ShuriZma
/
melonDS
mirror of https://github.com/melonDS-emu/melonDS.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

optimize oob rendering further

This commit is contained in:
Jaklyy 2024-07-10 15:19:59 -04:00
parent c7b92df113
commit 660d30baad
2 changed files with 85 additions and 66 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

71
src/GPU3D_Soft.cpp
View File

@ -617,16 +617,17 @@ void SoftRenderer::SetupPolygonLeftEdge(SoftRenderer::RendererPolygon* rp, s32 y
// note: if the end or current position in a slope is above the start point // note: if the end or current position in a slope is above the start point
// it seems to seek forwards(?) until the value overflows at 256 // it seems to seek forwards(?) until the value overflows at 256
// this can be emulated by just adding 256 to them // this can be emulated by just adding 256 to them
s32 y1 = polygon->Vertices[rp->NextVL]->FinalPosition[1];
if constexpr (oob)
{
if (y < polygon->Vertices[rp->CurVL]->FinalPosition[1]) if (y < polygon->Vertices[rp->CurVL]->FinalPosition[1])
y += 256; y += 256;
s32 y1 = polygon->Vertices[rp->NextVL]->FinalPosition[1];
if (y1 < polygon->Vertices[rp->CurVL]->FinalPosition[1]) if (y1 < polygon->Vertices[rp->CurVL]->FinalPosition[1])
y1 += 256; y1 += 256;
}
rp->XL = rp->SlopeL.Setup<oob>(polygon->Vertices[rp->CurVL]->FinalPosition[0], polygon->Vertices[rp->NextVL]->FinalPosition[0], rp->XL = rp->SlopeL.Setup<oob>(polygon->Vertices[rp->CurVL]->FinalPosition[0], polygon->Vertices[rp->NextVL]->FinalPosition[0],
polygon->Vertices[rp->CurVL]->FinalPosition[1], y1, polygon->Vertices[rp->CurVL]->FinalPosition[1], y1, polygon->FinalW[rp->CurVL], polygon->FinalW[rp->NextVL], y);
polygon->FinalW[rp->CurVL], polygon->FinalW[rp->NextVL], y);
} }
template <bool oob> template <bool oob>
@ -655,18 +656,20 @@ void SoftRenderer::SetupPolygonRightEdge(SoftRenderer::RendererPolygon* rp, s32
// note: if the end or current position in a slope is above the start point // note: if the end or current position in a slope is above the start point
// it seems to seek forwards(?) until the value overflows at 256 // it seems to seek forwards(?) until the value overflows at 256
// this can be emulated by just adding 256 to them // this can be emulated by just adding 256 to them
s32 y1 = polygon->Vertices[rp->NextVR]->FinalPosition[1];
if constexpr (oob)
{
if (y < polygon->Vertices[rp->CurVR]->FinalPosition[1]) if (y < polygon->Vertices[rp->CurVR]->FinalPosition[1])
y += 256; y += 256;
s32 y1 = polygon->Vertices[rp->NextVR]->FinalPosition[1];
if (y1 < polygon->Vertices[rp->CurVR]->FinalPosition[1]) if (y1 < polygon->Vertices[rp->CurVR]->FinalPosition[1])
y1 += 256; y1 += 256;
}
rp->XR = rp->SlopeR.Setup<oob>(polygon->Vertices[rp->CurVR]->FinalPosition[0], polygon->Vertices[rp->NextVR]->FinalPosition[0], rp->XR = rp->SlopeR.Setup<oob>(polygon->Vertices[rp->CurVR]->FinalPosition[0], polygon->Vertices[rp->NextVR]->FinalPosition[0],
polygon->Vertices[rp->CurVR]->FinalPosition[1], y1, polygon->Vertices[rp->CurVR]->FinalPosition[1], y1, polygon->FinalW[rp->CurVR], polygon->FinalW[rp->NextVR], y);
polygon->FinalW[rp->CurVR], polygon->FinalW[rp->NextVR], y);
} }
template <bool oob>
void SoftRenderer::SetupPolygon(SoftRenderer::RendererPolygon* rp, Polygon* polygon) const void SoftRenderer::SetupPolygon(SoftRenderer::RendererPolygon* rp, Polygon* polygon) const
{ {
u32 nverts = polygon->NumVertices; u32 nverts = polygon->NumVertices;
@ -717,29 +720,34 @@ void SoftRenderer::SetupPolygon(SoftRenderer::RendererPolygon* rp, Polygon* poly
} }
else else
{ {
// note: if the end or current position in a slope is above the start point
// it seems to seek forwards(?) until the value overflows at 256
// this can be emulated by just adding 256 to them
s32 y = ytop; s32 y = ytop;
s32 y1 = polygon->Vertices[rp->NextVL]->FinalPosition[1];
if (oob)
{
if (y < polygon->Vertices[rp->CurVL]->FinalPosition[1]) if (y < polygon->Vertices[rp->CurVL]->FinalPosition[1])
y += 256; y += 256;
s32 y1 = polygon->Vertices[rp->NextVL]->FinalPosition[1];
if (y1 < polygon->Vertices[rp->CurVL]->FinalPosition[1]) if (y1 < polygon->Vertices[rp->CurVL]->FinalPosition[1])
y1 += 256; y1 += 256;
}
rp->XL = rp->SlopeL.Setup<oob>(polygon->Vertices[rp->CurVL]->FinalPosition[0], polygon->Vertices[rp->NextVL]->FinalPosition[0],
polygon->Vertices[rp->CurVL]->FinalPosition[1], y1, polygon->FinalW[rp->CurVL], polygon->FinalW[rp->NextVL], y);
rp->XL = rp->SlopeL.Setup<true>(polygon->Vertices[rp->CurVL]->FinalPosition[0], polygon->Vertices[rp->NextVL]->FinalPosition[0], y1 = polygon->Vertices[rp->NextVR]->FinalPosition[1];
polygon->Vertices[rp->CurVL]->FinalPosition[1], y1, if constexpr (oob)
polygon->FinalW[rp->CurVL], polygon->FinalW[rp->NextVL], y); {
y = ytop; y = ytop;
if (y < polygon->Vertices[rp->CurVR]->FinalPosition[1]) if (y < polygon->Vertices[rp->CurVR]->FinalPosition[1])
y += 256; y += 256;
y1 = polygon->Vertices[rp->NextVR]->FinalPosition[1];
if (y1 < polygon->Vertices[rp->CurVR]->FinalPosition[1]) if (y1 < polygon->Vertices[rp->CurVR]->FinalPosition[1])
y1 += 256; y1 += 256;
}
rp->XR = rp->SlopeR.Setup<true>(polygon->Vertices[rp->CurVR]->FinalPosition[0], polygon->Vertices[rp->NextVR]->FinalPosition[0], rp->XR = rp->SlopeR.Setup<oob>(polygon->Vertices[rp->CurVR]->FinalPosition[0], polygon->Vertices[rp->NextVR]->FinalPosition[0],
polygon->Vertices[rp->CurVR]->FinalPosition[1], y1, polygon->Vertices[rp->CurVR]->FinalPosition[1], y1, polygon->FinalW[rp->CurVR], polygon->FinalW[rp->NextVR], y);
polygon->FinalW[rp->CurVR], polygon->FinalW[rp->NextVR], y);
} }
} }
@ -769,12 +777,12 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
if (polygon->YTop != polygon->YBottom) if (polygon->YTop != polygon->YBottom)
{ {
if ((y >= polygon->SlopePosition[rp->NextVL][1] || y == polygon->Vertices[rp->CurVL]->FinalPosition[1]) && rp->CurVL != polygon->VBottom) if ((y >= polygon->SlopePosition[rp->NextVL][1] || (oob && y == polygon->Vertices[rp->CurVL]->FinalPosition[1])) && rp->CurVL != polygon->VBottom)
{ {
SetupPolygonLeftEdge<oob>(rp, y); SetupPolygonLeftEdge<oob>(rp, y);
} }
if ((y >= polygon->SlopePosition[rp->NextVR][1] || y == polygon->Vertices[rp->CurVR]->FinalPosition[1]) && rp->CurVR != polygon->VBottom) if ((y >= polygon->SlopePosition[rp->NextVR][1] || (oob && y == polygon->Vertices[rp->CurVR]->FinalPosition[1])) && rp->CurVR != polygon->VBottom)
{ {
SetupPolygonRightEdge<oob>(rp, y); SetupPolygonRightEdge<oob>(rp, y);
} }
@ -882,6 +890,8 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
xend += 1; xend += 1;
Interpolator<0, oob> interpX(xstart, xend, wl, wr); Interpolator<0, oob> interpX(xstart, xend, wl, wr);
if constexpr (oob)
{
// CHECKME: should the unclamped values be used for timings? // CHECKME: should the unclamped values be used for timings?
// negative values are clamped to 0 // negative values are clamped to 0
if (xstart < 0) if (xstart < 0)
@ -891,14 +901,15 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
xstart = 0; xstart = 0;
if (xend < 1) xend = 1; if (xend < 1) xend = 1;
} }
s32 x = xstart;
// too big values are clamped to 511 // too big values are clamped to 511
if (xend > 511) if (xend > 511)
{ {
r_edgelen += 256 - xend; r_edgelen += 256 - xend;
xend = 511; xend = 511;
} }
}
s32 x = xstart;
s32 xlimit; s32 xlimit;
// for shadow masks: set stencil bits where the depth test fails. // for shadow masks: set stencil bits where the depth test fails.
@ -1011,12 +1022,12 @@ void SoftRenderer::RenderPolygonScanline(const GPU& gpu, RendererPolygon* rp, s3
if (polygon->YTop != polygon->YBottom) if (polygon->YTop != polygon->YBottom)
{ {
if ((y >= polygon->SlopePosition[rp->NextVL][1] || y == polygon->Vertices[rp->CurVL]->FinalPosition[1]) && rp->CurVL != polygon->VBottom) if ((y >= polygon->SlopePosition[rp->NextVL][1] || (oob && y == polygon->Vertices[rp->CurVL]->FinalPosition[1])) && rp->CurVL != polygon->VBottom)
{ {
SetupPolygonLeftEdge<oob>(rp, y); SetupPolygonLeftEdge<oob>(rp, y);
} }
if ((y >= polygon->SlopePosition[rp->NextVR][1] || y == polygon->Vertices[rp->CurVR]->FinalPosition[1]) && rp->CurVR != polygon->VBottom) if ((y >= polygon->SlopePosition[rp->NextVR][1] || (oob && y == polygon->Vertices[rp->CurVR]->FinalPosition[1])) && rp->CurVR != polygon->VBottom)
{ {
SetupPolygonRightEdge<oob>(rp, y); SetupPolygonRightEdge<oob>(rp, y);
} }
@ -1149,6 +1160,8 @@ void SoftRenderer::RenderPolygonScanline(const GPU& gpu, RendererPolygon* rp, s3
xend+=1; xend+=1;
Interpolator<0, oob> interpX(xstart, xend, wl, wr); Interpolator<0, oob> interpX(xstart, xend, wl, wr);
if constexpr (oob)
{
// CHECKME: should the unclamped values be used for timings? // CHECKME: should the unclamped values be used for timings?
// negative values are clamped to 0 // negative values are clamped to 0
if (xstart < 0) if (xstart < 0)
@ -1158,13 +1171,15 @@ void SoftRenderer::RenderPolygonScanline(const GPU& gpu, RendererPolygon* rp, s3
xstart = 0; xstart = 0;
if (xend < 1) xend = 1; if (xend < 1) xend = 1;
} }
s32 x = xstart;
// too big values are clamped to 511 // too big values are clamped to 511
if (xend > 511) if (xend > 511)
{ {
r_edgelen += 256 - xend; r_edgelen += 256 - xend;
xend = 511; xend = 511;
} }
}
s32 x = xstart;
s32 xlimit; s32 xlimit;
s32 xcov = 0; s32 xcov = 0;
@ -1471,7 +1486,7 @@ void SoftRenderer::RenderScanline(const GPU& gpu, s32 y, int npolys)
if (y >= polygon->YTop && (y < polygon->YBottom || (y == polygon->YTop && polygon->YBottom == polygon->YTop))) if (y >= polygon->YTop && (y < polygon->YBottom || (y == polygon->YTop && polygon->YBottom == polygon->YTop)))
{ {
if (polygon->OOBRendering) if (polygon->OOBRendering) [[unlikely]] // enable some extra handling for polygons that might render out of ordinary bounds
{ {
if (polygon->IsShadowMask) if (polygon->IsShadowMask)
RenderShadowMaskScanline<true>(gpu.GPU3D, rp, y); RenderShadowMaskScanline<true>(gpu.GPU3D, rp, y);
@ -1793,7 +1808,11 @@ void SoftRenderer::RenderPolygons(const GPU& gpu, bool threaded, Polygon** polyg
for (int i = 0; i < npolys; i++) for (int i = 0; i < npolys; i++)
{ {
if (polygons[i]->Degenerate) continue; if (polygons[i]->Degenerate) continue;
SetupPolygon(&PolygonList[j++], polygons[i]);
if (polygons[i]->OOBRendering) [[unlikely]] // enable some extra handling for polygons that might render out of ordinary bounds
SetupPolygon<true>(&PolygonList[j++], polygons[i]);
else
SetupPolygon<false>(&PolygonList[j++], polygons[i]);
} }
RenderScanline(gpu, 0, j); RenderScanline(gpu, 0, j);

4
src/GPU3D_Soft.h
View File

@ -293,7 +293,7 @@ private:
// note: for some reason, x/y isn't calculated directly, // note: for some reason, x/y isn't calculated directly,
// instead, 1/y is calculated and then multiplied by x // instead, 1/y is calculated and then multiplied by x
// TODO: this is still not perfect (see for example x=169 y=33) // TODO: this is still not perfect (see for example x=169 y=33)
if (oob && ylen == 0) // this case *should* only be triggered by glitched polygons that try to render oob if (ylen == 0)
Increment = xlen << 18; Increment = xlen << 18;
else if (ylen == xlen && xlen != 1) else if (ylen == xlen && xlen != 1)
Increment = 0x40000; Increment = 0x40000;
@ -468,7 +468,7 @@ private:
void PlotTranslucentPixel(const GPU3D& gpu3d, u32 pixeladdr, u32 color, u32 z, u32 polyattr, u32 shadow); void PlotTranslucentPixel(const GPU3D& gpu3d, u32 pixeladdr, u32 color, u32 z, u32 polyattr, u32 shadow);
template<bool oob> void SetupPolygonLeftEdge(RendererPolygon* rp, s32 y) const; template<bool oob> void SetupPolygonLeftEdge(RendererPolygon* rp, s32 y) const;
template<bool oob> void SetupPolygonRightEdge(RendererPolygon* rp, s32 y) const; template<bool oob> void SetupPolygonRightEdge(RendererPolygon* rp, s32 y) const;
void SetupPolygon(RendererPolygon* rp, Polygon* polygon) const; template<bool oob> void SetupPolygon(RendererPolygon* rp, Polygon* polygon) const;
template<bool oob> void RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon* rp, s32 y); template<bool oob> void RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon* rp, s32 y);
template<bool oob> void RenderPolygonScanline(const GPU& gpu, RendererPolygon* rp, s32 y); template<bool oob> void RenderPolygonScanline(const GPU& gpu, RendererPolygon* rp, s32 y);
void RenderScanline(const GPU& gpu, s32 y, int npolys); void RenderScanline(const GPU& gpu, s32 y, int npolys);