thirdparty/ppsspp
2
0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Align some vertex arrays to page size. It's said to possibly be beneficial.

Browse source
This commit is contained in:
Henrik Rydgard 2013-01-29 00:48:13 +01:00
parent 630c025fd5
commit 366583d34f
5 changed files with 41 additions and 39 deletions
Download patch file Download diff file Expand all files Collapse all files

4
GPU/GLES/DisplayListInterpreter.cpp
View file

@ -163,13 +163,11 @@ const int flushBeforeCommandList[] = {
*/ */
}; };
GLES_GPU::GLES_GPU(int renderWidth, int renderHeight) GLES_GPU::GLES_GPU()
: interruptsEnabled_(true), : interruptsEnabled_(true),
displayFramebufPtr_(0), displayFramebufPtr_(0),
prevDisplayFramebuf_(0), prevDisplayFramebuf_(0),
prevPrevDisplayFramebuf_(0), prevPrevDisplayFramebuf_(0),
renderWidth_(renderWidth),
renderHeight_(renderHeight),
resized_(false) resized_(false)
{ {
shaderManager_ = new ShaderManager(); shaderManager_ = new ShaderManager();

5
GPU/GLES/DisplayListInterpreter.h
View file

@ -32,7 +32,7 @@ class LinkedShader;
class GLES_GPU : public GPUCommon class GLES_GPU : public GPUCommon
{ {
public: public:
GLES_GPU(int renderWidth, int renderHeight); GLES_GPU();
~GLES_GPU(); ~GLES_GPU();
virtual void InitClear(); virtual void InitClear();
virtual void PreExecuteOp(u32 op, u32 diff); virtual void PreExecuteOp(u32 op, u32 diff);
@ -81,9 +81,6 @@ private:
u32 displayStride_; u32 displayStride_;
int displayFormat_; int displayFormat_;
int renderWidth_;
int renderHeight_;
struct CmdProcessorState { struct CmdProcessorState {
u32 pc; u32 pc;
u32 stallAddr; u32 stallAddr;

29
GPU/GLES/TransformPipeline.cpp
View file

@ -17,6 +17,7 @@
#include "base/timeutil.h" #include "base/timeutil.h"
#include "Common/MemoryUtil.h"
#include "../../Core/MemMap.h" #include "../../Core/MemMap.h"
#include "../../Core/Host.h" #include "../../Core/Host.h"
#include "../../Core/System.h" #include "../../Core/System.h"
@ -45,6 +46,12 @@ const GLuint glprim[8] = {
GL_TRIANGLES, // With OpenGL ES we have to expand sprites into triangles, tripling the data instead of doubling. sigh. OpenGL ES, Y U NO SUPPORT GL_QUADS? GL_TRIANGLES, // With OpenGL ES we have to expand sprites into triangles, tripling the data instead of doubling. sigh. OpenGL ES, Y U NO SUPPORT GL_QUADS?
}; };
enum {
DECODED_VERTEX_BUFFER_SIZE = 65536 * 48,
DECODED_INDEX_BUFFER_SIZE = 65536 * 2,
TRANSFORMED_VERTEX_BUFFER_SIZE = 65536 * sizeof(TransformedVertex)
};
TransformDrawEngine::TransformDrawEngine() TransformDrawEngine::TransformDrawEngine()
: numDrawCalls(0), : numDrawCalls(0),
collectedVerts(0), collectedVerts(0),
@ -52,10 +59,13 @@ TransformDrawEngine::TransformDrawEngine()
lastVType_(-1), lastVType_(-1),
curVbo_(0), curVbo_(0),
shaderManager_(0) { shaderManager_(0) {
decoded = new u8[65536 * 48]; // Allocate nicely aligned memory. Maybe graphics drivers will
decIndex = new u16[65536]; // appreciate it.
transformed = new TransformedVertex[65536]; // All this is a LOT of memory, need to see if we can cut down somehow.
transformedExpanded = new TransformedVertex[65536 * 3]; decoded = (u8 *)AllocateMemoryPages(DECODED_VERTEX_BUFFER_SIZE);
decIndex = (u16 *)AllocateMemoryPages(DECODED_INDEX_BUFFER_SIZE);
transformed = (TransformedVertex *)AllocateMemoryPages(TRANSFORMED_VERTEX_BUFFER_SIZE);
transformedExpanded = (TransformedVertex *)AllocateMemoryPages(3 * TRANSFORMED_VERTEX_BUFFER_SIZE);
memset(vbo_, 0, sizeof(vbo_)); memset(vbo_, 0, sizeof(vbo_));
memset(ebo_, 0, sizeof(ebo_)); memset(ebo_, 0, sizeof(ebo_));
indexGen.Setup(decIndex); indexGen.Setup(decIndex);
@ -65,10 +75,10 @@ TransformDrawEngine::TransformDrawEngine()
TransformDrawEngine::~TransformDrawEngine() { TransformDrawEngine::~TransformDrawEngine() {
DestroyDeviceObjects(); DestroyDeviceObjects();
delete [] decoded; FreeMemoryPages(decoded, DECODED_VERTEX_BUFFER_SIZE);
delete [] decIndex; FreeMemoryPages(decIndex, DECODED_INDEX_BUFFER_SIZE);
delete [] transformed; FreeMemoryPages(transformed, TRANSFORMED_VERTEX_BUFFER_SIZE);
delete [] transformedExpanded; FreeMemoryPages(transformedExpanded, 3 * TRANSFORMED_VERTEX_BUFFER_SIZE);
unregister_gl_resource_holder(this); unregister_gl_resource_holder(this);
} }
@ -293,7 +303,7 @@ struct GlTypeInfo {
GLboolean normalized; GLboolean normalized;
}; };
const GlTypeInfo GLComp[8] = { const GlTypeInfo GLComp[] = {
{0}, // DEC_NONE, {0}, // DEC_NONE,
{GL_FLOAT, 1, GL_FALSE}, // DEC_FLOAT_1, {GL_FLOAT, 1, GL_FALSE}, // DEC_FLOAT_1,
{GL_FLOAT, 2, GL_FALSE}, // DEC_FLOAT_2, {GL_FLOAT, 2, GL_FALSE}, // DEC_FLOAT_2,
@ -302,6 +312,7 @@ const GlTypeInfo GLComp[8] = {
{GL_BYTE, 4, GL_TRUE}, // DEC_S8_3, {GL_BYTE, 4, GL_TRUE}, // DEC_S8_3,
{GL_SHORT, 4, GL_TRUE},// DEC_S16_3, {GL_SHORT, 4, GL_TRUE},// DEC_S16_3,
{GL_UNSIGNED_BYTE, 4, GL_TRUE},// DEC_U8_4, {GL_UNSIGNED_BYTE, 4, GL_TRUE},// DEC_U8_4,
{GL_UNSIGNED_BYTE, 3, GL_TRUE},// DEC_U8_3,
}; };
static inline void VertexAttribSetup(int attrib, int fmt, int stride, u8 *ptr) { static inline void VertexAttribSetup(int attrib, int fmt, int stride, u8 *ptr) {

40
GPU/GLES/VertexDecoder.cpp
View file

@ -125,8 +125,7 @@ void VertexDecoder::Step_WeightsFloat() const
{ {
float *wt = (float *)(decoded_ + decFmt.w0off); float *wt = (float *)(decoded_ + decFmt.w0off);
const float *wdata = (const float*)(ptr_); const float *wdata = (const float*)(ptr_);
for (int j = 0; j < nweights; j++) memcpy(wt, wdata, nweights * sizeof(float));
wt[j] = wdata[j];
} }
void VertexDecoder::Step_TcU8() const void VertexDecoder::Step_TcU8() const
@ -157,8 +156,7 @@ void VertexDecoder::Step_TcFloat() const
{ {
float *uv = (float *)(decoded_ + decFmt.uvoff); float *uv = (float *)(decoded_ + decFmt.uvoff);
const float *uvdata = (const float*)(ptr_ + tcoff); const float *uvdata = (const float*)(ptr_ + tcoff);
uv[0] = uvdata[0]; memcpy(uv, uvdata, sizeof(float) * 2);
uv[1] = uvdata[1];
} }
void VertexDecoder::Step_TcFloatThrough() const void VertexDecoder::Step_TcFloatThrough() const
@ -377,10 +375,9 @@ void VertexDecoder::Step_PosS16() const
void VertexDecoder::Step_PosFloat() const void VertexDecoder::Step_PosFloat() const
{ {
float *v = (float *)(decoded_ + decFmt.posoff); u8 *v = (u8 *)(decoded_ + decFmt.posoff);
const float *fv = (const float*)(ptr_ + posoff); const u8 *fv = (const u8*)(ptr_ + posoff);
for (int j = 0; j < 3; j++) memcpy(v, fv, 12);
v[j] = fv[j];
} }
void VertexDecoder::Step_PosS8Through() const void VertexDecoder::Step_PosS8Through() const
@ -405,10 +402,9 @@ void VertexDecoder::Step_PosS16Through() const
void VertexDecoder::Step_PosFloatThrough() const void VertexDecoder::Step_PosFloatThrough() const
{ {
float *v = (float *)(decoded_ + decFmt.posoff); u8 *v = (u8 *)(decoded_ + decFmt.posoff);
const float *fv = (const float*)(ptr_ + posoff); const u8 *fv = (const u8*)(ptr_ + posoff);
for (int j = 0; j < 3; j++) memcpy(v, fv, 12);
v[j] = fv[j];
} }
void VertexDecoder::Step_PosS8Morph() const void VertexDecoder::Step_PosS8Morph() const
@ -445,21 +441,21 @@ void VertexDecoder::Step_PosFloatMorph() const
} }
} }
const StepFunction wtstep[4] = { static const StepFunction wtstep[4] = {
0, 0,
&VertexDecoder::Step_WeightsU8, &VertexDecoder::Step_WeightsU8,
&VertexDecoder::Step_WeightsU16, &VertexDecoder::Step_WeightsU16,
&VertexDecoder::Step_WeightsFloat, &VertexDecoder::Step_WeightsFloat,
}; };
const StepFunction tcstep[4] = { static const StepFunction tcstep[4] = {
0, 0,
&VertexDecoder::Step_TcU8, &VertexDecoder::Step_TcU8,
&VertexDecoder::Step_TcU16, &VertexDecoder::Step_TcU16,
&VertexDecoder::Step_TcFloat, &VertexDecoder::Step_TcFloat,
}; };
const StepFunction tcstep_through[4] = { static const StepFunction tcstep_through[4] = {
0, 0,
&VertexDecoder::Step_TcU8, &VertexDecoder::Step_TcU8,
&VertexDecoder::Step_TcU16Through, &VertexDecoder::Step_TcU16Through,
@ -468,7 +464,7 @@ const StepFunction tcstep_through[4] = {
// TODO: Tc Morph // TODO: Tc Morph
const StepFunction colstep[8] = { static const StepFunction colstep[8] = {
0, 0, 0, 0, 0, 0, 0, 0,
&VertexDecoder::Step_Color565, &VertexDecoder::Step_Color565,
&VertexDecoder::Step_Color5551, &VertexDecoder::Step_Color5551,
@ -476,7 +472,7 @@ const StepFunction colstep[8] = {
&VertexDecoder::Step_Color8888, &VertexDecoder::Step_Color8888,
}; };
const StepFunction colstep_morph[8] = { static const StepFunction colstep_morph[8] = {
0, 0, 0, 0, 0, 0, 0, 0,
&VertexDecoder::Step_Color565Morph, &VertexDecoder::Step_Color565Morph,
&VertexDecoder::Step_Color5551Morph, &VertexDecoder::Step_Color5551Morph,
@ -484,35 +480,35 @@ const StepFunction colstep_morph[8] = {
&VertexDecoder::Step_Color8888Morph, &VertexDecoder::Step_Color8888Morph,
}; };
const StepFunction nrmstep[4] = { static const StepFunction nrmstep[4] = {
0, 0,
&VertexDecoder::Step_NormalS8, &VertexDecoder::Step_NormalS8,
&VertexDecoder::Step_NormalS16, &VertexDecoder::Step_NormalS16,
&VertexDecoder::Step_NormalFloat, &VertexDecoder::Step_NormalFloat,
}; };
const StepFunction nrmstep_morph[4] = { static const StepFunction nrmstep_morph[4] = {
0, 0,
&VertexDecoder::Step_NormalS8Morph, &VertexDecoder::Step_NormalS8Morph,
&VertexDecoder::Step_NormalS16Morph, &VertexDecoder::Step_NormalS16Morph,
&VertexDecoder::Step_NormalFloatMorph, &VertexDecoder::Step_NormalFloatMorph,
}; };
const StepFunction posstep[4] = { static const StepFunction posstep[4] = {
0, 0,
&VertexDecoder::Step_PosS8, &VertexDecoder::Step_PosS8,
&VertexDecoder::Step_PosS16, &VertexDecoder::Step_PosS16,
&VertexDecoder::Step_PosFloat, &VertexDecoder::Step_PosFloat,
}; };
const StepFunction posstep_morph[4] = { static const StepFunction posstep_morph[4] = {
0, 0,
&VertexDecoder::Step_PosS8Morph, &VertexDecoder::Step_PosS8Morph,
&VertexDecoder::Step_PosS16Morph, &VertexDecoder::Step_PosS16Morph,
&VertexDecoder::Step_PosFloatMorph, &VertexDecoder::Step_PosFloatMorph,
}; };
const StepFunction posstep_through[4] = { static const StepFunction posstep_through[4] = {
0, 0,
&VertexDecoder::Step_PosS8Through, &VertexDecoder::Step_PosS8Through,
&VertexDecoder::Step_PosS16Through, &VertexDecoder::Step_PosS16Through,

2
GPU/GPUState.cpp
View file

@ -61,7 +61,7 @@ void InitGfxState()
gpu = new NullGPU(); gpu = new NullGPU();
break; break;
case GPU_GLES: case GPU_GLES:
gpu = new GLES_GPU(PSP_CoreParameter().renderWidth, PSP_CoreParameter().renderHeight); gpu = new GLES_GPU();
break; break;
} }
} }