Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 1 | // SwiftShader Software Renderer
|
| 2 | //
|
| 3 | // Copyright(c) 2015 Google Inc.
|
| 4 | //
|
| 5 | // All rights reserved. No part of this software may be copied, distributed, transmitted,
|
| 6 | // transcribed, stored in a retrieval system, translated into any human or computer
|
| 7 | // language by any means, or disclosed to third parties without the explicit written
|
| 8 | // agreement of Google Inc. Without such an agreement, no rights or licenses, express
|
| 9 | // or implied, including but not limited to any patent rights, are granted to you.
|
| 10 | //
|
| 11 |
|
| 12 | #include "PixelProgram.hpp"
|
| 13 | #include "Primitive.hpp"
|
| 14 | #include "Renderer.hpp"
|
| 15 | #include "SamplerCore.hpp"
|
| 16 |
|
| 17 | namespace sw
|
| 18 | {
|
| 19 | extern bool postBlendSRGB;
|
| 20 | extern bool booleanFaceRegister; |
| 21 | extern bool halfIntegerCoordinates; // Pixel centers are not at integer coordinates |
| 22 | extern bool fullPixelPositionRegister;
|
| 23 |
|
| 24 | void PixelProgram::setBuiltins(PixelRoutine::Registers &rBase, Int &x, Int &y, Float4(&z)[4], Float4 &w)
|
| 25 | {
|
| 26 | Registers& r = *static_cast<Registers*>(&rBase);
|
| 27 |
|
| 28 | if(shader->getVersion() >= 0x0300) |
| 29 | { |
| 30 | if(shader->vPosDeclared) |
| 31 | { |
| 32 | if(!halfIntegerCoordinates) |
| 33 | { |
| 34 | r.vPos.x = Float4(Float(x)) + Float4(0, 1, 0, 1); |
| 35 | r.vPos.y = Float4(Float(y)) + Float4(0, 0, 1, 1); |
| 36 | } |
| 37 | else |
| 38 | { |
| 39 | r.vPos.x = Float4(Float(x)) + Float4(0.5f, 1.5f, 0.5f, 1.5f); |
| 40 | r.vPos.y = Float4(Float(y)) + Float4(0.5f, 0.5f, 1.5f, 1.5f); |
| 41 | } |
| 42 | |
| 43 | if(fullPixelPositionRegister) |
| 44 | { |
| 45 | r.vPos.z = z[0]; // FIXME: Centroid? |
| 46 | r.vPos.w = w; // FIXME: Centroid? |
| 47 | } |
| 48 | } |
| 49 | |
| 50 | if(shader->vFaceDeclared) |
| 51 | { |
| 52 | Float4 area = *Pointer<Float>(r.primitive + OFFSET(Primitive, area)); |
| 53 | Float4 face = booleanFaceRegister ? Float4(As<Float4>(CmpNLT(area, Float4(0.0f)))) : area; |
| 54 | |
| 55 | r.vFace.x = face; |
| 56 | r.vFace.y = face; |
| 57 | r.vFace.z = face; |
| 58 | r.vFace.w = face; |
| 59 | } |
| 60 | }
|
| 61 | }
|
| 62 |
|
| 63 | void PixelProgram::applyShader(PixelRoutine::Registers &rBase, Int cMask[4]) |
| 64 | {
|
| 65 | Registers& r = *static_cast<Registers*>(&rBase); |
| 66 | |
| 67 | r.enableIndex = 0; |
| 68 | r.stackIndex = 0; |
| 69 | |
| 70 | if(shader->containsLeaveInstruction()) |
| 71 | { |
| 72 | r.enableLeave = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF); |
| 73 | } |
| 74 | |
| 75 | bool out[4][4] = { false }; |
| 76 | |
| 77 | // Create all call site return blocks up front |
| 78 | for(size_t i = 0; i < shader->getLength(); i++) |
| 79 | { |
| 80 | const Shader::Instruction *instruction = shader->getInstruction(i); |
| 81 | Shader::Opcode opcode = instruction->opcode; |
| 82 | |
| 83 | if(opcode == Shader::OPCODE_CALL || opcode == Shader::OPCODE_CALLNZ) |
| 84 | { |
| 85 | const Dst &dst = instruction->dst; |
| 86 | |
| 87 | ASSERT(callRetBlock[dst.label].size() == dst.callSite); |
| 88 | callRetBlock[dst.label].push_back(Nucleus::createBasicBlock()); |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | for(size_t i = 0; i < shader->getLength(); i++) |
| 93 | { |
| 94 | const Shader::Instruction *instruction = shader->getInstruction(i); |
| 95 | Shader::Opcode opcode = instruction->opcode; |
| 96 | |
| 97 | if(opcode == Shader::OPCODE_DCL || opcode == Shader::OPCODE_DEF || opcode == Shader::OPCODE_DEFI || opcode == Shader::OPCODE_DEFB) |
| 98 | { |
| 99 | continue; |
| 100 | } |
| 101 | |
| 102 | const Dst &dst = instruction->dst; |
| 103 | const Src &src0 = instruction->src[0]; |
| 104 | const Src &src1 = instruction->src[1]; |
| 105 | const Src &src2 = instruction->src[2]; |
| 106 | const Src &src3 = instruction->src[3]; |
| 107 | |
| 108 | bool predicate = instruction->predicate; |
| 109 | Control control = instruction->control; |
| 110 | bool pp = dst.partialPrecision; |
| 111 | bool project = instruction->project; |
| 112 | bool bias = instruction->bias; |
| 113 | |
| 114 | Vector4f d; |
| 115 | Vector4f s0; |
| 116 | Vector4f s1; |
| 117 | Vector4f s2; |
| 118 | Vector4f s3; |
| 119 | |
| 120 | if(opcode == Shader::OPCODE_TEXKILL) // Takes destination as input |
| 121 | { |
| 122 | if(dst.type == Shader::PARAMETER_TEXTURE) |
| 123 | { |
| 124 | d.x = r.vf[2 + dst.index].x; |
| 125 | d.y = r.vf[2 + dst.index].y; |
| 126 | d.z = r.vf[2 + dst.index].z; |
| 127 | d.w = r.vf[2 + dst.index].w; |
| 128 | } |
| 129 | else |
| 130 | { |
| 131 | d = r.rf[dst.index]; |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | if(src0.type != Shader::PARAMETER_VOID) s0 = fetchRegisterF(r, src0); |
| 136 | if(src1.type != Shader::PARAMETER_VOID) s1 = fetchRegisterF(r, src1); |
| 137 | if(src2.type != Shader::PARAMETER_VOID) s2 = fetchRegisterF(r, src2); |
| 138 | if(src3.type != Shader::PARAMETER_VOID) s3 = fetchRegisterF(r, src3); |
| 139 | |
| 140 | switch(opcode) |
| 141 | { |
| 142 | case Shader::OPCODE_PS_2_0: break; |
| 143 | case Shader::OPCODE_PS_2_x: break; |
| 144 | case Shader::OPCODE_PS_3_0: break; |
| 145 | case Shader::OPCODE_DEF: break; |
| 146 | case Shader::OPCODE_DCL: break; |
| 147 | case Shader::OPCODE_NOP: break; |
| 148 | case Shader::OPCODE_MOV: mov(d, s0); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 149 | case Shader::OPCODE_NEG: neg(d, s0); break; |
| 150 | case Shader::OPCODE_INEG: ineg(d, s0); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 151 | case Shader::OPCODE_F2B: f2b(d, s0); break; |
| 152 | case Shader::OPCODE_B2F: b2f(d, s0); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 153 | case Shader::OPCODE_F2I: f2i(d, s0); break; |
| 154 | case Shader::OPCODE_I2F: i2f(d, s0); break; |
| 155 | case Shader::OPCODE_F2U: f2u(d, s0); break; |
| 156 | case Shader::OPCODE_U2F: u2f(d, s0); break; |
| 157 | case Shader::OPCODE_I2B: i2b(d, s0); break; |
| 158 | case Shader::OPCODE_B2I: b2i(d, s0); break; |
| 159 | case Shader::OPCODE_U2B: u2b(d, s0); break; |
| 160 | case Shader::OPCODE_B2U: b2u(d, s0); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 161 | case Shader::OPCODE_ADD: add(d, s0, s1); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 162 | case Shader::OPCODE_IADD: iadd(d, s0, s1); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 163 | case Shader::OPCODE_SUB: sub(d, s0, s1); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 164 | case Shader::OPCODE_ISUB: isub(d, s0, s1); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 165 | case Shader::OPCODE_MUL: mul(d, s0, s1); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 166 | case Shader::OPCODE_IMUL: imul(d, s0, s1); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 167 | case Shader::OPCODE_MAD: mad(d, s0, s1, s2); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 168 | case Shader::OPCODE_IMAD: imad(d, s0, s1, s2); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 169 | case Shader::OPCODE_DP1: dp1(d, s0, s1); break; |
| 170 | case Shader::OPCODE_DP2: dp2(d, s0, s1); break; |
| 171 | case Shader::OPCODE_DP2ADD: dp2add(d, s0, s1, s2); break; |
| 172 | case Shader::OPCODE_DP3: dp3(d, s0, s1); break; |
| 173 | case Shader::OPCODE_DP4: dp4(d, s0, s1); break; |
| 174 | case Shader::OPCODE_CMP0: cmp0(d, s0, s1, s2); break; |
| 175 | case Shader::OPCODE_ICMP: icmp(d, s0, s1, control); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 176 | case Shader::OPCODE_UCMP: ucmp(d, s0, s1, control); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 177 | case Shader::OPCODE_SELECT: select(d, s0, s1, s2); break; |
| 178 | case Shader::OPCODE_EXTRACT: extract(d.x, s0, s1.x); break; |
| 179 | case Shader::OPCODE_INSERT: insert(d, s0, s1.x, s2.x); break; |
| 180 | case Shader::OPCODE_FRC: frc(d, s0); break; |
| 181 | case Shader::OPCODE_TRUNC: trunc(d, s0); break; |
| 182 | case Shader::OPCODE_FLOOR: floor(d, s0); break; |
| 183 | case Shader::OPCODE_ROUND: round(d, s0); break; |
| 184 | case Shader::OPCODE_ROUNDEVEN: roundEven(d, s0); break; |
| 185 | case Shader::OPCODE_CEIL: ceil(d, s0); break; |
| 186 | case Shader::OPCODE_EXP2X: exp2x(d, s0, pp); break; |
| 187 | case Shader::OPCODE_EXP2: exp2(d, s0, pp); break; |
| 188 | case Shader::OPCODE_LOG2X: log2x(d, s0, pp); break; |
| 189 | case Shader::OPCODE_LOG2: log2(d, s0, pp); break; |
| 190 | case Shader::OPCODE_EXP: exp(d, s0, pp); break; |
| 191 | case Shader::OPCODE_LOG: log(d, s0, pp); break; |
| 192 | case Shader::OPCODE_RCPX: rcpx(d, s0, pp); break; |
| 193 | case Shader::OPCODE_DIV: div(d, s0, s1); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 194 | case Shader::OPCODE_IDIV: idiv(d, s0, s1); break; |
| 195 | case Shader::OPCODE_UDIV: udiv(d, s0, s1); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 196 | case Shader::OPCODE_MOD: mod(d, s0, s1); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 197 | case Shader::OPCODE_IMOD: imod(d, s0, s1); break; |
| 198 | case Shader::OPCODE_UMOD: umod(d, s0, s1); break; |
| 199 | case Shader::OPCODE_SHL: shl(d, s0, s1); break; |
| 200 | case Shader::OPCODE_ISHR: ishr(d, s0, s1); break; |
| 201 | case Shader::OPCODE_USHR: ushr(d, s0, s1); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 202 | case Shader::OPCODE_RSQX: rsqx(d, s0, pp); break; |
| 203 | case Shader::OPCODE_SQRT: sqrt(d, s0, pp); break; |
| 204 | case Shader::OPCODE_RSQ: rsq(d, s0, pp); break; |
| 205 | case Shader::OPCODE_LEN2: len2(d.x, s0, pp); break; |
| 206 | case Shader::OPCODE_LEN3: len3(d.x, s0, pp); break; |
| 207 | case Shader::OPCODE_LEN4: len4(d.x, s0, pp); break; |
| 208 | case Shader::OPCODE_DIST1: dist1(d.x, s0, s1, pp); break; |
| 209 | case Shader::OPCODE_DIST2: dist2(d.x, s0, s1, pp); break; |
| 210 | case Shader::OPCODE_DIST3: dist3(d.x, s0, s1, pp); break; |
| 211 | case Shader::OPCODE_DIST4: dist4(d.x, s0, s1, pp); break; |
| 212 | case Shader::OPCODE_MIN: min(d, s0, s1); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 213 | case Shader::OPCODE_IMIN: imin(d, s0, s1); break; |
| 214 | case Shader::OPCODE_UMIN: umin(d, s0, s1); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 215 | case Shader::OPCODE_MAX: max(d, s0, s1); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 216 | case Shader::OPCODE_IMAX: imax(d, s0, s1); break; |
| 217 | case Shader::OPCODE_UMAX: umax(d, s0, s1); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 218 | case Shader::OPCODE_LRP: lrp(d, s0, s1, s2); break; |
| 219 | case Shader::OPCODE_STEP: step(d, s0, s1); break; |
| 220 | case Shader::OPCODE_SMOOTH: smooth(d, s0, s1, s2); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 221 | case Shader::OPCODE_FLOATBITSTOINT: |
| 222 | case Shader::OPCODE_FLOATBITSTOUINT: |
| 223 | case Shader::OPCODE_INTBITSTOFLOAT: |
| 224 | case Shader::OPCODE_UINTBITSTOFLOAT: d = s0; break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 225 | case Shader::OPCODE_POWX: powx(d, s0, s1, pp); break; |
| 226 | case Shader::OPCODE_POW: pow(d, s0, s1, pp); break; |
| 227 | case Shader::OPCODE_SGN: sgn(d, s0); break; |
| 228 | case Shader::OPCODE_CRS: crs(d, s0, s1); break; |
| 229 | case Shader::OPCODE_FORWARD1: forward1(d, s0, s1, s2); break; |
| 230 | case Shader::OPCODE_FORWARD2: forward2(d, s0, s1, s2); break; |
| 231 | case Shader::OPCODE_FORWARD3: forward3(d, s0, s1, s2); break; |
| 232 | case Shader::OPCODE_FORWARD4: forward4(d, s0, s1, s2); break; |
| 233 | case Shader::OPCODE_REFLECT1: reflect1(d, s0, s1); break; |
| 234 | case Shader::OPCODE_REFLECT2: reflect2(d, s0, s1); break; |
| 235 | case Shader::OPCODE_REFLECT3: reflect3(d, s0, s1); break; |
| 236 | case Shader::OPCODE_REFLECT4: reflect4(d, s0, s1); break; |
| 237 | case Shader::OPCODE_REFRACT1: refract1(d, s0, s1, s2.x); break; |
| 238 | case Shader::OPCODE_REFRACT2: refract2(d, s0, s1, s2.x); break; |
| 239 | case Shader::OPCODE_REFRACT3: refract3(d, s0, s1, s2.x); break; |
| 240 | case Shader::OPCODE_REFRACT4: refract4(d, s0, s1, s2.x); break; |
| 241 | case Shader::OPCODE_NRM2: nrm2(d, s0, pp); break; |
| 242 | case Shader::OPCODE_NRM3: nrm3(d, s0, pp); break; |
| 243 | case Shader::OPCODE_NRM4: nrm4(d, s0, pp); break; |
| 244 | case Shader::OPCODE_ABS: abs(d, s0); break; |
| 245 | case Shader::OPCODE_SINCOS: sincos(d, s0, pp); break; |
| 246 | case Shader::OPCODE_COS: cos(d, s0, pp); break; |
| 247 | case Shader::OPCODE_SIN: sin(d, s0, pp); break; |
| 248 | case Shader::OPCODE_TAN: tan(d, s0, pp); break; |
| 249 | case Shader::OPCODE_ACOS: acos(d, s0, pp); break; |
| 250 | case Shader::OPCODE_ASIN: asin(d, s0, pp); break; |
| 251 | case Shader::OPCODE_ATAN: atan(d, s0, pp); break; |
| 252 | case Shader::OPCODE_ATAN2: atan2(d, s0, s1, pp); break; |
| 253 | case Shader::OPCODE_COSH: cosh(d, s0, pp); break; |
| 254 | case Shader::OPCODE_SINH: sinh(d, s0, pp); break; |
| 255 | case Shader::OPCODE_TANH: tanh(d, s0, pp); break; |
| 256 | case Shader::OPCODE_ACOSH: acosh(d, s0, pp); break; |
| 257 | case Shader::OPCODE_ASINH: asinh(d, s0, pp); break; |
| 258 | case Shader::OPCODE_ATANH: atanh(d, s0, pp); break; |
| 259 | case Shader::OPCODE_M4X4: M4X4(r, d, s0, src1); break; |
| 260 | case Shader::OPCODE_M4X3: M4X3(r, d, s0, src1); break; |
| 261 | case Shader::OPCODE_M3X4: M3X4(r, d, s0, src1); break; |
| 262 | case Shader::OPCODE_M3X3: M3X3(r, d, s0, src1); break; |
| 263 | case Shader::OPCODE_M3X2: M3X2(r, d, s0, src1); break; |
| 264 | case Shader::OPCODE_TEX: TEXLD(r, d, s0, src1, project, bias); break; |
| 265 | case Shader::OPCODE_TEXLDD: TEXLDD(r, d, s0, src1, s2, s3, project, bias); break; |
| 266 | case Shader::OPCODE_TEXLDL: TEXLDL(r, d, s0, src1, project, bias); break; |
| 267 | case Shader::OPCODE_TEXKILL: TEXKILL(cMask, d, dst.mask); break; |
| 268 | case Shader::OPCODE_DISCARD: DISCARD(r, cMask, instruction); break; |
| 269 | case Shader::OPCODE_DFDX: DFDX(d, s0); break; |
| 270 | case Shader::OPCODE_DFDY: DFDY(d, s0); break; |
| 271 | case Shader::OPCODE_FWIDTH: FWIDTH(d, s0); break; |
| 272 | case Shader::OPCODE_BREAK: BREAK(r); break; |
| 273 | case Shader::OPCODE_BREAKC: BREAKC(r, s0, s1, control); break; |
| 274 | case Shader::OPCODE_BREAKP: BREAKP(r, src0); break; |
| 275 | case Shader::OPCODE_CONTINUE: CONTINUE(r); break; |
| 276 | case Shader::OPCODE_TEST: TEST(); break; |
| 277 | case Shader::OPCODE_CALL: CALL(r, dst.label, dst.callSite); break; |
| 278 | case Shader::OPCODE_CALLNZ: CALLNZ(r, dst.label, dst.callSite, src0); break; |
| 279 | case Shader::OPCODE_ELSE: ELSE(r); break; |
| 280 | case Shader::OPCODE_ENDIF: ENDIF(r); break; |
| 281 | case Shader::OPCODE_ENDLOOP: ENDLOOP(r); break; |
| 282 | case Shader::OPCODE_ENDREP: ENDREP(r); break; |
| 283 | case Shader::OPCODE_ENDWHILE: ENDWHILE(r); break; |
| 284 | case Shader::OPCODE_IF: IF(r, src0); break; |
| 285 | case Shader::OPCODE_IFC: IFC(r, s0, s1, control); break; |
| 286 | case Shader::OPCODE_LABEL: LABEL(dst.index); break; |
| 287 | case Shader::OPCODE_LOOP: LOOP(r, src1); break; |
| 288 | case Shader::OPCODE_REP: REP(r, src0); break; |
| 289 | case Shader::OPCODE_WHILE: WHILE(r, src0); break; |
| 290 | case Shader::OPCODE_RET: RET(r); break; |
| 291 | case Shader::OPCODE_LEAVE: LEAVE(r); break; |
| 292 | case Shader::OPCODE_CMP: cmp(d, s0, s1, control); break; |
| 293 | case Shader::OPCODE_ALL: all(d.x, s0); break; |
| 294 | case Shader::OPCODE_ANY: any(d.x, s0); break; |
| 295 | case Shader::OPCODE_NOT: not(d, s0); break; |
Alexis Hetu | c4f2c29 | 2015-08-18 15:43:09 -0400 | [diff] [blame] | 296 | case Shader::OPCODE_OR: or(d, s0, s1); break; |
| 297 | case Shader::OPCODE_XOR: xor(d, s0, s1); break; |
| 298 | case Shader::OPCODE_AND: and(d, s0, s1); break; |
Alexis Hetu | 8d78cf7 | 2015-08-28 14:24:45 -0400 | [diff] [blame^] | 299 | case Shader::OPCODE_EQ: equal(d, s0, s1); break; |
| 300 | case Shader::OPCODE_NE: notEqual(d, s0, s1); break; |
Alexis Hetu | f2a8c37 | 2015-07-13 11:08:41 -0400 | [diff] [blame] | 301 | case Shader::OPCODE_END: break; |
| 302 | default: |
| 303 | ASSERT(false); |
| 304 | } |
| 305 | |
| 306 | if(dst.type != Shader::PARAMETER_VOID && dst.type != Shader::PARAMETER_LABEL && opcode != Shader::OPCODE_TEXKILL && opcode != Shader::OPCODE_NOP) |
| 307 | { |
| 308 | if(dst.integer) |
| 309 | { |
| 310 | switch(opcode) |
| 311 | { |
| 312 | case Shader::OPCODE_DIV: |
| 313 | if(dst.x) d.x = Trunc(d.x); |
| 314 | if(dst.y) d.y = Trunc(d.y); |
| 315 | if(dst.z) d.z = Trunc(d.z); |
| 316 | if(dst.w) d.w = Trunc(d.w); |
| 317 | break; |
| 318 | default: |
| 319 | break; // No truncation to integer required when arguments are integer |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | if(dst.saturate) |
| 324 | { |
| 325 | if(dst.x) d.x = Max(d.x, Float4(0.0f)); |
| 326 | if(dst.y) d.y = Max(d.y, Float4(0.0f)); |
| 327 | if(dst.z) d.z = Max(d.z, Float4(0.0f)); |
| 328 | if(dst.w) d.w = Max(d.w, Float4(0.0f)); |
| 329 | |
| 330 | if(dst.x) d.x = Min(d.x, Float4(1.0f)); |
| 331 | if(dst.y) d.y = Min(d.y, Float4(1.0f)); |
| 332 | if(dst.z) d.z = Min(d.z, Float4(1.0f)); |
| 333 | if(dst.w) d.w = Min(d.w, Float4(1.0f)); |
| 334 | } |
| 335 | |
| 336 | if(instruction->isPredicated()) |
| 337 | { |
| 338 | Vector4f pDst; // FIXME: Rename |
| 339 | |
| 340 | switch(dst.type) |
| 341 | { |
| 342 | case Shader::PARAMETER_TEMP: |
| 343 | if(dst.rel.type == Shader::PARAMETER_VOID) |
| 344 | { |
| 345 | if(dst.x) pDst.x = r.rf[dst.index].x; |
| 346 | if(dst.y) pDst.y = r.rf[dst.index].y; |
| 347 | if(dst.z) pDst.z = r.rf[dst.index].z; |
| 348 | if(dst.w) pDst.w = r.rf[dst.index].w; |
| 349 | } |
| 350 | else |
| 351 | { |
| 352 | Int a = relativeAddress(r, dst); |
| 353 | |
| 354 | if(dst.x) pDst.x = r.rf[dst.index + a].x; |
| 355 | if(dst.y) pDst.y = r.rf[dst.index + a].y; |
| 356 | if(dst.z) pDst.z = r.rf[dst.index + a].z; |
| 357 | if(dst.w) pDst.w = r.rf[dst.index + a].w; |
| 358 | } |
| 359 | break; |
| 360 | case Shader::PARAMETER_COLOROUT: |
| 361 | ASSERT(dst.rel.type == Shader::PARAMETER_VOID); |
| 362 | if(dst.x) pDst.x = r.oC[dst.index].x; |
| 363 | if(dst.y) pDst.y = r.oC[dst.index].y; |
| 364 | if(dst.z) pDst.z = r.oC[dst.index].z; |
| 365 | if(dst.w) pDst.w = r.oC[dst.index].w; |
| 366 | break; |
| 367 | case Shader::PARAMETER_PREDICATE: |
| 368 | if(dst.x) pDst.x = r.p0.x; |
| 369 | if(dst.y) pDst.y = r.p0.y; |
| 370 | if(dst.z) pDst.z = r.p0.z; |
| 371 | if(dst.w) pDst.w = r.p0.w; |
| 372 | break; |
| 373 | case Shader::PARAMETER_DEPTHOUT: |
| 374 | pDst.x = r.oDepth; |
| 375 | break; |
| 376 | default: |
| 377 | ASSERT(false); |
| 378 | } |
| 379 | |
| 380 | Int4 enable = enableMask(r, instruction); |
| 381 | |
| 382 | Int4 xEnable = enable; |
| 383 | Int4 yEnable = enable; |
| 384 | Int4 zEnable = enable; |
| 385 | Int4 wEnable = enable; |
| 386 | |
| 387 | if(predicate) |
| 388 | { |
| 389 | unsigned char pSwizzle = instruction->predicateSwizzle; |
| 390 | |
| 391 | Float4 xPredicate = r.p0[(pSwizzle >> 0) & 0x03]; |
| 392 | Float4 yPredicate = r.p0[(pSwizzle >> 2) & 0x03]; |
| 393 | Float4 zPredicate = r.p0[(pSwizzle >> 4) & 0x03]; |
| 394 | Float4 wPredicate = r.p0[(pSwizzle >> 6) & 0x03]; |
| 395 | |
| 396 | if(!instruction->predicateNot) |
| 397 | { |
| 398 | if(dst.x) xEnable = xEnable & As<Int4>(xPredicate); |
| 399 | if(dst.y) yEnable = yEnable & As<Int4>(yPredicate); |
| 400 | if(dst.z) zEnable = zEnable & As<Int4>(zPredicate); |
| 401 | if(dst.w) wEnable = wEnable & As<Int4>(wPredicate); |
| 402 | } |
| 403 | else |
| 404 | { |
| 405 | if(dst.x) xEnable = xEnable & ~As<Int4>(xPredicate); |
| 406 | if(dst.y) yEnable = yEnable & ~As<Int4>(yPredicate); |
| 407 | if(dst.z) zEnable = zEnable & ~As<Int4>(zPredicate); |
| 408 | if(dst.w) wEnable = wEnable & ~As<Int4>(wPredicate); |
| 409 | } |
| 410 | } |
| 411 | |
| 412 | if(dst.x) d.x = As<Float4>(As<Int4>(d.x) & xEnable); |
| 413 | if(dst.y) d.y = As<Float4>(As<Int4>(d.y) & yEnable); |
| 414 | if(dst.z) d.z = As<Float4>(As<Int4>(d.z) & zEnable); |
| 415 | if(dst.w) d.w = As<Float4>(As<Int4>(d.w) & wEnable); |
| 416 | |
| 417 | if(dst.x) d.x = As<Float4>(As<Int4>(d.x) | (As<Int4>(pDst.x) & ~xEnable)); |
| 418 | if(dst.y) d.y = As<Float4>(As<Int4>(d.y) | (As<Int4>(pDst.y) & ~yEnable)); |
| 419 | if(dst.z) d.z = As<Float4>(As<Int4>(d.z) | (As<Int4>(pDst.z) & ~zEnable)); |
| 420 | if(dst.w) d.w = As<Float4>(As<Int4>(d.w) | (As<Int4>(pDst.w) & ~wEnable)); |
| 421 | } |
| 422 | |
| 423 | switch(dst.type) |
| 424 | { |
| 425 | case Shader::PARAMETER_TEMP: |
| 426 | if(dst.rel.type == Shader::PARAMETER_VOID) |
| 427 | { |
| 428 | if(dst.x) r.rf[dst.index].x = d.x; |
| 429 | if(dst.y) r.rf[dst.index].y = d.y; |
| 430 | if(dst.z) r.rf[dst.index].z = d.z; |
| 431 | if(dst.w) r.rf[dst.index].w = d.w; |
| 432 | } |
| 433 | else |
| 434 | { |
| 435 | Int a = relativeAddress(r, dst); |
| 436 | |
| 437 | if(dst.x) r.rf[dst.index + a].x = d.x; |
| 438 | if(dst.y) r.rf[dst.index + a].y = d.y; |
| 439 | if(dst.z) r.rf[dst.index + a].z = d.z; |
| 440 | if(dst.w) r.rf[dst.index + a].w = d.w; |
| 441 | } |
| 442 | break; |
| 443 | case Shader::PARAMETER_COLOROUT: |
| 444 | ASSERT(dst.rel.type == Shader::PARAMETER_VOID); |
| 445 | if(dst.x) { r.oC[dst.index].x = d.x; out[dst.index][0] = true; } |
| 446 | if(dst.y) { r.oC[dst.index].y = d.y; out[dst.index][1] = true; } |
| 447 | if(dst.z) { r.oC[dst.index].z = d.z; out[dst.index][2] = true; } |
| 448 | if(dst.w) { r.oC[dst.index].w = d.w; out[dst.index][3] = true; } |
| 449 | break; |
| 450 | case Shader::PARAMETER_PREDICATE: |
| 451 | if(dst.x) r.p0.x = d.x; |
| 452 | if(dst.y) r.p0.y = d.y; |
| 453 | if(dst.z) r.p0.z = d.z; |
| 454 | if(dst.w) r.p0.w = d.w; |
| 455 | break; |
| 456 | case Shader::PARAMETER_DEPTHOUT: |
| 457 | r.oDepth = d.x; |
| 458 | break; |
| 459 | default: |
| 460 | ASSERT(false); |
| 461 | } |
| 462 | } |
| 463 | } |
| 464 | |
| 465 | if(currentLabel != -1) |
| 466 | { |
| 467 | Nucleus::setInsertBlock(returnBlock); |
| 468 | } |
| 469 | |
| 470 | for(int i = 0; i < 4; i++) |
| 471 | { |
| 472 | if(state.targetFormat[i] != FORMAT_NULL) |
| 473 | { |
| 474 | if(!out[i][0]) r.oC[i].x = Float4(0.0f); |
| 475 | if(!out[i][1]) r.oC[i].y = Float4(0.0f); |
| 476 | if(!out[i][2]) r.oC[i].z = Float4(0.0f); |
| 477 | if(!out[i][3]) r.oC[i].w = Float4(0.0f); |
| 478 | } |
| 479 | } |
| 480 | } |
| 481 |
|
| 482 | Bool PixelProgram::alphaTest(PixelRoutine::Registers &rBase, Int cMask[4])
|
| 483 | {
|
| 484 | Registers& r = *static_cast<Registers*>(&rBase);
|
| 485 |
|
| 486 | clampColor(r.oC);
|
| 487 |
|
| 488 | if(!state.alphaTestActive()) |
| 489 | { |
| 490 | return true; |
| 491 | } |
| 492 | |
| 493 | Int aMask; |
| 494 | |
| 495 | if(state.transparencyAntialiasing == TRANSPARENCY_NONE) |
| 496 | { |
| 497 | Short4 alpha = RoundShort4(r.oC[0].w * Float4(0x1000)); |
| 498 | |
| 499 | PixelRoutine::alphaTest(r, aMask, alpha); |
| 500 | |
| 501 | for(unsigned int q = 0; q < state.multiSample; q++) |
| 502 | { |
| 503 | cMask[q] &= aMask; |
| 504 | } |
| 505 | } |
| 506 | else if(state.transparencyAntialiasing == TRANSPARENCY_ALPHA_TO_COVERAGE) |
| 507 | { |
| 508 | alphaToCoverage(r, cMask, r.oC[0].w); |
| 509 | } |
| 510 | else ASSERT(false); |
| 511 | |
| 512 | Int pass = cMask[0]; |
| 513 | |
| 514 | for(unsigned int q = 1; q < state.multiSample; q++) |
| 515 | { |
| 516 | pass = pass | cMask[q]; |
| 517 | } |
| 518 | |
| 519 | return pass != 0x0;
|
| 520 | }
|
| 521 |
|
| 522 | void PixelProgram::rasterOperation(PixelRoutine::Registers &rBase, Float4 &fog, Pointer<Byte> cBuffer[4], Int &x, Int sMask[4], Int zMask[4], Int cMask[4])
|
| 523 | {
|
| 524 | Registers& r = *static_cast<Registers*>(&rBase);
|
| 525 |
|
| 526 | for(int index = 0; index < 4; index++) |
| 527 | { |
| 528 | if(!state.colorWriteActive(index)) |
| 529 | { |
| 530 | continue; |
| 531 | } |
| 532 | |
| 533 | if(!postBlendSRGB && state.writeSRGB) |
| 534 | { |
| 535 | r.oC[index].x = linearToSRGB(r.oC[index].x); |
| 536 | r.oC[index].y = linearToSRGB(r.oC[index].y); |
| 537 | r.oC[index].z = linearToSRGB(r.oC[index].z); |
| 538 | } |
| 539 | |
| 540 | if(index == 0) |
| 541 | { |
| 542 | fogBlend(r, r.oC[index], fog, r.z[0], r.rhw); |
| 543 | } |
| 544 | |
| 545 | switch(state.targetFormat[index]) |
| 546 | { |
| 547 | case FORMAT_R5G6B5: |
| 548 | case FORMAT_X8R8G8B8: |
| 549 | case FORMAT_X8B8G8R8: |
| 550 | case FORMAT_A8R8G8B8: |
| 551 | case FORMAT_A8B8G8R8: |
| 552 | case FORMAT_A8: |
| 553 | case FORMAT_G16R16: |
| 554 | case FORMAT_A16B16G16R16: |
| 555 | for(unsigned int q = 0; q < state.multiSample; q++) |
| 556 | { |
| 557 | Pointer<Byte> buffer = cBuffer[index] + q * *Pointer<Int>(r.data + OFFSET(DrawData, colorSliceB[index])); |
| 558 | Vector4s color; |
| 559 | |
| 560 | color.x = convertFixed16(r.oC[index].x, false); |
| 561 | color.y = convertFixed16(r.oC[index].y, false); |
| 562 | color.z = convertFixed16(r.oC[index].z, false); |
| 563 | color.w = convertFixed16(r.oC[index].w, false); |
| 564 | |
| 565 | if(state.multiSampleMask & (1 << q)) |
| 566 | { |
| 567 | alphaBlend(r, index, buffer, color, x); |
| 568 | logicOperation(r, index, buffer, color, x); |
| 569 | writeColor(r, index, buffer, x, color, sMask[q], zMask[q], cMask[q]); |
| 570 | } |
| 571 | } |
| 572 | break; |
| 573 | case FORMAT_R32F: |
| 574 | case FORMAT_G32R32F: |
| 575 | case FORMAT_A32B32G32R32F: |
| 576 | for(unsigned int q = 0; q < state.multiSample; q++) |
| 577 | { |
| 578 | Pointer<Byte> buffer = cBuffer[index] + q * *Pointer<Int>(r.data + OFFSET(DrawData, colorSliceB[index])); |
| 579 | Vector4f color = r.oC[index]; |
| 580 | |
| 581 | if(state.multiSampleMask & (1 << q)) |
| 582 | { |
| 583 | alphaBlend(r, index, buffer, color, x); |
| 584 | writeColor(r, index, buffer, x, color, sMask[q], zMask[q], cMask[q]); |
| 585 | } |
| 586 | } |
| 587 | break; |
| 588 | default: |
| 589 | ASSERT(false); |
| 590 | } |
| 591 | } |
| 592 | }
|
| 593 | |
| 594 | void PixelProgram::sampleTexture(Registers &r, Vector4f &c, const Src &sampler, Float4 &u, Float4 &v, Float4 &w, Float4 &q, Vector4f &dsx, Vector4f &dsy, bool project, bool bias, bool gradients, bool lodProvided) |
| 595 | { |
| 596 | if(sampler.type == Shader::PARAMETER_SAMPLER && sampler.rel.type == Shader::PARAMETER_VOID) |
| 597 | { |
| 598 | sampleTexture(r, c, sampler.index, u, v, w, q, dsx, dsy, project, bias, gradients, lodProvided); |
| 599 | } |
| 600 | else |
| 601 | { |
| 602 | Int index = As<Int>(Float(fetchRegisterF(r, sampler).x.x)); |
| 603 | |
| 604 | for(int i = 0; i < TEXTURE_IMAGE_UNITS; i++) |
| 605 | { |
| 606 | if(shader->usesSampler(i)) |
| 607 | { |
| 608 | If(index == i) |
| 609 | { |
| 610 | sampleTexture(r, c, i, u, v, w, q, dsx, dsy, project, bias, gradients, lodProvided); |
| 611 | // FIXME: When the sampler states are the same, we could use one sampler and just index the texture |
| 612 | } |
| 613 | } |
| 614 | } |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | void PixelProgram::sampleTexture(Registers &r, Vector4f &c, int stage, Float4 &u, Float4 &v, Float4 &w, Float4 &q, Vector4f &dsx, Vector4f &dsy, bool project, bool bias, bool gradients, bool lodProvided) |
| 619 | { |
| 620 | #if PERF_PROFILE |
| 621 | Long texTime = Ticks(); |
| 622 | #endif |
| 623 | |
| 624 | Pointer<Byte> texture = r.data + OFFSET(DrawData, mipmap) + stage * sizeof(Texture); |
| 625 | |
| 626 | if(!project) |
| 627 | { |
| 628 | sampler[stage]->sampleTexture(texture, c, u, v, w, q, dsx, dsy, bias, gradients, lodProvided); |
| 629 | } |
| 630 | else |
| 631 | { |
| 632 | Float4 rq = reciprocal(q); |
| 633 | |
| 634 | Float4 u_q = u * rq; |
| 635 | Float4 v_q = v * rq; |
| 636 | Float4 w_q = w * rq; |
| 637 | |
| 638 | sampler[stage]->sampleTexture(texture, c, u_q, v_q, w_q, q, dsx, dsy, bias, gradients, lodProvided); |
| 639 | } |
| 640 | |
| 641 | #if PERF_PROFILE |
| 642 | r.cycles[PERF_TEX] += Ticks() - texTime; |
| 643 | #endif |
| 644 | } |
| 645 | |
| 646 | void PixelProgram::clampColor(Vector4f oC[4]) |
| 647 | { |
| 648 | for(int index = 0; index < 4; index++) |
| 649 | { |
| 650 | if(!state.colorWriteActive(index) && !(index == 0 && state.alphaTestActive())) |
| 651 | { |
| 652 | continue; |
| 653 | } |
| 654 | |
| 655 | switch(state.targetFormat[index]) |
| 656 | { |
| 657 | case FORMAT_NULL: |
| 658 | break; |
| 659 | case FORMAT_R5G6B5: |
| 660 | case FORMAT_A8R8G8B8: |
| 661 | case FORMAT_A8B8G8R8: |
| 662 | case FORMAT_X8R8G8B8: |
| 663 | case FORMAT_X8B8G8R8: |
| 664 | case FORMAT_A8: |
| 665 | case FORMAT_G16R16: |
| 666 | case FORMAT_A16B16G16R16: |
| 667 | oC[index].x = Max(oC[index].x, Float4(0.0f)); oC[index].x = Min(oC[index].x, Float4(1.0f)); |
| 668 | oC[index].y = Max(oC[index].y, Float4(0.0f)); oC[index].y = Min(oC[index].y, Float4(1.0f)); |
| 669 | oC[index].z = Max(oC[index].z, Float4(0.0f)); oC[index].z = Min(oC[index].z, Float4(1.0f)); |
| 670 | oC[index].w = Max(oC[index].w, Float4(0.0f)); oC[index].w = Min(oC[index].w, Float4(1.0f)); |
| 671 | break; |
| 672 | case FORMAT_R32F: |
| 673 | case FORMAT_G32R32F: |
| 674 | case FORMAT_A32B32G32R32F: |
| 675 | break; |
| 676 | default: |
| 677 | ASSERT(false); |
| 678 | } |
| 679 | } |
| 680 | } |
| 681 | |
| 682 | Int4 PixelProgram::enableMask(Registers &r, const Shader::Instruction *instruction) |
| 683 | { |
| 684 | Int4 enable = instruction->analysisBranch ? Int4(r.enableStack[r.enableIndex]) : Int4(0xFFFFFFFF); |
| 685 | |
| 686 | if(!whileTest) |
| 687 | { |
| 688 | if(shader->containsBreakInstruction() && instruction->analysisBreak) |
| 689 | { |
| 690 | enable &= r.enableBreak; |
| 691 | } |
| 692 | |
| 693 | if(shader->containsContinueInstruction() && instruction->analysisContinue) |
| 694 | { |
| 695 | enable &= r.enableContinue; |
| 696 | } |
| 697 | |
| 698 | if(shader->containsLeaveInstruction() && instruction->analysisLeave) |
| 699 | { |
| 700 | enable &= r.enableLeave; |
| 701 | } |
| 702 | } |
| 703 | |
| 704 | return enable; |
| 705 | } |
| 706 | |
| 707 | Vector4f PixelProgram::fetchRegisterF(Registers &r, const Src &src, int offset) |
| 708 | { |
| 709 | Vector4f reg; |
| 710 | int i = src.index + offset; |
| 711 | |
| 712 | switch(src.type) |
| 713 | { |
| 714 | case Shader::PARAMETER_TEMP: |
| 715 | if(src.rel.type == Shader::PARAMETER_VOID) |
| 716 | { |
| 717 | reg = r.rf[i]; |
| 718 | } |
| 719 | else |
| 720 | { |
| 721 | Int a = relativeAddress(r, src); |
| 722 | |
| 723 | reg = r.rf[i + a]; |
| 724 | } |
| 725 | break; |
| 726 | case Shader::PARAMETER_INPUT: |
| 727 | { |
| 728 | if(src.rel.type == Shader::PARAMETER_VOID) // Not relative |
| 729 | { |
| 730 | reg = r.vf[i]; |
| 731 | } |
| 732 | else if(src.rel.type == Shader::PARAMETER_LOOP) |
| 733 | { |
| 734 | Int aL = r.aL[r.loopDepth]; |
| 735 | |
| 736 | reg = r.vf[i + aL]; |
| 737 | } |
| 738 | else |
| 739 | { |
| 740 | Int a = relativeAddress(r, src); |
| 741 | |
| 742 | reg = r.vf[i + a]; |
| 743 | } |
| 744 | } |
| 745 | break; |
| 746 | case Shader::PARAMETER_CONST: |
| 747 | reg = readConstant(r, src, offset); |
| 748 | break; |
| 749 | case Shader::PARAMETER_TEXTURE: |
| 750 | reg = r.vf[2 + i]; |
| 751 | break; |
| 752 | case Shader::PARAMETER_MISCTYPE: |
| 753 | if(src.index == 0) reg = r.vPos; |
| 754 | if(src.index == 1) reg = r.vFace; |
| 755 | break; |
| 756 | case Shader::PARAMETER_SAMPLER: |
| 757 | if(src.rel.type == Shader::PARAMETER_VOID) |
| 758 | { |
| 759 | reg.x = As<Float4>(Int4(i)); |
| 760 | } |
| 761 | else if(src.rel.type == Shader::PARAMETER_TEMP) |
| 762 | { |
| 763 | reg.x = As<Float4>(Int4(i) + RoundInt(r.rf[src.rel.index].x)); |
| 764 | } |
| 765 | return reg; |
| 766 | case Shader::PARAMETER_PREDICATE: return reg; // Dummy |
| 767 | case Shader::PARAMETER_VOID: return reg; // Dummy |
| 768 | case Shader::PARAMETER_FLOAT4LITERAL: |
| 769 | reg.x = Float4(src.value[0]); |
| 770 | reg.y = Float4(src.value[1]); |
| 771 | reg.z = Float4(src.value[2]); |
| 772 | reg.w = Float4(src.value[3]); |
| 773 | break; |
| 774 | case Shader::PARAMETER_CONSTINT: return reg; // Dummy |
| 775 | case Shader::PARAMETER_CONSTBOOL: return reg; // Dummy |
| 776 | case Shader::PARAMETER_LOOP: return reg; // Dummy |
| 777 | case Shader::PARAMETER_COLOROUT: |
| 778 | reg = r.oC[i]; |
| 779 | break; |
| 780 | case Shader::PARAMETER_DEPTHOUT: |
| 781 | reg.x = r.oDepth; |
| 782 | break; |
| 783 | default: |
| 784 | ASSERT(false); |
| 785 | } |
| 786 | |
| 787 | const Float4 &x = reg[(src.swizzle >> 0) & 0x3]; |
| 788 | const Float4 &y = reg[(src.swizzle >> 2) & 0x3]; |
| 789 | const Float4 &z = reg[(src.swizzle >> 4) & 0x3]; |
| 790 | const Float4 &w = reg[(src.swizzle >> 6) & 0x3]; |
| 791 | |
| 792 | Vector4f mod; |
| 793 | |
| 794 | switch(src.modifier) |
| 795 | { |
| 796 | case Shader::MODIFIER_NONE: |
| 797 | mod.x = x; |
| 798 | mod.y = y; |
| 799 | mod.z = z; |
| 800 | mod.w = w; |
| 801 | break; |
| 802 | case Shader::MODIFIER_NEGATE: |
| 803 | mod.x = -x; |
| 804 | mod.y = -y; |
| 805 | mod.z = -z; |
| 806 | mod.w = -w; |
| 807 | break; |
| 808 | case Shader::MODIFIER_ABS: |
| 809 | mod.x = Abs(x); |
| 810 | mod.y = Abs(y); |
| 811 | mod.z = Abs(z); |
| 812 | mod.w = Abs(w); |
| 813 | break; |
| 814 | case Shader::MODIFIER_ABS_NEGATE: |
| 815 | mod.x = -Abs(x); |
| 816 | mod.y = -Abs(y); |
| 817 | mod.z = -Abs(z); |
| 818 | mod.w = -Abs(w); |
| 819 | break; |
| 820 | case Shader::MODIFIER_NOT: |
| 821 | mod.x = As<Float4>(As<Int4>(x) ^ Int4(0xFFFFFFFF)); |
| 822 | mod.y = As<Float4>(As<Int4>(y) ^ Int4(0xFFFFFFFF)); |
| 823 | mod.z = As<Float4>(As<Int4>(z) ^ Int4(0xFFFFFFFF)); |
| 824 | mod.w = As<Float4>(As<Int4>(w) ^ Int4(0xFFFFFFFF)); |
| 825 | break; |
| 826 | default: |
| 827 | ASSERT(false); |
| 828 | } |
| 829 | |
| 830 | return mod; |
| 831 | } |
| 832 | |
| 833 | Vector4f PixelProgram::readConstant(Registers &r, const Src &src, int offset) |
| 834 | { |
| 835 | Vector4f c; |
| 836 | |
| 837 | int i = src.index + offset; |
| 838 | |
| 839 | if(src.rel.type == Shader::PARAMETER_VOID) // Not relative |
| 840 | { |
| 841 | c.x = c.y = c.z = c.w = *Pointer<Float4>(r.data + OFFSET(DrawData, ps.c[i])); |
| 842 | |
| 843 | c.x = c.x.xxxx; |
| 844 | c.y = c.y.yyyy; |
| 845 | c.z = c.z.zzzz; |
| 846 | c.w = c.w.wwww; |
| 847 | |
| 848 | if(shader->containsDefineInstruction()) // Constant may be known at compile time |
| 849 | { |
| 850 | for(size_t j = 0; j < shader->getLength(); j++) |
| 851 | { |
| 852 | const Shader::Instruction &instruction = *shader->getInstruction(j); |
| 853 | |
| 854 | if(instruction.opcode == Shader::OPCODE_DEF) |
| 855 | { |
| 856 | if(instruction.dst.index == i) |
| 857 | { |
| 858 | c.x = Float4(instruction.src[0].value[0]); |
| 859 | c.y = Float4(instruction.src[0].value[1]); |
| 860 | c.z = Float4(instruction.src[0].value[2]); |
| 861 | c.w = Float4(instruction.src[0].value[3]); |
| 862 | |
| 863 | break; |
| 864 | } |
| 865 | } |
| 866 | } |
| 867 | } |
| 868 | } |
| 869 | else if(src.rel.type == Shader::PARAMETER_LOOP) |
| 870 | { |
| 871 | Int loopCounter = r.aL[r.loopDepth]; |
| 872 | |
| 873 | c.x = c.y = c.z = c.w = *Pointer<Float4>(r.data + OFFSET(DrawData, ps.c[i]) + loopCounter * 16); |
| 874 | |
| 875 | c.x = c.x.xxxx; |
| 876 | c.y = c.y.yyyy; |
| 877 | c.z = c.z.zzzz; |
| 878 | c.w = c.w.wwww; |
| 879 | } |
| 880 | else |
| 881 | { |
| 882 | Int a = relativeAddress(r, src); |
| 883 | |
| 884 | c.x = c.y = c.z = c.w = *Pointer<Float4>(r.data + OFFSET(DrawData, ps.c[i]) + a * 16); |
| 885 | |
| 886 | c.x = c.x.xxxx; |
| 887 | c.y = c.y.yyyy; |
| 888 | c.z = c.z.zzzz; |
| 889 | c.w = c.w.wwww; |
| 890 | } |
| 891 | |
| 892 | return c; |
| 893 | } |
| 894 | |
| 895 | Int PixelProgram::relativeAddress(Registers &r, const Shader::Parameter &var) |
| 896 | { |
| 897 | ASSERT(var.rel.deterministic); |
| 898 | |
| 899 | if(var.rel.type == Shader::PARAMETER_TEMP) |
| 900 | { |
| 901 | return RoundInt(Extract(r.rf[var.rel.index].x, 0)) * var.rel.scale; |
| 902 | } |
| 903 | else if(var.rel.type == Shader::PARAMETER_INPUT) |
| 904 | { |
| 905 | return RoundInt(Extract(r.vf[var.rel.index].x, 0)) * var.rel.scale; |
| 906 | } |
| 907 | else if(var.rel.type == Shader::PARAMETER_OUTPUT) |
| 908 | { |
| 909 | return RoundInt(Extract(r.oC[var.rel.index].x, 0)) * var.rel.scale; |
| 910 | } |
| 911 | else if(var.rel.type == Shader::PARAMETER_CONST) |
| 912 | { |
| 913 | RValue<Float4> c = *Pointer<Float4>(r.data + OFFSET(DrawData, ps.c[var.rel.index])); |
| 914 | |
| 915 | return RoundInt(Extract(c, 0)) * var.rel.scale; |
| 916 | } |
| 917 | else ASSERT(false); |
| 918 | |
| 919 | return 0; |
| 920 | } |
| 921 | |
| 922 | Float4 PixelProgram::linearToSRGB(const Float4 &x) // Approximates x^(1.0/2.2) |
| 923 | { |
| 924 | Float4 sqrtx = Rcp_pp(RcpSqrt_pp(x)); |
| 925 | Float4 sRGB = sqrtx * Float4(1.14f) - x * Float4(0.14f); |
| 926 | |
| 927 | return Min(Max(sRGB, Float4(0.0f)), Float4(1.0f)); |
| 928 | } |
| 929 | |
| 930 | void PixelProgram::M3X2(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1) |
| 931 | { |
| 932 | Vector4f row0 = fetchRegisterF(r, src1, 0); |
| 933 | Vector4f row1 = fetchRegisterF(r, src1, 1); |
| 934 | |
| 935 | dst.x = dot3(src0, row0); |
| 936 | dst.y = dot3(src0, row1); |
| 937 | } |
| 938 | |
| 939 | void PixelProgram::M3X3(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1) |
| 940 | { |
| 941 | Vector4f row0 = fetchRegisterF(r, src1, 0); |
| 942 | Vector4f row1 = fetchRegisterF(r, src1, 1); |
| 943 | Vector4f row2 = fetchRegisterF(r, src1, 2); |
| 944 | |
| 945 | dst.x = dot3(src0, row0); |
| 946 | dst.y = dot3(src0, row1); |
| 947 | dst.z = dot3(src0, row2); |
| 948 | } |
| 949 | |
| 950 | void PixelProgram::M3X4(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1) |
| 951 | { |
| 952 | Vector4f row0 = fetchRegisterF(r, src1, 0); |
| 953 | Vector4f row1 = fetchRegisterF(r, src1, 1); |
| 954 | Vector4f row2 = fetchRegisterF(r, src1, 2); |
| 955 | Vector4f row3 = fetchRegisterF(r, src1, 3); |
| 956 | |
| 957 | dst.x = dot3(src0, row0); |
| 958 | dst.y = dot3(src0, row1); |
| 959 | dst.z = dot3(src0, row2); |
| 960 | dst.w = dot3(src0, row3); |
| 961 | } |
| 962 | |
| 963 | void PixelProgram::M4X3(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1) |
| 964 | { |
| 965 | Vector4f row0 = fetchRegisterF(r, src1, 0); |
| 966 | Vector4f row1 = fetchRegisterF(r, src1, 1); |
| 967 | Vector4f row2 = fetchRegisterF(r, src1, 2); |
| 968 | |
| 969 | dst.x = dot4(src0, row0); |
| 970 | dst.y = dot4(src0, row1); |
| 971 | dst.z = dot4(src0, row2); |
| 972 | } |
| 973 | |
| 974 | void PixelProgram::M4X4(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1) |
| 975 | { |
| 976 | Vector4f row0 = fetchRegisterF(r, src1, 0); |
| 977 | Vector4f row1 = fetchRegisterF(r, src1, 1); |
| 978 | Vector4f row2 = fetchRegisterF(r, src1, 2); |
| 979 | Vector4f row3 = fetchRegisterF(r, src1, 3); |
| 980 | |
| 981 | dst.x = dot4(src0, row0); |
| 982 | dst.y = dot4(src0, row1); |
| 983 | dst.z = dot4(src0, row2); |
| 984 | dst.w = dot4(src0, row3); |
| 985 | } |
| 986 | |
| 987 | void PixelProgram::TEXLD(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1, bool project, bool bias) |
| 988 | { |
| 989 | Vector4f tmp; |
| 990 | sampleTexture(r, tmp, src1, src0.x, src0.y, src0.z, src0.w, src0, src0, project, bias); |
| 991 | |
| 992 | dst.x = tmp[(src1.swizzle >> 0) & 0x3]; |
| 993 | dst.y = tmp[(src1.swizzle >> 2) & 0x3]; |
| 994 | dst.z = tmp[(src1.swizzle >> 4) & 0x3]; |
| 995 | dst.w = tmp[(src1.swizzle >> 6) & 0x3]; |
| 996 | } |
| 997 | |
| 998 | void PixelProgram::TEXLDD(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1, Vector4f &src2, Vector4f &src3, bool project, bool bias) |
| 999 | { |
| 1000 | Vector4f tmp; |
| 1001 | sampleTexture(r, tmp, src1, src0.x, src0.y, src0.z, src0.w, src2, src3, project, bias, true); |
| 1002 | |
| 1003 | dst.x = tmp[(src1.swizzle >> 0) & 0x3]; |
| 1004 | dst.y = tmp[(src1.swizzle >> 2) & 0x3]; |
| 1005 | dst.z = tmp[(src1.swizzle >> 4) & 0x3]; |
| 1006 | dst.w = tmp[(src1.swizzle >> 6) & 0x3]; |
| 1007 | } |
| 1008 | |
| 1009 | void PixelProgram::TEXLDL(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1, bool project, bool bias) |
| 1010 | { |
| 1011 | Vector4f tmp; |
| 1012 | sampleTexture(r, tmp, src1, src0.x, src0.y, src0.z, src0.w, src0, src0, project, bias, false, true); |
| 1013 | |
| 1014 | dst.x = tmp[(src1.swizzle >> 0) & 0x3]; |
| 1015 | dst.y = tmp[(src1.swizzle >> 2) & 0x3]; |
| 1016 | dst.z = tmp[(src1.swizzle >> 4) & 0x3]; |
| 1017 | dst.w = tmp[(src1.swizzle >> 6) & 0x3]; |
| 1018 | } |
| 1019 | |
| 1020 | void PixelProgram::TEXKILL(Int cMask[4], Vector4f &src, unsigned char mask) |
| 1021 | { |
| 1022 | Int kill = -1; |
| 1023 | |
| 1024 | if(mask & 0x1) kill &= SignMask(CmpNLT(src.x, Float4(0.0f))); |
| 1025 | if(mask & 0x2) kill &= SignMask(CmpNLT(src.y, Float4(0.0f))); |
| 1026 | if(mask & 0x4) kill &= SignMask(CmpNLT(src.z, Float4(0.0f))); |
| 1027 | if(mask & 0x8) kill &= SignMask(CmpNLT(src.w, Float4(0.0f))); |
| 1028 | |
| 1029 | // FIXME: Dynamic branching affects TEXKILL? |
| 1030 | // if(shader->containsDynamicBranching()) |
| 1031 | // { |
| 1032 | // kill = ~SignMask(enableMask(r)); |
| 1033 | // } |
| 1034 | |
| 1035 | for(unsigned int q = 0; q < state.multiSample; q++) |
| 1036 | { |
| 1037 | cMask[q] &= kill; |
| 1038 | } |
| 1039 | |
| 1040 | // FIXME: Branch to end of shader if all killed? |
| 1041 | } |
| 1042 | |
| 1043 | void PixelProgram::DISCARD(Registers &r, Int cMask[4], const Shader::Instruction *instruction) |
| 1044 | { |
| 1045 | Int kill = 0; |
| 1046 | |
| 1047 | if(shader->containsDynamicBranching()) |
| 1048 | { |
| 1049 | kill = ~SignMask(enableMask(r, instruction)); |
| 1050 | } |
| 1051 | |
| 1052 | for(unsigned int q = 0; q < state.multiSample; q++) |
| 1053 | { |
| 1054 | cMask[q] &= kill; |
| 1055 | } |
| 1056 | |
| 1057 | // FIXME: Branch to end of shader if all killed? |
| 1058 | } |
| 1059 | |
| 1060 | void PixelProgram::DFDX(Vector4f &dst, Vector4f &src) |
| 1061 | { |
| 1062 | dst.x = src.x.yyww - src.x.xxzz; |
| 1063 | dst.y = src.y.yyww - src.y.xxzz; |
| 1064 | dst.z = src.z.yyww - src.z.xxzz; |
| 1065 | dst.w = src.w.yyww - src.w.xxzz; |
| 1066 | } |
| 1067 | |
| 1068 | void PixelProgram::DFDY(Vector4f &dst, Vector4f &src) |
| 1069 | { |
| 1070 | dst.x = src.x.zwzw - src.x.xyxy; |
| 1071 | dst.y = src.y.zwzw - src.y.xyxy; |
| 1072 | dst.z = src.z.zwzw - src.z.xyxy; |
| 1073 | dst.w = src.w.zwzw - src.w.xyxy; |
| 1074 | } |
| 1075 | |
| 1076 | void PixelProgram::FWIDTH(Vector4f &dst, Vector4f &src) |
| 1077 | { |
| 1078 | // abs(dFdx(src)) + abs(dFdy(src)); |
| 1079 | dst.x = Abs(src.x.yyww - src.x.xxzz) + Abs(src.x.zwzw - src.x.xyxy); |
| 1080 | dst.y = Abs(src.y.yyww - src.y.xxzz) + Abs(src.y.zwzw - src.y.xyxy); |
| 1081 | dst.z = Abs(src.z.yyww - src.z.xxzz) + Abs(src.z.zwzw - src.z.xyxy); |
| 1082 | dst.w = Abs(src.w.yyww - src.w.xxzz) + Abs(src.w.zwzw - src.w.xyxy); |
| 1083 | } |
| 1084 | |
| 1085 | void PixelProgram::BREAK(Registers &r) |
| 1086 | { |
| 1087 | llvm::BasicBlock *deadBlock = Nucleus::createBasicBlock(); |
| 1088 | llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1]; |
| 1089 | |
| 1090 | if(breakDepth == 0) |
| 1091 | { |
| 1092 | r.enableIndex = r.enableIndex - breakDepth; |
| 1093 | Nucleus::createBr(endBlock); |
| 1094 | } |
| 1095 | else |
| 1096 | { |
| 1097 | r.enableBreak = r.enableBreak & ~r.enableStack[r.enableIndex]; |
| 1098 | Bool allBreak = SignMask(r.enableBreak) == 0x0; |
| 1099 | |
| 1100 | r.enableIndex = r.enableIndex - breakDepth; |
| 1101 | branch(allBreak, endBlock, deadBlock); |
| 1102 | } |
| 1103 | |
| 1104 | Nucleus::setInsertBlock(deadBlock); |
| 1105 | r.enableIndex = r.enableIndex + breakDepth; |
| 1106 | } |
| 1107 | |
| 1108 | void PixelProgram::BREAKC(Registers &r, Vector4f &src0, Vector4f &src1, Control control) |
| 1109 | { |
| 1110 | Int4 condition; |
| 1111 | |
| 1112 | switch(control) |
| 1113 | { |
| 1114 | case Shader::CONTROL_GT: condition = CmpNLE(src0.x, src1.x); break; |
| 1115 | case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x); break; |
| 1116 | case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x); break; |
| 1117 | case Shader::CONTROL_LT: condition = CmpLT(src0.x, src1.x); break; |
| 1118 | case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x); break; |
| 1119 | case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x); break; |
| 1120 | default: |
| 1121 | ASSERT(false); |
| 1122 | } |
| 1123 | |
| 1124 | BREAK(r, condition); |
| 1125 | } |
| 1126 | |
| 1127 | void PixelProgram::BREAKP(Registers &r, const Src &predicateRegister) // FIXME: Factor out parts common with BREAKC |
| 1128 | { |
| 1129 | Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]); |
| 1130 | |
| 1131 | if(predicateRegister.modifier == Shader::MODIFIER_NOT) |
| 1132 | { |
| 1133 | condition = ~condition; |
| 1134 | } |
| 1135 | |
| 1136 | BREAK(r, condition); |
| 1137 | } |
| 1138 | |
| 1139 | void PixelProgram::BREAK(Registers &r, Int4 &condition) |
| 1140 | { |
| 1141 | condition &= r.enableStack[r.enableIndex]; |
| 1142 | |
| 1143 | llvm::BasicBlock *continueBlock = Nucleus::createBasicBlock(); |
| 1144 | llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1]; |
| 1145 | |
| 1146 | r.enableBreak = r.enableBreak & ~condition; |
| 1147 | Bool allBreak = SignMask(r.enableBreak) == 0x0; |
| 1148 | |
| 1149 | r.enableIndex = r.enableIndex - breakDepth; |
| 1150 | branch(allBreak, endBlock, continueBlock); |
| 1151 | |
| 1152 | Nucleus::setInsertBlock(continueBlock); |
| 1153 | r.enableIndex = r.enableIndex + breakDepth; |
| 1154 | } |
| 1155 | |
| 1156 | void PixelProgram::CONTINUE(Registers &r) |
| 1157 | { |
| 1158 | r.enableContinue = r.enableContinue & ~r.enableStack[r.enableIndex]; |
| 1159 | } |
| 1160 | |
| 1161 | void PixelProgram::TEST() |
| 1162 | { |
| 1163 | whileTest = true; |
| 1164 | } |
| 1165 | |
| 1166 | void PixelProgram::CALL(Registers &r, int labelIndex, int callSiteIndex) |
| 1167 | { |
| 1168 | if(!labelBlock[labelIndex]) |
| 1169 | { |
| 1170 | labelBlock[labelIndex] = Nucleus::createBasicBlock(); |
| 1171 | } |
| 1172 | |
| 1173 | if(callRetBlock[labelIndex].size() > 1) |
| 1174 | { |
| 1175 | r.callStack[r.stackIndex++] = UInt(callSiteIndex); |
| 1176 | } |
| 1177 | |
| 1178 | Int4 restoreLeave = r.enableLeave; |
| 1179 | |
| 1180 | Nucleus::createBr(labelBlock[labelIndex]); |
| 1181 | Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]); |
| 1182 | |
| 1183 | r.enableLeave = restoreLeave; |
| 1184 | } |
| 1185 | |
| 1186 | void PixelProgram::CALLNZ(Registers &r, int labelIndex, int callSiteIndex, const Src &src) |
| 1187 | { |
| 1188 | if(src.type == Shader::PARAMETER_CONSTBOOL) |
| 1189 | { |
| 1190 | CALLNZb(r, labelIndex, callSiteIndex, src); |
| 1191 | } |
| 1192 | else if(src.type == Shader::PARAMETER_PREDICATE) |
| 1193 | { |
| 1194 | CALLNZp(r, labelIndex, callSiteIndex, src); |
| 1195 | } |
| 1196 | else ASSERT(false); |
| 1197 | } |
| 1198 | |
| 1199 | void PixelProgram::CALLNZb(Registers &r, int labelIndex, int callSiteIndex, const Src &boolRegister) |
| 1200 | { |
| 1201 | Bool condition = (*Pointer<Byte>(r.data + OFFSET(DrawData, ps.b[boolRegister.index])) != Byte(0)); // FIXME |
| 1202 | |
| 1203 | if(boolRegister.modifier == Shader::MODIFIER_NOT) |
| 1204 | { |
| 1205 | condition = !condition; |
| 1206 | } |
| 1207 | |
| 1208 | if(!labelBlock[labelIndex]) |
| 1209 | { |
| 1210 | labelBlock[labelIndex] = Nucleus::createBasicBlock(); |
| 1211 | } |
| 1212 | |
| 1213 | if(callRetBlock[labelIndex].size() > 1) |
| 1214 | { |
| 1215 | r.callStack[r.stackIndex++] = UInt(callSiteIndex); |
| 1216 | } |
| 1217 | |
| 1218 | Int4 restoreLeave = r.enableLeave; |
| 1219 | |
| 1220 | branch(condition, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]); |
| 1221 | Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]); |
| 1222 | |
| 1223 | r.enableLeave = restoreLeave; |
| 1224 | } |
| 1225 | |
| 1226 | void PixelProgram::CALLNZp(Registers &r, int labelIndex, int callSiteIndex, const Src &predicateRegister) |
| 1227 | { |
| 1228 | Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]); |
| 1229 | |
| 1230 | if(predicateRegister.modifier == Shader::MODIFIER_NOT) |
| 1231 | { |
| 1232 | condition = ~condition; |
| 1233 | } |
| 1234 | |
| 1235 | condition &= r.enableStack[r.enableIndex]; |
| 1236 | |
| 1237 | if(!labelBlock[labelIndex]) |
| 1238 | { |
| 1239 | labelBlock[labelIndex] = Nucleus::createBasicBlock(); |
| 1240 | } |
| 1241 | |
| 1242 | if(callRetBlock[labelIndex].size() > 1) |
| 1243 | { |
| 1244 | r.callStack[r.stackIndex++] = UInt(callSiteIndex); |
| 1245 | } |
| 1246 | |
| 1247 | r.enableIndex++; |
| 1248 | r.enableStack[r.enableIndex] = condition; |
| 1249 | Int4 restoreLeave = r.enableLeave; |
| 1250 | |
| 1251 | Bool notAllFalse = SignMask(condition) != 0; |
| 1252 | branch(notAllFalse, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]); |
| 1253 | Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]); |
| 1254 | |
| 1255 | r.enableIndex--; |
| 1256 | r.enableLeave = restoreLeave; |
| 1257 | } |
| 1258 | |
| 1259 | void PixelProgram::ELSE(Registers &r) |
| 1260 | { |
| 1261 | ifDepth--; |
| 1262 | |
| 1263 | llvm::BasicBlock *falseBlock = ifFalseBlock[ifDepth]; |
| 1264 | llvm::BasicBlock *endBlock = Nucleus::createBasicBlock(); |
| 1265 | |
| 1266 | if(isConditionalIf[ifDepth]) |
| 1267 | { |
| 1268 | Int4 condition = ~r.enableStack[r.enableIndex] & r.enableStack[r.enableIndex - 1]; |
| 1269 | Bool notAllFalse = SignMask(condition) != 0; |
| 1270 | |
| 1271 | branch(notAllFalse, falseBlock, endBlock); |
| 1272 | |
| 1273 | r.enableStack[r.enableIndex] = ~r.enableStack[r.enableIndex] & r.enableStack[r.enableIndex - 1]; |
| 1274 | } |
| 1275 | else |
| 1276 | { |
| 1277 | Nucleus::createBr(endBlock); |
| 1278 | Nucleus::setInsertBlock(falseBlock); |
| 1279 | } |
| 1280 | |
| 1281 | ifFalseBlock[ifDepth] = endBlock; |
| 1282 | |
| 1283 | ifDepth++; |
| 1284 | } |
| 1285 | |
| 1286 | void PixelProgram::ENDIF(Registers &r) |
| 1287 | { |
| 1288 | ifDepth--; |
| 1289 | |
| 1290 | llvm::BasicBlock *endBlock = ifFalseBlock[ifDepth]; |
| 1291 | |
| 1292 | Nucleus::createBr(endBlock); |
| 1293 | Nucleus::setInsertBlock(endBlock); |
| 1294 | |
| 1295 | if(isConditionalIf[ifDepth]) |
| 1296 | { |
| 1297 | breakDepth--; |
| 1298 | r.enableIndex--; |
| 1299 | } |
| 1300 | } |
| 1301 | |
| 1302 | void PixelProgram::ENDLOOP(Registers &r) |
| 1303 | { |
| 1304 | loopRepDepth--; |
| 1305 | |
| 1306 | r.aL[r.loopDepth] = r.aL[r.loopDepth] + r.increment[r.loopDepth]; // FIXME: += |
| 1307 | |
| 1308 | llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth]; |
| 1309 | llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth]; |
| 1310 | |
| 1311 | Nucleus::createBr(testBlock); |
| 1312 | Nucleus::setInsertBlock(endBlock); |
| 1313 | |
| 1314 | r.loopDepth--; |
| 1315 | r.enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF); |
| 1316 | } |
| 1317 | |
| 1318 | void PixelProgram::ENDREP(Registers &r) |
| 1319 | { |
| 1320 | loopRepDepth--; |
| 1321 | |
| 1322 | llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth]; |
| 1323 | llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth]; |
| 1324 | |
| 1325 | Nucleus::createBr(testBlock); |
| 1326 | Nucleus::setInsertBlock(endBlock); |
| 1327 | |
| 1328 | r.loopDepth--; |
| 1329 | r.enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF); |
| 1330 | } |
| 1331 | |
| 1332 | void PixelProgram::ENDWHILE(Registers &r) |
| 1333 | { |
| 1334 | loopRepDepth--; |
| 1335 | |
| 1336 | llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth]; |
| 1337 | llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth]; |
| 1338 | |
| 1339 | Nucleus::createBr(testBlock); |
| 1340 | Nucleus::setInsertBlock(endBlock); |
| 1341 | |
| 1342 | r.enableIndex--; |
| 1343 | r.enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF); |
| 1344 | whileTest = false; |
| 1345 | } |
| 1346 | |
| 1347 | void PixelProgram::IF(Registers &r, const Src &src) |
| 1348 | { |
| 1349 | if(src.type == Shader::PARAMETER_CONSTBOOL) |
| 1350 | { |
| 1351 | IFb(r, src); |
| 1352 | } |
| 1353 | else if(src.type == Shader::PARAMETER_PREDICATE) |
| 1354 | { |
| 1355 | IFp(r, src); |
| 1356 | } |
| 1357 | else |
| 1358 | { |
| 1359 | Int4 condition = As<Int4>(fetchRegisterF(r, src).x); |
| 1360 | IF(r, condition); |
| 1361 | } |
| 1362 | } |
| 1363 | |
| 1364 | void PixelProgram::IFb(Registers &r, const Src &boolRegister) |
| 1365 | { |
| 1366 | ASSERT(ifDepth < 24 + 4); |
| 1367 | |
| 1368 | Bool condition = (*Pointer<Byte>(r.data + OFFSET(DrawData, ps.b[boolRegister.index])) != Byte(0)); // FIXME |
| 1369 | |
| 1370 | if(boolRegister.modifier == Shader::MODIFIER_NOT) |
| 1371 | { |
| 1372 | condition = !condition; |
| 1373 | } |
| 1374 | |
| 1375 | llvm::BasicBlock *trueBlock = Nucleus::createBasicBlock(); |
| 1376 | llvm::BasicBlock *falseBlock = Nucleus::createBasicBlock(); |
| 1377 | |
| 1378 | branch(condition, trueBlock, falseBlock); |
| 1379 | |
| 1380 | isConditionalIf[ifDepth] = false; |
| 1381 | ifFalseBlock[ifDepth] = falseBlock; |
| 1382 | |
| 1383 | ifDepth++; |
| 1384 | } |
| 1385 | |
| 1386 | void PixelProgram::IFp(Registers &r, const Src &predicateRegister) |
| 1387 | { |
| 1388 | Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]); |
| 1389 | |
| 1390 | if(predicateRegister.modifier == Shader::MODIFIER_NOT) |
| 1391 | { |
| 1392 | condition = ~condition; |
| 1393 | } |
| 1394 | |
| 1395 | IF(r, condition); |
| 1396 | } |
| 1397 | |
| 1398 | void PixelProgram::IFC(Registers &r, Vector4f &src0, Vector4f &src1, Control control) |
| 1399 | { |
| 1400 | Int4 condition; |
| 1401 | |
| 1402 | switch(control) |
| 1403 | { |
| 1404 | case Shader::CONTROL_GT: condition = CmpNLE(src0.x, src1.x); break; |
| 1405 | case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x); break; |
| 1406 | case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x); break; |
| 1407 | case Shader::CONTROL_LT: condition = CmpLT(src0.x, src1.x); break; |
| 1408 | case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x); break; |
| 1409 | case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x); break; |
| 1410 | default: |
| 1411 | ASSERT(false); |
| 1412 | } |
| 1413 | |
| 1414 | IF(r, condition); |
| 1415 | } |
| 1416 | |
| 1417 | void PixelProgram::IF(Registers &r, Int4 &condition) |
| 1418 | { |
| 1419 | condition &= r.enableStack[r.enableIndex]; |
| 1420 | |
| 1421 | r.enableIndex++; |
| 1422 | r.enableStack[r.enableIndex] = condition; |
| 1423 | |
| 1424 | llvm::BasicBlock *trueBlock = Nucleus::createBasicBlock(); |
| 1425 | llvm::BasicBlock *falseBlock = Nucleus::createBasicBlock(); |
| 1426 | |
| 1427 | Bool notAllFalse = SignMask(condition) != 0; |
| 1428 | |
| 1429 | branch(notAllFalse, trueBlock, falseBlock); |
| 1430 | |
| 1431 | isConditionalIf[ifDepth] = true; |
| 1432 | ifFalseBlock[ifDepth] = falseBlock; |
| 1433 | |
| 1434 | ifDepth++; |
| 1435 | breakDepth++; |
| 1436 | } |
| 1437 | |
| 1438 | void PixelProgram::LABEL(int labelIndex) |
| 1439 | { |
| 1440 | if(!labelBlock[labelIndex]) |
| 1441 | { |
| 1442 | labelBlock[labelIndex] = Nucleus::createBasicBlock(); |
| 1443 | } |
| 1444 | |
| 1445 | Nucleus::setInsertBlock(labelBlock[labelIndex]); |
| 1446 | currentLabel = labelIndex; |
| 1447 | } |
| 1448 | |
| 1449 | void PixelProgram::LOOP(Registers &r, const Src &integerRegister) |
| 1450 | { |
| 1451 | r.loopDepth++; |
| 1452 | |
| 1453 | r.iteration[r.loopDepth] = *Pointer<Int>(r.data + OFFSET(DrawData, ps.i[integerRegister.index][0])); |
| 1454 | r.aL[r.loopDepth] = *Pointer<Int>(r.data + OFFSET(DrawData, ps.i[integerRegister.index][1])); |
| 1455 | r.increment[r.loopDepth] = *Pointer<Int>(r.data + OFFSET(DrawData, ps.i[integerRegister.index][2])); |
| 1456 | |
| 1457 | // If(r.increment[r.loopDepth] == 0) |
| 1458 | // { |
| 1459 | // r.increment[r.loopDepth] = 1; |
| 1460 | // } |
| 1461 | |
| 1462 | llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock(); |
| 1463 | llvm::BasicBlock *testBlock = Nucleus::createBasicBlock(); |
| 1464 | llvm::BasicBlock *endBlock = Nucleus::createBasicBlock(); |
| 1465 | |
| 1466 | loopRepTestBlock[loopRepDepth] = testBlock; |
| 1467 | loopRepEndBlock[loopRepDepth] = endBlock; |
| 1468 | |
| 1469 | // FIXME: jump(testBlock) |
| 1470 | Nucleus::createBr(testBlock); |
| 1471 | Nucleus::setInsertBlock(testBlock); |
| 1472 | |
| 1473 | branch(r.iteration[r.loopDepth] > 0, loopBlock, endBlock); |
| 1474 | Nucleus::setInsertBlock(loopBlock); |
| 1475 | |
| 1476 | r.iteration[r.loopDepth] = r.iteration[r.loopDepth] - 1; // FIXME: -- |
| 1477 | |
| 1478 | loopRepDepth++; |
| 1479 | breakDepth = 0; |
| 1480 | } |
| 1481 | |
| 1482 | void PixelProgram::REP(Registers &r, const Src &integerRegister) |
| 1483 | { |
| 1484 | r.loopDepth++; |
| 1485 | |
| 1486 | r.iteration[r.loopDepth] = *Pointer<Int>(r.data + OFFSET(DrawData, ps.i[integerRegister.index][0])); |
| 1487 | r.aL[r.loopDepth] = r.aL[r.loopDepth - 1]; |
| 1488 | |
| 1489 | llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock(); |
| 1490 | llvm::BasicBlock *testBlock = Nucleus::createBasicBlock(); |
| 1491 | llvm::BasicBlock *endBlock = Nucleus::createBasicBlock(); |
| 1492 | |
| 1493 | loopRepTestBlock[loopRepDepth] = testBlock; |
| 1494 | loopRepEndBlock[loopRepDepth] = endBlock; |
| 1495 | |
| 1496 | // FIXME: jump(testBlock) |
| 1497 | Nucleus::createBr(testBlock); |
| 1498 | Nucleus::setInsertBlock(testBlock); |
| 1499 | |
| 1500 | branch(r.iteration[r.loopDepth] > 0, loopBlock, endBlock); |
| 1501 | Nucleus::setInsertBlock(loopBlock); |
| 1502 | |
| 1503 | r.iteration[r.loopDepth] = r.iteration[r.loopDepth] - 1; // FIXME: -- |
| 1504 | |
| 1505 | loopRepDepth++; |
| 1506 | breakDepth = 0; |
| 1507 | } |
| 1508 | |
| 1509 | void PixelProgram::WHILE(Registers &r, const Src &temporaryRegister) |
| 1510 | { |
| 1511 | r.enableIndex++; |
| 1512 | |
| 1513 | llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock(); |
| 1514 | llvm::BasicBlock *testBlock = Nucleus::createBasicBlock(); |
| 1515 | llvm::BasicBlock *endBlock = Nucleus::createBasicBlock(); |
| 1516 | |
| 1517 | loopRepTestBlock[loopRepDepth] = testBlock; |
| 1518 | loopRepEndBlock[loopRepDepth] = endBlock; |
| 1519 | |
| 1520 | Int4 restoreBreak = r.enableBreak; |
| 1521 | Int4 restoreContinue = r.enableContinue; |
| 1522 | |
| 1523 | // FIXME: jump(testBlock) |
| 1524 | Nucleus::createBr(testBlock); |
| 1525 | Nucleus::setInsertBlock(testBlock); |
| 1526 | r.enableContinue = restoreContinue; |
| 1527 | |
| 1528 | const Vector4f &src = fetchRegisterF(r, temporaryRegister); |
| 1529 | Int4 condition = As<Int4>(src.x); |
| 1530 | condition &= r.enableStack[r.enableIndex - 1]; |
| 1531 | r.enableStack[r.enableIndex] = condition; |
| 1532 | |
| 1533 | Bool notAllFalse = SignMask(condition) != 0; |
| 1534 | branch(notAllFalse, loopBlock, endBlock); |
| 1535 | |
| 1536 | Nucleus::setInsertBlock(endBlock); |
| 1537 | r.enableBreak = restoreBreak; |
| 1538 | |
| 1539 | Nucleus::setInsertBlock(loopBlock); |
| 1540 | |
| 1541 | loopRepDepth++; |
| 1542 | breakDepth = 0; |
| 1543 | } |
| 1544 | |
| 1545 | void PixelProgram::RET(Registers &r) |
| 1546 | { |
| 1547 | if(currentLabel == -1) |
| 1548 | { |
| 1549 | returnBlock = Nucleus::createBasicBlock(); |
| 1550 | Nucleus::createBr(returnBlock); |
| 1551 | } |
| 1552 | else |
| 1553 | { |
| 1554 | llvm::BasicBlock *unreachableBlock = Nucleus::createBasicBlock(); |
| 1555 | |
| 1556 | if(callRetBlock[currentLabel].size() > 1) // Pop the return destination from the call stack |
| 1557 | { |
| 1558 | // FIXME: Encapsulate |
| 1559 | UInt index = r.callStack[--r.stackIndex]; |
| 1560 | |
| 1561 | llvm::Value *value = index.loadValue(); |
| 1562 | llvm::Value *switchInst = Nucleus::createSwitch(value, unreachableBlock, (int)callRetBlock[currentLabel].size()); |
| 1563 | |
| 1564 | for(unsigned int i = 0; i < callRetBlock[currentLabel].size(); i++) |
| 1565 | { |
| 1566 | Nucleus::addSwitchCase(switchInst, i, callRetBlock[currentLabel][i]); |
| 1567 | } |
| 1568 | } |
| 1569 | else if(callRetBlock[currentLabel].size() == 1) // Jump directly to the unique return destination |
| 1570 | { |
| 1571 | Nucleus::createBr(callRetBlock[currentLabel][0]); |
| 1572 | } |
| 1573 | else // Function isn't called |
| 1574 | { |
| 1575 | Nucleus::createBr(unreachableBlock); |
| 1576 | } |
| 1577 | |
| 1578 | Nucleus::setInsertBlock(unreachableBlock); |
| 1579 | Nucleus::createUnreachable(); |
| 1580 | } |
| 1581 | } |
| 1582 | |
| 1583 | void PixelProgram::LEAVE(Registers &r) |
| 1584 | { |
| 1585 | r.enableLeave = r.enableLeave & ~r.enableStack[r.enableIndex]; |
| 1586 | |
| 1587 | // FIXME: Return from function if all instances left |
| 1588 | // FIXME: Use enableLeave in other control-flow constructs |
| 1589 | } |
| 1590 |
|
| 1591 | }
|