piwalker/melonDS
1
0
Fork 0
mirror of https://github.com/melonDS-emu/melonDS.git synced 2024-11-14 13:27:41 -07:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

dont use templates

Browse Source 
bigger code <<< slower code
This commit is contained in:
Jaklyy 2023-12-10 13:03:54 -05:00
parent 63a39b130e
commit 785fab024f
2 changed files with 26 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

47
src/GPU3D_Soft.cpp
View File

@ -112,10 +112,9 @@ void SoftRenderer::SetThreaded(bool threaded) noexcept
} }
} }
template <bool odd> bool SoftRenderer::DoTimings(s32 cycles, bool odd)
bool SoftRenderer::DoTimings(s32 cycles)
{ {
if constexpr (odd) if (odd)
{ {
RasterTimingCounterOdd += cycles; RasterTimingCounterOdd += cycles;
if ((RasterTimingCounterOdd + RasterTimingCounterPrev) < RasterTimingCap) return false; if ((RasterTimingCounterOdd + RasterTimingCounterPrev) < RasterTimingCap) return false;
@ -130,10 +129,9 @@ bool SoftRenderer::DoTimings(s32 cycles)
return true; return true;
} }
template <bool odd> void SoftRenderer::EndScanline(bool odd)
void SoftRenderer::EndScanline()
{ {
if constexpr (!odd) if (!odd)
{ {
RasterTimingCounterPrev += std::max(RasterTimingCounterOdd, RasterTimingCounterEven); RasterTimingCounterPrev += std::max(RasterTimingCounterOdd, RasterTimingCounterEven);
RasterTimingCounterPrev -= PerScanlineRecup; // wip RasterTimingCounterPrev -= PerScanlineRecup; // wip
@ -962,8 +960,7 @@ void SoftRenderer::RenderShadowMaskScanline(RendererPolygon* rp, s32 y)
rp->XR = rp->SlopeR.Step(); rp->XR = rp->SlopeR.Step();
} }
template <bool odd> bool SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y, bool odd)
bool SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
{ {
Polygon* polygon = rp->PolyData; Polygon* polygon = rp->PolyData;
u32 polyattr = (polygon->Attr & 0x3F008000); u32 polyattr = (polygon->Attr & 0x3F008000);
@ -984,7 +981,7 @@ bool SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
CheckSlope(rp, y); CheckSlope(rp, y);
if (DoTimings<odd>(PerPolyTiming)) return Step(rp, true); if (DoTimings(PerPolyTiming, odd)) return Step(rp, true);
Vertex *vlcur, *vlnext, *vrcur, *vrnext; Vertex *vlcur, *vlnext, *vrcur, *vrnext;
s32 xstart, xend; s32 xstart, xend;
@ -1133,7 +1130,7 @@ bool SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
for (; x < xlimit; x++) for (; x < xlimit; x++)
{ {
if (pixelsrendered >= 4 && DoTimings<odd>(PerPixelTiming)) return Step(rp, true); if (pixelsrendered >= 4 && DoTimings(PerPixelTiming, odd)) return Step(rp, true);
pixelsrendered++; pixelsrendered++;
if (!l_filledge) continue; if (!l_filledge) continue;
@ -1232,7 +1229,7 @@ bool SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
for (; x < xlimit; x++) for (; x < xlimit; x++)
{ {
if (pixelsrendered >= 4 && DoTimings<odd>(PerPixelTiming)) return Step(rp, true); if (pixelsrendered >= 4 && DoTimings(PerPixelTiming, odd)) return Step(rp, true);
pixelsrendered++; pixelsrendered++;
if (wireframe && !edge) continue; if (wireframe && !edge) continue;
@ -1328,7 +1325,7 @@ bool SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
//if (rp->SlopeR.Increment != 0 && DoTimings(PerRightSlope, odd)) return Step(rp, true); // should be fine to not immediately return? might be wrong //if (rp->SlopeR.Increment != 0 && DoTimings(PerRightSlope, odd)) return Step(rp, true); // should be fine to not immediately return? might be wrong
for (; x < xlimit; x++) for (; x < xlimit; x++)
{ {
if (pixelsrendered >= 4 && DoTimings<odd>(PerPixelTiming)) return Step(rp, true); if (pixelsrendered >= 4 && DoTimings(PerPixelTiming, odd)) return Step(rp, true);
pixelsrendered++; pixelsrendered++;
if (!r_filledge) continue; if (!r_filledge) continue;
@ -1421,8 +1418,7 @@ bool SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
return Step(rp, false); return Step(rp, false);
} }
template <bool odd> bool SoftRenderer::RenderScanline(s32 y, int npolys, bool odd)
bool SoftRenderer::RenderScanline(s32 y, int npolys)
{ {
bool abort = false; bool abort = false;
for (int i = 0; i < npolys; i++) for (int i = 0; i < npolys; i++)
@ -1437,19 +1433,19 @@ bool SoftRenderer::RenderScanline(s32 y, int npolys)
} }
else if (y == polygon->YBottom && y != polygon->YTop) else if (y == polygon->YBottom && y != polygon->YTop)
{ {
if (DoTimings<odd>(EmptyPolyScanline)) abort = true; if (DoTimings(EmptyPolyScanline, odd)) abort = true;
} }
else if (y >= polygon->YTop && (y < polygon->YBottom || (y == polygon->YTop && polygon->YBottom == polygon->YTop))) else if (y >= polygon->YTop && (y < polygon->YBottom || (y == polygon->YTop && polygon->YBottom == polygon->YTop)))
{ {
if (polygon->IsShadowMask) if (polygon->IsShadowMask)
;//RenderShadowMaskScanline(rp, y); ;//RenderShadowMaskScanline(rp, y);
else else
if (RenderPolygonScanline<odd>(rp, y)) abort = true; if (RenderPolygonScanline(rp, y, odd)) abort = true;
} }
} }
BufferOffset = (BufferOffset + 1) & 0x7; // loop if == 8 BufferOffset = (BufferOffset + 1) & 0x7; // loop if == 8
EndScanline<odd>(); EndScanline(odd);
return abort; return abort;
} }
@ -1492,14 +1488,13 @@ u32 SoftRenderer::CalculateFogDensity(u32 pixeladdr)
return density; return density;
} }
template<bool odd, bool finish> void SoftRenderer::ScanlineFinalPass(s32 y, u8 rdbufferoffset, bool late, bool odd, bool finish)
void SoftRenderer::ScanlineFinalPass(s32 y, u8 rdbufferoffset, bool late)
{ {
// to consider: // to consider:
// clearing all polygon fog flags if the master flag isn't set? // clearing all polygon fog flags if the master flag isn't set?
// merging all final pass loops into one? // merging all final pass loops into one?
u8 tempoffset; u8 tempoffset;
if constexpr (finish) if (finish)
tempoffset = (BufferOffset - 2 + (!odd)); tempoffset = (BufferOffset - 2 + (!odd));
else else
tempoffset = (BufferOffset - 4 + (!odd)); tempoffset = (BufferOffset - 4 + (!odd));
@ -1761,8 +1756,8 @@ void SoftRenderer::RenderPolygons(bool threaded, Polygon** polygons, int npolys)
for (u8 bufferline = 0; bufferline < 48; bufferline += 2) for (u8 bufferline = 0; bufferline < 48; bufferline += 2)
{ {
ClearBuffers(y); ClearBuffers(y);
latebuffer[y] = RenderScanline<true>(y, j); latebuffer[y] = RenderScanline(y, j, true);
latebuffer[y+1] = RenderScanline<false>(y+1, j); latebuffer[y+1] = RenderScanline(y+1, j, false);
if (prevbufferline >= 0) if (prevbufferline >= 0)
@ -1771,8 +1766,8 @@ void SoftRenderer::RenderPolygons(bool threaded, Polygon** polygons, int npolys)
if (!latebuffer[y-1]) latebuffer[y+1] = false; if (!latebuffer[y-1]) latebuffer[y+1] = false;
ScanlineFinalPass<true, false>(y-2, prevbufferline, latebuffer[y-2]); ScanlineFinalPass(y-2, prevbufferline, latebuffer[y-2], true, false);
ScanlineFinalPass<false, false>(y-1, prevbufferline+1, latebuffer[y-1]); ScanlineFinalPass(y-1, prevbufferline+1, latebuffer[y-1], false, false);
} }
y += 2; y += 2;
@ -1782,8 +1777,8 @@ void SoftRenderer::RenderPolygons(bool threaded, Polygon** polygons, int npolys)
Platform::Semaphore_Post(Sema_ScanlineCount); Platform::Semaphore_Post(Sema_ScanlineCount);
} }
ScanlineFinalPass<true, true>(190, prevbufferline, latebuffer[190]); ScanlineFinalPass(190, prevbufferline, latebuffer[190], true, true);
ScanlineFinalPass<false, true>(191, prevbufferline+1, latebuffer[191]); ScanlineFinalPass(191, prevbufferline+1, latebuffer[191], false, true);
if (threaded) if (threaded)
Platform::Semaphore_Post(Sema_ScanlineCount); Platform::Semaphore_Post(Sema_ScanlineCount);

10
src/GPU3D_Soft.h
View File

@ -454,8 +454,8 @@ private:
melonDS::GPU& GPU; melonDS::GPU& GPU;
RendererPolygon PolygonList[2048]; RendererPolygon PolygonList[2048];
template <bool odd> bool DoTimings(s32 cycles); bool DoTimings(s32 cycles, bool odd);
template <bool odd> void EndScanline(); void EndScanline(bool odd);
void TextureLookup(u32 texparam, u32 texpal, s16 s, s16 t, u16* color, u8* alpha); void TextureLookup(u32 texparam, u32 texpal, s16 s, s16 t, u16* color, u8* alpha);
u32 RenderPixel(Polygon* polygon, u8 vr, u8 vg, u8 vb, s16 s, s16 t); u32 RenderPixel(Polygon* polygon, u8 vr, u8 vg, u8 vb, s16 s, s16 t);
void PlotTranslucentPixel(u32 pixeladdr, u32 color, u32 z, u32 polyattr, u32 shadow); void PlotTranslucentPixel(u32 pixeladdr, u32 color, u32 z, u32 polyattr, u32 shadow);
@ -465,10 +465,10 @@ private:
bool Step(RendererPolygon* rp, bool abortscanline); bool Step(RendererPolygon* rp, bool abortscanline);
void CheckSlope(RendererPolygon* rp, s32 y); void CheckSlope(RendererPolygon* rp, s32 y);
void RenderShadowMaskScanline(RendererPolygon* rp, s32 y); void RenderShadowMaskScanline(RendererPolygon* rp, s32 y);
template <bool odd> bool RenderPolygonScanline(RendererPolygon* rp, s32 y); bool RenderPolygonScanline(RendererPolygon* rp, s32 y, bool odd);
template <bool odd> bool RenderScanline(s32 y, int npolys); bool RenderScanline(s32 y, int npolys, bool odd);
u32 CalculateFogDensity(u32 pixeladdr); u32 CalculateFogDensity(u32 pixeladdr);
template<bool odd, bool finish> void ScanlineFinalPass(s32 y, u8 rdbufferoffset, bool late); void ScanlineFinalPass(s32 y, u8 rdbufferoffset, bool late, bool odd, bool finish);
void ClearBuffers(s32 y); void ClearBuffers(s32 y);
void RenderPolygons(bool threaded, Polygon** polygons, int npolys); void RenderPolygons(bool threaded, Polygon** polygons, int npolys);