Skip to content

S
swiftshader
Commit 3e1fd3a7 by Alexis Hetu Committed by Alexis Hétu
Options

Minor PixelRoutine cleanup

- Removed redundant parameters from fogBlend() - Moved computation registers and color output registers from PixelRoutine to PixelProgram. - Made many PixelRoutine member functions private rather than protected when possible. Change-Id: I748333626d993c0d46b369991d74bc3a22c972c2 Reviewed-on: https://swiftshader-review.googlesource.com/3850Reviewed-by: 's avatarAlexis Hétu <sugoi@google.com> Reviewed-by: 's avatarNicolas Capens <capn@google.com> Tested-by: 's avatarAlexis Hétu <sugoi@google.com>
parent 3ed16305
Showing with 107 additions and 96 deletions
src/Shader/PixelPipeline.cpp
......@@ -21,15 +21,15 @@ namespace sw
{
Registers& r = *static_cast<Registers*>(&rBase);
if(state.color[0].component & 0x1) r.diffuse.x = convertFixed12(r.vf[0].x); else r.diffuse.x = Short4(0x1000);
if(state.color[0].component & 0x2) r.diffuse.y = convertFixed12(r.vf[0].y); else r.diffuse.y = Short4(0x1000);
if(state.color[0].component & 0x4) r.diffuse.z = convertFixed12(r.vf[0].z); else r.diffuse.z = Short4(0x1000);
if(state.color[0].component & 0x8) r.diffuse.w = convertFixed12(r.vf[0].w); else r.diffuse.w = Short4(0x1000);
if(state.color[0].component & 0x1) r.diffuse.x = convertFixed12(r.v[0].x); else r.diffuse.x = Short4(0x1000);
if(state.color[0].component & 0x2) r.diffuse.y = convertFixed12(r.v[0].y); else r.diffuse.y = Short4(0x1000);
if(state.color[0].component & 0x4) r.diffuse.z = convertFixed12(r.v[0].z); else r.diffuse.z = Short4(0x1000);
if(state.color[0].component & 0x8) r.diffuse.w = convertFixed12(r.v[0].w); else r.diffuse.w = Short4(0x1000);
if(state.color[1].component & 0x1) r.specular.x = convertFixed12(r.vf[1].x); else r.specular.x = Short4(0x0000, 0x0000, 0x0000, 0x0000);
if(state.color[1].component & 0x2) r.specular.y = convertFixed12(r.vf[1].y); else r.specular.y = Short4(0x0000, 0x0000, 0x0000, 0x0000);
if(state.color[1].component & 0x4) r.specular.z = convertFixed12(r.vf[1].z); else r.specular.z = Short4(0x0000, 0x0000, 0x0000, 0x0000);
if(state.color[1].component & 0x8) r.specular.w = convertFixed12(r.vf[1].w); else r.specular.w = Short4(0x0000, 0x0000, 0x0000, 0x0000);
if(state.color[1].component & 0x1) r.specular.x = convertFixed12(r.v[1].x); else r.specular.x = Short4(0x0000, 0x0000, 0x0000, 0x0000);
if(state.color[1].component & 0x2) r.specular.y = convertFixed12(r.v[1].y); else r.specular.y = Short4(0x0000, 0x0000, 0x0000, 0x0000);
if(state.color[1].component & 0x4) r.specular.z = convertFixed12(r.v[1].z); else r.specular.z = Short4(0x0000, 0x0000, 0x0000, 0x0000);
if(state.color[1].component & 0x8) r.specular.w = convertFixed12(r.v[1].w); else r.specular.w = Short4(0x0000, 0x0000, 0x0000, 0x0000);
}
void PixelPipeline::fixedFunction(Registers& r)
......@@ -102,10 +102,10 @@ namespace sw
if(src1.type != Shader::PARAMETER_VOID) s1 = fetchRegisterS(r, src1);
if(src2.type != Shader::PARAMETER_VOID) s2 = fetchRegisterS(r, src2);
Float4 u = version < 0x0104 ? r.vf[2 + dst.index].x : r.vf[2 + src0.index].x;
Float4 v = version < 0x0104 ? r.vf[2 + dst.index].y : r.vf[2 + src0.index].y;
Float4 s = version < 0x0104 ? r.vf[2 + dst.index].z : r.vf[2 + src0.index].z;
Float4 t = version < 0x0104 ? r.vf[2 + dst.index].w : r.vf[2 + src0.index].w;
Float4 u = version < 0x0104 ? r.v[2 + dst.index].x : r.v[2 + src0.index].x;
Float4 v = version < 0x0104 ? r.v[2 + dst.index].y : r.v[2 + src0.index].y;
Float4 s = version < 0x0104 ? r.v[2 + dst.index].z : r.v[2 + src0.index].z;
Float4 t = version < 0x0104 ? r.v[2 + dst.index].w : r.v[2 + src0.index].w;
switch(opcode)
{
......@@ -342,7 +342,7 @@ namespace sw
r.current.z &= Short4(0xF800u);
}
fogBlend(r, r.current, fog, r.z[0], r.rhw);
fogBlend(r, r.current, fog);
for(unsigned int q = 0; q < state.multiSample; q++)
{
......@@ -361,7 +361,7 @@ namespace sw
case FORMAT_G32R32F:
case FORMAT_A32B32G32R32F:
convertSigned12(oC, r.current);
PixelRoutine::fogBlend(r, oC, fog, r.z[0], r.rhw);
PixelRoutine::fogBlend(r, oC, fog);
for(unsigned int q = 0; q < state.multiSample; q++)
{
......@@ -1174,7 +1174,7 @@ namespace sw
}
}
void PixelPipeline::fogBlend(Registers &r, Vector4s &current, Float4 &f, Float4 &z, Float4 &rhw)
void PixelPipeline::fogBlend(Registers &r, Vector4s &current, Float4 &f)
{
if(!state.fogActive)
{
......@@ -1183,7 +1183,7 @@ namespace sw
if(state.pixelFogMode != FOG_NONE)
{
pixelFog(r, f, z, rhw);
pixelFog(r, f);
}
UShort4 fog = convertFixed16(f, true);
......@@ -1213,10 +1213,10 @@ namespace sw
void PixelPipeline::sampleTexture(Registers &r, Vector4s &c, int coordinates, int stage, bool project)
{
Float4 u = r.vf[2 + coordinates].x;
Float4 v = r.vf[2 + coordinates].y;
Float4 w = r.vf[2 + coordinates].z;
Float4 q = r.vf[2 + coordinates].w;
Float4 u = r.v[2 + coordinates].x;
Float4 v = r.v[2 + coordinates].y;
Float4 w = r.v[2 + coordinates].z;
Float4 q = r.v[2 + coordinates].w;
if(perturbate)
{
......@@ -1887,9 +1887,9 @@ namespace sw
Float4 E[3]; // Eye vector
E[0] = r.vf[2 + stage - 2].w;
E[1] = r.vf[2 + stage - 1].w;
E[2] = r.vf[2 + stage - 0].w;
E[0] = r.v[2 + stage - 2].w;
E[1] = r.v[2 + stage - 1].w;
E[2] = r.v[2 + stage - 0].w;
// Reflection
Float4 u__;
......
src/Shader/PixelPipeline.hpp
......@@ -59,7 +59,7 @@ namespace sw
void fixedFunction(Registers& r);
void blendTexture(Registers &r, Vector4s &temp, Vector4s &texture, int stage);
void fogBlend(Registers &r, Vector4s &current, Float4 &fog, Float4 &z, Float4 &rhw);
void fogBlend(Registers &r, Vector4s &current, Float4 &fog);
void specularPixel(Vector4s &current, Vector4s &specular);
void sampleTexture(Registers &r, Vector4s &c, int coordinates, int sampler, bool project = false);
......
src/Shader/PixelProgram.cpp
......@@ -121,14 +121,14 @@ namespace sw
{
if(dst.type == Shader::PARAMETER_TEXTURE)
{
d.x = r.vf[2 + dst.index].x;
d.y = r.vf[2 + dst.index].y;
d.z = r.vf[2 + dst.index].z;
d.w = r.vf[2 + dst.index].w;
d.x = r.v[2 + dst.index].x;
d.y = r.v[2 + dst.index].y;
d.z = r.v[2 + dst.index].z;
d.w = r.v[2 + dst.index].w;
}
else
{
d = r.rf[dst.index];
d = r.r[dst.index];
}
}
......@@ -342,19 +342,19 @@ namespace sw
case Shader::PARAMETER_TEMP:
if(dst.rel.type == Shader::PARAMETER_VOID)
{
if(dst.x) pDst.x = r.rf[dst.index].x;
if(dst.y) pDst.y = r.rf[dst.index].y;
if(dst.z) pDst.z = r.rf[dst.index].z;
if(dst.w) pDst.w = r.rf[dst.index].w;
if(dst.x) pDst.x = r.r[dst.index].x;
if(dst.y) pDst.y = r.r[dst.index].y;
if(dst.z) pDst.z = r.r[dst.index].z;
if(dst.w) pDst.w = r.r[dst.index].w;
}
else
{
Int a = relativeAddress(r, dst);
if(dst.x) pDst.x = r.rf[dst.index + a].x;
if(dst.y) pDst.y = r.rf[dst.index + a].y;
if(dst.z) pDst.z = r.rf[dst.index + a].z;
if(dst.w) pDst.w = r.rf[dst.index + a].w;
if(dst.x) pDst.x = r.r[dst.index + a].x;
if(dst.y) pDst.y = r.r[dst.index + a].y;
if(dst.z) pDst.z = r.r[dst.index + a].z;
if(dst.w) pDst.w = r.r[dst.index + a].w;
}
break;
case Shader::PARAMETER_COLOROUT:
......@@ -425,19 +425,19 @@ namespace sw
case Shader::PARAMETER_TEMP:
if(dst.rel.type == Shader::PARAMETER_VOID)
{
if(dst.x) r.rf[dst.index].x = d.x;
if(dst.y) r.rf[dst.index].y = d.y;
if(dst.z) r.rf[dst.index].z = d.z;
if(dst.w) r.rf[dst.index].w = d.w;
if(dst.x) r.r[dst.index].x = d.x;
if(dst.y) r.r[dst.index].y = d.y;
if(dst.z) r.r[dst.index].z = d.z;
if(dst.w) r.r[dst.index].w = d.w;
}
else
{
Int a = relativeAddress(r, dst);
if(dst.x) r.rf[dst.index + a].x = d.x;
if(dst.y) r.rf[dst.index + a].y = d.y;
if(dst.z) r.rf[dst.index + a].z = d.z;
if(dst.w) r.rf[dst.index + a].w = d.w;
if(dst.x) r.r[dst.index + a].x = d.x;
if(dst.y) r.r[dst.index + a].y = d.y;
if(dst.z) r.r[dst.index + a].z = d.z;
if(dst.w) r.r[dst.index + a].w = d.w;
}
break;
case Shader::PARAMETER_COLOROUT:
......@@ -539,7 +539,7 @@ namespace sw
if(index == 0)
{
fogBlend(r, r.oC[index], fog, r.z[0], r.rhw);
fogBlend(r, r.oC[index], fog);
}
switch(state.targetFormat[index])
......@@ -714,32 +714,32 @@ namespace sw
case Shader::PARAMETER_TEMP:
if(src.rel.type == Shader::PARAMETER_VOID)
{
reg = r.rf[i];
reg = r.r[i];
}
else
{
Int a = relativeAddress(r, src);
reg = r.rf[i + a];
reg = r.r[i + a];
}
break;
case Shader::PARAMETER_INPUT:
{
if(src.rel.type == Shader::PARAMETER_VOID) // Not relative
{
reg = r.vf[i];
reg = r.v[i];
}
else if(src.rel.type == Shader::PARAMETER_LOOP)
{
Int aL = r.aL[r.loopDepth];
reg = r.vf[i + aL];
reg = r.v[i + aL];
}
else
{
Int a = relativeAddress(r, src);
reg = r.vf[i + a];
reg = r.v[i + a];
}
}
break;
......@@ -747,7 +747,7 @@ namespace sw
reg = readConstant(r, src, offset);
break;
case Shader::PARAMETER_TEXTURE:
reg = r.vf[2 + i];
reg = r.v[2 + i];
break;
case Shader::PARAMETER_MISCTYPE:
if(src.index == 0) reg = r.vPos;
......@@ -760,7 +760,7 @@ namespace sw
}
else if(src.rel.type == Shader::PARAMETER_TEMP)
{
reg.x = As<Float4>(Int4(i) + As<Int4>(r.rf[src.rel.index].x));
reg.x = As<Float4>(Int4(i) + As<Int4>(r.r[src.rel.index].x));
}
return reg;
case Shader::PARAMETER_PREDICATE: return reg; // Dummy
......@@ -898,11 +898,11 @@ namespace sw
if(var.rel.type == Shader::PARAMETER_TEMP)
{
return As<Int>(Extract(r.rf[var.rel.index].x, 0)) * var.rel.scale;
return As<Int>(Extract(r.r[var.rel.index].x, 0)) * var.rel.scale;
}
else if(var.rel.type == Shader::PARAMETER_INPUT)
{
return As<Int>(Extract(r.vf[var.rel.index].x, 0)) * var.rel.scale;
return As<Int>(Extract(r.v[var.rel.index].x, 0)) * var.rel.scale;
}
else if(var.rel.type == Shader::PARAMETER_OUTPUT)
{
......
src/Shader/PixelProgram.hpp
......@@ -38,7 +38,10 @@ namespace sw
private:
struct Registers : public PixelRoutine::Registers
{
Registers(const PixelShader *shader) : PixelRoutine::Registers(shader), loopDepth(-1)
Registers(const PixelShader *shader) :
PixelRoutine::Registers(shader),
r(shader && shader->dynamicallyIndexedTemporaries),
loopDepth(-1)
{
enableStack[0] = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
......@@ -53,6 +56,12 @@ namespace sw
}
}
// Temporary registers
RegisterArray<4096> r;
// Color outputs
Vector4f oC[4];
// Shader variables
Vector4f vPos;
Vector4f vFace;
......
src/Shader/PixelRoutine.cpp
......@@ -29,17 +29,16 @@ namespace sw
PixelRoutine::Registers::Registers(const PixelShader *shader) :
QuadRasterizer::Registers(),
rf(shader && shader->dynamicallyIndexedTemporaries),
vf(shader && shader->dynamicallyIndexedInput)
v(shader && shader->dynamicallyIndexedInput)
{
if(!shader || shader->getVersion() < 0x0200 || forceClearRegisters)
{
for(int i = 0; i < 10; i++)
{
vf[i].x = Float4(0.0f);
vf[i].y = Float4(0.0f);
vf[i].z = Float4(0.0f);
vf[i].w = Float4(0.0f);
v[i].x = Float4(0.0f);
v[i].y = Float4(0.0f);
v[i].z = Float4(0.0f);
v[i].w = Float4(0.0f);
}
}
}
......@@ -169,11 +168,11 @@ namespace sw
{
if(!state.interpolant[interpolant].centroid)
{
r.vf[interpolant][component] = interpolate(xxxx, r.Dv[interpolant][component], rhw, r.primitive + OFFSET(Primitive, V[interpolant][component]), (state.interpolant[interpolant].flat & (1 << component)) != 0, state.perspective);
r.v[interpolant][component] = interpolate(xxxx, r.Dv[interpolant][component], rhw, r.primitive + OFFSET(Primitive, V[interpolant][component]), (state.interpolant[interpolant].flat & (1 << component)) != 0, state.perspective);
}
else
{
r.vf[interpolant][component] = interpolateCentroid(XXXX, YYYY, rhwCentroid, r.primitive + OFFSET(Primitive, V[interpolant][component]), (state.interpolant[interpolant].flat & (1 << component)) != 0, state.perspective);
r.v[interpolant][component] = interpolateCentroid(XXXX, YYYY, rhwCentroid, r.primitive + OFFSET(Primitive, V[interpolant][component]), (state.interpolant[interpolant].flat & (1 << component)) != 0, state.perspective);
}
}
}
......@@ -185,19 +184,19 @@ namespace sw
case 0:
break;
case 1:
rcp = reciprocal(r.vf[interpolant].y);
r.vf[interpolant].x = r.vf[interpolant].x * rcp;
rcp = reciprocal(r.v[interpolant].y);
r.v[interpolant].x = r.v[interpolant].x * rcp;
break;
case 2:
rcp = reciprocal(r.vf[interpolant].z);
r.vf[interpolant].x = r.vf[interpolant].x * rcp;
r.vf[interpolant].y = r.vf[interpolant].y * rcp;
rcp = reciprocal(r.v[interpolant].z);
r.v[interpolant].x = r.v[interpolant].x * rcp;
r.v[interpolant].y = r.v[interpolant].y * rcp;
break;
case 3:
rcp = reciprocal(r.vf[interpolant].w);
r.vf[interpolant].x = r.vf[interpolant].x * rcp;
r.vf[interpolant].y = r.vf[interpolant].y * rcp;
r.vf[interpolant].z = r.vf[interpolant].z * rcp;
rcp = reciprocal(r.v[interpolant].w);
r.v[interpolant].x = r.v[interpolant].x * rcp;
r.v[interpolant].y = r.v[interpolant].y * rcp;
r.v[interpolant].z = r.v[interpolant].z * rcp;
break;
}
}
......@@ -602,7 +601,7 @@ namespace sw
cMask[3] &= aMask3;
}
void PixelRoutine::fogBlend(Registers &r, Vector4f &c0, Float4 &fog, Float4 &z, Float4 &rhw)
void PixelRoutine::fogBlend(Registers &r, Vector4f &c0, Float4 &fog)
{
if(!state.fogActive)
{
......@@ -611,7 +610,7 @@ namespace sw
if(state.pixelFogMode != FOG_NONE)
{
pixelFog(r, fog, z, rhw);
pixelFog(r, fog);
fog = Min(fog, Float4(1.0f));
fog = Max(fog, Float4(0.0f));
......@@ -630,7 +629,7 @@ namespace sw
c0.z += *Pointer<Float4>(r.data + OFFSET(DrawData,fog.colorF[2]));
}
void PixelRoutine::pixelFog(Registers &r, Float4 &visibility, Float4 &z, Float4 &rhw)
void PixelRoutine::pixelFog(Registers &r, Float4 &visibility)
{
Float4 &zw = visibility;
......@@ -638,17 +637,17 @@ namespace sw
{
if(state.wBasedFog)
{
zw = rhw;
zw = r.rhw;
}
else
{
if(complementaryDepthBuffer)
{
zw = Float4(1.0f) - z;
zw = Float4(1.0f) - r.z[0];
}
else
{
zw = z;
zw = r.z[0];
}
}
}
......
src/Shader/PixelRoutine.hpp
......@@ -37,11 +37,9 @@ namespace sw
Float4 w; // Used as is
Float4 rhw; // Reciprocal w
RegisterArray<4096> rf; // Computation registers
RegisterArray<10> vf; // Varying registers
RegisterArray<10> v; // Varying registers
// Outputs
Vector4f oC[4];
// Depth output
Float4 oDepth;
};
......@@ -56,41 +54,46 @@ namespace sw
virtual void quad(QuadRasterizer::Registers &r, Pointer<Byte> cBuffer[4], Pointer<Byte> &zBuffer, Pointer<Byte> &sBuffer, Int cMask[4], Int &x, Int &y);
void alphaTest(Registers &r, Int &aMask, Short4 &alpha);
void alphaToCoverage(Registers &r, Int cMask[4], Float4 &alpha);
void fogBlend(Registers &r, Vector4f &c0, Float4 &fog);
void pixelFog(Registers &r, Float4 &visibility);
// Raster operations
void alphaBlend(Registers &r, int index, Pointer<Byte> &cBuffer, Vector4s &current, Int &x);
void logicOperation(Registers &r, int index, Pointer<Byte> &cBuffer, Vector4s &current, Int &x);
void writeColor(Registers &r, int index, Pointer<Byte> &cBuffer, Int &i, Vector4s &current, Int &sMask, Int &zMask, Int &cMask);
void alphaBlend(Registers &r, int index, Pointer<Byte> &cBuffer, Vector4f &oC, Int &x);
void writeColor(Registers &r, int index, Pointer<Byte> &cBuffer, Int &i, Vector4f &oC, Int &sMask, Int &zMask, Int &cMask);
UShort4 convertFixed16(Float4 &cf, bool saturate = true);
void linearToSRGB12_16(Registers &r, Vector4s &c);
SamplerCore *sampler[TEXTURE_IMAGE_UNITS];
private:
Float4 interpolateCentroid(Float4 &x, Float4 &y, Float4 &rhw, Pointer<Byte> planeEquation, bool flat, bool perspective);
void stencilTest(Registers &r, Pointer<Byte> &sBuffer, int q, Int &x, Int &sMask, Int &cMask);
void stencilTest(Registers &r, Byte8 &value, StencilCompareMode stencilCompareMode, bool CCW);
void stencilOperation(Registers &r, Byte8 &newValue, Byte8 &bufferValue, StencilOperation stencilPassOperation, StencilOperation stencilZFailOperation, StencilOperation stencilFailOperation, bool CCW, Int &zMask, Int &sMask);
void stencilOperation(Registers &r, Byte8 &output, Byte8 &bufferValue, StencilOperation operation, bool CCW);
Bool depthTest(Registers &r, Pointer<Byte> &zBuffer, int q, Int &x, Float4 &z, Int &sMask, Int &zMask, Int &cMask);
void alphaTest(Registers &r, Int &aMask, Short4 &alpha);
void alphaToCoverage(Registers &r, Int cMask[4], Float4 &alpha);
void fogBlend(Registers &r, Vector4f &c0, Float4 &fog, Float4 &z, Float4 &rhw);
void pixelFog(Registers &r, Float4 &visibility, Float4 &z, Float4 &rhw);
// Raster operations
void blendFactor(Registers &r, const Vector4s &blendFactor, const Vector4s &current, const Vector4s &pixel, BlendFactor blendFactorActive);
void blendFactorAlpha(Registers &r, const Vector4s &blendFactor, const Vector4s &current, const Vector4s &pixel, BlendFactor blendFactorAlphaActive);
void readPixel(Registers &r, int index, Pointer<Byte> &cBuffer, Int &x, Vector4s &pixel);
void alphaBlend(Registers &r, int index, Pointer<Byte> &cBuffer, Vector4s &current, Int &x);
void logicOperation(Registers &r, int index, Pointer<Byte> &cBuffer, Vector4s &current, Int &x);
void writeColor(Registers &r, int index, Pointer<Byte> &cBuffer, Int &i, Vector4s &current, Int &sMask, Int &zMask, Int &cMask);
void blendFactor(Registers &r, const Vector4f &blendFactor, const Vector4f &oC, const Vector4f &pixel, BlendFactor blendFactorActive);
void blendFactorAlpha(Registers &r, const Vector4f &blendFactor, const Vector4f &oC, const Vector4f &pixel, BlendFactor blendFactorAlphaActive);
void alphaBlend(Registers &r, int index, Pointer<Byte> &cBuffer, Vector4f &oC, Int &x);
void writeColor(Registers &r, int index, Pointer<Byte> &cBuffer, Int &i, Vector4f &oC, Int &sMask, Int &zMask, Int &cMask);
void writeStencil(Registers &r, Pointer<Byte> &sBuffer, int q, Int &x, Int &sMask, Int &zMask, Int &cMask);
void writeDepth(Registers &r, Pointer<Byte> &zBuffer, int q, Int &x, Float4 &z, Int &zMask);
UShort4 convertFixed16(Float4 &cf, bool saturate = true);
void sRGBtoLinear16_12_16(Registers &r, Vector4s &c);
void sRGBtoLinear12_16(Registers &r, Vector4s &c);
void linearToSRGB16_12_16(Registers &r, Vector4s &c);
void linearToSRGB12_16(Registers &r, Vector4s &c);
Float4 sRGBtoLinear(const Float4 &x);
bool colorUsed();
SamplerCore *sampler[TEXTURE_IMAGE_UNITS];
};
}
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment