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 337 additions and 337 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.cpp
...@@ -42,7 +42,7 @@ namespace sw ...@@ -42,7 +42,7 @@ namespace sw
} }
} }
void VertexProgram::pipeline(Registers &r) void VertexProgram::pipeline()
{ {
for(int i = 0; i < VERTEX_TEXTURE_IMAGE_UNITS; i++) for(int i = 0; i < VERTEX_TEXTURE_IMAGE_UNITS; i++)
{ {
...@@ -51,15 +51,15 @@ namespace sw ...@@ -51,15 +51,15 @@ namespace sw
if(!state.preTransformed) if(!state.preTransformed)
{ {
program(r); program();
} }
else else
{ {
passThrough(r); passThrough();
} }
} }
void VertexProgram::program(Registers &r) void VertexProgram::program()
{ {
// shader->print("VertexShader-%0.8X.txt", state.shaderID); // shader->print("VertexShader-%0.8X.txt", state.shaderID);
...@@ -87,7 +87,7 @@ namespace sw ...@@ -87,7 +87,7 @@ namespace sw
callRetBlock[dst.label].push_back(Nucleus::createBasicBlock()); callRetBlock[dst.label].push_back(Nucleus::createBasicBlock());
} }
} }
for(size_t i = 0; i < shader->getLength(); i++) for(size_t i = 0; i < shader->getLength(); i++)
{ {
const Shader::Instruction *instruction = shader->getInstruction(i); const Shader::Instruction *instruction = shader->getInstruction(i);
...@@ -117,187 +117,187 @@ namespace sw ...@@ -117,187 +117,187 @@ namespace sw
Vector4f s3; Vector4f s3;
Vector4f s4; Vector4f s4;
if(src0.type != Shader::PARAMETER_VOID) s0 = fetchRegisterF(r, src0); if(src0.type != Shader::PARAMETER_VOID) s0 = fetchRegisterF(src0);
if(src1.type != Shader::PARAMETER_VOID) s1 = fetchRegisterF(r, src1); if(src1.type != Shader::PARAMETER_VOID) s1 = fetchRegisterF(src1);
if(src2.type != Shader::PARAMETER_VOID) s2 = fetchRegisterF(r, src2); if(src2.type != Shader::PARAMETER_VOID) s2 = fetchRegisterF(src2);
if(src3.type != Shader::PARAMETER_VOID) s3 = fetchRegisterF(r, src3); if(src3.type != Shader::PARAMETER_VOID) s3 = fetchRegisterF(src3);
if(src4.type != Shader::PARAMETER_VOID) s4 = fetchRegisterF(r, src4); if(src4.type != Shader::PARAMETER_VOID) s4 = fetchRegisterF(src4);
switch(opcode) switch(opcode)
{ {
case Shader::OPCODE_VS_1_0: break; case Shader::OPCODE_VS_1_0: break;
case Shader::OPCODE_VS_1_1: break; case Shader::OPCODE_VS_1_1: break;
case Shader::OPCODE_VS_2_0: break; case Shader::OPCODE_VS_2_0: break;
case Shader::OPCODE_VS_2_x: break; case Shader::OPCODE_VS_2_x: break;
case Shader::OPCODE_VS_2_sw: break; case Shader::OPCODE_VS_2_sw: break;
case Shader::OPCODE_VS_3_0: break; case Shader::OPCODE_VS_3_0: break;
case Shader::OPCODE_VS_3_sw: break; case Shader::OPCODE_VS_3_sw: break;
case Shader::OPCODE_DCL: break; case Shader::OPCODE_DCL: break;
case Shader::OPCODE_DEF: break; case Shader::OPCODE_DEF: break;
case Shader::OPCODE_DEFI: break; case Shader::OPCODE_DEFI: break;
case Shader::OPCODE_DEFB: break; case Shader::OPCODE_DEFB: break;
case Shader::OPCODE_NOP: break; case Shader::OPCODE_NOP: break;
case Shader::OPCODE_ABS: abs(d, s0); break; case Shader::OPCODE_ABS: abs(d, s0); break;
case Shader::OPCODE_ADD: add(d, s0, s1); break; case Shader::OPCODE_ADD: add(d, s0, s1); break;
case Shader::OPCODE_IADD: iadd(d, s0, s1); break; case Shader::OPCODE_IADD: iadd(d, s0, s1); break;
case Shader::OPCODE_CRS: crs(d, s0, s1); break; case Shader::OPCODE_CRS: crs(d, s0, s1); break;
case Shader::OPCODE_FORWARD1: forward1(d, s0, s1, s2); break; case Shader::OPCODE_FORWARD1: forward1(d, s0, s1, s2); break;
case Shader::OPCODE_FORWARD2: forward2(d, s0, s1, s2); break; case Shader::OPCODE_FORWARD2: forward2(d, s0, s1, s2); break;
case Shader::OPCODE_FORWARD3: forward3(d, s0, s1, s2); break; case Shader::OPCODE_FORWARD3: forward3(d, s0, s1, s2); break;
case Shader::OPCODE_FORWARD4: forward4(d, s0, s1, s2); break; case Shader::OPCODE_FORWARD4: forward4(d, s0, s1, s2); break;
case Shader::OPCODE_REFLECT1: reflect1(d, s0, s1); break; case Shader::OPCODE_REFLECT1: reflect1(d, s0, s1); break;
case Shader::OPCODE_REFLECT2: reflect2(d, s0, s1); break; case Shader::OPCODE_REFLECT2: reflect2(d, s0, s1); break;
case Shader::OPCODE_REFLECT3: reflect3(d, s0, s1); break; case Shader::OPCODE_REFLECT3: reflect3(d, s0, s1); break;
case Shader::OPCODE_REFLECT4: reflect4(d, s0, s1); break; case Shader::OPCODE_REFLECT4: reflect4(d, s0, s1); break;
case Shader::OPCODE_REFRACT1: refract1(d, s0, s1, s2.x); break; case Shader::OPCODE_REFRACT1: refract1(d, s0, s1, s2.x); break;
case Shader::OPCODE_REFRACT2: refract2(d, s0, s1, s2.x); break; case Shader::OPCODE_REFRACT2: refract2(d, s0, s1, s2.x); break;
case Shader::OPCODE_REFRACT3: refract3(d, s0, s1, s2.x); break; case Shader::OPCODE_REFRACT3: refract3(d, s0, s1, s2.x); break;
case Shader::OPCODE_REFRACT4: refract4(d, s0, s1, s2.x); break; case Shader::OPCODE_REFRACT4: refract4(d, s0, s1, s2.x); break;
case Shader::OPCODE_DP1: dp1(d, s0, s1); break; case Shader::OPCODE_DP1: dp1(d, s0, s1); break;
case Shader::OPCODE_DP2: dp2(d, s0, s1); break; case Shader::OPCODE_DP2: dp2(d, s0, s1); break;
case Shader::OPCODE_DP3: dp3(d, s0, s1); break; case Shader::OPCODE_DP3: dp3(d, s0, s1); break;
case Shader::OPCODE_DP4: dp4(d, s0, s1); break; case Shader::OPCODE_DP4: dp4(d, s0, s1); break;
case Shader::OPCODE_DET2: det2(d, s0, s1); break; case Shader::OPCODE_DET2: det2(d, s0, s1); break;
case Shader::OPCODE_DET3: det3(d, s0, s1, s2); break; case Shader::OPCODE_DET3: det3(d, s0, s1, s2); break;
case Shader::OPCODE_DET4: det4(d, s0, s1, s2, s3); break; case Shader::OPCODE_DET4: det4(d, s0, s1, s2, s3); break;
case Shader::OPCODE_ATT: att(d, s0, s1); break; case Shader::OPCODE_ATT: att(d, s0, s1); break;
case Shader::OPCODE_EXP2X: exp2x(d, s0, pp); break; case Shader::OPCODE_EXP2X: exp2x(d, s0, pp); break;
case Shader::OPCODE_EXP2: exp2(d, s0, pp); break; case Shader::OPCODE_EXP2: exp2(d, s0, pp); break;
case Shader::OPCODE_EXPP: expp(d, s0, version); break; case Shader::OPCODE_EXPP: expp(d, s0, version); break;
case Shader::OPCODE_EXP: exp(d, s0, pp); break; case Shader::OPCODE_EXP: exp(d, s0, pp); break;
case Shader::OPCODE_FRC: frc(d, s0); break; case Shader::OPCODE_FRC: frc(d, s0); break;
case Shader::OPCODE_TRUNC: trunc(d, s0); break; case Shader::OPCODE_TRUNC: trunc(d, s0); break;
case Shader::OPCODE_FLOOR: floor(d, s0); break; case Shader::OPCODE_FLOOR: floor(d, s0); break;
case Shader::OPCODE_ROUND: round(d, s0); break; case Shader::OPCODE_ROUND: round(d, s0); break;
case Shader::OPCODE_ROUNDEVEN: roundEven(d, s0); break; case Shader::OPCODE_ROUNDEVEN: roundEven(d, s0); break;
case Shader::OPCODE_CEIL: ceil(d, s0); break; case Shader::OPCODE_CEIL: ceil(d, s0); break;
case Shader::OPCODE_LIT: lit(d, s0); break; case Shader::OPCODE_LIT: lit(d, s0); break;
case Shader::OPCODE_LOG2X: log2x(d, s0, pp); break; case Shader::OPCODE_LOG2X: log2x(d, s0, pp); break;
case Shader::OPCODE_LOG2: log2(d, s0, pp); break; case Shader::OPCODE_LOG2: log2(d, s0, pp); break;
case Shader::OPCODE_LOGP: logp(d, s0, version); break; case Shader::OPCODE_LOGP: logp(d, s0, version); break;
case Shader::OPCODE_LOG: log(d, s0, pp); break; case Shader::OPCODE_LOG: log(d, s0, pp); break;
case Shader::OPCODE_LRP: lrp(d, s0, s1, s2); break; case Shader::OPCODE_LRP: lrp(d, s0, s1, s2); break;
case Shader::OPCODE_STEP: step(d, s0, s1); break; case Shader::OPCODE_STEP: step(d, s0, s1); break;
case Shader::OPCODE_SMOOTH: smooth(d, s0, s1, s2); break; case Shader::OPCODE_SMOOTH: smooth(d, s0, s1, s2); break;
case Shader::OPCODE_FLOATBITSTOINT: case Shader::OPCODE_FLOATBITSTOINT:
case Shader::OPCODE_FLOATBITSTOUINT: case Shader::OPCODE_FLOATBITSTOUINT:
case Shader::OPCODE_INTBITSTOFLOAT: case Shader::OPCODE_INTBITSTOFLOAT:
case Shader::OPCODE_UINTBITSTOFLOAT: d = s0; break; case Shader::OPCODE_UINTBITSTOFLOAT: d = s0; break;
case Shader::OPCODE_M3X2: M3X2(r, d, s0, src1); break; case Shader::OPCODE_M3X2: M3X2(d, s0, src1); break;
case Shader::OPCODE_M3X3: M3X3(r, d, s0, src1); break; case Shader::OPCODE_M3X3: M3X3(d, s0, src1); break;
case Shader::OPCODE_M3X4: M3X4(r, d, s0, src1); break; case Shader::OPCODE_M3X4: M3X4(d, s0, src1); break;
case Shader::OPCODE_M4X3: M4X3(r, d, s0, src1); break; case Shader::OPCODE_M4X3: M4X3(d, s0, src1); break;
case Shader::OPCODE_M4X4: M4X4(r, d, s0, src1); break; case Shader::OPCODE_M4X4: M4X4(d, s0, src1); break;
case Shader::OPCODE_MAD: mad(d, s0, s1, s2); break; case Shader::OPCODE_MAD: mad(d, s0, s1, s2); break;
case Shader::OPCODE_IMAD: imad(d, s0, s1, s2); break; case Shader::OPCODE_IMAD: imad(d, s0, s1, s2); break;
case Shader::OPCODE_MAX: max(d, s0, s1); break; case Shader::OPCODE_MAX: max(d, s0, s1); break;
case Shader::OPCODE_IMAX: imax(d, s0, s1); break; case Shader::OPCODE_IMAX: imax(d, s0, s1); break;
case Shader::OPCODE_UMAX: umax(d, s0, s1); break; case Shader::OPCODE_UMAX: umax(d, s0, s1); break;
case Shader::OPCODE_MIN: min(d, s0, s1); break; case Shader::OPCODE_MIN: min(d, s0, s1); break;
case Shader::OPCODE_IMIN: imin(d, s0, s1); break; case Shader::OPCODE_IMIN: imin(d, s0, s1); break;
case Shader::OPCODE_UMIN: umin(d, s0, s1); break; case Shader::OPCODE_UMIN: umin(d, s0, s1); break;
case Shader::OPCODE_MOV: mov(d, s0, integer); break; case Shader::OPCODE_MOV: mov(d, s0, integer); break;
case Shader::OPCODE_MOVA: mov(d, s0, true); break; case Shader::OPCODE_MOVA: mov(d, s0, true); break;
case Shader::OPCODE_NEG: neg(d, s0); break; case Shader::OPCODE_NEG: neg(d, s0); break;
case Shader::OPCODE_INEG: ineg(d, s0); break; case Shader::OPCODE_INEG: ineg(d, s0); break;
case Shader::OPCODE_F2B: f2b(d, s0); break; case Shader::OPCODE_F2B: f2b(d, s0); break;
case Shader::OPCODE_B2F: b2f(d, s0); break; case Shader::OPCODE_B2F: b2f(d, s0); break;
case Shader::OPCODE_F2I: f2i(d, s0); break; case Shader::OPCODE_F2I: f2i(d, s0); break;
case Shader::OPCODE_I2F: i2f(d, s0); break; case Shader::OPCODE_I2F: i2f(d, s0); break;
case Shader::OPCODE_F2U: f2u(d, s0); break; case Shader::OPCODE_F2U: f2u(d, s0); break;
case Shader::OPCODE_U2F: u2f(d, s0); break; case Shader::OPCODE_U2F: u2f(d, s0); break;
case Shader::OPCODE_I2B: i2b(d, s0); break; case Shader::OPCODE_I2B: i2b(d, s0); break;
case Shader::OPCODE_B2I: b2i(d, s0); break; case Shader::OPCODE_B2I: b2i(d, s0); break;
case Shader::OPCODE_MUL: mul(d, s0, s1); break; case Shader::OPCODE_MUL: mul(d, s0, s1); break;
case Shader::OPCODE_IMUL: imul(d, s0, s1); break; case Shader::OPCODE_IMUL: imul(d, s0, s1); break;
case Shader::OPCODE_NRM2: nrm2(d, s0, pp); break; case Shader::OPCODE_NRM2: nrm2(d, s0, pp); break;
case Shader::OPCODE_NRM3: nrm3(d, s0, pp); break; case Shader::OPCODE_NRM3: nrm3(d, s0, pp); break;
case Shader::OPCODE_NRM4: nrm4(d, s0, pp); break; case Shader::OPCODE_NRM4: nrm4(d, s0, pp); break;
case Shader::OPCODE_POWX: powx(d, s0, s1, pp); break; case Shader::OPCODE_POWX: powx(d, s0, s1, pp); break;
case Shader::OPCODE_POW: pow(d, s0, s1, pp); break; case Shader::OPCODE_POW: pow(d, s0, s1, pp); break;
case Shader::OPCODE_RCPX: rcpx(d, s0, pp); break; case Shader::OPCODE_RCPX: rcpx(d, s0, pp); break;
case Shader::OPCODE_DIV: div(d, s0, s1); break; case Shader::OPCODE_DIV: div(d, s0, s1); break;
case Shader::OPCODE_IDIV: idiv(d, s0, s1); break; case Shader::OPCODE_IDIV: idiv(d, s0, s1); break;
case Shader::OPCODE_UDIV: udiv(d, s0, s1); break; case Shader::OPCODE_UDIV: udiv(d, s0, s1); break;
case Shader::OPCODE_MOD: mod(d, s0, s1); break; case Shader::OPCODE_MOD: mod(d, s0, s1); break;
case Shader::OPCODE_IMOD: imod(d, s0, s1); break; case Shader::OPCODE_IMOD: imod(d, s0, s1); break;
case Shader::OPCODE_UMOD: umod(d, s0, s1); break; case Shader::OPCODE_UMOD: umod(d, s0, s1); break;
case Shader::OPCODE_SHL: shl(d, s0, s1); break; case Shader::OPCODE_SHL: shl(d, s0, s1); break;
case Shader::OPCODE_ISHR: ishr(d, s0, s1); break; case Shader::OPCODE_ISHR: ishr(d, s0, s1); break;
case Shader::OPCODE_USHR: ushr(d, s0, s1); break; case Shader::OPCODE_USHR: ushr(d, s0, s1); break;
case Shader::OPCODE_RSQX: rsqx(d, s0, pp); break; case Shader::OPCODE_RSQX: rsqx(d, s0, pp); break;
case Shader::OPCODE_SQRT: sqrt(d, s0, pp); break; case Shader::OPCODE_SQRT: sqrt(d, s0, pp); break;
case Shader::OPCODE_RSQ: rsq(d, s0, pp); break; case Shader::OPCODE_RSQ: rsq(d, s0, pp); break;
case Shader::OPCODE_LEN2: len2(d.x, s0, pp); break; case Shader::OPCODE_LEN2: len2(d.x, s0, pp); break;
case Shader::OPCODE_LEN3: len3(d.x, s0, pp); break; case Shader::OPCODE_LEN3: len3(d.x, s0, pp); break;
case Shader::OPCODE_LEN4: len4(d.x, s0, pp); break; case Shader::OPCODE_LEN4: len4(d.x, s0, pp); break;
case Shader::OPCODE_DIST1: dist1(d.x, s0, s1, pp); break; case Shader::OPCODE_DIST1: dist1(d.x, s0, s1, pp); break;
case Shader::OPCODE_DIST2: dist2(d.x, s0, s1, pp); break; case Shader::OPCODE_DIST2: dist2(d.x, s0, s1, pp); break;
case Shader::OPCODE_DIST3: dist3(d.x, s0, s1, pp); break; case Shader::OPCODE_DIST3: dist3(d.x, s0, s1, pp); break;
case Shader::OPCODE_DIST4: dist4(d.x, s0, s1, pp); break; case Shader::OPCODE_DIST4: dist4(d.x, s0, s1, pp); break;
case Shader::OPCODE_SGE: step(d, s1, s0); break; case Shader::OPCODE_SGE: step(d, s1, s0); break;
case Shader::OPCODE_SGN: sgn(d, s0); break; case Shader::OPCODE_SGN: sgn(d, s0); break;
case Shader::OPCODE_SINCOS: sincos(d, s0, pp); break; case Shader::OPCODE_SINCOS: sincos(d, s0, pp); break;
case Shader::OPCODE_COS: cos(d, s0, pp); break; case Shader::OPCODE_COS: cos(d, s0, pp); break;
case Shader::OPCODE_SIN: sin(d, s0, pp); break; case Shader::OPCODE_SIN: sin(d, s0, pp); break;
case Shader::OPCODE_TAN: tan(d, s0); break; case Shader::OPCODE_TAN: tan(d, s0); break;
case Shader::OPCODE_ACOS: acos(d, s0); break; case Shader::OPCODE_ACOS: acos(d, s0); break;
case Shader::OPCODE_ASIN: asin(d, s0); break; case Shader::OPCODE_ASIN: asin(d, s0); break;
case Shader::OPCODE_ATAN: atan(d, s0); break; case Shader::OPCODE_ATAN: atan(d, s0); break;
case Shader::OPCODE_ATAN2: atan2(d, s0, s1); break; case Shader::OPCODE_ATAN2: atan2(d, s0, s1); break;
case Shader::OPCODE_COSH: cosh(d, s0, pp); break; case Shader::OPCODE_COSH: cosh(d, s0, pp); break;
case Shader::OPCODE_SINH: sinh(d, s0, pp); break; case Shader::OPCODE_SINH: sinh(d, s0, pp); break;
case Shader::OPCODE_TANH: tanh(d, s0, pp); break; case Shader::OPCODE_TANH: tanh(d, s0, pp); break;
case Shader::OPCODE_ACOSH: acosh(d, s0, pp); break; case Shader::OPCODE_ACOSH: acosh(d, s0, pp); break;
case Shader::OPCODE_ASINH: asinh(d, s0, pp); break; case Shader::OPCODE_ASINH: asinh(d, s0, pp); break;
case Shader::OPCODE_ATANH: atanh(d, s0, pp); break; case Shader::OPCODE_ATANH: atanh(d, s0, pp); break;
case Shader::OPCODE_SLT: slt(d, s0, s1); break; case Shader::OPCODE_SLT: slt(d, s0, s1); break;
case Shader::OPCODE_SUB: sub(d, s0, s1); break; case Shader::OPCODE_SUB: sub(d, s0, s1); break;
case Shader::OPCODE_ISUB: isub(d, s0, s1); break; case Shader::OPCODE_ISUB: isub(d, s0, s1); break;
case Shader::OPCODE_BREAK: BREAK(r); break; case Shader::OPCODE_BREAK: BREAK(); break;
case Shader::OPCODE_BREAKC: BREAKC(r, s0, s1, control); break; case Shader::OPCODE_BREAKC: BREAKC(s0, s1, control); break;
case Shader::OPCODE_BREAKP: BREAKP(r, src0); break; case Shader::OPCODE_BREAKP: BREAKP(src0); break;
case Shader::OPCODE_CONTINUE: CONTINUE(r); break; case Shader::OPCODE_CONTINUE: CONTINUE(); break;
case Shader::OPCODE_TEST: TEST(); break; case Shader::OPCODE_TEST: TEST(); break;
case Shader::OPCODE_CALL: CALL(r, dst.label, dst.callSite); break; case Shader::OPCODE_CALL: CALL(dst.label, dst.callSite); break;
case Shader::OPCODE_CALLNZ: CALLNZ(r, dst.label, dst.callSite, src0); break; case Shader::OPCODE_CALLNZ: CALLNZ(dst.label, dst.callSite, src0); break;
case Shader::OPCODE_ELSE: ELSE(r); break; case Shader::OPCODE_ELSE: ELSE(); break;
case Shader::OPCODE_ENDIF: ENDIF(r); break; case Shader::OPCODE_ENDIF: ENDIF(); break;
case Shader::OPCODE_ENDLOOP: ENDLOOP(r); break; case Shader::OPCODE_ENDLOOP: ENDLOOP(); break;
case Shader::OPCODE_ENDREP: ENDREP(r); break; case Shader::OPCODE_ENDREP: ENDREP(); break;
case Shader::OPCODE_ENDWHILE: ENDWHILE(r); break; case Shader::OPCODE_ENDWHILE: ENDWHILE(); break;
case Shader::OPCODE_IF: IF(r, src0); break; case Shader::OPCODE_IF: IF(src0); break;
case Shader::OPCODE_IFC: IFC(r, s0, s1, control); break; case Shader::OPCODE_IFC: IFC(s0, s1, control); break;
case Shader::OPCODE_LABEL: LABEL(dst.index); break; case Shader::OPCODE_LABEL: LABEL(dst.index); break;
case Shader::OPCODE_LOOP: LOOP(r, src1); break; case Shader::OPCODE_LOOP: LOOP(src1); break;
case Shader::OPCODE_REP: REP(r, src0); break; case Shader::OPCODE_REP: REP(src0); break;
case Shader::OPCODE_WHILE: WHILE(r, src0); break; case Shader::OPCODE_WHILE: WHILE(src0); break;
case Shader::OPCODE_RET: RET(r); break; case Shader::OPCODE_RET: RET(); break;
case Shader::OPCODE_LEAVE: LEAVE(r); break; case Shader::OPCODE_LEAVE: LEAVE(); break;
case Shader::OPCODE_CMP: cmp(d, s0, s1, control); break; case Shader::OPCODE_CMP: cmp(d, s0, s1, control); break;
case Shader::OPCODE_ICMP: icmp(d, s0, s1, control); break; case Shader::OPCODE_ICMP: icmp(d, s0, s1, control); break;
case Shader::OPCODE_UCMP: ucmp(d, s0, s1, control); break; case Shader::OPCODE_UCMP: ucmp(d, s0, s1, control); break;
case Shader::OPCODE_SELECT: select(d, s0, s1, s2); break; case Shader::OPCODE_SELECT: select(d, s0, s1, s2); break;
case Shader::OPCODE_EXTRACT: extract(d.x, s0, s1.x); break; case Shader::OPCODE_EXTRACT: extract(d.x, s0, s1.x); break;
case Shader::OPCODE_INSERT: insert(d, s0, s1.x, s2.x); break; case Shader::OPCODE_INSERT: insert(d, s0, s1.x, s2.x); break;
case Shader::OPCODE_ALL: all(d.x, s0); break; case Shader::OPCODE_ALL: all(d.x, s0); break;
case Shader::OPCODE_ANY: any(d.x, s0); break; case Shader::OPCODE_ANY: any(d.x, s0); break;
case Shader::OPCODE_NOT: not(d, s0); break; case Shader::OPCODE_NOT: not(d, s0); break;
case Shader::OPCODE_OR: or(d, s0, s1); break; case Shader::OPCODE_OR: or(d, s0, s1); break;
case Shader::OPCODE_XOR: xor(d, s0, s1); break; case Shader::OPCODE_XOR: xor(d, s0, s1); break;
case Shader::OPCODE_AND: and(d, s0, s1); break; case Shader::OPCODE_AND: and(d, s0, s1); break;
case Shader::OPCODE_EQ: equal(d, s0, s1); break; case Shader::OPCODE_EQ: equal(d, s0, s1); break;
case Shader::OPCODE_NE: notEqual(d, s0, s1); break; case Shader::OPCODE_NE: notEqual(d, s0, s1); break;
case Shader::OPCODE_TEXLDL: TEXLDL(r, d, s0, src1); break; case Shader::OPCODE_TEXLDL: TEXLDL(d, s0, src1); break;
case Shader::OPCODE_TEX: TEX(r, d, s0, src1); break; case Shader::OPCODE_TEX: TEX(d, s0, src1); break;
case Shader::OPCODE_TEXOFFSET: TEXOFFSET(r, d, s0, src1, s2, s3); break; case Shader::OPCODE_TEXOFFSET: TEXOFFSET(d, s0, src1, s2, s3); break;
case Shader::OPCODE_TEXLDLOFFSET: TEXLDL(r, d, s0, src1, s2); break; case Shader::OPCODE_TEXLDLOFFSET: TEXLDL(d, s0, src1, s2); break;
case Shader::OPCODE_TEXELFETCH: TEXELFETCH(r, d, s0, src1, s2); break; case Shader::OPCODE_TEXELFETCH: TEXELFETCH(d, s0, src1, s2); break;
case Shader::OPCODE_TEXELFETCHOFFSET: TEXELFETCH(r, d, s0, src1, s2, s3); break; case Shader::OPCODE_TEXELFETCHOFFSET: TEXELFETCH(d, s0, src1, s2, s3); break;
case Shader::OPCODE_TEXGRAD: TEXGRAD(r, d, s0, src1, s2, s3); break; case Shader::OPCODE_TEXGRAD: TEXGRAD(d, s0, src1, s2, s3); break;
case Shader::OPCODE_TEXGRADOFFSET: TEXGRAD(r, d, s0, src1, s2, s3, s4); break; case Shader::OPCODE_TEXGRADOFFSET: TEXGRAD(d, s0, src1, s2, s3, s4); break;
case Shader::OPCODE_TEXSIZE: TEXSIZE(r, d, s0.x, src1); break; case Shader::OPCODE_TEXSIZE: TEXSIZE(d, s0.x, src1); break;
case Shader::OPCODE_END: break; case Shader::OPCODE_END: break;
default: default:
ASSERT(false); ASSERT(false);
} }
...@@ -338,7 +338,7 @@ namespace sw ...@@ -338,7 +338,7 @@ namespace sw
switch(dst.type) switch(dst.type)
{ {
case Shader::PARAMETER_VOID: break; case Shader::PARAMETER_VOID: break;
case Shader::PARAMETER_TEMP: case Shader::PARAMETER_TEMP:
if(dst.rel.type == Shader::PARAMETER_VOID) if(dst.rel.type == Shader::PARAMETER_VOID)
{ {
...@@ -349,7 +349,7 @@ namespace sw ...@@ -349,7 +349,7 @@ namespace sw
} }
else else
{ {
Int a = relativeAddress(r, dst); Int a = relativeAddress(dst);
if(dst.x) pDst.x = r.r[dst.index + a].x; if(dst.x) pDst.x = r.r[dst.index + a].x;
if(dst.y) pDst.y = r.r[dst.index + a].y; if(dst.y) pDst.y = r.r[dst.index + a].y;
...@@ -357,7 +357,7 @@ namespace sw ...@@ -357,7 +357,7 @@ namespace sw
if(dst.w) pDst.w = r.r[dst.index + a].w; if(dst.w) pDst.w = r.r[dst.index + a].w;
} }
break; break;
case Shader::PARAMETER_ADDR: pDst = r.a0; break; case Shader::PARAMETER_ADDR: pDst = r.a0; break;
case Shader::PARAMETER_RASTOUT: case Shader::PARAMETER_RASTOUT:
switch(dst.index) switch(dst.index)
{ {
...@@ -412,7 +412,7 @@ namespace sw ...@@ -412,7 +412,7 @@ namespace sw
} }
else else
{ {
Int a = relativeAddress(r, dst); Int a = relativeAddress(dst);
if(dst.x) pDst.x = r.o[dst.index + a].x; if(dst.x) pDst.x = r.o[dst.index + a].x;
if(dst.y) pDst.y = r.o[dst.index + a].y; if(dst.y) pDst.y = r.o[dst.index + a].y;
...@@ -421,14 +421,14 @@ namespace sw ...@@ -421,14 +421,14 @@ namespace sw
} }
} }
break; break;
case Shader::PARAMETER_LABEL: break; case Shader::PARAMETER_LABEL: break;
case Shader::PARAMETER_PREDICATE: pDst = r.p0; break; case Shader::PARAMETER_PREDICATE: pDst = r.p0; break;
case Shader::PARAMETER_INPUT: break; case Shader::PARAMETER_INPUT: break;
default: default:
ASSERT(false); ASSERT(false);
} }
Int4 enable = enableMask(r, instruction); Int4 enable = enableMask(instruction);
Int4 xEnable = enable; Int4 xEnable = enable;
Int4 yEnable = enable; Int4 yEnable = enable;
...@@ -485,7 +485,7 @@ namespace sw ...@@ -485,7 +485,7 @@ namespace sw
} }
else else
{ {
Int a = relativeAddress(r, dst); Int a = relativeAddress(dst);
if(dst.x) r.r[dst.index + a].x = d.x; if(dst.x) r.r[dst.index + a].x = d.x;
if(dst.y) r.r[dst.index + a].y = d.y; if(dst.y) r.r[dst.index + a].y = d.y;
...@@ -511,13 +511,13 @@ namespace sw ...@@ -511,13 +511,13 @@ namespace sw
case 1: case 1:
r.o[Fog].x = d.x; r.o[Fog].x = d.x;
break; break;
case 2: case 2:
r.o[Pts].y = d.x; r.o[Pts].y = d.x;
break; break;
default: ASSERT(false); default: ASSERT(false);
} }
break; break;
case Shader::PARAMETER_ATTROUT: case Shader::PARAMETER_ATTROUT:
if(dst.x) r.o[D0 + dst.index].x = d.x; if(dst.x) r.o[D0 + dst.index].x = d.x;
if(dst.y) r.o[D0 + dst.index].y = d.y; if(dst.y) r.o[D0 + dst.index].y = d.y;
if(dst.z) r.o[D0 + dst.index].z = d.z; if(dst.z) r.o[D0 + dst.index].z = d.z;
...@@ -552,7 +552,7 @@ namespace sw ...@@ -552,7 +552,7 @@ namespace sw
} }
else else
{ {
Int a = relativeAddress(r, dst); Int a = relativeAddress(dst);
if(dst.x) r.o[dst.index + a].x = d.x; if(dst.x) r.o[dst.index + a].x = d.x;
if(dst.y) r.o[dst.index + a].y = d.y; if(dst.y) r.o[dst.index + a].y = d.y;
...@@ -561,9 +561,9 @@ namespace sw ...@@ -561,9 +561,9 @@ namespace sw
} }
} }
break; break;
case Shader::PARAMETER_LABEL: break; case Shader::PARAMETER_LABEL: break;
case Shader::PARAMETER_PREDICATE: r.p0 = d; break; case Shader::PARAMETER_PREDICATE: r.p0 = d; break;
case Shader::PARAMETER_INPUT: break; case Shader::PARAMETER_INPUT: break;
default: default:
ASSERT(false); ASSERT(false);
} }
...@@ -576,7 +576,7 @@ namespace sw ...@@ -576,7 +576,7 @@ namespace sw
} }
} }
void VertexProgram::passThrough(Registers &r) void VertexProgram::passThrough()
{ {
if(shader) if(shader)
{ {
...@@ -644,7 +644,7 @@ namespace sw ...@@ -644,7 +644,7 @@ namespace sw
} }
} }
Vector4f VertexProgram::fetchRegisterF(Registers &r, const Src &src, unsigned int offset) Vector4f VertexProgram::fetchRegisterF(const Src &src, unsigned int offset)
{ {
Vector4f reg; Vector4f reg;
unsigned int i = src.index + offset; unsigned int i = src.index + offset;
...@@ -658,11 +658,11 @@ namespace sw ...@@ -658,11 +658,11 @@ namespace sw
} }
else else
{ {
reg = r.r[i + relativeAddress(r, src)]; reg = r.r[i + relativeAddress(src)];
} }
break; break;
case Shader::PARAMETER_CONST: case Shader::PARAMETER_CONST:
reg = readConstant(r, src, offset); reg = readConstant(src, offset);
break; break;
case Shader::PARAMETER_INPUT: case Shader::PARAMETER_INPUT:
if(src.rel.type == Shader::PARAMETER_VOID) if(src.rel.type == Shader::PARAMETER_VOID)
...@@ -671,7 +671,7 @@ namespace sw ...@@ -671,7 +671,7 @@ namespace sw
} }
else else
{ {
reg = r.v[i + relativeAddress(r, src)]; reg = r.v[i + relativeAddress(src)];
} }
break; break;
case Shader::PARAMETER_VOID: return r.r[0]; // Dummy case Shader::PARAMETER_VOID: return r.r[0]; // Dummy
...@@ -703,7 +703,7 @@ namespace sw ...@@ -703,7 +703,7 @@ namespace sw
} }
else else
{ {
reg = r.o[i + relativeAddress(r, src)]; reg = r.o[i + relativeAddress(src)];
} }
break; break;
case Shader::PARAMETER_MISCTYPE: case Shader::PARAMETER_MISCTYPE:
...@@ -759,7 +759,7 @@ namespace sw ...@@ -759,7 +759,7 @@ namespace sw
return mod; return mod;
} }
Vector4f VertexProgram::readConstant(Registers &r, const Src &src, unsigned int offset) Vector4f VertexProgram::readConstant(const Src &src, unsigned int offset)
{ {
Vector4f c; Vector4f c;
unsigned int i = src.index + offset; unsigned int i = src.index + offset;
...@@ -809,8 +809,8 @@ namespace sw ...@@ -809,8 +809,8 @@ namespace sw
{ {
if(src.rel.deterministic) if(src.rel.deterministic)
{ {
Int a = relativeAddress(r, src); Int a = relativeAddress(src);
c.x = c.y = c.z = c.w = *Pointer<Float4>(r.data + OFFSET(DrawData,vs.c[i]) + a * 16); c.x = c.y = c.z = c.w = *Pointer<Float4>(r.data + OFFSET(DrawData,vs.c[i]) + a * 16);
c.x = c.x.xxxx; c.x = c.x.xxxx;
...@@ -836,7 +836,7 @@ namespace sw ...@@ -836,7 +836,7 @@ namespace sw
Int4 index = Int4(i) + As<Int4>(a) * Int4(src.rel.scale); Int4 index = Int4(i) + As<Int4>(a) * Int4(src.rel.scale);
index = Min(As<UInt4>(index), UInt4(VERTEX_UNIFORM_VECTORS)); // Clamp to constant register range, c[VERTEX_UNIFORM_VECTORS] = {0, 0, 0, 0} index = Min(As<UInt4>(index), UInt4(VERTEX_UNIFORM_VECTORS)); // Clamp to constant register range, c[VERTEX_UNIFORM_VECTORS] = {0, 0, 0, 0}
Int index0 = Extract(index, 0); Int index0 = Extract(index, 0);
Int index1 = Extract(index, 1); Int index1 = Extract(index, 1);
Int index2 = Extract(index, 2); Int index2 = Extract(index, 2);
...@@ -854,7 +854,7 @@ namespace sw ...@@ -854,7 +854,7 @@ namespace sw
return c; return c;
} }
Int VertexProgram::relativeAddress(Registers &r, const Shader::Parameter &var) Int VertexProgram::relativeAddress(const Shader::Parameter &var)
{ {
ASSERT(var.rel.deterministic); ASSERT(var.rel.deterministic);
...@@ -881,10 +881,10 @@ namespace sw ...@@ -881,10 +881,10 @@ namespace sw
return 0; return 0;
} }
Int4 VertexProgram::enableMask(Registers &r, const Shader::Instruction *instruction) Int4 VertexProgram::enableMask(const Shader::Instruction *instruction)
{ {
Int4 enable = instruction->analysisBranch ? Int4(r.enableStack[r.enableIndex]) : Int4(0xFFFFFFFF); Int4 enable = instruction->analysisBranch ? Int4(r.enableStack[r.enableIndex]) : Int4(0xFFFFFFFF);
if(!whileTest) if(!whileTest)
{ {
if(shader->containsBreakInstruction() && instruction->analysisBreak) if(shader->containsBreakInstruction() && instruction->analysisBreak)
...@@ -906,32 +906,32 @@ namespace sw ...@@ -906,32 +906,32 @@ namespace sw
return enable; return enable;
} }
void VertexProgram::M3X2(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1) void VertexProgram::M3X2(Vector4f &dst, Vector4f &src0, Src &src1)
{ {
Vector4f row0 = fetchRegisterF(r, src1, 0); Vector4f row0 = fetchRegisterF(src1, 0);
Vector4f row1 = fetchRegisterF(r, src1, 1); Vector4f row1 = fetchRegisterF(src1, 1);
dst.x = dot3(src0, row0); dst.x = dot3(src0, row0);
dst.y = dot3(src0, row1); dst.y = dot3(src0, row1);
} }
void VertexProgram::M3X3(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1) void VertexProgram::M3X3(Vector4f &dst, Vector4f &src0, Src &src1)
{ {
Vector4f row0 = fetchRegisterF(r, src1, 0); Vector4f row0 = fetchRegisterF(src1, 0);
Vector4f row1 = fetchRegisterF(r, src1, 1); Vector4f row1 = fetchRegisterF(src1, 1);
Vector4f row2 = fetchRegisterF(r, src1, 2); Vector4f row2 = fetchRegisterF(src1, 2);
dst.x = dot3(src0, row0); dst.x = dot3(src0, row0);
dst.y = dot3(src0, row1); dst.y = dot3(src0, row1);
dst.z = dot3(src0, row2); dst.z = dot3(src0, row2);
} }
void VertexProgram::M3X4(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1) void VertexProgram::M3X4(Vector4f &dst, Vector4f &src0, Src &src1)
{ {
Vector4f row0 = fetchRegisterF(r, src1, 0); Vector4f row0 = fetchRegisterF(src1, 0);
Vector4f row1 = fetchRegisterF(r, src1, 1); Vector4f row1 = fetchRegisterF(src1, 1);
Vector4f row2 = fetchRegisterF(r, src1, 2); Vector4f row2 = fetchRegisterF(src1, 2);
Vector4f row3 = fetchRegisterF(r, src1, 3); Vector4f row3 = fetchRegisterF(src1, 3);
dst.x = dot3(src0, row0); dst.x = dot3(src0, row0);
dst.y = dot3(src0, row1); dst.y = dot3(src0, row1);
...@@ -939,23 +939,23 @@ namespace sw ...@@ -939,23 +939,23 @@ namespace sw
dst.w = dot3(src0, row3); dst.w = dot3(src0, row3);
} }
void VertexProgram::M4X3(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1) void VertexProgram::M4X3(Vector4f &dst, Vector4f &src0, Src &src1)
{ {
Vector4f row0 = fetchRegisterF(r, src1, 0); Vector4f row0 = fetchRegisterF(src1, 0);
Vector4f row1 = fetchRegisterF(r, src1, 1); Vector4f row1 = fetchRegisterF(src1, 1);
Vector4f row2 = fetchRegisterF(r, src1, 2); Vector4f row2 = fetchRegisterF(src1, 2);
dst.x = dot4(src0, row0); dst.x = dot4(src0, row0);
dst.y = dot4(src0, row1); dst.y = dot4(src0, row1);
dst.z = dot4(src0, row2); dst.z = dot4(src0, row2);
} }
void VertexProgram::M4X4(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1) void VertexProgram::M4X4(Vector4f &dst, Vector4f &src0, Src &src1)
{ {
Vector4f row0 = fetchRegisterF(r, src1, 0); Vector4f row0 = fetchRegisterF(src1, 0);
Vector4f row1 = fetchRegisterF(r, src1, 1); Vector4f row1 = fetchRegisterF(src1, 1);
Vector4f row2 = fetchRegisterF(r, src1, 2); Vector4f row2 = fetchRegisterF(src1, 2);
Vector4f row3 = fetchRegisterF(r, src1, 3); Vector4f row3 = fetchRegisterF(src1, 3);
dst.x = dot4(src0, row0); dst.x = dot4(src0, row0);
dst.y = dot4(src0, row1); dst.y = dot4(src0, row1);
...@@ -963,7 +963,7 @@ namespace sw ...@@ -963,7 +963,7 @@ namespace sw
dst.w = dot4(src0, row3); dst.w = dot4(src0, row3);
} }
void VertexProgram::BREAK(Registers &r) void VertexProgram::BREAK()
{ {
llvm::BasicBlock *deadBlock = Nucleus::createBasicBlock(); llvm::BasicBlock *deadBlock = Nucleus::createBasicBlock();
llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1]; llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
...@@ -986,7 +986,7 @@ namespace sw ...@@ -986,7 +986,7 @@ namespace sw
r.enableIndex = r.enableIndex + breakDepth; r.enableIndex = r.enableIndex + breakDepth;
} }
void VertexProgram::BREAKC(Registers &r, Vector4f &src0, Vector4f &src1, Control control) void VertexProgram::BREAKC(Vector4f &src0, Vector4f &src1, Control control)
{ {
Int4 condition; Int4 condition;
...@@ -1002,10 +1002,10 @@ namespace sw ...@@ -1002,10 +1002,10 @@ namespace sw
ASSERT(false); ASSERT(false);
} }
BREAK(r, condition); BREAK(condition);
} }
void VertexProgram::BREAKP(Registers &r, const Src &predicateRegister) // FIXME: Factor out parts common with BREAKC void VertexProgram::BREAKP(const Src &predicateRegister) // FIXME: Factor out parts common with BREAKC
{ {
Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]); Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]);
...@@ -1014,10 +1014,10 @@ namespace sw ...@@ -1014,10 +1014,10 @@ namespace sw
condition = ~condition; condition = ~condition;
} }
BREAK(r, condition); BREAK(condition);
} }
void VertexProgram::BREAK(Registers &r, Int4 &condition) void VertexProgram::BREAK(Int4 &condition)
{ {
condition &= r.enableStack[r.enableIndex]; condition &= r.enableStack[r.enableIndex];
...@@ -1034,7 +1034,7 @@ namespace sw ...@@ -1034,7 +1034,7 @@ namespace sw
r.enableIndex = r.enableIndex + breakDepth; r.enableIndex = r.enableIndex + breakDepth;
} }
void VertexProgram::CONTINUE(Registers &r) void VertexProgram::CONTINUE()
{ {
r.enableContinue = r.enableContinue & ~r.enableStack[r.enableIndex]; r.enableContinue = r.enableContinue & ~r.enableStack[r.enableIndex];
} }
...@@ -1044,7 +1044,7 @@ namespace sw ...@@ -1044,7 +1044,7 @@ namespace sw
whileTest = true; whileTest = true;
} }
void VertexProgram::CALL(Registers &r, int labelIndex, int callSiteIndex) void VertexProgram::CALL(int labelIndex, int callSiteIndex)
{ {
if(!labelBlock[labelIndex]) if(!labelBlock[labelIndex])
{ {
...@@ -1064,26 +1064,26 @@ namespace sw ...@@ -1064,26 +1064,26 @@ namespace sw
r.enableLeave = restoreLeave; r.enableLeave = restoreLeave;
} }
void VertexProgram::CALLNZ(Registers &r, int labelIndex, int callSiteIndex, const Src &src) void VertexProgram::CALLNZ(int labelIndex, int callSiteIndex, const Src &src)
{ {
if(src.type == Shader::PARAMETER_CONSTBOOL) if(src.type == Shader::PARAMETER_CONSTBOOL)
{ {
CALLNZb(r, labelIndex, callSiteIndex, src); CALLNZb(labelIndex, callSiteIndex, src);
} }
else if(src.type == Shader::PARAMETER_PREDICATE) else if(src.type == Shader::PARAMETER_PREDICATE)
{ {
CALLNZp(r, labelIndex, callSiteIndex, src); CALLNZp(labelIndex, callSiteIndex, src);
} }
else ASSERT(false); else ASSERT(false);
} }
void VertexProgram::CALLNZb(Registers &r, int labelIndex, int callSiteIndex, const Src &boolRegister) void VertexProgram::CALLNZb(int labelIndex, int callSiteIndex, const Src &boolRegister)
{ {
Bool condition = (*Pointer<Byte>(r.data + OFFSET(DrawData,vs.b[boolRegister.index])) != Byte(0)); // FIXME Bool condition = (*Pointer<Byte>(r.data + OFFSET(DrawData,vs.b[boolRegister.index])) != Byte(0)); // FIXME
if(boolRegister.modifier == Shader::MODIFIER_NOT) if(boolRegister.modifier == Shader::MODIFIER_NOT)
{ {
condition = !condition; condition = !condition;
} }
if(!labelBlock[labelIndex]) if(!labelBlock[labelIndex])
...@@ -1104,7 +1104,7 @@ namespace sw ...@@ -1104,7 +1104,7 @@ namespace sw
r.enableLeave = restoreLeave; r.enableLeave = restoreLeave;
} }
void VertexProgram::CALLNZp(Registers &r, int labelIndex, int callSiteIndex, const Src &predicateRegister) void VertexProgram::CALLNZp(int labelIndex, int callSiteIndex, const Src &predicateRegister)
{ {
Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]); Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]);
...@@ -1137,7 +1137,7 @@ namespace sw ...@@ -1137,7 +1137,7 @@ namespace sw
r.enableLeave = restoreLeave; r.enableLeave = restoreLeave;
} }
void VertexProgram::ELSE(Registers &r) void VertexProgram::ELSE()
{ {
ifDepth--; ifDepth--;
...@@ -1164,7 +1164,7 @@ namespace sw ...@@ -1164,7 +1164,7 @@ namespace sw
ifDepth++; ifDepth++;
} }
void VertexProgram::ENDIF(Registers &r) void VertexProgram::ENDIF()
{ {
ifDepth--; ifDepth--;
...@@ -1180,7 +1180,7 @@ namespace sw ...@@ -1180,7 +1180,7 @@ namespace sw
} }
} }
void VertexProgram::ENDLOOP(Registers &r) void VertexProgram::ENDLOOP()
{ {
loopRepDepth--; loopRepDepth--;
...@@ -1196,7 +1196,7 @@ namespace sw ...@@ -1196,7 +1196,7 @@ namespace sw
r.enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF); r.enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
} }
void VertexProgram::ENDREP(Registers &r) void VertexProgram::ENDREP()
{ {
loopRepDepth--; loopRepDepth--;
...@@ -1210,7 +1210,7 @@ namespace sw ...@@ -1210,7 +1210,7 @@ namespace sw
r.enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF); r.enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
} }
void VertexProgram::ENDWHILE(Registers &r) void VertexProgram::ENDWHILE()
{ {
loopRepDepth--; loopRepDepth--;
...@@ -1225,24 +1225,24 @@ namespace sw ...@@ -1225,24 +1225,24 @@ namespace sw
whileTest = false; whileTest = false;
} }
void VertexProgram::IF(Registers &r, const Src &src) void VertexProgram::IF(const Src &src)
{ {
if(src.type == Shader::PARAMETER_CONSTBOOL) if(src.type == Shader::PARAMETER_CONSTBOOL)
{ {
IFb(r, src); IFb(src);
} }
else if(src.type == Shader::PARAMETER_PREDICATE) else if(src.type == Shader::PARAMETER_PREDICATE)
{ {
IFp(r, src); IFp(src);
} }
else else
{ {
Int4 condition = As<Int4>(fetchRegisterF(r, src).x); Int4 condition = As<Int4>(fetchRegisterF(src).x);
IF(r, condition); IF(condition);
} }
} }
void VertexProgram::IFb(Registers &r, const Src &boolRegister) void VertexProgram::IFb(const Src &boolRegister)
{ {
ASSERT(ifDepth < 24 + 4); ASSERT(ifDepth < 24 + 4);
...@@ -1264,7 +1264,7 @@ namespace sw ...@@ -1264,7 +1264,7 @@ namespace sw
ifDepth++; ifDepth++;
} }
void VertexProgram::IFp(Registers &r, const Src &predicateRegister) void VertexProgram::IFp(const Src &predicateRegister)
{ {
Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]); Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]);
...@@ -1273,10 +1273,10 @@ namespace sw ...@@ -1273,10 +1273,10 @@ namespace sw
condition = ~condition; condition = ~condition;
} }
IF(r, condition); IF(condition);
} }
void VertexProgram::IFC(Registers &r, Vector4f &src0, Vector4f &src1, Control control) void VertexProgram::IFC(Vector4f &src0, Vector4f &src1, Control control)
{ {
Int4 condition; Int4 condition;
...@@ -1292,10 +1292,10 @@ namespace sw ...@@ -1292,10 +1292,10 @@ namespace sw
ASSERT(false); ASSERT(false);
} }
IF(r, condition); IF(condition);
} }
void VertexProgram::IF(Registers &r, Int4 &condition) void VertexProgram::IF(Int4 &condition)
{ {
condition &= r.enableStack[r.enableIndex]; condition &= r.enableStack[r.enableIndex];
...@@ -1327,7 +1327,7 @@ namespace sw ...@@ -1327,7 +1327,7 @@ namespace sw
currentLabel = labelIndex; currentLabel = labelIndex;
} }
void VertexProgram::LOOP(Registers &r, const Src &integerRegister) void VertexProgram::LOOP(const Src &integerRegister)
{ {
r.loopDepth++; r.loopDepth++;
...@@ -1356,12 +1356,12 @@ namespace sw ...@@ -1356,12 +1356,12 @@ namespace sw
Nucleus::setInsertBlock(loopBlock); Nucleus::setInsertBlock(loopBlock);
r.iteration[r.loopDepth] = r.iteration[r.loopDepth] - 1; // FIXME: -- r.iteration[r.loopDepth] = r.iteration[r.loopDepth] - 1; // FIXME: --
loopRepDepth++; loopRepDepth++;
breakDepth = 0; breakDepth = 0;
} }
void VertexProgram::REP(Registers &r, const Src &integerRegister) void VertexProgram::REP(const Src &integerRegister)
{ {
r.loopDepth++; r.loopDepth++;
...@@ -1388,14 +1388,14 @@ namespace sw ...@@ -1388,14 +1388,14 @@ namespace sw
breakDepth = 0; breakDepth = 0;
} }
void VertexProgram::WHILE(Registers &r, const Src &temporaryRegister) void VertexProgram::WHILE(const Src &temporaryRegister)
{ {
r.enableIndex++; r.enableIndex++;
llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock(); llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
llvm::BasicBlock *testBlock = Nucleus::createBasicBlock(); llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
llvm::BasicBlock *endBlock = Nucleus::createBasicBlock(); llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
loopRepTestBlock[loopRepDepth] = testBlock; loopRepTestBlock[loopRepDepth] = testBlock;
loopRepEndBlock[loopRepDepth] = endBlock; loopRepEndBlock[loopRepDepth] = endBlock;
...@@ -1407,24 +1407,24 @@ namespace sw ...@@ -1407,24 +1407,24 @@ namespace sw
Nucleus::setInsertBlock(testBlock); Nucleus::setInsertBlock(testBlock);
r.enableContinue = restoreContinue; r.enableContinue = restoreContinue;
const Vector4f &src = fetchRegisterF(r, temporaryRegister); const Vector4f &src = fetchRegisterF(temporaryRegister);
Int4 condition = As<Int4>(src.x); Int4 condition = As<Int4>(src.x);
condition &= r.enableStack[r.enableIndex - 1]; condition &= r.enableStack[r.enableIndex - 1];
r.enableStack[r.enableIndex] = condition; r.enableStack[r.enableIndex] = condition;
Bool notAllFalse = SignMask(condition) != 0; Bool notAllFalse = SignMask(condition) != 0;
branch(notAllFalse, loopBlock, endBlock); branch(notAllFalse, loopBlock, endBlock);
Nucleus::setInsertBlock(endBlock); Nucleus::setInsertBlock(endBlock);
r.enableBreak = restoreBreak; r.enableBreak = restoreBreak;
Nucleus::setInsertBlock(loopBlock); Nucleus::setInsertBlock(loopBlock);
loopRepDepth++; loopRepDepth++;
breakDepth = 0; breakDepth = 0;
} }
void VertexProgram::RET(Registers &r) void VertexProgram::RET()
{ {
if(currentLabel == -1) if(currentLabel == -1)
{ {
...@@ -1439,7 +1439,7 @@ namespace sw ...@@ -1439,7 +1439,7 @@ namespace sw
{ {
// FIXME: Encapsulate // FIXME: Encapsulate
UInt index = r.callStack[--r.stackIndex]; UInt index = r.callStack[--r.stackIndex];
llvm::Value *value = index.loadValue(); llvm::Value *value = index.loadValue();
llvm::Value *switchInst = Nucleus::createSwitch(value, unreachableBlock, (int)callRetBlock[currentLabel].size()); llvm::Value *switchInst = Nucleus::createSwitch(value, unreachableBlock, (int)callRetBlock[currentLabel].size());
...@@ -1462,7 +1462,7 @@ namespace sw ...@@ -1462,7 +1462,7 @@ namespace sw
} }
} }
void VertexProgram::LEAVE(Registers &r) void VertexProgram::LEAVE()
{ {
r.enableLeave = r.enableLeave & ~r.enableStack[r.enableIndex]; r.enableLeave = r.enableLeave & ~r.enableStack[r.enableIndex];
...@@ -1470,10 +1470,10 @@ namespace sw ...@@ -1470,10 +1470,10 @@ namespace sw
// FIXME: Use enableLeave in other control-flow constructs // FIXME: Use enableLeave in other control-flow constructs
} }
void VertexProgram::TEXLDL(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1) void VertexProgram::TEXLDL(Vector4f &dst, Vector4f &src0, const Src &src1)
{ {
Vector4f tmp; Vector4f tmp;
sampleTexture(r, tmp, src1, src0.x, src0.y, src0.z, src0.w); sampleTexture(tmp, src1, src0.x, src0.y, src0.z, src0.w);
dst.x = tmp[(src1.swizzle >> 0) & 0x3]; dst.x = tmp[(src1.swizzle >> 0) & 0x3];
dst.y = tmp[(src1.swizzle >> 2) & 0x3]; dst.y = tmp[(src1.swizzle >> 2) & 0x3];
...@@ -1481,11 +1481,11 @@ namespace sw ...@@ -1481,11 +1481,11 @@ namespace sw
dst.w = tmp[(src1.swizzle >> 6) & 0x3]; dst.w = tmp[(src1.swizzle >> 6) & 0x3];
} }
void VertexProgram::TEX(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1) void VertexProgram::TEX(Vector4f &dst, Vector4f &src0, const Src &src1)
{ {
Float4 lod = Float4(0.0f); Float4 lod = Float4(0.0f);
Vector4f tmp; Vector4f tmp;
sampleTexture(r, tmp, src1, src0.x, src0.y, src0.z, lod); sampleTexture(tmp, src1, src0.x, src0.y, src0.z, lod);
dst.x = tmp[(src1.swizzle >> 0) & 0x3]; dst.x = tmp[(src1.swizzle >> 0) & 0x3];
dst.y = tmp[(src1.swizzle >> 2) & 0x3]; dst.y = tmp[(src1.swizzle >> 2) & 0x3];
...@@ -1493,37 +1493,37 @@ namespace sw ...@@ -1493,37 +1493,37 @@ namespace sw
dst.w = tmp[(src1.swizzle >> 6) & 0x3]; dst.w = tmp[(src1.swizzle >> 6) & 0x3];
} }
void VertexProgram::TEXOFFSET(Registers &r, Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3) void VertexProgram::TEXOFFSET(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3)
{ {
UNIMPLEMENTED(); UNIMPLEMENTED();
} }
void VertexProgram::TEXLDL(Registers &r, Vector4f &dst, Vector4f &src, const Src&, Vector4f &offset) void VertexProgram::TEXLDL(Vector4f &dst, Vector4f &src, const Src&, Vector4f &offset)
{ {
UNIMPLEMENTED(); UNIMPLEMENTED();
} }
void VertexProgram::TEXELFETCH(Registers &r, Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2) void VertexProgram::TEXELFETCH(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2)
{ {
UNIMPLEMENTED(); UNIMPLEMENTED();
} }
void VertexProgram::TEXELFETCH(Registers &r, Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &offset) void VertexProgram::TEXELFETCH(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &offset)
{ {
UNIMPLEMENTED(); UNIMPLEMENTED();
} }
void VertexProgram::TEXGRAD(Registers &r, Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3) void VertexProgram::TEXGRAD(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3)
{ {
UNIMPLEMENTED(); UNIMPLEMENTED();
} }
void VertexProgram::TEXGRAD(Registers &r, Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3, Vector4f &offset) void VertexProgram::TEXGRAD(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3, Vector4f &offset)
{ {
UNIMPLEMENTED(); UNIMPLEMENTED();
} }
void VertexProgram::TEXSIZE(Registers &r, Vector4f &dst, Float4 &lod, const Src &src1) void VertexProgram::TEXSIZE(Vector4f &dst, Float4 &lod, const Src &src1)
{ {
Pointer<Byte> textureMipmap = r.data + OFFSET(DrawData, mipmap[16]) + src1.index * sizeof(Texture) + OFFSET(Texture, mipmap); Pointer<Byte> textureMipmap = r.data + OFFSET(DrawData, mipmap[16]) + src1.index * sizeof(Texture) + OFFSET(Texture, mipmap);
for(int i = 0; i < 4; ++i) for(int i = 0; i < 4; ++i)
...@@ -1535,16 +1535,16 @@ namespace sw ...@@ -1535,16 +1535,16 @@ namespace sw
} }
} }
void VertexProgram::sampleTexture(Registers &r, Vector4f &c, const Src &s, Float4 &u, Float4 &v, Float4 &w, Float4 &q) void VertexProgram::sampleTexture(Vector4f &c, const Src &s, Float4 &u, Float4 &v, Float4 &w, Float4 &q)
{ {
if(s.type == Shader::PARAMETER_SAMPLER && s.rel.type == Shader::PARAMETER_VOID) if(s.type == Shader::PARAMETER_SAMPLER && s.rel.type == Shader::PARAMETER_VOID)
{ {
Pointer<Byte> texture = r.data + OFFSET(DrawData,mipmap[16]) + s.index * sizeof(Texture); Pointer<Byte> texture = r.data + OFFSET(DrawData,mipmap[16]) + s.index * sizeof(Texture);
sampler[s.index]->sampleTexture(texture, c, u, v, w, q, r.a0, r.a0, false, false, true); sampler[s.index]->sampleTexture(texture, c, u, v, w, q, r.a0, r.a0, false, false, true);
} }
else else
{ {
Int index = As<Int>(Float(fetchRegisterF(r, s).x.x)); Int index = As<Int>(Float(fetchRegisterF(s).x.x));
for(int i = 0; i < 16; i++) for(int i = 0; i < 16; i++)
{ {
......
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