preserve-emulation/ppsspp
1
0 Fork 0
mirror of https://github.com/hrydgard/ppsspp.git synced 2026-11-11 08:39:28 +01:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #9906 from hrydgard/fastrunloop-optimize

Browse Source 
GPU FastRunLoop optimizations
This commit is contained in:
Henrik Rydgård authored and GitHub committed 2017-08-17 16:25:15 +02:00
2 parents e50cb4f395 319df97bc7
commit d60bb27cd9
7 files changed
+156 -88
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

Common/Vulkan/VulkanContext.h
-2
View File
@@ -314,7 +314,6 @@ public:
const VkPhysicalDeviceFeatures &GetFeaturesEnabled() const { return featuresEnabled_; }
const VulkanPhysicalDeviceInfo &GetDeviceInfo() const { return deviceInfo_; }
private:
VkSemaphore acquireSemaphore;
VkSemaphore renderingCompleteSemaphore;
@@ -381,7 +380,6 @@ private:
VulkanDeleteList deleteList;
};
FrameData frame_[2];
int curFrame_;
GPU/D3D11/GPU_D3D11.cpp
+33 -15
View File
@@ -73,6 +73,7 @@ struct D3D11CommandTableEntry {
GPU_D3D11::CmdFunc func;
};
// This table gets crunched into a faster form by init.
static const D3D11CommandTableEntry commandTable[] = {
// Changes that dirty the current texture.
{ GE_CMD_TEXSIZE0, FLAG_FLUSHBEFOREONCHANGE | FLAG_EXECUTE, 0, &GPU_D3D11::Execute_TexSize0 },
@@ -88,7 +89,6 @@ static const D3D11CommandTableEntry commandTable[] = {
// Changes that trigger data copies. Only flushing on change for LOADCLUT must be a bit of a hack...
{ GE_CMD_LOADCLUT, FLAG_FLUSHBEFOREONCHANGE | FLAG_EXECUTE, 0, &GPU_D3D11::Execute_LoadClut },
{ GE_CMD_TRANSFERSTART, FLAG_FLUSHBEFORE | FLAG_EXECUTE | FLAG_READS_PC, 0, &GPUCommon::Execute_BlockTransferStart },
};
GPU_D3D11::CommandInfo GPU_D3D11::cmdInfo_[256]{};
@@ -389,27 +389,38 @@ void GPU_D3D11::CopyDisplayToOutputInternal() {
void GPU_D3D11::FastRunLoop(DisplayList &list) {
PROFILE_THIS_SCOPE("gpuloop");
const CommandInfo *cmdInfo = cmdInfo_;
for (; downcount > 0; --downcount) {
int dc = downcount;
for (; dc > 0; --dc) {
// We know that display list PCs have the upper nibble == 0 - no need to mask the pointer
const u32 op = *(const u32 *)(Memory::base + list.pc);
const u32 cmd = op >> 24;
const CommandInfo info = cmdInfo[cmd];
const u8 cmdFlags = info.flags; // If we stashed the cmdFlags in the top bits of the cmdmem, we could get away with one table lookup instead of two
const CommandInfo &info = cmdInfo[cmd];
const u32 diff = op ^ gstate.cmdmem[cmd];
// Inlined CheckFlushOp here to get rid of the dumpThisFrame_ check.
if ((cmdFlags & FLAG_FLUSHBEFORE) || (diff && (cmdFlags & FLAG_FLUSHBEFOREONCHANGE))) {
drawEngine_.Flush();
}
gstate.cmdmem[cmd] = op; // TODO: no need to write if diff==0...
if ((cmdFlags & FLAG_EXECUTE) || (diff && (cmdFlags & FLAG_EXECUTEONCHANGE))) {
(this->*info.func)(op, diff);
} else if (diff) {
uint64_t dirty = info.flags >> 8;
if (dirty)
gstate_c.Dirty(dirty);
if (diff == 0) {
if (info.flags & FLAG_EXECUTE) {
downcount = dc;
(this->*info.func)(op, diff);
dc = downcount;
}
} else {
uint64_t flags = info.flags;
if (flags & FLAG_FLUSHBEFOREONCHANGE) {
drawEngine_.Flush();
}
gstate.cmdmem[cmd] = op;
if (flags & (FLAG_EXECUTE | FLAG_EXECUTEONCHANGE)) {
downcount = dc;
(this->*info.func)(op, diff);
dc = downcount;
} else {
uint64_t dirty = flags >> 8;
if (dirty)
gstate_c.Dirty(dirty);
}
}
list.pc += 4;
}
downcount = 0;
}
void GPU_D3D11::FinishDeferred() {
@@ -533,6 +544,7 @@ void GPU_D3D11::Execute_Prim(u32 op, u32 diff) {
#endif
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitPrim(verts, inds, prim, count, vertexType, &bytesRead);
int vertexCost = EstimatePerVertexCost() * count;
@@ -546,6 +558,8 @@ void GPU_D3D11::Execute_Prim(u32 op, u32 diff) {
}
void GPU_D3D11::Execute_Bezier(u32 op, u32 diff) {
Flush();
// We don't dirty on normal changes anymore as we prescale, but it's needed for splines/bezier.
gstate_c.Dirty(DIRTY_UVSCALEOFFSET);
@@ -596,6 +610,7 @@ void GPU_D3D11::Execute_Bezier(u32 op, u32 diff) {
}
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitBezier(control_points, indices, gstate.getPatchDivisionU(), gstate.getPatchDivisionV(), bz_ucount, bz_vcount, patchPrim, computeNormals, patchFacing, gstate.vertType, &bytesRead);
if (gstate_c.bezier)
@@ -608,6 +623,8 @@ void GPU_D3D11::Execute_Bezier(u32 op, u32 diff) {
}
void GPU_D3D11::Execute_Spline(u32 op, u32 diff) {
Flush();
// We don't dirty on normal changes anymore as we prescale, but it's needed for splines/bezier.
gstate_c.Dirty(DIRTY_UVSCALEOFFSET);
@@ -664,6 +681,7 @@ void GPU_D3D11::Execute_Spline(u32 op, u32 diff) {
}
}
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitSpline(control_points, indices, gstate.getPatchDivisionU(), gstate.getPatchDivisionV(), sp_ucount, sp_vcount, sp_utype, sp_vtype, patchPrim, computeNormals, patchFacing, vertType, &bytesRead);
if (gstate_c.spline)
GPU/Directx9/GPU_DX9.cpp
+34 -16
View File
@@ -57,6 +57,7 @@ struct D3D9CommandTableEntry {
GPU_DX9::CmdFunc func;
};
// This table gets crunched into a faster form by init.
static const D3D9CommandTableEntry commandTable[] = {
// Changes that dirty the current texture.
{ GE_CMD_TEXSIZE0, FLAG_FLUSHBEFOREONCHANGE | FLAG_EXECUTE, 0, &GPU_DX9::Execute_TexSize0 },
@@ -72,7 +73,6 @@ static const D3D9CommandTableEntry commandTable[] = {
// Changes that trigger data copies. Only flushing on change for LOADCLUT must be a bit of a hack...
{ GE_CMD_LOADCLUT, FLAG_FLUSHBEFOREONCHANGE | FLAG_EXECUTE, 0, &GPU_DX9::Execute_LoadClut },
{ GE_CMD_TRANSFERSTART, FLAG_FLUSHBEFORE | FLAG_EXECUTE | FLAG_READS_PC, 0, &GPUCommon::Execute_BlockTransferStart },
};
GPU_DX9::CommandInfo GPU_DX9::cmdInfo_[256];
@@ -356,27 +356,38 @@ void GPU_DX9::CopyDisplayToOutputInternal() {
void GPU_DX9::FastRunLoop(DisplayList &list) {
PROFILE_THIS_SCOPE("gpuloop");
const CommandInfo *cmdInfo = cmdInfo_;
for (; downcount > 0; --downcount) {
int dc = downcount;
for (; dc > 0; --dc) {
// We know that display list PCs have the upper nibble == 0 - no need to mask the pointer
const u32 op = *(const u32 *)(Memory::base + list.pc);
const u32 cmd = op >> 24;
const CommandInfo info = cmdInfo[cmd];
const u8 cmdFlags = info.flags; // If we stashed the cmdFlags in the top bits of the cmdmem, we could get away with one table lookup instead of two
const CommandInfo &info = cmdInfo[cmd];
const u32 diff = op ^ gstate.cmdmem[cmd];
// Inlined CheckFlushOp here to get rid of the dumpThisFrame_ check.
if ((cmdFlags & FLAG_FLUSHBEFORE) || (diff && (cmdFlags & FLAG_FLUSHBEFOREONCHANGE))) {
drawEngine_.Flush();
}
gstate.cmdmem[cmd] = op; // TODO: no need to write if diff==0...
if ((cmdFlags & FLAG_EXECUTE) || (diff && (cmdFlags & FLAG_EXECUTEONCHANGE))) {
(this->*info.func)(op, diff);
} else if (diff) {
uint64_t dirty = info.flags >> 8;
if (dirty)
gstate_c.Dirty(dirty);
if (diff == 0) {
if (info.flags & FLAG_EXECUTE) {
downcount = dc;
(this->*info.func)(op, diff);
dc = downcount;
}
} else {
uint64_t flags = info.flags;
if (flags & FLAG_FLUSHBEFOREONCHANGE) {
drawEngine_.Flush();
}
gstate.cmdmem[cmd] = op;
if (flags & (FLAG_EXECUTE | FLAG_EXECUTEONCHANGE)) {
downcount = dc;
(this->*info.func)(op, diff);
dc = downcount;
} else {
uint64_t dirty = flags >> 8;
if (dirty)
gstate_c.Dirty(dirty);
}
}
list.pc += 4;
}
downcount = 0;
}
void GPU_DX9::FinishDeferred() {
@@ -386,7 +397,7 @@ void GPU_DX9::FinishDeferred() {
inline void GPU_DX9::CheckFlushOp(int cmd, u32 diff) {
const u8 cmdFlags = cmdInfo_[cmd].flags;
if ((cmdFlags & FLAG_FLUSHBEFORE) || (diff && (cmdFlags & FLAG_FLUSHBEFOREONCHANGE))) {
if (diff && (cmdFlags & FLAG_FLUSHBEFOREONCHANGE)) {
if (dumpThisFrame_) {
NOTICE_LOG(G3D, "================ FLUSH ================");
}
@@ -499,6 +510,7 @@ void GPU_DX9::Execute_Prim(u32 op, u32 diff) {
#endif
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitPrim(verts, inds, prim, count, vertexType, &bytesRead);
int vertexCost = EstimatePerVertexCost() * count;
@@ -512,6 +524,8 @@ void GPU_DX9::Execute_Prim(u32 op, u32 diff) {
}
void GPU_DX9::Execute_Bezier(u32 op, u32 diff) {
Flush();
// We don't dirty on normal changes anymore as we prescale, but it's needed for splines/bezier.
gstate_c.Dirty(DIRTY_UVSCALEOFFSET);
@@ -551,6 +565,7 @@ void GPU_DX9::Execute_Bezier(u32 op, u32 diff) {
bool computeNormals = gstate.isLightingEnabled();
bool patchFacing = gstate.patchfacing & 1;
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitBezier(control_points, indices, gstate.getPatchDivisionU(), gstate.getPatchDivisionV(), bz_ucount, bz_vcount, patchPrim, computeNormals, patchFacing, gstate.vertType, &bytesRead);
// After drawing, we advance pointers - see SubmitPrim which does the same.
@@ -559,6 +574,8 @@ void GPU_DX9::Execute_Bezier(u32 op, u32 diff) {
}
void GPU_DX9::Execute_Spline(u32 op, u32 diff) {
Flush();
// We don't dirty on normal changes anymore as we prescale, but it's needed for splines/bezier.
gstate_c.Dirty(DIRTY_UVSCALEOFFSET);
@@ -601,6 +618,7 @@ void GPU_DX9::Execute_Spline(u32 op, u32 diff) {
bool patchFacing = gstate.patchfacing & 1;
u32 vertType = gstate.vertType;
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitSpline(control_points, indices, gstate.getPatchDivisionU(), gstate.getPatchDivisionV(), sp_ucount, sp_vcount, sp_utype, sp_vtype, patchPrim, computeNormals, patchFacing, vertType, &bytesRead);
// After drawing, we advance pointers - see SubmitPrim which does the same.
GPU/GLES/GPU_GLES.cpp
+30 -17
View File
@@ -76,7 +76,6 @@ static const GLESCommandTableEntry commandTable[] = {
// Changes that trigger data copies. Only flushing on change for LOADCLUT must be a bit of a hack...
{ GE_CMD_LOADCLUT, FLAG_FLUSHBEFOREONCHANGE | FLAG_EXECUTE, 0, &GPU_GLES::Execute_LoadClut },
{ GE_CMD_TRANSFERSTART, FLAG_FLUSHBEFORE | FLAG_EXECUTE | FLAG_READS_PC, 0, &GPUCommon::Execute_BlockTransferStart },
};
GPU_GLES::CommandInfo GPU_GLES::cmdInfo_[256];
@@ -567,22 +566,29 @@ void GPU_GLES::FastRunLoop(DisplayList &list) {
// We know that display list PCs have the upper nibble == 0 - no need to mask the pointer
const u32 op = *(const u32 *)(Memory::base + list.pc);
const u32 cmd = op >> 24;
const CommandInfo info = cmdInfo[cmd];
const u8 cmdFlags = info.flags; // If we stashed the cmdFlags in the top bits of the cmdmem, we could get away with one table lookup instead of two
const CommandInfo &info = cmdInfo[cmd];
const u32 diff = op ^ gstate.cmdmem[cmd];
// Inlined CheckFlushOp here to get rid of the dumpThisFrame_ check.
if ((cmdFlags & FLAG_FLUSHBEFORE) || (diff && (cmdFlags & FLAG_FLUSHBEFOREONCHANGE))) {
drawEngine_.Flush();
}
gstate.cmdmem[cmd] = op; // TODO: no need to write if diff==0...
if ((cmdFlags & FLAG_EXECUTE) || (diff && (cmdFlags & FLAG_EXECUTEONCHANGE))) {
downcount = dc;
(this->*info.func)(op, diff);
dc = downcount;
} else if (diff) {
uint64_t dirty = info.flags >> 8;
if (dirty)
gstate_c.Dirty(dirty);
if (diff == 0) {
if (info.flags & FLAG_EXECUTE) {
downcount = dc;
(this->*info.func)(op, diff);
dc = downcount;
}
} else {
uint64_t flags = info.flags;
if (flags & FLAG_FLUSHBEFOREONCHANGE) {
drawEngine_.Flush();
}
gstate.cmdmem[cmd] = op;
if (flags & (FLAG_EXECUTE | FLAG_EXECUTEONCHANGE)) {
downcount = dc;
(this->*info.func)(op, diff);
dc = downcount;
} else {
uint64_t dirty = flags >> 8;
if (dirty)
gstate_c.Dirty(dirty);
}
}
list.pc += 4;
}
@@ -596,7 +602,7 @@ void GPU_GLES::FinishDeferred() {
inline void GPU_GLES::CheckFlushOp(int cmd, u32 diff) {
const u8 cmdFlags = cmdInfo_[cmd].flags;
if ((cmdFlags & FLAG_FLUSHBEFORE) || (diff && (cmdFlags & FLAG_FLUSHBEFOREONCHANGE))) {
if (diff && (cmdFlags & FLAG_FLUSHBEFOREONCHANGE)) {
if (dumpThisFrame_) {
NOTICE_LOG(G3D, "================ FLUSH ================");
}
@@ -676,6 +682,7 @@ void GPU_GLES::Execute_Prim(u32 op, u32 diff) {
#endif
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitPrim(verts, inds, prim, count, gstate.vertType, &bytesRead);
int vertexCost = EstimatePerVertexCost();
@@ -720,6 +727,8 @@ void GPU_GLES::Execute_VertexTypeSkinning(u32 op, u32 diff) {
}
void GPU_GLES::Execute_Bezier(u32 op, u32 diff) {
Flush();
// We don't dirty on normal changes anymore as we prescale, but it's needed for splines/bezier.
gstate_c.Dirty(DIRTY_UVSCALEOFFSET);
@@ -770,6 +779,7 @@ void GPU_GLES::Execute_Bezier(u32 op, u32 diff) {
}
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitBezier(control_points, indices, gstate.getPatchDivisionU(), gstate.getPatchDivisionV(), bz_ucount, bz_vcount, patchPrim, computeNormals, patchFacing, gstate.vertType, &bytesRead);
if (gstate_c.bezier)
@@ -782,6 +792,8 @@ void GPU_GLES::Execute_Bezier(u32 op, u32 diff) {
}
void GPU_GLES::Execute_Spline(u32 op, u32 diff) {
Flush();
// We don't dirty on normal changes anymore as we prescale, but it's needed for splines/bezier.
gstate_c.Dirty(DIRTY_UVSCALEOFFSET);
@@ -839,6 +851,7 @@ void GPU_GLES::Execute_Spline(u32 op, u32 diff) {
}
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitSpline(control_points, indices, gstate.getPatchDivisionU(), gstate.getPatchDivisionV(), sp_ucount, sp_vcount, sp_utype, sp_vtype, patchPrim, computeNormals, patchFacing, vertType, &bytesRead);
if (gstate_c.spline)
GPU/GPUCommon.cpp
+12 -20
View File
@@ -111,10 +111,12 @@ const CommonCommandTableEntry commonCommandTable[] = {
{ GE_CMD_LOGICOPENABLE, FLAG_FLUSHBEFOREONCHANGE, DIRTY_BLEND_STATE | DIRTY_FRAGMENTSHADER_STATE },
{ GE_CMD_TEXMAPMODE, FLAG_FLUSHBEFOREONCHANGE, DIRTY_VERTEXSHADER_STATE | DIRTY_FRAGMENTSHADER_STATE },
{ GE_CMD_TEXSCALEU, FLAG_EXECUTEONCHANGE, 0, &GPUCommon::Execute_TexScaleU },
{ GE_CMD_TEXSCALEV, FLAG_EXECUTEONCHANGE, 0, &GPUCommon::Execute_TexScaleV },
{ GE_CMD_TEXOFFSETU, FLAG_EXECUTEONCHANGE, 0, &GPUCommon::Execute_TexOffsetU },
{ GE_CMD_TEXOFFSETV, FLAG_EXECUTEONCHANGE, 0, &GPUCommon::Execute_TexOffsetV },
// These are read on every SubmitPrim, no need for dirtying or flushing.
{ GE_CMD_TEXSCALEU },
{ GE_CMD_TEXSCALEV },
{ GE_CMD_TEXOFFSETU },
{ GE_CMD_TEXOFFSETV },
// TEXSIZE0 is handled by each backend.
{ GE_CMD_TEXSIZE1, FLAG_FLUSHBEFOREONCHANGE, DIRTY_TEXTURE_PARAMS },
@@ -278,6 +280,7 @@ const CommonCommandTableEntry commonCommandTable[] = {
{ GE_CMD_TRANSFERSRCPOS, 0 },
{ GE_CMD_TRANSFERDSTPOS, 0 },
{ GE_CMD_TRANSFERSIZE, 0 },
{ GE_CMD_TRANSFERSTART, FLAG_FLUSHBEFORE | FLAG_EXECUTE | FLAG_READS_PC, 0, &GPUCommon::Execute_BlockTransferStart },
// We don't use the dither table.
{ GE_CMD_DITH0 },
@@ -1267,6 +1270,8 @@ void GPUCommon::Execute_Ret(u32 op, u32 diff) {
}
void GPUCommon::Execute_End(u32 op, u32 diff) {
Flush();
easy_guard guard(listLock);
const u32 prev = Memory::ReadUnchecked_U32(currentList->pc - 4);
UpdatePC(currentList->pc, currentList->pc);
@@ -1423,22 +1428,6 @@ void GPUCommon::Execute_End(u32 op, u32 diff) {
}
}
void GPUCommon::Execute_TexScaleU(u32 op, u32 diff) {
gstate_c.uv.uScale = getFloat24(op);
}
void GPUCommon::Execute_TexScaleV(u32 op, u32 diff) {
gstate_c.uv.vScale = getFloat24(op);
}
void GPUCommon::Execute_TexOffsetU(u32 op, u32 diff) {
gstate_c.uv.uOff = getFloat24(op);
}
void GPUCommon::Execute_TexOffsetV(u32 op, u32 diff) {
gstate_c.uv.vOff = getFloat24(op);
}
void GPUCommon::Execute_TexLevel(u32 op, u32 diff) {
if (diff == 0xFFFFFFFF) return;
@@ -1451,6 +1440,7 @@ void GPUCommon::Execute_TexLevel(u32 op, u32 diff) {
}
void GPUCommon::Execute_Bezier(u32 op, u32 diff) {
Flush();
// This also make skipping drawing very effective.
framebufferManager_->SetRenderFrameBuffer(gstate_c.IsDirty(DIRTY_FRAMEBUF), gstate_c.skipDrawReason);
if (gstate_c.skipDrawReason & (SKIPDRAW_SKIPFRAME | SKIPDRAW_NON_DISPLAYED_FB)) {
@@ -1494,6 +1484,7 @@ void GPUCommon::Execute_Bezier(u32 op, u32 diff) {
}
void GPUCommon::Execute_Spline(u32 op, u32 diff) {
Flush();
// This also make skipping drawing very effective.
framebufferManager_->SetRenderFrameBuffer(gstate_c.IsDirty(DIRTY_FRAMEBUF), gstate_c.skipDrawReason);
if (gstate_c.skipDrawReason & (SKIPDRAW_SKIPFRAME | SKIPDRAW_NON_DISPLAYED_FB)) {
@@ -1568,6 +1559,7 @@ void GPUCommon::Execute_BoundingBox(u32 op, u32 diff) {
}
void GPUCommon::Execute_BlockTransferStart(u32 op, u32 diff) {
Flush();
// and take appropriate action. This is a block transfer between RAM and VRAM, or vice versa.
// Can we skip this on SkipDraw?
DoBlockTransfer(gstate_c.skipDrawReason);
GPU/GPUCommon.h
+18 -1
View File
@@ -9,7 +9,9 @@
#if defined(__ANDROID__)
#include <atomic>
#elif defined(_M_SSE)
#endif
#if defined(_M_SSE)
#include <emmintrin.h>
#endif
@@ -196,6 +198,21 @@ public:
GPUgstate GetGState() override;
void SetCmdValue(u32 op) override;
void UpdateUVScaleOffset() {
#ifdef _M_SSE
__m128i values = _mm_slli_epi32(_mm_load_si128((const __m128i *)&gstate.texscaleu), 8);
_mm_storeu_si128((__m128i *)&gstate_c.uv, values);
#elif PPSSPP_PLATFORM(ARM_NEON)
const uint32x4_t values = vshlq_n_u32(vld1q_u32(&gstate.texscaleu), 8);
vst1q_u32(&gstate_c.uv, values);
#else
gstate_c.uv.uScale = getFloat24(gstate.texscaleu);
gstate_c.uv.vScale = getFloat24(gstate.texscalev);
gstate_c.uv.uOff = getFloat24(gstate.texoffsetu);
gstate_c.uv.vOff = getFloat24(gstate.texoffsetv);
#endif
}
DisplayList* getList(int listid) override {
return &dls[listid];
}
GPU/Vulkan/GPU_Vulkan.cpp
+29 -17
View File
@@ -56,7 +56,6 @@ struct VulkanCommandTableEntry {
GPU_Vulkan::CommandInfo GPU_Vulkan::cmdInfo_[256];
// This table gets crunched into a faster form by init.
// TODO: Share this table between the backends. Will have to make another indirection for the function pointers though..
static const VulkanCommandTableEntry commandTable[] = {
// Changes that dirty the current texture.
{ GE_CMD_TEXSIZE0, FLAG_FLUSHBEFOREONCHANGE | FLAG_EXECUTE, 0, &GPU_Vulkan::Execute_TexSize0 },
@@ -72,7 +71,6 @@ static const VulkanCommandTableEntry commandTable[] = {
// Changes that trigger data copies. Only flushing on change for LOADCLUT must be a bit of a hack...
{ GE_CMD_LOADCLUT, FLAG_FLUSHBEFOREONCHANGE | FLAG_EXECUTE, 0, &GPU_Vulkan::Execute_LoadClut },
{ GE_CMD_TRANSFERSTART, FLAG_FLUSHBEFORE | FLAG_EXECUTE | FLAG_READS_PC, 0, &GPUCommon::Execute_BlockTransferStart },
};
GPU_Vulkan::GPU_Vulkan(GraphicsContext *gfxCtx, Draw::DrawContext *draw)
@@ -398,22 +396,29 @@ void GPU_Vulkan::FastRunLoop(DisplayList &list) {
// We know that display list PCs have the upper nibble == 0 - no need to mask the pointer
const u32 op = *(const u32 *)(Memory::base + list.pc);
const u32 cmd = op >> 24;
const CommandInfo info = cmdInfo[cmd];
const u8 cmdFlags = info.flags; // If we stashed the cmdFlags in the top bits of the cmdmem, we could get away with one table lookup instead of two
const CommandInfo &info = cmdInfo[cmd];
const u32 diff = op ^ gstate.cmdmem[cmd];
// Inlined CheckFlushOp here to get rid of the dumpThisFrame_ check.
if ((cmdFlags & FLAG_FLUSHBEFORE) || (diff && (cmdFlags & FLAG_FLUSHBEFOREONCHANGE))) {
drawEngine_.Flush();
}
gstate.cmdmem[cmd] = op; // TODO: no need to write if diff==0...
if ((cmdFlags & FLAG_EXECUTE) || (diff && (cmdFlags & FLAG_EXECUTEONCHANGE))) {
downcount = dc;
(this->*info.func)(op, diff);
dc = downcount;
} else if (diff) {
uint64_t dirty = info.flags >> 8;
if (dirty)
gstate_c.Dirty(dirty);
if (diff == 0) {
if (info.flags & FLAG_EXECUTE) {
downcount = dc;
(this->*info.func)(op, diff);
dc = downcount;
}
} else {
uint64_t flags = info.flags;
if (flags & FLAG_FLUSHBEFOREONCHANGE) {
drawEngine_.Flush();
}
gstate.cmdmem[cmd] = op;
if (flags & (FLAG_EXECUTE | FLAG_EXECUTEONCHANGE)) {
downcount = dc;
(this->*info.func)(op, diff);
dc = downcount;
} else {
uint64_t dirty = flags >> 8;
if (dirty)
gstate_c.Dirty(dirty);
}
}
list.pc += 4;
}
@@ -507,6 +512,7 @@ void GPU_Vulkan::Execute_Prim(u32 op, u32 diff) {
#endif
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitPrim(verts, inds, prim, count, gstate.vertType, &bytesRead);
int vertexCost = EstimatePerVertexCost() * count;
@@ -551,6 +557,8 @@ void GPU_Vulkan::Execute_VertexTypeSkinning(u32 op, u32 diff) {
}
void GPU_Vulkan::Execute_Bezier(u32 op, u32 diff) {
Flush();
// We don't dirty on normal changes anymore as we prescale, but it's needed for splines/bezier.
gstate_c.Dirty(DIRTY_UVSCALEOFFSET);
@@ -601,6 +609,7 @@ void GPU_Vulkan::Execute_Bezier(u32 op, u32 diff) {
}
}
UpdateUVScaleOffset();
drawEngine_.SubmitBezier(control_points, indices, gstate.getPatchDivisionU(), gstate.getPatchDivisionV(), bz_ucount, bz_vcount, patchPrim, computeNormals, patchFacing, gstate.vertType, &bytesRead);
if (gstate_c.bezier)
@@ -613,6 +622,8 @@ void GPU_Vulkan::Execute_Bezier(u32 op, u32 diff) {
}
void GPU_Vulkan::Execute_Spline(u32 op, u32 diff) {
Flush();
// We don't dirty on normal changes anymore as we prescale, but it's needed for splines/bezier.
gstate_c.Dirty(DIRTY_UVSCALEOFFSET);
@@ -670,6 +681,7 @@ void GPU_Vulkan::Execute_Spline(u32 op, u32 diff) {
}
int bytesRead = 0;
UpdateUVScaleOffset();
drawEngine_.SubmitSpline(control_points, indices, gstate.getPatchDivisionU(), gstate.getPatchDivisionV(), sp_ucount, sp_vcount, sp_utype, sp_vtype, patchPrim, computeNormals, patchFacing, vertType, &bytesRead);
if (gstate_c.spline)