_H = 0x1a,
PE_PERF_ZCOMP_OUTPUT_ZCOMPLOC_L = 0x1c,
PE_PERF_ZCOMP_OUTPUT_ZCOMPLOC_H = 0x1e,
PE_PERF_ZCOMP_INPUT_L = 0x20,
PE_PERF_ZCOMP_INPUT_H = 0x22,
PE_PERF_ZCOMP_OUTPUT_L = 0x24,
PE_PERF_ZCOMP_OUTPUT_H = 0x26,
PE_PERF_BLEND_INPUT_L = 0x28,
PE_PERF_BLEND_INPUT_H = 0x2a,
PE_PERF_EFB_COPY_CLOCKS_L = 0x2c,
PE_PERF_EFB_COPY_CLOCKS_H = 0x2e,
};
namespace PixelEngine

View File

@ -52,6 +52,15 @@ public:
Renderer();
virtual ~Renderer();
enum PixelPerfQuery {
PP_ZCOMP_INPUT_ZCOMPLOC,
PP_ZCOMP_OUTPUT_ZCOMPLOC,
PP_ZCOMP_INPUT,
PP_ZCOMP_OUTPUT,
PP_BLEND_INPUT,
PP_EFB_COPY_CLOCKS
};
virtual void SetColorMask() = 0;
virtual void SetBlendMode(bool forceUpdate) = 0;
virtual void SetScissorRect(const TargetRectangle& rc) = 0;

View File

@ -257,7 +257,9 @@ void VertexManager::Flush()
//if (g_nativeVertexFmt)
g_nativeVertexFmt->SetupVertexPointers();
g_renderer->ResumePixelPerf(false);
g_vertex_manager->Draw(stride, false);
g_renderer->PausePixelPerf(false);
// run through vertex groups again to set alpha
if (false == g_ActiveConfig.bDstAlphaPass && bpmem.dstalpha.enable && bpmem.blendmode.alphaupdate)

View File

@ -115,7 +115,7 @@ struct VideoConfig
int iAnaglyphStereoSeparation;
int iAnaglyphFocalAngle;
bool b3DVision;
// Hacks
bool bEFBAccessEnable;
bool bDlistCachingEnable;

View File

@ -196,6 +196,7 @@
<ClCompile Include="Src\OpcodeDecoding.cpp" />
<ClCompile Include="Src\OpenCL.cpp" />
<ClCompile Include="Src\OpenCL\OCLTextureDecoder.cpp" />
<ClCompile Include="Src\PerfQueryBase.cpp" />
<ClCompile Include="Src\PixelEngine.cpp" />
<ClCompile Include="Src\PixelShaderGen.cpp" />
<ClCompile Include="Src\PixelShaderManager.cpp" />
@ -244,6 +245,7 @@
<ClInclude Include="Src\OpcodeDecoding.h" />
<ClInclude Include="Src\OpenCL.h" />
<ClInclude Include="Src\OpenCL\OCLTextureDecoder.h" />
<ClInclude Include="Src\PerfQueryBase.h" />
<ClInclude Include="Src\PixelEngine.h" />
<ClInclude Include="Src\PixelShaderGen.h" />
<ClInclude Include="Src\PixelShaderManager.h" />

View File

@ -101,6 +101,9 @@
<ClCompile Include="Src\MainBase.cpp">
<Filter>Base</Filter>
</ClCompile>
<ClCompile Include="Src\PerfQueryBase.cpp">
<Filter>Base</Filter>
</ClCompile>
<ClCompile Include="Src\RenderBase.cpp">
<Filter>Base</Filter>
</ClCompile>
@ -237,6 +240,9 @@
<ClInclude Include="Src\MainBase.h">
<Filter>Base</Filter>
</ClInclude>
<ClInclude Include="Src\PerfQueryBase.h">
<Filter>Base</Filter>
</ClInclude>
<ClInclude Include="Src\RenderBase.h">
<Filter>Base</Filter>
</ClInclude>
@ -285,4 +291,4 @@
<UniqueIdentifier>{e2a527a2-ccc8-4ab8-a93e-dd2628c0f3b6}</UniqueIdentifier>
</Filter>
</ItemGroup>
</Project>
</Project>

View File

@ -199,6 +199,7 @@
<ClCompile Include="Src\LineGeometryShader.cpp" />
<ClCompile Include="Src\main.cpp" />
<ClCompile Include="Src\NativeVertexFormat.cpp" />
<ClCompile Include="Src\PerfQuery.cpp" />
<ClCompile Include="Src\PixelShaderCache.cpp" />
<ClCompile Include="Src\PointGeometryShader.cpp" />
<ClCompile Include="Src\PSTextureEncoder.cpp" />
@ -228,6 +229,7 @@
<ClInclude Include="Src\Globals.h" />
<ClInclude Include="Src\LineGeometryShader.h" />
<ClInclude Include="Src\main.h" />
<ClInclude Include="Src\PerfQuery.h" />
<ClInclude Include="Src\PixelShaderCache.h" />
<ClInclude Include="Src\PointGeometryShader.h" />
<ClInclude Include="Src\PSTextureEncoder.h" />

View File

@ -57,6 +57,9 @@
<ClCompile Include="Src\PointGeometryShader.cpp">
<Filter>Render</Filter>
</ClCompile>
<ClCompile Include="Src\PerfQuery.cpp">
<Filter>Render</Filter>
</ClCompile>
</ItemGroup>
<ItemGroup>
<ClInclude Include="Src\Globals.h" />
@ -117,6 +120,9 @@
<ClInclude Include="Src\PointGeometryShader.h">
<Filter>Render</Filter>
</ClInclude>
<ClInclude Include="Src\PerfQuery.h">
<Filter>Render</Filter>
</ClInclude>
</ItemGroup>
<ItemGroup>
<Filter Include="D3D">

View File

@ -0,0 +1,150 @@
#include "RenderBase.h"
#include "D3DBase.h"
#include "PerfQuery.h"
namespace DX11 {
PerfQuery::PerfQuery()
: m_query_read_pos()
, m_query_count()
{
for (int i = 0; i != ARRAYSIZE(m_query_buffer); ++i)
{
D3D11_QUERY_DESC qdesc = CD3D11_QUERY_DESC(D3D11_QUERY_OCCLUSION, 0);
D3D::device->CreateQuery(&qdesc, &m_query_buffer[i].query);
}
ResetQuery();
}
PerfQuery::~PerfQuery()
{
for (int i = 0; i != ARRAYSIZE(m_query_buffer); ++i)
{
// TODO: EndQuery?
m_query_buffer[i].query->Release();
}
}
void PerfQuery::EnableQuery(PerfQueryGroup type)
{
// Is this sane?
if (m_query_count > ARRAYSIZE(m_query_buffer) / 2)
WeakFlush();
if (ARRAYSIZE(m_query_buffer) == m_query_count)
{
// TODO
FlushOne();
ERROR_LOG(VIDEO, "flushed query buffer early!");
}
// start query
if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
{
auto& entry = m_query_buffer[(m_query_read_pos + m_query_count) % ARRAYSIZE(m_query_buffer)];
D3D::context->Begin(entry.query);
entry.query_type = type;
++m_query_count;
}
}
void PerfQuery::DisableQuery(PerfQueryGroup type)
{
// stop query
if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
{
auto& entry = m_query_buffer[(m_query_read_pos + m_query_count + ARRAYSIZE(m_query_buffer)-1) % ARRAYSIZE(m_query_buffer)];
D3D::context->End(entry.query);
}
}
void PerfQuery::ResetQuery()
{
m_query_count = 0;
std::fill_n(m_results, ARRAYSIZE(m_results), 0);
}
u32 PerfQuery::GetQueryResult(PerfQueryType type)
{
u32 result = 0;
if (type == PQ_ZCOMP_INPUT_ZCOMPLOC || type == PQ_ZCOMP_OUTPUT_ZCOMPLOC)
{
result = m_results[PQG_ZCOMP_ZCOMPLOC];
}
else if (type == PQ_ZCOMP_INPUT || type == PQ_ZCOMP_OUTPUT)
{
result = m_results[PQG_ZCOMP];
}
else if (type == PQ_BLEND_INPUT)
{
result = m_results[PQG_ZCOMP] + m_results[PQG_ZCOMP_ZCOMPLOC];
}
else if (type == PQ_EFB_COPY_CLOCKS)
{
result = m_results[PQG_EFB_COPY_CLOCKS];
}
return result / 4;
}
void PerfQuery::FlushOne()
{
auto& entry = m_query_buffer[m_query_read_pos];
UINT64 result = 0;
HRESULT hr = S_FALSE;
while (hr != S_OK)
{
// TODO: Might cause us to be stuck in an infinite loop!
hr = D3D::context->GetData(entry.query, &result, sizeof(result), 0);
}
// NOTE: Reported pixel metrics should be referenced to native resolution
m_results[entry.query_type] += (u64)result * EFB_WIDTH / g_renderer->GetTargetWidth() * EFB_HEIGHT / g_renderer->GetTargetHeight();
m_query_read_pos = (m_query_read_pos + 1) % ARRAYSIZE(m_query_buffer);
--m_query_count;
}
// TODO: could selectively flush things, but I don't think that will do much
void PerfQuery::FlushResults()
{
while (!IsFlushed())
FlushOne();
}
void PerfQuery::WeakFlush()
{
while (!IsFlushed())
{
auto& entry = m_query_buffer[m_query_read_pos];
UINT64 result = 0;
HRESULT hr = D3D::context->GetData(entry.query, &result, sizeof(result), D3D11_ASYNC_GETDATA_DONOTFLUSH);
if (hr == S_OK)
{
// NOTE: Reported pixel metrics should be referenced to native resolution
m_results[entry.query_type] += (u64)result * EFB_WIDTH / g_renderer->GetTargetWidth() * EFB_HEIGHT / g_renderer->GetTargetHeight();
m_query_read_pos = (m_query_read_pos + 1) % ARRAYSIZE(m_query_buffer);
--m_query_count;
}
else
{
break;
}
}
}
bool PerfQuery::IsFlushed() const
{
return 0 == m_query_count;
}
} // namespace

View File

@ -0,0 +1,46 @@
#ifndef _PERFQUERY_H_
#define _PERFQUERY_H_
#include "PerfQueryBase.h"
namespace DX11 {
class PerfQuery : public PerfQueryBase
{
public:
PerfQuery();
~PerfQuery();
void EnableQuery(PerfQueryGroup type);
void DisableQuery(PerfQueryGroup type);
void ResetQuery();
u32 GetQueryResult(PerfQueryType type);
void FlushResults();
bool IsFlushed() const;
private:
struct ActiveQuery
{
ID3D11Query* query;
PerfQueryGroup query_type;
};
void WeakFlush();
// Only use when non-empty
void FlushOne();
// when testing in SMS: 64 was too small, 128 was ok
static const int PERF_QUERY_BUFFER_SIZE = 512;
ActiveQuery m_query_buffer[PERF_QUERY_BUFFER_SIZE];
int m_query_read_pos;
// TODO: sloppy
volatile int m_query_count;
volatile u32 m_results[PQG_NUM_MEMBERS];
};
} // namespace
#endif // _PERFQUERY_H_

View File

@ -65,6 +65,7 @@ ID3D11RasterizerState* resetraststate = NULL;
static ID3D11Texture2D* s_screenshot_texture = NULL;
// GX pipeline state
struct
{

View File

@ -208,7 +208,6 @@ void VertexManager::Draw(UINT stride)
if (IndexGenerator::GetNumLines() > 0 || IndexGenerator::GetNumPoints() > 0)
((DX11::Renderer*)g_renderer)->RestoreCull();
}
void VertexManager::vFlush()
{
if (LocalVBuffer == s_pCurBufferPointer) return;
@ -274,8 +273,10 @@ void VertexManager::vFlush()
unsigned int stride = g_nativeVertexFmt->GetVertexStride();
g_nativeVertexFmt->SetupVertexPointers();
g_renderer->ApplyState(useDstAlpha);
g_perf_query->EnableQuery(bpmem.zcontrol.early_ztest ? PQG_ZCOMP_ZCOMPLOC : PQG_ZCOMP);
Draw(stride);
g_perf_query->DisableQuery(bpmem.zcontrol.early_ztest ? PQG_ZCOMP_ZCOMPLOC : PQG_ZCOMP);
GFX_DEBUGGER_PAUSE_AT(NEXT_FLUSH, true);

View File

@ -42,6 +42,7 @@
#include "D3DUtil.h"
#include "D3DBase.h"
#include "PerfQuery.h"
#include "PixelShaderCache.h"
#include "TextureCache.h"
#include "VertexManager.h"
@ -185,6 +186,7 @@ void VideoBackend::Video_Prepare()
g_renderer = new Renderer;
g_texture_cache = new TextureCache;
g_vertex_manager = new VertexManager;
g_perf_query = new PerfQuery;
VertexShaderCache::Init();
PixelShaderCache::Init();
D3D::InitUtils();
@ -227,6 +229,7 @@ void VideoBackend::Shutdown()
D3D::ShutdownUtils();
PixelShaderCache::Shutdown();
VertexShaderCache::Shutdown();
delete g_perf_query;
delete g_vertex_manager;
delete g_texture_cache;
delete g_renderer;

View File

@ -57,6 +57,7 @@
#include "ConfigManager.h"
#include "VideoBackend.h"
#include "PerfQueryBase.h"
namespace DX9
{
@ -97,8 +98,6 @@ void InitBackendInfo()
g_Config.backend_info.bSupports3DVision = true;
g_Config.backend_info.bSupportsDualSourceBlend = false;
g_Config.backend_info.bSupportsFormatReinterpretation = true;
g_Config.backend_info.bSupportsPixelLighting = C_PLIGHTS + 40 <= maxConstants && C_PMATERIALS + 4 <= maxConstants;
// adapters
@ -172,6 +171,7 @@ void VideoBackend::Video_Prepare()
g_vertex_manager = new VertexManager;
g_renderer = new Renderer;
g_texture_cache = new TextureCache;
g_perf_query = new PerfQueryBase;
// VideoCommon
BPInit();
Fifo_Init();
@ -209,6 +209,7 @@ void VideoBackend::Shutdown()
// internal interfaces
PixelShaderCache::Shutdown();
VertexShaderCache::Shutdown();
delete g_perf_query;
delete g_texture_cache;
delete g_renderer;
delete g_vertex_manager;

View File

@ -2,6 +2,7 @@ set(SRCS Src/FramebufferManager.cpp
Src/GLUtil.cpp
Src/main.cpp
Src/NativeVertexFormat.cpp
Src/PerfQuery.cpp
Src/PixelShaderCache.cpp
Src/PostProcessing.cpp
Src/RasterFont.cpp

View File

@ -200,6 +200,7 @@
<ClCompile Include="Src\GLUtil.cpp" />
<ClCompile Include="Src\main.cpp" />
<ClCompile Include="Src\NativeVertexFormat.cpp" />
<ClCompile Include="Src\PerfQuery.cpp" />
<ClCompile Include="Src\PixelShaderCache.cpp" />
<ClCompile Include="Src\PostProcessing.cpp" />
<ClCompile Include="Src\RasterFont.cpp" />
@ -222,6 +223,7 @@
<ClInclude Include="Src\Globals.h" />
<ClInclude Include="Src\GLUtil.h" />
<ClInclude Include="Src\main.h" />
<ClInclude Include="Src\PerfQuery.h" />
<ClInclude Include="Src\PixelShaderCache.h" />
<ClInclude Include="Src\PostProcessing.h" />
<ClInclude Include="Src\RasterFont.h" />

View File

@ -36,6 +36,9 @@
<ClCompile Include="Src\VertexShaderCache.cpp">
<Filter>Render</Filter>
</ClCompile>
<ClCompile Include="Src\PerfQuery.cpp">
<Filter>Render</Filter>
</ClCompile>
</ItemGroup>
<ItemGroup>
<ClInclude Include="Src\Globals.h" />
@ -72,6 +75,9 @@
<ClInclude Include="Src\VertexShaderCache.h">
<Filter>Render</Filter>
</ClInclude>
<ClInclude Include="Src\PerfQuery.h">
<Filter>Render</Filter>
</ClInclude>
</ItemGroup>
<ItemGroup>
<None Include="CMakeLists.txt" />

View File

@ -0,0 +1,133 @@
#include "RenderBase.h"
#include "GLUtil.h"
#include "PerfQuery.h"
namespace OGL
{
PerfQuery::PerfQuery()
: m_query_read_pos()
, m_query_count()
{
for (int i = 0; i != ARRAYSIZE(m_query_buffer); ++i)
glGenQueries(1, &m_query_buffer[i].query_id);
ResetQuery();
}
PerfQuery::~PerfQuery()
{
for (int i = 0; i != ARRAYSIZE(m_query_buffer); ++i)
glDeleteQueries(1, &m_query_buffer[i].query_id);
}
void PerfQuery::EnableQuery(PerfQueryGroup type)
{
// Is this sane?
if (m_query_count > ARRAYSIZE(m_query_buffer) / 2)
WeakFlush();
if (ARRAYSIZE(m_query_buffer) == m_query_count)
{
FlushOne();
//ERROR_LOG(VIDEO, "flushed query buffer early!");
}
// start query
if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
{
auto& entry = m_query_buffer[(m_query_read_pos + m_query_count) % ARRAYSIZE(m_query_buffer)];
glBeginQuery(GL_SAMPLES_PASSED, entry.query_id);
entry.query_type = type;
++m_query_count;
}
}
void PerfQuery::DisableQuery(PerfQueryGroup type)
{
// stop query
if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
{
glEndQuery(GL_SAMPLES_PASSED);
}
}
bool PerfQuery::IsFlushed() const
{
return 0 == m_query_count;
}
void PerfQuery::FlushOne()
{
auto& entry = m_query_buffer[m_query_read_pos];
GLuint result = 0;
glGetQueryObjectuiv(entry.query_id, GL_QUERY_RESULT, &result);
// NOTE: Reported pixel metrics should be referenced to native resolution
m_results[entry.query_type] += (u64)result * EFB_WIDTH / g_renderer->GetTargetWidth() * EFB_HEIGHT / g_renderer->GetTargetHeight();
m_query_read_pos = (m_query_read_pos + 1) % ARRAYSIZE(m_query_buffer);
--m_query_count;
}
// TODO: could selectively flush things, but I don't think that will do much
void PerfQuery::FlushResults()
{
while (!IsFlushed())
FlushOne();
}
void PerfQuery::WeakFlush()
{
while (!IsFlushed())
{
auto& entry = m_query_buffer[m_query_read_pos];
GLuint result = GL_FALSE;
glGetQueryObjectuiv(entry.query_id, GL_QUERY_RESULT_AVAILABLE, &result);
if (GL_TRUE == result)
{
FlushOne();
}
else
{
break;
}
}
}
void PerfQuery::ResetQuery()
{
m_query_count = 0;
std::fill_n(m_results, ARRAYSIZE(m_results), 0);
}
u32 PerfQuery::GetQueryResult(PerfQueryType type)
{
u32 result = 0;
if (type == PQ_ZCOMP_INPUT_ZCOMPLOC || type == PQ_ZCOMP_OUTPUT_ZCOMPLOC)
{
result = m_results[PQG_ZCOMP_ZCOMPLOC];
}
else if (type == PQ_ZCOMP_INPUT || type == PQ_ZCOMP_OUTPUT)
{
result = m_results[PQG_ZCOMP];
}
else if (type == PQ_BLEND_INPUT)
{
result = m_results[PQG_ZCOMP] + m_results[PQG_ZCOMP_ZCOMPLOC];
}
else if (type == PQ_EFB_COPY_CLOCKS)
{
result = m_results[PQG_EFB_COPY_CLOCKS];
}
return result / 4;
}
} // namespace

View File

@ -0,0 +1,46 @@
#ifndef _PERFQUERY_H_
#define _PERFQUERY_H_
#include "PerfQueryBase.h"
namespace OGL {
class PerfQuery : public PerfQueryBase
{
public:
PerfQuery();
~PerfQuery();
void EnableQuery(PerfQueryGroup type);
void DisableQuery(PerfQueryGroup type);
void ResetQuery();
u32 GetQueryResult(PerfQueryType type);
void FlushResults();
bool IsFlushed() const;
private:
struct ActiveQuery
{
GLuint query_id;
PerfQueryGroup query_type;
};
// when testing in SMS: 64 was too small, 128 was ok
static const int PERF_QUERY_BUFFER_SIZE = 512;
void WeakFlush();
// Only use when non-empty
void FlushOne();
// This contains gl query objects with unretrieved results.
ActiveQuery m_query_buffer[PERF_QUERY_BUFFER_SIZE];
int m_query_read_pos;
// TODO: sloppy
volatile int m_query_count;
volatile u32 m_results[PQG_NUM_MEMBERS];
};
} // namespace
#endif // _PERFQUERY_H_

View File

@ -1419,6 +1419,7 @@ void Renderer::SetDepthMode()
else
{
// if the test is disabled write is disabled too
// TODO: When PE performance metrics are being emulated via occlusion queries, we should (probably?) enable depth test with depth function ALWAYS here
glDisable(GL_DEPTH_TEST);
glDepthMask(GL_FALSE);
}

View File

@ -40,6 +40,7 @@
#include "OpcodeDecoding.h"
#include "FileUtil.h"
#include "Debugger.h"
#include "PerfQueryBase.h"
#include "main.h"
@ -217,7 +218,10 @@ void VertexManager::vFlush()
if (ps) PixelShaderCache::SetCurrentShader(ps->glprogid); // Lego Star Wars crashes here.
if (vs) VertexShaderCache::SetCurrentShader(vs->glprogid);
g_perf_query->EnableQuery(bpmem.zcontrol.early_ztest ? PQG_ZCOMP_ZCOMPLOC : PQG_ZCOMP);
Draw();
g_perf_query->DisableQuery(bpmem.zcontrol.early_ztest ? PQG_ZCOMP_ZCOMPLOC : PQG_ZCOMP);
//ERROR_LOG(VIDEO, "PerfQuery result: %d", g_perf_query->GetQueryResult(bpmem.zcontrol.early_ztest ? PQ_ZCOMP_OUTPUT_ZCOMPLOC : PQ_ZCOMP_OUTPUT));
// run through vertex groups again to set alpha
if (useDstAlpha && !dualSourcePossible)

View File

@ -92,6 +92,7 @@ Make AA apply instantly during gameplay if possible
#include "FramebufferManager.h"
#include "Core.h"
#include "Host.h"
#include "PerfQuery.h"
#include "VideoState.h"
#include "VideoBackend.h"
@ -194,6 +195,7 @@ void VideoBackend::Video_Prepare()
BPInit();
g_vertex_manager = new VertexManager;
g_perf_query = new PerfQuery;
Fifo_Init(); // must be done before OpcodeDecoder_Init()
OpcodeDecoder_Init();
VertexShaderCache::Init();

View File

@ -90,6 +90,21 @@ void SWBPWritten(int address, int newvalue)
SWPixelEngine::pereg.boxBottom = newvalue >> 10;
SWPixelEngine::pereg.boxTop = newvalue & 0x3ff;
break;
case BPMEM_CLEAR_PIXEL_PERF:
// TODO: I didn't test if the value written to this register affects the amount of cleared registers
SWPixelEngine::pereg.perfZcompInputZcomplocLo = 0;
SWPixelEngine::pereg.perfZcompInputZcomplocHi = 0;
SWPixelEngine::pereg.perfZcompOutputZcomplocLo = 0;
SWPixelEngine::pereg.perfZcompOutputZcomplocHi = 0;
SWPixelEngine::pereg.perfZcompInputLo = 0;
SWPixelEngine::pereg.perfZcompInputHi = 0;
SWPixelEngine::pereg.perfZcompOutputLo = 0;
SWPixelEngine::pereg.perfZcompOutputHi = 0;
SWPixelEngine::pereg.perfBlendInputLo = 0;
SWPixelEngine::pereg.perfBlendInputHi = 0;
SWPixelEngine::pereg.perfEfbCopyClocksLo = 0;
SWPixelEngine::pereg.perfEfbCopyClocksHi = 0;
break;
case BPMEM_LOADTLUT0: // This one updates bpmem.tlutXferSrc, no need to do anything here.
break;
case BPMEM_LOADTLUT1: // Load a Texture Look Up Table

View File

@ -23,6 +23,7 @@
#include "BPMemLoader.h"
#include "XFMemLoader.h"
#include "Tev.h"
#include "SWPixelEngine.h"
#include "SWStatistics.h"
#include "SWVideoConfig.h"
@ -149,9 +150,15 @@ inline void Draw(s32 x, s32 y, s32 xi, s32 yi)
if (bpmem.zcontrol.early_ztest && bpmem.zmode.testenable && g_SWVideoConfig.bZComploc)
{
// early z
if (!EfbInterface::ZCompare(x, y, z))
return;
// TODO: Test if perf regs are incremented even if test is disabled
SWPixelEngine::pereg.IncZInputQuadCount(true);
if (bpmem.zmode.testenable)
{
// early z
if (!EfbInterface::ZCompare(x, y, z))
return;
}
SWPixelEngine::pereg.IncZOutputQuadCount(true);
}
RasterBlockPixel& pixel = rasterBlock.Pixel[xi][yi];

View File

@ -79,7 +79,7 @@ void Read16(u16& _uReturnValue, const u32 _iAddress)
u16 address = _iAddress & 0xFFF;
if (address <= 0x16)
if (address <= 0x2e)
_uReturnValue = ((u16*)&pereg)[address >> 1];
}
@ -111,7 +111,7 @@ void Write16(const u16 _iValue, const u32 _iAddress)
}
break;
default:
if (address <= 0x16)
if (address <= 0x2e)
((u16*)&pereg)[address >> 1] = _iValue;
break;
}

View File

@ -38,6 +38,21 @@ namespace SWPixelEngine
PE_BBOX_RIGHT = 0x012, // Flip Right
PE_BBOX_TOP = 0x014, // Flip Top
PE_BBOX_BOTTOM = 0x016, // Flip Bottom
// NOTE: Order not verified
// These indicate the number of quads that are being used as input/output for each particular stage
PE_PERF_ZCOMP_INPUT_ZCOMPLOC_L = 0x18,
PE_PERF_ZCOMP_INPUT_ZCOMPLOC_H = 0x1a,
PE_PERF_ZCOMP_OUTPUT_ZCOMPLOC_L = 0x1c,
PE_PERF_ZCOMP_OUTPUT_ZCOMPLOC_H = 0x1e,
PE_PERF_ZCOMP_INPUT_L = 0x20,
PE_PERF_ZCOMP_INPUT_H = 0x22,
PE_PERF_ZCOMP_OUTPUT_L = 0x24,
PE_PERF_ZCOMP_OUTPUT_H = 0x26,
PE_PERF_BLEND_INPUT_L = 0x28,
PE_PERF_BLEND_INPUT_H = 0x2a,
PE_PERF_EFB_COPY_CLOCKS_L = 0x2c,
PE_PERF_EFB_COPY_CLOCKS_H = 0x2e,
};
union UPEZConfReg
@ -125,10 +140,72 @@ namespace SWPixelEngine
UPECtrlReg ctrl;
u16 unk0;
u16 token;
u16 boxLeft;
u16 boxRight;
u16 boxTop;
u16 boxBottom;
u16 perfZcompInputZcomplocLo;
u16 perfZcompInputZcomplocHi;
u16 perfZcompOutputZcomplocLo;
u16 perfZcompOutputZcomplocHi;
u16 perfZcompInputLo;
u16 perfZcompInputHi;
u16 perfZcompOutputLo;
u16 perfZcompOutputHi;
u16 perfBlendInputLo;
u16 perfBlendInputHi;
u16 perfEfbCopyClocksLo;
u16 perfEfbCopyClocksHi;
// NOTE: hardware doesn't process individual pixels but quads instead. Current software renderer architecture works on pixels though, so we have this "quad" hack here to only increment the registers on every fourth rendered pixel
void IncZInputQuadCount(bool early_ztest)
{
static int quad = 0;
if (++quad != 3)
return;
quad = 0;
if (early_ztest)
{
if (++perfZcompInputZcomplocLo == 0)
perfZcompInputZcomplocHi++;
}
else
{
if (++perfZcompInputLo == 0)
perfZcompInputHi++;
}
}
void IncZOutputQuadCount(bool early_ztest)
{
static int quad = 0;
if (++quad != 3)
return;
quad = 0;
if (early_ztest)
{
if (++perfZcompOutputZcomplocLo == 0)
perfZcompOutputZcomplocHi++;
}
else
{
if (++perfZcompOutputLo == 0)
perfZcompOutputHi++;
}
}
void IncBlendInputQuadCount()
{
static int quad = 0;
if (++quad != 3)
return;
quad = 0;
if (++perfBlendInputLo == 0)
perfBlendInputHi++;
}
};
extern PEReg pereg;

View File

@ -225,6 +225,12 @@ u32 VideoSoftware::Video_AccessEFB(EFBAccessType type, u32 x, u32 y, u32 InputDa
return value;
}
u32 VideoSoftware::Video_GetQueryResult(PerfQueryType type)
{
// TODO:
return 0;
}
bool VideoSoftware::Video_Screenshot(const char *_szFilename)
{
return false;

View File

@ -20,6 +20,7 @@
#include "Tev.h"
#include "EfbInterface.h"
#include "TextureSampler.h"
#include "SWPixelEngine.h"
#include "SWStatistics.h"
#include "SWVideoConfig.h"
#include "DebugUtil.h"
@ -787,8 +788,13 @@ void Tev::Draw()
bool late_ztest = !bpmem.zcontrol.early_ztest || !g_SWVideoConfig.bZComploc;
if (late_ztest && bpmem.zmode.testenable)
{
if (!EfbInterface::ZCompare(Position[0], Position[1], Position[2]))
return;
// TODO: Check against hw if these values get incremented even if depth testing is disabled
SWPixelEngine::pereg.IncZInputQuadCount(false);
if (!EfbInterface::ZCompare(Position[0], Position[1], Position[2]))
return;
SWPixelEngine::pereg.IncZOutputQuadCount(false);
}
#if ALLOW_TEV_DUMPS
@ -812,6 +818,7 @@ void Tev::Draw()
#endif
INCSTAT(swstats.thisFrame.tevPixelsOut);
SWPixelEngine::pereg.IncBlendInputQuadCount();
EfbInterface::BlendTev(Position[0], Position[1], output);
}

View File

@ -26,7 +26,9 @@ class VideoSoftware : public VideoBackend
void Video_ExitLoop();
void Video_BeginField(u32, FieldType, u32, u32);
void Video_EndField();
u32 Video_AccessEFB(EFBAccessType, u32, u32, u32);
u32 Video_GetQueryResult(PerfQueryType type);
void Video_AddMessage(const char* pstr, unsigned int milliseconds);
void Video_ClearMessages();