Merge branch 'mesa_7_6_branch' of git+ssh://agd5f@git.freedesktop.org/git/mesa/mesa
[mesa.git] / src / mesa / drivers / dri / r300 / compiler / r500_fragprog_emit.c
1 /*
2 * Copyright (C) 2005 Ben Skeggs.
3 *
4 * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
5 * Adaptation and modification for ATI/AMD Radeon R500 GPU chipsets.
6 *
7 * All Rights Reserved.
8 *
9 * Permission is hereby granted, free of charge, to any person obtaining
10 * a copy of this software and associated documentation files (the
11 * "Software"), to deal in the Software without restriction, including
12 * without limitation the rights to use, copy, modify, merge, publish,
13 * distribute, sublicense, and/or sell copies of the Software, and to
14 * permit persons to whom the Software is furnished to do so, subject to
15 * the following conditions:
16 *
17 * The above copyright notice and this permission notice (including the
18 * next paragraph) shall be included in all copies or substantial
19 * portions of the Software.
20 *
21 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
22 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
23 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
24 * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
25 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
26 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
27 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
28 *
29 */
30
31 /**
32 * \file
33 *
34 * \author Ben Skeggs <darktama@iinet.net.au>
35 *
36 * \author Jerome Glisse <j.glisse@gmail.com>
37 *
38 * \author Corbin Simpson <MostAwesomeDude@gmail.com>
39 *
40 */
41
42 #include "r500_fragprog.h"
43
44 #include "../r300_reg.h"
45
46 #include "radeon_program_pair.h"
47
48
49 #define PROG_CODE \
50 struct r500_fragment_program_code *code = &c->code->code.r500
51
52 #define error(fmt, args...) do { \
53 rc_error(&c->Base, "%s::%s(): " fmt "\n", \
54 __FILE__, __FUNCTION__, ##args); \
55 } while(0)
56
57
58 struct branch_info {
59 int If;
60 int Else;
61 int Endif;
62 };
63
64 struct emit_state {
65 struct radeon_compiler * C;
66 struct r500_fragment_program_code * Code;
67
68 struct branch_info * Branches;
69 unsigned int CurrentBranchDepth;
70 unsigned int BranchesReserved;
71
72 unsigned int MaxBranchDepth;
73 };
74
75 static unsigned int translate_rgb_op(struct r300_fragment_program_compiler *c, rc_opcode opcode)
76 {
77 switch(opcode) {
78 case RC_OPCODE_CMP: return R500_ALU_RGBA_OP_CMP;
79 case RC_OPCODE_DDX: return R500_ALU_RGBA_OP_MDH;
80 case RC_OPCODE_DDY: return R500_ALU_RGBA_OP_MDV;
81 case RC_OPCODE_DP3: return R500_ALU_RGBA_OP_DP3;
82 case RC_OPCODE_DP4: return R500_ALU_RGBA_OP_DP4;
83 case RC_OPCODE_FRC: return R500_ALU_RGBA_OP_FRC;
84 default:
85 error("translate_rgb_op(%d): unknown opcode\n", opcode);
86 /* fall through */
87 case RC_OPCODE_NOP:
88 /* fall through */
89 case RC_OPCODE_MAD: return R500_ALU_RGBA_OP_MAD;
90 case RC_OPCODE_MAX: return R500_ALU_RGBA_OP_MAX;
91 case RC_OPCODE_MIN: return R500_ALU_RGBA_OP_MIN;
92 case RC_OPCODE_REPL_ALPHA: return R500_ALU_RGBA_OP_SOP;
93 }
94 }
95
96 static unsigned int translate_alpha_op(struct r300_fragment_program_compiler *c, rc_opcode opcode)
97 {
98 switch(opcode) {
99 case RC_OPCODE_CMP: return R500_ALPHA_OP_CMP;
100 case RC_OPCODE_COS: return R500_ALPHA_OP_COS;
101 case RC_OPCODE_DDX: return R500_ALPHA_OP_MDH;
102 case RC_OPCODE_DDY: return R500_ALPHA_OP_MDV;
103 case RC_OPCODE_DP3: return R500_ALPHA_OP_DP;
104 case RC_OPCODE_DP4: return R500_ALPHA_OP_DP;
105 case RC_OPCODE_EX2: return R500_ALPHA_OP_EX2;
106 case RC_OPCODE_FRC: return R500_ALPHA_OP_FRC;
107 case RC_OPCODE_LG2: return R500_ALPHA_OP_LN2;
108 default:
109 error("translate_alpha_op(%d): unknown opcode\n", opcode);
110 /* fall through */
111 case RC_OPCODE_NOP:
112 /* fall through */
113 case RC_OPCODE_MAD: return R500_ALPHA_OP_MAD;
114 case RC_OPCODE_MAX: return R500_ALPHA_OP_MAX;
115 case RC_OPCODE_MIN: return R500_ALPHA_OP_MIN;
116 case RC_OPCODE_RCP: return R500_ALPHA_OP_RCP;
117 case RC_OPCODE_RSQ: return R500_ALPHA_OP_RSQ;
118 case RC_OPCODE_SIN: return R500_ALPHA_OP_SIN;
119 }
120 }
121
122 static unsigned int fix_hw_swizzle(unsigned int swz)
123 {
124 if (swz == 5) swz = 6;
125 if (swz == RC_SWIZZLE_UNUSED) swz = 4;
126 return swz;
127 }
128
129 static unsigned int translate_arg_rgb(struct rc_pair_instruction *inst, int arg)
130 {
131 unsigned int t = inst->RGB.Arg[arg].Source;
132 int comp;
133 t |= inst->RGB.Arg[arg].Negate << 11;
134 t |= inst->RGB.Arg[arg].Abs << 12;
135
136 for(comp = 0; comp < 3; ++comp)
137 t |= fix_hw_swizzle(GET_SWZ(inst->RGB.Arg[arg].Swizzle, comp)) << (3*comp + 2);
138
139 return t;
140 }
141
142 static unsigned int translate_arg_alpha(struct rc_pair_instruction *inst, int i)
143 {
144 unsigned int t = inst->Alpha.Arg[i].Source;
145 t |= fix_hw_swizzle(inst->Alpha.Arg[i].Swizzle) << 2;
146 t |= inst->Alpha.Arg[i].Negate << 5;
147 t |= inst->Alpha.Arg[i].Abs << 6;
148 return t;
149 }
150
151 static uint32_t translate_alu_result_op(struct r300_fragment_program_compiler * c, rc_compare_func func)
152 {
153 switch(func) {
154 case RC_COMPARE_FUNC_EQUAL: return R500_INST_ALU_RESULT_OP_EQ;
155 case RC_COMPARE_FUNC_LESS: return R500_INST_ALU_RESULT_OP_LT;
156 case RC_COMPARE_FUNC_GEQUAL: return R500_INST_ALU_RESULT_OP_GE;
157 case RC_COMPARE_FUNC_NOTEQUAL: return R500_INST_ALU_RESULT_OP_NE;
158 default:
159 rc_error(&c->Base, "%s: unsupported compare func %i\n", __FUNCTION__, func);
160 return 0;
161 }
162 }
163
164 static void use_temporary(struct r500_fragment_program_code* code, unsigned int index)
165 {
166 if (index > code->max_temp_idx)
167 code->max_temp_idx = index;
168 }
169
170 static unsigned int use_source(struct r500_fragment_program_code* code, struct radeon_pair_instruction_source src)
171 {
172 if (src.File == RC_FILE_CONSTANT) {
173 return src.Index | 0x100;
174 } else if (src.File == RC_FILE_TEMPORARY) {
175 use_temporary(code, src.Index);
176 return src.Index;
177 }
178
179 return 0;
180 }
181
182
183 /**
184 * Emit a paired ALU instruction.
185 */
186 static void emit_paired(struct r300_fragment_program_compiler *c, struct rc_pair_instruction *inst)
187 {
188 PROG_CODE;
189
190 if (code->inst_end >= 511) {
191 error("emit_alu: Too many instructions");
192 return;
193 }
194
195 int ip = ++code->inst_end;
196
197 code->inst[ip].inst5 = translate_rgb_op(c, inst->RGB.Opcode);
198 code->inst[ip].inst4 = translate_alpha_op(c, inst->Alpha.Opcode);
199
200 if (inst->RGB.OutputWriteMask || inst->Alpha.OutputWriteMask || inst->Alpha.DepthWriteMask) {
201 code->inst[ip].inst0 = R500_INST_TYPE_OUT;
202 if (inst->WriteALUResult) {
203 error("%s: cannot write output and ALU result at the same time");
204 return;
205 }
206 } else {
207 code->inst[ip].inst0 = R500_INST_TYPE_ALU;
208 }
209 code->inst[ip].inst0 |= R500_INST_TEX_SEM_WAIT;
210
211 code->inst[ip].inst0 |= (inst->RGB.WriteMask << 11) | (inst->Alpha.WriteMask << 14);
212 code->inst[ip].inst0 |= (inst->RGB.OutputWriteMask << 15) | (inst->Alpha.OutputWriteMask << 18);
213 if (inst->Alpha.DepthWriteMask) {
214 code->inst[ip].inst4 |= R500_ALPHA_W_OMASK;
215 c->code->writes_depth = 1;
216 }
217
218 code->inst[ip].inst4 |= R500_ALPHA_ADDRD(inst->Alpha.DestIndex);
219 code->inst[ip].inst5 |= R500_ALU_RGBA_ADDRD(inst->RGB.DestIndex);
220 use_temporary(code, inst->Alpha.DestIndex);
221 use_temporary(code, inst->RGB.DestIndex);
222
223 if (inst->RGB.Saturate)
224 code->inst[ip].inst0 |= R500_INST_RGB_CLAMP;
225 if (inst->Alpha.Saturate)
226 code->inst[ip].inst0 |= R500_INST_ALPHA_CLAMP;
227
228 code->inst[ip].inst1 |= R500_RGB_ADDR0(use_source(code, inst->RGB.Src[0]));
229 code->inst[ip].inst1 |= R500_RGB_ADDR1(use_source(code, inst->RGB.Src[1]));
230 code->inst[ip].inst1 |= R500_RGB_ADDR2(use_source(code, inst->RGB.Src[2]));
231
232 code->inst[ip].inst2 |= R500_ALPHA_ADDR0(use_source(code, inst->Alpha.Src[0]));
233 code->inst[ip].inst2 |= R500_ALPHA_ADDR1(use_source(code, inst->Alpha.Src[1]));
234 code->inst[ip].inst2 |= R500_ALPHA_ADDR2(use_source(code, inst->Alpha.Src[2]));
235
236 code->inst[ip].inst3 |= translate_arg_rgb(inst, 0) << R500_ALU_RGB_SEL_A_SHIFT;
237 code->inst[ip].inst3 |= translate_arg_rgb(inst, 1) << R500_ALU_RGB_SEL_B_SHIFT;
238 code->inst[ip].inst5 |= translate_arg_rgb(inst, 2) << R500_ALU_RGBA_SEL_C_SHIFT;
239
240 code->inst[ip].inst4 |= translate_arg_alpha(inst, 0) << R500_ALPHA_SEL_A_SHIFT;
241 code->inst[ip].inst4 |= translate_arg_alpha(inst, 1) << R500_ALPHA_SEL_B_SHIFT;
242 code->inst[ip].inst5 |= translate_arg_alpha(inst, 2) << R500_ALU_RGBA_ALPHA_SEL_C_SHIFT;
243
244 if (inst->WriteALUResult) {
245 code->inst[ip].inst3 |= R500_ALU_RGB_WMASK;
246
247 if (inst->WriteALUResult == RC_ALURESULT_X)
248 code->inst[ip].inst0 |= R500_INST_ALU_RESULT_SEL_RED;
249 else
250 code->inst[ip].inst0 |= R500_INST_ALU_RESULT_SEL_ALPHA;
251
252 code->inst[ip].inst0 |= translate_alu_result_op(c, inst->ALUResultCompare);
253 }
254 }
255
256 static unsigned int translate_strq_swizzle(unsigned int swizzle)
257 {
258 unsigned int swiz = 0;
259 int i;
260 for (i = 0; i < 4; i++)
261 swiz |= (GET_SWZ(swizzle, i) & 0x3) << i*2;
262 return swiz;
263 }
264
265 /**
266 * Emit a single TEX instruction
267 */
268 static int emit_tex(struct r300_fragment_program_compiler *c, struct rc_sub_instruction *inst)
269 {
270 PROG_CODE;
271
272 if (code->inst_end >= 511) {
273 error("emit_tex: Too many instructions");
274 return 0;
275 }
276
277 int ip = ++code->inst_end;
278
279 code->inst[ip].inst0 = R500_INST_TYPE_TEX
280 | (inst->DstReg.WriteMask << 11)
281 | R500_INST_TEX_SEM_WAIT;
282 code->inst[ip].inst1 = R500_TEX_ID(inst->TexSrcUnit)
283 | R500_TEX_SEM_ACQUIRE | R500_TEX_IGNORE_UNCOVERED;
284
285 if (inst->TexSrcTarget == RC_TEXTURE_RECT)
286 code->inst[ip].inst1 |= R500_TEX_UNSCALED;
287
288 switch (inst->Opcode) {
289 case RC_OPCODE_KIL:
290 code->inst[ip].inst1 |= R500_TEX_INST_TEXKILL;
291 break;
292 case RC_OPCODE_TEX:
293 code->inst[ip].inst1 |= R500_TEX_INST_LD;
294 break;
295 case RC_OPCODE_TXB:
296 code->inst[ip].inst1 |= R500_TEX_INST_LODBIAS;
297 break;
298 case RC_OPCODE_TXP:
299 code->inst[ip].inst1 |= R500_TEX_INST_PROJ;
300 break;
301 default:
302 error("emit_tex can't handle opcode %x\n", inst->Opcode);
303 }
304
305 use_temporary(code, inst->SrcReg[0].Index);
306 if (inst->Opcode != RC_OPCODE_KIL)
307 use_temporary(code, inst->DstReg.Index);
308
309 code->inst[ip].inst2 = R500_TEX_SRC_ADDR(inst->SrcReg[0].Index)
310 | (translate_strq_swizzle(inst->SrcReg[0].Swizzle) << 8)
311 | R500_TEX_DST_ADDR(inst->DstReg.Index)
312 | R500_TEX_DST_R_SWIZ_R | R500_TEX_DST_G_SWIZ_G
313 | R500_TEX_DST_B_SWIZ_B | R500_TEX_DST_A_SWIZ_A;
314
315 return 1;
316 }
317
318 static void grow_branches(struct emit_state * s)
319 {
320 unsigned int newreserved = s->BranchesReserved * 2;
321 struct branch_info * newbranches;
322
323 if (!newreserved)
324 newreserved = 4;
325
326 newbranches = memory_pool_malloc(&s->C->Pool, newreserved*sizeof(struct branch_info));
327 memcpy(newbranches, s->Branches, s->CurrentBranchDepth*sizeof(struct branch_info));
328
329 s->Branches = newbranches;
330 s->BranchesReserved = newreserved;
331 }
332
333 static void emit_flowcontrol(struct emit_state * s, struct rc_instruction * inst)
334 {
335 if (s->Code->inst_end >= 511) {
336 rc_error(s->C, "emit_tex: Too many instructions");
337 return;
338 }
339
340 unsigned int newip = ++s->Code->inst_end;
341
342 s->Code->inst[newip].inst0 = R500_INST_TYPE_FC | R500_INST_ALU_WAIT;
343
344 if (inst->U.I.Opcode == RC_OPCODE_IF) {
345 if (s->CurrentBranchDepth >= 32) {
346 rc_error(s->C, "Branch depth exceeds hardware limit");
347 return;
348 }
349
350 if (s->CurrentBranchDepth >= s->BranchesReserved)
351 grow_branches(s);
352
353 struct branch_info * branch = &s->Branches[s->CurrentBranchDepth++];
354 branch->If = newip;
355 branch->Else = -1;
356 branch->Endif = -1;
357
358 if (s->CurrentBranchDepth > s->MaxBranchDepth)
359 s->MaxBranchDepth = s->CurrentBranchDepth;
360
361 /* actual instruction is filled in at ENDIF time */
362 } else if (inst->U.I.Opcode == RC_OPCODE_ELSE) {
363 if (!s->CurrentBranchDepth) {
364 rc_error(s->C, "%s: got ELSE outside a branch", __FUNCTION__);
365 return;
366 }
367
368 struct branch_info * branch = &s->Branches[s->CurrentBranchDepth - 1];
369 branch->Else = newip;
370
371 /* actual instruction is filled in at ENDIF time */
372 } else if (inst->U.I.Opcode == RC_OPCODE_ENDIF) {
373 if (!s->CurrentBranchDepth) {
374 rc_error(s->C, "%s: got ELSE outside a branch", __FUNCTION__);
375 return;
376 }
377
378 struct branch_info * branch = &s->Branches[s->CurrentBranchDepth - 1];
379 branch->Endif = newip;
380
381 s->Code->inst[branch->If].inst2 = R500_FC_OP_JUMP
382 | R500_FC_A_OP_NONE /* no address stack */
383 | R500_FC_JUMP_FUNC(0x0f) /* jump if ALU result is false */
384 | R500_FC_B_OP0_INCR /* increment branch counter if stay */
385 ;
386
387 if (branch->Else >= 0) {
388 /* increment branch counter also if jump */
389 s->Code->inst[branch->If].inst2 |= R500_FC_B_OP1_INCR;
390 s->Code->inst[branch->If].inst3 = R500_FC_JUMP_ADDR(branch->Else + 1);
391
392 s->Code->inst[branch->Else].inst2 = R500_FC_OP_JUMP
393 | R500_FC_A_OP_NONE /* no address stack */
394 | R500_FC_B_ELSE /* all active pixels want to jump */
395 | R500_FC_B_OP0_NONE /* no counter op if stay */
396 | R500_FC_B_OP1_DECR /* decrement branch counter if jump */
397 | R500_FC_B_POP_CNT(1)
398 ;
399 s->Code->inst[branch->Else].inst3 = R500_FC_JUMP_ADDR(branch->Endif + 1);
400 } else {
401 /* don't touch branch counter on jump */
402 s->Code->inst[branch->If].inst2 |= R500_FC_B_OP1_NONE;
403 s->Code->inst[branch->If].inst3 = R500_FC_JUMP_ADDR(branch->Endif + 1);
404 }
405
406 s->Code->inst[branch->Endif].inst2 = R500_FC_OP_JUMP
407 | R500_FC_A_OP_NONE /* no address stack */
408 | R500_FC_JUMP_ANY /* docs says set this, but I don't understand why */
409 | R500_FC_B_OP0_DECR /* decrement branch counter if stay */
410 | R500_FC_B_OP1_NONE /* no branch counter if stay */
411 | R500_FC_B_POP_CNT(1)
412 ;
413 s->Code->inst[branch->Endif].inst3 = R500_FC_JUMP_ADDR(branch->Endif + 1);
414
415 s->CurrentBranchDepth--;
416 } else {
417 rc_error(s->C, "%s: unknown opcode %i\n", __FUNCTION__, inst->U.I.Opcode);
418 }
419 }
420
421 void r500BuildFragmentProgramHwCode(struct r300_fragment_program_compiler *compiler)
422 {
423 struct emit_state s;
424 struct r500_fragment_program_code *code = &compiler->code->code.r500;
425
426 memset(&s, 0, sizeof(s));
427 s.C = &compiler->Base;
428 s.Code = code;
429
430 memset(code, 0, sizeof(*code));
431 code->max_temp_idx = 1;
432 code->inst_end = -1;
433
434 for(struct rc_instruction * inst = compiler->Base.Program.Instructions.Next;
435 inst != &compiler->Base.Program.Instructions && !compiler->Base.Error;
436 inst = inst->Next) {
437 if (inst->Type == RC_INSTRUCTION_NORMAL) {
438 const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
439
440 if (opcode->IsFlowControl) {
441 emit_flowcontrol(&s, inst);
442 } else if (inst->U.I.Opcode == RC_OPCODE_BEGIN_TEX) {
443 continue;
444 } else {
445 emit_tex(compiler, &inst->U.I);
446 }
447 } else {
448 emit_paired(compiler, &inst->U.P);
449 }
450 }
451
452 if (code->max_temp_idx >= 128)
453 rc_error(&compiler->Base, "Too many hardware temporaries used");
454
455 if (compiler->Base.Error)
456 return;
457
458 if ((code->inst[code->inst_end].inst0 & R500_INST_TYPE_MASK) != R500_INST_TYPE_OUT) {
459 /* This may happen when dead-code elimination is disabled or
460 * when most of the fragment program logic is leading to a KIL */
461 if (code->inst_end >= 511) {
462 rc_error(&compiler->Base, "Introducing fake OUT: Too many instructions");
463 return;
464 }
465
466 int ip = ++code->inst_end;
467 code->inst[ip].inst0 = R500_INST_TYPE_OUT | R500_INST_TEX_SEM_WAIT;
468 }
469
470 if (s.MaxBranchDepth >= 4) {
471 if (code->max_temp_idx < 1)
472 code->max_temp_idx = 1;
473
474 code->us_fc_ctrl |= R500_FC_FULL_FC_EN;
475 }
476 }