1/*
2 * Copyright 2011 Christoph Bumiller
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17 * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
18 * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
19 * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
20 * SOFTWARE.
21 */
22
23#include "nv50_ir_target_nv50.h"
24
25namespace nv50_ir {
26
27Target *getTargetNV50(unsigned int chipset)
28{
29   return new TargetNV50(chipset);
30}
31
32TargetNV50::TargetNV50(unsigned int card) : Target(true, false)
33{
34   chipset = card;
35
36   wposMask = 0;
37   for (unsigned int i = 0; i <= SV_LAST; ++i)
38      sysvalLocation[i] = ~0;
39
40   initOpInfo();
41}
42
43#if 0
44// BULTINS / LIBRARY FUNCTIONS:
45
46// TODO
47static const uint32_t nvc0_builtin_code[] =
48{
49};
50
51static const uint16_t nvc0_builtin_offsets[NV50_BUILTIN_COUNT] =
52{
53};
54#endif
55
56void
57TargetNV50::getBuiltinCode(const uint32_t **code, uint32_t *size) const
58{
59   *code = NULL;
60   *size = 0;
61}
62
63uint32_t
64TargetNV50::getBuiltinOffset(int builtin) const
65{
66   return 0;
67}
68
69struct opProperties
70{
71   operation op;
72   unsigned int mNeg    : 4;
73   unsigned int mAbs    : 4;
74   unsigned int mNot    : 4;
75   unsigned int mSat    : 4;
76   unsigned int fConst  : 3;
77   unsigned int fShared : 3;
78   unsigned int fAttrib : 3;
79   unsigned int fImm    : 3;
80};
81
82static const struct opProperties _initProps[] =
83{
84   //           neg  abs  not  sat  c[]  s[], a[], imm
85   { OP_ADD,    0x3, 0x0, 0x0, 0x8, 0x2, 0x1, 0x1, 0x2 },
86   { OP_SUB,    0x3, 0x0, 0x0, 0x0, 0x2, 0x1, 0x1, 0x2 },
87   { OP_MUL,    0x3, 0x0, 0x0, 0x0, 0x2, 0x1, 0x1, 0x2 },
88   { OP_MAX,    0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 },
89   { OP_MIN,    0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 },
90   { OP_MAD,    0x7, 0x0, 0x0, 0x0, 0x6, 0x1, 0x1, 0x0 }, // special constraint
91   { OP_ABS,    0x0, 0x0, 0x0, 0x0, 0x0, 0x1, 0x1, 0x0 },
92   { OP_NEG,    0x0, 0x1, 0x0, 0x0, 0x0, 0x1, 0x1, 0x0 },
93   { OP_CVT,    0x1, 0x1, 0x0, 0x8, 0x0, 0x1, 0x1, 0x0 },
94   { OP_AND,    0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 },
95   { OP_OR,     0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 },
96   { OP_XOR,    0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 },
97   { OP_SHL,    0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x2 },
98   { OP_SHR,    0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x2 },
99   { OP_SET,    0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 },
100   { OP_PREEX2, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
101   { OP_PRESIN, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
102   { OP_LG2,    0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
103   { OP_RCP,    0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
104   { OP_RSQ,    0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
105   { OP_DFDX,   0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
106   { OP_DFDY,   0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
107};
108
109void TargetNV50::initOpInfo()
110{
111   unsigned int i, j;
112
113   static const uint32_t commutative[(OP_LAST + 31) / 32] =
114   {
115      // ADD,MAD,MUL,AND,OR,XOR,MAX,MIN
116      0x0670ca00, 0x0000003f, 0x00000000
117   };
118   static const uint32_t shortForm[(OP_LAST + 31) / 32] =
119   {
120      // MOV,ADD,SUB,MUL,SAD,L/PINTERP,RCP,TEX,TXF
121      0x00010e40, 0x00000040, 0x00000498
122   };
123   static const operation noDestList[] =
124   {
125      OP_STORE, OP_WRSV, OP_EXPORT, OP_BRA, OP_CALL, OP_RET, OP_EXIT,
126      OP_DISCARD, OP_CONT, OP_BREAK, OP_PRECONT, OP_PREBREAK, OP_PRERET,
127      OP_JOIN, OP_JOINAT, OP_BRKPT, OP_MEMBAR, OP_EMIT, OP_RESTART,
128      OP_QUADON, OP_QUADPOP
129   };
130   static const operation noPredList[] =
131   {
132      OP_CALL, OP_PREBREAK, OP_PRERET, OP_QUADON, OP_QUADPOP, OP_JOINAT
133   };
134
135   for (i = 0; i < DATA_FILE_COUNT; ++i)
136      nativeFileMap[i] = (DataFile)i;
137   nativeFileMap[FILE_PREDICATE] = FILE_FLAGS;
138
139   for (i = 0; i < OP_LAST; ++i) {
140      opInfo[i].variants = NULL;
141      opInfo[i].op = (operation)i;
142      opInfo[i].srcTypes = 1 << (int)TYPE_F32;
143      opInfo[i].dstTypes = 1 << (int)TYPE_F32;
144      opInfo[i].immdBits = 0xffffffff;
145      opInfo[i].srcNr = operationSrcNr[i];
146
147      for (j = 0; j < opInfo[i].srcNr; ++j) {
148         opInfo[i].srcMods[j] = 0;
149         opInfo[i].srcFiles[j] = 1 << (int)FILE_GPR;
150      }
151      opInfo[i].dstMods = 0;
152      opInfo[i].dstFiles = 1 << (int)FILE_GPR;
153
154      opInfo[i].hasDest = 1;
155      opInfo[i].vector = (i >= OP_TEX && i <= OP_TEXCSAA);
156      opInfo[i].commutative = (commutative[i / 32] >> (i % 32)) & 1;
157      opInfo[i].pseudo = (i < OP_MOV);
158      opInfo[i].predicate = !opInfo[i].pseudo;
159      opInfo[i].flow = (i >= OP_BRA && i <= OP_JOIN);
160      opInfo[i].minEncSize = (shortForm[i / 32] & (1 << (i % 32))) ? 4 : 8;
161   }
162   for (i = 0; i < sizeof(noDestList) / sizeof(noDestList[0]); ++i)
163      opInfo[noDestList[i]].hasDest = 0;
164   for (i = 0; i < sizeof(noPredList) / sizeof(noPredList[0]); ++i)
165      opInfo[noPredList[i]].predicate = 0;
166
167   for (i = 0; i < sizeof(_initProps) / sizeof(_initProps[0]); ++i) {
168      const struct opProperties *prop = &_initProps[i];
169
170      for (int s = 0; s < 3; ++s) {
171         if (prop->mNeg & (1 << s))
172            opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NEG;
173         if (prop->mAbs & (1 << s))
174            opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_ABS;
175         if (prop->mNot & (1 << s))
176            opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NOT;
177         if (prop->fConst & (1 << s))
178            opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_MEMORY_CONST;
179         if (prop->fShared & (1 << s))
180            opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_MEMORY_SHARED;
181         if (prop->fAttrib & (1 << s))
182            opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_SHADER_INPUT;
183         if (prop->fImm & (1 << s))
184            opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_IMMEDIATE;
185      }
186      if (prop->mSat & 8)
187         opInfo[prop->op].dstMods = NV50_IR_MOD_SAT;
188   }
189}
190
191unsigned int
192TargetNV50::getFileSize(DataFile file) const
193{
194   switch (file) {
195   case FILE_NULL:          return 0;
196   case FILE_GPR:           return 256; // in 16-bit units **
197   case FILE_PREDICATE:     return 0;
198   case FILE_FLAGS:         return 4;
199   case FILE_ADDRESS:       return 4;
200   case FILE_IMMEDIATE:     return 0;
201   case FILE_MEMORY_CONST:  return 65536;
202   case FILE_SHADER_INPUT:  return 0x200;
203   case FILE_SHADER_OUTPUT: return 0x200;
204   case FILE_MEMORY_GLOBAL: return 0xffffffff;
205   case FILE_MEMORY_SHARED: return 16 << 10;
206   case FILE_MEMORY_LOCAL:  return 48 << 10;
207   case FILE_SYSTEM_VALUE:  return 16;
208   default:
209      assert(!"invalid file");
210      return 0;
211   }
212   // ** only first 128 units encodable for 16-bit regs
213}
214
215unsigned int
216TargetNV50::getFileUnit(DataFile file) const
217{
218   if (file == FILE_GPR || file == FILE_ADDRESS)
219      return 1;
220   if (file == FILE_SYSTEM_VALUE)
221      return 2;
222   return 0;
223}
224
225uint32_t
226TargetNV50::getSVAddress(DataFile shaderFile, const Symbol *sym) const
227{
228   switch (sym->reg.data.sv.sv) {
229   case SV_FACE:
230      return 0x3fc;
231   case SV_POSITION:
232   {
233      uint32_t addr = sysvalLocation[sym->reg.data.sv.sv];
234      for (int c = 0; c < sym->reg.data.sv.index; ++c)
235         if (wposMask & (1 << c))
236            addr += 4;
237      return addr;
238   }
239   case SV_NCTAID:
240      return 0x8 + 2 * sym->reg.data.sv.index;
241   case SV_CTAID:
242      return 0xc + 2 * sym->reg.data.sv.index;
243   case SV_NTID:
244      return 0x2 + 2 * sym->reg.data.sv.index;
245   case SV_TID:
246      return 0;
247   default:
248      return sysvalLocation[sym->reg.data.sv.sv];
249   }
250}
251
252// long:  rrr, arr, rcr, acr, rrc, arc, gcr, grr
253// short: rr, ar, rc, gr
254// immd:  ri, gi
255bool
256TargetNV50::insnCanLoad(const Instruction *i, int s,
257                        const Instruction *ld) const
258{
259   DataFile sf = ld->src(0).getFile();
260
261   if (sf == FILE_IMMEDIATE && (i->predSrc >= 0 || i->flagsDef >= 0))
262      return false;
263   if (s >= opInfo[i->op].srcNr)
264      return false;
265   if (!(opInfo[i->op].srcFiles[s] & (1 << (int)sf)))
266      return false;
267   if (s == 2 && i->src(1).getFile() != FILE_GPR)
268      return false;
269
270   // NOTE: don't rely on flagsDef
271   for (int d = 0; i->defExists(d); ++d)
272      if (i->def(d).getFile() == FILE_FLAGS)
273         return false;
274
275   unsigned mode = 0;
276
277   for (int z = 0; z < Target::operationSrcNr[i->op]; ++z) {
278      DataFile zf = (z == s) ? sf : i->src(z).getFile();
279      switch (zf) {
280      case FILE_GPR:
281         break;
282      case FILE_MEMORY_SHARED:
283      case FILE_SHADER_INPUT:
284         mode |= 1 << (z * 2);
285         break;
286      case FILE_MEMORY_CONST:
287         mode |= 2 << (z * 2);
288         break;
289      case FILE_IMMEDIATE:
290         mode |= 3 << (z * 2);
291      default:
292         break;
293      }
294   }
295
296   switch (mode) {
297   case 0x00:
298   case 0x01:
299   case 0x03:
300   case 0x08:
301   case 0x09:
302   case 0x0c:
303   case 0x20:
304   case 0x21:
305      break;
306   case 0x0d:
307      if (ld->bb->getProgram()->getType() != Program::TYPE_GEOMETRY)
308         return false;
309   default:
310      return false;
311   }
312
313   uint8_t ldSize;
314
315   if ((i->op == OP_MUL || i->op == OP_MAD) && !isFloatType(i->dType)) {
316      // 32-bit MUL will be split into 16-bit MULs
317      if (ld->src(0).isIndirect(0))
318         return false;
319      if (sf == FILE_IMMEDIATE)
320         return false;
321      ldSize = 2;
322   } else {
323      ldSize = typeSizeof(ld->dType);
324   }
325
326   if (sf == FILE_IMMEDIATE)
327      return true;
328
329
330   // Check if memory access is encodable:
331
332   if (ldSize < 4 && sf == FILE_SHADER_INPUT) // no < 4-byte aligned a[] access
333      return false;
334   if (ld->getSrc(0)->reg.data.offset > (int32_t)(127 * ldSize))
335      return false;
336
337   if (ld->src(0).isIndirect(0)) {
338      for (int z = 0; i->srcExists(z); ++z)
339         if (i->src(z).isIndirect(0))
340            return false;
341
342      // s[] access only possible in CP, $aX always applies
343      if (sf == FILE_MEMORY_SHARED)
344         return true;
345      if (!ld->bb) // can't check type ...
346         return false;
347      Program::Type pt = ld->bb->getProgram()->getType();
348
349      // $aX applies to c[] only in VP, FP, GP if p[] is not accessed
350      if (pt == Program::TYPE_COMPUTE)
351         return false;
352      if (pt == Program::TYPE_GEOMETRY) {
353         if (sf == FILE_MEMORY_CONST)
354            return i->src(s).getFile() != FILE_SHADER_INPUT;
355         return sf == FILE_SHADER_INPUT;
356      }
357      return sf == FILE_MEMORY_CONST;
358   }
359   return true;
360}
361
362bool
363TargetNV50::isAccessSupported(DataFile file, DataType ty) const
364{
365   if (ty == TYPE_B96 || ty == TYPE_NONE)
366      return false;
367   if (typeSizeof(ty) > 4)
368      return (file == FILE_MEMORY_LOCAL) || (file == FILE_MEMORY_GLOBAL);
369   return true;
370}
371
372bool
373TargetNV50::isOpSupported(operation op, DataType ty) const
374{
375   if (ty == TYPE_F64 && chipset < 0xa0)
376      return false;
377
378   switch (op) {
379   case OP_PRERET:
380      return chipset >= 0xa0;
381   case OP_TXG:
382      return chipset >= 0xa3;
383   case OP_POW:
384   case OP_SQRT:
385   case OP_DIV:
386   case OP_MOD:
387   case OP_SET_AND:
388   case OP_SET_OR:
389   case OP_SET_XOR:
390   case OP_SLCT:
391   case OP_SELP:
392   case OP_POPCNT:
393   case OP_INSBF:
394   case OP_EXTBF:
395   case OP_EXIT: // want exit modifier instead (on NOP if required)
396      return false;
397   case OP_SAD:
398      return ty == TYPE_S32;
399   default:
400      return true;
401   }
402}
403
404bool
405TargetNV50::isModSupported(const Instruction *insn, int s, Modifier mod) const
406{
407   if (!isFloatType(insn->dType)) {
408      switch (insn->op) {
409      case OP_ABS:
410      case OP_NEG:
411      case OP_CVT:
412      case OP_CEIL:
413      case OP_FLOOR:
414      case OP_TRUNC:
415      case OP_AND:
416      case OP_OR:
417      case OP_XOR:
418         break;
419      case OP_ADD:
420         if (insn->src(s ? 0 : 1).mod.neg())
421            return false;
422         break;
423      case OP_SUB:
424         if (s == 0)
425            return insn->src(1).mod.neg() ? false : true;
426         break;
427      case OP_SET:
428         if (insn->sType != TYPE_F32)
429            return false;
430         break;
431      default:
432         return false;
433      }
434   }
435   if (s > 3)
436      return false;
437   return (mod & Modifier(opInfo[insn->op].srcMods[s])) == mod;
438}
439
440bool
441TargetNV50::mayPredicate(const Instruction *insn, const Value *pred) const
442{
443   if (insn->getPredicate() || insn->flagsSrc >= 0)
444      return false;
445   for (int s = 0; insn->srcExists(s); ++s)
446      if (insn->src(s).getFile() == FILE_IMMEDIATE)
447         return false;
448   return opInfo[insn->op].predicate;
449}
450
451bool
452TargetNV50::isSatSupported(const Instruction *insn) const
453{
454   if (insn->op == OP_CVT)
455      return true;
456   if (insn->dType != TYPE_F32)
457      return false;
458   return opInfo[insn->op].dstMods & NV50_IR_MOD_SAT;
459}
460
461int TargetNV50::getLatency(const Instruction *i) const
462{
463   // TODO: tune these values
464   if (i->op == OP_LOAD) {
465      switch (i->src(0).getFile()) {
466      case FILE_MEMORY_LOCAL:
467      case FILE_MEMORY_GLOBAL:
468         return 100; // really 400 to 800
469      default:
470         return 22;
471      }
472   }
473   return 22;
474}
475
476// These are "inverse" throughput values, i.e. the number of cycles required
477// to issue a specific instruction for a full warp (32 threads).
478//
479// Assuming we have more than 1 warp in flight, a higher issue latency results
480// in a lower result latency since the MP will have spent more time with other
481// warps.
482// This also helps to determine the number of cycles between instructions in
483// a single warp.
484//
485int TargetNV50::getThroughput(const Instruction *i) const
486{
487   // TODO: tune these values
488   if (i->dType == TYPE_F32) {
489      switch (i->op) {
490      case OP_RCP:
491      case OP_RSQ:
492      case OP_LG2:
493      case OP_SIN:
494      case OP_COS:
495      case OP_PRESIN:
496      case OP_PREEX2:
497         return 16;
498      default:
499         return 4;
500      }
501   } else
502   if (i->dType == TYPE_U32 || i->dType == TYPE_S32) {
503      return 4;
504   } else
505   if (i->dType == TYPE_F64) {
506      return 32;
507   } else {
508      return 1;
509   }
510}
511
512static void
513recordLocation(uint16_t *locs, uint8_t *masks,
514               const struct nv50_ir_varying *var)
515{
516   uint16_t addr = var->slot[0] * 4;
517
518   switch (var->sn) {
519   case TGSI_SEMANTIC_POSITION: locs[SV_POSITION] = addr; break;
520   case TGSI_SEMANTIC_INSTANCEID: locs[SV_INSTANCE_ID] = addr; break;
521   case TGSI_SEMANTIC_VERTEXID: locs[SV_VERTEX_ID] = addr; break;
522   case TGSI_SEMANTIC_PRIMID: locs[SV_PRIMITIVE_ID] = addr; break;
523   case NV50_SEMANTIC_LAYER: locs[SV_LAYER] = addr; break;
524   case NV50_SEMANTIC_VIEWPORTINDEX: locs[SV_VIEWPORT_INDEX] = addr; break;
525   default:
526      break;
527   }
528   if (var->sn == TGSI_SEMANTIC_POSITION && masks)
529      masks[0] = var->mask;
530}
531
532void
533TargetNV50::parseDriverInfo(const struct nv50_ir_prog_info *info)
534{
535   unsigned int i;
536   for (i = 0; i < info->numOutputs; ++i)
537      recordLocation(sysvalLocation, NULL, &info->out[i]);
538   for (i = 0; i < info->numInputs; ++i)
539      recordLocation(sysvalLocation, &wposMask, &info->in[i]);
540   for (i = 0; i < info->numSysVals; ++i)
541      recordLocation(sysvalLocation, NULL, &info->sv[i]);
542
543   if (sysvalLocation[SV_POSITION] >= 0x200) {
544      // not assigned by driver, but we need it internally
545      wposMask = 0x8;
546      sysvalLocation[SV_POSITION] = 0;
547   }
548}
549
550} // namespace nv50_ir
551