OpenGL based software renderer hardware accel now works for most things

This commit is contained in:
Magnus Norddahl 2016-10-11 00:03:46 +02:00
parent f4308b3184
commit 682b040b97
4 changed files with 99 additions and 44 deletions

View File

@ -70,6 +70,7 @@
#include "gl/utility/gl_clock.h" #include "gl/utility/gl_clock.h"
#include "gl/utility/gl_templates.h" #include "gl/utility/gl_templates.h"
#include "gl/gl_functions.h" #include "gl/gl_functions.h"
#include "gl_debug.h"
CVAR(Int, gl_showpacks, 0, 0) CVAR(Int, gl_showpacks, 0, 0)
#ifndef WIN32 // Defined in fb_d3d9 for Windows #ifndef WIN32 // Defined in fb_d3d9 for Windows
@ -131,6 +132,9 @@ OpenGLSWFrameBuffer::OpenGLSWFrameBuffer(void *hMonitor, int width, int height,
gl_LoadExtensions(); gl_LoadExtensions();
Super::InitializeState(); Super::InitializeState();
Debug = std::make_shared<FGLDebug>();
Debug->Update();
// SetVSync needs to be at the very top to workaround a bug in Nvidia's OpenGL driver. // SetVSync needs to be at the very top to workaround a bug in Nvidia's OpenGL driver.
// If wglSwapIntervalEXT is called after glBindFramebuffer in a frame the setting is not changed! // If wglSwapIntervalEXT is called after glBindFramebuffer in a frame the setting is not changed!
//SetVSync(vid_vsync); //SetVSync(vid_vsync);
@ -160,6 +164,7 @@ OpenGLSWFrameBuffer::OpenGLSWFrameBuffer(void *hMonitor, int width, int height,
ScreenWipe = nullptr; ScreenWipe = nullptr;
InScene = false; InScene = false;
QuadExtra = new BufferedTris[MAX_QUAD_BATCH]; QuadExtra = new BufferedTris[MAX_QUAD_BATCH];
memset(QuadExtra, 0, sizeof(BufferedTris) * MAX_QUAD_BATCH);
Atlases = nullptr; Atlases = nullptr;
PixelDoubling = 0; PixelDoubling = 0;
@ -288,10 +293,10 @@ bool OpenGLSWFrameBuffer::CreatePixelShader(FString vertexsrc, FString fragments
if (status != GL_FALSE) { errorShader = shader->FragmentShader; glGetShaderiv(shader->FragmentShader, GL_COMPILE_STATUS, &status); } if (status != GL_FALSE) { errorShader = shader->FragmentShader; glGetShaderiv(shader->FragmentShader, GL_COMPILE_STATUS, &status); }
if (status == GL_FALSE) if (status == GL_FALSE)
{ {
static char buffer[10000]; /*static char buffer[10000];
GLsizei length = 0; GLsizei length = 0;
buffer[0] = 0; buffer[0] = 0;
glGetShaderInfoLog(errorShader, 10000, &length, buffer); glGetShaderInfoLog(errorShader, 10000, &length, buffer);*/
*outShader = nullptr; *outShader = nullptr;
return false; return false;
@ -307,10 +312,19 @@ bool OpenGLSWFrameBuffer::CreatePixelShader(FString vertexsrc, FString fragments
*outShader = nullptr; *outShader = nullptr;
return false; return false;
} }
glBindAttribLocation(shader->Program, 0, "Position"); glBindAttribLocation(shader->Program, 0, "AttrPosition");
glBindAttribLocation(shader->Program, 1, "Color0"); glBindAttribLocation(shader->Program, 1, "AttrColor0");
glBindAttribLocation(shader->Program, 2, "Color1"); glBindAttribLocation(shader->Program, 2, "AttrColor1");
glBindAttribLocation(shader->Program, 3, "TexCoord"); glBindAttribLocation(shader->Program, 3, "AttrTexCoord0");
shader->ConstantLocations[PSCONST_Desaturation] = glGetUniformLocation(shader->Program, "Desaturation");
shader->ConstantLocations[PSCONST_PaletteMod] = glGetUniformLocation(shader->Program, "PaletteMod");
shader->ConstantLocations[PSCONST_Weights] = glGetUniformLocation(shader->Program, "Weights");
shader->ConstantLocations[PSCONST_Gamma] = glGetUniformLocation(shader->Program, "Gamma");
shader->ImageLocation = glGetUniformLocation(shader->Program, "Image");
shader->PaletteLocation = glGetUniformLocation(shader->Program, "Palette");
shader->NewScreenLocation = glGetUniformLocation(shader->Program, "NewScreen");
shader->BurnLocation = glGetUniformLocation(shader->Program, "Burn");
*outShader = shader.release(); *outShader = shader.release();
return true; return true;
@ -361,7 +375,7 @@ bool OpenGLSWFrameBuffer::CreateIndexBuffer(int size, HWIndexBuffer **outIndexBu
return true; return true;
} }
bool OpenGLSWFrameBuffer::CreateTexture(int width, int height, int levels, int format, HWTexture **outTexture) bool OpenGLSWFrameBuffer::CreateTexture(const FString &name, int width, int height, int levels, int format, HWTexture **outTexture)
{ {
auto obj = std::make_unique<HWTexture>(); auto obj = std::make_unique<HWTexture>();
@ -389,6 +403,8 @@ bool OpenGLSWFrameBuffer::CreateTexture(int width, int height, int levels, int f
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
FGLDebug::LabelObject(GL_TEXTURE, obj->Texture, name);
glBindTexture(GL_TEXTURE_2D, oldBinding); glBindTexture(GL_TEXTURE_2D, oldBinding);
*outTexture = obj.release(); *outTexture = obj.release();
@ -401,15 +417,32 @@ void OpenGLSWFrameBuffer::SetGammaRamp(const GammaRamp *ramp)
void OpenGLSWFrameBuffer::SetPixelShaderConstantF(int uniformIndex, const float *data, int vec4fcount) void OpenGLSWFrameBuffer::SetPixelShaderConstantF(int uniformIndex, const float *data, int vec4fcount)
{ {
glUniform4fv(uniformIndex, vec4fcount, data); assert(uniformIndex < 4 && vec4fcount == 1); // This emulation of d3d9 only works for very simple stuff
for (int i = 0; i < 4; i++)
ShaderConstants[uniformIndex * 4 + i] = data[i];
if (CurrentShader && CurrentShader->ConstantLocations[uniformIndex] != -1)
glUniform4fv(CurrentShader->ConstantLocations[uniformIndex], vec4fcount, data);
} }
void OpenGLSWFrameBuffer::SetHWPixelShader(HWPixelShader *shader) void OpenGLSWFrameBuffer::SetHWPixelShader(HWPixelShader *shader)
{ {
if (shader) if (shader != CurrentShader)
glUseProgram(shader->Program); {
else if (shader)
glUseProgram(0); {
glUseProgram(shader->Program);
for (int i = 0; i < 4; i++)
{
if (shader->ConstantLocations[i] != -1)
glUniform4fv(shader->ConstantLocations[i], 1, &ShaderConstants[i * 4]);
}
}
else
{
glUseProgram(0);
}
}
CurrentShader = shader;
} }
void OpenGLSWFrameBuffer::SetStreamSource(HWVertexBuffer *vertexBuffer) void OpenGLSWFrameBuffer::SetStreamSource(HWVertexBuffer *vertexBuffer)
@ -430,6 +463,8 @@ void OpenGLSWFrameBuffer::SetIndices(HWIndexBuffer *indexBuffer)
void OpenGLSWFrameBuffer::DrawTriangleFans(int count, const FBVERTEX *vertices) void OpenGLSWFrameBuffer::DrawTriangleFans(int count, const FBVERTEX *vertices)
{ {
count = 2 + count;
GLint oldBinding = 0; GLint oldBinding = 0;
glGetIntegerv(GL_VERTEX_ARRAY_BINDING, &oldBinding); glGetIntegerv(GL_VERTEX_ARRAY_BINDING, &oldBinding);
@ -500,7 +535,7 @@ void OpenGLSWFrameBuffer::DrawPoints(int count, const FBVERTEX *vertices)
void OpenGLSWFrameBuffer::DrawLineList(int count) void OpenGLSWFrameBuffer::DrawLineList(int count)
{ {
glDrawArrays(GL_LINES, 0, count); glDrawArrays(GL_LINES, 0, count * 2);
} }
void OpenGLSWFrameBuffer::DrawTriangleList(int minIndex, int numVertices, int startIndex, int primitiveCount) void OpenGLSWFrameBuffer::DrawTriangleList(int minIndex, int numVertices, int startIndex, int primitiveCount)
@ -511,6 +546,8 @@ void OpenGLSWFrameBuffer::DrawTriangleList(int minIndex, int numVertices, int st
void OpenGLSWFrameBuffer::Present() void OpenGLSWFrameBuffer::Present()
{ {
SwapBuffers(); SwapBuffers();
glViewport(0, 0, GetClientWidth(), GetClientHeight());
Debug->Update();
} }
//========================================================================== //==========================================================================
@ -611,6 +648,13 @@ bool OpenGLSWFrameBuffer::LoadShaders()
{ {
break; break;
} }
glUseProgram(Shaders[i]->Program);
if (Shaders[i]->ImageLocation != -1) glUniform1i(Shaders[i]->ImageLocation, 0);
if (Shaders[i]->PaletteLocation != -1) glUniform1i(Shaders[i]->PaletteLocation, 1);
if (Shaders[i]->NewScreenLocation != -1) glUniform1i(Shaders[i]->NewScreenLocation, 0);
if (Shaders[i]->BurnLocation != -1) glUniform1i(Shaders[i]->BurnLocation, 1);
glUseProgram(0);
} }
if (i == NUM_SHADERS) if (i == NUM_SHADERS)
{ // Success! { // Success!
@ -711,7 +755,7 @@ void OpenGLSWFrameBuffer::KillNativeTexs()
bool OpenGLSWFrameBuffer::CreateFBTexture() bool OpenGLSWFrameBuffer::CreateFBTexture()
{ {
CreateTexture(Width, Height, 1, GL_R8, &FBTexture); CreateTexture("FBTexture", Width, Height, 1, GL_R8, &FBTexture);
FBWidth = Width; FBWidth = Width;
FBHeight = Height; FBHeight = Height;
return true; return true;
@ -725,7 +769,7 @@ bool OpenGLSWFrameBuffer::CreateFBTexture()
bool OpenGLSWFrameBuffer::CreatePaletteTexture() bool OpenGLSWFrameBuffer::CreatePaletteTexture()
{ {
if (!CreateTexture(256, 1, 1, GL_RGBA8, &PaletteTexture)) if (!CreateTexture("PaletteTexture", 256, 1, 1, GL_RGBA8, &PaletteTexture))
{ {
return false; return false;
} }
@ -1118,7 +1162,7 @@ void OpenGLSWFrameBuffer::Draw3DPart(bool copy3d)
GLint oldBinding = 0; GLint oldBinding = 0;
glGetIntegerv(GL_TEXTURE_BINDING_2D, &oldBinding); glGetIntegerv(GL_TEXTURE_BINDING_2D, &oldBinding);
glBindTexture(GL_TEXTURE_2D, FBTexture->Texture); glBindTexture(GL_TEXTURE_2D, FBTexture->Texture);
glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, Width, Height, GL_R8, GL_UNSIGNED_BYTE, 0); glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, Width, Height, GL_RED, GL_UNSIGNED_BYTE, 0);
glBindTexture(GL_TEXTURE_2D, oldBinding); glBindTexture(GL_TEXTURE_2D, oldBinding);
} }
@ -1623,7 +1667,7 @@ OpenGLSWFrameBuffer::Atlas::Atlas(OpenGLSWFrameBuffer *fb, int w, int h, int for
} }
*prev = this; *prev = this;
fb->CreateTexture(w, h, 1, format, &Tex); fb->CreateTexture("Atlas", w, h, 1, format, &Tex);
Width = w; Width = w;
Height = h; Height = h;
} }
@ -1907,7 +1951,7 @@ bool OpenGLSWFrameBuffer::OpenGLTex::Update()
glBindTexture(GL_TEXTURE_2D, Box->Owner->Tex->Texture); glBindTexture(GL_TEXTURE_2D, Box->Owner->Tex->Texture);
glTexSubImage2D(GL_TEXTURE_2D, 0, rect.left, rect.top, rect.right - rect.left, rect.bottom - rect.top, format == GL_RGBA8 ? GL_BGRA : GL_RED, GL_UNSIGNED_BYTE, 0); glTexSubImage2D(GL_TEXTURE_2D, 0, rect.left, rect.top, rect.right - rect.left, rect.bottom - rect.top, format == GL_RGBA8 ? GL_BGRA : GL_RED, GL_UNSIGNED_BYTE, 0);
glBindTexture(GL_TEXTURE_2D, oldBinding); glBindTexture(GL_TEXTURE_2D, oldBinding);
glBindTexture(GL_PIXEL_UNPACK_BUFFER, 0); glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
return true; return true;
} }
@ -1996,7 +2040,7 @@ OpenGLSWFrameBuffer::OpenGLPal::OpenGLPal(FRemapTable *remap, OpenGLSWFrameBuffe
BorderColor = 0; BorderColor = 0;
RoundedPaletteSize = count; RoundedPaletteSize = count;
if (fb->CreateTexture(count, 1, 1, GL_RGBA8, &Tex)) if (fb->CreateTexture("Pal", count, 1, 1, GL_RGBA8, &Tex))
{ {
if (!Update()) if (!Update())
{ {
@ -3312,7 +3356,7 @@ void OpenGLSWFrameBuffer::SetColorOverlay(uint32_t color, float alpha, uint32_t
} }
} }
void OpenGLSWFrameBuffer::EnableAlphaTest(BOOL enabled) void OpenGLSWFrameBuffer::EnableAlphaTest(bool enabled)
{ {
if (enabled != AlphaTestEnabled) if (enabled != AlphaTestEnabled)
{ {

View File

@ -136,12 +136,18 @@ private:
int Program = 0; int Program = 0;
int VertexShader = 0; int VertexShader = 0;
int FragmentShader = 0; int FragmentShader = 0;
int ConstantLocations[4];
int ImageLocation = -1;
int PaletteLocation = -1;
int NewScreenLocation = -1;
int BurnLocation = -1;
}; };
bool CreatePixelShader(FString vertexsrc, FString fragmentsrc, const FString &defines, HWPixelShader **outShader); bool CreatePixelShader(FString vertexsrc, FString fragmentsrc, const FString &defines, HWPixelShader **outShader);
bool CreateVertexBuffer(int size, HWVertexBuffer **outVertexBuffer); bool CreateVertexBuffer(int size, HWVertexBuffer **outVertexBuffer);
bool CreateIndexBuffer(int size, HWIndexBuffer **outIndexBuffer); bool CreateIndexBuffer(int size, HWIndexBuffer **outIndexBuffer);
bool CreateTexture(int width, int height, int levels, int format, HWTexture **outTexture); bool CreateTexture(const FString &name, int width, int height, int levels, int format, HWTexture **outTexture);
void SetGammaRamp(const GammaRamp *ramp); void SetGammaRamp(const GammaRamp *ramp);
void SetPixelShaderConstantF(int uniformIndex, const float *data, int vec4fcount); void SetPixelShaderConstantF(int uniformIndex, const float *data, int vec4fcount);
void SetHWPixelShader(HWPixelShader *shader); void SetHWPixelShader(HWPixelShader *shader);
@ -154,7 +160,7 @@ private:
void Present(); void Present();
static uint32_t ColorARGB(uint32_t a, uint32_t r, uint32_t g, uint32_t b) { return ((a & 0xff) << 24) | ((r & 0xff) << 16) | ((g & 0xff) << 8) | ((b) & 0xff); } static uint32_t ColorARGB(uint32_t a, uint32_t r, uint32_t g, uint32_t b) { return ((a & 0xff) << 24) | ((r & 0xff) << 16) | ((g & 0xff) << 8) | ((b) & 0xff); }
static uint32_t ColorRGBA(uint32_t a, uint32_t r, uint32_t g, uint32_t b) { return ColorARGB(a, r, g, b); } static uint32_t ColorRGBA(uint32_t r, uint32_t g, uint32_t b, uint32_t a) { return ColorARGB(a, r, g, b); }
static uint32_t ColorXRGB(uint32_t r, uint32_t g, uint32_t b) { return ColorARGB(0xff, r, g, b); } static uint32_t ColorXRGB(uint32_t r, uint32_t g, uint32_t b) { return ColorARGB(0xff, r, g, b); }
static uint32_t ColorValue(float r, float g, float b, float a) { return ColorRGBA((uint32_t)(r * 255.0f), (uint32_t)(g * 255.0f), (uint32_t)(b * 255.0f), (uint32_t)(a * 255.0f)); } static uint32_t ColorValue(float r, float g, float b, float a) { return ColorRGBA((uint32_t)(r * 255.0f), (uint32_t)(g * 255.0f), (uint32_t)(b * 255.0f), (uint32_t)(a * 255.0f)); }
@ -297,10 +303,10 @@ private:
enum enum
{ {
PSCONST_Desaturation = 1, PSCONST_Desaturation = 0,
PSCONST_PaletteMod = 2, PSCONST_PaletteMod = 1,
PSCONST_Weights = 6, PSCONST_Weights = 2,
PSCONST_Gamma = 7, PSCONST_Gamma = 3,
}; };
enum enum
{ {
@ -366,7 +372,7 @@ private:
void EndBatch(); void EndBatch();
// State // State
void EnableAlphaTest(BOOL enabled); void EnableAlphaTest(bool enabled);
void SetAlphaBlend(int op, int srcblend = 0, int destblend = 0); void SetAlphaBlend(int op, int srcblend = 0, int destblend = 0);
void SetConstant(int cnum, float r, float g, float b, float a); void SetConstant(int cnum, float r, float g, float b, float a);
void SetPixelShader(HWPixelShader *shader); void SetPixelShader(HWPixelShader *shader);
@ -377,13 +383,17 @@ private:
template<typename T> static void SafeRelease(T &x) { if (x != nullptr) { delete x; x = nullptr; } } template<typename T> static void SafeRelease(T &x) { if (x != nullptr) { delete x; x = nullptr; } }
std::unique_ptr<HWVertexBuffer> StreamVertexBuffer; std::shared_ptr<FGLDebug> Debug;
BOOL AlphaTestEnabled; std::unique_ptr<HWVertexBuffer> StreamVertexBuffer;
BOOL AlphaBlendEnabled; float ShaderConstants[16];
int AlphaBlendOp; HWPixelShader *CurrentShader = nullptr;
int AlphaSrcBlend;
int AlphaDestBlend; bool AlphaTestEnabled = false;
bool AlphaBlendEnabled = false;
int AlphaBlendOp = 0;
int AlphaSrcBlend = 0;
int AlphaDestBlend = 0;
float Constant[3][4]; float Constant[3][4];
uint32_t CurBorderColor; uint32_t CurBorderColor;
HWPixelShader *CurPixelShader; HWPixelShader *CurPixelShader;

View File

@ -5,15 +5,15 @@ in vec4 PixelTexCoord0;
out vec4 FragColor; out vec4 FragColor;
uniform sampler2D Image;// : register(s0); uniform sampler2D Image;
uniform sampler2D Palette;// : register(s1); uniform sampler2D Palette;
uniform sampler2D NewScreen;// : register(s0); uniform sampler2D NewScreen;
uniform sampler2D Burn;// : register(s1); uniform sampler2D Burn;
uniform vec4 Desaturation;// : register(c1); // { Desat, 1 - Desat } uniform vec4 Desaturation; // { Desat, 1 - Desat }
uniform vec4 PaletteMod;// : register(c2); uniform vec4 PaletteMod;
uniform vec4 Weights;// : register(c6); // RGB->Gray weighting { 77/256.0, 143/256.0, 37/256.0, 1 } uniform vec4 Weights; // RGB->Gray weighting { 77/256.0, 143/256.0, 37/256.0, 1 }
uniform vec4 Gamma;// : register(c7); uniform vec4 Gamma;
vec4 TextureLookup(vec2 tex_coord) vec4 TextureLookup(vec2 tex_coord)
{ {

View File

@ -10,8 +10,9 @@ out vec4 PixelTexCoord0;
void main() void main()
{ {
gl_Position = AttrPosition; gl_Position = vec4(AttrPosition.xy / vec2(1920*2,1080*2) * 2.0 - 1.0, 1.0, 1.0);
PixelColor0 = AttrColor0; gl_Position.y = -gl_Position.y;
PixelColor1 = AttrColor1; PixelColor0 = AttrColor0.bgra;
PixelColor1 = AttrColor1.bgra;
PixelTexCoord0 = AttrTexCoord0; PixelTexCoord0 = AttrTexCoord0;
} }