Emulator-Archive/ppsspp
1
0
Fork 0
mirror of https://github.com/hrydgard/ppsspp.git synced 2024-11-26 23:10:38 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Cache vertex decoders. No big boost now, but opens for optimizing them more, and lets us keep stats of the most common ones.

Browse Source
This commit is contained in:
Henrik Rydgard 2013-03-24 12:28:42 +01:00
parent 15e3a7750f
commit d96c588288
6 changed files with 98 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

62
GPU/GLES/TransformPipeline.cpp
View File

@ -55,6 +55,7 @@ enum {
TransformDrawEngine::TransformDrawEngine()
: collectedVerts(0),
prevPrim_(-1),
dec_(0),
lastVType_(-1),
curVbo_(0),
shaderManager_(0),
@ -142,8 +143,9 @@ void TransformDrawEngine::DrawBezier(int ucount, int vcount) {
}
if (!(gstate.vertType & GE_VTYPE_TC_MASK)) {
dec.SetVertexType(gstate.vertType);
u32 newVertType = dec.InjectUVs(decoded2, Memory::GetPointer(gstate_c.vertexAddr), customUV, 16);
VertexDecoder *dec = GetVertexDecoder(gstate.vertType);
dec->SetVertexType(gstate.vertType);
u32 newVertType = dec->InjectUVs(decoded2, Memory::GetPointer(gstate_c.vertexAddr), customUV, 16);
SubmitPrim(decoded2, &indices[0], GE_PRIM_TRIANGLES, c, newVertType, GE_VTYPE_IDX_16BIT, 0);
} else {
SubmitPrim(Memory::GetPointer(gstate_c.vertexAddr), &indices[0], GE_PRIM_TRIANGLES, c, gstate.vertType, GE_VTYPE_IDX_16BIT, 0);
@ -184,8 +186,9 @@ void TransformDrawEngine::DrawSpline(int ucount, int vcount, int utype, int vtyp
}
if (!(gstate.vertType & GE_VTYPE_TC_MASK)) {
dec.SetVertexType(gstate.vertType);
u32 newVertType = dec.InjectUVs(decoded2, Memory::GetPointer(gstate_c.vertexAddr), customUV, 16);
VertexDecoder *dec = GetVertexDecoder(gstate.vertType);
dec->SetVertexType(gstate.vertType);
u32 newVertType = dec->InjectUVs(decoded2, Memory::GetPointer(gstate_c.vertexAddr), customUV, 16);
SubmitPrim(decoded2, &indices[0], GE_PRIM_TRIANGLES, c, newVertType, GE_VTYPE_IDX_16BIT, 0);
} else {
SubmitPrim(Memory::GetPointer(gstate_c.vertexAddr), &indices[0], GE_PRIM_TRIANGLES, c, gstate.vertType, GE_VTYPE_IDX_16BIT, 0);
@ -762,6 +765,16 @@ void TransformDrawEngine::SoftwareTransformAndDraw(
}
}
VertexDecoder *TransformDrawEngine::GetVertexDecoder(u32 vtype) {
auto iter = decoderMap_.find(vtype);
if (iter != decoderMap_.end())
return iter->second;
VertexDecoder *dec = new VertexDecoder();
dec->SetVertexType(vtype);
decoderMap_[vtype] = dec;
return dec;
}
void TransformDrawEngine::SubmitPrim(void *verts, void *inds, int prim, int vertexCount, u32 vertType, int forceIndexType, int *bytesRead) {
if (vertexCount == 0)
return; // we ignore zero-sized draw calls.
@ -772,16 +785,14 @@ void TransformDrawEngine::SubmitPrim(void *verts, void *inds, int prim, int vert
// If vtype has changed, setup the vertex decoder.
// TODO: Simply cache the setup decoders instead.
if (vertType != lastVType_) {
dec.SetVertexType(vertType);
dec_ = GetVertexDecoder(vertType);
lastVType_ = vertType;
}
if (bytesRead)
*bytesRead = vertexCount * dec.VertexSize();
dec_->IncrementStat(STAT_VERTSSUBMITTED, vertexCount);
if (!indexGen.Empty()) {
gpuStats.numJoins++;
}
if (bytesRead)
*bytesRead = vertexCount * dec_->VertexSize();
gpuStats.numDrawCalls++;
gpuStats.numVertsSubmitted += vertexCount;
@ -812,7 +823,7 @@ void TransformDrawEngine::DecodeVerts() {
void *inds = dc.inds;
if (indexType == GE_VTYPE_IDX_NONE >> GE_VTYPE_IDX_SHIFT) {
// Decode the verts and apply morphing. Simple.
dec.DecodeVerts(decoded + collectedVerts * (int)dec.GetDecVtxFmt().stride,
dec_->DecodeVerts(decoded + collectedVerts * (int)dec_->GetDecVtxFmt().stride,
dc.verts, indexLowerBound, indexUpperBound);
collectedVerts += indexUpperBound - indexLowerBound + 1;
indexGen.AddPrim(dc.prim, dc.vertexCount);
@ -848,7 +859,7 @@ void TransformDrawEngine::DecodeVerts() {
}
// 3. Decode that range of vertex data.
dec.DecodeVerts(decoded + collectedVerts * (int)dec.GetDecVtxFmt().stride,
dec_->DecodeVerts(decoded + collectedVerts * (int)dec_->GetDecVtxFmt().stride,
dc.verts, indexLowerBound, indexUpperBound);
collectedVerts += indexUpperBound - indexLowerBound + 1;
@ -861,7 +872,7 @@ void TransformDrawEngine::DecodeVerts() {
u32 TransformDrawEngine::ComputeHash() {
u32 fullhash = 0;
int vertexSize = dec.GetDecVtxFmt().stride;
int vertexSize = dec_->GetDecVtxFmt().stride;
// TODO: Add some caps both for numDrawCalls and num verts to check?
for (int i = 0; i < numDrawCalls; i++) {
@ -870,7 +881,7 @@ u32 TransformDrawEngine::ComputeHash() {
} else {
fullhash += CityHash32((const char *)drawCalls[i].verts + vertexSize * drawCalls[i].indexLowerBound,
vertexSize * (drawCalls[i].indexUpperBound - drawCalls[i].indexLowerBound));
int indexSize = (dec.VertexType() & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT ? 2 : 1;
int indexSize = (dec_->VertexType() & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT ? 2 : 1;
fullhash += CityHash32((const char *)drawCalls[i].inds, indexSize * drawCalls[i].vertexCount);
}
}
@ -913,6 +924,17 @@ void TransformDrawEngine::DecimateTrackedVertexArrays() {
else
++iter;
}
// Enable if you want to see vertex decoders in the log output. Need a better way.
#if 0
char buffer[16384];
for (std::map<u32, VertexDecoder*>::iterator dec = decoderMap_.begin(); dec != decoderMap_.end(); ++dec) {
char *ptr = buffer;
ptr += dec->second->ToString(ptr);
// *ptr++ = '\n';
NOTICE_LOG(HLE, buffer);
}
#endif
}
VertexArrayInfo::~VertexArrayInfo() {
@ -952,7 +974,7 @@ void TransformDrawEngine::Flush() {
vai = iter->second;
} else {
vai = new VertexArrayInfo();
vai->decFmt = dec.GetDecVtxFmt();
vai->decFmt = dec_->GetDecVtxFmt();
vai_[id] = vai;
}
@ -1015,7 +1037,7 @@ void TransformDrawEngine::Flush() {
glGenBuffers(1, &vai->vbo);
glBindBuffer(GL_ARRAY_BUFFER, vai->vbo);
glBufferData(GL_ARRAY_BUFFER, dec.GetDecVtxFmt().stride * indexGen.MaxIndex(), decoded, GL_STATIC_DRAW);
glBufferData(GL_ARRAY_BUFFER, dec_->GetDecVtxFmt().stride * indexGen.MaxIndex(), decoded, GL_STATIC_DRAW);
// If there's only been one primitive type, and it's either TRIANGLES, LINES or POINTS,
// there is no need for the index buffer we built. We can then use glDrawArrays instead
// for a very minor speed boost.
@ -1087,7 +1109,7 @@ rotateVBO:
if (curVbo_ == NUM_VBOS)
curVbo_ = 0;
glBindBuffer(GL_ARRAY_BUFFER, vbo);
glBufferData(GL_ARRAY_BUFFER, dec.GetDecVtxFmt().stride * indexGen.MaxIndex(), decoded, GL_STREAM_DRAW);
glBufferData(GL_ARRAY_BUFFER, dec_->GetDecVtxFmt().stride * indexGen.MaxIndex(), decoded, GL_STREAM_DRAW);
if (useElements) {
glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, ebo);
glBufferData(GL_ELEMENT_ARRAY_BUFFER, sizeof(short) * vertexCount, (GLvoid *)decIndex, GL_STREAM_DRAW);
@ -1101,7 +1123,7 @@ rotateVBO:
DEBUG_LOG(G3D, "Flush prim %i! %i verts in one go", prim, vertexCount);
SetupDecFmtForDraw(program, dec.GetDecVtxFmt(), vbo ? 0 : decoded);
SetupDecFmtForDraw(program, dec_->GetDecVtxFmt(), vbo ? 0 : decoded);
if (useElements) {
glDrawElements(glprim[prim], vertexCount, GL_UNSIGNED_SHORT, ebo ? 0 : (GLvoid*)decIndex);
if (ebo)
@ -1117,7 +1139,9 @@ rotateVBO:
prim = indexGen.Prim();
DEBUG_LOG(G3D, "Flush prim %i SW! %i verts in one go", prim, indexGen.VertexCount());
SoftwareTransformAndDraw(prim, decoded, program, indexGen.VertexCount(), dec.VertexType(), (void *)decIndex, GE_VTYPE_IDX_16BIT, dec.GetDecVtxFmt(),
SoftwareTransformAndDraw(
prim, decoded, program, indexGen.VertexCount(),
dec_->VertexType(), (void *)decIndex, GE_VTYPE_IDX_16BIT, dec_->GetDecVtxFmt(),
indexGen.MaxIndex());
}

11
GPU/GLES/TransformPipeline.h
View File

@ -17,6 +17,8 @@
#pragma once
#include <map>
#include "IndexGenerator.h"
#include "VertexDecoder.h"
#include "gfx/gl_lost_manager.h"
@ -125,6 +127,8 @@ private:
u32 ComputeFastDCID();
u32 ComputeHash(); // Reads deferred vertex data.
VertexDecoder *GetVertexDecoder(u32 vtype);
// Defer all vertex decoding to a Flush, so that we can hash and cache the
// generated buffers without having to redecode them every time.
struct DeferredDrawCall {
@ -143,9 +147,12 @@ private:
int collectedVerts;
int prevPrim_;
// Vertex collector buffers
VertexDecoder dec;
// Cached vertex decoders
std::map<u32, VertexDecoder *> decoderMap_;
VertexDecoder *dec_;
u32 lastVType_;
// Vertex collector buffers
u8 *decoded;
u16 *decIndex;

23
GPU/GLES/VertexDecoder.cpp
View File

@ -750,3 +750,26 @@ u32 VertexDecoder::InjectUVs(u8 *decoded, const void *verts, float *customuv, in
}
return customVertType;
}
int VertexDecoder::ToString(char *output) const {
char * start = output;
output += sprintf(output, "P: %i ", pos);
if (nrm)
output += sprintf(output, "N: %i ", nrm);
if (col)
output += sprintf(output, "C: %i ", col);
if (tc)
output += sprintf(output, "T: %i ", tc);
if (weighttype)
output += sprintf(output, "W: %i ", weighttype);
if (idx)
output += sprintf(output, "I: %i ", idx);
if (morphcount > 1)
output += sprintf(output, "Morph: %i ", morphcount);
output += sprintf(output, "Verts: %i ", stats_[STAT_VERTSSUBMITTED]);
if (throughmode)
output += sprintf(output, " (through)");
output += sprintf(output, " (size: %i)", VertexSize());
return output - start;
}

19
GPU/GLES/VertexDecoder.h
View File

@ -76,6 +76,11 @@ typedef void (VertexDecoder::*StepFunction)() const;
void GetIndexBounds(void *inds, int count, u32 vertType, u16 *indexLowerBound, u16 *indexUpperBound);
enum {
STAT_VERTSSUBMITTED = 0,
NUM_VERTEX_DECODER_STATS = 1
};
// Right now
// - only contains computed information
// - does decoding in nasty branchfilled loops
@ -147,6 +152,18 @@ public:
void Step_PosS16Through() const;
void Step_PosFloatThrough() const;
void ResetStats() {
memset(stats_, 0, sizeof(stats_));
}
void IncrementStat(int stat, int amount) {
stats_[stat] += amount;
}
// output must be big for safety.
// Returns number of chars written.
// Ugly for speed.
int ToString(char *output) const;
// Mutable decoder state
mutable u8 *decoded_;
@ -180,6 +197,8 @@ public:
int idx;
int morphcount;
int nweights;
int stats_[NUM_VERTEX_DECODER_STATS];
};
// Reads decoded vertex formats in a convenient way. For software transform and debugging.

2
GPU/GPUState.h
View File

@ -282,7 +282,6 @@ struct GPUStatistics
memset(this, 0, sizeof(*this));
}
void resetFrame() {
numJoins = 0;
numDrawCalls = 0;
numCachedDrawCalls = 0;
numVertsSubmitted = 0;
@ -298,7 +297,6 @@ struct GPUStatistics
}
// Per frame statistics
int numJoins;
int numDrawCalls;
int numCachedDrawCalls;
int numFlushes;

5
GPU/GeDisasm.cpp
View File

@ -52,7 +52,10 @@ void GeDisassembleOp(u32 pc, u32 op, u32 prev, char *buffer) {
"TRIANGLE_FAN",
"RECTANGLES",
};
sprintf(buffer, "DrawPrim type: %s count: %i vaddr= %08x, iaddr= %08x", type < 7 ? types[type] : "INVALID", count, gstate_c.vertexAddr, gstate_c.indexAddr);
if (gstate.vertType & GE_VTYPE_IDX_MASK)
sprintf(buffer, "DrawPrim indexed type: %s count: %i vaddr= %08x, iaddr= %08x", type < 7 ? types[type] : "INVALID", count, gstate_c.vertexAddr, gstate_c.indexAddr);
else
sprintf(buffer, "DrawPrim type: %s count: %i vaddr= %08x", type < 7 ? types[type] : "INVALID", count, gstate_c.vertexAddr);
}
break;