Skip to content

S
swiftshader
Commit b4fb3678 by Nicolas Capens
Options

Eliminate redundant register set argument passing.

Bug 22652760 Change-Id: If6bf124c3218847ecc4af0ae16102452a6b344d9 Reviewed-on: https://swiftshader-review.googlesource.com/4558Tested-by: 's avatarNicolas Capens <capn@google.com> Reviewed-by: 's avatarAlexis Hétu <sugoi@google.com> Reviewed-by: 's avatarNicolas Capens <capn@google.com>
parent 2ca19030
Showing with 87 additions and 87 deletions
src/Shader/VertexPipeline.cpp
...@@ -34,7 +34,7 @@ namespace sw ...@@ -34,7 +34,7 @@ namespace sw
{ {
} }
Vector4f VertexPipeline::transformBlend(Registers &r, const Register &src, const Pointer<Byte> &matrix, bool homogeneous) Vector4f VertexPipeline::transformBlend(const Register &src, const Pointer<Byte> &matrix, bool homogeneous)
{ {
Vector4f dst; Vector4f dst;
...@@ -55,7 +55,7 @@ namespace sw ...@@ -55,7 +55,7 @@ namespace sw
{ {
Float4 B = r.v[BlendIndices].x; Float4 B = r.v[BlendIndices].x;
UInt indices; UInt indices;
switch(i) switch(i)
{ {
case 0: indices = As<UInt>(Float(B.x)); break; case 0: indices = As<UInt>(Float(B.x)); break;
...@@ -155,14 +155,14 @@ namespace sw ...@@ -155,14 +155,14 @@ namespace sw
return dst; return dst;
} }
void VertexPipeline::pipeline(Registers &r) void VertexPipeline::pipeline()
{ {
Vector4f position; Vector4f position;
Vector4f normal; Vector4f normal;
if(!state.preTransformed) if(!state.preTransformed)
{ {
position = transformBlend(r, r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.transformT)), true); position = transformBlend(r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.transformT)), true);
} }
else else
{ {
...@@ -174,11 +174,11 @@ namespace sw ...@@ -174,11 +174,11 @@ namespace sw
r.o[Pos].z = position.z; r.o[Pos].z = position.z;
r.o[Pos].w = position.w; r.o[Pos].w = position.w;
Vector4f vertexPosition = transformBlend(r, r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.cameraTransformT)), true); Vector4f vertexPosition = transformBlend(r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.cameraTransformT)), true);
if(state.vertexNormalActive) if(state.vertexNormalActive)
{ {
normal = transformBlend(r, r.v[Normal], Pointer<Byte>(r.data + OFFSET(DrawData,ff.normalTransformT)), false); normal = transformBlend(r.v[Normal], Pointer<Byte>(r.data + OFFSET(DrawData,ff.normalTransformT)), false);
if(state.normalizeNormals) if(state.normalizeNormals)
{ {
...@@ -532,13 +532,13 @@ namespace sw ...@@ -532,13 +532,13 @@ namespace sw
for(int stage = 0; stage < 8; stage++) for(int stage = 0; stage < 8; stage++)
{ {
processTextureCoordinate(r, stage, normal, position); processTextureCoordinate(stage, normal, position);
} }
processPointSize(r); processPointSize();
} }
void VertexPipeline::processTextureCoordinate(Registers &r, int stage, Vector4f &normal, Vector4f &position) void VertexPipeline::processTextureCoordinate(int stage, Vector4f &normal, Vector4f &position)
{ {
if(state.output[T0 + stage].write) if(state.output[T0 + stage].write)
{ {
...@@ -605,7 +605,7 @@ namespace sw ...@@ -605,7 +605,7 @@ namespace sw
} }
Nc.w = Float4(1.0f); Nc.w = Float4(1.0f);
r.o[T0 + stage].x = Nc.x; r.o[T0 + stage].x = Nc.x;
r.o[T0 + stage].y = Nc.y; r.o[T0 + stage].y = Nc.y;
r.o[T0 + stage].z = Nc.z; r.o[T0 + stage].z = Nc.z;
...@@ -614,10 +614,10 @@ namespace sw ...@@ -614,10 +614,10 @@ namespace sw
break; break;
case TEXGEN_POSITION: case TEXGEN_POSITION:
{ {
Vector4f Pn = transformBlend(r, r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.cameraTransformT)), true); // Position in camera space Vector4f Pn = transformBlend(r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.cameraTransformT)), true); // Position in camera space
Pn.w = Float4(1.0f); Pn.w = Float4(1.0f);
r.o[T0 + stage].x = Pn.x; r.o[T0 + stage].x = Pn.x;
r.o[T0 + stage].y = Pn.y; r.o[T0 + stage].y = Pn.y;
r.o[T0 + stage].z = Pn.z; r.o[T0 + stage].z = Pn.z;
...@@ -639,7 +639,7 @@ namespace sw ...@@ -639,7 +639,7 @@ namespace sw
Vector4f Ec; // Eye vector in camera space Vector4f Ec; // Eye vector in camera space
Vector4f N2; Vector4f N2;
Ec = transformBlend(r, r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.cameraTransformT)), true); Ec = transformBlend(r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.cameraTransformT)), true);
Ec = normalize(Ec); Ec = normalize(Ec);
// R = E - 2 * N * (E . N) // R = E - 2 * N * (E . N)
...@@ -690,12 +690,12 @@ namespace sw ...@@ -690,12 +690,12 @@ namespace sw
Vector4f Ec; // Eye vector in camera space Vector4f Ec; // Eye vector in camera space
Vector4f N2; Vector4f N2;
Ec = transformBlend(r, r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.cameraTransformT)), true); Ec = transformBlend(r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.cameraTransformT)), true);
Ec = normalize(Ec); Ec = normalize(Ec);
// R = E - 2 * N * (E . N) // R = E - 2 * N * (E . N)
Float4 dot = Float4(2.0f) * dot3(Ec, Nc); Float4 dot = Float4(2.0f) * dot3(Ec, Nc);
R.x = Ec.x - Nc.x * dot; R.x = Ec.x - Nc.x * dot;
R.y = Ec.y - Nc.y * dot; R.y = Ec.y - Nc.y * dot;
R.z = Ec.z - Nc.z * dot; R.z = Ec.z - Nc.z * dot;
...@@ -792,7 +792,7 @@ namespace sw ...@@ -792,7 +792,7 @@ namespace sw
} }
} }
void VertexPipeline::processPointSize(Registers &r) void VertexPipeline::processPointSize()
{ {
if(!state.pointSizeActive) if(!state.pointSizeActive)
{ {
...@@ -810,7 +810,7 @@ namespace sw ...@@ -810,7 +810,7 @@ namespace sw
if(state.pointScaleActive && !state.preTransformed) if(state.pointScaleActive && !state.preTransformed)
{ {
Vector4f p = transformBlend(r, r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.cameraTransformT)), true); Vector4f p = transformBlend(r.v[Position], Pointer<Byte>(r.data + OFFSET(DrawData,ff.cameraTransformT)), true);
Float4 d = Sqrt(dot3(p, p)); // FIXME: length(p); Float4 d = Sqrt(dot3(p, p)); // FIXME: length(p);
...@@ -923,7 +923,7 @@ namespace sw ...@@ -923,7 +923,7 @@ namespace sw
Vector4f dst; Vector4f dst;
Float4 rcpLength = RcpSqrt_pp(dot3(src, src)); Float4 rcpLength = RcpSqrt_pp(dot3(src, src));
dst.x = src.x * rcpLength; dst.x = src.x * rcpLength;
dst.y = src.y * rcpLength; dst.y = src.y * rcpLength;
dst.z = src.z * rcpLength; dst.z = src.z * rcpLength;
...@@ -934,11 +934,11 @@ namespace sw ...@@ -934,11 +934,11 @@ namespace sw
Float4 VertexPipeline::power(Float4 &src0, Float4 &src1) Float4 VertexPipeline::power(Float4 &src0, Float4 &src1)
{ {
Float4 dst = src0; Float4 dst = src0;
dst = dst * dst; dst = dst * dst;
dst = dst * dst; dst = dst * dst;
dst = Float4(As<Int4>(dst) - As<Int4>(Float4(1.0f))); dst = Float4(As<Int4>(dst) - As<Int4>(Float4(1.0f)));
dst *= src1; dst *= src1;
dst = As<Float4>(Int4(dst) + As<Int4>(Float4(1.0f))); dst = As<Float4>(Int4(dst) + As<Int4>(Float4(1.0f)));
......
src/Shader/VertexPipeline.hpp
...@@ -27,11 +27,11 @@ namespace sw ...@@ -27,11 +27,11 @@ namespace sw
virtual ~VertexPipeline(); virtual ~VertexPipeline();
private: private:
void pipeline(Registers &r); void pipeline() override;
void processTextureCoordinate(Registers &r, int stage, Vector4f &normal, Vector4f &position); void processTextureCoordinate(int stage, Vector4f &normal, Vector4f &position);
void processPointSize(Registers &r); void processPointSize();
Vector4f transformBlend(Registers &r, const Register &src, const Pointer<Byte> &matrix, bool homogenous); Vector4f transformBlend(const Register &src, const Pointer<Byte> &matrix, bool homogenous);
Vector4f transform(const Register &src, const Pointer<Byte> &matrix, bool homogenous); Vector4f transform(const Register &src, const Pointer<Byte> &matrix, bool homogenous);
Vector4f transform(const Register &src, const Pointer<Byte> &matrix, UInt index[4], bool homogenous); Vector4f transform(const Register &src, const Pointer<Byte> &matrix, UInt index[4], bool homogenous);
Vector4f normalize(Vector4f &src); Vector4f normalize(Vector4f &src);
......
src/Shader/VertexProgram.hpp
...@@ -37,57 +37,57 @@ namespace sw ...@@ -37,57 +37,57 @@ namespace sw
typedef Shader::Control Control; typedef Shader::Control Control;
typedef Shader::Usage Usage; typedef Shader::Usage Usage;
void pipeline(Registers &r); void pipeline() override;
void program(Registers &r); void program();
void passThrough(Registers &r); void passThrough();
Vector4f fetchRegisterF(Registers &r, const Src &src, unsigned int offset = 0); Vector4f fetchRegisterF(const Src &src, unsigned int offset = 0);
Vector4f readConstant(Registers &r, const Src &src, unsigned int offset = 0); Vector4f readConstant(const Src &src, unsigned int offset = 0);
Int relativeAddress(Registers &r, const Shader::Parameter &var); Int relativeAddress(const Shader::Parameter &var);
Int4 enableMask(Registers &r, const Shader::Instruction *instruction); Int4 enableMask(const Shader::Instruction *instruction);
void M3X2(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1); void M3X2(Vector4f &dst, Vector4f &src0, Src &src1);
void M3X3(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1); void M3X3(Vector4f &dst, Vector4f &src0, Src &src1);
void M3X4(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1); void M3X4(Vector4f &dst, Vector4f &src0, Src &src1);
void M4X3(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1); void M4X3(Vector4f &dst, Vector4f &src0, Src &src1);
void M4X4(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1); void M4X4(Vector4f &dst, Vector4f &src0, Src &src1);
void BREAK(Registers &r); void BREAK();
void BREAKC(Registers &r, Vector4f &src0, Vector4f &src1, Control); void BREAKC(Vector4f &src0, Vector4f &src1, Control);
void BREAKP(Registers &r, const Src &predicateRegister); void BREAKP(const Src &predicateRegister);
void BREAK(Registers &r, Int4 &condition); void BREAK(Int4 &condition);
void CONTINUE(Registers &r); void CONTINUE();
void TEST(); void TEST();
void CALL(Registers &r, int labelIndex, int callSiteIndex); void CALL(int labelIndex, int callSiteIndex);
void CALLNZ(Registers &r, int labelIndex, int callSiteIndex, const Src &src); void CALLNZ(int labelIndex, int callSiteIndex, const Src &src);
void CALLNZb(Registers &r, int labelIndex, int callSiteIndex, const Src &boolRegister); void CALLNZb(int labelIndex, int callSiteIndex, const Src &boolRegister);
void CALLNZp(Registers &r, int labelIndex, int callSiteIndex, const Src &predicateRegister); void CALLNZp(int labelIndex, int callSiteIndex, const Src &predicateRegister);
void ELSE(Registers &r); void ELSE();
void ENDIF(Registers &r); void ENDIF();
void ENDLOOP(Registers &r); void ENDLOOP();
void ENDREP(Registers &r); void ENDREP();
void ENDWHILE(Registers &r); void ENDWHILE();
void IF(Registers &r, const Src &src); void IF(const Src &src);
void IFb(Registers &r, const Src &boolRegister); void IFb(const Src &boolRegister);
void IFp(Registers &r, const Src &predicateRegister); void IFp(const Src &predicateRegister);
void IFC(Registers &r, Vector4f &src0, Vector4f &src1, Control); void IFC(Vector4f &src0, Vector4f &src1, Control);
void IF(Registers &r, Int4 &condition); void IF(Int4 &condition);
void LABEL(int labelIndex); void LABEL(int labelIndex);
void LOOP(Registers &r, const Src &integerRegister); void LOOP(const Src &integerRegister);
void REP(Registers &r, const Src &integerRegister); void REP(const Src &integerRegister);
void WHILE(Registers &r, const Src &temporaryRegister); void WHILE(const Src &temporaryRegister);
void RET(Registers &r); void RET();
void LEAVE(Registers &r); void LEAVE();
void TEXLDL(Registers &r, Vector4f &dst, Vector4f &src, const Src&); void TEXLDL(Vector4f &dst, Vector4f &src, const Src&);
void TEX(Registers &r, Vector4f &dst, Vector4f &src, const Src&); void TEX(Vector4f &dst, Vector4f &src, const Src&);
void TEXOFFSET(Registers &r, Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2, Vector4f &src3); void TEXOFFSET(Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2, Vector4f &src3);
void TEXLDL(Registers &r, Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2); void TEXLDL(Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2);
void TEXELFETCH(Registers &r, Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2); void TEXELFETCH(Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2);
void TEXELFETCH(Registers &r, Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2, Vector4f &src3); void TEXELFETCH(Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2, Vector4f &src3);
void TEXGRAD(Registers &r, Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2, Vector4f &src3); void TEXGRAD(Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2, Vector4f &src3);
void TEXGRAD(Registers &r, Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2, Vector4f &src3, Vector4f &src4); void TEXGRAD(Vector4f &dst, Vector4f &src, const Src&, Vector4f &src2, Vector4f &src3, Vector4f &src4);
void TEXSIZE(Registers &r, Vector4f &dst, Float4 &lod, const Src&); void TEXSIZE(Vector4f &dst, Float4 &lod, const Src&);
void sampleTexture(Registers &r, Vector4f &c, const Src &s, Float4 &u, Float4 &v, Float4 &w, Float4 &q); void sampleTexture(Vector4f &c, const Src &s, Float4 &u, Float4 &v, Float4 &w, Float4 &q);
SamplerCore *sampler[VERTEX_TEXTURE_IMAGE_UNITS]; SamplerCore *sampler[VERTEX_TEXTURE_IMAGE_UNITS];
......
src/Shader/VertexRoutine.cpp
...@@ -63,13 +63,13 @@ namespace sw ...@@ -63,13 +63,13 @@ namespace sw
{ {
*Pointer<UInt>(tagCache + tagIndex) = indexQ; *Pointer<UInt>(tagCache + tagIndex) = indexQ;
readInput(r, indexQ); readInput(indexQ);
pipeline(r); pipeline();
postTransform(r); postTransform();
computeClipFlags(r); computeClipFlags();
Pointer<Byte> cacheLine0 = vertexCache + tagIndex * UInt((int)sizeof(Vertex)); Pointer<Byte> cacheLine0 = vertexCache + tagIndex * UInt((int)sizeof(Vertex));
writeCache(cacheLine0, r); writeCache(cacheLine0);
} }
UInt cacheIndex = index & 0x0000003F; UInt cacheIndex = index & 0x0000003F;
...@@ -85,18 +85,18 @@ namespace sw ...@@ -85,18 +85,18 @@ namespace sw
Return(); Return();
} }
void VertexRoutine::readInput(Registers &r, UInt &index) void VertexRoutine::readInput(UInt &index)
{ {
for(int i = 0; i < VERTEX_ATTRIBUTES; i++) for(int i = 0; i < VERTEX_ATTRIBUTES; i++)
{ {
Pointer<Byte> input = *Pointer<Pointer<Byte> >(r.data + OFFSET(DrawData,input) + sizeof(void*) * i); Pointer<Byte> input = *Pointer<Pointer<Byte> >(r.data + OFFSET(DrawData,input) + sizeof(void*) * i);
UInt stride = *Pointer<UInt>(r.data + OFFSET(DrawData,stride) + sizeof(unsigned int) * i); UInt stride = *Pointer<UInt>(r.data + OFFSET(DrawData,stride) + sizeof(unsigned int) * i);
r.v[i] = readStream(r, input, stride, state.input[i], index); r.v[i] = readStream(input, stride, state.input[i], index);
} }
} }
void VertexRoutine::computeClipFlags(Registers &r) void VertexRoutine::computeClipFlags()
{ {
int pos = state.positionRegister; int pos = state.positionRegister;
...@@ -136,7 +136,7 @@ namespace sw ...@@ -136,7 +136,7 @@ namespace sw
} }
} }
Vector4f VertexRoutine::readStream(Registers &r, Pointer<Byte> &buffer, UInt &stride, const Stream &stream, const UInt &index) Vector4f VertexRoutine::readStream(Pointer<Byte> &buffer, UInt &stride, const Stream &stream, const UInt &index)
{ {
const bool texldl = state.shaderContainsTexldl; const bool texldl = state.shaderContainsTexldl;
...@@ -444,7 +444,7 @@ namespace sw ...@@ -444,7 +444,7 @@ namespace sw
return v; return v;
} }
void VertexRoutine::postTransform(Registers &r) void VertexRoutine::postTransform()
{ {
int pos = state.positionRegister; int pos = state.positionRegister;
...@@ -482,7 +482,7 @@ namespace sw ...@@ -482,7 +482,7 @@ namespace sw
} }
} }
void VertexRoutine::writeCache(Pointer<Byte> &cacheLine, Registers &r) void VertexRoutine::writeCache(Pointer<Byte> &cacheLine)
{ {
Vector4f v; Vector4f v;
......
src/Shader/VertexRoutine.hpp
...@@ -85,15 +85,15 @@ namespace sw ...@@ -85,15 +85,15 @@ namespace sw
const VertexShader *const shader; const VertexShader *const shader;
private: private:
virtual void pipeline(Registers &r) = 0; virtual void pipeline() = 0;
typedef VertexProcessor::State::Input Stream; typedef VertexProcessor::State::Input Stream;
Vector4f readStream(Registers &r, Pointer<Byte> &buffer, UInt &stride, const Stream &stream, const UInt &index); Vector4f readStream(Pointer<Byte> &buffer, UInt &stride, const Stream &stream, const UInt &index);
void readInput(Registers &r, UInt &index); void readInput(UInt &index);
void computeClipFlags(Registers &r); void computeClipFlags();
void postTransform(Registers &r); void postTransform();
void writeCache(Pointer<Byte> &cacheLine, Registers &r); void writeCache(Pointer<Byte> &cacheLine);
void writeVertex(Pointer<Byte> &vertex, Pointer<Byte> &cacheLine); void writeVertex(Pointer<Byte> &vertex, Pointer<Byte> &cacheLine);
}; };
} }
......
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