• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  */
22 
23 #include "codegen/nv50_ir_target_nvc0.h"
24 
25 namespace nv50_ir {
26 
getTargetNVC0(unsigned int chipset)27 Target *getTargetNVC0(unsigned int chipset)
28 {
29    return new TargetNVC0(chipset);
30 }
31 
TargetNVC0(unsigned int card)32 TargetNVC0::TargetNVC0(unsigned int card) :
33    Target(card < 0x110, false, card >= 0xe4)
34 {
35    chipset = card;
36    initOpInfo();
37 }
38 
39 // BULTINS / LIBRARY FUNCTIONS:
40 
41 // lazyness -> will just hardcode everything for the time being
42 
43 #include "lib/gf100.asm.h"
44 #include "lib/gk104.asm.h"
45 #include "lib/gk110.asm.h"
46 
47 void
getBuiltinCode(const uint32_t ** code,uint32_t * size) const48 TargetNVC0::getBuiltinCode(const uint32_t **code, uint32_t *size) const
49 {
50    switch (chipset & ~0xf) {
51    case 0xe0:
52       if (chipset < NVISA_GK20A_CHIPSET) {
53          *code = (const uint32_t *)&gk104_builtin_code[0];
54          *size = sizeof(gk104_builtin_code);
55          break;
56       }
57       /* fall-through for GK20A */
58    case 0xf0:
59    case 0x100:
60       *code = (const uint32_t *)&gk110_builtin_code[0];
61       *size = sizeof(gk110_builtin_code);
62       break;
63    default:
64       *code = (const uint32_t *)&gf100_builtin_code[0];
65       *size = sizeof(gf100_builtin_code);
66       break;
67    }
68 }
69 
70 uint32_t
getBuiltinOffset(int builtin) const71 TargetNVC0::getBuiltinOffset(int builtin) const
72 {
73    assert(builtin < NVC0_BUILTIN_COUNT);
74 
75    switch (chipset & ~0xf) {
76    case 0xe0:
77       if (chipset < NVISA_GK20A_CHIPSET)
78          return gk104_builtin_offsets[builtin];
79       /* fall-through for GK20A */
80    case 0xf0:
81    case 0x100:
82       return gk110_builtin_offsets[builtin];
83    default:
84       return gf100_builtin_offsets[builtin];
85    }
86 }
87 
88 struct opProperties
89 {
90    operation op;
91    unsigned int mNeg   : 4;
92    unsigned int mAbs   : 4;
93    unsigned int mNot   : 4;
94    unsigned int mSat   : 4;
95    unsigned int fConst : 3;
96    unsigned int fImmd  : 4; // last bit indicates if full immediate is suppoted
97 };
98 
99 static const struct opProperties _initProps[] =
100 {
101    //           neg  abs  not  sat  c[]  imm
102    { OP_ADD,    0x3, 0x3, 0x0, 0x8, 0x2, 0x2 | 0x8 },
103    { OP_SUB,    0x3, 0x3, 0x0, 0x0, 0x2, 0x2 | 0x8 },
104    { OP_MUL,    0x3, 0x0, 0x0, 0x8, 0x2, 0x2 | 0x8 },
105    { OP_MAX,    0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
106    { OP_MIN,    0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
107    { OP_MAD,    0x7, 0x0, 0x0, 0x8, 0x6, 0x2 }, // special c[] constraint
108    { OP_FMA,    0x7, 0x0, 0x0, 0x8, 0x6, 0x2 }, // keep the same as OP_MAD
109    { OP_SHLADD, 0x5, 0x0, 0x0, 0x0, 0x4, 0x6 },
110    { OP_MADSP,  0x0, 0x0, 0x0, 0x0, 0x6, 0x2 },
111    { OP_ABS,    0x0, 0x0, 0x0, 0x0, 0x1, 0x0 },
112    { OP_NEG,    0x0, 0x1, 0x0, 0x0, 0x1, 0x0 },
113    { OP_CVT,    0x1, 0x1, 0x0, 0x8, 0x1, 0x0 },
114    { OP_CEIL,   0x1, 0x1, 0x0, 0x8, 0x1, 0x0 },
115    { OP_FLOOR,  0x1, 0x1, 0x0, 0x8, 0x1, 0x0 },
116    { OP_TRUNC,  0x1, 0x1, 0x0, 0x8, 0x1, 0x0 },
117    { OP_AND,    0x0, 0x0, 0x3, 0x0, 0x2, 0x2 | 0x8 },
118    { OP_OR,     0x0, 0x0, 0x3, 0x0, 0x2, 0x2 | 0x8 },
119    { OP_XOR,    0x0, 0x0, 0x3, 0x0, 0x2, 0x2 | 0x8 },
120    { OP_SHL,    0x0, 0x0, 0x0, 0x0, 0x2, 0x2 },
121    { OP_SHR,    0x0, 0x0, 0x0, 0x0, 0x2, 0x2 },
122    { OP_SET,    0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
123    { OP_SLCT,   0x4, 0x0, 0x0, 0x0, 0x6, 0x2 }, // special c[] constraint
124    { OP_PREEX2, 0x1, 0x1, 0x0, 0x0, 0x1, 0x1 },
125    { OP_PRESIN, 0x1, 0x1, 0x0, 0x0, 0x1, 0x1 },
126    { OP_COS,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
127    { OP_SIN,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
128    { OP_EX2,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
129    { OP_LG2,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
130    { OP_RCP,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
131    { OP_RSQ,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
132    { OP_DFDX,   0x1, 0x0, 0x0, 0x0, 0x0, 0x0 },
133    { OP_DFDY,   0x1, 0x0, 0x0, 0x0, 0x0, 0x0 },
134    { OP_CALL,   0x0, 0x0, 0x0, 0x0, 0x1, 0x0 },
135    { OP_POPCNT, 0x0, 0x0, 0x3, 0x0, 0x2, 0x2 },
136    { OP_INSBF,  0x0, 0x0, 0x0, 0x0, 0x6, 0x2 },
137    { OP_EXTBF,  0x0, 0x0, 0x0, 0x0, 0x2, 0x2 },
138    { OP_BFIND,  0x0, 0x0, 0x1, 0x0, 0x1, 0x1 },
139    { OP_PERMT,  0x0, 0x0, 0x0, 0x0, 0x6, 0x2 },
140    { OP_SET_AND, 0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
141    { OP_SET_OR, 0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
142    { OP_SET_XOR, 0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
143    // saturate only:
144    { OP_LINTERP, 0x0, 0x0, 0x0, 0x8, 0x0, 0x0 },
145    { OP_PINTERP, 0x0, 0x0, 0x0, 0x8, 0x0, 0x0 },
146    // nve4 ops:
147    { OP_SULDB,   0x0, 0x0, 0x0, 0x0, 0x2, 0x0 },
148    { OP_SUSTB,   0x0, 0x0, 0x0, 0x0, 0x2, 0x0 },
149    { OP_SUSTP,   0x0, 0x0, 0x0, 0x0, 0x2, 0x0 },
150    { OP_SUCLAMP, 0x0, 0x0, 0x0, 0x0, 0x2, 0x2 },
151    { OP_SUBFM,   0x0, 0x0, 0x0, 0x0, 0x6, 0x2 },
152    { OP_SUEAU,   0x0, 0x0, 0x0, 0x0, 0x6, 0x2 }
153 };
154 
initOpInfo()155 void TargetNVC0::initOpInfo()
156 {
157    unsigned int i, j;
158 
159    static const uint32_t commutative[(OP_LAST + 31) / 32] =
160    {
161       // ADD, MUL, MAD, FMA, AND, OR, XOR, MAX, MIN, SET_AND, SET_OR, SET_XOR,
162       // SET, SELP, SLCT
163       0x0ce0ca00, 0x0000007e, 0x00000000, 0x00000000
164    };
165 
166    static const uint32_t shortForm[(OP_LAST + 31) / 32] =
167    {
168       // ADD, MUL, MAD, FMA, AND, OR, XOR, MAX, MIN
169       0x0ce0ca00, 0x00000000, 0x00000000, 0x00000000
170    };
171 
172    static const operation noDest[] =
173    {
174       OP_STORE, OP_WRSV, OP_EXPORT, OP_BRA, OP_CALL, OP_RET, OP_EXIT,
175       OP_DISCARD, OP_CONT, OP_BREAK, OP_PRECONT, OP_PREBREAK, OP_PRERET,
176       OP_JOIN, OP_JOINAT, OP_BRKPT, OP_MEMBAR, OP_EMIT, OP_RESTART,
177       OP_QUADON, OP_QUADPOP, OP_TEXBAR, OP_SUSTB, OP_SUSTP, OP_SUREDP,
178       OP_SUREDB, OP_BAR
179    };
180 
181    static const operation noPred[] =
182    {
183       OP_CALL, OP_PRERET, OP_QUADON, OP_QUADPOP,
184       OP_JOINAT, OP_PREBREAK, OP_PRECONT, OP_BRKPT
185    };
186 
187    for (i = 0; i < DATA_FILE_COUNT; ++i)
188       nativeFileMap[i] = (DataFile)i;
189    nativeFileMap[FILE_ADDRESS] = FILE_GPR;
190 
191    for (i = 0; i < OP_LAST; ++i) {
192       opInfo[i].variants = NULL;
193       opInfo[i].op = (operation)i;
194       opInfo[i].srcTypes = 1 << (int)TYPE_F32;
195       opInfo[i].dstTypes = 1 << (int)TYPE_F32;
196       opInfo[i].immdBits = 0;
197       opInfo[i].srcNr = operationSrcNr[i];
198 
199       for (j = 0; j < opInfo[i].srcNr; ++j) {
200          opInfo[i].srcMods[j] = 0;
201          opInfo[i].srcFiles[j] = 1 << (int)FILE_GPR;
202       }
203       opInfo[i].dstMods = 0;
204       opInfo[i].dstFiles = 1 << (int)FILE_GPR;
205 
206       opInfo[i].hasDest = 1;
207       opInfo[i].vector = (i >= OP_TEX && i <= OP_TEXCSAA);
208       opInfo[i].commutative = (commutative[i / 32] >> (i % 32)) & 1;
209       opInfo[i].pseudo = (i < OP_MOV);
210       opInfo[i].predicate = !opInfo[i].pseudo;
211       opInfo[i].flow = (i >= OP_BRA && i <= OP_JOIN);
212       opInfo[i].minEncSize = (shortForm[i / 32] & (1 << (i % 32))) ? 4 : 8;
213    }
214    for (i = 0; i < sizeof(noDest) / sizeof(noDest[0]); ++i)
215       opInfo[noDest[i]].hasDest = 0;
216    for (i = 0; i < sizeof(noPred) / sizeof(noPred[0]); ++i)
217       opInfo[noPred[i]].predicate = 0;
218 
219    for (i = 0; i < sizeof(_initProps) / sizeof(_initProps[0]); ++i) {
220       const struct opProperties *prop = &_initProps[i];
221 
222       for (int s = 0; s < 3; ++s) {
223          if (prop->mNeg & (1 << s))
224             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NEG;
225          if (prop->mAbs & (1 << s))
226             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_ABS;
227          if (prop->mNot & (1 << s))
228             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NOT;
229          if (prop->fConst & (1 << s))
230             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_MEMORY_CONST;
231          if (prop->fImmd & (1 << s))
232             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_IMMEDIATE;
233          if (prop->fImmd & 8)
234             opInfo[prop->op].immdBits = 0xffffffff;
235       }
236       if (prop->mSat & 8)
237          opInfo[prop->op].dstMods = NV50_IR_MOD_SAT;
238    }
239 }
240 
241 unsigned int
getFileSize(DataFile file) const242 TargetNVC0::getFileSize(DataFile file) const
243 {
244    const unsigned int gprs = (chipset >= NVISA_GK20A_CHIPSET) ? 255 : 63;
245    const unsigned int smregs = (chipset >= NVISA_GK104_CHIPSET) ? 65536 : 32768;
246    switch (file) {
247    case FILE_NULL:          return 0;
248    case FILE_GPR:           return MIN2(gprs, smregs / threads);
249    case FILE_PREDICATE:     return 7;
250    case FILE_FLAGS:         return 1;
251    case FILE_ADDRESS:       return 0;
252    case FILE_IMMEDIATE:     return 0;
253    case FILE_MEMORY_CONST:  return 65536;
254    case FILE_SHADER_INPUT:  return 0x400;
255    case FILE_SHADER_OUTPUT: return 0x400;
256    case FILE_MEMORY_BUFFER: return 0xffffffff;
257    case FILE_MEMORY_GLOBAL: return 0xffffffff;
258    case FILE_MEMORY_SHARED: return 16 << 10;
259    case FILE_MEMORY_LOCAL:  return 48 << 10;
260    case FILE_SYSTEM_VALUE:  return 32;
261    default:
262       assert(!"invalid file");
263       return 0;
264    }
265 }
266 
267 unsigned int
getFileUnit(DataFile file) const268 TargetNVC0::getFileUnit(DataFile file) const
269 {
270    if (file == FILE_GPR || file == FILE_ADDRESS || file == FILE_SYSTEM_VALUE)
271       return 2;
272    return 0;
273 }
274 
275 uint32_t
getSVAddress(DataFile shaderFile,const Symbol * sym) const276 TargetNVC0::getSVAddress(DataFile shaderFile, const Symbol *sym) const
277 {
278    const int idx = sym->reg.data.sv.index;
279    const SVSemantic sv = sym->reg.data.sv.sv;
280 
281    const bool isInput = shaderFile == FILE_SHADER_INPUT;
282    const bool kepler = getChipset() >= NVISA_GK104_CHIPSET;
283 
284    switch (sv) {
285    case SV_POSITION:       return 0x070 + idx * 4;
286    case SV_INSTANCE_ID:    return 0x2f8;
287    case SV_VERTEX_ID:      return 0x2fc;
288    case SV_PRIMITIVE_ID:   return isInput ? 0x060 : 0x040;
289    case SV_LAYER:          return 0x064;
290    case SV_VIEWPORT_INDEX: return 0x068;
291    case SV_POINT_SIZE:     return 0x06c;
292    case SV_CLIP_DISTANCE:  return 0x2c0 + idx * 4;
293    case SV_POINT_COORD:    return 0x2e0 + idx * 4;
294    case SV_FACE:           return 0x3fc;
295    case SV_TESS_OUTER:     return 0x000 + idx * 4;
296    case SV_TESS_INNER:     return 0x010 + idx * 4;
297    case SV_TESS_COORD:     return 0x2f0 + idx * 4;
298    case SV_NTID:           return kepler ? (0x00 + idx * 4) : ~0;
299    case SV_NCTAID:         return kepler ? (0x0c + idx * 4) : ~0;
300    case SV_GRIDID:         return kepler ? 0x18 : ~0;
301    case SV_WORK_DIM:       return 0x1c;
302    case SV_SAMPLE_INDEX:   return 0;
303    case SV_SAMPLE_POS:     return 0;
304    case SV_SAMPLE_MASK:    return 0;
305    case SV_BASEVERTEX:     return 0;
306    case SV_BASEINSTANCE:   return 0;
307    case SV_DRAWID:         return 0;
308    default:
309       return 0xffffffff;
310    }
311 }
312 
313 bool
insnCanLoad(const Instruction * i,int s,const Instruction * ld) const314 TargetNVC0::insnCanLoad(const Instruction *i, int s,
315                         const Instruction *ld) const
316 {
317    DataFile sf = ld->src(0).getFile();
318 
319    // immediate 0 can be represented by GPR $r63/$r255
320    if (sf == FILE_IMMEDIATE && ld->getSrc(0)->reg.data.u64 == 0)
321       return (!i->isPseudo() &&
322               !i->asTex() &&
323               i->op != OP_EXPORT && i->op != OP_STORE);
324 
325    if (s >= opInfo[i->op].srcNr)
326       return false;
327    if (!(opInfo[i->op].srcFiles[s] & (1 << (int)sf)))
328       return false;
329 
330    // indirect loads can only be done by OP_LOAD/VFETCH/INTERP on nvc0
331    if (ld->src(0).isIndirect(0))
332       return false;
333    // these are implemented using shf.r and shf.l which can't load consts
334    if ((i->op == OP_SHL || i->op == OP_SHR) && typeSizeof(i->sType) == 8 &&
335        sf == FILE_MEMORY_CONST)
336       return false;
337 
338    for (int k = 0; i->srcExists(k); ++k) {
339       if (i->src(k).getFile() == FILE_IMMEDIATE) {
340          if (k == 2 && i->op == OP_SUCLAMP) // special case
341             continue;
342          if (k == 1 && i->op == OP_SHLADD) // special case
343             continue;
344          if (i->getSrc(k)->reg.data.u64 != 0)
345             return false;
346       } else
347       if (i->src(k).getFile() != FILE_GPR &&
348           i->src(k).getFile() != FILE_PREDICATE &&
349           i->src(k).getFile() != FILE_FLAGS) {
350          return false;
351       }
352    }
353 
354    // not all instructions support full 32 bit immediates
355    if (sf == FILE_IMMEDIATE) {
356       Storage &reg = ld->getSrc(0)->asImm()->reg;
357 
358       if (opInfo[i->op].immdBits != 0xffffffff || typeSizeof(i->sType) > 4) {
359          switch (i->sType) {
360          case TYPE_F64:
361             if (reg.data.u64 & 0x00000fffffffffffULL)
362                return false;
363             break;
364          case TYPE_F32:
365             if (reg.data.u32 & 0xfff)
366                return false;
367             break;
368          case TYPE_S32:
369          case TYPE_U32:
370             // with u32, 0xfffff counts as 0xffffffff as well
371             if (reg.data.s32 > 0x7ffff || reg.data.s32 < -0x80000)
372                return false;
373             break;
374          case TYPE_U8:
375          case TYPE_S8:
376          case TYPE_U16:
377          case TYPE_S16:
378          case TYPE_F16:
379             break;
380          default:
381             return false;
382          }
383       } else
384       if (i->op == OP_ADD && i->sType == TYPE_F32) {
385          // add f32 LIMM cannot saturate
386          if (i->saturate && (reg.data.u32 & 0xfff))
387             return false;
388       }
389    }
390 
391    return true;
392 }
393 
394 bool
insnCanLoadOffset(const Instruction * insn,int s,int offset) const395 TargetNVC0::insnCanLoadOffset(const Instruction *insn, int s, int offset) const
396 {
397    const ValueRef& ref = insn->src(s);
398    if (ref.getFile() == FILE_MEMORY_CONST &&
399        (insn->op != OP_LOAD || insn->subOp != NV50_IR_SUBOP_LDC_IS))
400       return offset >= -0x8000 && offset < 0x8000;
401    return true;
402 }
403 
404 bool
isAccessSupported(DataFile file,DataType ty) const405 TargetNVC0::isAccessSupported(DataFile file, DataType ty) const
406 {
407    if (ty == TYPE_NONE)
408       return false;
409    if (file == FILE_MEMORY_CONST) {
410       if (getChipset() >= NVISA_GM107_CHIPSET)
411          return typeSizeof(ty) <= 4;
412       else
413       if (getChipset() >= NVISA_GK104_CHIPSET) // wrong encoding ?
414          return typeSizeof(ty) <= 8;
415    }
416    if (ty == TYPE_B96)
417       return false;
418    return true;
419 }
420 
421 bool
isOpSupported(operation op,DataType ty) const422 TargetNVC0::isOpSupported(operation op, DataType ty) const
423 {
424    if (op == OP_SAD && ty != TYPE_S32 && ty != TYPE_U32)
425       return false;
426    if (op == OP_POW || op == OP_SQRT || op == OP_DIV || op == OP_MOD)
427       return false;
428    return true;
429 }
430 
431 bool
isModSupported(const Instruction * insn,int s,Modifier mod) const432 TargetNVC0::isModSupported(const Instruction *insn, int s, Modifier mod) const
433 {
434    if (!isFloatType(insn->dType)) {
435       switch (insn->op) {
436       case OP_ABS:
437       case OP_NEG:
438       case OP_CVT:
439       case OP_CEIL:
440       case OP_FLOOR:
441       case OP_TRUNC:
442       case OP_AND:
443       case OP_OR:
444       case OP_XOR:
445       case OP_POPCNT:
446       case OP_BFIND:
447          break;
448       case OP_SET:
449          if (insn->sType != TYPE_F32)
450             return false;
451          break;
452       case OP_ADD:
453          if (mod.abs())
454             return false;
455          if (insn->src(s ? 0 : 1).mod.neg())
456             return false;
457          break;
458       case OP_SUB:
459          if (s == 0)
460             return insn->src(1).mod.neg() ? false : true;
461          break;
462       case OP_SHLADD:
463          if (s == 1)
464             return false;
465          if (insn->src(s ? 0 : 2).mod.neg())
466             return false;
467          break;
468       default:
469          return false;
470       }
471    }
472    if (s >= opInfo[insn->op].srcNr || s >= 3)
473       return false;
474    return (mod & Modifier(opInfo[insn->op].srcMods[s])) == mod;
475 }
476 
477 bool
mayPredicate(const Instruction * insn,const Value * pred) const478 TargetNVC0::mayPredicate(const Instruction *insn, const Value *pred) const
479 {
480    if (insn->getPredicate())
481       return false;
482    return opInfo[insn->op].predicate;
483 }
484 
485 bool
isSatSupported(const Instruction * insn) const486 TargetNVC0::isSatSupported(const Instruction *insn) const
487 {
488    if (insn->op == OP_CVT)
489       return true;
490    if (!(opInfo[insn->op].dstMods & NV50_IR_MOD_SAT))
491       return false;
492 
493    if (insn->dType == TYPE_U32)
494       return (insn->op == OP_ADD) || (insn->op == OP_MAD);
495 
496    // add f32 LIMM cannot saturate
497    if (insn->op == OP_ADD && insn->sType == TYPE_F32) {
498       if (insn->getSrc(1)->asImm() &&
499           insn->getSrc(1)->reg.data.u32 & 0xfff)
500          return false;
501    }
502 
503    return insn->dType == TYPE_F32;
504 }
505 
506 bool
isPostMultiplySupported(operation op,float f,int & e) const507 TargetNVC0::isPostMultiplySupported(operation op, float f, int& e) const
508 {
509    if (op != OP_MUL)
510       return false;
511    f = fabsf(f);
512    e = static_cast<int>(log2f(f));
513    if (e < -3 || e > 3)
514       return false;
515    return f == exp2f(static_cast<float>(e));
516 }
517 
518 // TODO: better values
519 // this could be more precise, e.g. depending on the issue-to-read/write delay
520 // of the depending instruction, but it's good enough
getLatency(const Instruction * i) const521 int TargetNVC0::getLatency(const Instruction *i) const
522 {
523    if (chipset >= 0xe4) {
524       if (i->dType == TYPE_F64 || i->sType == TYPE_F64)
525          return 20;
526       switch (i->op) {
527       case OP_LINTERP:
528       case OP_PINTERP:
529          return 15;
530       case OP_LOAD:
531          if (i->src(0).getFile() == FILE_MEMORY_CONST)
532             return 9;
533          // fall through
534       case OP_VFETCH:
535          return 24;
536       default:
537          if (Target::getOpClass(i->op) == OPCLASS_TEXTURE)
538             return 17;
539          if (i->op == OP_MUL && i->dType != TYPE_F32)
540             return 15;
541          return 9;
542       }
543    } else {
544       if (i->op == OP_LOAD) {
545          if (i->cache == CACHE_CV)
546             return 700;
547          return 48;
548       }
549       return 24;
550    }
551    return 32;
552 }
553 
554 // These are "inverse" throughput values, i.e. the number of cycles required
555 // to issue a specific instruction for a full warp (32 threads).
556 //
557 // Assuming we have more than 1 warp in flight, a higher issue latency results
558 // in a lower result latency since the MP will have spent more time with other
559 // warps.
560 // This also helps to determine the number of cycles between instructions in
561 // a single warp.
562 //
getThroughput(const Instruction * i) const563 int TargetNVC0::getThroughput(const Instruction *i) const
564 {
565    // TODO: better values
566    if (i->dType == TYPE_F32) {
567       switch (i->op) {
568       case OP_ADD:
569       case OP_MUL:
570       case OP_MAD:
571       case OP_FMA:
572          return 1;
573       case OP_CVT:
574       case OP_CEIL:
575       case OP_FLOOR:
576       case OP_TRUNC:
577       case OP_SET:
578       case OP_SLCT:
579       case OP_MIN:
580       case OP_MAX:
581          return 2;
582       case OP_RCP:
583       case OP_RSQ:
584       case OP_LG2:
585       case OP_SIN:
586       case OP_COS:
587       case OP_PRESIN:
588       case OP_PREEX2:
589       default:
590          return 8;
591       }
592    } else
593    if (i->dType == TYPE_U32 || i->dType == TYPE_S32) {
594       switch (i->op) {
595       case OP_ADD:
596       case OP_AND:
597       case OP_OR:
598       case OP_XOR:
599       case OP_NOT:
600          return 1;
601       case OP_MUL:
602       case OP_MAD:
603       case OP_CVT:
604       case OP_SET:
605       case OP_SLCT:
606       case OP_SHL:
607       case OP_SHR:
608       case OP_NEG:
609       case OP_ABS:
610       case OP_MIN:
611       case OP_MAX:
612       default:
613          return 2;
614       }
615    } else
616    if (i->dType == TYPE_F64) {
617       return 2;
618    } else {
619       return 1;
620    }
621 }
622 
canDualIssue(const Instruction * a,const Instruction * b) const623 bool TargetNVC0::canDualIssue(const Instruction *a, const Instruction *b) const
624 {
625    const OpClass clA = operationClass[a->op];
626    const OpClass clB = operationClass[b->op];
627 
628    if (getChipset() >= 0xe4) {
629       // not texturing
630       // not if the 2nd instruction isn't necessarily executed
631       if (clA == OPCLASS_TEXTURE || clA == OPCLASS_FLOW)
632          return false;
633 
634       // Check that a and b don't write to the same sources, nor that b reads
635       // anything that a writes.
636       if (!a->canCommuteDefDef(b) || !a->canCommuteDefSrc(b))
637          return false;
638 
639       // anything with MOV
640       if (a->op == OP_MOV || b->op == OP_MOV)
641          return true;
642       if (clA == clB) {
643          switch (clA) {
644          // there might be more
645          case OPCLASS_COMPARE:
646             if ((a->op == OP_MIN || a->op == OP_MAX) &&
647                 (b->op == OP_MIN || b->op == OP_MAX))
648                break;
649             return false;
650          case OPCLASS_ARITH:
651             break;
652          default:
653             return false;
654          }
655          // only F32 arith or integer additions
656          return (a->dType == TYPE_F32 || a->op == OP_ADD ||
657                  b->dType == TYPE_F32 || b->op == OP_ADD);
658       }
659       // nothing with TEXBAR
660       if (a->op == OP_TEXBAR || b->op == OP_TEXBAR)
661          return false;
662       // no loads and stores accessing the same space
663       if ((clA == OPCLASS_LOAD && clB == OPCLASS_STORE) ||
664           (clB == OPCLASS_LOAD && clA == OPCLASS_STORE))
665          if (a->src(0).getFile() == b->src(0).getFile())
666             return false;
667       // no > 32-bit ops
668       if (typeSizeof(a->dType) > 4 || typeSizeof(b->dType) > 4 ||
669           typeSizeof(a->sType) > 4 || typeSizeof(b->sType) > 4)
670          return false;
671       return true;
672    } else {
673       return false; // info not needed (yet)
674    }
675 }
676 
677 } // namespace nv50_ir
678