ShuriZma
/
dolphin
mirror of https://github.com/dolphin-emu/dolphin.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'primitive_restart'

This commit is contained in:
degasus 2013-04-11 14:01:58 +02:00
parent 3c87512180 26b428539a
commit 7e630ba920
12 changed files with 232 additions and 80 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

182
Source/Core/VideoCommon/Src/IndexGenerator.cpp
View File

@ -18,17 +18,9 @@
#include <cstddef>
#include "Common.h"
#include "VideoConfig.h"
#include "IndexGenerator.h"
/*
*
QUAD simulator
0 1 4 5
3 2 7 6
012023 147172 ...
*/
//Init
u16 *IndexGenerator::Tptr;
u16 *IndexGenerator::BASETptr;
@ -41,6 +33,32 @@ u32 IndexGenerator::numL;
u32 IndexGenerator::numP;
u32 IndexGenerator::index;
static const u16 s_primitive_restart = -1;
static void (*primitive_table[8])(u32);
void IndexGenerator::Init()
{
if(g_Config.backend_info.bSupportsPrimitiveRestart)
{
primitive_table[0] = IndexGenerator::AddQuads<true>;
primitive_table[2] = IndexGenerator::AddList<true>;
primitive_table[3] = IndexGenerator::AddStrip<true>;
primitive_table[4] = IndexGenerator::AddFan<true>;
}
else
{
primitive_table[0] = IndexGenerator::AddQuads<false>;
primitive_table[2] = IndexGenerator::AddList<false>;
primitive_table[3] = IndexGenerator::AddStrip<false>;
primitive_table[4] = IndexGenerator::AddFan<false>;
}
primitive_table[1] = NULL;
primitive_table[5] = &IndexGenerator::AddLineList;
primitive_table[6] = &IndexGenerator::AddLineStrip;
primitive_table[7] = &IndexGenerator::AddPoints;
}
void IndexGenerator::Start(u16* Triangleptr, u16* Lineptr, u16* Pointptr)
{
Tptr = Triangleptr;
@ -57,58 +75,49 @@ void IndexGenerator::Start(u16* Triangleptr, u16* Lineptr, u16* Pointptr)
void IndexGenerator::AddIndices(int primitive, u32 numVerts)
{
//switch (primitive)
//{
//case GX_DRAW_QUADS: IndexGenerator::AddQuads(numVerts); break;
//case GX_DRAW_TRIANGLES: IndexGenerator::AddList(numVerts); break;
//case GX_DRAW_TRIANGLE_STRIP: IndexGenerator::AddStrip(numVerts); break;
//case GX_DRAW_TRIANGLE_FAN: IndexGenerator::AddFan(numVerts); break;
//case GX_DRAW_LINES: IndexGenerator::AddLineList(numVerts); break;
//case GX_DRAW_LINE_STRIP: IndexGenerator::AddLineStrip(numVerts); break;
//case GX_DRAW_POINTS: IndexGenerator::AddPoints(numVerts); break;
//}
static void (*const primitive_table[])(u32) =
{
IndexGenerator::AddQuads,
NULL,
IndexGenerator::AddList,
IndexGenerator::AddStrip,
IndexGenerator::AddFan,
IndexGenerator::AddLineList,
IndexGenerator::AddLineStrip,
IndexGenerator::AddPoints,
};
primitive_table[primitive](numVerts);
index += numVerts;
}
// Triangles
__forceinline void IndexGenerator::WriteTriangle(u32 index1, u32 index2, u32 index3)
template <bool pr> __forceinline void IndexGenerator::WriteTriangle(u32 index1, u32 index2, u32 index3)
{
*Tptr++ = index1;
*Tptr++ = index2;
*Tptr++ = index3;
if(pr)
*Tptr++ = s_primitive_restart;
++numT;
}
void IndexGenerator::AddList(u32 const numVerts)
template <bool pr> void IndexGenerator::AddList(u32 const numVerts)
{
auto const numTris = numVerts / 3;
for (u32 i = 0; i != numTris; ++i)
{
WriteTriangle(index + i * 3, index + i * 3 + 1, index + i * 3 + 2);
WriteTriangle<pr>(index + i * 3, index + i * 3 + 1, index + i * 3 + 2);
}
u32 remainingVerts = numVerts - numTris*3;
if(remainingVerts)
ERROR_LOG(VIDEO, "AddList: unknown count of vertices found");
}
void IndexGenerator::AddStrip(u32 const numVerts)
template <bool pr> void IndexGenerator::AddStrip(u32 const numVerts)
{
if(pr) {
for (u32 i = 0; i < numVerts; ++i)
{
*Tptr++ = index + i;
}
*Tptr++ = s_primitive_restart;
numT += numVerts - 2;
} else {
bool wind = false;
for (u32 i = 2; i < numVerts; ++i)
{
WriteTriangle(
WriteTriangle<pr>(
index + i - 2,
index + i - !wind,
index + i - wind);
@ -116,24 +125,101 @@ void IndexGenerator::AddStrip(u32 const numVerts)
wind ^= true;
}
}
}
void IndexGenerator::AddFan(u32 numVerts)
/**
* FAN simulator:
*
* 2---3
* / \ / \
* 1---0---4
*
* would generate this triangles:
* 012, 023, 034
*
* rotated (for better striping):
* 120, 302, 034
*
* as odd ones have to winded, following strip is fine:
* 12034
*
* so we use 6 indices for 3 triangles
*/
template <bool pr> void IndexGenerator::AddFan(u32 numVerts)
{
for (u32 i = 2; i < numVerts; ++i)
{
WriteTriangle(index, index + i - 1, index + i);
u32 i = 2;
if(pr) {
for(; i<=numVerts-3; i+=3) {
*Tptr++ = index + i - 1;
*Tptr++ = index + i + 0;
*Tptr++ = index;
*Tptr++ = index + i + 1;
*Tptr++ = index + i + 2;
*Tptr++ = s_primitive_restart;
numT += 3;
}
for(; i<=numVerts-2; i+=2) {
*Tptr++ = index + i - 1;
*Tptr++ = index + i + 0;
*Tptr++ = index;
*Tptr++ = index + i + 1;
*Tptr++ = s_primitive_restart;
numT += 2;
}
}
void IndexGenerator::AddQuads(u32 numVerts)
for (; i < numVerts; ++i)
{
WriteTriangle<pr>(index, index + i - 1, index + i);
}
}
/*
* QUAD simulator
*
* 0---1 4---5
* |\ | |\ |
* | \ | | \ |
* | \| | \|
* 3---2 7---6
*
* 012,023, 456,467 ...
* or 120,302, 564,746
* or as strip: 1203, 5647
*
* Warning:
* A simple triangle has to be rendered for three vertices.
* ZWW do this for sun rays
*/
template <bool pr> void IndexGenerator::AddQuads(u32 numVerts)
{
auto const numQuads = numVerts / 4;
for (u32 i = 0; i != numQuads; ++i)
{
WriteTriangle(index + i * 4, index + i * 4 + 1, index + i * 4 + 2);
WriteTriangle(index + i * 4, index + i * 4 + 2, index + i * 4 + 3);
if(pr) {
*Tptr++ = index + i * 4 + 1;
*Tptr++ = index + i * 4 + 2;
*Tptr++ = index + i * 4 + 0;
*Tptr++ = index + i * 4 + 3;
*Tptr++ = s_primitive_restart;
numT += 2;
} else {
WriteTriangle<pr>(index + i * 4, index + i * 4 + 1, index + i * 4 + 2);
WriteTriangle<pr>(index + i * 4, index + i * 4 + 2, index + i * 4 + 3);
}
}
// three vertices remaining, so render a triangle
u32 remainingVerts = numVerts - numQuads*4;
if(remainingVerts == 3)
{
WriteTriangle<pr>(index+numVerts-3, index+numVerts-2, index+numVerts-1);
}
else if(remainingVerts)
ERROR_LOG(VIDEO, "AddQuads: unknown count of vertices found");
}
// Lines
void IndexGenerator::AddLineList(u32 numVerts)
@ -145,8 +231,14 @@ void IndexGenerator::AddLineList(u32 numVerts)
*Lptr++ = index + i * 2 + 1;
++numL;
}
u32 remainingVerts = numVerts - numLines*2;
if(remainingVerts)
ERROR_LOG(VIDEO, "AddLineList: unknown count of vertices found");
}
// shouldn't be used as strips as LineLists are much more common
// so converting them to lists
void IndexGenerator::AddLineStrip(u32 numVerts)
{
for (u32 i = 1; i < numVerts; ++i)
@ -170,6 +262,6 @@ void IndexGenerator::AddPoints(u32 numVerts)
u32 IndexGenerator::GetRemainingIndices()
{
u32 max_index = 65535;
u32 max_index = 65534; // -1 is reserved for primitive restart (ogl + dx11)
return max_index - index;
}

11
Source/Core/VideoCommon/Src/IndexGenerator.h
View File

@ -26,6 +26,7 @@ class IndexGenerator
{
public:
// Init
static void Init();
static void Start(u16 *Triangleptr,u16 *Lineptr,u16 *Pointptr);
static void AddIndices(int primitive, u32 numVertices);
@ -54,10 +55,10 @@ public:
*/
private:
// Triangles
static void AddList(u32 numVerts);
static void AddStrip(u32 numVerts);
static void AddFan(u32 numVerts);
static void AddQuads(u32 numVerts);
template <bool pr> static void AddList(u32 numVerts);
template <bool pr> static void AddStrip(u32 numVerts);
template <bool pr> static void AddFan(u32 numVerts);
template <bool pr> static void AddQuads(u32 numVerts);
// Lines
static void AddLineList(u32 numVerts);
@ -66,7 +67,7 @@ private:
// Points
static void AddPoints(u32 numVerts);
static void WriteTriangle(u32 index1, u32 index2, u32 index3);
template <bool pr> static void WriteTriangle(u32 index1, u32 index2, u32 index3);
static u16 *Tptr;
static u16 *BASETptr;

4
Source/Core/VideoCommon/Src/VertexLoader.cpp
View File

@ -659,6 +659,10 @@ void VertexLoader::SetVAT(u32 _group0, u32 _group1, u32 _group2)
m_VtxAttr.texCoord[7].Elements = vat.g2.Tex7CoordElements;
m_VtxAttr.texCoord[7].Format = vat.g2.Tex7CoordFormat;
m_VtxAttr.texCoord[7].Frac = vat.g2.Tex7Frac;
if(!m_VtxAttr.ByteDequant) {
ERROR_LOG(VIDEO, "ByteDequant is set to zero");
}
};
void VertexLoader::AppendToString(std::string *dest) const

26
Source/Core/VideoCommon/Src/VertexManagerBase.cpp
View File

@ -74,6 +74,31 @@ bool VertexManager::IsFlushed() const
u32 VertexManager::GetRemainingIndices(int primitive)
{
if(g_Config.backend_info.bSupportsPrimitiveRestart) {
switch (primitive)
{
case GX_DRAW_QUADS:
return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 5 * 4;
case GX_DRAW_TRIANGLES:
return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 4 * 3;
case GX_DRAW_TRIANGLE_STRIP:
return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 1 - 1;
case GX_DRAW_TRIANGLE_FAN:
return (MAXIBUFFERSIZE - IndexGenerator::GetTriangleindexLen()) / 6 * 4 + 1;
case GX_DRAW_LINES:
return (MAXIBUFFERSIZE - IndexGenerator::GetLineindexLen());
case GX_DRAW_LINE_STRIP:
return (MAXIBUFFERSIZE - IndexGenerator::GetLineindexLen()) / 2 + 1;
case GX_DRAW_POINTS:
return (MAXIBUFFERSIZE - IndexGenerator::GetPointindexLen());
default:
return 0;
}
} else {
switch (primitive)
{
case GX_DRAW_QUADS:
@ -97,6 +122,7 @@ u32 VertexManager::GetRemainingIndices(int primitive)
return 0;
}
}
}
void VertexManager::AddVertices(int primitive, u32 numVertices)
{

1
Source/Core/VideoCommon/Src/VideoConfig.h
View File

@ -165,6 +165,7 @@ struct VideoConfig
bool bSupportsDualSourceBlend; // only supported by D3D11 and OpenGL
bool bSupportsFormatReinterpretation;
bool bSupportsPixelLighting;
bool bSupportsPrimitiveRestart;
bool bSupportsSeparateAlphaFunction;
bool bSupportsGLSLUBO; // needed by pixelShaderGen, so must stay in videoCommon
} backend_info;

2
Source/Plugins/Plugin_VideoDX11/Src/VertexManager.cpp
View File

@ -154,7 +154,7 @@ void VertexManager::Draw(UINT stride)
if (IndexGenerator::GetNumTriangles() > 0)
{
D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
D3D::context->DrawIndexed(IndexGenerator::GetTriangleindexLen(), m_triangle_draw_index, 0);
INCSTAT(stats.thisFrame.numIndexedDrawCalls);
}

3
Source/Plugins/Plugin_VideoDX11/Src/main.cpp
View File

@ -35,6 +35,7 @@
#include "Debugger/DebuggerPanel.h"
#include "DLCache.h"
#include "EmuWindow.h"
#include "IndexGenerator.h"
#include "FileUtil.h"
#include "Globals.h"
#include "IniFile.h"
@ -96,6 +97,7 @@ void InitBackendInfo()
g_Config.backend_info.bSupportsDualSourceBlend = true;
g_Config.backend_info.bSupportsFormatReinterpretation = true;
g_Config.backend_info.bSupportsPixelLighting = true;
g_Config.backend_info.bSupportsPrimitiveRestart = true;
IDXGIFactory* factory;
IDXGIAdapter* ad;
@ -192,6 +194,7 @@ void VideoBackend::Video_Prepare()
// VideoCommon
BPInit();
Fifo_Init();
IndexGenerator::Init();
VertexLoaderManager::Init();
OpcodeDecoder_Init();
VertexShaderManager::Init();

3
Source/Plugins/Plugin_VideoDX9/Src/main.cpp
View File

@ -51,6 +51,7 @@
#include "VideoState.h"
#include "Render.h"
#include "DLCache.h"
#include "IndexGenerator.h"
#include "IniFile.h"
#include "Core.h"
#include "Host.h"
@ -102,6 +103,7 @@ void InitBackendInfo()
g_Config.backend_info.bUseRGBATextures = false;
g_Config.backend_info.bUseMinimalMipCount = true;
g_Config.backend_info.bSupports3DVision = true;
g_Config.backend_info.bSupportsPrimitiveRestart = false; // TODO: figure out if it does
g_Config.backend_info.bSupportsSeparateAlphaFunction = device_caps.PrimitiveMiscCaps & D3DPMISCCAPS_SEPARATEALPHABLEND;
// Dual source blend disabled by default until a proper method to test for support is found
g_Config.backend_info.bSupportsDualSourceBlend = false;
@ -183,6 +185,7 @@ void VideoBackend::Video_Prepare()
// VideoCommon
BPInit();
Fifo_Init();
IndexGenerator::Init();
VertexLoaderManager::Init();
OpcodeDecoder_Init();
VertexShaderManager::Init();

19
Source/Plugins/Plugin_VideoOGL/Src/Render.cpp
View File

@ -338,6 +338,7 @@ Renderer::Renderer()
g_Config.backend_info.bSupportsDualSourceBlend = GLEW_ARB_blend_func_extended;
g_Config.backend_info.bSupportsGLSLUBO = GLEW_ARB_uniform_buffer_object;
g_Config.backend_info.bSupportsPrimitiveRestart = GLEW_VERSION_3_1 || GLEW_NV_primitive_restart;
g_ogl_config.bSupportsGLSLCache = GLEW_ARB_get_program_binary;
g_ogl_config.bSupportsGLPinnedMemory = GLEW_AMD_pinned_memory;
@ -345,6 +346,7 @@ Renderer::Renderer()
g_ogl_config.bSupportsGLBaseVertex = GLEW_ARB_draw_elements_base_vertex;
g_ogl_config.bSupportCoverageMSAA = GLEW_NV_framebuffer_multisample_coverage;
g_ogl_config.bSupportSampleShading = GLEW_ARB_sample_shading;
g_ogl_config.bSupportOGL31 = GLEW_VERSION_3_1;
g_ogl_config.gl_vendor = (const char*)glGetString(GL_VENDOR);
g_ogl_config.gl_renderer = (const char*)glGetString(GL_RENDERER);
@ -397,9 +399,10 @@ Renderer::Renderer()
g_ogl_config.gl_renderer,
g_ogl_config.gl_version).c_str(), 5000);
WARN_LOG(VIDEO,"Missing OGL Extensions: %s%s%s%s%s%s%s%s",
WARN_LOG(VIDEO,"Missing OGL Extensions: %s%s%s%s%s%s%s%s%s",
g_ActiveConfig.backend_info.bSupportsDualSourceBlend ? "" : "DualSourceBlend ",
g_ActiveConfig.backend_info.bSupportsGLSLUBO ? "" : "UniformBuffer ",
g_ActiveConfig.backend_info.bSupportsPrimitiveRestart ? "" : "PrimitiveRestart ",
g_ogl_config.bSupportsGLPinnedMemory ? "" : "PinnedMemory ",
g_ogl_config.bSupportsGLSLCache ? "" : "ShaderCache ",
g_ogl_config.bSupportsGLBaseVertex ? "" : "BaseVertex ",
@ -476,6 +479,20 @@ Renderer::Renderer()
glBlendColor(0, 0, 0, 0.5f);
glClearDepth(1.0f);
if(g_ActiveConfig.backend_info.bSupportsPrimitiveRestart)
{
if(g_ogl_config.bSupportOGL31)
{
glEnable(GL_PRIMITIVE_RESTART);
glPrimitiveRestartIndex(65535);
}
else
{
glEnableClientState(GL_PRIMITIVE_RESTART_NV);
glPrimitiveRestartIndexNV(65535);
}
}
UpdateActiveConfig();
}

1
Source/Plugins/Plugin_VideoOGL/Src/Render.h
View File

@ -24,6 +24,7 @@ extern struct VideoConfig {
bool bSupportCoverageMSAA;
bool bSupportSampleShading;
GLSL_VERSION eSupportedGLSLVersion;
bool bSupportOGL31;
const char *gl_vendor;
const char *gl_renderer;

6
Source/Plugins/Plugin_VideoOGL/Src/VertexManager.cpp
View File

@ -124,10 +124,12 @@ void VertexManager::Draw(u32 stride)
u32 triangle_index_size = IndexGenerator::GetTriangleindexLen();
u32 line_index_size = IndexGenerator::GetLineindexLen();
u32 point_index_size = IndexGenerator::GetPointindexLen();
GLenum triangle_mode = g_ActiveConfig.backend_info.bSupportsPrimitiveRestart?GL_TRIANGLE_STRIP:GL_TRIANGLES;
if(g_ogl_config.bSupportsGLBaseVertex) {
if (triangle_index_size > 0)
{
glDrawElementsBaseVertex(GL_TRIANGLES, triangle_index_size, GL_UNSIGNED_SHORT, (u8*)NULL+s_offset[0], s_baseVertex);
glDrawElementsBaseVertex(triangle_mode, triangle_index_size, GL_UNSIGNED_SHORT, (u8*)NULL+s_offset[0], s_baseVertex);
INCSTAT(stats.thisFrame.numIndexedDrawCalls);
}
if (line_index_size > 0)
@ -143,7 +145,7 @@ void VertexManager::Draw(u32 stride)
} else {
if (triangle_index_size > 0)
{
glDrawElements(GL_TRIANGLES, triangle_index_size, GL_UNSIGNED_SHORT, (u8*)NULL+s_offset[0]);
glDrawElements(triangle_mode, triangle_index_size, GL_UNSIGNED_SHORT, (u8*)NULL+s_offset[0]);
INCSTAT(stats.thisFrame.numIndexedDrawCalls);
}
if (line_index_size > 0)

2
Source/Plugins/Plugin_VideoOGL/Src/main.cpp
View File

@ -96,6 +96,7 @@ Make AA apply instantly during gameplay if possible
#include "PerfQuery.h"
#include "VideoState.h"
#include "IndexGenerator.h"
#include "VideoBackend.h"
#include "ConfigManager.h"
@ -199,6 +200,7 @@ void VideoBackend::Video_Prepare()
g_perf_query = new PerfQuery;
Fifo_Init(); // must be done before OpcodeDecoder_Init()
OpcodeDecoder_Init();
IndexGenerator::Init();
VertexShaderManager::Init();
PixelShaderManager::Init();
ProgramShaderCache::Init();