OSDN Git Service

gl_VertexID implementation
[android-x86/external-swiftshader.git] / src / Shader / PixelProgram.cpp
1 // Copyright 2016 The SwiftShader Authors. All Rights Reserved.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //    http://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14
15 #include "PixelProgram.hpp"
16 #include "Primitive.hpp"
17 #include "Renderer.hpp"
18 #include "SamplerCore.hpp"
19
20 namespace sw
21 {
22         extern bool postBlendSRGB;
23         extern bool booleanFaceRegister;
24         extern bool halfIntegerCoordinates;     // Pixel centers are not at integer coordinates
25         extern bool fullPixelPositionRegister;
26
27         void PixelProgram::setBuiltins(Int &x, Int &y, Float4(&z)[4], Float4 &w)
28         {
29                 if(shader->getVersion() >= 0x0300)
30                 {
31                         if(shader->isVPosDeclared())
32                         {
33                                 if(!halfIntegerCoordinates)
34                                 {
35                                         vPos.x = Float4(Float(x)) + Float4(0, 1, 0, 1);
36                                         vPos.y = Float4(Float(y)) + Float4(0, 0, 1, 1);
37                                 }
38                                 else
39                                 {
40                                         vPos.x = Float4(Float(x)) + Float4(0.5f, 1.5f, 0.5f, 1.5f);
41                                         vPos.y = Float4(Float(y)) + Float4(0.5f, 0.5f, 1.5f, 1.5f);
42                                 }
43
44                                 if(fullPixelPositionRegister)
45                                 {
46                                         vPos.z = z[0]; // FIXME: Centroid?
47                                         vPos.w = w;    // FIXME: Centroid?
48                                 }
49                         }
50
51                         if(shader->isVFaceDeclared())
52                         {
53                                 Float4 area = *Pointer<Float>(primitive + OFFSET(Primitive, area));
54                                 Float4 face = booleanFaceRegister ? Float4(As<Float4>(CmpNLT(area, Float4(0.0f)))) : area;
55
56                                 vFace.x = face;
57                                 vFace.y = face;
58                                 vFace.z = face;
59                                 vFace.w = face;
60                         }
61                 }
62         }
63
64         void PixelProgram::applyShader(Int cMask[4])
65         {
66                 enableIndex = 0;
67                 stackIndex = 0;
68
69                 if(shader->containsLeaveInstruction())
70                 {
71                         enableLeave = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
72                 }
73
74                 for(int i = 0; i < RENDERTARGETS; i++)
75                 {
76                         if(state.targetFormat[i] != FORMAT_NULL)
77                         {
78                                 oC[i] = Vector4f(0.0f, 0.0f, 0.0f, 0.0f);
79                         }
80                 }
81
82                 // Create all call site return blocks up front
83                 for(size_t i = 0; i < shader->getLength(); i++)
84                 {
85                         const Shader::Instruction *instruction = shader->getInstruction(i);
86                         Shader::Opcode opcode = instruction->opcode;
87
88                         if(opcode == Shader::OPCODE_CALL || opcode == Shader::OPCODE_CALLNZ)
89                         {
90                                 const Dst &dst = instruction->dst;
91
92                                 ASSERT(callRetBlock[dst.label].size() == dst.callSite);
93                                 callRetBlock[dst.label].push_back(Nucleus::createBasicBlock());
94                         }
95                 }
96
97                 bool broadcastColor0 = true;
98
99                 for(size_t i = 0; i < shader->getLength(); i++)
100                 {
101                         const Shader::Instruction *instruction = shader->getInstruction(i);
102                         Shader::Opcode opcode = instruction->opcode;
103
104                         if(opcode == Shader::OPCODE_DCL || opcode == Shader::OPCODE_DEF || opcode == Shader::OPCODE_DEFI || opcode == Shader::OPCODE_DEFB)
105                         {
106                                 continue;
107                         }
108
109                         const Dst &dst = instruction->dst;
110                         const Src &src0 = instruction->src[0];
111                         const Src &src1 = instruction->src[1];
112                         const Src &src2 = instruction->src[2];
113                         const Src &src3 = instruction->src[3];
114                         const Src &src4 = instruction->src[4];
115
116                         bool predicate = instruction->predicate;
117                         Control control = instruction->control;
118                         bool pp = dst.partialPrecision;
119                         bool project = instruction->project;
120                         bool bias = instruction->bias;
121
122                         Vector4f d;
123                         Vector4f s0;
124                         Vector4f s1;
125                         Vector4f s2;
126                         Vector4f s3;
127                         Vector4f s4;
128
129                         if(opcode == Shader::OPCODE_TEXKILL)   // Takes destination as input
130                         {
131                                 if(dst.type == Shader::PARAMETER_TEXTURE)
132                                 {
133                                         d.x = v[2 + dst.index].x;
134                                         d.y = v[2 + dst.index].y;
135                                         d.z = v[2 + dst.index].z;
136                                         d.w = v[2 + dst.index].w;
137                                 }
138                                 else
139                                 {
140                                         d = r[dst.index];
141                                 }
142                         }
143
144                         if(src0.type != Shader::PARAMETER_VOID) s0 = fetchRegister(src0);
145                         if(src1.type != Shader::PARAMETER_VOID) s1 = fetchRegister(src1);
146                         if(src2.type != Shader::PARAMETER_VOID) s2 = fetchRegister(src2);
147                         if(src3.type != Shader::PARAMETER_VOID) s3 = fetchRegister(src3);
148                         if(src4.type != Shader::PARAMETER_VOID) s4 = fetchRegister(src4);
149
150                         switch(opcode)
151                         {
152                         case Shader::OPCODE_PS_2_0:                                                    break;
153                         case Shader::OPCODE_PS_2_x:                                                    break;
154                         case Shader::OPCODE_PS_3_0:                                                    break;
155                         case Shader::OPCODE_DEF:                                                       break;
156                         case Shader::OPCODE_DCL:                                                       break;
157                         case Shader::OPCODE_NOP:                                                       break;
158                         case Shader::OPCODE_MOV:        mov(d, s0);                                    break;
159                         case Shader::OPCODE_NEG:        neg(d, s0);                                    break;
160                         case Shader::OPCODE_INEG:       ineg(d, s0);                                   break;
161                         case Shader::OPCODE_F2B:        f2b(d, s0);                                    break;
162                         case Shader::OPCODE_B2F:        b2f(d, s0);                                    break;
163                         case Shader::OPCODE_F2I:        f2i(d, s0);                                    break;
164                         case Shader::OPCODE_I2F:        i2f(d, s0);                                    break;
165                         case Shader::OPCODE_F2U:        f2u(d, s0);                                    break;
166                         case Shader::OPCODE_U2F:        u2f(d, s0);                                    break;
167                         case Shader::OPCODE_I2B:        i2b(d, s0);                                    break;
168                         case Shader::OPCODE_B2I:        b2i(d, s0);                                    break;
169                         case Shader::OPCODE_ADD:        add(d, s0, s1);                                break;
170                         case Shader::OPCODE_IADD:       iadd(d, s0, s1);                               break;
171                         case Shader::OPCODE_SUB:        sub(d, s0, s1);                                break;
172                         case Shader::OPCODE_ISUB:       isub(d, s0, s1);                               break;
173                         case Shader::OPCODE_MUL:        mul(d, s0, s1);                                break;
174                         case Shader::OPCODE_IMUL:       imul(d, s0, s1);                               break;
175                         case Shader::OPCODE_MAD:        mad(d, s0, s1, s2);                            break;
176                         case Shader::OPCODE_IMAD:       imad(d, s0, s1, s2);                           break;
177                         case Shader::OPCODE_DP1:        dp1(d, s0, s1);                                break;
178                         case Shader::OPCODE_DP2:        dp2(d, s0, s1);                                break;
179                         case Shader::OPCODE_DP2ADD:     dp2add(d, s0, s1, s2);                         break;
180                         case Shader::OPCODE_DP3:        dp3(d, s0, s1);                                break;
181                         case Shader::OPCODE_DP4:        dp4(d, s0, s1);                                break;
182                         case Shader::OPCODE_DET2:       det2(d, s0, s1);                               break;
183                         case Shader::OPCODE_DET3:       det3(d, s0, s1, s2);                           break;
184                         case Shader::OPCODE_DET4:       det4(d, s0, s1, s2, s3);                       break;
185                         case Shader::OPCODE_CMP0:       cmp0(d, s0, s1, s2);                           break;
186                         case Shader::OPCODE_ICMP:       icmp(d, s0, s1, control);                      break;
187                         case Shader::OPCODE_UCMP:       ucmp(d, s0, s1, control);                      break;
188                         case Shader::OPCODE_SELECT:     select(d, s0, s1, s2);                         break;
189                         case Shader::OPCODE_EXTRACT:    extract(d.x, s0, s1.x);                        break;
190                         case Shader::OPCODE_INSERT:     insert(d, s0, s1.x, s2.x);                     break;
191                         case Shader::OPCODE_FRC:        frc(d, s0);                                    break;
192                         case Shader::OPCODE_TRUNC:      trunc(d, s0);                                  break;
193                         case Shader::OPCODE_FLOOR:      floor(d, s0);                                  break;
194                         case Shader::OPCODE_ROUND:      round(d, s0);                                  break;
195                         case Shader::OPCODE_ROUNDEVEN:  roundEven(d, s0);                              break;
196                         case Shader::OPCODE_CEIL:       ceil(d, s0);                                   break;
197                         case Shader::OPCODE_EXP2X:      exp2x(d, s0, pp);                              break;
198                         case Shader::OPCODE_EXP2:       exp2(d, s0, pp);                               break;
199                         case Shader::OPCODE_LOG2X:      log2x(d, s0, pp);                              break;
200                         case Shader::OPCODE_LOG2:       log2(d, s0, pp);                               break;
201                         case Shader::OPCODE_EXP:        exp(d, s0, pp);                                break;
202                         case Shader::OPCODE_LOG:        log(d, s0, pp);                                break;
203                         case Shader::OPCODE_RCPX:       rcpx(d, s0, pp);                               break;
204                         case Shader::OPCODE_DIV:        div(d, s0, s1);                                break;
205                         case Shader::OPCODE_IDIV:       idiv(d, s0, s1);                               break;
206                         case Shader::OPCODE_UDIV:       udiv(d, s0, s1);                               break;
207                         case Shader::OPCODE_MOD:        mod(d, s0, s1);                                break;
208                         case Shader::OPCODE_IMOD:       imod(d, s0, s1);                               break;
209                         case Shader::OPCODE_UMOD:       umod(d, s0, s1);                               break;
210                         case Shader::OPCODE_SHL:        shl(d, s0, s1);                                break;
211                         case Shader::OPCODE_ISHR:       ishr(d, s0, s1);                               break;
212                         case Shader::OPCODE_USHR:       ushr(d, s0, s1);                               break;
213                         case Shader::OPCODE_RSQX:       rsqx(d, s0, pp);                               break;
214                         case Shader::OPCODE_SQRT:       sqrt(d, s0, pp);                               break;
215                         case Shader::OPCODE_RSQ:        rsq(d, s0, pp);                                break;
216                         case Shader::OPCODE_LEN2:       len2(d.x, s0, pp);                             break;
217                         case Shader::OPCODE_LEN3:       len3(d.x, s0, pp);                             break;
218                         case Shader::OPCODE_LEN4:       len4(d.x, s0, pp);                             break;
219                         case Shader::OPCODE_DIST1:      dist1(d.x, s0, s1, pp);                        break;
220                         case Shader::OPCODE_DIST2:      dist2(d.x, s0, s1, pp);                        break;
221                         case Shader::OPCODE_DIST3:      dist3(d.x, s0, s1, pp);                        break;
222                         case Shader::OPCODE_DIST4:      dist4(d.x, s0, s1, pp);                        break;
223                         case Shader::OPCODE_MIN:        min(d, s0, s1);                                break;
224                         case Shader::OPCODE_IMIN:       imin(d, s0, s1);                               break;
225                         case Shader::OPCODE_UMIN:       umin(d, s0, s1);                               break;
226                         case Shader::OPCODE_MAX:        max(d, s0, s1);                                break;
227                         case Shader::OPCODE_IMAX:       imax(d, s0, s1);                               break;
228                         case Shader::OPCODE_UMAX:       umax(d, s0, s1);                               break;
229                         case Shader::OPCODE_LRP:        lrp(d, s0, s1, s2);                            break;
230                         case Shader::OPCODE_STEP:       step(d, s0, s1);                               break;
231                         case Shader::OPCODE_SMOOTH:     smooth(d, s0, s1, s2);                         break;
232                         case Shader::OPCODE_FLOATBITSTOINT:
233                         case Shader::OPCODE_FLOATBITSTOUINT:
234                         case Shader::OPCODE_INTBITSTOFLOAT:
235                         case Shader::OPCODE_UINTBITSTOFLOAT: d = s0;                                   break;
236                         case Shader::OPCODE_PACKSNORM2x16:   packSnorm2x16(d, s0);                     break;
237                         case Shader::OPCODE_PACKUNORM2x16:   packUnorm2x16(d, s0);                     break;
238                         case Shader::OPCODE_PACKHALF2x16:    packHalf2x16(d, s0);                      break;
239                         case Shader::OPCODE_UNPACKSNORM2x16: unpackSnorm2x16(d, s0);                   break;
240                         case Shader::OPCODE_UNPACKUNORM2x16: unpackUnorm2x16(d, s0);                   break;
241                         case Shader::OPCODE_UNPACKHALF2x16:  unpackHalf2x16(d, s0);                    break;
242                         case Shader::OPCODE_POWX:       powx(d, s0, s1, pp);                           break;
243                         case Shader::OPCODE_POW:        pow(d, s0, s1, pp);                            break;
244                         case Shader::OPCODE_SGN:        sgn(d, s0);                                    break;
245                         case Shader::OPCODE_ISGN:       isgn(d, s0);                                   break;
246                         case Shader::OPCODE_CRS:        crs(d, s0, s1);                                break;
247                         case Shader::OPCODE_FORWARD1:   forward1(d, s0, s1, s2);                       break;
248                         case Shader::OPCODE_FORWARD2:   forward2(d, s0, s1, s2);                       break;
249                         case Shader::OPCODE_FORWARD3:   forward3(d, s0, s1, s2);                       break;
250                         case Shader::OPCODE_FORWARD4:   forward4(d, s0, s1, s2);                       break;
251                         case Shader::OPCODE_REFLECT1:   reflect1(d, s0, s1);                           break;
252                         case Shader::OPCODE_REFLECT2:   reflect2(d, s0, s1);                           break;
253                         case Shader::OPCODE_REFLECT3:   reflect3(d, s0, s1);                           break;
254                         case Shader::OPCODE_REFLECT4:   reflect4(d, s0, s1);                           break;
255                         case Shader::OPCODE_REFRACT1:   refract1(d, s0, s1, s2.x);                     break;
256                         case Shader::OPCODE_REFRACT2:   refract2(d, s0, s1, s2.x);                     break;
257                         case Shader::OPCODE_REFRACT3:   refract3(d, s0, s1, s2.x);                     break;
258                         case Shader::OPCODE_REFRACT4:   refract4(d, s0, s1, s2.x);                     break;
259                         case Shader::OPCODE_NRM2:       nrm2(d, s0, pp);                               break;
260                         case Shader::OPCODE_NRM3:       nrm3(d, s0, pp);                               break;
261                         case Shader::OPCODE_NRM4:       nrm4(d, s0, pp);                               break;
262                         case Shader::OPCODE_ABS:        abs(d, s0);                                    break;
263                         case Shader::OPCODE_IABS:       iabs(d, s0);                                   break;
264                         case Shader::OPCODE_SINCOS:     sincos(d, s0, pp);                             break;
265                         case Shader::OPCODE_COS:        cos(d, s0, pp);                                break;
266                         case Shader::OPCODE_SIN:        sin(d, s0, pp);                                break;
267                         case Shader::OPCODE_TAN:        tan(d, s0, pp);                                break;
268                         case Shader::OPCODE_ACOS:       acos(d, s0, pp);                               break;
269                         case Shader::OPCODE_ASIN:       asin(d, s0, pp);                               break;
270                         case Shader::OPCODE_ATAN:       atan(d, s0, pp);                               break;
271                         case Shader::OPCODE_ATAN2:      atan2(d, s0, s1, pp);                          break;
272                         case Shader::OPCODE_COSH:       cosh(d, s0, pp);                               break;
273                         case Shader::OPCODE_SINH:       sinh(d, s0, pp);                               break;
274                         case Shader::OPCODE_TANH:       tanh(d, s0, pp);                               break;
275                         case Shader::OPCODE_ACOSH:      acosh(d, s0, pp);                              break;
276                         case Shader::OPCODE_ASINH:      asinh(d, s0, pp);                              break;
277                         case Shader::OPCODE_ATANH:      atanh(d, s0, pp);                              break;
278                         case Shader::OPCODE_M4X4:       M4X4(d, s0, src1);                             break;
279                         case Shader::OPCODE_M4X3:       M4X3(d, s0, src1);                             break;
280                         case Shader::OPCODE_M3X4:       M3X4(d, s0, src1);                             break;
281                         case Shader::OPCODE_M3X3:       M3X3(d, s0, src1);                             break;
282                         case Shader::OPCODE_M3X2:       M3X2(d, s0, src1);                             break;
283                         case Shader::OPCODE_TEX:        TEXLD(d, s0, src1, project, bias);             break;
284                         case Shader::OPCODE_TEXLDD:     TEXLDD(d, s0, src1, s2, s3);                   break;
285                         case Shader::OPCODE_TEXLDL:     TEXLDL(d, s0, src1);                           break;
286                         case Shader::OPCODE_TEXSIZE:    TEXSIZE(d, s0.x, src1);                        break;
287                         case Shader::OPCODE_TEXKILL:    TEXKILL(cMask, d, dst.mask);                   break;
288                         case Shader::OPCODE_TEXOFFSET:  TEXOFFSET(d, s0, src1, s2, bias);              break;
289                         case Shader::OPCODE_TEXLDLOFFSET: TEXLDL(d, s0, src1, s2, bias);               break;
290                         case Shader::OPCODE_TEXELFETCH: TEXELFETCH(d, s0, src1);                       break;
291                         case Shader::OPCODE_TEXELFETCHOFFSET: TEXELFETCH(d, s0, src1, s2);             break;
292                         case Shader::OPCODE_TEXGRAD:    TEXGRAD(d, s0, src1, s2, s3);                  break;
293                         case Shader::OPCODE_TEXGRADOFFSET: TEXGRAD(d, s0, src1, s2, s3, s4);           break;
294                         case Shader::OPCODE_DISCARD:    DISCARD(cMask, instruction);                   break;
295                         case Shader::OPCODE_DFDX:       DFDX(d, s0);                                   break;
296                         case Shader::OPCODE_DFDY:       DFDY(d, s0);                                   break;
297                         case Shader::OPCODE_FWIDTH:     FWIDTH(d, s0);                                 break;
298                         case Shader::OPCODE_BREAK:      BREAK();                                       break;
299                         case Shader::OPCODE_BREAKC:     BREAKC(s0, s1, control);                       break;
300                         case Shader::OPCODE_BREAKP:     BREAKP(src0);                                  break;
301                         case Shader::OPCODE_CONTINUE:   CONTINUE();                                    break;
302                         case Shader::OPCODE_TEST:       TEST();                                        break;
303                         case Shader::OPCODE_CALL:       CALL(dst.label, dst.callSite);                 break;
304                         case Shader::OPCODE_CALLNZ:     CALLNZ(dst.label, dst.callSite, src0);         break;
305                         case Shader::OPCODE_ELSE:       ELSE();                                        break;
306                         case Shader::OPCODE_ENDIF:      ENDIF();                                       break;
307                         case Shader::OPCODE_ENDLOOP:    ENDLOOP();                                     break;
308                         case Shader::OPCODE_ENDREP:     ENDREP();                                      break;
309                         case Shader::OPCODE_ENDWHILE:   ENDWHILE();                                    break;
310                         case Shader::OPCODE_ENDSWITCH:  ENDSWITCH();                                   break;
311                         case Shader::OPCODE_IF:         IF(src0);                                      break;
312                         case Shader::OPCODE_IFC:        IFC(s0, s1, control);                          break;
313                         case Shader::OPCODE_LABEL:      LABEL(dst.index);                              break;
314                         case Shader::OPCODE_LOOP:       LOOP(src1);                                    break;
315                         case Shader::OPCODE_REP:        REP(src0);                                     break;
316                         case Shader::OPCODE_WHILE:      WHILE(src0);                                   break;
317                         case Shader::OPCODE_SWITCH:     SWITCH();                                      break;
318                         case Shader::OPCODE_RET:        RET();                                         break;
319                         case Shader::OPCODE_LEAVE:      LEAVE();                                       break;
320                         case Shader::OPCODE_CMP:        cmp(d, s0, s1, control);                       break;
321                         case Shader::OPCODE_ALL:        all(d.x, s0);                                  break;
322                         case Shader::OPCODE_ANY:        any(d.x, s0);                                  break;
323                         case Shader::OPCODE_NOT:        bitwise_not(d, s0);                            break;
324                         case Shader::OPCODE_OR:         bitwise_or(d, s0, s1);                         break;
325                         case Shader::OPCODE_XOR:        bitwise_xor(d, s0, s1);                        break;
326                         case Shader::OPCODE_AND:        bitwise_and(d, s0, s1);                        break;
327                         case Shader::OPCODE_EQ:         equal(d, s0, s1);                              break;
328                         case Shader::OPCODE_NE:         notEqual(d, s0, s1);                           break;
329                         case Shader::OPCODE_END:                                                       break;
330                         default:
331                                 ASSERT(false);
332                         }
333
334                         if(dst.type != Shader::PARAMETER_VOID && dst.type != Shader::PARAMETER_LABEL && opcode != Shader::OPCODE_TEXKILL && opcode != Shader::OPCODE_NOP)
335                         {
336                                 if(dst.integer)
337                                 {
338                                         switch(opcode)
339                                         {
340                                         case Shader::OPCODE_DIV:
341                                                 if(dst.x) d.x = Trunc(d.x);
342                                                 if(dst.y) d.y = Trunc(d.y);
343                                                 if(dst.z) d.z = Trunc(d.z);
344                                                 if(dst.w) d.w = Trunc(d.w);
345                                                 break;
346                                         default:
347                                                 break;   // No truncation to integer required when arguments are integer
348                                         }
349                                 }
350
351                                 if(dst.saturate)
352                                 {
353                                         if(dst.x) d.x = Max(d.x, Float4(0.0f));
354                                         if(dst.y) d.y = Max(d.y, Float4(0.0f));
355                                         if(dst.z) d.z = Max(d.z, Float4(0.0f));
356                                         if(dst.w) d.w = Max(d.w, Float4(0.0f));
357
358                                         if(dst.x) d.x = Min(d.x, Float4(1.0f));
359                                         if(dst.y) d.y = Min(d.y, Float4(1.0f));
360                                         if(dst.z) d.z = Min(d.z, Float4(1.0f));
361                                         if(dst.w) d.w = Min(d.w, Float4(1.0f));
362                                 }
363
364                                 if(instruction->isPredicated())
365                                 {
366                                         Vector4f pDst;   // FIXME: Rename
367
368                                         switch(dst.type)
369                                         {
370                                         case Shader::PARAMETER_TEMP:
371                                                 if(dst.rel.type == Shader::PARAMETER_VOID)
372                                                 {
373                                                         if(dst.x) pDst.x = r[dst.index].x;
374                                                         if(dst.y) pDst.y = r[dst.index].y;
375                                                         if(dst.z) pDst.z = r[dst.index].z;
376                                                         if(dst.w) pDst.w = r[dst.index].w;
377                                                 }
378                                                 else
379                                                 {
380                                                         Int a = relativeAddress(dst);
381
382                                                         if(dst.x) pDst.x = r[dst.index + a].x;
383                                                         if(dst.y) pDst.y = r[dst.index + a].y;
384                                                         if(dst.z) pDst.z = r[dst.index + a].z;
385                                                         if(dst.w) pDst.w = r[dst.index + a].w;
386                                                 }
387                                                 break;
388                                         case Shader::PARAMETER_COLOROUT:
389                                                 if(dst.rel.type == Shader::PARAMETER_VOID)
390                                                 {
391                                                         if(dst.x) pDst.x = oC[dst.index].x;
392                                                         if(dst.y) pDst.y = oC[dst.index].y;
393                                                         if(dst.z) pDst.z = oC[dst.index].z;
394                                                         if(dst.w) pDst.w = oC[dst.index].w;
395                                                 }
396                                                 else
397                                                 {
398                                                         Int a = relativeAddress(dst) + dst.index;
399
400                                                         if(dst.x) pDst.x = oC[a].x;
401                                                         if(dst.y) pDst.y = oC[a].y;
402                                                         if(dst.z) pDst.z = oC[a].z;
403                                                         if(dst.w) pDst.w = oC[a].w;
404                                                 }
405                                                 break;
406                                         case Shader::PARAMETER_PREDICATE:
407                                                 if(dst.x) pDst.x = p0.x;
408                                                 if(dst.y) pDst.y = p0.y;
409                                                 if(dst.z) pDst.z = p0.z;
410                                                 if(dst.w) pDst.w = p0.w;
411                                                 break;
412                                         case Shader::PARAMETER_DEPTHOUT:
413                                                 pDst.x = oDepth;
414                                                 break;
415                                         default:
416                                                 ASSERT(false);
417                                         }
418
419                                         Int4 enable = enableMask(instruction);
420
421                                         Int4 xEnable = enable;
422                                         Int4 yEnable = enable;
423                                         Int4 zEnable = enable;
424                                         Int4 wEnable = enable;
425
426                                         if(predicate)
427                                         {
428                                                 unsigned char pSwizzle = instruction->predicateSwizzle;
429
430                                                 Float4 xPredicate = p0[(pSwizzle >> 0) & 0x03];
431                                                 Float4 yPredicate = p0[(pSwizzle >> 2) & 0x03];
432                                                 Float4 zPredicate = p0[(pSwizzle >> 4) & 0x03];
433                                                 Float4 wPredicate = p0[(pSwizzle >> 6) & 0x03];
434
435                                                 if(!instruction->predicateNot)
436                                                 {
437                                                         if(dst.x) xEnable = xEnable & As<Int4>(xPredicate);
438                                                         if(dst.y) yEnable = yEnable & As<Int4>(yPredicate);
439                                                         if(dst.z) zEnable = zEnable & As<Int4>(zPredicate);
440                                                         if(dst.w) wEnable = wEnable & As<Int4>(wPredicate);
441                                                 }
442                                                 else
443                                                 {
444                                                         if(dst.x) xEnable = xEnable & ~As<Int4>(xPredicate);
445                                                         if(dst.y) yEnable = yEnable & ~As<Int4>(yPredicate);
446                                                         if(dst.z) zEnable = zEnable & ~As<Int4>(zPredicate);
447                                                         if(dst.w) wEnable = wEnable & ~As<Int4>(wPredicate);
448                                                 }
449                                         }
450
451                                         if(dst.x) d.x = As<Float4>(As<Int4>(d.x) & xEnable);
452                                         if(dst.y) d.y = As<Float4>(As<Int4>(d.y) & yEnable);
453                                         if(dst.z) d.z = As<Float4>(As<Int4>(d.z) & zEnable);
454                                         if(dst.w) d.w = As<Float4>(As<Int4>(d.w) & wEnable);
455
456                                         if(dst.x) d.x = As<Float4>(As<Int4>(d.x) | (As<Int4>(pDst.x) & ~xEnable));
457                                         if(dst.y) d.y = As<Float4>(As<Int4>(d.y) | (As<Int4>(pDst.y) & ~yEnable));
458                                         if(dst.z) d.z = As<Float4>(As<Int4>(d.z) | (As<Int4>(pDst.z) & ~zEnable));
459                                         if(dst.w) d.w = As<Float4>(As<Int4>(d.w) | (As<Int4>(pDst.w) & ~wEnable));
460                                 }
461
462                                 switch(dst.type)
463                                 {
464                                 case Shader::PARAMETER_TEMP:
465                                         if(dst.rel.type == Shader::PARAMETER_VOID)
466                                         {
467                                                 if(dst.x) r[dst.index].x = d.x;
468                                                 if(dst.y) r[dst.index].y = d.y;
469                                                 if(dst.z) r[dst.index].z = d.z;
470                                                 if(dst.w) r[dst.index].w = d.w;
471                                         }
472                                         else
473                                         {
474                                                 Int a = relativeAddress(dst);
475
476                                                 if(dst.x) r[dst.index + a].x = d.x;
477                                                 if(dst.y) r[dst.index + a].y = d.y;
478                                                 if(dst.z) r[dst.index + a].z = d.z;
479                                                 if(dst.w) r[dst.index + a].w = d.w;
480                                         }
481                                         break;
482                                 case Shader::PARAMETER_COLOROUT:
483                                         if(dst.rel.type == Shader::PARAMETER_VOID)
484                                         {
485                                                 broadcastColor0 = (dst.index == 0) && broadcastColor0;
486
487                                                 if(dst.x) { oC[dst.index].x = d.x; }
488                                                 if(dst.y) { oC[dst.index].y = d.y; }
489                                                 if(dst.z) { oC[dst.index].z = d.z; }
490                                                 if(dst.w) { oC[dst.index].w = d.w; }
491                                         }
492                                         else
493                                         {
494                                                 broadcastColor0 = false;
495                                                 Int a = relativeAddress(dst) + dst.index;
496
497                                                 if(dst.x) { oC[a].x = d.x; }
498                                                 if(dst.y) { oC[a].y = d.y; }
499                                                 if(dst.z) { oC[a].z = d.z; }
500                                                 if(dst.w) { oC[a].w = d.w; }
501                                         }
502                                         break;
503                                 case Shader::PARAMETER_PREDICATE:
504                                         if(dst.x) p0.x = d.x;
505                                         if(dst.y) p0.y = d.y;
506                                         if(dst.z) p0.z = d.z;
507                                         if(dst.w) p0.w = d.w;
508                                         break;
509                                 case Shader::PARAMETER_DEPTHOUT:
510                                         oDepth = d.x;
511                                         break;
512                                 default:
513                                         ASSERT(false);
514                                 }
515                         }
516                 }
517
518                 if(currentLabel != -1)
519                 {
520                         Nucleus::setInsertBlock(returnBlock);
521                 }
522
523                 if(broadcastColor0)
524                 {
525                         for(int i = 0; i < RENDERTARGETS; i++)
526                         {
527                                 c[i] = oC[0];
528                         }
529                 }
530                 else
531                 {
532                         for(int i = 0; i < RENDERTARGETS; i++)
533                         {
534                                 c[i] = oC[i];
535                         }
536                 }
537         }
538
539         Bool PixelProgram::alphaTest(Int cMask[4])
540         {
541                 clampColor(c);
542
543                 if(!state.alphaTestActive())
544                 {
545                         return true;
546                 }
547
548                 Int aMask;
549
550                 if(state.transparencyAntialiasing == TRANSPARENCY_NONE)
551                 {
552                         Short4 alpha = RoundShort4(c[0].w * Float4(0x1000));
553
554                         PixelRoutine::alphaTest(aMask, alpha);
555
556                         for(unsigned int q = 0; q < state.multiSample; q++)
557                         {
558                                 cMask[q] &= aMask;
559                         }
560                 }
561                 else if(state.transparencyAntialiasing == TRANSPARENCY_ALPHA_TO_COVERAGE)
562                 {
563                         alphaToCoverage(cMask, c[0].w);
564                 }
565                 else ASSERT(false);
566
567                 Int pass = cMask[0];
568
569                 for(unsigned int q = 1; q < state.multiSample; q++)
570                 {
571                         pass = pass | cMask[q];
572                 }
573
574                 return pass != 0x0;
575         }
576
577         void PixelProgram::rasterOperation(Float4 &fog, Pointer<Byte> cBuffer[4], Int &x, Int sMask[4], Int zMask[4], Int cMask[4])
578         {
579                 for(int index = 0; index < RENDERTARGETS; index++)
580                 {
581                         if(!state.colorWriteActive(index))
582                         {
583                                 continue;
584                         }
585
586                         if(!postBlendSRGB && state.writeSRGB && !isSRGB(index))
587                         {
588                                 c[index].x = linearToSRGB(c[index].x);
589                                 c[index].y = linearToSRGB(c[index].y);
590                                 c[index].z = linearToSRGB(c[index].z);
591                         }
592
593                         if(index == 0)
594                         {
595                                 fogBlend(c[index], fog);
596                         }
597
598                         switch(state.targetFormat[index])
599                         {
600                         case FORMAT_R5G6B5:
601                         case FORMAT_X8R8G8B8:
602                         case FORMAT_X8B8G8R8:
603                         case FORMAT_A8R8G8B8:
604                         case FORMAT_A8B8G8R8:
605                         case FORMAT_SRGB8_X8:
606                         case FORMAT_SRGB8_A8:
607                         case FORMAT_G8R8:
608                         case FORMAT_R8:
609                         case FORMAT_A8:
610                         case FORMAT_G16R16:
611                         case FORMAT_A16B16G16R16:
612                                 for(unsigned int q = 0; q < state.multiSample; q++)
613                                 {
614                                         Pointer<Byte> buffer = cBuffer[index] + q * *Pointer<Int>(data + OFFSET(DrawData, colorSliceB[index]));
615                                         Vector4s color;
616
617                                         if(state.targetFormat[index] == FORMAT_R5G6B5)
618                                         {
619                                                 color.x = UShort4(c[index].x * Float4(0xFBFF), false);
620                                                 color.y = UShort4(c[index].y * Float4(0xFDFF), false);
621                                                 color.z = UShort4(c[index].z * Float4(0xFBFF), false);
622                                                 color.w = UShort4(c[index].w * Float4(0xFFFF), false);
623                                         }
624                                         else
625                                         {
626                                                 color.x = convertFixed16(c[index].x, false);
627                                                 color.y = convertFixed16(c[index].y, false);
628                                                 color.z = convertFixed16(c[index].z, false);
629                                                 color.w = convertFixed16(c[index].w, false);
630                                         }
631
632                                         if(state.multiSampleMask & (1 << q))
633                                         {
634                                                 alphaBlend(index, buffer, color, x);
635                                                 logicOperation(index, buffer, color, x);
636                                                 writeColor(index, buffer, x, color, sMask[q], zMask[q], cMask[q]);
637                                         }
638                                 }
639                                 break;
640                         case FORMAT_R32F:
641                         case FORMAT_G32R32F:
642                         case FORMAT_X32B32G32R32F:
643                         case FORMAT_A32B32G32R32F:
644                         case FORMAT_R32I:
645                         case FORMAT_G32R32I:
646                         case FORMAT_A32B32G32R32I:
647                         case FORMAT_R32UI:
648                         case FORMAT_G32R32UI:
649                         case FORMAT_A32B32G32R32UI:
650                         case FORMAT_R16I:
651                         case FORMAT_G16R16I:
652                         case FORMAT_A16B16G16R16I:
653                         case FORMAT_R16UI:
654                         case FORMAT_G16R16UI:
655                         case FORMAT_A16B16G16R16UI:
656                         case FORMAT_R8I:
657                         case FORMAT_G8R8I:
658                         case FORMAT_A8B8G8R8I:
659                         case FORMAT_R8UI:
660                         case FORMAT_G8R8UI:
661                         case FORMAT_A8B8G8R8UI:
662                                 for(unsigned int q = 0; q < state.multiSample; q++)
663                                 {
664                                         Pointer<Byte> buffer = cBuffer[index] + q * *Pointer<Int>(data + OFFSET(DrawData, colorSliceB[index]));
665                                         Vector4f color = c[index];
666
667                                         if(state.multiSampleMask & (1 << q))
668                                         {
669                                                 alphaBlend(index, buffer, color, x);
670                                                 writeColor(index, buffer, x, color, sMask[q], zMask[q], cMask[q]);
671                                         }
672                                 }
673                                 break;
674                         default:
675                                 ASSERT(false);
676                         }
677                 }
678         }
679
680         void PixelProgram::sampleTexture(Vector4f &c, const Src &sampler, Vector4f &uvwq, Vector4f &dsx, Vector4f &dsy, Vector4f &offset, SamplerFunction function)
681         {
682                 Vector4f tmp;
683
684                 if(sampler.type == Shader::PARAMETER_SAMPLER && sampler.rel.type == Shader::PARAMETER_VOID)
685                 {
686                         sampleTexture(tmp, sampler.index, uvwq, dsx, dsy, offset, function);
687                 }
688                 else
689                 {
690                         Int index = As<Int>(Float(fetchRegister(sampler).x.x));
691
692                         for(int i = 0; i < TEXTURE_IMAGE_UNITS; i++)
693                         {
694                                 if(shader->usesSampler(i))
695                                 {
696                                         If(index == i)
697                                         {
698                                                 sampleTexture(tmp, i, uvwq, dsx, dsy, offset, function);
699                                                 // FIXME: When the sampler states are the same, we could use one sampler and just index the texture
700                                         }
701                                 }
702                         }
703                 }
704
705                 c.x = tmp[(sampler.swizzle >> 0) & 0x3];
706                 c.y = tmp[(sampler.swizzle >> 2) & 0x3];
707                 c.z = tmp[(sampler.swizzle >> 4) & 0x3];
708                 c.w = tmp[(sampler.swizzle >> 6) & 0x3];
709         }
710
711         void PixelProgram::sampleTexture(Vector4f &c, int samplerIndex, Vector4f &uvwq, Vector4f &dsx, Vector4f &dsy, Vector4f &offset, SamplerFunction function)
712         {
713                 #if PERF_PROFILE
714                         Long texTime = Ticks();
715                 #endif
716
717                 Pointer<Byte> texture = data + OFFSET(DrawData, mipmap) + samplerIndex * sizeof(Texture);
718                 sampler[samplerIndex]->sampleTexture(texture, c, uvwq.x, uvwq.y, uvwq.z, uvwq.w, dsx, dsy, offset, function);
719
720                 #if PERF_PROFILE
721                         cycles[PERF_TEX] += Ticks() - texTime;
722                 #endif
723         }
724
725         void PixelProgram::clampColor(Vector4f oC[RENDERTARGETS])
726         {
727                 for(int index = 0; index < RENDERTARGETS; index++)
728                 {
729                         if(!state.colorWriteActive(index) && !(index == 0 && state.alphaTestActive()))
730                         {
731                                 continue;
732                         }
733
734                         switch(state.targetFormat[index])
735                         {
736                         case FORMAT_NULL:
737                                 break;
738                         case FORMAT_R5G6B5:
739                         case FORMAT_A8R8G8B8:
740                         case FORMAT_A8B8G8R8:
741                         case FORMAT_X8R8G8B8:
742                         case FORMAT_X8B8G8R8:
743                         case FORMAT_SRGB8_X8:
744                         case FORMAT_SRGB8_A8:
745                         case FORMAT_G8R8:
746                         case FORMAT_R8:
747                         case FORMAT_A8:
748                         case FORMAT_G16R16:
749                         case FORMAT_A16B16G16R16:
750                                 oC[index].x = Max(oC[index].x, Float4(0.0f)); oC[index].x = Min(oC[index].x, Float4(1.0f));
751                                 oC[index].y = Max(oC[index].y, Float4(0.0f)); oC[index].y = Min(oC[index].y, Float4(1.0f));
752                                 oC[index].z = Max(oC[index].z, Float4(0.0f)); oC[index].z = Min(oC[index].z, Float4(1.0f));
753                                 oC[index].w = Max(oC[index].w, Float4(0.0f)); oC[index].w = Min(oC[index].w, Float4(1.0f));
754                                 break;
755                         case FORMAT_R32F:
756                         case FORMAT_G32R32F:
757                         case FORMAT_X32B32G32R32F:
758                         case FORMAT_A32B32G32R32F:
759                         case FORMAT_R32I:
760                         case FORMAT_G32R32I:
761                         case FORMAT_A32B32G32R32I:
762                         case FORMAT_R32UI:
763                         case FORMAT_G32R32UI:
764                         case FORMAT_A32B32G32R32UI:
765                         case FORMAT_R16I:
766                         case FORMAT_G16R16I:
767                         case FORMAT_A16B16G16R16I:
768                         case FORMAT_R16UI:
769                         case FORMAT_G16R16UI:
770                         case FORMAT_A16B16G16R16UI:
771                         case FORMAT_R8I:
772                         case FORMAT_G8R8I:
773                         case FORMAT_A8B8G8R8I:
774                         case FORMAT_R8UI:
775                         case FORMAT_G8R8UI:
776                         case FORMAT_A8B8G8R8UI:
777                                 break;
778                         default:
779                                 ASSERT(false);
780                         }
781                 }
782         }
783
784         Int4 PixelProgram::enableMask(const Shader::Instruction *instruction)
785         {
786                 Int4 enable = instruction->analysisBranch ? Int4(enableStack[enableIndex]) : Int4(0xFFFFFFFF);
787
788                 if(!whileTest)
789                 {
790                         if(shader->containsBreakInstruction() && instruction->analysisBreak)
791                         {
792                                 enable &= enableBreak;
793                         }
794
795                         if(shader->containsContinueInstruction() && instruction->analysisContinue)
796                         {
797                                 enable &= enableContinue;
798                         }
799
800                         if(shader->containsLeaveInstruction() && instruction->analysisLeave)
801                         {
802                                 enable &= enableLeave;
803                         }
804                 }
805
806                 return enable;
807         }
808
809         Vector4f PixelProgram::fetchRegister(const Src &src, unsigned int offset)
810         {
811                 Vector4f reg;
812                 unsigned int i = src.index + offset;
813
814                 switch(src.type)
815                 {
816                 case Shader::PARAMETER_TEMP:
817                         if(src.rel.type == Shader::PARAMETER_VOID)
818                         {
819                                 reg = r[i];
820                         }
821                         else
822                         {
823                                 Int a = relativeAddress(src, src.bufferIndex);
824
825                                 reg = r[i + a];
826                         }
827                         break;
828                 case Shader::PARAMETER_INPUT:
829                         {
830                                 if(src.rel.type == Shader::PARAMETER_VOID)   // Not relative
831                                 {
832                                         reg = v[i];
833                                 }
834                                 else
835                                 {
836                                         Int a = relativeAddress(src, src.bufferIndex);
837
838                                         reg = v[i + a];
839                                 }
840                         }
841                         break;
842                 case Shader::PARAMETER_CONST:
843                         reg = readConstant(src, offset);
844                         break;
845                 case Shader::PARAMETER_TEXTURE:
846                         reg = v[2 + i];
847                         break;
848                 case Shader::PARAMETER_MISCTYPE:
849                         if(src.index == Shader::VPosIndex) reg = vPos;
850                         if(src.index == Shader::VFaceIndex) reg = vFace;
851                         break;
852                 case Shader::PARAMETER_SAMPLER:
853                         if(src.rel.type == Shader::PARAMETER_VOID)
854                         {
855                                 reg.x = As<Float4>(Int4(i));
856                         }
857                         else if(src.rel.type == Shader::PARAMETER_TEMP)
858                         {
859                                 reg.x = As<Float4>(Int4(i) + As<Int4>(r[src.rel.index].x));
860                         }
861                         return reg;
862                 case Shader::PARAMETER_PREDICATE:   return reg; // Dummy
863                 case Shader::PARAMETER_VOID:        return reg; // Dummy
864                 case Shader::PARAMETER_FLOAT4LITERAL:
865                         reg.x = Float4(src.value[0]);
866                         reg.y = Float4(src.value[1]);
867                         reg.z = Float4(src.value[2]);
868                         reg.w = Float4(src.value[3]);
869                         break;
870                 case Shader::PARAMETER_CONSTINT:    return reg; // Dummy
871                 case Shader::PARAMETER_CONSTBOOL:   return reg; // Dummy
872                 case Shader::PARAMETER_LOOP:        return reg; // Dummy
873                 case Shader::PARAMETER_COLOROUT:
874                         if(src.rel.type == Shader::PARAMETER_VOID)   // Not relative
875                         {
876                                 reg = oC[i];
877                         }
878                         else
879                         {
880                                 Int a = relativeAddress(src, src.bufferIndex);
881
882                                 reg = oC[i + a];
883                         }
884                         break;
885                 case Shader::PARAMETER_DEPTHOUT:
886                         reg.x = oDepth;
887                         break;
888                 default:
889                         ASSERT(false);
890                 }
891
892                 const Float4 &x = reg[(src.swizzle >> 0) & 0x3];
893                 const Float4 &y = reg[(src.swizzle >> 2) & 0x3];
894                 const Float4 &z = reg[(src.swizzle >> 4) & 0x3];
895                 const Float4 &w = reg[(src.swizzle >> 6) & 0x3];
896
897                 Vector4f mod;
898
899                 switch(src.modifier)
900                 {
901                 case Shader::MODIFIER_NONE:
902                         mod.x = x;
903                         mod.y = y;
904                         mod.z = z;
905                         mod.w = w;
906                         break;
907                 case Shader::MODIFIER_NEGATE:
908                         mod.x = -x;
909                         mod.y = -y;
910                         mod.z = -z;
911                         mod.w = -w;
912                         break;
913                 case Shader::MODIFIER_ABS:
914                         mod.x = Abs(x);
915                         mod.y = Abs(y);
916                         mod.z = Abs(z);
917                         mod.w = Abs(w);
918                         break;
919                 case Shader::MODIFIER_ABS_NEGATE:
920                         mod.x = -Abs(x);
921                         mod.y = -Abs(y);
922                         mod.z = -Abs(z);
923                         mod.w = -Abs(w);
924                         break;
925                 case Shader::MODIFIER_NOT:
926                         mod.x = As<Float4>(As<Int4>(x) ^ Int4(0xFFFFFFFF));
927                         mod.y = As<Float4>(As<Int4>(y) ^ Int4(0xFFFFFFFF));
928                         mod.z = As<Float4>(As<Int4>(z) ^ Int4(0xFFFFFFFF));
929                         mod.w = As<Float4>(As<Int4>(w) ^ Int4(0xFFFFFFFF));
930                         break;
931                 default:
932                         ASSERT(false);
933                 }
934
935                 return mod;
936         }
937
938         RValue<Pointer<Byte>> PixelProgram::uniformAddress(int bufferIndex, unsigned int index)
939         {
940                 if(bufferIndex == -1)
941                 {
942                         return data + OFFSET(DrawData, ps.c[index]);
943                 }
944                 else
945                 {
946                         return *Pointer<Pointer<Byte>>(data + OFFSET(DrawData, ps.u[bufferIndex])) + index;
947                 }
948         }
949
950         RValue<Pointer<Byte>> PixelProgram::uniformAddress(int bufferIndex, unsigned int index, Int& offset)
951         {
952                 return uniformAddress(bufferIndex, index) + offset * sizeof(float4);
953         }
954
955         Vector4f PixelProgram::readConstant(const Src &src, unsigned int offset)
956         {
957                 Vector4f c;
958                 unsigned int i = src.index + offset;
959
960                 if(src.rel.type == Shader::PARAMETER_VOID)   // Not relative
961                 {
962                         c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i));
963
964                         c.x = c.x.xxxx;
965                         c.y = c.y.yyyy;
966                         c.z = c.z.zzzz;
967                         c.w = c.w.wwww;
968
969                         if(shader->containsDefineInstruction())   // Constant may be known at compile time
970                         {
971                                 for(size_t j = 0; j < shader->getLength(); j++)
972                                 {
973                                         const Shader::Instruction &instruction = *shader->getInstruction(j);
974
975                                         if(instruction.opcode == Shader::OPCODE_DEF)
976                                         {
977                                                 if(instruction.dst.index == i)
978                                                 {
979                                                         c.x = Float4(instruction.src[0].value[0]);
980                                                         c.y = Float4(instruction.src[0].value[1]);
981                                                         c.z = Float4(instruction.src[0].value[2]);
982                                                         c.w = Float4(instruction.src[0].value[3]);
983
984                                                         break;
985                                                 }
986                                         }
987                                 }
988                         }
989                 }
990                 else if(src.rel.type == Shader::PARAMETER_LOOP)
991                 {
992                         Int loopCounter = aL[loopDepth];
993
994                         c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i, loopCounter));
995
996                         c.x = c.x.xxxx;
997                         c.y = c.y.yyyy;
998                         c.z = c.z.zzzz;
999                         c.w = c.w.wwww;
1000                 }
1001                 else
1002                 {
1003                         Int a = relativeAddress(src, src.bufferIndex);
1004
1005                         c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i, a));
1006
1007                         c.x = c.x.xxxx;
1008                         c.y = c.y.yyyy;
1009                         c.z = c.z.zzzz;
1010                         c.w = c.w.wwww;
1011                 }
1012
1013                 return c;
1014         }
1015
1016         Int PixelProgram::relativeAddress(const Shader::Parameter &var, int bufferIndex)
1017         {
1018                 ASSERT(var.rel.deterministic);
1019
1020                 if(var.rel.type == Shader::PARAMETER_TEMP)
1021                 {
1022                         return As<Int>(Extract(r[var.rel.index].x, 0)) * var.rel.scale;
1023                 }
1024                 else if(var.rel.type == Shader::PARAMETER_INPUT)
1025                 {
1026                         return As<Int>(Extract(v[var.rel.index].x, 0)) * var.rel.scale;
1027                 }
1028                 else if(var.rel.type == Shader::PARAMETER_OUTPUT)
1029                 {
1030                         return As<Int>(Extract(oC[var.rel.index].x, 0)) * var.rel.scale;
1031                 }
1032                 else if(var.rel.type == Shader::PARAMETER_CONST)
1033                 {
1034                         return *Pointer<Int>(uniformAddress(bufferIndex, var.rel.index)) * var.rel.scale;
1035                 }
1036                 else if(var.rel.type == Shader::PARAMETER_LOOP)
1037                 {
1038                         return aL[loopDepth];
1039                 }
1040                 else ASSERT(false);
1041
1042                 return 0;
1043         }
1044
1045         Float4 PixelProgram::linearToSRGB(const Float4 &x)   // Approximates x^(1.0/2.2)
1046         {
1047                 Float4 sqrtx = Rcp_pp(RcpSqrt_pp(x));
1048                 Float4 sRGB = sqrtx * Float4(1.14f) - x * Float4(0.14f);
1049
1050                 return Min(Max(sRGB, Float4(0.0f)), Float4(1.0f));
1051         }
1052
1053         void PixelProgram::M3X2(Vector4f &dst, Vector4f &src0, const Src &src1)
1054         {
1055                 Vector4f row0 = fetchRegister(src1, 0);
1056                 Vector4f row1 = fetchRegister(src1, 1);
1057
1058                 dst.x = dot3(src0, row0);
1059                 dst.y = dot3(src0, row1);
1060         }
1061
1062         void PixelProgram::M3X3(Vector4f &dst, Vector4f &src0, const Src &src1)
1063         {
1064                 Vector4f row0 = fetchRegister(src1, 0);
1065                 Vector4f row1 = fetchRegister(src1, 1);
1066                 Vector4f row2 = fetchRegister(src1, 2);
1067
1068                 dst.x = dot3(src0, row0);
1069                 dst.y = dot3(src0, row1);
1070                 dst.z = dot3(src0, row2);
1071         }
1072
1073         void PixelProgram::M3X4(Vector4f &dst, Vector4f &src0, const Src &src1)
1074         {
1075                 Vector4f row0 = fetchRegister(src1, 0);
1076                 Vector4f row1 = fetchRegister(src1, 1);
1077                 Vector4f row2 = fetchRegister(src1, 2);
1078                 Vector4f row3 = fetchRegister(src1, 3);
1079
1080                 dst.x = dot3(src0, row0);
1081                 dst.y = dot3(src0, row1);
1082                 dst.z = dot3(src0, row2);
1083                 dst.w = dot3(src0, row3);
1084         }
1085
1086         void PixelProgram::M4X3(Vector4f &dst, Vector4f &src0, const Src &src1)
1087         {
1088                 Vector4f row0 = fetchRegister(src1, 0);
1089                 Vector4f row1 = fetchRegister(src1, 1);
1090                 Vector4f row2 = fetchRegister(src1, 2);
1091
1092                 dst.x = dot4(src0, row0);
1093                 dst.y = dot4(src0, row1);
1094                 dst.z = dot4(src0, row2);
1095         }
1096
1097         void PixelProgram::M4X4(Vector4f &dst, Vector4f &src0, const Src &src1)
1098         {
1099                 Vector4f row0 = fetchRegister(src1, 0);
1100                 Vector4f row1 = fetchRegister(src1, 1);
1101                 Vector4f row2 = fetchRegister(src1, 2);
1102                 Vector4f row3 = fetchRegister(src1, 3);
1103
1104                 dst.x = dot4(src0, row0);
1105                 dst.y = dot4(src0, row1);
1106                 dst.z = dot4(src0, row2);
1107                 dst.w = dot4(src0, row3);
1108         }
1109
1110         void PixelProgram::TEXLD(Vector4f &dst, Vector4f &src0, const Src &src1, bool project, bool bias)
1111         {
1112                 if(project)
1113                 {
1114                         Vector4f proj;
1115                         Float4 rw = reciprocal(src0.w);
1116                         proj.x = src0.x * rw;
1117                         proj.y = src0.y * rw;
1118                         proj.z = src0.z * rw;
1119
1120                         sampleTexture(dst, src1, proj, src0, src0, src0, Implicit);
1121                 }
1122                 else
1123                 {
1124                         sampleTexture(dst, src1, src0, src0, src0, src0, bias ? Bias : Implicit);
1125                 }
1126         }
1127
1128         void PixelProgram::TEXOFFSET(Vector4f &dst, Vector4f &src0, const Src &src1, Vector4f &src2, bool bias)
1129         {
1130                 sampleTexture(dst, src1, src0, src0, src0, src2, {bias ? Bias : Implicit, Offset});
1131         }
1132
1133         void PixelProgram::TEXLDL(Vector4f &dst, Vector4f &src0, const Src &src1, Vector4f &offset, bool bias)
1134         {
1135                 sampleTexture(dst, src1, src0, src0, src0, offset, {Lod, Offset});
1136         }
1137
1138         void PixelProgram::TEXELFETCH(Vector4f &dst, Vector4f &src0, const Src& src1)
1139         {
1140                 sampleTexture(dst, src1, src0, src0, src0, src0, Fetch);
1141         }
1142
1143         void PixelProgram::TEXELFETCH(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &offset)
1144         {
1145                 sampleTexture(dst, src1, src0, src0, src0, offset, {Fetch, Offset});
1146         }
1147
1148         void PixelProgram::TEXGRAD(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3)
1149         {
1150                 sampleTexture(dst, src1, src0, src2, src3, src0, Grad);
1151         }
1152
1153         void PixelProgram::TEXGRAD(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3, Vector4f &offset)
1154         {
1155                 sampleTexture(dst, src1, src0, src2, src3, offset, {Grad, Offset});
1156         }
1157
1158         void PixelProgram::TEXLDD(Vector4f &dst, Vector4f &src0, const Src &src1, Vector4f &src2, Vector4f &src3)
1159         {
1160                 sampleTexture(dst, src1, src0, src2, src3, src0, Grad);
1161         }
1162
1163         void PixelProgram::TEXLDL(Vector4f &dst, Vector4f &src0, const Src &src1)
1164         {
1165                 sampleTexture(dst, src1, src0, src0, src0, src0, Lod);
1166         }
1167
1168         void PixelProgram::TEXSIZE(Vector4f &dst, Float4 &lod, const Src &src1)
1169         {
1170                 Pointer<Byte> texture = data + OFFSET(DrawData, mipmap) + src1.index * sizeof(Texture);
1171                 sampler[src1.index]->textureSize(texture, dst, lod);
1172         }
1173
1174         void PixelProgram::TEXKILL(Int cMask[4], Vector4f &src, unsigned char mask)
1175         {
1176                 Int kill = -1;
1177
1178                 if(mask & 0x1) kill &= SignMask(CmpNLT(src.x, Float4(0.0f)));
1179                 if(mask & 0x2) kill &= SignMask(CmpNLT(src.y, Float4(0.0f)));
1180                 if(mask & 0x4) kill &= SignMask(CmpNLT(src.z, Float4(0.0f)));
1181                 if(mask & 0x8) kill &= SignMask(CmpNLT(src.w, Float4(0.0f)));
1182
1183                 // FIXME: Dynamic branching affects TEXKILL?
1184                 //      if(shader->containsDynamicBranching())
1185                 //      {
1186                 //              kill = ~SignMask(enableMask());
1187                 //      }
1188
1189                 for(unsigned int q = 0; q < state.multiSample; q++)
1190                 {
1191                         cMask[q] &= kill;
1192                 }
1193
1194                 // FIXME: Branch to end of shader if all killed?
1195         }
1196
1197         void PixelProgram::DISCARD(Int cMask[4], const Shader::Instruction *instruction)
1198         {
1199                 Int kill = 0;
1200
1201                 if(shader->containsDynamicBranching())
1202                 {
1203                         kill = ~SignMask(enableMask(instruction));
1204                 }
1205
1206                 for(unsigned int q = 0; q < state.multiSample; q++)
1207                 {
1208                         cMask[q] &= kill;
1209                 }
1210
1211                 // FIXME: Branch to end of shader if all killed?
1212         }
1213
1214         void PixelProgram::DFDX(Vector4f &dst, Vector4f &src)
1215         {
1216                 dst.x = src.x.yyww - src.x.xxzz;
1217                 dst.y = src.y.yyww - src.y.xxzz;
1218                 dst.z = src.z.yyww - src.z.xxzz;
1219                 dst.w = src.w.yyww - src.w.xxzz;
1220         }
1221
1222         void PixelProgram::DFDY(Vector4f &dst, Vector4f &src)
1223         {
1224                 dst.x = src.x.zwzw - src.x.xyxy;
1225                 dst.y = src.y.zwzw - src.y.xyxy;
1226                 dst.z = src.z.zwzw - src.z.xyxy;
1227                 dst.w = src.w.zwzw - src.w.xyxy;
1228         }
1229
1230         void PixelProgram::FWIDTH(Vector4f &dst, Vector4f &src)
1231         {
1232                 // abs(dFdx(src)) + abs(dFdy(src));
1233                 dst.x = Abs(src.x.yyww - src.x.xxzz) + Abs(src.x.zwzw - src.x.xyxy);
1234                 dst.y = Abs(src.y.yyww - src.y.xxzz) + Abs(src.y.zwzw - src.y.xyxy);
1235                 dst.z = Abs(src.z.yyww - src.z.xxzz) + Abs(src.z.zwzw - src.z.xyxy);
1236                 dst.w = Abs(src.w.yyww - src.w.xxzz) + Abs(src.w.zwzw - src.w.xyxy);
1237         }
1238
1239         void PixelProgram::BREAK()
1240         {
1241                 BasicBlock *deadBlock = Nucleus::createBasicBlock();
1242                 BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
1243
1244                 if(breakDepth == 0)
1245                 {
1246                         enableIndex = enableIndex - breakDepth;
1247                         Nucleus::createBr(endBlock);
1248                 }
1249                 else
1250                 {
1251                         enableBreak = enableBreak & ~enableStack[enableIndex];
1252                         Bool allBreak = SignMask(enableBreak) == 0x0;
1253
1254                         enableIndex = enableIndex - breakDepth;
1255                         branch(allBreak, endBlock, deadBlock);
1256                 }
1257
1258                 Nucleus::setInsertBlock(deadBlock);
1259                 enableIndex = enableIndex + breakDepth;
1260         }
1261
1262         void PixelProgram::BREAKC(Vector4f &src0, Vector4f &src1, Control control)
1263         {
1264                 Int4 condition;
1265
1266                 switch(control)
1267                 {
1268                 case Shader::CONTROL_GT: condition = CmpNLE(src0.x, src1.x); break;
1269                 case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x);  break;
1270                 case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x); break;
1271                 case Shader::CONTROL_LT: condition = CmpLT(src0.x, src1.x);  break;
1272                 case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x); break;
1273                 case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x);  break;
1274                 default:
1275                         ASSERT(false);
1276                 }
1277
1278                 BREAK(condition);
1279         }
1280
1281         void PixelProgram::BREAKP(const Src &predicateRegister)   // FIXME: Factor out parts common with BREAKC
1282         {
1283                 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
1284
1285                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1286                 {
1287                         condition = ~condition;
1288                 }
1289
1290                 BREAK(condition);
1291         }
1292
1293         void PixelProgram::BREAK(Int4 &condition)
1294         {
1295                 condition &= enableStack[enableIndex];
1296
1297                 BasicBlock *continueBlock = Nucleus::createBasicBlock();
1298                 BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
1299
1300                 enableBreak = enableBreak & ~condition;
1301                 Bool allBreak = SignMask(enableBreak) == 0x0;
1302
1303                 enableIndex = enableIndex - breakDepth;
1304                 branch(allBreak, endBlock, continueBlock);
1305
1306                 Nucleus::setInsertBlock(continueBlock);
1307                 enableIndex = enableIndex + breakDepth;
1308         }
1309
1310         void PixelProgram::CONTINUE()
1311         {
1312                 enableContinue = enableContinue & ~enableStack[enableIndex];
1313         }
1314
1315         void PixelProgram::TEST()
1316         {
1317                 whileTest = true;
1318         }
1319
1320         void PixelProgram::CALL(int labelIndex, int callSiteIndex)
1321         {
1322                 if(!labelBlock[labelIndex])
1323                 {
1324                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1325                 }
1326
1327                 if(callRetBlock[labelIndex].size() > 1)
1328                 {
1329                         callStack[stackIndex++] = UInt(callSiteIndex);
1330                 }
1331
1332                 Int4 restoreLeave = enableLeave;
1333
1334                 Nucleus::createBr(labelBlock[labelIndex]);
1335                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1336
1337                 enableLeave = restoreLeave;
1338         }
1339
1340         void PixelProgram::CALLNZ(int labelIndex, int callSiteIndex, const Src &src)
1341         {
1342                 if(src.type == Shader::PARAMETER_CONSTBOOL)
1343                 {
1344                         CALLNZb(labelIndex, callSiteIndex, src);
1345                 }
1346                 else if(src.type == Shader::PARAMETER_PREDICATE)
1347                 {
1348                         CALLNZp(labelIndex, callSiteIndex, src);
1349                 }
1350                 else ASSERT(false);
1351         }
1352
1353         void PixelProgram::CALLNZb(int labelIndex, int callSiteIndex, const Src &boolRegister)
1354         {
1355                 Bool condition = (*Pointer<Byte>(data + OFFSET(DrawData, ps.b[boolRegister.index])) != Byte(0));   // FIXME
1356
1357                 if(boolRegister.modifier == Shader::MODIFIER_NOT)
1358                 {
1359                         condition = !condition;
1360                 }
1361
1362                 if(!labelBlock[labelIndex])
1363                 {
1364                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1365                 }
1366
1367                 if(callRetBlock[labelIndex].size() > 1)
1368                 {
1369                         callStack[stackIndex++] = UInt(callSiteIndex);
1370                 }
1371
1372                 Int4 restoreLeave = enableLeave;
1373
1374                 branch(condition, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]);
1375                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1376
1377                 enableLeave = restoreLeave;
1378         }
1379
1380         void PixelProgram::CALLNZp(int labelIndex, int callSiteIndex, const Src &predicateRegister)
1381         {
1382                 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
1383
1384                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1385                 {
1386                         condition = ~condition;
1387                 }
1388
1389                 condition &= enableStack[enableIndex];
1390
1391                 if(!labelBlock[labelIndex])
1392                 {
1393                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1394                 }
1395
1396                 if(callRetBlock[labelIndex].size() > 1)
1397                 {
1398                         callStack[stackIndex++] = UInt(callSiteIndex);
1399                 }
1400
1401                 enableIndex++;
1402                 enableStack[enableIndex] = condition;
1403                 Int4 restoreLeave = enableLeave;
1404
1405                 Bool notAllFalse = SignMask(condition) != 0;
1406                 branch(notAllFalse, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]);
1407                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1408
1409                 enableIndex--;
1410                 enableLeave = restoreLeave;
1411         }
1412
1413         void PixelProgram::ELSE()
1414         {
1415                 ifDepth--;
1416
1417                 BasicBlock *falseBlock = ifFalseBlock[ifDepth];
1418                 BasicBlock *endBlock = Nucleus::createBasicBlock();
1419
1420                 if(isConditionalIf[ifDepth])
1421                 {
1422                         Int4 condition = ~enableStack[enableIndex] & enableStack[enableIndex - 1];
1423                         Bool notAllFalse = SignMask(condition) != 0;
1424
1425                         branch(notAllFalse, falseBlock, endBlock);
1426
1427                         enableStack[enableIndex] = ~enableStack[enableIndex] & enableStack[enableIndex - 1];
1428                 }
1429                 else
1430                 {
1431                         Nucleus::createBr(endBlock);
1432                         Nucleus::setInsertBlock(falseBlock);
1433                 }
1434
1435                 ifFalseBlock[ifDepth] = endBlock;
1436
1437                 ifDepth++;
1438         }
1439
1440         void PixelProgram::ENDIF()
1441         {
1442                 ifDepth--;
1443
1444                 BasicBlock *endBlock = ifFalseBlock[ifDepth];
1445
1446                 Nucleus::createBr(endBlock);
1447                 Nucleus::setInsertBlock(endBlock);
1448
1449                 if(isConditionalIf[ifDepth])
1450                 {
1451                         breakDepth--;
1452                         enableIndex--;
1453                 }
1454         }
1455
1456         void PixelProgram::ENDLOOP()
1457         {
1458                 loopRepDepth--;
1459
1460                 aL[loopDepth] = aL[loopDepth] + increment[loopDepth];   // FIXME: +=
1461
1462                 BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1463                 BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1464
1465                 Nucleus::createBr(testBlock);
1466                 Nucleus::setInsertBlock(endBlock);
1467
1468                 loopDepth--;
1469                 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1470         }
1471
1472         void PixelProgram::ENDREP()
1473         {
1474                 loopRepDepth--;
1475
1476                 BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1477                 BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1478
1479                 Nucleus::createBr(testBlock);
1480                 Nucleus::setInsertBlock(endBlock);
1481
1482                 loopDepth--;
1483                 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1484         }
1485
1486         void PixelProgram::ENDWHILE()
1487         {
1488                 loopRepDepth--;
1489
1490                 BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1491                 BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1492
1493                 Nucleus::createBr(testBlock);
1494                 Nucleus::setInsertBlock(endBlock);
1495
1496                 enableIndex--;
1497                 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1498                 whileTest = false;
1499         }
1500
1501         void PixelProgram::ENDSWITCH()
1502         {
1503                 loopRepDepth--;
1504
1505                 BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1506
1507                 Nucleus::createBr(loopRepEndBlock[loopRepDepth]);
1508                 Nucleus::setInsertBlock(endBlock);
1509
1510                 enableIndex--;
1511                 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1512         }
1513
1514         void PixelProgram::IF(const Src &src)
1515         {
1516                 if(src.type == Shader::PARAMETER_CONSTBOOL)
1517                 {
1518                         IFb(src);
1519                 }
1520                 else if(src.type == Shader::PARAMETER_PREDICATE)
1521                 {
1522                         IFp(src);
1523                 }
1524                 else
1525                 {
1526                         Int4 condition = As<Int4>(fetchRegister(src).x);
1527                         IF(condition);
1528                 }
1529         }
1530
1531         void PixelProgram::IFb(const Src &boolRegister)
1532         {
1533                 ASSERT(ifDepth < 24 + 4);
1534
1535                 Bool condition = (*Pointer<Byte>(data + OFFSET(DrawData, ps.b[boolRegister.index])) != Byte(0));   // FIXME
1536
1537                 if(boolRegister.modifier == Shader::MODIFIER_NOT)
1538                 {
1539                         condition = !condition;
1540                 }
1541
1542                 BasicBlock *trueBlock = Nucleus::createBasicBlock();
1543                 BasicBlock *falseBlock = Nucleus::createBasicBlock();
1544
1545                 branch(condition, trueBlock, falseBlock);
1546
1547                 isConditionalIf[ifDepth] = false;
1548                 ifFalseBlock[ifDepth] = falseBlock;
1549
1550                 ifDepth++;
1551         }
1552
1553         void PixelProgram::IFp(const Src &predicateRegister)
1554         {
1555                 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
1556
1557                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1558                 {
1559                         condition = ~condition;
1560                 }
1561
1562                 IF(condition);
1563         }
1564
1565         void PixelProgram::IFC(Vector4f &src0, Vector4f &src1, Control control)
1566         {
1567                 Int4 condition;
1568
1569                 switch(control)
1570                 {
1571                 case Shader::CONTROL_GT: condition = CmpNLE(src0.x, src1.x); break;
1572                 case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x);  break;
1573                 case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x); break;
1574                 case Shader::CONTROL_LT: condition = CmpLT(src0.x, src1.x);  break;
1575                 case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x); break;
1576                 case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x);  break;
1577                 default:
1578                         ASSERT(false);
1579                 }
1580
1581                 IF(condition);
1582         }
1583
1584         void PixelProgram::IF(Int4 &condition)
1585         {
1586                 condition &= enableStack[enableIndex];
1587
1588                 enableIndex++;
1589                 enableStack[enableIndex] = condition;
1590
1591                 BasicBlock *trueBlock = Nucleus::createBasicBlock();
1592                 BasicBlock *falseBlock = Nucleus::createBasicBlock();
1593
1594                 Bool notAllFalse = SignMask(condition) != 0;
1595
1596                 branch(notAllFalse, trueBlock, falseBlock);
1597
1598                 isConditionalIf[ifDepth] = true;
1599                 ifFalseBlock[ifDepth] = falseBlock;
1600
1601                 ifDepth++;
1602                 breakDepth++;
1603         }
1604
1605         void PixelProgram::LABEL(int labelIndex)
1606         {
1607                 if(!labelBlock[labelIndex])
1608                 {
1609                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1610                 }
1611
1612                 Nucleus::setInsertBlock(labelBlock[labelIndex]);
1613                 currentLabel = labelIndex;
1614         }
1615
1616         void PixelProgram::LOOP(const Src &integerRegister)
1617         {
1618                 loopDepth++;
1619
1620                 iteration[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][0]));
1621                 aL[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][1]));
1622                 increment[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][2]));
1623
1624                 //      If(increment[loopDepth] == 0)
1625                 //      {
1626                 //              increment[loopDepth] = 1;
1627                 //      }
1628
1629                 BasicBlock *loopBlock = Nucleus::createBasicBlock();
1630                 BasicBlock *testBlock = Nucleus::createBasicBlock();
1631                 BasicBlock *endBlock = Nucleus::createBasicBlock();
1632
1633                 loopRepTestBlock[loopRepDepth] = testBlock;
1634                 loopRepEndBlock[loopRepDepth] = endBlock;
1635
1636                 // FIXME: jump(testBlock)
1637                 Nucleus::createBr(testBlock);
1638                 Nucleus::setInsertBlock(testBlock);
1639
1640                 branch(iteration[loopDepth] > 0, loopBlock, endBlock);
1641                 Nucleus::setInsertBlock(loopBlock);
1642
1643                 iteration[loopDepth] = iteration[loopDepth] - 1;   // FIXME: --
1644
1645                 loopRepDepth++;
1646                 breakDepth = 0;
1647         }
1648
1649         void PixelProgram::REP(const Src &integerRegister)
1650         {
1651                 loopDepth++;
1652
1653                 iteration[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][0]));
1654                 aL[loopDepth] = aL[loopDepth - 1];
1655
1656                 BasicBlock *loopBlock = Nucleus::createBasicBlock();
1657                 BasicBlock *testBlock = Nucleus::createBasicBlock();
1658                 BasicBlock *endBlock = Nucleus::createBasicBlock();
1659
1660                 loopRepTestBlock[loopRepDepth] = testBlock;
1661                 loopRepEndBlock[loopRepDepth] = endBlock;
1662
1663                 // FIXME: jump(testBlock)
1664                 Nucleus::createBr(testBlock);
1665                 Nucleus::setInsertBlock(testBlock);
1666
1667                 branch(iteration[loopDepth] > 0, loopBlock, endBlock);
1668                 Nucleus::setInsertBlock(loopBlock);
1669
1670                 iteration[loopDepth] = iteration[loopDepth] - 1;   // FIXME: --
1671
1672                 loopRepDepth++;
1673                 breakDepth = 0;
1674         }
1675
1676         void PixelProgram::WHILE(const Src &temporaryRegister)
1677         {
1678                 enableIndex++;
1679
1680                 BasicBlock *loopBlock = Nucleus::createBasicBlock();
1681                 BasicBlock *testBlock = Nucleus::createBasicBlock();
1682                 BasicBlock *endBlock = Nucleus::createBasicBlock();
1683
1684                 loopRepTestBlock[loopRepDepth] = testBlock;
1685                 loopRepEndBlock[loopRepDepth] = endBlock;
1686
1687                 Int4 restoreBreak = enableBreak;
1688                 Int4 restoreContinue = enableContinue;
1689
1690                 // FIXME: jump(testBlock)
1691                 Nucleus::createBr(testBlock);
1692                 Nucleus::setInsertBlock(testBlock);
1693                 enableContinue = restoreContinue;
1694
1695                 const Vector4f &src = fetchRegister(temporaryRegister);
1696                 Int4 condition = As<Int4>(src.x);
1697                 condition &= enableStack[enableIndex - 1];
1698                 if(shader->containsLeaveInstruction()) condition &= enableLeave;
1699                 enableStack[enableIndex] = condition;
1700
1701                 Bool notAllFalse = SignMask(condition) != 0;
1702                 branch(notAllFalse, loopBlock, endBlock);
1703
1704                 Nucleus::setInsertBlock(endBlock);
1705                 enableBreak = restoreBreak;
1706
1707                 Nucleus::setInsertBlock(loopBlock);
1708
1709                 loopRepDepth++;
1710                 breakDepth = 0;
1711         }
1712
1713         void PixelProgram::SWITCH()
1714         {
1715                 enableIndex++;
1716                 enableStack[enableIndex] = Int4(0xFFFFFFFF);
1717
1718                 BasicBlock *endBlock = Nucleus::createBasicBlock();
1719
1720                 loopRepTestBlock[loopRepDepth] = nullptr;
1721                 loopRepEndBlock[loopRepDepth] = endBlock;
1722
1723                 loopRepDepth++;
1724                 breakDepth = 0;
1725         }
1726
1727         void PixelProgram::RET()
1728         {
1729                 if(currentLabel == -1)
1730                 {
1731                         returnBlock = Nucleus::createBasicBlock();
1732                         Nucleus::createBr(returnBlock);
1733                 }
1734                 else
1735                 {
1736                         BasicBlock *unreachableBlock = Nucleus::createBasicBlock();
1737
1738                         if(callRetBlock[currentLabel].size() > 1)   // Pop the return destination from the call stack
1739                         {
1740                                 // FIXME: Encapsulate
1741                                 UInt index = callStack[--stackIndex];
1742
1743                                 Value *value = index.loadValue();
1744                                 SwitchCases *switchCases = Nucleus::createSwitch(value, unreachableBlock, (int)callRetBlock[currentLabel].size());
1745
1746                                 for(unsigned int i = 0; i < callRetBlock[currentLabel].size(); i++)
1747                                 {
1748                                         Nucleus::addSwitchCase(switchCases, i, callRetBlock[currentLabel][i]);
1749                                 }
1750                         }
1751                         else if(callRetBlock[currentLabel].size() == 1)   // Jump directly to the unique return destination
1752                         {
1753                                 Nucleus::createBr(callRetBlock[currentLabel][0]);
1754                         }
1755                         else   // Function isn't called
1756                         {
1757                                 Nucleus::createBr(unreachableBlock);
1758                         }
1759
1760                         Nucleus::setInsertBlock(unreachableBlock);
1761                         Nucleus::createUnreachable();
1762                 }
1763         }
1764
1765         void PixelProgram::LEAVE()
1766         {
1767                 enableLeave = enableLeave & ~enableStack[enableIndex];
1768
1769                 // FIXME: Return from function if all instances left
1770                 // FIXME: Use enableLeave in other control-flow constructs
1771         }
1772 }