2 * Copyright 2011 Christoph Bumiller
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
23 #include "codegen/nv50_ir.h"
24 #include "codegen/nv50_ir_build_util.h"
26 #include "codegen/nv50_ir_target_nvc0.h"
27 #include "codegen/nv50_ir_lowering_nvc0.h"
39 #define QUADOP(q, r, s, t) \
40 ((QOP_##q << 6) | (QOP_##r << 4) | \
41 (QOP_##s << 2) | (QOP_##t << 0))
44 NVC0LegalizeSSA::handleDIV(Instruction *i)
46 FlowInstruction *call;
50 bld.setPosition(i, false);
51 def[0] = bld.mkMovToReg(0, i->getSrc(0))->getDef(0);
52 def[1] = bld.mkMovToReg(1, i->getSrc(1))->getDef(0);
54 case TYPE_U32: builtin = NVC0_BUILTIN_DIV_U32; break;
55 case TYPE_S32: builtin = NVC0_BUILTIN_DIV_S32; break;
59 call = bld.mkFlow(OP_CALL, NULL, CC_ALWAYS, NULL);
60 bld.mkMov(i->getDef(0), def[(i->op == OP_DIV) ? 0 : 1]);
61 bld.mkClobber(FILE_GPR, (i->op == OP_DIV) ? 0xe : 0xd, 2);
62 bld.mkClobber(FILE_PREDICATE, (i->dType == TYPE_S32) ? 0xf : 0x3, 0);
65 call->absolute = call->builtin = 1;
66 call->target.builtin = builtin;
67 delete_Instruction(prog, i);
71 NVC0LegalizeSSA::handleRCPRSQ(Instruction *i)
73 assert(i->dType == TYPE_F64);
74 // There are instructions that will compute the high 32 bits of the 64-bit
75 // float. We will just stick 0 in the bottom 32 bits.
77 bld.setPosition(i, false);
79 // 1. Take the source and it up.
80 Value *src[2], *dst[2], *def = i->getDef(0);
81 bld.mkSplit(src, 4, i->getSrc(0));
83 // 2. We don't care about the low 32 bits of the destination. Stick a 0 in.
84 dst[0] = bld.loadImm(NULL, 0);
85 dst[1] = bld.getSSA();
87 // 3. The new version of the instruction takes the high 32 bits of the
88 // source and outputs the high 32 bits of the destination.
92 i->subOp = NV50_IR_SUBOP_RCPRSQ_64H;
94 // 4. Recombine the two dst pieces back into the original destination.
95 bld.setPosition(i, true);
96 bld.mkOp2(OP_MERGE, TYPE_U64, def, dst[0], dst[1]);
100 NVC0LegalizeSSA::handleFTZ(Instruction *i)
102 // Only want to flush float inputs
103 assert(i->sType == TYPE_F32);
105 // If we're already flushing denorms (and NaN's) to zero, no need for this.
109 // Only certain classes of operations can flush
110 OpClass cls = prog->getTarget()->getOpClass(i->op);
111 if (cls != OPCLASS_ARITH && cls != OPCLASS_COMPARE &&
112 cls != OPCLASS_CONVERT)
119 NVC0LegalizeSSA::visit(Function *fn)
121 bld.setProgram(fn->getProgram());
126 NVC0LegalizeSSA::visit(BasicBlock *bb)
129 for (Instruction *i = bb->getEntry(); i; i = next) {
131 if (i->sType == TYPE_F32) {
132 if (prog->getType() != Program::TYPE_COMPUTE)
143 if (i->dType == TYPE_F64)
153 NVC0LegalizePostRA::NVC0LegalizePostRA(const Program *prog)
157 needTexBar(prog->getTarget()->getChipset() >= 0xe0)
162 NVC0LegalizePostRA::insnDominatedBy(const Instruction *later,
163 const Instruction *early) const
165 if (early->bb == later->bb)
166 return early->serial < later->serial;
167 return later->bb->dominatedBy(early->bb);
171 NVC0LegalizePostRA::addTexUse(std::list<TexUse> &uses,
172 Instruction *usei, const Instruction *texi)
175 for (std::list<TexUse>::iterator it = uses.begin();
177 if (insnDominatedBy(usei, it->insn)) {
181 if (insnDominatedBy(it->insn, usei))
187 uses.push_back(TexUse(usei, texi));
190 // While it might be tempting to use the an algorithm that just looks at tex
191 // uses, not all texture results are guaranteed to be used on all paths. In
192 // the case where along some control flow path a texture result is never used,
193 // we might reuse that register for something else, creating a
194 // write-after-write hazard. So we have to manually look through all
195 // instructions looking for ones that reference the registers in question.
197 NVC0LegalizePostRA::findFirstUses(
198 Instruction *texi, std::list<TexUse> &uses)
200 int minGPR = texi->def(0).rep()->reg.data.id;
201 int maxGPR = minGPR + texi->def(0).rep()->reg.size / 4 - 1;
203 unordered_set<const BasicBlock *> visited;
204 findFirstUsesBB(minGPR, maxGPR, texi->next, texi, uses, visited);
208 NVC0LegalizePostRA::findFirstUsesBB(
209 int minGPR, int maxGPR, Instruction *start,
210 const Instruction *texi, std::list<TexUse> &uses,
211 unordered_set<const BasicBlock *> &visited)
213 const BasicBlock *bb = start->bb;
215 // We don't process the whole bb the first time around. This is correct,
216 // however we might be in a loop and hit this BB again, and need to process
217 // the full thing. So only mark a bb as visited if we processed it from the
219 if (start == bb->getEntry()) {
220 if (visited.find(bb) != visited.end())
225 for (Instruction *insn = start; insn != bb->getExit(); insn = insn->next) {
229 for (int d = 0; insn->defExists(d); ++d) {
230 const Value *def = insn->def(d).rep();
231 if (insn->def(d).getFile() != FILE_GPR ||
232 def->reg.data.id + def->reg.size / 4 - 1 < minGPR ||
233 def->reg.data.id > maxGPR)
235 addTexUse(uses, insn, texi);
239 for (int s = 0; insn->srcExists(s); ++s) {
240 const Value *src = insn->src(s).rep();
241 if (insn->src(s).getFile() != FILE_GPR ||
242 src->reg.data.id + src->reg.size / 4 - 1 < minGPR ||
243 src->reg.data.id > maxGPR)
245 addTexUse(uses, insn, texi);
250 for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
251 findFirstUsesBB(minGPR, maxGPR, BasicBlock::get(ei.getNode())->getEntry(),
252 texi, uses, visited);
257 // This pass is a bit long and ugly and can probably be optimized.
259 // 1. obtain a list of TEXes and their outputs' first use(s)
260 // 2. calculate the barrier level of each first use (minimal number of TEXes,
261 // over all paths, between the TEX and the use in question)
262 // 3. for each barrier, if all paths from the source TEX to that barrier
263 // contain a barrier of lesser level, it can be culled
265 NVC0LegalizePostRA::insertTextureBarriers(Function *fn)
267 std::list<TexUse> *uses;
268 std::vector<Instruction *> texes;
269 std::vector<int> bbFirstTex;
270 std::vector<int> bbFirstUse;
271 std::vector<int> texCounts;
272 std::vector<TexUse> useVec;
275 fn->orderInstructions(insns);
277 texCounts.resize(fn->allBBlocks.getSize(), 0);
278 bbFirstTex.resize(fn->allBBlocks.getSize(), insns.getSize());
279 bbFirstUse.resize(fn->allBBlocks.getSize(), insns.getSize());
281 // tag BB CFG nodes by their id for later
282 for (ArrayList::Iterator i = fn->allBBlocks.iterator(); !i.end(); i.next()) {
283 BasicBlock *bb = reinterpret_cast<BasicBlock *>(i.get());
285 bb->cfg.tag = bb->getId();
288 // gather the first uses for each TEX
289 for (int i = 0; i < insns.getSize(); ++i) {
290 Instruction *tex = reinterpret_cast<Instruction *>(insns.get(i));
291 if (isTextureOp(tex->op)) {
292 texes.push_back(tex);
293 if (!texCounts.at(tex->bb->getId()))
294 bbFirstTex[tex->bb->getId()] = texes.size() - 1;
295 texCounts[tex->bb->getId()]++;
301 uses = new std::list<TexUse>[texes.size()];
304 for (size_t i = 0; i < texes.size(); ++i) {
305 findFirstUses(texes[i], uses[i]);
308 // determine the barrier level at each use
309 for (size_t i = 0; i < texes.size(); ++i) {
310 for (std::list<TexUse>::iterator u = uses[i].begin(); u != uses[i].end();
312 BasicBlock *tb = texes[i]->bb;
313 BasicBlock *ub = u->insn->bb;
316 for (size_t j = i + 1; j < texes.size() &&
317 texes[j]->bb == tb && texes[j]->serial < u->insn->serial;
321 u->level = fn->cfg.findLightestPathWeight(&tb->cfg,
322 &ub->cfg, texCounts);
324 WARN("Failed to find path TEX -> TEXBAR\n");
328 // this counted all TEXes in the origin block, correct that
329 u->level -= i - bbFirstTex.at(tb->getId()) + 1 /* this TEX */;
330 // and did not count the TEXes in the destination block, add those
331 for (size_t j = bbFirstTex.at(ub->getId()); j < texes.size() &&
332 texes[j]->bb == ub && texes[j]->serial < u->insn->serial;
336 assert(u->level >= 0);
337 useVec.push_back(*u);
342 // insert the barriers
343 for (size_t i = 0; i < useVec.size(); ++i) {
344 Instruction *prev = useVec[i].insn->prev;
345 if (useVec[i].level < 0)
347 if (prev && prev->op == OP_TEXBAR) {
348 if (prev->subOp > useVec[i].level)
349 prev->subOp = useVec[i].level;
350 prev->setSrc(prev->srcCount(), useVec[i].tex->getDef(0));
352 Instruction *bar = new_Instruction(func, OP_TEXBAR, TYPE_NONE);
354 bar->subOp = useVec[i].level;
355 // make use explicit to ease latency calculation
356 bar->setSrc(bar->srcCount(), useVec[i].tex->getDef(0));
357 useVec[i].insn->bb->insertBefore(useVec[i].insn, bar);
361 if (fn->getProgram()->optLevel < 3)
364 std::vector<Limits> limitT, limitB, limitS; // entry, exit, single
366 limitT.resize(fn->allBBlocks.getSize(), Limits(0, 0));
367 limitB.resize(fn->allBBlocks.getSize(), Limits(0, 0));
368 limitS.resize(fn->allBBlocks.getSize());
370 // cull unneeded barriers (should do that earlier, but for simplicity)
371 IteratorRef bi = fn->cfg.iteratorCFG();
372 // first calculate min/max outstanding TEXes for each BB
373 for (bi->reset(); !bi->end(); bi->next()) {
374 Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
375 BasicBlock *bb = BasicBlock::get(n);
377 int max = std::numeric_limits<int>::max();
378 for (Instruction *i = bb->getFirst(); i; i = i->next) {
379 if (isTextureOp(i->op)) {
381 if (max < std::numeric_limits<int>::max())
384 if (i->op == OP_TEXBAR) {
385 min = MIN2(min, i->subOp);
386 max = MIN2(max, i->subOp);
389 // limits when looking at an isolated block
390 limitS[bb->getId()].min = min;
391 limitS[bb->getId()].max = max;
393 // propagate the min/max values
394 for (unsigned int l = 0; l <= fn->loopNestingBound; ++l) {
395 for (bi->reset(); !bi->end(); bi->next()) {
396 Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
397 BasicBlock *bb = BasicBlock::get(n);
398 const int bbId = bb->getId();
399 for (Graph::EdgeIterator ei = n->incident(); !ei.end(); ei.next()) {
400 BasicBlock *in = BasicBlock::get(ei.getNode());
401 const int inId = in->getId();
402 limitT[bbId].min = MAX2(limitT[bbId].min, limitB[inId].min);
403 limitT[bbId].max = MAX2(limitT[bbId].max, limitB[inId].max);
405 // I just hope this is correct ...
406 if (limitS[bbId].max == std::numeric_limits<int>::max()) {
408 limitB[bbId].min = limitT[bbId].min + limitS[bbId].min;
409 limitB[bbId].max = limitT[bbId].max + limitS[bbId].min;
411 // block contained a barrier
412 limitB[bbId].min = MIN2(limitS[bbId].max,
413 limitT[bbId].min + limitS[bbId].min);
414 limitB[bbId].max = MIN2(limitS[bbId].max,
415 limitT[bbId].max + limitS[bbId].min);
419 // finally delete unnecessary barriers
420 for (bi->reset(); !bi->end(); bi->next()) {
421 Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
422 BasicBlock *bb = BasicBlock::get(n);
423 Instruction *prev = NULL;
425 int max = limitT[bb->getId()].max;
426 for (Instruction *i = bb->getFirst(); i; i = next) {
428 if (i->op == OP_TEXBAR) {
429 if (i->subOp >= max) {
430 delete_Instruction(prog, i);
434 if (prev && prev->op == OP_TEXBAR && prev->subOp >= max) {
435 delete_Instruction(prog, prev);
440 if (isTextureOp(i->op)) {
443 if (i && !i->isNop())
451 NVC0LegalizePostRA::visit(Function *fn)
454 insertTextureBarriers(fn);
456 rZero = new_LValue(fn, FILE_GPR);
457 pOne = new_LValue(fn, FILE_PREDICATE);
458 carry = new_LValue(fn, FILE_FLAGS);
460 rZero->reg.data.id = prog->getTarget()->getFileSize(FILE_GPR);
461 carry->reg.data.id = 0;
462 pOne->reg.data.id = 7;
468 NVC0LegalizePostRA::replaceZero(Instruction *i)
470 for (int s = 0; i->srcExists(s); ++s) {
471 if (s == 2 && i->op == OP_SUCLAMP)
473 ImmediateValue *imm = i->getSrc(s)->asImm();
475 if (i->op == OP_SELP && s == 2) {
477 if (imm->reg.data.u64 == 0)
478 i->src(s).mod = i->src(s).mod ^ Modifier(NV50_IR_MOD_NOT);
479 } else if (imm->reg.data.u64 == 0) {
486 // replace CONT with BRA for single unconditional continue
488 NVC0LegalizePostRA::tryReplaceContWithBra(BasicBlock *bb)
490 if (bb->cfg.incidentCount() != 2 || bb->getEntry()->op != OP_PRECONT)
492 Graph::EdgeIterator ei = bb->cfg.incident();
493 if (ei.getType() != Graph::Edge::BACK)
495 if (ei.getType() != Graph::Edge::BACK)
497 BasicBlock *contBB = BasicBlock::get(ei.getNode());
499 if (!contBB->getExit() || contBB->getExit()->op != OP_CONT ||
500 contBB->getExit()->getPredicate())
502 contBB->getExit()->op = OP_BRA;
503 bb->remove(bb->getEntry()); // delete PRECONT
506 assert(ei.end() || ei.getType() != Graph::Edge::BACK);
510 // replace branches to join blocks with join ops
512 NVC0LegalizePostRA::propagateJoin(BasicBlock *bb)
514 if (bb->getEntry()->op != OP_JOIN || bb->getEntry()->asFlow()->limit)
516 for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
517 BasicBlock *in = BasicBlock::get(ei.getNode());
518 Instruction *exit = in->getExit();
520 in->insertTail(new FlowInstruction(func, OP_JOIN, bb));
521 // there should always be a terminator instruction
522 WARN("inserted missing terminator in BB:%i\n", in->getId());
524 if (exit->op == OP_BRA) {
526 exit->asFlow()->limit = 1; // must-not-propagate marker
529 bb->remove(bb->getEntry());
533 NVC0LegalizePostRA::visit(BasicBlock *bb)
535 Instruction *i, *next;
537 // remove pseudo operations and non-fixed no-ops, split 64 bit operations
538 for (i = bb->getFirst(); i; i = next) {
540 if (i->op == OP_EMIT || i->op == OP_RESTART) {
541 if (!i->getDef(0)->refCount())
543 if (i->src(0).getFile() == FILE_IMMEDIATE)
544 i->setSrc(0, rZero); // initial value must be 0
550 if (i->op == OP_BAR && i->subOp == NV50_IR_SUBOP_BAR_SYNC &&
551 prog->getType() != Program::TYPE_COMPUTE) {
552 // It seems like barriers are never required for tessellation since
553 // the warp size is 32, and there are always at most 32 tcs threads.
556 if (i->op == OP_LOAD && i->subOp == NV50_IR_SUBOP_LDC_IS) {
557 int offset = i->src(0).get()->reg.data.offset;
558 if (abs(offset) > 0x10000)
559 i->src(0).get()->reg.fileIndex += offset >> 16;
560 i->src(0).get()->reg.data.offset = (int)(short)offset;
562 // TODO: Move this to before register allocation for operations that
563 // need the $c register !
564 if (typeSizeof(i->dType) == 8) {
566 hi = BuildUtil::split64BitOpPostRA(func, i, rZero, carry);
571 if (i->op != OP_MOV && i->op != OP_PFETCH)
578 if (!tryReplaceContWithBra(bb))
584 NVC0LoweringPass::NVC0LoweringPass(Program *prog) : targ(prog->getTarget())
586 bld.setProgram(prog);
591 NVC0LoweringPass::visit(Function *fn)
593 if (prog->getType() == Program::TYPE_GEOMETRY) {
594 assert(!strncmp(fn->getName(), "MAIN", 4));
595 // TODO: when we generate actual functions pass this value along somehow
596 bld.setPosition(BasicBlock::get(fn->cfg.getRoot()), false);
597 gpEmitAddress = bld.loadImm(NULL, 0)->asLValue();
599 bld.setPosition(BasicBlock::get(fn->cfgExit)->getExit(), false);
600 bld.mkMovToReg(0, gpEmitAddress);
607 NVC0LoweringPass::visit(BasicBlock *bb)
613 NVC0LoweringPass::loadTexHandle(Value *ptr, unsigned int slot)
615 uint8_t b = prog->driver->io.auxCBSlot;
616 uint32_t off = prog->driver->io.texBindBase + slot * 4;
618 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
621 // move array source to first slot, convert to u16, add indirections
623 NVC0LoweringPass::handleTEX(TexInstruction *i)
625 const int dim = i->tex.target.getDim() + i->tex.target.isCube();
626 const int arg = i->tex.target.getArgCount();
627 const int lyr = arg - (i->tex.target.isMS() ? 2 : 1);
628 const int chipset = prog->getTarget()->getChipset();
630 /* Only normalize in the non-explicit derivatives case. For explicit
631 * derivatives, this is handled in handleManualTXD.
633 if (i->tex.target.isCube() && i->dPdx[0].get() == NULL) {
636 for (c = 0; c < 3; ++c)
637 src[c] = bld.mkOp1v(OP_ABS, TYPE_F32, bld.getSSA(), i->getSrc(c));
638 val = bld.getScratch();
639 bld.mkOp2(OP_MAX, TYPE_F32, val, src[0], src[1]);
640 bld.mkOp2(OP_MAX, TYPE_F32, val, src[2], val);
641 bld.mkOp1(OP_RCP, TYPE_F32, val, val);
642 for (c = 0; c < 3; ++c) {
643 i->setSrc(c, bld.mkOp2v(OP_MUL, TYPE_F32, bld.getSSA(),
648 // Arguments to the TEX instruction are a little insane. Even though the
649 // encoding is identical between SM20 and SM30, the arguments mean
650 // different things between Fermi and Kepler+. A lot of arguments are
651 // optional based on flags passed to the instruction. This summarizes the
661 // - tg4: 8 bits each, either 2 (1 offset reg) or 8 (2 offset reg)
662 // - other: 4 bits each, single reg
666 // array (+ offsets for txd in upper 16 bits)
671 // offsets (same as fermi, except txd which takes it with array)
688 if (chipset >= NVISA_GK104_CHIPSET) {
689 if (i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
690 // XXX this ignores tsc, and assumes a 1:1 mapping
691 assert(i->tex.rIndirectSrc >= 0);
692 Value *hnd = loadTexHandle(
693 bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
694 i->getIndirectR(), bld.mkImm(2)),
698 i->setIndirectR(hnd);
699 i->setIndirectS(NULL);
700 } else if (i->tex.r == i->tex.s || i->op == OP_TXF) {
701 i->tex.r += prog->driver->io.texBindBase / 4;
702 i->tex.s = 0; // only a single cX[] value possible here
704 Value *hnd = bld.getScratch();
705 Value *rHnd = loadTexHandle(NULL, i->tex.r);
706 Value *sHnd = loadTexHandle(NULL, i->tex.s);
708 bld.mkOp3(OP_INSBF, TYPE_U32, hnd, rHnd, bld.mkImm(0x1400), sHnd);
710 i->tex.r = 0; // not used for indirect tex
712 i->setIndirectR(hnd);
714 if (i->tex.target.isArray()) {
715 LValue *layer = new_LValue(func, FILE_GPR);
716 Value *src = i->getSrc(lyr);
717 const int sat = (i->op == OP_TXF) ? 1 : 0;
718 DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32;
719 bld.mkCvt(OP_CVT, TYPE_U16, layer, sTy, src)->saturate = sat;
720 if (i->op != OP_TXD || chipset < NVISA_GM107_CHIPSET) {
721 for (int s = dim; s >= 1; --s)
722 i->setSrc(s, i->getSrc(s - 1));
725 i->setSrc(dim, layer);
728 // Move the indirect reference to the first place
729 if (i->tex.rIndirectSrc >= 0 && (
730 i->op == OP_TXD || chipset < NVISA_GM107_CHIPSET)) {
731 Value *hnd = i->getIndirectR();
733 i->setIndirectR(NULL);
734 i->moveSources(0, 1);
736 i->tex.rIndirectSrc = 0;
737 i->tex.sIndirectSrc = -1;
740 // (nvc0) generate and move the tsc/tic/array source to the front
741 if (i->tex.target.isArray() || i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
742 LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa
744 Value *ticRel = i->getIndirectR();
745 Value *tscRel = i->getIndirectS();
748 i->setSrc(i->tex.rIndirectSrc, NULL);
750 ticRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
751 ticRel, bld.mkImm(i->tex.r));
754 i->setSrc(i->tex.sIndirectSrc, NULL);
756 tscRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
757 tscRel, bld.mkImm(i->tex.s));
760 Value *arrayIndex = i->tex.target.isArray() ? i->getSrc(lyr) : NULL;
762 for (int s = dim; s >= 1; --s)
763 i->setSrc(s, i->getSrc(s - 1));
764 i->setSrc(0, arrayIndex);
766 i->moveSources(0, 1);
770 int sat = (i->op == OP_TXF) ? 1 : 0;
771 DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32;
772 bld.mkCvt(OP_CVT, TYPE_U16, src, sTy, arrayIndex)->saturate = sat;
778 bld.mkOp3(OP_INSBF, TYPE_U32, src, ticRel, bld.mkImm(0x0917), src);
780 bld.mkOp3(OP_INSBF, TYPE_U32, src, tscRel, bld.mkImm(0x0710), src);
785 // For nvc0, the sample id has to be in the second operand, as the offset
786 // does. Right now we don't know how to pass both in, and this case can't
787 // happen with OpenGL. On nve0, the sample id is part of the texture
788 // coordinate argument.
789 assert(chipset >= NVISA_GK104_CHIPSET ||
790 !i->tex.useOffsets || !i->tex.target.isMS());
792 // offset is between lod and dc
793 if (i->tex.useOffsets) {
795 int s = i->srcCount(0xff, true);
796 if (i->op != OP_TXD || chipset < NVISA_GK104_CHIPSET) {
797 if (i->tex.target.isShadow())
799 if (i->srcExists(s)) // move potential predicate out of the way
800 i->moveSources(s, 1);
801 if (i->tex.useOffsets == 4 && i->srcExists(s + 1))
802 i->moveSources(s + 1, 1);
804 if (i->op == OP_TXG) {
805 // Either there is 1 offset, which goes into the 2 low bytes of the
806 // first source, or there are 4 offsets, which go into 2 sources (8
807 // values, 1 byte each).
808 Value *offs[2] = {NULL, NULL};
809 for (n = 0; n < i->tex.useOffsets; n++) {
810 for (c = 0; c < 2; ++c) {
811 if ((n % 2) == 0 && c == 0)
812 offs[n / 2] = i->offset[n][c].get();
814 bld.mkOp3(OP_INSBF, TYPE_U32,
816 i->offset[n][c].get(),
817 bld.mkImm(0x800 | ((n * 16 + c * 8) % 32)),
821 i->setSrc(s, offs[0]);
823 i->setSrc(s + 1, offs[1]);
826 assert(i->tex.useOffsets == 1);
827 for (c = 0; c < 3; ++c) {
829 if (!i->offset[0][c].getImmediate(val))
830 assert(!"non-immediate offset passed to non-TXG");
831 imm |= (val.reg.data.u32 & 0xf) << (c * 4);
833 if (i->op == OP_TXD && chipset >= NVISA_GK104_CHIPSET) {
834 // The offset goes into the upper 16 bits of the array index. So
835 // create it if it's not already there, and INSBF it if it already
837 s = (i->tex.rIndirectSrc >= 0) ? 1 : 0;
838 if (chipset >= NVISA_GM107_CHIPSET)
840 if (i->tex.target.isArray()) {
841 bld.mkOp3(OP_INSBF, TYPE_U32, i->getSrc(s),
842 bld.loadImm(NULL, imm), bld.mkImm(0xc10),
845 i->moveSources(s, 1);
846 i->setSrc(s, bld.loadImm(NULL, imm << 16));
849 i->setSrc(s, bld.loadImm(NULL, imm));
854 if (chipset >= NVISA_GK104_CHIPSET) {
856 // If TEX requires more than 4 sources, the 2nd register tuple must be
857 // aligned to 4, even if it consists of just a single 4-byte register.
859 // XXX HACK: We insert 0 sources to avoid the 5 or 6 regs case.
861 int s = i->srcCount(0xff, true);
862 if (s > 4 && s < 7) {
863 if (i->srcExists(s)) // move potential predicate out of the way
864 i->moveSources(s, 7 - s);
866 i->setSrc(s++, bld.loadImm(NULL, 0));
874 NVC0LoweringPass::handleManualTXD(TexInstruction *i)
876 static const uint8_t qOps[4][2] =
878 { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(MOV2, MOV2, ADD, ADD) }, // l0
879 { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(MOV2, MOV2, ADD, ADD) }, // l1
880 { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l2
881 { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l3
886 Value *zero = bld.loadImm(bld.getSSA(), 0);
888 const int dim = i->tex.target.getDim() + i->tex.target.isCube();
890 // This function is invoked after handleTEX lowering, so we have to expect
891 // the arguments in the order that the hw wants them. For Fermi, array and
892 // indirect are both in the leading arg, while for Kepler, array and
893 // indirect are separate (and both precede the coordinates). Maxwell is
894 // handled in a separate function.
896 if (targ->getChipset() < NVISA_GK104_CHIPSET)
897 array = i->tex.target.isArray() || i->tex.rIndirectSrc >= 0;
899 array = i->tex.target.isArray() + (i->tex.rIndirectSrc >= 0);
901 i->op = OP_TEX; // no need to clone dPdx/dPdy later
903 for (c = 0; c < dim; ++c)
904 crd[c] = bld.getScratch();
906 bld.mkOp(OP_QUADON, TYPE_NONE, NULL);
907 for (l = 0; l < 4; ++l) {
909 // mov coordinates from lane l to all lanes
910 for (c = 0; c < dim; ++c)
911 bld.mkQuadop(0x00, crd[c], l, i->getSrc(c + array), zero);
912 // add dPdx from lane l to lanes dx
913 for (c = 0; c < dim; ++c)
914 bld.mkQuadop(qOps[l][0], crd[c], l, i->dPdx[c].get(), crd[c]);
915 // add dPdy from lane l to lanes dy
916 for (c = 0; c < dim; ++c)
917 bld.mkQuadop(qOps[l][1], crd[c], l, i->dPdy[c].get(), crd[c]);
918 // normalize cube coordinates
919 if (i->tex.target.isCube()) {
920 for (c = 0; c < 3; ++c)
921 src[c] = bld.mkOp1v(OP_ABS, TYPE_F32, bld.getSSA(), crd[c]);
922 val = bld.getScratch();
923 bld.mkOp2(OP_MAX, TYPE_F32, val, src[0], src[1]);
924 bld.mkOp2(OP_MAX, TYPE_F32, val, src[2], val);
925 bld.mkOp1(OP_RCP, TYPE_F32, val, val);
926 for (c = 0; c < 3; ++c)
927 src[c] = bld.mkOp2v(OP_MUL, TYPE_F32, bld.getSSA(), crd[c], val);
929 for (c = 0; c < dim; ++c)
933 bld.insert(tex = cloneForward(func, i));
934 for (c = 0; c < dim; ++c)
935 tex->setSrc(c + array, src[c]);
937 for (c = 0; i->defExists(c); ++c) {
939 def[c][l] = bld.getSSA();
940 mov = bld.mkMov(def[c][l], tex->getDef(c));
945 bld.mkOp(OP_QUADPOP, TYPE_NONE, NULL);
947 for (c = 0; i->defExists(c); ++c) {
948 Instruction *u = bld.mkOp(OP_UNION, TYPE_U32, i->getDef(c));
949 for (l = 0; l < 4; ++l)
950 u->setSrc(l, def[c][l]);
958 NVC0LoweringPass::handleTXD(TexInstruction *txd)
960 int dim = txd->tex.target.getDim() + txd->tex.target.isCube();
961 unsigned arg = txd->tex.target.getArgCount();
962 unsigned expected_args = arg;
963 const int chipset = prog->getTarget()->getChipset();
965 if (chipset >= NVISA_GK104_CHIPSET) {
966 if (!txd->tex.target.isArray() && txd->tex.useOffsets)
968 if (txd->tex.rIndirectSrc >= 0 || txd->tex.sIndirectSrc >= 0)
971 if (txd->tex.useOffsets)
973 if (!txd->tex.target.isArray() && (
974 txd->tex.rIndirectSrc >= 0 || txd->tex.sIndirectSrc >= 0))
978 if (expected_args > 4 ||
980 txd->tex.target.isShadow())
984 while (txd->srcExists(arg))
987 txd->tex.derivAll = true;
988 if (txd->op == OP_TEX)
989 return handleManualTXD(txd);
991 assert(arg == expected_args);
992 for (int c = 0; c < dim; ++c) {
993 txd->setSrc(arg + c * 2 + 0, txd->dPdx[c]);
994 txd->setSrc(arg + c * 2 + 1, txd->dPdy[c]);
995 txd->dPdx[c].set(NULL);
996 txd->dPdy[c].set(NULL);
999 // In this case we have fewer than 4 "real" arguments, which means that
1000 // handleTEX didn't apply any padding. However we have to make sure that
1001 // the second "group" of arguments still gets padded up to 4.
1002 if (chipset >= NVISA_GK104_CHIPSET) {
1003 int s = arg + 2 * dim;
1004 if (s >= 4 && s < 7) {
1005 if (txd->srcExists(s)) // move potential predicate out of the way
1006 txd->moveSources(s, 7 - s);
1008 txd->setSrc(s++, bld.loadImm(NULL, 0));
1016 NVC0LoweringPass::handleTXQ(TexInstruction *txq)
1018 const int chipset = prog->getTarget()->getChipset();
1019 if (chipset >= NVISA_GK104_CHIPSET && txq->tex.rIndirectSrc < 0)
1020 txq->tex.r += prog->driver->io.texBindBase / 4;
1022 if (txq->tex.rIndirectSrc < 0)
1025 Value *ticRel = txq->getIndirectR();
1027 txq->setIndirectS(NULL);
1028 txq->tex.sIndirectSrc = -1;
1032 if (chipset < NVISA_GK104_CHIPSET) {
1033 LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa
1035 txq->setSrc(txq->tex.rIndirectSrc, NULL);
1037 ticRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
1038 ticRel, bld.mkImm(txq->tex.r));
1040 bld.mkOp2(OP_SHL, TYPE_U32, src, ticRel, bld.mkImm(0x17));
1042 txq->moveSources(0, 1);
1043 txq->setSrc(0, src);
1045 Value *hnd = loadTexHandle(
1046 bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
1047 txq->getIndirectR(), bld.mkImm(2)),
1052 txq->setIndirectR(NULL);
1053 txq->moveSources(0, 1);
1054 txq->setSrc(0, hnd);
1055 txq->tex.rIndirectSrc = 0;
1062 NVC0LoweringPass::handleTXLQ(TexInstruction *i)
1064 /* The outputs are inverted compared to what the TGSI instruction
1065 * expects. Take that into account in the mask.
1067 assert((i->tex.mask & ~3) == 0);
1068 if (i->tex.mask == 1)
1070 else if (i->tex.mask == 2)
1073 bld.setPosition(i, true);
1075 /* The returned values are not quite what we want:
1076 * (a) convert from s16/u16 to f32
1077 * (b) multiply by 1/256
1079 for (int def = 0; def < 2; ++def) {
1080 if (!i->defExists(def))
1082 enum DataType type = TYPE_S16;
1083 if (i->tex.mask == 2 || def > 0)
1085 bld.mkCvt(OP_CVT, TYPE_F32, i->getDef(def), type, i->getDef(def));
1086 bld.mkOp2(OP_MUL, TYPE_F32, i->getDef(def),
1087 i->getDef(def), bld.loadImm(NULL, 1.0f / 256));
1089 if (i->tex.mask == 3) {
1090 LValue *t = new_LValue(func, FILE_GPR);
1091 bld.mkMov(t, i->getDef(0));
1092 bld.mkMov(i->getDef(0), i->getDef(1));
1093 bld.mkMov(i->getDef(1), t);
1099 NVC0LoweringPass::handleBUFQ(Instruction *bufq)
1102 bufq->setSrc(0, loadBufLength32(bufq->getIndirect(0, 1),
1103 bufq->getSrc(0)->reg.fileIndex * 16));
1104 bufq->setIndirect(0, 0, NULL);
1105 bufq->setIndirect(0, 1, NULL);
1110 NVC0LoweringPass::handleSharedATOMNVE4(Instruction *atom)
1112 assert(atom->src(0).getFile() == FILE_MEMORY_SHARED);
1114 BasicBlock *currBB = atom->bb;
1115 BasicBlock *tryLockBB = atom->bb->splitBefore(atom, false);
1116 BasicBlock *joinBB = atom->bb->splitAfter(atom);
1117 BasicBlock *setAndUnlockBB = new BasicBlock(func);
1118 BasicBlock *failLockBB = new BasicBlock(func);
1120 bld.setPosition(currBB, true);
1121 assert(!currBB->joinAt);
1122 currBB->joinAt = bld.mkFlow(OP_JOINAT, joinBB, CC_ALWAYS, NULL);
1124 CmpInstruction *pred =
1125 bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
1126 TYPE_U32, bld.mkImm(0), bld.mkImm(1));
1128 bld.mkFlow(OP_BRA, tryLockBB, CC_ALWAYS, NULL);
1129 currBB->cfg.attach(&tryLockBB->cfg, Graph::Edge::TREE);
1131 bld.setPosition(tryLockBB, true);
1134 bld.mkLoad(TYPE_U32, atom->getDef(0), atom->getSrc(0)->asSym(),
1135 atom->getIndirect(0, 0));
1136 ld->setDef(1, bld.getSSA(1, FILE_PREDICATE));
1137 ld->subOp = NV50_IR_SUBOP_LOAD_LOCKED;
1139 bld.mkFlow(OP_BRA, setAndUnlockBB, CC_P, ld->getDef(1));
1140 bld.mkFlow(OP_BRA, failLockBB, CC_ALWAYS, NULL);
1141 tryLockBB->cfg.attach(&failLockBB->cfg, Graph::Edge::CROSS);
1142 tryLockBB->cfg.attach(&setAndUnlockBB->cfg, Graph::Edge::TREE);
1144 tryLockBB->cfg.detach(&joinBB->cfg);
1147 bld.setPosition(setAndUnlockBB, true);
1149 if (atom->subOp == NV50_IR_SUBOP_ATOM_EXCH) {
1150 // Read the old value, and write the new one.
1151 stVal = atom->getSrc(1);
1152 } else if (atom->subOp == NV50_IR_SUBOP_ATOM_CAS) {
1153 CmpInstruction *set =
1154 bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(),
1155 TYPE_U32, ld->getDef(0), atom->getSrc(1));
1157 bld.mkCmp(OP_SLCT, CC_NE, TYPE_U32, (stVal = bld.getSSA()),
1158 TYPE_U32, atom->getSrc(2), ld->getDef(0), set->getDef(0));
1162 switch (atom->subOp) {
1163 case NV50_IR_SUBOP_ATOM_ADD:
1166 case NV50_IR_SUBOP_ATOM_AND:
1169 case NV50_IR_SUBOP_ATOM_OR:
1172 case NV50_IR_SUBOP_ATOM_XOR:
1175 case NV50_IR_SUBOP_ATOM_MIN:
1178 case NV50_IR_SUBOP_ATOM_MAX:
1186 stVal = bld.mkOp2v(op, atom->dType, bld.getSSA(), ld->getDef(0),
1191 bld.mkStore(OP_STORE, TYPE_U32, atom->getSrc(0)->asSym(),
1192 atom->getIndirect(0, 0), stVal);
1193 st->setDef(0, pred->getDef(0));
1194 st->subOp = NV50_IR_SUBOP_STORE_UNLOCKED;
1196 bld.mkFlow(OP_BRA, failLockBB, CC_ALWAYS, NULL);
1197 setAndUnlockBB->cfg.attach(&failLockBB->cfg, Graph::Edge::TREE);
1199 // Lock until the store has not been performed.
1200 bld.setPosition(failLockBB, true);
1201 bld.mkFlow(OP_BRA, tryLockBB, CC_NOT_P, pred->getDef(0));
1202 bld.mkFlow(OP_BRA, joinBB, CC_ALWAYS, NULL);
1203 failLockBB->cfg.attach(&tryLockBB->cfg, Graph::Edge::BACK);
1204 failLockBB->cfg.attach(&joinBB->cfg, Graph::Edge::TREE);
1206 bld.setPosition(joinBB, false);
1207 bld.mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1;
1211 NVC0LoweringPass::handleSharedATOM(Instruction *atom)
1213 assert(atom->src(0).getFile() == FILE_MEMORY_SHARED);
1215 BasicBlock *currBB = atom->bb;
1216 BasicBlock *tryLockAndSetBB = atom->bb->splitBefore(atom, false);
1217 BasicBlock *joinBB = atom->bb->splitAfter(atom);
1219 bld.setPosition(currBB, true);
1220 assert(!currBB->joinAt);
1221 currBB->joinAt = bld.mkFlow(OP_JOINAT, joinBB, CC_ALWAYS, NULL);
1223 bld.mkFlow(OP_BRA, tryLockAndSetBB, CC_ALWAYS, NULL);
1224 currBB->cfg.attach(&tryLockAndSetBB->cfg, Graph::Edge::TREE);
1226 bld.setPosition(tryLockAndSetBB, true);
1229 bld.mkLoad(TYPE_U32, atom->getDef(0), atom->getSrc(0)->asSym(),
1230 atom->getIndirect(0, 0));
1231 ld->setDef(1, bld.getSSA(1, FILE_PREDICATE));
1232 ld->subOp = NV50_IR_SUBOP_LOAD_LOCKED;
1235 if (atom->subOp == NV50_IR_SUBOP_ATOM_EXCH) {
1236 // Read the old value, and write the new one.
1237 stVal = atom->getSrc(1);
1238 } else if (atom->subOp == NV50_IR_SUBOP_ATOM_CAS) {
1239 CmpInstruction *set =
1240 bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
1241 TYPE_U32, ld->getDef(0), atom->getSrc(1));
1242 set->setPredicate(CC_P, ld->getDef(1));
1245 bld.mkOp3(OP_SELP, TYPE_U32, bld.getSSA(), ld->getDef(0),
1246 atom->getSrc(2), set->getDef(0));
1247 selp->src(2).mod = Modifier(NV50_IR_MOD_NOT);
1248 selp->setPredicate(CC_P, ld->getDef(1));
1250 stVal = selp->getDef(0);
1254 switch (atom->subOp) {
1255 case NV50_IR_SUBOP_ATOM_ADD:
1258 case NV50_IR_SUBOP_ATOM_AND:
1261 case NV50_IR_SUBOP_ATOM_OR:
1264 case NV50_IR_SUBOP_ATOM_XOR:
1267 case NV50_IR_SUBOP_ATOM_MIN:
1270 case NV50_IR_SUBOP_ATOM_MAX:
1279 bld.mkOp2(op, atom->dType, bld.getSSA(), ld->getDef(0),
1281 i->setPredicate(CC_P, ld->getDef(1));
1283 stVal = i->getDef(0);
1287 bld.mkStore(OP_STORE, TYPE_U32, atom->getSrc(0)->asSym(),
1288 atom->getIndirect(0, 0), stVal);
1289 st->setPredicate(CC_P, ld->getDef(1));
1290 st->subOp = NV50_IR_SUBOP_STORE_UNLOCKED;
1292 // Loop until the lock is acquired.
1293 bld.mkFlow(OP_BRA, tryLockAndSetBB, CC_NOT_P, ld->getDef(1));
1294 tryLockAndSetBB->cfg.attach(&tryLockAndSetBB->cfg, Graph::Edge::BACK);
1295 tryLockAndSetBB->cfg.attach(&joinBB->cfg, Graph::Edge::CROSS);
1296 bld.mkFlow(OP_BRA, joinBB, CC_ALWAYS, NULL);
1300 bld.setPosition(joinBB, false);
1301 bld.mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1;
1305 NVC0LoweringPass::handleATOM(Instruction *atom)
1308 Value *ptr = atom->getIndirect(0, 0), *ind = atom->getIndirect(0, 1), *base;
1310 switch (atom->src(0).getFile()) {
1311 case FILE_MEMORY_LOCAL:
1314 case FILE_MEMORY_SHARED:
1315 // For Fermi/Kepler, we have to use ld lock/st unlock to perform atomic
1316 // operations on shared memory. For Maxwell, ATOMS is enough.
1317 if (targ->getChipset() < NVISA_GK104_CHIPSET)
1318 handleSharedATOM(atom);
1319 else if (targ->getChipset() < NVISA_GM107_CHIPSET)
1320 handleSharedATOMNVE4(atom);
1323 assert(atom->src(0).getFile() == FILE_MEMORY_BUFFER);
1324 base = loadBufInfo64(ind, atom->getSrc(0)->reg.fileIndex * 16);
1325 assert(base->reg.size == 8);
1327 base = bld.mkOp2v(OP_ADD, TYPE_U64, base, base, ptr);
1328 assert(base->reg.size == 8);
1329 atom->setIndirect(0, 0, base);
1330 atom->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
1332 // Harden against out-of-bounds accesses
1333 Value *offset = bld.loadImm(NULL, atom->getSrc(0)->reg.data.offset + typeSizeof(atom->sType));
1334 Value *length = loadBufLength32(ind, atom->getSrc(0)->reg.fileIndex * 16);
1335 Value *pred = new_LValue(func, FILE_PREDICATE);
1337 bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, ptr);
1338 bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
1339 atom->setPredicate(CC_NOT_P, pred);
1340 if (atom->defExists(0)) {
1341 Value *zero, *dst = atom->getDef(0);
1342 atom->setDef(0, bld.getSSA());
1344 bld.setPosition(atom, true);
1345 bld.mkMov((zero = bld.getSSA()), bld.mkImm(0))
1346 ->setPredicate(CC_P, pred);
1347 bld.mkOp2(OP_UNION, TYPE_U32, dst, atom->getDef(0), zero);
1353 bld.mkOp1v(OP_RDSV, TYPE_U32, bld.getScratch(), bld.mkSysVal(sv, 0));
1355 atom->setSrc(0, cloneShallow(func, atom->getSrc(0)));
1356 atom->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
1358 base = bld.mkOp2v(OP_ADD, TYPE_U32, base, base, ptr);
1359 atom->setIndirect(0, 1, NULL);
1360 atom->setIndirect(0, 0, base);
1366 NVC0LoweringPass::handleCasExch(Instruction *cas, bool needCctl)
1368 if (targ->getChipset() < NVISA_GM107_CHIPSET) {
1369 if (cas->src(0).getFile() == FILE_MEMORY_SHARED) {
1370 // ATOM_CAS and ATOM_EXCH are handled in handleSharedATOM().
1375 if (cas->subOp != NV50_IR_SUBOP_ATOM_CAS &&
1376 cas->subOp != NV50_IR_SUBOP_ATOM_EXCH)
1378 bld.setPosition(cas, true);
1381 Instruction *cctl = bld.mkOp1(OP_CCTL, TYPE_NONE, NULL, cas->getSrc(0));
1382 cctl->setIndirect(0, 0, cas->getIndirect(0, 0));
1384 cctl->subOp = NV50_IR_SUBOP_CCTL_IV;
1385 if (cas->isPredicated())
1386 cctl->setPredicate(cas->cc, cas->getPredicate());
1389 if (cas->subOp == NV50_IR_SUBOP_ATOM_CAS) {
1390 // CAS is crazy. It's 2nd source is a double reg, and the 3rd source
1391 // should be set to the high part of the double reg or bad things will
1392 // happen elsewhere in the universe.
1393 // Also, it sometimes returns the new value instead of the old one
1394 // under mysterious circumstances.
1395 Value *dreg = bld.getSSA(8);
1396 bld.setPosition(cas, false);
1397 bld.mkOp2(OP_MERGE, TYPE_U64, dreg, cas->getSrc(1), cas->getSrc(2));
1398 cas->setSrc(1, dreg);
1399 cas->setSrc(2, dreg);
1406 NVC0LoweringPass::loadResInfo32(Value *ptr, uint32_t off, uint16_t base)
1408 uint8_t b = prog->driver->io.auxCBSlot;
1412 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
1416 NVC0LoweringPass::loadResInfo64(Value *ptr, uint32_t off, uint16_t base)
1418 uint8_t b = prog->driver->io.auxCBSlot;
1422 ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4));
1425 mkLoadv(TYPE_U64, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off), ptr);
1429 NVC0LoweringPass::loadResLength32(Value *ptr, uint32_t off, uint16_t base)
1431 uint8_t b = prog->driver->io.auxCBSlot;
1435 ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4));
1438 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off + 8), ptr);
1442 NVC0LoweringPass::loadSuInfo32(Value *ptr, uint32_t off)
1444 return loadResInfo32(ptr, off, prog->driver->io.suInfoBase);
1448 NVC0LoweringPass::loadSuInfo64(Value *ptr, uint32_t off)
1450 return loadResInfo64(ptr, off, prog->driver->io.suInfoBase);
1454 NVC0LoweringPass::loadSuLength32(Value *ptr, uint32_t off)
1456 return loadResLength32(ptr, off, prog->driver->io.suInfoBase);
1460 NVC0LoweringPass::loadBufInfo32(Value *ptr, uint32_t off)
1462 return loadResInfo32(ptr, off, prog->driver->io.bufInfoBase);
1466 NVC0LoweringPass::loadBufInfo64(Value *ptr, uint32_t off)
1468 return loadResInfo64(ptr, off, prog->driver->io.bufInfoBase);
1472 NVC0LoweringPass::loadBufLength32(Value *ptr, uint32_t off)
1474 return loadResLength32(ptr, off, prog->driver->io.bufInfoBase);
1478 NVC0LoweringPass::loadUboInfo32(Value *ptr, uint32_t off)
1480 return loadResInfo32(ptr, off, prog->driver->io.uboInfoBase);
1484 NVC0LoweringPass::loadUboInfo64(Value *ptr, uint32_t off)
1486 return loadResInfo64(ptr, off, prog->driver->io.uboInfoBase);
1490 NVC0LoweringPass::loadUboLength32(Value *ptr, uint32_t off)
1492 return loadResLength32(ptr, off, prog->driver->io.uboInfoBase);
1496 NVC0LoweringPass::loadMsInfo32(Value *ptr, uint32_t off)
1498 uint8_t b = prog->driver->io.msInfoCBSlot;
1499 off += prog->driver->io.msInfoBase;
1501 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
1504 /* On nvc0, surface info is obtained via the surface binding points passed
1505 * to the SULD/SUST instructions.
1506 * On nve4, surface info is stored in c[] and is used by various special
1507 * instructions, e.g. for clamping coordiantes or generating an address.
1508 * They couldn't just have added an equivalent to TIC now, couldn't they ?
1510 #define NVE4_SU_INFO_ADDR 0x00
1511 #define NVE4_SU_INFO_FMT 0x04
1512 #define NVE4_SU_INFO_DIM_X 0x08
1513 #define NVE4_SU_INFO_PITCH 0x0c
1514 #define NVE4_SU_INFO_DIM_Y 0x10
1515 #define NVE4_SU_INFO_ARRAY 0x14
1516 #define NVE4_SU_INFO_DIM_Z 0x18
1517 #define NVE4_SU_INFO_UNK1C 0x1c
1518 #define NVE4_SU_INFO_WIDTH 0x20
1519 #define NVE4_SU_INFO_HEIGHT 0x24
1520 #define NVE4_SU_INFO_DEPTH 0x28
1521 #define NVE4_SU_INFO_TARGET 0x2c
1522 #define NVE4_SU_INFO_BSIZE 0x30
1523 #define NVE4_SU_INFO_RAW_X 0x34
1524 #define NVE4_SU_INFO_MS_X 0x38
1525 #define NVE4_SU_INFO_MS_Y 0x3c
1527 #define NVE4_SU_INFO__STRIDE 0x40
1529 #define NVE4_SU_INFO_DIM(i) (0x08 + (i) * 8)
1530 #define NVE4_SU_INFO_SIZE(i) (0x20 + (i) * 4)
1531 #define NVE4_SU_INFO_MS(i) (0x38 + (i) * 4)
1533 static inline uint16_t getSuClampSubOp(const TexInstruction *su, int c)
1535 switch (su->tex.target.getEnum()) {
1536 case TEX_TARGET_BUFFER: return NV50_IR_SUBOP_SUCLAMP_PL(0, 1);
1537 case TEX_TARGET_RECT: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1538 case TEX_TARGET_1D: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1539 case TEX_TARGET_1D_ARRAY: return (c == 1) ?
1540 NV50_IR_SUBOP_SUCLAMP_PL(0, 2) :
1541 NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1542 case TEX_TARGET_2D: return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
1543 case TEX_TARGET_2D_MS: return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
1544 case TEX_TARGET_2D_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1545 case TEX_TARGET_2D_MS_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1546 case TEX_TARGET_3D: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1547 case TEX_TARGET_CUBE: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1548 case TEX_TARGET_CUBE_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1556 NVC0LoweringPass::handleSUQ(TexInstruction *suq)
1558 int dim = suq->tex.target.getDim();
1559 int arg = dim + (suq->tex.target.isArray() || suq->tex.target.isCube());
1560 uint8_t s = prog->driver->io.auxCBSlot;
1561 Value *ind = suq->getIndirectR();
1565 base = prog->driver->io.suInfoBase + suq->tex.r * NVE4_SU_INFO__STRIDE;
1568 ind = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(),
1571 for (c = 0; c < arg; ++c) {
1572 if (suq->defExists(c)) {
1575 if (c == 1 && suq->tex.target == TEX_TARGET_1D_ARRAY) {
1576 offset = base + NVE4_SU_INFO_SIZE(2);
1578 offset = base + NVE4_SU_INFO_SIZE(c);
1580 bld.mkLoad(TYPE_U32, suq->getDef(c),
1581 bld.mkSymbol(FILE_MEMORY_CONST, s, TYPE_U32, offset), ind);
1585 if (suq->tex.target.isCube()) {
1586 if (suq->defExists(2)) {
1587 bld.mkOp2(OP_DIV, TYPE_U32, suq->getDef(2), suq->getDef(2),
1588 bld.loadImm(NULL, 6));
1592 if (suq->defExists(3)) {
1593 // .w contains the number of samples for multi-sampled images but we
1594 // don't support them for now.
1595 bld.mkMov(suq->getDef(3), bld.loadImm(NULL, 1));
1603 NVC0LoweringPass::adjustCoordinatesMS(TexInstruction *tex)
1605 const uint16_t base = tex->tex.r * NVE4_SU_INFO__STRIDE;
1606 const int arg = tex->tex.target.getArgCount();
1608 if (tex->tex.target == TEX_TARGET_2D_MS)
1609 tex->tex.target = TEX_TARGET_2D;
1611 if (tex->tex.target == TEX_TARGET_2D_MS_ARRAY)
1612 tex->tex.target = TEX_TARGET_2D_ARRAY;
1616 Value *x = tex->getSrc(0);
1617 Value *y = tex->getSrc(1);
1618 Value *s = tex->getSrc(arg - 1);
1620 Value *tx = bld.getSSA(), *ty = bld.getSSA(), *ts = bld.getSSA();
1623 if (tex->tex.rIndirectSrc >= 0) {
1624 assert(tex->tex.r == 0);
1625 // FIXME: out of bounds
1626 ind = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
1627 tex->getIndirectR(), bld.mkImm(6));
1630 Value *ms_x = loadSuInfo32(ind, base + NVE4_SU_INFO_MS(0));
1631 Value *ms_y = loadSuInfo32(ind, base + NVE4_SU_INFO_MS(1));
1633 bld.mkOp2(OP_SHL, TYPE_U32, tx, x, ms_x);
1634 bld.mkOp2(OP_SHL, TYPE_U32, ty, y, ms_y);
1636 s = bld.mkOp2v(OP_AND, TYPE_U32, ts, s, bld.loadImm(NULL, 0x7));
1637 s = bld.mkOp2v(OP_SHL, TYPE_U32, ts, ts, bld.mkImm(3));
1639 Value *dx = loadMsInfo32(ts, 0x0);
1640 Value *dy = loadMsInfo32(ts, 0x4);
1642 bld.mkOp2(OP_ADD, TYPE_U32, tx, tx, dx);
1643 bld.mkOp2(OP_ADD, TYPE_U32, ty, ty, dy);
1647 tex->moveSources(arg, -1);
1650 // Sets 64-bit "generic address", predicate and format sources for SULD/SUST.
1651 // They're computed from the coordinates using the surface info in c[] space.
1653 NVC0LoweringPass::processSurfaceCoordsNVE4(TexInstruction *su)
1656 const bool atom = su->op == OP_SUREDB || su->op == OP_SUREDP;
1658 su->op == OP_SULDB || su->op == OP_SUSTB || su->op == OP_SUREDB;
1659 const int idx = su->tex.r;
1660 const int dim = su->tex.target.getDim();
1661 const int arg = dim + (su->tex.target.isArray() || su->tex.target.isCube());
1662 const uint16_t base = idx * NVE4_SU_INFO__STRIDE;
1664 Value *zero = bld.mkImm(0);
1668 Value *bf, *eau, *off;
1672 off = bld.getScratch(4);
1673 bf = bld.getScratch(4);
1674 addr = bld.getSSA(8);
1675 pred = bld.getScratch(1, FILE_PREDICATE);
1677 bld.setPosition(su, false);
1679 adjustCoordinatesMS(su);
1681 if (su->tex.rIndirectSrc >= 0) {
1682 ind = su->getIndirectR();
1683 if (su->tex.r > 0) {
1684 ind = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), ind,
1685 bld.loadImm(NULL, su->tex.r));
1687 ind = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ind, bld.mkImm(7));
1688 ind = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), ind, bld.mkImm(6));
1691 // calculate clamped coordinates
1692 for (c = 0; c < arg; ++c) {
1695 if (c == 1 && su->tex.target == TEX_TARGET_1D_ARRAY) {
1696 // The array index is stored in the Z component for 1D arrays.
1700 src[c] = bld.getScratch();
1702 v = loadSuInfo32(ind, base + NVE4_SU_INFO_RAW_X);
1704 v = loadSuInfo32(ind, base + NVE4_SU_INFO_DIM(dimc));
1705 bld.mkOp3(OP_SUCLAMP, TYPE_S32, src[c], su->getSrc(c), v, zero)
1706 ->subOp = getSuClampSubOp(su, dimc);
1711 // set predicate output
1712 if (su->tex.target == TEX_TARGET_BUFFER) {
1713 src[0]->getInsn()->setFlagsDef(1, pred);
1715 if (su->tex.target.isArray() || su->tex.target.isCube()) {
1716 p1 = bld.getSSA(1, FILE_PREDICATE);
1717 src[dim]->getInsn()->setFlagsDef(1, p1);
1720 // calculate pixel offset
1722 if (su->tex.target != TEX_TARGET_BUFFER)
1723 bld.mkOp2(OP_AND, TYPE_U32, off, src[0], bld.loadImm(NULL, 0xffff));
1726 v = loadSuInfo32(ind, base + NVE4_SU_INFO_UNK1C);
1727 bld.mkOp3(OP_MADSP, TYPE_U32, off, src[2], v, src[1])
1728 ->subOp = NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
1730 v = loadSuInfo32(ind, base + NVE4_SU_INFO_PITCH);
1731 bld.mkOp3(OP_MADSP, TYPE_U32, off, off, v, src[0])
1732 ->subOp = NV50_IR_SUBOP_MADSP(0,2,8); // u32 u16l u16l
1735 v = loadSuInfo32(ind, base + NVE4_SU_INFO_PITCH);
1736 bld.mkOp3(OP_MADSP, TYPE_U32, off, src[1], v, src[0])
1737 ->subOp = (su->tex.target.isArray() || su->tex.target.isCube()) ?
1738 NV50_IR_SUBOP_MADSP_SD : NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
1741 // calculate effective address part 1
1742 if (su->tex.target == TEX_TARGET_BUFFER) {
1746 v = loadSuInfo32(ind, base + NVE4_SU_INFO_FMT);
1747 bld.mkOp3(OP_VSHL, TYPE_U32, bf, src[0], v, zero)
1748 ->subOp = NV50_IR_SUBOP_V1(7,6,8|2);
1762 if (!su->tex.target.isArray() && !su->tex.target.isCube()) {
1763 z = loadSuInfo32(ind, base + NVE4_SU_INFO_UNK1C);
1764 subOp = NV50_IR_SUBOP_SUBFM_3D;
1768 subOp = NV50_IR_SUBOP_SUBFM_3D;
1772 insn = bld.mkOp3(OP_SUBFM, TYPE_U32, bf, src[0], y, z);
1773 insn->subOp = subOp;
1774 insn->setFlagsDef(1, pred);
1778 v = loadSuInfo32(ind, base + NVE4_SU_INFO_ADDR);
1780 if (su->tex.target == TEX_TARGET_BUFFER) {
1783 eau = bld.mkOp3v(OP_SUEAU, TYPE_U32, bld.getScratch(4), off, bf, v);
1785 // add array layer offset
1786 if (su->tex.target.isArray() || su->tex.target.isCube()) {
1787 v = loadSuInfo32(ind, base + NVE4_SU_INFO_ARRAY);
1789 bld.mkOp3(OP_MADSP, TYPE_U32, eau, src[1], v, eau)
1790 ->subOp = NV50_IR_SUBOP_MADSP(4,0,0); // u16 u24 u32
1792 bld.mkOp3(OP_MADSP, TYPE_U32, eau, v, src[2], eau)
1793 ->subOp = NV50_IR_SUBOP_MADSP(0,0,0); // u32 u24 u32
1794 // combine predicates
1796 bld.mkOp2(OP_OR, TYPE_U8, pred, pred, p1);
1801 if (su->tex.target == TEX_TARGET_BUFFER) {
1805 // bf == g[] address & 0xff
1806 // eau == g[] address >> 8
1807 bld.mkOp3(OP_PERMT, TYPE_U32, bf, lo, bld.loadImm(NULL, 0x6540), eau);
1808 bld.mkOp3(OP_PERMT, TYPE_U32, eau, zero, bld.loadImm(NULL, 0x0007), eau);
1810 if (su->op == OP_SULDP && su->tex.target == TEX_TARGET_BUFFER) {
1811 // Convert from u32 to u8 address format, which is what the library code
1812 // doing SULDP currently uses.
1813 // XXX: can SUEAU do this ?
1814 // XXX: does it matter that we don't mask high bytes in bf ?
1816 bld.mkOp2(OP_SHR, TYPE_U32, off, bf, bld.mkImm(8));
1817 bld.mkOp2(OP_ADD, TYPE_U32, eau, eau, off);
1820 bld.mkOp2(OP_MERGE, TYPE_U64, addr, bf, eau);
1822 if (atom && su->tex.target == TEX_TARGET_BUFFER)
1823 bld.mkOp2(OP_ADD, TYPE_U64, addr, addr, off);
1825 // let's just set it 0 for raw access and hope it works
1827 bld.mkImm(0) : loadSuInfo32(ind, base + NVE4_SU_INFO_FMT);
1829 // get rid of old coordinate sources, make space for fmt info and predicate
1830 su->moveSources(arg, 3 - arg);
1831 // set 64 bit address and 32-bit format sources
1832 su->setSrc(0, addr);
1834 su->setSrc(2, pred);
1836 // prevent read fault when the image is not actually bound
1837 CmpInstruction *pred1 =
1838 bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
1839 TYPE_U32, bld.mkImm(0),
1840 loadSuInfo32(ind, base + NVE4_SU_INFO_ADDR));
1842 if (su->op != OP_SUSTP && su->tex.format) {
1843 const TexInstruction::ImgFormatDesc *format = su->tex.format;
1844 int blockwidth = format->bits[0] + format->bits[1] +
1845 format->bits[2] + format->bits[3];
1847 // make sure that the format doesn't mismatch
1848 assert(format->components != 0);
1849 bld.mkCmp(OP_SET_OR, CC_NE, TYPE_U32, pred1->getDef(0),
1850 TYPE_U32, bld.loadImm(NULL, blockwidth / 8),
1851 loadSuInfo32(ind, base + NVE4_SU_INFO_BSIZE),
1854 su->setPredicate(CC_NOT_P, pred1->getDef(0));
1856 // TODO: initialize def values to 0 when the surface operation is not
1857 // performed (not needed for stores). Also, fix the "address bounds test"
1858 // subtests from arb_shader_image_load_store-invalid for buffers, because it
1859 // seems like that the predicate is not correctly set by suclamp.
1863 getSrcType(const TexInstruction::ImgFormatDesc *t, int c)
1866 case FLOAT: return t->bits[c] == 16 ? TYPE_F16 : TYPE_F32;
1867 case UNORM: return t->bits[c] == 8 ? TYPE_U8 : TYPE_U16;
1868 case SNORM: return t->bits[c] == 8 ? TYPE_S8 : TYPE_S16;
1870 return (t->bits[c] == 8 ? TYPE_U8 :
1871 (t->bits[c] == 16 ? TYPE_U16 : TYPE_U32));
1873 return (t->bits[c] == 8 ? TYPE_S8 :
1874 (t->bits[c] == 16 ? TYPE_S16 : TYPE_S32));
1880 getDestType(const ImgType type) {
1891 assert(!"Impossible type");
1897 NVC0LoweringPass::convertSurfaceFormat(TexInstruction *su)
1899 const TexInstruction::ImgFormatDesc *format = su->tex.format;
1900 int width = format->bits[0] + format->bits[1] +
1901 format->bits[2] + format->bits[3];
1902 Value *untypedDst[4] = {};
1903 Value *typedDst[4] = {};
1905 // We must convert this to a generic load.
1908 su->dType = typeOfSize(width / 8);
1909 su->sType = TYPE_U8;
1911 for (int i = 0; i < width / 32; i++)
1912 untypedDst[i] = bld.getSSA();
1914 untypedDst[0] = bld.getSSA();
1916 for (int i = 0; i < 4; i++) {
1917 typedDst[i] = su->getDef(i);
1920 // Set the untyped dsts as the su's destinations
1921 for (int i = 0; i < 4; i++)
1922 su->setDef(i, untypedDst[i]);
1924 bld.setPosition(su, true);
1926 // Unpack each component into the typed dsts
1928 for (int i = 0; i < 4; bits += format->bits[i], i++) {
1931 if (i >= format->components) {
1932 if (format->type == FLOAT ||
1933 format->type == UNORM ||
1934 format->type == SNORM)
1935 bld.loadImm(typedDst[i], i == 3 ? 1.0f : 0.0f);
1937 bld.loadImm(typedDst[i], i == 3 ? 1 : 0);
1941 // Get just that component's data into the relevant place
1942 if (format->bits[i] == 32)
1943 bld.mkMov(typedDst[i], untypedDst[i]);
1944 else if (format->bits[i] == 16)
1945 bld.mkCvt(OP_CVT, getDestType(format->type), typedDst[i],
1946 getSrcType(format, i), untypedDst[i / 2])
1947 ->subOp = (i & 1) << (format->type == FLOAT ? 0 : 1);
1948 else if (format->bits[i] == 8)
1949 bld.mkCvt(OP_CVT, getDestType(format->type), typedDst[i],
1950 getSrcType(format, i), untypedDst[0])->subOp = i;
1952 bld.mkOp2(OP_EXTBF, TYPE_U32, typedDst[i], untypedDst[bits / 32],
1953 bld.mkImm((bits % 32) | (format->bits[i] << 8)));
1954 if (format->type == UNORM || format->type == SNORM)
1955 bld.mkCvt(OP_CVT, TYPE_F32, typedDst[i], getSrcType(format, i), typedDst[i]);
1958 // Normalize / convert as necessary
1959 if (format->type == UNORM)
1960 bld.mkOp2(OP_MUL, TYPE_F32, typedDst[i], typedDst[i], bld.loadImm(NULL, 1.0f / ((1 << format->bits[i]) - 1)));
1961 else if (format->type == SNORM)
1962 bld.mkOp2(OP_MUL, TYPE_F32, typedDst[i], typedDst[i], bld.loadImm(NULL, 1.0f / ((1 << (format->bits[i] - 1)) - 1)));
1963 else if (format->type == FLOAT && format->bits[i] < 16) {
1964 bld.mkOp2(OP_SHL, TYPE_U32, typedDst[i], typedDst[i], bld.loadImm(NULL, 15 - format->bits[i]));
1965 bld.mkCvt(OP_CVT, TYPE_F32, typedDst[i], TYPE_F16, typedDst[i]);
1971 NVC0LoweringPass::handleSurfaceOpNVE4(TexInstruction *su)
1973 processSurfaceCoordsNVE4(su);
1975 if (su->op == OP_SULDP)
1976 convertSurfaceFormat(su);
1978 if (su->op == OP_SUREDB || su->op == OP_SUREDP) {
1979 Value *pred = su->getSrc(2);
1980 CondCode cc = CC_NOT_P;
1981 if (su->getPredicate()) {
1982 pred = bld.getScratch(1, FILE_PREDICATE);
1984 if (cc == CC_NOT_P) {
1985 bld.mkOp2(OP_OR, TYPE_U8, pred, su->getPredicate(), su->getSrc(2));
1987 bld.mkOp2(OP_AND, TYPE_U8, pred, su->getPredicate(), su->getSrc(2));
1988 pred->getInsn()->src(1).mod = Modifier(NV50_IR_MOD_NOT);
1991 Instruction *red = bld.mkOp(OP_ATOM, su->dType, bld.getSSA());
1992 red->subOp = su->subOp;
1994 gMemBase = bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, TYPE_U32, 0);
1995 red->setSrc(0, gMemBase);
1996 red->setSrc(1, su->getSrc(3));
1997 if (su->subOp == NV50_IR_SUBOP_ATOM_CAS)
1998 red->setSrc(2, su->getSrc(4));
1999 red->setIndirect(0, 0, su->getSrc(0));
2001 // make sure to initialize dst value when the atomic operation is not
2003 Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));
2005 assert(cc == CC_NOT_P);
2006 red->setPredicate(cc, pred);
2007 mov->setPredicate(CC_P, pred);
2009 bld.mkOp2(OP_UNION, TYPE_U32, su->getDef(0),
2010 red->getDef(0), mov->getDef(0));
2012 delete_Instruction(bld.getProgram(), su);
2013 handleCasExch(red, true);
2016 if (su->op == OP_SUSTB || su->op == OP_SUSTP)
2017 su->sType = (su->tex.target == TEX_TARGET_BUFFER) ? TYPE_U32 : TYPE_U8;
2021 NVC0LoweringPass::processSurfaceCoordsNVC0(TexInstruction *su)
2023 const int idx = su->tex.r;
2024 const int dim = su->tex.target.getDim();
2025 const int arg = dim + (su->tex.target.isArray() || su->tex.target.isCube());
2026 const uint16_t base = idx * NVE4_SU_INFO__STRIDE;
2028 Value *zero = bld.mkImm(0);
2033 if (su->tex.rIndirectSrc >= 0) {
2034 ind = su->getIndirectR();
2035 if (su->tex.r > 0) {
2036 ind = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), ind,
2037 bld.loadImm(NULL, su->tex.r));
2039 ind = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ind, bld.mkImm(7));
2040 ind = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), ind, bld.mkImm(6));
2043 // get surface coordinates
2044 for (c = 0; c < arg; ++c)
2045 src[c] = su->getSrc(c);
2049 // calculate pixel offset
2050 if (su->op == OP_SULDP || su->op == OP_SUREDP) {
2051 v = loadSuInfo32(ind, base + NVE4_SU_INFO_BSIZE);
2052 su->setSrc(0, bld.mkOp2v(OP_MUL, TYPE_U32, bld.getSSA(), src[0], v));
2055 // add array layer offset
2056 if (su->tex.target.isArray() || su->tex.target.isCube()) {
2057 v = loadSuInfo32(ind, base + NVE4_SU_INFO_ARRAY);
2059 su->setSrc(2, bld.mkOp2v(OP_MUL, TYPE_U32, bld.getSSA(), src[2], v));
2062 // prevent read fault when the image is not actually bound
2063 CmpInstruction *pred =
2064 bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
2065 TYPE_U32, bld.mkImm(0),
2066 loadSuInfo32(ind, base + NVE4_SU_INFO_ADDR));
2067 if (su->op != OP_SUSTP && su->tex.format) {
2068 const TexInstruction::ImgFormatDesc *format = su->tex.format;
2069 int blockwidth = format->bits[0] + format->bits[1] +
2070 format->bits[2] + format->bits[3];
2072 assert(format->components != 0);
2073 // make sure that the format doesn't mismatch when it's not FMT_NONE
2074 bld.mkCmp(OP_SET_OR, CC_NE, TYPE_U32, pred->getDef(0),
2075 TYPE_U32, bld.loadImm(NULL, blockwidth / 8),
2076 loadSuInfo32(ind, base + NVE4_SU_INFO_BSIZE),
2079 su->setPredicate(CC_NOT_P, pred->getDef(0));
2083 NVC0LoweringPass::handleSurfaceOpNVC0(TexInstruction *su)
2085 if (su->tex.target == TEX_TARGET_1D_ARRAY) {
2086 /* As 1d arrays also need 3 coordinates, switching to TEX_TARGET_2D_ARRAY
2087 * will simplify the lowering pass and the texture constraints. */
2088 su->moveSources(1, 1);
2089 su->setSrc(1, bld.loadImm(NULL, 0));
2090 su->tex.target = TEX_TARGET_2D_ARRAY;
2093 processSurfaceCoordsNVC0(su);
2095 if (su->op == OP_SULDP)
2096 convertSurfaceFormat(su);
2098 if (su->op == OP_SUREDB || su->op == OP_SUREDP) {
2099 const int dim = su->tex.target.getDim();
2100 const int arg = dim + (su->tex.target.isArray() || su->tex.target.isCube());
2101 LValue *addr = bld.getSSA(8);
2102 Value *def = su->getDef(0);
2106 // Set the destination to the address
2107 su->dType = TYPE_U64;
2108 su->setDef(0, addr);
2109 su->setDef(1, su->getPredicate());
2111 bld.setPosition(su, true);
2113 // Perform the atomic op
2114 Instruction *red = bld.mkOp(OP_ATOM, su->sType, bld.getSSA());
2115 red->subOp = su->subOp;
2116 red->setSrc(0, bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, su->sType, 0));
2117 red->setSrc(1, su->getSrc(arg));
2118 if (red->subOp == NV50_IR_SUBOP_ATOM_CAS)
2119 red->setSrc(2, su->getSrc(arg + 1));
2120 red->setIndirect(0, 0, addr);
2122 // make sure to initialize dst value when the atomic operation is not
2124 Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));
2126 assert(su->cc == CC_NOT_P);
2127 red->setPredicate(su->cc, su->getPredicate());
2128 mov->setPredicate(CC_P, su->getPredicate());
2130 bld.mkOp2(OP_UNION, TYPE_U32, def, red->getDef(0), mov->getDef(0));
2132 handleCasExch(red, false);
2137 NVC0LoweringPass::handleWRSV(Instruction *i)
2143 // must replace, $sreg are not writeable
2144 addr = targ->getSVAddress(FILE_SHADER_OUTPUT, i->getSrc(0)->asSym());
2147 sym = bld.mkSymbol(FILE_SHADER_OUTPUT, 0, i->sType, addr);
2149 st = bld.mkStore(OP_EXPORT, i->dType, sym, i->getIndirect(0, 0),
2151 st->perPatch = i->perPatch;
2153 bld.getBB()->remove(i);
2158 NVC0LoweringPass::handleLDST(Instruction *i)
2160 if (i->src(0).getFile() == FILE_SHADER_INPUT) {
2161 if (prog->getType() == Program::TYPE_COMPUTE) {
2162 i->getSrc(0)->reg.file = FILE_MEMORY_CONST;
2163 i->getSrc(0)->reg.fileIndex = 0;
2165 if (prog->getType() == Program::TYPE_GEOMETRY &&
2166 i->src(0).isIndirect(0)) {
2167 // XXX: this assumes vec4 units
2168 Value *ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
2169 i->getIndirect(0, 0), bld.mkImm(4));
2170 i->setIndirect(0, 0, ptr);
2174 assert(prog->getType() != Program::TYPE_FRAGMENT); // INTERP
2176 } else if (i->src(0).getFile() == FILE_MEMORY_CONST) {
2177 if (targ->getChipset() >= NVISA_GK104_CHIPSET &&
2178 prog->getType() == Program::TYPE_COMPUTE) {
2179 // The launch descriptor only allows to set up 8 CBs, but OpenGL
2180 // requires at least 12 UBOs. To bypass this limitation, we store the
2181 // addrs into the driver constbuf and we directly load from the global
2183 int8_t fileIndex = i->getSrc(0)->reg.fileIndex - 1;
2184 Value *ind = i->getIndirect(0, 1);
2186 // TODO: clamp the offset to the maximum number of const buf.
2187 if (i->src(0).isIndirect(1)) {
2188 Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
2189 Value *ptr = loadUboInfo64(ind, fileIndex * 16);
2190 Value *length = loadUboLength32(ind, fileIndex * 16);
2191 Value *pred = new_LValue(func, FILE_PREDICATE);
2192 if (i->src(0).isIndirect(0)) {
2193 bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
2194 bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
2196 i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
2197 i->setIndirect(0, 1, NULL);
2198 i->setIndirect(0, 0, ptr);
2199 bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
2200 i->setPredicate(CC_NOT_P, pred);
2201 if (i->defExists(0)) {
2202 bld.mkMov(i->getDef(0), bld.mkImm(0));
2204 } else if (fileIndex >= 0) {
2205 Value *ptr = loadUboInfo64(ind, fileIndex * 16);
2206 if (i->src(0).isIndirect(0)) {
2207 bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
2209 i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
2210 i->setIndirect(0, 1, NULL);
2211 i->setIndirect(0, 0, ptr);
2213 } else if (i->src(0).isIndirect(1)) {
2215 if (i->src(0).isIndirect(0))
2216 ptr = bld.mkOp3v(OP_INSBF, TYPE_U32, bld.getSSA(),
2217 i->getIndirect(0, 1), bld.mkImm(0x1010),
2218 i->getIndirect(0, 0));
2220 ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
2221 i->getIndirect(0, 1), bld.mkImm(16));
2222 i->setIndirect(0, 1, NULL);
2223 i->setIndirect(0, 0, ptr);
2224 i->subOp = NV50_IR_SUBOP_LDC_IS;
2226 } else if (i->src(0).getFile() == FILE_SHADER_OUTPUT) {
2227 assert(prog->getType() == Program::TYPE_TESSELLATION_CONTROL);
2229 } else if (i->src(0).getFile() == FILE_MEMORY_BUFFER) {
2230 Value *ind = i->getIndirect(0, 1);
2231 Value *ptr = loadBufInfo64(ind, i->getSrc(0)->reg.fileIndex * 16);
2232 // XXX come up with a way not to do this for EVERY little access but
2233 // rather to batch these up somehow. Unfortunately we've lost the
2234 // information about the field width by the time we get here.
2235 Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
2236 Value *length = loadBufLength32(ind, i->getSrc(0)->reg.fileIndex * 16);
2237 Value *pred = new_LValue(func, FILE_PREDICATE);
2238 if (i->src(0).isIndirect(0)) {
2239 bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
2240 bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
2242 i->setIndirect(0, 1, NULL);
2243 i->setIndirect(0, 0, ptr);
2244 i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
2245 bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
2246 i->setPredicate(CC_NOT_P, pred);
2247 if (i->defExists(0)) {
2248 Value *zero, *dst = i->getDef(0);
2249 i->setDef(0, bld.getSSA());
2251 bld.setPosition(i, true);
2252 bld.mkMov((zero = bld.getSSA()), bld.mkImm(0))
2253 ->setPredicate(CC_P, pred);
2254 bld.mkOp2(OP_UNION, TYPE_U32, dst, i->getDef(0), zero);
2260 NVC0LoweringPass::readTessCoord(LValue *dst, int c)
2262 Value *laneid = bld.getSSA();
2265 bld.mkOp1(OP_RDSV, TYPE_U32, laneid, bld.mkSysVal(SV_LANEID, 0));
2276 if (prog->driver->prop.tp.domain != PIPE_PRIM_TRIANGLES) {
2277 bld.mkMov(dst, bld.loadImm(NULL, 0));
2284 bld.mkFetch(x, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f0, NULL, laneid);
2286 bld.mkFetch(y, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f4, NULL, laneid);
2289 bld.mkOp2(OP_ADD, TYPE_F32, dst, x, y);
2290 bld.mkOp2(OP_SUB, TYPE_F32, dst, bld.loadImm(NULL, 1.0f), dst);
2295 NVC0LoweringPass::handleRDSV(Instruction *i)
2297 Symbol *sym = i->getSrc(0)->asSym();
2298 const SVSemantic sv = sym->reg.data.sv.sv;
2301 uint32_t addr = targ->getSVAddress(FILE_SHADER_INPUT, sym);
2303 if (addr >= 0x400) {
2305 if (sym->reg.data.sv.index == 3) {
2306 // TGSI backend may use 4th component of TID,NTID,CTAID,NCTAID
2308 i->setSrc(0, bld.mkImm((sv == SV_NTID || sv == SV_NCTAID) ? 1 : 0));
2310 if (sv == SV_VERTEX_COUNT) {
2311 bld.setPosition(i, true);
2312 bld.mkOp2(OP_EXTBF, TYPE_U32, i->getDef(0), i->getDef(0), bld.mkImm(0x808));
2319 assert(prog->getType() == Program::TYPE_FRAGMENT);
2320 if (i->srcExists(1)) {
2321 // Pass offset through to the interpolation logic
2322 ld = bld.mkInterp(NV50_IR_INTERP_LINEAR | NV50_IR_INTERP_OFFSET,
2323 i->getDef(0), addr, NULL);
2324 ld->setSrc(1, i->getSrc(1));
2326 bld.mkInterp(NV50_IR_INTERP_LINEAR, i->getDef(0), addr, NULL);
2331 Value *face = i->getDef(0);
2332 bld.mkInterp(NV50_IR_INTERP_FLAT, face, addr, NULL);
2333 if (i->dType == TYPE_F32) {
2334 bld.mkOp2(OP_OR, TYPE_U32, face, face, bld.mkImm(0x00000001));
2335 bld.mkOp1(OP_NEG, TYPE_S32, face, face);
2336 bld.mkCvt(OP_CVT, TYPE_F32, face, TYPE_S32, face);
2341 assert(prog->getType() == Program::TYPE_TESSELLATION_EVAL);
2342 readTessCoord(i->getDef(0)->asLValue(), i->getSrc(0)->reg.data.sv.index);
2347 assert(targ->getChipset() >= NVISA_GK104_CHIPSET); // mov $sreg otherwise
2348 if (sym->reg.data.sv.index == 3) {
2350 i->setSrc(0, bld.mkImm(sv == SV_GRIDID ? 0 : 1));
2353 addr += prog->driver->prop.cp.gridInfoBase;
2354 bld.mkLoad(TYPE_U32, i->getDef(0),
2355 bld.mkSymbol(FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
2356 TYPE_U32, addr), NULL);
2358 case SV_SAMPLE_INDEX:
2359 // TODO: Properly pass source as an address in the PIX address space
2360 // (which can be of the form [r0+offset]). But this is currently
2362 ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
2363 ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
2365 case SV_SAMPLE_POS: {
2366 Value *off = new_LValue(func, FILE_GPR);
2367 ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
2368 ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
2369 bld.mkOp2(OP_SHL, TYPE_U32, off, i->getDef(0), bld.mkImm(3));
2370 bld.mkLoad(TYPE_F32,
2373 FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
2374 TYPE_U32, prog->driver->io.sampleInfoBase +
2375 4 * sym->reg.data.sv.index),
2379 case SV_SAMPLE_MASK: {
2380 ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
2381 ld->subOp = NV50_IR_SUBOP_PIXLD_COVMASK;
2382 Instruction *sampleid =
2383 bld.mkOp1(OP_PIXLD, TYPE_U32, bld.getSSA(), bld.mkImm(0));
2384 sampleid->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
2386 bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ld->getDef(0),
2387 bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
2388 bld.loadImm(NULL, 1), sampleid->getDef(0)));
2389 if (prog->driver->prop.fp.persampleInvocation) {
2390 bld.mkMov(i->getDef(0), masked);
2392 bld.mkOp3(OP_SELP, TYPE_U32, i->getDef(0), ld->getDef(0), masked,
2399 case SV_BASEINSTANCE:
2401 ld = bld.mkLoad(TYPE_U32, i->getDef(0),
2402 bld.mkSymbol(FILE_MEMORY_CONST,
2403 prog->driver->io.auxCBSlot,
2405 prog->driver->io.drawInfoBase +
2406 4 * (sv - SV_BASEVERTEX)),
2410 if (prog->getType() == Program::TYPE_TESSELLATION_EVAL && !i->perPatch)
2411 vtx = bld.mkOp1v(OP_PFETCH, TYPE_U32, bld.getSSA(), bld.mkImm(0));
2412 ld = bld.mkFetch(i->getDef(0), i->dType,
2413 FILE_SHADER_INPUT, addr, i->getIndirect(0, 0), vtx);
2414 ld->perPatch = i->perPatch;
2417 bld.getBB()->remove(i);
2422 NVC0LoweringPass::handleDIV(Instruction *i)
2424 if (!isFloatType(i->dType))
2426 bld.setPosition(i, false);
2427 Instruction *rcp = bld.mkOp1(OP_RCP, i->dType, bld.getSSA(typeSizeof(i->dType)), i->getSrc(1));
2429 i->setSrc(1, rcp->getDef(0));
2434 NVC0LoweringPass::handleMOD(Instruction *i)
2436 if (!isFloatType(i->dType))
2438 LValue *value = bld.getScratch(typeSizeof(i->dType));
2439 bld.mkOp1(OP_RCP, i->dType, value, i->getSrc(1));
2440 bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(0), value);
2441 bld.mkOp1(OP_TRUNC, i->dType, value, value);
2442 bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(1), value);
2444 i->setSrc(1, value);
2449 NVC0LoweringPass::handleSQRT(Instruction *i)
2451 if (i->dType == TYPE_F64) {
2452 Value *pred = bld.getSSA(1, FILE_PREDICATE);
2453 Value *zero = bld.loadImm(NULL, 0.0);
2454 Value *dst = bld.getSSA(8);
2455 bld.mkOp1(OP_RSQ, i->dType, dst, i->getSrc(0));
2456 bld.mkCmp(OP_SET, CC_LE, i->dType, pred, i->dType, i->getSrc(0), zero);
2457 bld.mkOp3(OP_SELP, TYPE_U64, dst, zero, dst, pred);
2460 // TODO: Handle this properly with a library function
2462 bld.setPosition(i, true);
2464 bld.mkOp1(OP_RCP, i->dType, i->getDef(0), i->getDef(0));
2471 NVC0LoweringPass::handlePOW(Instruction *i)
2473 LValue *val = bld.getScratch();
2475 bld.mkOp1(OP_LG2, TYPE_F32, val, i->getSrc(0));
2476 bld.mkOp2(OP_MUL, TYPE_F32, val, i->getSrc(1), val)->dnz = 1;
2477 bld.mkOp1(OP_PREEX2, TYPE_F32, val, val);
2487 NVC0LoweringPass::handleEXPORT(Instruction *i)
2489 if (prog->getType() == Program::TYPE_FRAGMENT) {
2490 int id = i->getSrc(0)->reg.data.offset / 4;
2492 if (i->src(0).isIndirect(0)) // TODO, ugly
2495 i->subOp = NV50_IR_SUBOP_MOV_FINAL;
2496 i->src(0).set(i->src(1));
2498 i->setDef(0, new_LValue(func, FILE_GPR));
2499 i->getDef(0)->reg.data.id = id;
2501 prog->maxGPR = MAX2(prog->maxGPR, id);
2503 if (prog->getType() == Program::TYPE_GEOMETRY) {
2504 i->setIndirect(0, 1, gpEmitAddress);
2510 NVC0LoweringPass::handleOUT(Instruction *i)
2512 Instruction *prev = i->prev;
2513 ImmediateValue stream, prevStream;
2515 // Only merge if the stream ids match. Also, note that the previous
2516 // instruction would have already been lowered, so we take arg1 from it.
2517 if (i->op == OP_RESTART && prev && prev->op == OP_EMIT &&
2518 i->src(0).getImmediate(stream) &&
2519 prev->src(1).getImmediate(prevStream) &&
2520 stream.reg.data.u32 == prevStream.reg.data.u32) {
2521 i->prev->subOp = NV50_IR_SUBOP_EMIT_RESTART;
2522 delete_Instruction(prog, i);
2524 assert(gpEmitAddress);
2525 i->setDef(0, gpEmitAddress);
2526 i->setSrc(1, i->getSrc(0));
2527 i->setSrc(0, gpEmitAddress);
2532 // Generate a binary predicate if an instruction is predicated by
2533 // e.g. an f32 value.
2535 NVC0LoweringPass::checkPredicate(Instruction *insn)
2537 Value *pred = insn->getPredicate();
2540 if (!pred || pred->reg.file == FILE_PREDICATE)
2542 pdst = new_LValue(func, FILE_PREDICATE);
2544 // CAUTION: don't use pdst->getInsn, the definition might not be unique,
2545 // delay turning PSET(FSET(x,y),0) into PSET(x,y) to a later pass
2547 bld.mkCmp(OP_SET, CC_NEU, insn->dType, pdst, insn->dType, bld.mkImm(0), pred);
2549 insn->setPredicate(insn->cc, pdst);
2553 // - add quadop dance for texturing
2554 // - put FP outputs in GPRs
2555 // - convert instruction sequences
2558 NVC0LoweringPass::visit(Instruction *i)
2561 bld.setPosition(i, false);
2563 if (i->cc != CC_ALWAYS)
2572 return handleTEX(i->asTex());
2574 return handleTXD(i->asTex());
2576 return handleTXLQ(i->asTex());
2578 return handleTXQ(i->asTex());
2580 bld.mkOp1(OP_PREEX2, TYPE_F32, i->getDef(0), i->getSrc(0));
2581 i->setSrc(0, i->getDef(0));
2584 return handlePOW(i);
2586 return handleDIV(i);
2588 return handleMOD(i);
2590 return handleSQRT(i);
2592 ret = handleEXPORT(i);
2596 return handleOUT(i);
2598 return handleRDSV(i);
2600 return handleWRSV(i);
2607 const bool cctl = i->src(0).getFile() == FILE_MEMORY_BUFFER;
2609 handleCasExch(i, cctl);
2618 if (targ->getChipset() >= NVISA_GK104_CHIPSET)
2619 handleSurfaceOpNVE4(i->asTex());
2621 handleSurfaceOpNVC0(i->asTex());
2624 handleSUQ(i->asTex());
2633 /* Kepler+ has a special opcode to compute a new base address to be used
2634 * for indirect loads.
2636 if (targ->getChipset() >= NVISA_GK104_CHIPSET && !i->perPatch &&
2637 (i->op == OP_VFETCH || i->op == OP_EXPORT) && i->src(0).isIndirect(0)) {
2638 Instruction *afetch = bld.mkOp1(OP_AFETCH, TYPE_U32, bld.getSSA(),
2639 cloneShallow(func, i->getSrc(0)));
2640 afetch->setIndirect(0, 0, i->getIndirect(0, 0));
2641 i->src(0).get()->reg.data.offset = 0;
2642 i->setIndirect(0, 0, afetch->getDef(0));
2649 TargetNVC0::runLegalizePass(Program *prog, CGStage stage) const
2651 if (stage == CG_STAGE_PRE_SSA) {
2652 NVC0LoweringPass pass(prog);
2653 return pass.run(prog, false, true);
2655 if (stage == CG_STAGE_POST_RA) {
2656 NVC0LegalizePostRA pass(prog);
2657 return pass.run(prog, false, true);
2659 if (stage == CG_STAGE_SSA) {
2660 NVC0LegalizeSSA pass;
2661 return pass.run(prog, false, true);
2666 } // namespace nv50_ir