Emulation/ppsspp
1
0
Fork 0
mirror of https://github.com/hrydgard/ppsspp.git synced 2025-04-02 11:01:50 -04:00
Code Issues Projects Releases Packages Wiki Activity

Merge pull request #19789 from hrydgard/depth-raster-low-quality

Browse source 
Implement the low-quality depth raster mode, default to it on Android/iOS
This commit is contained in:
Henrik Rydgård 2024-12-31 13:10:54 +01:00 • committed by GitHub
commit f88b6abb5d
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
5 changed files with 69 additions and 19 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
Common/Math/CrossSIMD.h
View file

@ -160,7 +160,7 @@ struct Vec4S32 {
Vec4S32 Mul(Vec4S32 other) const { return *this * other; }
template<int imm>
Vec4S32 Shl() const { return Vec4S32{ _mm_slli_epi32(v, imm) }; }
Vec4S32 Shl() const { return Vec4S32{ imm == 0 ? v : _mm_slli_epi32(v, imm) }; }
// NOTE: May be slow.
int operator[](size_t index) const { return ((int *)&v)[index]; }

10
Core/Config.cpp
View file

@ -139,6 +139,14 @@ const char *DefaultLangRegion() {
return defaultLangRegion.c_str();
}
static int DefaultDepthRaster() {
#if PPSSPP_PLATFORM(ANDROID) || PPSSPP_PLATFORM(IOS)
return (int)DepthRasterMode::LOW_QUALITY;
#else
return (int)DepthRasterMode::DEFAULT;
#endif
}
std::string CreateRandMAC() {
std::stringstream randStream;
srand(time(nullptr));
@ -616,7 +624,7 @@ static const ConfigSetting graphicsSettings[] = {
ConfigSetting("UseGeometryShader", &g_Config.bUseGeometryShader, false, CfgFlag::PER_GAME),
ConfigSetting("SkipBufferEffects", &g_Config.bSkipBufferEffects, false, CfgFlag::PER_GAME | CfgFlag::REPORT),
ConfigSetting("DisableRangeCulling", &g_Config.bDisableRangeCulling, false, CfgFlag::PER_GAME | CfgFlag::REPORT),
ConfigSetting("DepthRasterMode", &g_Config.iDepthRasterMode, 0, CfgFlag::PER_GAME | CfgFlag::REPORT),
ConfigSetting("DepthRasterMode", &g_Config.iDepthRasterMode, &DefaultDepthRaster, CfgFlag::PER_GAME | CfgFlag::REPORT),
ConfigSetting("SoftwareRenderer", &g_Config.bSoftwareRendering, false, CfgFlag::PER_GAME),
ConfigSetting("SoftwareRendererJit", &g_Config.bSoftwareRenderingJit, true, CfgFlag::PER_GAME),
ConfigSetting("HardwareTransform", &g_Config.bHardwareTransform, true, CfgFlag::PER_GAME | CfgFlag::REPORT),

71
GPU/Common/DepthRaster.cpp
View file

@ -87,12 +87,6 @@ static void DepthRasterRect(uint16_t *dest, int stride, const DepthScissor sciss
alignas(16) static const int zero123[4] = {0, 1, 2, 3};
constexpr int stepXSize = 4;
constexpr int stepYSize = 1;
constexpr int stepXShift = 2;
constexpr int stepYShift = 0;
enum class TriangleStat {
OK,
NoPixels,
@ -102,7 +96,7 @@ enum class TriangleStat {
constexpr int MIN_TWICE_TRI_AREA = 10;
// A mix of ideas from Intel's sample and ryg's rasterizer blog series.
template<ZCompareMode compareMode>
template<ZCompareMode compareMode, bool lowQ>
void DepthRaster4Triangles(int stats[3], uint16_t *depthBuf, int stride, DepthScissor scissor, const int *tx, const int *ty, const float *tz) {
// Triangle setup. This is done using SIMD, four triangles at a time.
// 16x16->32 multiplications are doable on SSE2, which should be all we need.
@ -117,6 +111,12 @@ void DepthRaster4Triangles(int stats[3], uint16_t *depthBuf, int stride, DepthSc
Vec4S32 x2 = Vec4S32::LoadAligned(tx + 8);
Vec4S32 y2 = Vec4S32::LoadAligned(ty + 8);
if (lowQ) {
y0 &= Vec4S32::Splat(~1);
y1 &= Vec4S32::Splat(~1);
y2 &= Vec4S32::Splat(~1);
}
// FixupAfterMinMax is just 16->32 sign extension, in case the current platform (like SSE2) just has 16-bit min/max operations.
Vec4S32 minX = x0.Min16(x1).Min16(x2).Max16(Vec4S32::Splat(scissor.x1)).FixupAfterMinMax();
Vec4S32 maxX = x0.Max16(x1).Max16(x2).Min16(Vec4S32::Splat(scissor.x2)).FixupAfterMinMax();
@ -138,6 +138,12 @@ void DepthRaster4Triangles(int stats[3], uint16_t *depthBuf, int stride, DepthSc
Vec4S32 B01 = x1 - x0;
Vec4S32 C01 = x0.Mul16(y1) - y0.Mul16(x1);
constexpr int stepXSize = 4;
constexpr int stepYSize = lowQ ? 2 : 1;
constexpr int stepXShift = 2;
constexpr int stepYShift = lowQ ? 1 : 0;
// Step deltas
Vec4S32 stepX12 = A12.Shl<stepXShift>();
Vec4S32 stepY12 = B12.Shl<stepYShift>();
@ -229,22 +235,27 @@ void DepthRaster4Triangles(int stats[3], uint16_t *depthBuf, int stride, DepthSc
Vec4U16 shortZ = Vec4U16::FromVec4F32(zs);
// This switch is on a templated constant, so should collapse away.
Vec4U16 writeVal;
switch (compareMode) {
case ZCompareMode::Greater:
// To implement the greater/greater-than comparison, we can combine mask and max.
// Unfortunately there's no unsigned max on SSE2, it's synthesized by xoring 0x8000 on input and output.
// We use AndNot to zero out Z results, before doing Max with the buffer.
shortZ.AndNot(shortMaskInv).Max(bufferValues).Store(rowPtr + x);
writeVal = shortZ.AndNot(shortMaskInv).Max(bufferValues);
break;
case ZCompareMode::Less:
// This time, we OR the mask and use .Min.
(shortZ | shortMaskInv).Min(bufferValues).Store(rowPtr + x);
writeVal = (shortZ | shortMaskInv).Min(bufferValues);
break;
case ZCompareMode::Always: // UNTESTED
// This could be replaced with a vblend operation.
((bufferValues & shortMaskInv) | shortZ.AndNot(shortMaskInv)).Store(rowPtr + x);
writeVal = ((bufferValues & shortMaskInv) | shortZ.AndNot(shortMaskInv));
break;
}
writeVal.Store(rowPtr + x);
if (lowQ) {
writeVal.Store(rowPtr + stride + x);
}
}
}
@ -544,7 +555,7 @@ int DepthRasterClipIndexedTriangles(int *tx, int *ty, float *tz, const float *tr
}
// Rasterizes screen-space vertices.
void DepthRasterScreenVerts(uint16_t *depth, int depthStride, const int *tx, const int *ty, const float *tz, int count, const DepthDraw &draw, const DepthScissor scissor) {
void DepthRasterScreenVerts(uint16_t *depth, int depthStride, const int *tx, const int *ty, const float *tz, int count, const DepthDraw &draw, const DepthScissor scissor, bool lowQ) {
// Prim should now be either TRIANGLES or RECTs.
_dbg_assert_(draw.prim == GE_PRIM_RECTANGLES || draw.prim == GE_PRIM_TRIANGLES);
@ -562,21 +573,51 @@ void DepthRasterScreenVerts(uint16_t *depth, int depthStride, const int *tx, con
{
int stats[3]{};
// Batches of 4 triangles, as output by the clip function.
for (int i = 0; i < count; i += 12) {
if (lowQ) {
switch (draw.compareMode) {
case ZCompareMode::Greater:
{
DepthRaster4Triangles<ZCompareMode::Greater>(stats, depth, depthStride, scissor, &tx[i], &ty[i], &tz[i]);
for (int i = 0; i < count; i += 12) {
DepthRaster4Triangles<ZCompareMode::Greater, true>(stats, depth, depthStride, scissor, &tx[i], &ty[i], &tz[i]);
}
break;
}
case ZCompareMode::Less:
{
DepthRaster4Triangles<ZCompareMode::Less>(stats, depth, depthStride, scissor, &tx[i], &ty[i], &tz[i]);
for (int i = 0; i < count; i += 12) {
DepthRaster4Triangles<ZCompareMode::Less, true>(stats, depth, depthStride, scissor, &tx[i], &ty[i], &tz[i]);
}
break;
}
case ZCompareMode::Always:
{
DepthRaster4Triangles<ZCompareMode::Always>(stats, depth, depthStride, scissor, &tx[i], &ty[i], &tz[i]);
for (int i = 0; i < count; i += 12) {
DepthRaster4Triangles<ZCompareMode::Always, true>(stats, depth, depthStride, scissor, &tx[i], &ty[i], &tz[i]);
}
break;
}
}
} else {
switch (draw.compareMode) {
case ZCompareMode::Greater:
{
for (int i = 0; i < count; i += 12) {
DepthRaster4Triangles<ZCompareMode::Greater, false>(stats, depth, depthStride, scissor, &tx[i], &ty[i], &tz[i]);
}
break;
}
case ZCompareMode::Less:
{
for (int i = 0; i < count; i += 12) {
DepthRaster4Triangles<ZCompareMode::Less, false>(stats, depth, depthStride, scissor, &tx[i], &ty[i], &tz[i]);
}
break;
}
case ZCompareMode::Always:
{
for (int i = 0; i < count; i += 12) {
DepthRaster4Triangles<ZCompareMode::Always, false>(stats, depth, depthStride, scissor, &tx[i], &ty[i], &tz[i]);
}
break;
}
}

2
GPU/Common/DepthRaster.h
View file

@ -56,4 +56,4 @@ int DepthRasterClipIndexedRectangles(int *tx, int *ty, float *tz, const float *t
void DecodeAndTransformForDepthRaster(float *dest, const float *worldviewproj, const void *vertexData, int indexLowerBound, int indexUpperBound, VertexDecoder *dec, u32 vertTypeID);
void TransformPredecodedForDepthRaster(float *dest, const float *worldviewproj, const void *decodedVertexData, VertexDecoder *dec, int count);
void ConvertPredecodedThroughForDepthRaster(float *dest, const void *decodedVertexData, VertexDecoder *dec, int count);
void DepthRasterScreenVerts(uint16_t *depth, int depthStride, const int *tx, const int *ty, const float *tz, int count, const DepthDraw &draw, const DepthScissor scissor);
void DepthRasterScreenVerts(uint16_t *depth, int depthStride, const int *tx, const int *ty, const float *tz, int count, const DepthDraw &draw, const DepthScissor scissor, bool lowQ);

3
GPU/Common/DrawEngineCommon.cpp
View file

@ -1098,6 +1098,7 @@ 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_;
@ -1127,7 +1128,7 @@ void DrawEngineCommon::FlushQueuedDepth() {
}
{
TimeCollector collectStat(&gpuStats.msRasterizeDepth, collectStats);
DepthRasterScreenVerts((uint16_t *)Memory::GetPointerWrite(draw.depthAddr), draw.depthStride, tx, ty, tz, outVertCount, draw, tileScissor);
DepthRasterScreenVerts((uint16_t *)Memory::GetPointerWrite(draw.depthAddr), draw.depthStride, tx, ty, tz, outVertCount, draw, tileScissor, lowQ);
}
}