From ed79b1772c5cfff95346b69f8b2ba1ac355b1637 Mon Sep 17 00:00:00 2001
From: Jaklyy <102590697+Jaklyy@users.noreply.github.com>
Date: Tue, 17 Oct 2023 18:53:50 -0400
Subject: [PATCH 1/4] invert coverage to fix xmajor aa
if the calculated y coordinate differs from the expected y coordinate, invert its coverage.
---
src/GPU3D_Soft.cpp | 12 ++----------
src/GPU3D_Soft.h | 12 ++++++++++--
2 files changed, 12 insertions(+), 12 deletions(-)
diff --git a/src/GPU3D_Soft.cpp b/src/GPU3D_Soft.cpp
index 03c6265e..c131dc0f 100644
--- a/src/GPU3D_Soft.cpp
+++ b/src/GPU3D_Soft.cpp
@@ -1070,11 +1070,7 @@ void SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
xlimit = xstart+l_edgelen;
if (xlimit > xend+1) xlimit = xend+1;
if (xlimit > 256) xlimit = 256;
- if (l_edgecov & (1<<31))
- {
- xcov = (l_edgecov >> 12) & 0x3FF;
- if (xcov == 0x3FF) xcov = 0;
- }
+ if (l_edgecov & (1<<31)) xcov = (l_edgecov >> 12) & 0x3FF;
if (!l_filledge) x = xlimit;
else
@@ -1259,11 +1255,7 @@ void SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
edge = yedge | 0x2;
xlimit = xend+1;
if (xlimit > 256) xlimit = 256;
- if (r_edgecov & (1<<31))
- {
- xcov = (r_edgecov >> 12) & 0x3FF;
- if (xcov == 0x3FF) xcov = 0;
- }
+ if (r_edgecov & (1<<31)) xcov = (r_edgecov >> 12) & 0x3FF;
if (r_filledge)
for (; x < xlimit; x++)
diff --git a/src/GPU3D_Soft.h b/src/GPU3D_Soft.h
index 2f5664e2..ec89cac5 100644
--- a/src/GPU3D_Soft.h
+++ b/src/GPU3D_Soft.h
@@ -236,6 +236,7 @@ private:
dx = 0;
this->x0 = x0;
+ this->y0 = y0;
this->xmin = x0;
this->xmax = x0;
@@ -253,6 +254,7 @@ private:
constexpr s32 Setup(s32 x0, s32 x1, s32 y0, s32 y1, s32 w0, s32 w1, s32 y)
{
this->x0 = x0;
+ this->y0 = y0;
this->y = y;
if (x1 > x0)
@@ -367,7 +369,13 @@ private:
if (side) startx = startx - *length + 1;
s32 startcov = (((startx << 10) + 0x1FF) * ylen) / xlen;
- *coverage = (1<<31) | ((startcov & 0x3FF) << 12) | (xcov_incr & 0x3FF);
+
+ // fix the y value for negative slopes
+ s32 ycoord = Negative ? (ylen << 10) - startcov >> 10 : startcov >> 10;
+ // if yvalue is not equal to actual y value, invert coverage value
+ if (ycoord != y - y0) startcov = 0x3FF - (startcov & 0x3FF);
+
+ *coverage = (1<<31) | (startcov << 12) | (xcov_incr & 0x3FF);
if constexpr (swapped) *length = 1;
}
@@ -419,7 +427,7 @@ private:
Interpolator<1> Interp;
private:
- s32 x0, xmin, xmax;
+ s32 x0, y0, xmin, xmax;
s32 xlen, ylen;
s32 dx;
s32 y;
From 0c8095c1bb0e43ce4db884500b372aeb853d0ca4 Mon Sep 17 00:00:00 2001
From: Jaklyy <102590697+Jaklyy@users.noreply.github.com>
Date: Wed, 18 Oct 2023 00:36:53 -0400
Subject: [PATCH 2/4] rewrite of aa code
should behave identically
---
src/GPU3D_Soft.cpp | 24 ++++++++----------------
src/GPU3D_Soft.h | 31 +++++++++++--------------------
2 files changed, 19 insertions(+), 36 deletions(-)
diff --git a/src/GPU3D_Soft.cpp b/src/GPU3D_Soft.cpp
index c131dc0f..1871caeb 100644
--- a/src/GPU3D_Soft.cpp
+++ b/src/GPU3D_Soft.cpp
@@ -1063,14 +1063,11 @@ void SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
if (x < 0) x = 0;
s32 xlimit;
- s32 xcov = 0;
-
// part 1: left edge
edge = yedge | 0x1;
xlimit = xstart+l_edgelen;
if (xlimit > xend+1) xlimit = xend+1;
if (xlimit > 256) xlimit = 256;
- if (l_edgecov & (1<<31)) xcov = (l_edgecov >> 12) & 0x3FF;
if (!l_filledge) x = xlimit;
else
@@ -1129,14 +1126,12 @@ void SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
// anti-aliasing: all edges are rendered
// calculate coverage
- s32 cov = l_edgecov;
- if (cov & (1<<31))
+ attr |= (l_edgecov >> 5) << 8;
+ if (x < xlimit-1)
{
- cov = xcov >> 5;
- if (cov > 31) cov = 31;
- xcov += (l_edgecov & 0x3FF);
+ l_edgecov += rp->SlopeL.XCov_Incr;
+ if (l_edgecov > 0x3FF) l_edgecov = 0x3FF;
}
- attr |= (cov << 8);
// push old pixel down if needed
if (pixeladdr < BufferSize)
@@ -1255,7 +1250,6 @@ void SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
edge = yedge | 0x2;
xlimit = xend+1;
if (xlimit > 256) xlimit = 256;
- if (r_edgecov & (1<<31)) xcov = (r_edgecov >> 12) & 0x3FF;
if (r_filledge)
for (; x < xlimit; x++)
@@ -1313,14 +1307,12 @@ void SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
// anti-aliasing: all edges are rendered
// calculate coverage
- s32 cov = r_edgecov;
- if (cov & (1<<31))
+ attr |= (r_edgecov >> 5) << 8;
+ if (x < xlimit-1)
{
- cov = 0x1F - (xcov >> 5);
- if (cov < 0) cov = 0;
- xcov += (r_edgecov & 0x3FF);
+ r_edgecov -= rp->SlopeR.XCov_Incr;
+ if (r_edgecov < 0) r_edgecov = 0;
}
- attr |= (cov << 8);
// push old pixel down if needed
if (pixeladdr < BufferSize)
diff --git a/src/GPU3D_Soft.h b/src/GPU3D_Soft.h
index ec89cac5..38de794c 100644
--- a/src/GPU3D_Soft.h
+++ b/src/GPU3D_Soft.h
@@ -246,7 +246,7 @@ private:
Interp.Setup(0, 0, 0, 0);
Interp.SetX(0);
- xcov_incr = 0;
+ XCov_Incr = 0;
return x0;
}
@@ -322,7 +322,7 @@ private:
Interp.SetX(y);
// used for calculating AA coverage
- if (XMajor) xcov_incr = (ylen << 10) / xlen;
+ if (XMajor) XCov_Incr = ((ylen << 10) / xlen) & 0x3FF;
return x;
}
@@ -373,9 +373,11 @@ private:
// fix the y value for negative slopes
s32 ycoord = Negative ? (ylen << 10) - startcov >> 10 : startcov >> 10;
// if yvalue is not equal to actual y value, invert coverage value
- if (ycoord != y - y0) startcov = 0x3FF - (startcov & 0x3FF);
+ startcov &= 0x3FF;
+ if (ycoord != y - y0) startcov = 0x3FF - startcov;
+ if (side ^ swapped) startcov = 0x3FF - startcov;
- *coverage = (1<<31) | (startcov << 12) | (xcov_incr & 0x3FF);
+ *coverage = startcov;
if constexpr (swapped) *length = 1;
}
@@ -389,26 +391,17 @@ private:
{
// for some reason vertical edges' aa values
// are inverted too when the edges are swapped
- if constexpr (swapped)
- *coverage = 0;
- else
- *coverage = 31;
+ *coverage = swapped ? 0 : 31 << 5;
}
else
{
s32 cov = ((dx >> 9) + (Increment >> 10)) >> 4;
if ((cov >> 5) != (dx >> 18)) cov = 31;
cov &= 0x1F;
- if constexpr (swapped)
- {
- if (side ^ Negative) cov = 0x1F - cov;
- }
- else
- {
- if (!(side ^ Negative)) cov = 0x1F - cov;
- }
+ if (side ^ !Negative ^ swapped) cov = 0x1F - cov;
- *coverage = cov;
+ // shift left 5 just to make it align with xmajor coverage values
+ *coverage = cov << 5;
}
}
@@ -425,15 +418,13 @@ private:
bool Negative;
bool XMajor;
Interpolator<1> Interp;
+ s32 XCov_Incr;
private:
s32 x0, y0, xmin, xmax;
s32 xlen, ylen;
s32 dx;
s32 y;
-
- s32 xcov_incr;
- s32 ycoverage, ycov_incr;
};
template <typename T>
From 3facbe90c981a77d1a4f5d618bdba47d7008adab Mon Sep 17 00:00:00 2001
From: Jaklyy <102590697+Jaklyy@users.noreply.github.com>
Date: Sun, 10 Dec 2023 15:40:58 -0500
Subject: [PATCH 3/4] remove templates
seems to just make things sliiightly slower on my machine
---
src/GPU3D_Soft.cpp | 18 +++++++++---------
src/GPU3D_Soft.h | 21 +++++++++------------
2 files changed, 18 insertions(+), 21 deletions(-)
diff --git a/src/GPU3D_Soft.cpp b/src/GPU3D_Soft.cpp
index 1871caeb..1166b770 100644
--- a/src/GPU3D_Soft.cpp
+++ b/src/GPU3D_Soft.cpp
@@ -741,8 +741,8 @@ void SoftRenderer::RenderShadowMaskScanline(RendererPolygon* rp, s32 y)
interp_start = &rp->SlopeR.Interp;
interp_end = &rp->SlopeL.Interp;
- rp->SlopeR.EdgeParams<true>(&l_edgelen, &l_edgecov);
- rp->SlopeL.EdgeParams<true>(&r_edgelen, &r_edgecov);
+ rp->SlopeR.EdgeParams(&l_edgelen, &l_edgecov, true);
+ rp->SlopeL.EdgeParams(&r_edgelen, &r_edgecov, true);
std::swap(xstart, xend);
std::swap(wl, wr);
@@ -773,9 +773,9 @@ void SoftRenderer::RenderShadowMaskScanline(RendererPolygon* rp, s32 y)
interp_start = &rp->SlopeL.Interp;
interp_end = &rp->SlopeR.Interp;
- rp->SlopeL.EdgeParams<false>(&l_edgelen, &l_edgecov);
- rp->SlopeR.EdgeParams<false>(&r_edgelen, &r_edgecov);
-
+ rp->SlopeL.EdgeParams(&l_edgelen, &l_edgecov, false);
+ rp->SlopeR.EdgeParams(&r_edgelen, &r_edgecov, false);
+
// CHECKME: edge fill rules for unswapped opaque shadow mask polygons
if ((GPU.GPU3D.RenderDispCnt & ((1<<4)|(1<<5))) || ((polyalpha < 31) && (GPU.GPU3D.RenderDispCnt & (1<<3))) || wireframe)
{
@@ -970,8 +970,8 @@ void SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
interp_start = &rp->SlopeR.Interp;
interp_end = &rp->SlopeL.Interp;
- rp->SlopeR.EdgeParams<true>(&l_edgelen, &l_edgecov);
- rp->SlopeL.EdgeParams<true>(&r_edgelen, &r_edgecov);
+ rp->SlopeR.EdgeParams(&l_edgelen, &l_edgecov, true);
+ rp->SlopeL.EdgeParams(&r_edgelen, &r_edgecov, true);
std::swap(xstart, xend);
std::swap(wl, wr);
@@ -1008,8 +1008,8 @@ void SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y)
interp_start = &rp->SlopeL.Interp;
interp_end = &rp->SlopeR.Interp;
- rp->SlopeL.EdgeParams<false>(&l_edgelen, &l_edgecov);
- rp->SlopeR.EdgeParams<false>(&r_edgelen, &r_edgecov);
+ rp->SlopeL.EdgeParams(&l_edgelen, &l_edgecov, false);
+ rp->SlopeR.EdgeParams(&r_edgelen, &r_edgecov, false);
// edge fill rules for unswapped opaque edges:
// * right edge is filled if slope > 1
diff --git a/src/GPU3D_Soft.h b/src/GPU3D_Soft.h
index 38de794c..9b284932 100644
--- a/src/GPU3D_Soft.h
+++ b/src/GPU3D_Soft.h
@@ -347,13 +347,12 @@ private:
else if (ret > xmax) ret = xmax;
return ret;
}
-
- template<bool swapped>
- constexpr void EdgeParams_XMajor(s32* length, s32* coverage) const
+
+ constexpr void EdgeParams_XMajor(s32* length, s32* coverage, bool swapped) const
{
// only do length calc for right side when swapped as it's
// only needed for aa calcs, as actual line spans are broken
- if constexpr (!swapped || side)
+ if (!swapped || side)
{
if (side ^ Negative)
*length = (dx >> 18) - ((dx-Increment) >> 18);
@@ -379,11 +378,10 @@ private:
*coverage = startcov;
- if constexpr (swapped) *length = 1;
+ if (swapped) *length = 1;
}
- template<bool swapped>
- constexpr void EdgeParams_YMajor(s32* length, s32* coverage) const
+ constexpr void EdgeParams_YMajor(s32* length, s32* coverage, bool swapped) const
{
*length = 1;
@@ -404,14 +402,13 @@ private:
*coverage = cov << 5;
}
}
-
- template<bool swapped>
- constexpr void EdgeParams(s32* length, s32* coverage) const
+
+ constexpr void EdgeParams(s32* length, s32* coverage, bool swapped) const
{
if (XMajor)
- return EdgeParams_XMajor<swapped>(length, coverage);
+ return EdgeParams_XMajor(length, coverage, swapped);
else
- return EdgeParams_YMajor<swapped>(length, coverage);
+ return EdgeParams_YMajor(length, coverage, swapped);
}
s32 Increment;
From 3268f94f4bd9c8d64b9afb71d1f4a611ef5cc70c Mon Sep 17 00:00:00 2001
From: Jaklyy <102590697+Jaklyy@users.noreply.github.com>
Date: Sun, 10 Dec 2023 15:52:28 -0500
Subject: [PATCH 4/4] remove an optimization for swapped polygons
why did i optimize swapped polygons?
they're not that common.
---
src/GPU3D_Soft.h | 11 ++++-------
1 file changed, 4 insertions(+), 7 deletions(-)
diff --git a/src/GPU3D_Soft.h b/src/GPU3D_Soft.h
index 9b284932..b1ea641c 100644
--- a/src/GPU3D_Soft.h
+++ b/src/GPU3D_Soft.h
@@ -352,13 +352,10 @@ private:
{
// only do length calc for right side when swapped as it's
// only needed for aa calcs, as actual line spans are broken
- if (!swapped || side)
- {
- if (side ^ Negative)
- *length = (dx >> 18) - ((dx-Increment) >> 18);
- else
- *length = ((dx+Increment) >> 18) - (dx >> 18);
- }
+ if (side ^ Negative)
+ *length = (dx >> 18) - ((dx-Increment) >> 18);
+ else
+ *length = ((dx+Increment) >> 18) - (dx >> 18);
// for X-major edges, we return the coverage
// for the first pixel, and the increment for