Merge pull request #21511 from hrydgard/ge-improvements

Depth raster: Handle triangles properly when backface culling is disabled
This commit is contained in:
Henrik Rydgård
2026-03-31 18:41:44 -06:00
committed by GitHub
5 changed files with 35 additions and 15 deletions
+13 -3
View File
@@ -409,6 +409,8 @@ int DepthRasterClipIndexedTriangles(int *tx, int *ty, float *tz, const float *tr
return true;
};
const bool cullDisabled = !draw.cullEnabled;
for (int i = 0; i < count; i += 3) {
// Collect valid triangles into buffer.
const float *v0 = transformed + indexBuffer[i] * 4;
@@ -420,6 +422,14 @@ int DepthRasterClipIndexedTriangles(int *tx, int *ty, float *tz, const float *tr
verts[collected] = v0;
verts[collected + 1] = v1;
verts[collected + 2] = v2;
if (cullDisabled) {
collected += 3;
// Add the reverse triangle too. We could alternatively handle culling later in the pipeline, but mainly
// Syphon Filter needs this (issue #21498) and this simplifies things.
verts[collected] = v0;
verts[collected + 1] = v2;
verts[collected + 2] = v1;
}
collected += 3;
} else {
planeCulled++;
@@ -433,12 +443,12 @@ int DepthRasterClipIndexedTriangles(int *tx, int *ty, float *tz, const float *tr
collected = 12;
}
if (collected != 12) {
if (collected < 12) {
// Fetch more!
continue;
}
collected = 0;
collected -= 12;
// These names are wrong .. until we transpose.
Vec4F32 x0 = Vec4F32::Load(verts[0]);
@@ -505,7 +515,7 @@ int DepthRasterClipIndexedTriangles(int *tx, int *ty, float *tz, const float *tr
x2 &= inGuardBand;
// Floating point double triangle area. Can't be reused for the integer-snapped raster reliably (though may work...)
// Still good for culling early and pretty cheap to compute.
// Still good for backface culling early and pretty cheap to compute.
Vec4F32 doubleTriArea = (x1 - x0) * (y2 - y0) - (x2 - x0) * (y1 - y0) - Vec4F32::Splat((float)(MIN_TWICE_TRI_AREA));
if (!AnyZeroSignBit(doubleTriArea)) {
gpuStats.numDepthRasterEarlySize += 4;
+6 -3
View File
@@ -1141,7 +1141,6 @@ void DrawEngineCommon::FlushQueuedDepth() {
const bool collectStats = coreCollectDebugStats;
const bool lowQ = g_Config.iDepthRasterMode == (int)DepthRasterMode::LOW_QUALITY;
for (const auto &draw : depthDraws_) {
int *tx = depthScreenVerts_;
int *ty = depthScreenVerts_ + DEPTH_SCREENVERTS_COMPONENT_COUNT;
@@ -1168,9 +1167,13 @@ void DrawEngineCommon::FlushQueuedDepth() {
break;
}
}
{
if (outVertCount > 0) {
TimeCollector collectStat(&gpuStats.msRasterizeDepth, collectStats);
DepthRasterScreenVerts((uint16_t *)Memory::GetPointerWrite(draw.depthAddr), draw.depthStride, tx, ty, tz, outVertCount, draw, tileScissor, lowQ);
if (!Memory::IsValid4AlignedAddress(draw.depthAddr)) {
continue;
}
u16 *depthPtr = (uint16_t *)Memory::GetPointerWriteUnchecked(draw.depthAddr);
DepthRasterScreenVerts(depthPtr, draw.depthStride, tx, ty, tz, outVertCount, draw, tileScissor, lowQ);
}
}
+3
View File
@@ -1344,6 +1344,9 @@ void DrawMediaDecodersView(ImConfig &cfg, ImControl &control) {
if (ctx->BufferState() == ATRAC_STATUS_ALL_DATA_LOADED) {
if (ImGui::Button("Save to disk...")) {
System_BrowseForFileSave(cfg.requesterToken, "Save AT3 file", "song.at3", BrowseFileType::ATRAC3, [=](const std::string &filename, int) {
if (!Memory::IsValidRange(info.buffer, info.bufferByte)) {
return;
}
const u8 *data = Memory::GetPointerRange(info.buffer, info.bufferByte);
if (!data) {
return;
+12 -8
View File
@@ -426,31 +426,34 @@ bool ImGePixelViewer::FormatValueAt(char *buf, size_t bufSize, int x, int y) con
// Go look directly in RAM.
int bpp = BufferFormatBytesPerPixel(format);
u32 pixelAddr = addr + (y * stride + x) * bpp;
if (!Memory::IsValidAddress(pixelAddr)) {
return false;
}
switch (format) {
case GE_FORMAT_8888:
snprintf(buf, bufSize, "%08x", Memory::Read_U32(pixelAddr));
snprintf(buf, bufSize, "%08x", Memory::ReadUnchecked_U32(pixelAddr));
break;
case GE_FORMAT_4444:
{
u16 raw = Memory::Read_U16(pixelAddr);
u16 raw = Memory::ReadUnchecked_U16(pixelAddr);
snprintf(buf, bufSize, "%08x (raw: %04x)", RGBA4444ToRGBA8888(raw), raw);
break;
}
case GE_FORMAT_565:
{
u16 raw = Memory::Read_U16(pixelAddr);
u16 raw = Memory::ReadUnchecked_U16(pixelAddr);
snprintf(buf, bufSize, "%08x (raw: %04x)", RGB565ToRGBA8888(raw), raw);
break;
}
case GE_FORMAT_5551:
{
u16 raw = Memory::Read_U16(pixelAddr);
u16 raw = Memory::ReadUnchecked_U16(pixelAddr);
snprintf(buf, bufSize, "%08x (raw: %04x)", RGBA5551ToRGBA8888(raw), raw);
break;
}
case GE_FORMAT_DEPTH16:
{
u16 raw = Memory::Read_U16(pixelAddr);
u16 raw = Memory::ReadUnchecked_U16(pixelAddr);
snprintf(buf, bufSize, "%0.4f (raw: %04x / %d)", (float)raw / 65535.0f, raw, raw);
break;
}
@@ -779,7 +782,7 @@ void ImGeDisasmView::Draw(GPUDebugInterface *gpuDebug) {
if (Memory::IsValid4AlignedAddress(addr)) {
draw_list->AddText(lineStart, 0xFFC0C0C0, addrBuffer);
u32 opcode = Memory::Read_U32(addr);
u32 opcode = Memory::ReadUnchecked_U32(addr);
GPUDebugOp op = gpuDebug->DisassembleOp(addr, opcode);
u32 color = 0xFFFFFFFF;
char temp[16];
@@ -842,7 +845,7 @@ void ImGeDisasmView::Draw(GPUDebugInterface *gpuDebug) {
}
} else if (Memory::IsValid4AlignedAddress(dragAddr_)) {
char buffer[64];
u32 opcode = Memory::Read_U32(dragAddr_);
u32 opcode = Memory::ReadUnchecked_U32(dragAddr_);
GPUDebugOp op = gpuDebug->DisassembleOp(dragAddr_, opcode);
// affect dragAddr_?
if (ImGui::MenuItem("Copy Address", NULL, false)) {
@@ -1182,7 +1185,8 @@ void ImGeDebuggerWindow::Draw(ImConfig &cfg, ImControl &control, GPUDebugInterfa
DisplayList list;
bool isOnBlockTransfer = false;
if (gpuDebug->GetCurrentDisplayList(list)) {
op = Memory::Read_U32(list.pc);
_dbg_assert_(Memory::IsValid4AlignedAddress(list.pc));
op = Memory::ReadUnchecked_U32(list.pc);
// TODO: Also add support for block transfer previews!
+1 -1
View File
@@ -998,7 +998,7 @@ void ImMemDumpWindow::Draw(ImConfig &cfg, MIPSDebugInterface *debug) {
errorMsg_ = "Couldn't open file for writing";
} else {
if (mode_ == MemDumpMode::Raw) {
const uint8_t *ptr = Memory::GetPointer(address_);
const uint8_t *ptr = Memory::GetPointerUnchecked(address_);
fwrite(ptr, 1, size_, file);
} else {
std::string disassembly = DisassembleRange(address_, size_, true, debug);