blob: 0286bc30a8c6f0cc46a9d83285f730c914c4c63d [file] [log] [blame]
Nicolas Capens5d961882016-01-01 23:18:14 -05001// SwiftShader Software Renderer
2//
3// Copyright(c) 2015 Google Inc.
4//
5// All rights reserved. No part of this software may be copied, distributed, transmitted,
6// transcribed, stored in a retrieval system, translated into any human or computer
7// language by any means, or disclosed to third parties without the explicit written
8// agreement of Google Inc. Without such an agreement, no rights or licenses, express
9// or implied, including but not limited to any patent rights, are granted to you.
10//
11
12#include "PixelProgram.hpp"
13#include "Primitive.hpp"
14#include "Renderer.hpp"
15#include "SamplerCore.hpp"
16
17namespace sw
18{
19 extern bool postBlendSRGB;
Alexis Hetuf2a8c372015-07-13 11:08:41 -040020 extern bool booleanFaceRegister;
21 extern bool halfIntegerCoordinates; // Pixel centers are not at integer coordinates
Nicolas Capens5d961882016-01-01 23:18:14 -050022 extern bool fullPixelPositionRegister;
23
Nicolas Capens4f172c72016-01-13 08:34:30 -050024 void PixelProgram::setBuiltins(Int &x, Int &y, Float4(&z)[4], Float4 &w)
Nicolas Capens5d961882016-01-01 23:18:14 -050025 {
Alexis Hetuf2a8c372015-07-13 11:08:41 -040026 if(shader->getVersion() >= 0x0300)
27 {
28 if(shader->vPosDeclared)
29 {
30 if(!halfIntegerCoordinates)
31 {
Nicolas Capens4f172c72016-01-13 08:34:30 -050032 vPos.x = Float4(Float(x)) + Float4(0, 1, 0, 1);
33 vPos.y = Float4(Float(y)) + Float4(0, 0, 1, 1);
Alexis Hetuf2a8c372015-07-13 11:08:41 -040034 }
35 else
36 {
Nicolas Capens4f172c72016-01-13 08:34:30 -050037 vPos.x = Float4(Float(x)) + Float4(0.5f, 1.5f, 0.5f, 1.5f);
38 vPos.y = Float4(Float(y)) + Float4(0.5f, 0.5f, 1.5f, 1.5f);
Alexis Hetuf2a8c372015-07-13 11:08:41 -040039 }
40
41 if(fullPixelPositionRegister)
42 {
Nicolas Capens4f172c72016-01-13 08:34:30 -050043 vPos.z = z[0]; // FIXME: Centroid?
44 vPos.w = w; // FIXME: Centroid?
Alexis Hetuf2a8c372015-07-13 11:08:41 -040045 }
46 }
47
48 if(shader->vFaceDeclared)
49 {
Nicolas Capens4f172c72016-01-13 08:34:30 -050050 Float4 area = *Pointer<Float>(primitive + OFFSET(Primitive, area));
Alexis Hetuf2a8c372015-07-13 11:08:41 -040051 Float4 face = booleanFaceRegister ? Float4(As<Float4>(CmpNLT(area, Float4(0.0f)))) : area;
52
Nicolas Capens4f172c72016-01-13 08:34:30 -050053 vFace.x = face;
54 vFace.y = face;
55 vFace.z = face;
56 vFace.w = face;
Alexis Hetuf2a8c372015-07-13 11:08:41 -040057 }
Nicolas Capens5d961882016-01-01 23:18:14 -050058 }
59 }
60
Nicolas Capens4f172c72016-01-13 08:34:30 -050061 void PixelProgram::applyShader(Int cMask[4])
Nicolas Capens5d961882016-01-01 23:18:14 -050062 {
Nicolas Capens4f172c72016-01-13 08:34:30 -050063 enableIndex = 0;
64 stackIndex = 0;
Alexis Hetuf2a8c372015-07-13 11:08:41 -040065
66 if(shader->containsLeaveInstruction())
67 {
Nicolas Capens4f172c72016-01-13 08:34:30 -050068 enableLeave = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
Alexis Hetuf2a8c372015-07-13 11:08:41 -040069 }
70
Alexis Hetu1edcd8b2015-11-05 11:12:41 -050071 for(int i = 0; i < RENDERTARGETS; i++)
72 {
73 if(state.targetFormat[i] != FORMAT_NULL)
74 {
Nicolas Capens4f172c72016-01-13 08:34:30 -050075 oC[i] = Vector4f(0.0f, 0.0f, 0.0f, 0.0f);
Alexis Hetu1edcd8b2015-11-05 11:12:41 -050076 }
77 }
Alexis Hetuf2a8c372015-07-13 11:08:41 -040078
79 // Create all call site return blocks up front
80 for(size_t i = 0; i < shader->getLength(); i++)
81 {
82 const Shader::Instruction *instruction = shader->getInstruction(i);
83 Shader::Opcode opcode = instruction->opcode;
84
85 if(opcode == Shader::OPCODE_CALL || opcode == Shader::OPCODE_CALLNZ)
86 {
87 const Dst &dst = instruction->dst;
88
89 ASSERT(callRetBlock[dst.label].size() == dst.callSite);
90 callRetBlock[dst.label].push_back(Nucleus::createBasicBlock());
91 }
92 }
93
94 for(size_t i = 0; i < shader->getLength(); i++)
95 {
96 const Shader::Instruction *instruction = shader->getInstruction(i);
97 Shader::Opcode opcode = instruction->opcode;
98
99 if(opcode == Shader::OPCODE_DCL || opcode == Shader::OPCODE_DEF || opcode == Shader::OPCODE_DEFI || opcode == Shader::OPCODE_DEFB)
100 {
101 continue;
102 }
103
104 const Dst &dst = instruction->dst;
105 const Src &src0 = instruction->src[0];
106 const Src &src1 = instruction->src[1];
107 const Src &src2 = instruction->src[2];
108 const Src &src3 = instruction->src[3];
Alexis Hetu25d47fc2015-10-22 13:58:32 -0400109 const Src &src4 = instruction->src[4];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400110
111 bool predicate = instruction->predicate;
112 Control control = instruction->control;
113 bool pp = dst.partialPrecision;
114 bool project = instruction->project;
115 bool bias = instruction->bias;
116
117 Vector4f d;
118 Vector4f s0;
119 Vector4f s1;
120 Vector4f s2;
121 Vector4f s3;
Alexis Hetu25d47fc2015-10-22 13:58:32 -0400122 Vector4f s4;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400123
124 if(opcode == Shader::OPCODE_TEXKILL) // Takes destination as input
125 {
126 if(dst.type == Shader::PARAMETER_TEXTURE)
127 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500128 d.x = v[2 + dst.index].x;
129 d.y = v[2 + dst.index].y;
130 d.z = v[2 + dst.index].z;
131 d.w = v[2 + dst.index].w;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400132 }
133 else
134 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500135 d = r[dst.index];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400136 }
137 }
138
Nicolas Capens4f172c72016-01-13 08:34:30 -0500139 if(src0.type != Shader::PARAMETER_VOID) s0 = fetchRegisterF(src0);
140 if(src1.type != Shader::PARAMETER_VOID) s1 = fetchRegisterF(src1);
141 if(src2.type != Shader::PARAMETER_VOID) s2 = fetchRegisterF(src2);
142 if(src3.type != Shader::PARAMETER_VOID) s3 = fetchRegisterF(src3);
143 if(src4.type != Shader::PARAMETER_VOID) s4 = fetchRegisterF(src4);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400144
145 switch(opcode)
146 {
147 case Shader::OPCODE_PS_2_0: break;
148 case Shader::OPCODE_PS_2_x: break;
149 case Shader::OPCODE_PS_3_0: break;
150 case Shader::OPCODE_DEF: break;
151 case Shader::OPCODE_DCL: break;
152 case Shader::OPCODE_NOP: break;
153 case Shader::OPCODE_MOV: mov(d, s0); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400154 case Shader::OPCODE_NEG: neg(d, s0); break;
155 case Shader::OPCODE_INEG: ineg(d, s0); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400156 case Shader::OPCODE_F2B: f2b(d, s0); break;
157 case Shader::OPCODE_B2F: b2f(d, s0); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400158 case Shader::OPCODE_F2I: f2i(d, s0); break;
159 case Shader::OPCODE_I2F: i2f(d, s0); break;
160 case Shader::OPCODE_F2U: f2u(d, s0); break;
161 case Shader::OPCODE_U2F: u2f(d, s0); break;
162 case Shader::OPCODE_I2B: i2b(d, s0); break;
163 case Shader::OPCODE_B2I: b2i(d, s0); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400164 case Shader::OPCODE_ADD: add(d, s0, s1); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400165 case Shader::OPCODE_IADD: iadd(d, s0, s1); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400166 case Shader::OPCODE_SUB: sub(d, s0, s1); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400167 case Shader::OPCODE_ISUB: isub(d, s0, s1); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400168 case Shader::OPCODE_MUL: mul(d, s0, s1); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400169 case Shader::OPCODE_IMUL: imul(d, s0, s1); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400170 case Shader::OPCODE_MAD: mad(d, s0, s1, s2); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400171 case Shader::OPCODE_IMAD: imad(d, s0, s1, s2); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400172 case Shader::OPCODE_DP1: dp1(d, s0, s1); break;
173 case Shader::OPCODE_DP2: dp2(d, s0, s1); break;
174 case Shader::OPCODE_DP2ADD: dp2add(d, s0, s1, s2); break;
175 case Shader::OPCODE_DP3: dp3(d, s0, s1); break;
176 case Shader::OPCODE_DP4: dp4(d, s0, s1); break;
Alexis Hetuc3d95f32015-09-23 12:27:32 -0400177 case Shader::OPCODE_DET2: det2(d, s0, s1); break;
178 case Shader::OPCODE_DET3: det3(d, s0, s1, s2); break;
179 case Shader::OPCODE_DET4: det4(d, s0, s1, s2, s3); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400180 case Shader::OPCODE_CMP0: cmp0(d, s0, s1, s2); break;
181 case Shader::OPCODE_ICMP: icmp(d, s0, s1, control); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400182 case Shader::OPCODE_UCMP: ucmp(d, s0, s1, control); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400183 case Shader::OPCODE_SELECT: select(d, s0, s1, s2); break;
184 case Shader::OPCODE_EXTRACT: extract(d.x, s0, s1.x); break;
185 case Shader::OPCODE_INSERT: insert(d, s0, s1.x, s2.x); break;
186 case Shader::OPCODE_FRC: frc(d, s0); break;
187 case Shader::OPCODE_TRUNC: trunc(d, s0); break;
188 case Shader::OPCODE_FLOOR: floor(d, s0); break;
189 case Shader::OPCODE_ROUND: round(d, s0); break;
190 case Shader::OPCODE_ROUNDEVEN: roundEven(d, s0); break;
191 case Shader::OPCODE_CEIL: ceil(d, s0); break;
192 case Shader::OPCODE_EXP2X: exp2x(d, s0, pp); break;
193 case Shader::OPCODE_EXP2: exp2(d, s0, pp); break;
194 case Shader::OPCODE_LOG2X: log2x(d, s0, pp); break;
195 case Shader::OPCODE_LOG2: log2(d, s0, pp); break;
196 case Shader::OPCODE_EXP: exp(d, s0, pp); break;
197 case Shader::OPCODE_LOG: log(d, s0, pp); break;
198 case Shader::OPCODE_RCPX: rcpx(d, s0, pp); break;
199 case Shader::OPCODE_DIV: div(d, s0, s1); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400200 case Shader::OPCODE_IDIV: idiv(d, s0, s1); break;
201 case Shader::OPCODE_UDIV: udiv(d, s0, s1); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400202 case Shader::OPCODE_MOD: mod(d, s0, s1); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400203 case Shader::OPCODE_IMOD: imod(d, s0, s1); break;
204 case Shader::OPCODE_UMOD: umod(d, s0, s1); break;
205 case Shader::OPCODE_SHL: shl(d, s0, s1); break;
206 case Shader::OPCODE_ISHR: ishr(d, s0, s1); break;
207 case Shader::OPCODE_USHR: ushr(d, s0, s1); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400208 case Shader::OPCODE_RSQX: rsqx(d, s0, pp); break;
209 case Shader::OPCODE_SQRT: sqrt(d, s0, pp); break;
210 case Shader::OPCODE_RSQ: rsq(d, s0, pp); break;
211 case Shader::OPCODE_LEN2: len2(d.x, s0, pp); break;
212 case Shader::OPCODE_LEN3: len3(d.x, s0, pp); break;
213 case Shader::OPCODE_LEN4: len4(d.x, s0, pp); break;
214 case Shader::OPCODE_DIST1: dist1(d.x, s0, s1, pp); break;
215 case Shader::OPCODE_DIST2: dist2(d.x, s0, s1, pp); break;
216 case Shader::OPCODE_DIST3: dist3(d.x, s0, s1, pp); break;
217 case Shader::OPCODE_DIST4: dist4(d.x, s0, s1, pp); break;
218 case Shader::OPCODE_MIN: min(d, s0, s1); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400219 case Shader::OPCODE_IMIN: imin(d, s0, s1); break;
220 case Shader::OPCODE_UMIN: umin(d, s0, s1); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400221 case Shader::OPCODE_MAX: max(d, s0, s1); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400222 case Shader::OPCODE_IMAX: imax(d, s0, s1); break;
223 case Shader::OPCODE_UMAX: umax(d, s0, s1); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400224 case Shader::OPCODE_LRP: lrp(d, s0, s1, s2); break;
225 case Shader::OPCODE_STEP: step(d, s0, s1); break;
226 case Shader::OPCODE_SMOOTH: smooth(d, s0, s1, s2); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400227 case Shader::OPCODE_FLOATBITSTOINT:
228 case Shader::OPCODE_FLOATBITSTOUINT:
229 case Shader::OPCODE_INTBITSTOFLOAT:
230 case Shader::OPCODE_UINTBITSTOFLOAT: d = s0; break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400231 case Shader::OPCODE_POWX: powx(d, s0, s1, pp); break;
232 case Shader::OPCODE_POW: pow(d, s0, s1, pp); break;
233 case Shader::OPCODE_SGN: sgn(d, s0); break;
Alexis Hetu0f448072016-03-18 10:56:08 -0400234 case Shader::OPCODE_ISGN: isgn(d, s0); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400235 case Shader::OPCODE_CRS: crs(d, s0, s1); break;
236 case Shader::OPCODE_FORWARD1: forward1(d, s0, s1, s2); break;
237 case Shader::OPCODE_FORWARD2: forward2(d, s0, s1, s2); break;
238 case Shader::OPCODE_FORWARD3: forward3(d, s0, s1, s2); break;
239 case Shader::OPCODE_FORWARD4: forward4(d, s0, s1, s2); break;
240 case Shader::OPCODE_REFLECT1: reflect1(d, s0, s1); break;
241 case Shader::OPCODE_REFLECT2: reflect2(d, s0, s1); break;
242 case Shader::OPCODE_REFLECT3: reflect3(d, s0, s1); break;
243 case Shader::OPCODE_REFLECT4: reflect4(d, s0, s1); break;
244 case Shader::OPCODE_REFRACT1: refract1(d, s0, s1, s2.x); break;
245 case Shader::OPCODE_REFRACT2: refract2(d, s0, s1, s2.x); break;
246 case Shader::OPCODE_REFRACT3: refract3(d, s0, s1, s2.x); break;
247 case Shader::OPCODE_REFRACT4: refract4(d, s0, s1, s2.x); break;
248 case Shader::OPCODE_NRM2: nrm2(d, s0, pp); break;
249 case Shader::OPCODE_NRM3: nrm3(d, s0, pp); break;
250 case Shader::OPCODE_NRM4: nrm4(d, s0, pp); break;
251 case Shader::OPCODE_ABS: abs(d, s0); break;
Alexis Hetu0f448072016-03-18 10:56:08 -0400252 case Shader::OPCODE_IABS: iabs(d, s0); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400253 case Shader::OPCODE_SINCOS: sincos(d, s0, pp); break;
254 case Shader::OPCODE_COS: cos(d, s0, pp); break;
255 case Shader::OPCODE_SIN: sin(d, s0, pp); break;
256 case Shader::OPCODE_TAN: tan(d, s0, pp); break;
257 case Shader::OPCODE_ACOS: acos(d, s0, pp); break;
258 case Shader::OPCODE_ASIN: asin(d, s0, pp); break;
259 case Shader::OPCODE_ATAN: atan(d, s0, pp); break;
260 case Shader::OPCODE_ATAN2: atan2(d, s0, s1, pp); break;
261 case Shader::OPCODE_COSH: cosh(d, s0, pp); break;
262 case Shader::OPCODE_SINH: sinh(d, s0, pp); break;
263 case Shader::OPCODE_TANH: tanh(d, s0, pp); break;
264 case Shader::OPCODE_ACOSH: acosh(d, s0, pp); break;
265 case Shader::OPCODE_ASINH: asinh(d, s0, pp); break;
266 case Shader::OPCODE_ATANH: atanh(d, s0, pp); break;
Nicolas Capens4f172c72016-01-13 08:34:30 -0500267 case Shader::OPCODE_M4X4: M4X4(d, s0, src1); break;
268 case Shader::OPCODE_M4X3: M4X3(d, s0, src1); break;
269 case Shader::OPCODE_M3X4: M3X4(d, s0, src1); break;
270 case Shader::OPCODE_M3X3: M3X3(d, s0, src1); break;
271 case Shader::OPCODE_M3X2: M3X2(d, s0, src1); break;
272 case Shader::OPCODE_TEX: TEXLD(d, s0, src1, project, bias); break;
273 case Shader::OPCODE_TEXLDD: TEXLDD(d, s0, src1, s2, s3, project, bias); break;
274 case Shader::OPCODE_TEXLDL: TEXLDL(d, s0, src1, project, bias); break;
275 case Shader::OPCODE_TEXSIZE: TEXSIZE(d, s0.x, src1); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400276 case Shader::OPCODE_TEXKILL: TEXKILL(cMask, d, dst.mask); break;
Nicolas Capens4f172c72016-01-13 08:34:30 -0500277 case Shader::OPCODE_TEXOFFSET: TEXOFFSET(d, s0, src1, s2, s3, project, bias); break;
278 case Shader::OPCODE_TEXLDLOFFSET: TEXLDL(d, s0, src1, s2, project, bias); break;
279 case Shader::OPCODE_TEXELFETCH: TEXELFETCH(d, s0, src1, s2); break;
280 case Shader::OPCODE_TEXELFETCHOFFSET: TEXELFETCH(d, s0, src1, s2, s3); break;
281 case Shader::OPCODE_TEXGRAD: TEXGRAD(d, s0, src1, s2, s3); break;
282 case Shader::OPCODE_TEXGRADOFFSET: TEXGRAD(d, s0, src1, s2, s3, s4); break;
283 case Shader::OPCODE_DISCARD: DISCARD(cMask, instruction); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400284 case Shader::OPCODE_DFDX: DFDX(d, s0); break;
285 case Shader::OPCODE_DFDY: DFDY(d, s0); break;
286 case Shader::OPCODE_FWIDTH: FWIDTH(d, s0); break;
Nicolas Capens4f172c72016-01-13 08:34:30 -0500287 case Shader::OPCODE_BREAK: BREAK(); break;
288 case Shader::OPCODE_BREAKC: BREAKC(s0, s1, control); break;
289 case Shader::OPCODE_BREAKP: BREAKP(src0); break;
290 case Shader::OPCODE_CONTINUE: CONTINUE(); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400291 case Shader::OPCODE_TEST: TEST(); break;
Nicolas Capens4f172c72016-01-13 08:34:30 -0500292 case Shader::OPCODE_CALL: CALL(dst.label, dst.callSite); break;
293 case Shader::OPCODE_CALLNZ: CALLNZ(dst.label, dst.callSite, src0); break;
294 case Shader::OPCODE_ELSE: ELSE(); break;
295 case Shader::OPCODE_ENDIF: ENDIF(); break;
296 case Shader::OPCODE_ENDLOOP: ENDLOOP(); break;
297 case Shader::OPCODE_ENDREP: ENDREP(); break;
298 case Shader::OPCODE_ENDWHILE: ENDWHILE(); break;
299 case Shader::OPCODE_IF: IF(src0); break;
300 case Shader::OPCODE_IFC: IFC(s0, s1, control); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400301 case Shader::OPCODE_LABEL: LABEL(dst.index); break;
Nicolas Capens4f172c72016-01-13 08:34:30 -0500302 case Shader::OPCODE_LOOP: LOOP(src1); break;
303 case Shader::OPCODE_REP: REP(src0); break;
304 case Shader::OPCODE_WHILE: WHILE(src0); break;
305 case Shader::OPCODE_RET: RET(); break;
306 case Shader::OPCODE_LEAVE: LEAVE(); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400307 case Shader::OPCODE_CMP: cmp(d, s0, s1, control); break;
308 case Shader::OPCODE_ALL: all(d.x, s0); break;
309 case Shader::OPCODE_ANY: any(d.x, s0); break;
310 case Shader::OPCODE_NOT: not(d, s0); break;
Alexis Hetuc4f2c292015-08-18 15:43:09 -0400311 case Shader::OPCODE_OR: or(d, s0, s1); break;
312 case Shader::OPCODE_XOR: xor(d, s0, s1); break;
313 case Shader::OPCODE_AND: and(d, s0, s1); break;
Alexis Hetu8d78cf72015-08-28 14:24:45 -0400314 case Shader::OPCODE_EQ: equal(d, s0, s1); break;
315 case Shader::OPCODE_NE: notEqual(d, s0, s1); break;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400316 case Shader::OPCODE_END: break;
317 default:
318 ASSERT(false);
319 }
320
321 if(dst.type != Shader::PARAMETER_VOID && dst.type != Shader::PARAMETER_LABEL && opcode != Shader::OPCODE_TEXKILL && opcode != Shader::OPCODE_NOP)
322 {
323 if(dst.integer)
324 {
325 switch(opcode)
326 {
327 case Shader::OPCODE_DIV:
328 if(dst.x) d.x = Trunc(d.x);
329 if(dst.y) d.y = Trunc(d.y);
330 if(dst.z) d.z = Trunc(d.z);
331 if(dst.w) d.w = Trunc(d.w);
332 break;
333 default:
334 break; // No truncation to integer required when arguments are integer
335 }
336 }
337
338 if(dst.saturate)
339 {
340 if(dst.x) d.x = Max(d.x, Float4(0.0f));
341 if(dst.y) d.y = Max(d.y, Float4(0.0f));
342 if(dst.z) d.z = Max(d.z, Float4(0.0f));
343 if(dst.w) d.w = Max(d.w, Float4(0.0f));
344
345 if(dst.x) d.x = Min(d.x, Float4(1.0f));
346 if(dst.y) d.y = Min(d.y, Float4(1.0f));
347 if(dst.z) d.z = Min(d.z, Float4(1.0f));
348 if(dst.w) d.w = Min(d.w, Float4(1.0f));
349 }
350
351 if(instruction->isPredicated())
352 {
353 Vector4f pDst; // FIXME: Rename
354
355 switch(dst.type)
356 {
357 case Shader::PARAMETER_TEMP:
358 if(dst.rel.type == Shader::PARAMETER_VOID)
359 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500360 if(dst.x) pDst.x = r[dst.index].x;
361 if(dst.y) pDst.y = r[dst.index].y;
362 if(dst.z) pDst.z = r[dst.index].z;
363 if(dst.w) pDst.w = r[dst.index].w;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400364 }
365 else
366 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500367 Int a = relativeAddress(dst);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400368
Nicolas Capens4f172c72016-01-13 08:34:30 -0500369 if(dst.x) pDst.x = r[dst.index + a].x;
370 if(dst.y) pDst.y = r[dst.index + a].y;
371 if(dst.z) pDst.z = r[dst.index + a].z;
372 if(dst.w) pDst.w = r[dst.index + a].w;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400373 }
374 break;
375 case Shader::PARAMETER_COLOROUT:
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500376 if(dst.rel.type == Shader::PARAMETER_VOID)
377 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500378 if(dst.x) pDst.x = oC[dst.index].x;
379 if(dst.y) pDst.y = oC[dst.index].y;
380 if(dst.z) pDst.z = oC[dst.index].z;
381 if(dst.w) pDst.w = oC[dst.index].w;
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500382 }
383 else
384 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500385 Int a = relativeAddress(dst) + dst.index;
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500386
Nicolas Capens4f172c72016-01-13 08:34:30 -0500387 if(dst.x) pDst.x = oC[a].x;
388 if(dst.y) pDst.y = oC[a].y;
389 if(dst.z) pDst.z = oC[a].z;
390 if(dst.w) pDst.w = oC[a].w;
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500391 }
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400392 break;
393 case Shader::PARAMETER_PREDICATE:
Nicolas Capens4f172c72016-01-13 08:34:30 -0500394 if(dst.x) pDst.x = p0.x;
395 if(dst.y) pDst.y = p0.y;
396 if(dst.z) pDst.z = p0.z;
397 if(dst.w) pDst.w = p0.w;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400398 break;
399 case Shader::PARAMETER_DEPTHOUT:
Nicolas Capens4f172c72016-01-13 08:34:30 -0500400 pDst.x = oDepth;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400401 break;
402 default:
403 ASSERT(false);
404 }
405
Nicolas Capens4f172c72016-01-13 08:34:30 -0500406 Int4 enable = enableMask(instruction);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400407
408 Int4 xEnable = enable;
409 Int4 yEnable = enable;
410 Int4 zEnable = enable;
411 Int4 wEnable = enable;
412
413 if(predicate)
414 {
415 unsigned char pSwizzle = instruction->predicateSwizzle;
416
Nicolas Capens4f172c72016-01-13 08:34:30 -0500417 Float4 xPredicate = p0[(pSwizzle >> 0) & 0x03];
418 Float4 yPredicate = p0[(pSwizzle >> 2) & 0x03];
419 Float4 zPredicate = p0[(pSwizzle >> 4) & 0x03];
420 Float4 wPredicate = p0[(pSwizzle >> 6) & 0x03];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400421
422 if(!instruction->predicateNot)
423 {
424 if(dst.x) xEnable = xEnable & As<Int4>(xPredicate);
425 if(dst.y) yEnable = yEnable & As<Int4>(yPredicate);
426 if(dst.z) zEnable = zEnable & As<Int4>(zPredicate);
427 if(dst.w) wEnable = wEnable & As<Int4>(wPredicate);
428 }
429 else
430 {
431 if(dst.x) xEnable = xEnable & ~As<Int4>(xPredicate);
432 if(dst.y) yEnable = yEnable & ~As<Int4>(yPredicate);
433 if(dst.z) zEnable = zEnable & ~As<Int4>(zPredicate);
434 if(dst.w) wEnable = wEnable & ~As<Int4>(wPredicate);
435 }
436 }
437
438 if(dst.x) d.x = As<Float4>(As<Int4>(d.x) & xEnable);
439 if(dst.y) d.y = As<Float4>(As<Int4>(d.y) & yEnable);
440 if(dst.z) d.z = As<Float4>(As<Int4>(d.z) & zEnable);
441 if(dst.w) d.w = As<Float4>(As<Int4>(d.w) & wEnable);
442
443 if(dst.x) d.x = As<Float4>(As<Int4>(d.x) | (As<Int4>(pDst.x) & ~xEnable));
444 if(dst.y) d.y = As<Float4>(As<Int4>(d.y) | (As<Int4>(pDst.y) & ~yEnable));
445 if(dst.z) d.z = As<Float4>(As<Int4>(d.z) | (As<Int4>(pDst.z) & ~zEnable));
446 if(dst.w) d.w = As<Float4>(As<Int4>(d.w) | (As<Int4>(pDst.w) & ~wEnable));
447 }
448
449 switch(dst.type)
450 {
451 case Shader::PARAMETER_TEMP:
452 if(dst.rel.type == Shader::PARAMETER_VOID)
453 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500454 if(dst.x) r[dst.index].x = d.x;
455 if(dst.y) r[dst.index].y = d.y;
456 if(dst.z) r[dst.index].z = d.z;
457 if(dst.w) r[dst.index].w = d.w;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400458 }
459 else
460 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500461 Int a = relativeAddress(dst);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400462
Nicolas Capens4f172c72016-01-13 08:34:30 -0500463 if(dst.x) r[dst.index + a].x = d.x;
464 if(dst.y) r[dst.index + a].y = d.y;
465 if(dst.z) r[dst.index + a].z = d.z;
466 if(dst.w) r[dst.index + a].w = d.w;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400467 }
468 break;
469 case Shader::PARAMETER_COLOROUT:
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500470 if(dst.rel.type == Shader::PARAMETER_VOID)
471 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500472 if(dst.x) { oC[dst.index].x = d.x; }
473 if(dst.y) { oC[dst.index].y = d.y; }
474 if(dst.z) { oC[dst.index].z = d.z; }
475 if(dst.w) { oC[dst.index].w = d.w; }
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500476 }
477 else
478 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500479 Int a = relativeAddress(dst) + dst.index;
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500480
Nicolas Capens4f172c72016-01-13 08:34:30 -0500481 if(dst.x) { oC[a].x = d.x; }
482 if(dst.y) { oC[a].y = d.y; }
483 if(dst.z) { oC[a].z = d.z; }
484 if(dst.w) { oC[a].w = d.w; }
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500485 }
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400486 break;
487 case Shader::PARAMETER_PREDICATE:
Nicolas Capens4f172c72016-01-13 08:34:30 -0500488 if(dst.x) p0.x = d.x;
489 if(dst.y) p0.y = d.y;
490 if(dst.z) p0.z = d.z;
491 if(dst.w) p0.w = d.w;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400492 break;
493 case Shader::PARAMETER_DEPTHOUT:
Nicolas Capens4f172c72016-01-13 08:34:30 -0500494 oDepth = d.x;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400495 break;
496 default:
497 ASSERT(false);
498 }
499 }
500 }
501
502 if(currentLabel != -1)
503 {
504 Nucleus::setInsertBlock(returnBlock);
505 }
506
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500507 for(int i = 0; i < RENDERTARGETS; i++)
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400508 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500509 c[i] = oC[i];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400510 }
511 }
Nicolas Capens5d961882016-01-01 23:18:14 -0500512
Nicolas Capens4f172c72016-01-13 08:34:30 -0500513 Bool PixelProgram::alphaTest(Int cMask[4])
Nicolas Capens5d961882016-01-01 23:18:14 -0500514 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500515 clampColor(c);
Nicolas Capens5d961882016-01-01 23:18:14 -0500516
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400517 if(!state.alphaTestActive())
518 {
519 return true;
520 }
521
522 Int aMask;
523
524 if(state.transparencyAntialiasing == TRANSPARENCY_NONE)
525 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500526 Short4 alpha = RoundShort4(c[0].w * Float4(0x1000));
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400527
Nicolas Capens4f172c72016-01-13 08:34:30 -0500528 PixelRoutine::alphaTest(aMask, alpha);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400529
530 for(unsigned int q = 0; q < state.multiSample; q++)
531 {
532 cMask[q] &= aMask;
533 }
534 }
535 else if(state.transparencyAntialiasing == TRANSPARENCY_ALPHA_TO_COVERAGE)
536 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500537 alphaToCoverage(cMask, c[0].w);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400538 }
539 else ASSERT(false);
540
541 Int pass = cMask[0];
542
543 for(unsigned int q = 1; q < state.multiSample; q++)
544 {
545 pass = pass | cMask[q];
546 }
547
Nicolas Capens5d961882016-01-01 23:18:14 -0500548 return pass != 0x0;
549 }
550
Nicolas Capens4f172c72016-01-13 08:34:30 -0500551 void PixelProgram::rasterOperation(Float4 &fog, Pointer<Byte> cBuffer[4], Int &x, Int sMask[4], Int zMask[4], Int cMask[4])
Nicolas Capens5d961882016-01-01 23:18:14 -0500552 {
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500553 for(int index = 0; index < RENDERTARGETS; index++)
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400554 {
555 if(!state.colorWriteActive(index))
556 {
557 continue;
Nicolas Capens5d961882016-01-01 23:18:14 -0500558 }
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400559
560 if(!postBlendSRGB && state.writeSRGB)
561 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500562 c[index].x = linearToSRGB(c[index].x);
563 c[index].y = linearToSRGB(c[index].y);
564 c[index].z = linearToSRGB(c[index].z);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400565 }
566
567 if(index == 0)
568 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500569 fogBlend(c[index], fog);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400570 }
571
572 switch(state.targetFormat[index])
573 {
574 case FORMAT_R5G6B5:
575 case FORMAT_X8R8G8B8:
576 case FORMAT_X8B8G8R8:
577 case FORMAT_A8R8G8B8:
578 case FORMAT_A8B8G8R8:
579 case FORMAT_A8:
580 case FORMAT_G16R16:
581 case FORMAT_A16B16G16R16:
582 for(unsigned int q = 0; q < state.multiSample; q++)
583 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500584 Pointer<Byte> buffer = cBuffer[index] + q * *Pointer<Int>(data + OFFSET(DrawData, colorSliceB[index]));
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400585 Vector4s color;
586
Nicolas Capens4f172c72016-01-13 08:34:30 -0500587 color.x = convertFixed16(c[index].x, false);
588 color.y = convertFixed16(c[index].y, false);
589 color.z = convertFixed16(c[index].z, false);
590 color.w = convertFixed16(c[index].w, false);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400591
592 if(state.multiSampleMask & (1 << q))
593 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500594 alphaBlend(index, buffer, color, x);
595 logicOperation(index, buffer, color, x);
596 writeColor(index, buffer, x, color, sMask[q], zMask[q], cMask[q]);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400597 }
598 }
599 break;
600 case FORMAT_R32F:
601 case FORMAT_G32R32F:
602 case FORMAT_A32B32G32R32F:
Alexis Hetu1abb6382016-02-08 11:21:16 -0500603 case FORMAT_R32I:
604 case FORMAT_G32R32I:
605 case FORMAT_A32B32G32R32I:
606 case FORMAT_R32UI:
607 case FORMAT_G32R32UI:
608 case FORMAT_A32B32G32R32UI:
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400609 for(unsigned int q = 0; q < state.multiSample; q++)
610 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500611 Pointer<Byte> buffer = cBuffer[index] + q * *Pointer<Int>(data + OFFSET(DrawData, colorSliceB[index]));
612 Vector4f color = c[index];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400613
614 if(state.multiSampleMask & (1 << q))
615 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500616 alphaBlend(index, buffer, color, x);
617 writeColor(index, buffer, x, color, sMask[q], zMask[q], cMask[q]);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400618 }
619 }
620 break;
621 default:
622 ASSERT(false);
623 }
624 }
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500625 }
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400626
Nicolas Capens4f172c72016-01-13 08:34:30 -0500627 void PixelProgram::sampleTexture(Vector4f &c, const Src &sampler, Float4 &u, Float4 &v, Float4 &w, Float4 &q, Vector4f &dsx, Vector4f &dsy, bool project, bool bias, bool gradients, bool lodProvided)
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400628 {
629 if(sampler.type == Shader::PARAMETER_SAMPLER && sampler.rel.type == Shader::PARAMETER_VOID)
630 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500631 sampleTexture(c, sampler.index, u, v, w, q, dsx, dsy, project, bias, gradients, lodProvided);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400632 }
633 else
634 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500635 Int index = As<Int>(Float(fetchRegisterF(sampler).x.x));
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400636
637 for(int i = 0; i < TEXTURE_IMAGE_UNITS; i++)
638 {
639 if(shader->usesSampler(i))
640 {
641 If(index == i)
642 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500643 sampleTexture(c, i, u, v, w, q, dsx, dsy, project, bias, gradients, lodProvided);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400644 // FIXME: When the sampler states are the same, we could use one sampler and just index the texture
645 }
646 }
647 }
648 }
649 }
650
Nicolas Capens4f172c72016-01-13 08:34:30 -0500651 void PixelProgram::sampleTexture(Vector4f &c, int stage, Float4 &u, Float4 &v, Float4 &w, Float4 &q, Vector4f &dsx, Vector4f &dsy, bool project, bool bias, bool gradients, bool lodProvided)
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400652 {
653#if PERF_PROFILE
654 Long texTime = Ticks();
655#endif
656
Nicolas Capens4f172c72016-01-13 08:34:30 -0500657 Pointer<Byte> texture = data + OFFSET(DrawData, mipmap) + stage * sizeof(Texture);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400658
659 if(!project)
660 {
661 sampler[stage]->sampleTexture(texture, c, u, v, w, q, dsx, dsy, bias, gradients, lodProvided);
662 }
663 else
664 {
665 Float4 rq = reciprocal(q);
666
667 Float4 u_q = u * rq;
668 Float4 v_q = v * rq;
669 Float4 w_q = w * rq;
670
671 sampler[stage]->sampleTexture(texture, c, u_q, v_q, w_q, q, dsx, dsy, bias, gradients, lodProvided);
672 }
673
674#if PERF_PROFILE
Nicolas Capens4f172c72016-01-13 08:34:30 -0500675 cycles[PERF_TEX] += Ticks() - texTime;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400676#endif
677 }
678
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500679 void PixelProgram::clampColor(Vector4f oC[RENDERTARGETS])
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400680 {
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500681 for(int index = 0; index < RENDERTARGETS; index++)
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400682 {
683 if(!state.colorWriteActive(index) && !(index == 0 && state.alphaTestActive()))
684 {
685 continue;
686 }
687
688 switch(state.targetFormat[index])
689 {
690 case FORMAT_NULL:
691 break;
692 case FORMAT_R5G6B5:
693 case FORMAT_A8R8G8B8:
694 case FORMAT_A8B8G8R8:
695 case FORMAT_X8R8G8B8:
696 case FORMAT_X8B8G8R8:
697 case FORMAT_A8:
698 case FORMAT_G16R16:
699 case FORMAT_A16B16G16R16:
700 oC[index].x = Max(oC[index].x, Float4(0.0f)); oC[index].x = Min(oC[index].x, Float4(1.0f));
701 oC[index].y = Max(oC[index].y, Float4(0.0f)); oC[index].y = Min(oC[index].y, Float4(1.0f));
702 oC[index].z = Max(oC[index].z, Float4(0.0f)); oC[index].z = Min(oC[index].z, Float4(1.0f));
703 oC[index].w = Max(oC[index].w, Float4(0.0f)); oC[index].w = Min(oC[index].w, Float4(1.0f));
704 break;
705 case FORMAT_R32F:
706 case FORMAT_G32R32F:
707 case FORMAT_A32B32G32R32F:
Alexis Hetu1abb6382016-02-08 11:21:16 -0500708 case FORMAT_R32I:
709 case FORMAT_G32R32I:
710 case FORMAT_A32B32G32R32I:
711 case FORMAT_R32UI:
712 case FORMAT_G32R32UI:
713 case FORMAT_A32B32G32R32UI:
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400714 break;
715 default:
716 ASSERT(false);
717 }
718 }
719 }
720
Nicolas Capens4f172c72016-01-13 08:34:30 -0500721 Int4 PixelProgram::enableMask(const Shader::Instruction *instruction)
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400722 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500723 Int4 enable = instruction->analysisBranch ? Int4(enableStack[enableIndex]) : Int4(0xFFFFFFFF);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400724
725 if(!whileTest)
726 {
727 if(shader->containsBreakInstruction() && instruction->analysisBreak)
728 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500729 enable &= enableBreak;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400730 }
731
732 if(shader->containsContinueInstruction() && instruction->analysisContinue)
733 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500734 enable &= enableContinue;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400735 }
736
737 if(shader->containsLeaveInstruction() && instruction->analysisLeave)
738 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500739 enable &= enableLeave;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400740 }
741 }
742
743 return enable;
744 }
745
Nicolas Capens4f172c72016-01-13 08:34:30 -0500746 Vector4f PixelProgram::fetchRegisterF(const Src &src, unsigned int offset)
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400747 {
748 Vector4f reg;
Nicolas Capens5d961882016-01-01 23:18:14 -0500749 unsigned int i = src.index + offset;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400750
751 switch(src.type)
752 {
753 case Shader::PARAMETER_TEMP:
754 if(src.rel.type == Shader::PARAMETER_VOID)
755 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500756 reg = r[i];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400757 }
758 else
759 {
Alexis Hetu2c2a7b22015-10-27 16:12:11 -0400760 Int a = relativeAddress(src, src.bufferIndex);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400761
Nicolas Capens4f172c72016-01-13 08:34:30 -0500762 reg = r[i + a];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400763 }
764 break;
765 case Shader::PARAMETER_INPUT:
766 {
767 if(src.rel.type == Shader::PARAMETER_VOID) // Not relative
768 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500769 reg = v[i];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400770 }
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400771 else
772 {
Alexis Hetu2c2a7b22015-10-27 16:12:11 -0400773 Int a = relativeAddress(src, src.bufferIndex);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400774
Nicolas Capens4f172c72016-01-13 08:34:30 -0500775 reg = v[i + a];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400776 }
777 }
778 break;
779 case Shader::PARAMETER_CONST:
Nicolas Capens4f172c72016-01-13 08:34:30 -0500780 reg = readConstant(src, offset);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400781 break;
782 case Shader::PARAMETER_TEXTURE:
Nicolas Capens4f172c72016-01-13 08:34:30 -0500783 reg = v[2 + i];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400784 break;
785 case Shader::PARAMETER_MISCTYPE:
Nicolas Capens4f172c72016-01-13 08:34:30 -0500786 if(src.index == 0) reg = vPos;
787 if(src.index == 1) reg = vFace;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400788 break;
789 case Shader::PARAMETER_SAMPLER:
790 if(src.rel.type == Shader::PARAMETER_VOID)
791 {
792 reg.x = As<Float4>(Int4(i));
793 }
794 else if(src.rel.type == Shader::PARAMETER_TEMP)
795 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500796 reg.x = As<Float4>(Int4(i) + As<Int4>(r[src.rel.index].x));
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400797 }
798 return reg;
799 case Shader::PARAMETER_PREDICATE: return reg; // Dummy
800 case Shader::PARAMETER_VOID: return reg; // Dummy
801 case Shader::PARAMETER_FLOAT4LITERAL:
802 reg.x = Float4(src.value[0]);
803 reg.y = Float4(src.value[1]);
804 reg.z = Float4(src.value[2]);
805 reg.w = Float4(src.value[3]);
806 break;
807 case Shader::PARAMETER_CONSTINT: return reg; // Dummy
808 case Shader::PARAMETER_CONSTBOOL: return reg; // Dummy
809 case Shader::PARAMETER_LOOP: return reg; // Dummy
810 case Shader::PARAMETER_COLOROUT:
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500811 if(src.rel.type == Shader::PARAMETER_VOID) // Not relative
812 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500813 reg = oC[i];
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500814 }
815 else
816 {
Alexis Hetu2c2a7b22015-10-27 16:12:11 -0400817 Int a = relativeAddress(src, src.bufferIndex);
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500818
Nicolas Capens4f172c72016-01-13 08:34:30 -0500819 reg = oC[i + a];
Alexis Hetu1edcd8b2015-11-05 11:12:41 -0500820 }
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400821 break;
822 case Shader::PARAMETER_DEPTHOUT:
Nicolas Capens4f172c72016-01-13 08:34:30 -0500823 reg.x = oDepth;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400824 break;
825 default:
826 ASSERT(false);
827 }
828
829 const Float4 &x = reg[(src.swizzle >> 0) & 0x3];
830 const Float4 &y = reg[(src.swizzle >> 2) & 0x3];
831 const Float4 &z = reg[(src.swizzle >> 4) & 0x3];
832 const Float4 &w = reg[(src.swizzle >> 6) & 0x3];
833
834 Vector4f mod;
835
836 switch(src.modifier)
837 {
838 case Shader::MODIFIER_NONE:
839 mod.x = x;
840 mod.y = y;
841 mod.z = z;
842 mod.w = w;
843 break;
844 case Shader::MODIFIER_NEGATE:
845 mod.x = -x;
846 mod.y = -y;
847 mod.z = -z;
848 mod.w = -w;
849 break;
850 case Shader::MODIFIER_ABS:
851 mod.x = Abs(x);
852 mod.y = Abs(y);
853 mod.z = Abs(z);
854 mod.w = Abs(w);
855 break;
856 case Shader::MODIFIER_ABS_NEGATE:
857 mod.x = -Abs(x);
858 mod.y = -Abs(y);
859 mod.z = -Abs(z);
860 mod.w = -Abs(w);
861 break;
862 case Shader::MODIFIER_NOT:
863 mod.x = As<Float4>(As<Int4>(x) ^ Int4(0xFFFFFFFF));
864 mod.y = As<Float4>(As<Int4>(y) ^ Int4(0xFFFFFFFF));
865 mod.z = As<Float4>(As<Int4>(z) ^ Int4(0xFFFFFFFF));
866 mod.w = As<Float4>(As<Int4>(w) ^ Int4(0xFFFFFFFF));
867 break;
868 default:
869 ASSERT(false);
870 }
871
872 return mod;
873 }
874
Alexis Hetu2c2a7b22015-10-27 16:12:11 -0400875 RValue<Pointer<Byte>> PixelProgram::uniformAddress(int bufferIndex, unsigned int index)
876 {
877 if(bufferIndex == -1)
878 {
879 return data + OFFSET(DrawData, ps.c[index]);
880 }
881 else
882 {
883 return *Pointer<Pointer<Byte>>(data + OFFSET(DrawData, ps.u[bufferIndex])) + index;
884 }
885 }
886
887 RValue<Pointer<Byte>> PixelProgram::uniformAddress(int bufferIndex, unsigned int index, Int& offset)
888 {
889 return uniformAddress(bufferIndex, index) + offset * sizeof(float4);
890 }
891
Nicolas Capens4f172c72016-01-13 08:34:30 -0500892 Vector4f PixelProgram::readConstant(const Src &src, unsigned int offset)
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400893 {
894 Vector4f c;
Nicolas Capens5d961882016-01-01 23:18:14 -0500895 unsigned int i = src.index + offset;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400896
897 if(src.rel.type == Shader::PARAMETER_VOID) // Not relative
898 {
Alexis Hetu2c2a7b22015-10-27 16:12:11 -0400899 c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i));
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400900
901 c.x = c.x.xxxx;
902 c.y = c.y.yyyy;
903 c.z = c.z.zzzz;
904 c.w = c.w.wwww;
905
906 if(shader->containsDefineInstruction()) // Constant may be known at compile time
907 {
908 for(size_t j = 0; j < shader->getLength(); j++)
909 {
910 const Shader::Instruction &instruction = *shader->getInstruction(j);
911
912 if(instruction.opcode == Shader::OPCODE_DEF)
913 {
914 if(instruction.dst.index == i)
915 {
916 c.x = Float4(instruction.src[0].value[0]);
917 c.y = Float4(instruction.src[0].value[1]);
918 c.z = Float4(instruction.src[0].value[2]);
919 c.w = Float4(instruction.src[0].value[3]);
920
921 break;
922 }
923 }
924 }
925 }
926 }
927 else if(src.rel.type == Shader::PARAMETER_LOOP)
928 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500929 Int loopCounter = aL[loopDepth];
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400930
Alexis Hetu2c2a7b22015-10-27 16:12:11 -0400931 c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i, loopCounter));
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400932
933 c.x = c.x.xxxx;
934 c.y = c.y.yyyy;
935 c.z = c.z.zzzz;
936 c.w = c.w.wwww;
937 }
938 else
939 {
Alexis Hetu2c2a7b22015-10-27 16:12:11 -0400940 Int a = relativeAddress(src, src.bufferIndex);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400941
Alexis Hetu2c2a7b22015-10-27 16:12:11 -0400942 c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i, a));
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400943
944 c.x = c.x.xxxx;
945 c.y = c.y.yyyy;
946 c.z = c.z.zzzz;
947 c.w = c.w.wwww;
948 }
949
950 return c;
951 }
952
Alexis Hetu2c2a7b22015-10-27 16:12:11 -0400953 Int PixelProgram::relativeAddress(const Shader::Parameter &var, int bufferIndex)
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400954 {
955 ASSERT(var.rel.deterministic);
956
957 if(var.rel.type == Shader::PARAMETER_TEMP)
958 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500959 return As<Int>(Extract(r[var.rel.index].x, 0)) * var.rel.scale;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400960 }
961 else if(var.rel.type == Shader::PARAMETER_INPUT)
962 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500963 return As<Int>(Extract(v[var.rel.index].x, 0)) * var.rel.scale;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400964 }
965 else if(var.rel.type == Shader::PARAMETER_OUTPUT)
966 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500967 return As<Int>(Extract(oC[var.rel.index].x, 0)) * var.rel.scale;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400968 }
969 else if(var.rel.type == Shader::PARAMETER_CONST)
970 {
Alexis Hetu48be7352016-02-10 14:32:34 -0500971 return *Pointer<Int>(uniformAddress(bufferIndex, var.rel.index)) * var.rel.scale;
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400972 }
Nicolas Capens907700d2016-01-20 17:09:28 -0500973 else if(var.rel.type == Shader::PARAMETER_LOOP)
974 {
975 return aL[loopDepth];
976 }
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400977 else ASSERT(false);
978
979 return 0;
980 }
981
982 Float4 PixelProgram::linearToSRGB(const Float4 &x) // Approximates x^(1.0/2.2)
983 {
984 Float4 sqrtx = Rcp_pp(RcpSqrt_pp(x));
985 Float4 sRGB = sqrtx * Float4(1.14f) - x * Float4(0.14f);
986
987 return Min(Max(sRGB, Float4(0.0f)), Float4(1.0f));
988 }
989
Nicolas Capens4f172c72016-01-13 08:34:30 -0500990 void PixelProgram::M3X2(Vector4f &dst, Vector4f &src0, const Src &src1)
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400991 {
Nicolas Capens4f172c72016-01-13 08:34:30 -0500992 Vector4f row0 = fetchRegisterF(src1, 0);
993 Vector4f row1 = fetchRegisterF(src1, 1);
Alexis Hetuf2a8c372015-07-13 11:08:41 -0400994
995 dst.x = dot3(src0, row0);
996 dst.y = dot3(src0, row1);
997 }
998
Nicolas Capens4f172c72016-01-13 08:34:30 -0500999 void PixelProgram::M3X3(Vector4f &dst, Vector4f &src0, const Src &src1)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001000 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001001 Vector4f row0 = fetchRegisterF(src1, 0);
1002 Vector4f row1 = fetchRegisterF(src1, 1);
1003 Vector4f row2 = fetchRegisterF(src1, 2);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001004
1005 dst.x = dot3(src0, row0);
1006 dst.y = dot3(src0, row1);
1007 dst.z = dot3(src0, row2);
1008 }
1009
Nicolas Capens4f172c72016-01-13 08:34:30 -05001010 void PixelProgram::M3X4(Vector4f &dst, Vector4f &src0, const Src &src1)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001011 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001012 Vector4f row0 = fetchRegisterF(src1, 0);
1013 Vector4f row1 = fetchRegisterF(src1, 1);
1014 Vector4f row2 = fetchRegisterF(src1, 2);
1015 Vector4f row3 = fetchRegisterF(src1, 3);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001016
1017 dst.x = dot3(src0, row0);
1018 dst.y = dot3(src0, row1);
1019 dst.z = dot3(src0, row2);
1020 dst.w = dot3(src0, row3);
1021 }
1022
Nicolas Capens4f172c72016-01-13 08:34:30 -05001023 void PixelProgram::M4X3(Vector4f &dst, Vector4f &src0, const Src &src1)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001024 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001025 Vector4f row0 = fetchRegisterF(src1, 0);
1026 Vector4f row1 = fetchRegisterF(src1, 1);
1027 Vector4f row2 = fetchRegisterF(src1, 2);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001028
1029 dst.x = dot4(src0, row0);
1030 dst.y = dot4(src0, row1);
1031 dst.z = dot4(src0, row2);
1032 }
1033
Nicolas Capens4f172c72016-01-13 08:34:30 -05001034 void PixelProgram::M4X4(Vector4f &dst, Vector4f &src0, const Src &src1)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001035 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001036 Vector4f row0 = fetchRegisterF(src1, 0);
1037 Vector4f row1 = fetchRegisterF(src1, 1);
1038 Vector4f row2 = fetchRegisterF(src1, 2);
1039 Vector4f row3 = fetchRegisterF(src1, 3);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001040
1041 dst.x = dot4(src0, row0);
1042 dst.y = dot4(src0, row1);
1043 dst.z = dot4(src0, row2);
1044 dst.w = dot4(src0, row3);
1045 }
1046
Nicolas Capens4f172c72016-01-13 08:34:30 -05001047 void PixelProgram::TEXLD(Vector4f &dst, Vector4f &src0, const Src &src1, bool project, bool bias)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001048 {
1049 Vector4f tmp;
Nicolas Capens4f172c72016-01-13 08:34:30 -05001050 sampleTexture(tmp, src1, src0.x, src0.y, src0.z, src0.w, src0, src0, project, bias);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001051
1052 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1053 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1054 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1055 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1056 }
1057
Nicolas Capens4f172c72016-01-13 08:34:30 -05001058 void PixelProgram::TEXOFFSET(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3, bool project, bool bias)
Alexis Hetu25d47fc2015-10-22 13:58:32 -04001059 {
1060 UNIMPLEMENTED();
1061 }
1062
Nicolas Capens4f172c72016-01-13 08:34:30 -05001063 void PixelProgram::TEXLDL(Vector4f &dst, Vector4f &src0, const Src &src1, Vector4f &offset, bool project, bool bias)
Alexis Hetu25d47fc2015-10-22 13:58:32 -04001064 {
1065 UNIMPLEMENTED();
1066 }
1067
Nicolas Capens4f172c72016-01-13 08:34:30 -05001068 void PixelProgram::TEXELFETCH(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2)
Alexis Hetu25d47fc2015-10-22 13:58:32 -04001069 {
1070 UNIMPLEMENTED();
1071 }
1072
Nicolas Capens4f172c72016-01-13 08:34:30 -05001073 void PixelProgram::TEXELFETCH(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &offset)
Alexis Hetu25d47fc2015-10-22 13:58:32 -04001074 {
1075 UNIMPLEMENTED();
1076 }
1077
Nicolas Capens4f172c72016-01-13 08:34:30 -05001078 void PixelProgram::TEXGRAD(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3)
Alexis Hetu25d47fc2015-10-22 13:58:32 -04001079 {
1080 UNIMPLEMENTED();
1081 }
1082
Nicolas Capens4f172c72016-01-13 08:34:30 -05001083 void PixelProgram::TEXGRAD(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3, Vector4f &offset)
Alexis Hetu25d47fc2015-10-22 13:58:32 -04001084 {
1085 UNIMPLEMENTED();
1086 }
1087
Nicolas Capens4f172c72016-01-13 08:34:30 -05001088 void PixelProgram::TEXLDD(Vector4f &dst, Vector4f &src0, const Src &src1, Vector4f &src2, Vector4f &src3, bool project, bool bias)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001089 {
1090 Vector4f tmp;
Nicolas Capens4f172c72016-01-13 08:34:30 -05001091 sampleTexture(tmp, src1, src0.x, src0.y, src0.z, src0.w, src2, src3, project, bias, true);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001092
1093 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1094 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1095 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1096 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1097 }
1098
Nicolas Capens4f172c72016-01-13 08:34:30 -05001099 void PixelProgram::TEXLDL(Vector4f &dst, Vector4f &src0, const Src &src1, bool project, bool bias)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001100 {
1101 Vector4f tmp;
Nicolas Capens4f172c72016-01-13 08:34:30 -05001102 sampleTexture(tmp, src1, src0.x, src0.y, src0.z, src0.w, src0, src0, project, bias, false, true);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001103
1104 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1105 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1106 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1107 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1108 }
1109
Nicolas Capens4f172c72016-01-13 08:34:30 -05001110 void PixelProgram::TEXSIZE(Vector4f &dst, Float4 &lod, const Src &src1)
Alexis Hetu9bcb31d2015-07-22 17:03:26 -04001111 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001112 Pointer<Byte> textureMipmap = data + OFFSET(DrawData, mipmap) + src1.index * sizeof(Texture) + OFFSET(Texture, mipmap);
Alexis Hetu9bcb31d2015-07-22 17:03:26 -04001113 for(int i = 0; i < 4; ++i)
1114 {
1115 Pointer<Byte> mipmap = textureMipmap + (As<Int>(Extract(lod, i)) + Int(1)) * sizeof(Mipmap);
1116 dst.x = Insert(dst.x, As<Float>(Int(*Pointer<Short>(mipmap + OFFSET(Mipmap, width)))), i);
1117 dst.y = Insert(dst.y, As<Float>(Int(*Pointer<Short>(mipmap + OFFSET(Mipmap, height)))), i);
1118 dst.z = Insert(dst.z, As<Float>(Int(*Pointer<Short>(mipmap + OFFSET(Mipmap, depth)))), i);
1119 }
1120 }
1121
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001122 void PixelProgram::TEXKILL(Int cMask[4], Vector4f &src, unsigned char mask)
1123 {
1124 Int kill = -1;
1125
1126 if(mask & 0x1) kill &= SignMask(CmpNLT(src.x, Float4(0.0f)));
1127 if(mask & 0x2) kill &= SignMask(CmpNLT(src.y, Float4(0.0f)));
1128 if(mask & 0x4) kill &= SignMask(CmpNLT(src.z, Float4(0.0f)));
1129 if(mask & 0x8) kill &= SignMask(CmpNLT(src.w, Float4(0.0f)));
1130
1131 // FIXME: Dynamic branching affects TEXKILL?
1132 // if(shader->containsDynamicBranching())
1133 // {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001134 // kill = ~SignMask(enableMask());
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001135 // }
1136
1137 for(unsigned int q = 0; q < state.multiSample; q++)
1138 {
1139 cMask[q] &= kill;
1140 }
1141
1142 // FIXME: Branch to end of shader if all killed?
1143 }
1144
Nicolas Capens4f172c72016-01-13 08:34:30 -05001145 void PixelProgram::DISCARD(Int cMask[4], const Shader::Instruction *instruction)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001146 {
1147 Int kill = 0;
1148
1149 if(shader->containsDynamicBranching())
1150 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001151 kill = ~SignMask(enableMask(instruction));
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001152 }
1153
1154 for(unsigned int q = 0; q < state.multiSample; q++)
1155 {
1156 cMask[q] &= kill;
1157 }
1158
1159 // FIXME: Branch to end of shader if all killed?
1160 }
1161
1162 void PixelProgram::DFDX(Vector4f &dst, Vector4f &src)
1163 {
1164 dst.x = src.x.yyww - src.x.xxzz;
1165 dst.y = src.y.yyww - src.y.xxzz;
1166 dst.z = src.z.yyww - src.z.xxzz;
1167 dst.w = src.w.yyww - src.w.xxzz;
1168 }
1169
1170 void PixelProgram::DFDY(Vector4f &dst, Vector4f &src)
1171 {
1172 dst.x = src.x.zwzw - src.x.xyxy;
1173 dst.y = src.y.zwzw - src.y.xyxy;
1174 dst.z = src.z.zwzw - src.z.xyxy;
1175 dst.w = src.w.zwzw - src.w.xyxy;
1176 }
1177
1178 void PixelProgram::FWIDTH(Vector4f &dst, Vector4f &src)
1179 {
1180 // abs(dFdx(src)) + abs(dFdy(src));
1181 dst.x = Abs(src.x.yyww - src.x.xxzz) + Abs(src.x.zwzw - src.x.xyxy);
1182 dst.y = Abs(src.y.yyww - src.y.xxzz) + Abs(src.y.zwzw - src.y.xyxy);
1183 dst.z = Abs(src.z.yyww - src.z.xxzz) + Abs(src.z.zwzw - src.z.xyxy);
1184 dst.w = Abs(src.w.yyww - src.w.xxzz) + Abs(src.w.zwzw - src.w.xyxy);
1185 }
1186
Nicolas Capens4f172c72016-01-13 08:34:30 -05001187 void PixelProgram::BREAK()
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001188 {
1189 llvm::BasicBlock *deadBlock = Nucleus::createBasicBlock();
1190 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
1191
1192 if(breakDepth == 0)
1193 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001194 enableIndex = enableIndex - breakDepth;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001195 Nucleus::createBr(endBlock);
1196 }
1197 else
1198 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001199 enableBreak = enableBreak & ~enableStack[enableIndex];
1200 Bool allBreak = SignMask(enableBreak) == 0x0;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001201
Nicolas Capens4f172c72016-01-13 08:34:30 -05001202 enableIndex = enableIndex - breakDepth;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001203 branch(allBreak, endBlock, deadBlock);
1204 }
1205
1206 Nucleus::setInsertBlock(deadBlock);
Nicolas Capens4f172c72016-01-13 08:34:30 -05001207 enableIndex = enableIndex + breakDepth;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001208 }
1209
Nicolas Capens4f172c72016-01-13 08:34:30 -05001210 void PixelProgram::BREAKC(Vector4f &src0, Vector4f &src1, Control control)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001211 {
1212 Int4 condition;
1213
1214 switch(control)
1215 {
1216 case Shader::CONTROL_GT: condition = CmpNLE(src0.x, src1.x); break;
1217 case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x); break;
1218 case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x); break;
1219 case Shader::CONTROL_LT: condition = CmpLT(src0.x, src1.x); break;
1220 case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x); break;
1221 case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x); break;
1222 default:
1223 ASSERT(false);
1224 }
1225
Nicolas Capens4f172c72016-01-13 08:34:30 -05001226 BREAK(condition);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001227 }
1228
Nicolas Capens4f172c72016-01-13 08:34:30 -05001229 void PixelProgram::BREAKP(const Src &predicateRegister) // FIXME: Factor out parts common with BREAKC
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001230 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001231 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001232
1233 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1234 {
1235 condition = ~condition;
1236 }
1237
Nicolas Capens4f172c72016-01-13 08:34:30 -05001238 BREAK(condition);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001239 }
1240
Nicolas Capens4f172c72016-01-13 08:34:30 -05001241 void PixelProgram::BREAK(Int4 &condition)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001242 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001243 condition &= enableStack[enableIndex];
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001244
1245 llvm::BasicBlock *continueBlock = Nucleus::createBasicBlock();
1246 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
1247
Nicolas Capens4f172c72016-01-13 08:34:30 -05001248 enableBreak = enableBreak & ~condition;
1249 Bool allBreak = SignMask(enableBreak) == 0x0;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001250
Nicolas Capens4f172c72016-01-13 08:34:30 -05001251 enableIndex = enableIndex - breakDepth;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001252 branch(allBreak, endBlock, continueBlock);
1253
1254 Nucleus::setInsertBlock(continueBlock);
Nicolas Capens4f172c72016-01-13 08:34:30 -05001255 enableIndex = enableIndex + breakDepth;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001256 }
1257
Nicolas Capens4f172c72016-01-13 08:34:30 -05001258 void PixelProgram::CONTINUE()
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001259 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001260 enableContinue = enableContinue & ~enableStack[enableIndex];
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001261 }
1262
1263 void PixelProgram::TEST()
1264 {
1265 whileTest = true;
1266 }
1267
Nicolas Capens4f172c72016-01-13 08:34:30 -05001268 void PixelProgram::CALL(int labelIndex, int callSiteIndex)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001269 {
1270 if(!labelBlock[labelIndex])
1271 {
1272 labelBlock[labelIndex] = Nucleus::createBasicBlock();
1273 }
1274
1275 if(callRetBlock[labelIndex].size() > 1)
1276 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001277 callStack[stackIndex++] = UInt(callSiteIndex);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001278 }
1279
Nicolas Capens4f172c72016-01-13 08:34:30 -05001280 Int4 restoreLeave = enableLeave;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001281
1282 Nucleus::createBr(labelBlock[labelIndex]);
1283 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1284
Nicolas Capens4f172c72016-01-13 08:34:30 -05001285 enableLeave = restoreLeave;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001286 }
1287
Nicolas Capens4f172c72016-01-13 08:34:30 -05001288 void PixelProgram::CALLNZ(int labelIndex, int callSiteIndex, const Src &src)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001289 {
1290 if(src.type == Shader::PARAMETER_CONSTBOOL)
1291 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001292 CALLNZb(labelIndex, callSiteIndex, src);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001293 }
1294 else if(src.type == Shader::PARAMETER_PREDICATE)
1295 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001296 CALLNZp(labelIndex, callSiteIndex, src);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001297 }
1298 else ASSERT(false);
1299 }
1300
Nicolas Capens4f172c72016-01-13 08:34:30 -05001301 void PixelProgram::CALLNZb(int labelIndex, int callSiteIndex, const Src &boolRegister)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001302 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001303 Bool condition = (*Pointer<Byte>(data + OFFSET(DrawData, ps.b[boolRegister.index])) != Byte(0)); // FIXME
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001304
1305 if(boolRegister.modifier == Shader::MODIFIER_NOT)
1306 {
1307 condition = !condition;
1308 }
1309
1310 if(!labelBlock[labelIndex])
1311 {
1312 labelBlock[labelIndex] = Nucleus::createBasicBlock();
1313 }
1314
1315 if(callRetBlock[labelIndex].size() > 1)
1316 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001317 callStack[stackIndex++] = UInt(callSiteIndex);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001318 }
1319
Nicolas Capens4f172c72016-01-13 08:34:30 -05001320 Int4 restoreLeave = enableLeave;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001321
1322 branch(condition, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]);
1323 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1324
Nicolas Capens4f172c72016-01-13 08:34:30 -05001325 enableLeave = restoreLeave;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001326 }
1327
Nicolas Capens4f172c72016-01-13 08:34:30 -05001328 void PixelProgram::CALLNZp(int labelIndex, int callSiteIndex, const Src &predicateRegister)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001329 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001330 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001331
1332 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1333 {
1334 condition = ~condition;
1335 }
1336
Nicolas Capens4f172c72016-01-13 08:34:30 -05001337 condition &= enableStack[enableIndex];
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001338
1339 if(!labelBlock[labelIndex])
1340 {
1341 labelBlock[labelIndex] = Nucleus::createBasicBlock();
1342 }
1343
1344 if(callRetBlock[labelIndex].size() > 1)
1345 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001346 callStack[stackIndex++] = UInt(callSiteIndex);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001347 }
1348
Nicolas Capens4f172c72016-01-13 08:34:30 -05001349 enableIndex++;
1350 enableStack[enableIndex] = condition;
1351 Int4 restoreLeave = enableLeave;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001352
1353 Bool notAllFalse = SignMask(condition) != 0;
1354 branch(notAllFalse, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]);
1355 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1356
Nicolas Capens4f172c72016-01-13 08:34:30 -05001357 enableIndex--;
1358 enableLeave = restoreLeave;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001359 }
1360
Nicolas Capens4f172c72016-01-13 08:34:30 -05001361 void PixelProgram::ELSE()
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001362 {
1363 ifDepth--;
1364
1365 llvm::BasicBlock *falseBlock = ifFalseBlock[ifDepth];
1366 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1367
1368 if(isConditionalIf[ifDepth])
1369 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001370 Int4 condition = ~enableStack[enableIndex] & enableStack[enableIndex - 1];
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001371 Bool notAllFalse = SignMask(condition) != 0;
1372
1373 branch(notAllFalse, falseBlock, endBlock);
1374
Nicolas Capens4f172c72016-01-13 08:34:30 -05001375 enableStack[enableIndex] = ~enableStack[enableIndex] & enableStack[enableIndex - 1];
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001376 }
1377 else
1378 {
1379 Nucleus::createBr(endBlock);
1380 Nucleus::setInsertBlock(falseBlock);
1381 }
1382
1383 ifFalseBlock[ifDepth] = endBlock;
1384
1385 ifDepth++;
1386 }
1387
Nicolas Capens4f172c72016-01-13 08:34:30 -05001388 void PixelProgram::ENDIF()
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001389 {
1390 ifDepth--;
1391
1392 llvm::BasicBlock *endBlock = ifFalseBlock[ifDepth];
1393
1394 Nucleus::createBr(endBlock);
1395 Nucleus::setInsertBlock(endBlock);
1396
1397 if(isConditionalIf[ifDepth])
1398 {
1399 breakDepth--;
Nicolas Capens4f172c72016-01-13 08:34:30 -05001400 enableIndex--;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001401 }
1402 }
1403
Nicolas Capens4f172c72016-01-13 08:34:30 -05001404 void PixelProgram::ENDLOOP()
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001405 {
1406 loopRepDepth--;
1407
Nicolas Capens4f172c72016-01-13 08:34:30 -05001408 aL[loopDepth] = aL[loopDepth] + increment[loopDepth]; // FIXME: +=
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001409
1410 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1411 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1412
1413 Nucleus::createBr(testBlock);
1414 Nucleus::setInsertBlock(endBlock);
1415
Nicolas Capens4f172c72016-01-13 08:34:30 -05001416 loopDepth--;
1417 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001418 }
1419
Nicolas Capens4f172c72016-01-13 08:34:30 -05001420 void PixelProgram::ENDREP()
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001421 {
1422 loopRepDepth--;
1423
1424 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1425 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1426
1427 Nucleus::createBr(testBlock);
1428 Nucleus::setInsertBlock(endBlock);
1429
Nicolas Capens4f172c72016-01-13 08:34:30 -05001430 loopDepth--;
1431 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001432 }
1433
Nicolas Capens4f172c72016-01-13 08:34:30 -05001434 void PixelProgram::ENDWHILE()
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001435 {
1436 loopRepDepth--;
1437
1438 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1439 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1440
1441 Nucleus::createBr(testBlock);
1442 Nucleus::setInsertBlock(endBlock);
1443
Nicolas Capens4f172c72016-01-13 08:34:30 -05001444 enableIndex--;
1445 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001446 whileTest = false;
1447 }
1448
Nicolas Capens4f172c72016-01-13 08:34:30 -05001449 void PixelProgram::IF(const Src &src)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001450 {
1451 if(src.type == Shader::PARAMETER_CONSTBOOL)
1452 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001453 IFb(src);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001454 }
1455 else if(src.type == Shader::PARAMETER_PREDICATE)
1456 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001457 IFp(src);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001458 }
1459 else
1460 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001461 Int4 condition = As<Int4>(fetchRegisterF(src).x);
1462 IF(condition);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001463 }
1464 }
1465
Nicolas Capens4f172c72016-01-13 08:34:30 -05001466 void PixelProgram::IFb(const Src &boolRegister)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001467 {
1468 ASSERT(ifDepth < 24 + 4);
1469
Nicolas Capens4f172c72016-01-13 08:34:30 -05001470 Bool condition = (*Pointer<Byte>(data + OFFSET(DrawData, ps.b[boolRegister.index])) != Byte(0)); // FIXME
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001471
1472 if(boolRegister.modifier == Shader::MODIFIER_NOT)
1473 {
1474 condition = !condition;
1475 }
1476
1477 llvm::BasicBlock *trueBlock = Nucleus::createBasicBlock();
1478 llvm::BasicBlock *falseBlock = Nucleus::createBasicBlock();
1479
1480 branch(condition, trueBlock, falseBlock);
1481
1482 isConditionalIf[ifDepth] = false;
1483 ifFalseBlock[ifDepth] = falseBlock;
1484
1485 ifDepth++;
1486 }
1487
Nicolas Capens4f172c72016-01-13 08:34:30 -05001488 void PixelProgram::IFp(const Src &predicateRegister)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001489 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001490 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001491
1492 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1493 {
1494 condition = ~condition;
1495 }
1496
Nicolas Capens4f172c72016-01-13 08:34:30 -05001497 IF(condition);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001498 }
1499
Nicolas Capens4f172c72016-01-13 08:34:30 -05001500 void PixelProgram::IFC(Vector4f &src0, Vector4f &src1, Control control)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001501 {
1502 Int4 condition;
1503
1504 switch(control)
1505 {
1506 case Shader::CONTROL_GT: condition = CmpNLE(src0.x, src1.x); break;
1507 case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x); break;
1508 case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x); break;
1509 case Shader::CONTROL_LT: condition = CmpLT(src0.x, src1.x); break;
1510 case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x); break;
1511 case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x); break;
1512 default:
1513 ASSERT(false);
1514 }
1515
Nicolas Capens4f172c72016-01-13 08:34:30 -05001516 IF(condition);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001517 }
1518
Nicolas Capens4f172c72016-01-13 08:34:30 -05001519 void PixelProgram::IF(Int4 &condition)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001520 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001521 condition &= enableStack[enableIndex];
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001522
Nicolas Capens4f172c72016-01-13 08:34:30 -05001523 enableIndex++;
1524 enableStack[enableIndex] = condition;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001525
1526 llvm::BasicBlock *trueBlock = Nucleus::createBasicBlock();
1527 llvm::BasicBlock *falseBlock = Nucleus::createBasicBlock();
1528
1529 Bool notAllFalse = SignMask(condition) != 0;
1530
1531 branch(notAllFalse, trueBlock, falseBlock);
1532
1533 isConditionalIf[ifDepth] = true;
1534 ifFalseBlock[ifDepth] = falseBlock;
1535
1536 ifDepth++;
1537 breakDepth++;
1538 }
1539
1540 void PixelProgram::LABEL(int labelIndex)
1541 {
1542 if(!labelBlock[labelIndex])
1543 {
1544 labelBlock[labelIndex] = Nucleus::createBasicBlock();
1545 }
1546
1547 Nucleus::setInsertBlock(labelBlock[labelIndex]);
1548 currentLabel = labelIndex;
1549 }
1550
Nicolas Capens4f172c72016-01-13 08:34:30 -05001551 void PixelProgram::LOOP(const Src &integerRegister)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001552 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001553 loopDepth++;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001554
Nicolas Capens4f172c72016-01-13 08:34:30 -05001555 iteration[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][0]));
1556 aL[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][1]));
1557 increment[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][2]));
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001558
Nicolas Capens4f172c72016-01-13 08:34:30 -05001559 // If(increment[loopDepth] == 0)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001560 // {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001561 // increment[loopDepth] = 1;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001562 // }
1563
1564 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1565 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1566 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1567
1568 loopRepTestBlock[loopRepDepth] = testBlock;
1569 loopRepEndBlock[loopRepDepth] = endBlock;
1570
1571 // FIXME: jump(testBlock)
1572 Nucleus::createBr(testBlock);
1573 Nucleus::setInsertBlock(testBlock);
1574
Nicolas Capens4f172c72016-01-13 08:34:30 -05001575 branch(iteration[loopDepth] > 0, loopBlock, endBlock);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001576 Nucleus::setInsertBlock(loopBlock);
1577
Nicolas Capens4f172c72016-01-13 08:34:30 -05001578 iteration[loopDepth] = iteration[loopDepth] - 1; // FIXME: --
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001579
1580 loopRepDepth++;
1581 breakDepth = 0;
1582 }
1583
Nicolas Capens4f172c72016-01-13 08:34:30 -05001584 void PixelProgram::REP(const Src &integerRegister)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001585 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001586 loopDepth++;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001587
Nicolas Capens4f172c72016-01-13 08:34:30 -05001588 iteration[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][0]));
1589 aL[loopDepth] = aL[loopDepth - 1];
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001590
1591 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1592 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1593 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1594
1595 loopRepTestBlock[loopRepDepth] = testBlock;
1596 loopRepEndBlock[loopRepDepth] = endBlock;
1597
1598 // FIXME: jump(testBlock)
1599 Nucleus::createBr(testBlock);
1600 Nucleus::setInsertBlock(testBlock);
1601
Nicolas Capens4f172c72016-01-13 08:34:30 -05001602 branch(iteration[loopDepth] > 0, loopBlock, endBlock);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001603 Nucleus::setInsertBlock(loopBlock);
1604
Nicolas Capens4f172c72016-01-13 08:34:30 -05001605 iteration[loopDepth] = iteration[loopDepth] - 1; // FIXME: --
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001606
1607 loopRepDepth++;
1608 breakDepth = 0;
1609 }
1610
Nicolas Capens4f172c72016-01-13 08:34:30 -05001611 void PixelProgram::WHILE(const Src &temporaryRegister)
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001612 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001613 enableIndex++;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001614
1615 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1616 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1617 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1618
1619 loopRepTestBlock[loopRepDepth] = testBlock;
1620 loopRepEndBlock[loopRepDepth] = endBlock;
1621
Nicolas Capens4f172c72016-01-13 08:34:30 -05001622 Int4 restoreBreak = enableBreak;
1623 Int4 restoreContinue = enableContinue;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001624
1625 // FIXME: jump(testBlock)
1626 Nucleus::createBr(testBlock);
1627 Nucleus::setInsertBlock(testBlock);
Nicolas Capens4f172c72016-01-13 08:34:30 -05001628 enableContinue = restoreContinue;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001629
Nicolas Capens4f172c72016-01-13 08:34:30 -05001630 const Vector4f &src = fetchRegisterF(temporaryRegister);
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001631 Int4 condition = As<Int4>(src.x);
Nicolas Capens4f172c72016-01-13 08:34:30 -05001632 condition &= enableStack[enableIndex - 1];
1633 enableStack[enableIndex] = condition;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001634
1635 Bool notAllFalse = SignMask(condition) != 0;
1636 branch(notAllFalse, loopBlock, endBlock);
1637
1638 Nucleus::setInsertBlock(endBlock);
Nicolas Capens4f172c72016-01-13 08:34:30 -05001639 enableBreak = restoreBreak;
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001640
1641 Nucleus::setInsertBlock(loopBlock);
1642
1643 loopRepDepth++;
1644 breakDepth = 0;
1645 }
1646
Nicolas Capens4f172c72016-01-13 08:34:30 -05001647 void PixelProgram::RET()
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001648 {
1649 if(currentLabel == -1)
1650 {
1651 returnBlock = Nucleus::createBasicBlock();
1652 Nucleus::createBr(returnBlock);
1653 }
1654 else
1655 {
1656 llvm::BasicBlock *unreachableBlock = Nucleus::createBasicBlock();
1657
1658 if(callRetBlock[currentLabel].size() > 1) // Pop the return destination from the call stack
1659 {
1660 // FIXME: Encapsulate
Nicolas Capens4f172c72016-01-13 08:34:30 -05001661 UInt index = callStack[--stackIndex];
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001662
1663 llvm::Value *value = index.loadValue();
1664 llvm::Value *switchInst = Nucleus::createSwitch(value, unreachableBlock, (int)callRetBlock[currentLabel].size());
1665
1666 for(unsigned int i = 0; i < callRetBlock[currentLabel].size(); i++)
1667 {
1668 Nucleus::addSwitchCase(switchInst, i, callRetBlock[currentLabel][i]);
1669 }
1670 }
1671 else if(callRetBlock[currentLabel].size() == 1) // Jump directly to the unique return destination
1672 {
1673 Nucleus::createBr(callRetBlock[currentLabel][0]);
1674 }
1675 else // Function isn't called
1676 {
1677 Nucleus::createBr(unreachableBlock);
1678 }
1679
1680 Nucleus::setInsertBlock(unreachableBlock);
1681 Nucleus::createUnreachable();
1682 }
1683 }
1684
Nicolas Capens4f172c72016-01-13 08:34:30 -05001685 void PixelProgram::LEAVE()
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001686 {
Nicolas Capens4f172c72016-01-13 08:34:30 -05001687 enableLeave = enableLeave & ~enableStack[enableIndex];
Alexis Hetuf2a8c372015-07-13 11:08:41 -04001688
1689 // FIXME: Return from function if all instances left
1690 // FIXME: Use enableLeave in other control-flow constructs
1691 }
Nicolas Capens5d961882016-01-01 23:18:14 -05001692}