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
)
156 needTexBar(prog
->getTarget()->getChipset() >= 0xe0)
161 NVC0LegalizePostRA::insnDominatedBy(const Instruction
*later
,
162 const Instruction
*early
) const
164 if (early
->bb
== later
->bb
)
165 return early
->serial
< later
->serial
;
166 return later
->bb
->dominatedBy(early
->bb
);
170 NVC0LegalizePostRA::addTexUse(std::list
<TexUse
> &uses
,
171 Instruction
*usei
, const Instruction
*texi
)
174 for (std::list
<TexUse
>::iterator it
= uses
.begin();
176 if (insnDominatedBy(usei
, it
->insn
)) {
180 if (insnDominatedBy(it
->insn
, usei
))
186 uses
.push_back(TexUse(usei
, texi
));
189 // While it might be tempting to use the an algorithm that just looks at tex
190 // uses, not all texture results are guaranteed to be used on all paths. In
191 // the case where along some control flow path a texture result is never used,
192 // we might reuse that register for something else, creating a
193 // write-after-write hazard. So we have to manually look through all
194 // instructions looking for ones that reference the registers in question.
196 NVC0LegalizePostRA::findFirstUses(
197 Instruction
*texi
, std::list
<TexUse
> &uses
)
199 int minGPR
= texi
->def(0).rep()->reg
.data
.id
;
200 int maxGPR
= minGPR
+ texi
->def(0).rep()->reg
.size
/ 4 - 1;
202 unordered_set
<const BasicBlock
*> visited
;
203 findFirstUsesBB(minGPR
, maxGPR
, texi
->next
, texi
, uses
, visited
);
207 NVC0LegalizePostRA::findFirstUsesBB(
208 int minGPR
, int maxGPR
, Instruction
*start
,
209 const Instruction
*texi
, std::list
<TexUse
> &uses
,
210 unordered_set
<const BasicBlock
*> &visited
)
212 const BasicBlock
*bb
= start
->bb
;
214 // We don't process the whole bb the first time around. This is correct,
215 // however we might be in a loop and hit this BB again, and need to process
216 // the full thing. So only mark a bb as visited if we processed it from the
218 if (start
== bb
->getEntry()) {
219 if (visited
.find(bb
) != visited
.end())
224 for (Instruction
*insn
= start
; insn
!= bb
->getExit(); insn
= insn
->next
) {
228 for (int d
= 0; insn
->defExists(d
); ++d
) {
229 if (insn
->def(d
).getFile() != FILE_GPR
||
230 insn
->def(d
).rep()->reg
.data
.id
< minGPR
||
231 insn
->def(d
).rep()->reg
.data
.id
> maxGPR
)
233 addTexUse(uses
, insn
, texi
);
237 for (int s
= 0; insn
->srcExists(s
); ++s
) {
238 if (insn
->src(s
).getFile() != FILE_GPR
||
239 insn
->src(s
).rep()->reg
.data
.id
< minGPR
||
240 insn
->src(s
).rep()->reg
.data
.id
> maxGPR
)
242 addTexUse(uses
, insn
, texi
);
247 for (Graph::EdgeIterator ei
= bb
->cfg
.outgoing(); !ei
.end(); ei
.next()) {
248 findFirstUsesBB(minGPR
, maxGPR
, BasicBlock::get(ei
.getNode())->getEntry(),
249 texi
, uses
, visited
);
254 // This pass is a bit long and ugly and can probably be optimized.
256 // 1. obtain a list of TEXes and their outputs' first use(s)
257 // 2. calculate the barrier level of each first use (minimal number of TEXes,
258 // over all paths, between the TEX and the use in question)
259 // 3. for each barrier, if all paths from the source TEX to that barrier
260 // contain a barrier of lesser level, it can be culled
262 NVC0LegalizePostRA::insertTextureBarriers(Function
*fn
)
264 std::list
<TexUse
> *uses
;
265 std::vector
<Instruction
*> texes
;
266 std::vector
<int> bbFirstTex
;
267 std::vector
<int> bbFirstUse
;
268 std::vector
<int> texCounts
;
269 std::vector
<TexUse
> useVec
;
272 fn
->orderInstructions(insns
);
274 texCounts
.resize(fn
->allBBlocks
.getSize(), 0);
275 bbFirstTex
.resize(fn
->allBBlocks
.getSize(), insns
.getSize());
276 bbFirstUse
.resize(fn
->allBBlocks
.getSize(), insns
.getSize());
278 // tag BB CFG nodes by their id for later
279 for (ArrayList::Iterator i
= fn
->allBBlocks
.iterator(); !i
.end(); i
.next()) {
280 BasicBlock
*bb
= reinterpret_cast<BasicBlock
*>(i
.get());
282 bb
->cfg
.tag
= bb
->getId();
285 // gather the first uses for each TEX
286 for (int i
= 0; i
< insns
.getSize(); ++i
) {
287 Instruction
*tex
= reinterpret_cast<Instruction
*>(insns
.get(i
));
288 if (isTextureOp(tex
->op
)) {
289 texes
.push_back(tex
);
290 if (!texCounts
.at(tex
->bb
->getId()))
291 bbFirstTex
[tex
->bb
->getId()] = texes
.size() - 1;
292 texCounts
[tex
->bb
->getId()]++;
298 uses
= new std::list
<TexUse
>[texes
.size()];
301 for (size_t i
= 0; i
< texes
.size(); ++i
) {
302 findFirstUses(texes
[i
], uses
[i
]);
305 // determine the barrier level at each use
306 for (size_t i
= 0; i
< texes
.size(); ++i
) {
307 for (std::list
<TexUse
>::iterator u
= uses
[i
].begin(); u
!= uses
[i
].end();
309 BasicBlock
*tb
= texes
[i
]->bb
;
310 BasicBlock
*ub
= u
->insn
->bb
;
313 for (size_t j
= i
+ 1; j
< texes
.size() &&
314 texes
[j
]->bb
== tb
&& texes
[j
]->serial
< u
->insn
->serial
;
318 u
->level
= fn
->cfg
.findLightestPathWeight(&tb
->cfg
,
319 &ub
->cfg
, texCounts
);
321 WARN("Failed to find path TEX -> TEXBAR\n");
325 // this counted all TEXes in the origin block, correct that
326 u
->level
-= i
- bbFirstTex
.at(tb
->getId()) + 1 /* this TEX */;
327 // and did not count the TEXes in the destination block, add those
328 for (size_t j
= bbFirstTex
.at(ub
->getId()); j
< texes
.size() &&
329 texes
[j
]->bb
== ub
&& texes
[j
]->serial
< u
->insn
->serial
;
333 assert(u
->level
>= 0);
334 useVec
.push_back(*u
);
339 // insert the barriers
340 for (size_t i
= 0; i
< useVec
.size(); ++i
) {
341 Instruction
*prev
= useVec
[i
].insn
->prev
;
342 if (useVec
[i
].level
< 0)
344 if (prev
&& prev
->op
== OP_TEXBAR
) {
345 if (prev
->subOp
> useVec
[i
].level
)
346 prev
->subOp
= useVec
[i
].level
;
347 prev
->setSrc(prev
->srcCount(), useVec
[i
].tex
->getDef(0));
349 Instruction
*bar
= new_Instruction(func
, OP_TEXBAR
, TYPE_NONE
);
351 bar
->subOp
= useVec
[i
].level
;
352 // make use explicit to ease latency calculation
353 bar
->setSrc(bar
->srcCount(), useVec
[i
].tex
->getDef(0));
354 useVec
[i
].insn
->bb
->insertBefore(useVec
[i
].insn
, bar
);
358 if (fn
->getProgram()->optLevel
< 3)
361 std::vector
<Limits
> limitT
, limitB
, limitS
; // entry, exit, single
363 limitT
.resize(fn
->allBBlocks
.getSize(), Limits(0, 0));
364 limitB
.resize(fn
->allBBlocks
.getSize(), Limits(0, 0));
365 limitS
.resize(fn
->allBBlocks
.getSize());
367 // cull unneeded barriers (should do that earlier, but for simplicity)
368 IteratorRef bi
= fn
->cfg
.iteratorCFG();
369 // first calculate min/max outstanding TEXes for each BB
370 for (bi
->reset(); !bi
->end(); bi
->next()) {
371 Graph::Node
*n
= reinterpret_cast<Graph::Node
*>(bi
->get());
372 BasicBlock
*bb
= BasicBlock::get(n
);
374 int max
= std::numeric_limits
<int>::max();
375 for (Instruction
*i
= bb
->getFirst(); i
; i
= i
->next
) {
376 if (isTextureOp(i
->op
)) {
378 if (max
< std::numeric_limits
<int>::max())
381 if (i
->op
== OP_TEXBAR
) {
382 min
= MIN2(min
, i
->subOp
);
383 max
= MIN2(max
, i
->subOp
);
386 // limits when looking at an isolated block
387 limitS
[bb
->getId()].min
= min
;
388 limitS
[bb
->getId()].max
= max
;
390 // propagate the min/max values
391 for (unsigned int l
= 0; l
<= fn
->loopNestingBound
; ++l
) {
392 for (bi
->reset(); !bi
->end(); bi
->next()) {
393 Graph::Node
*n
= reinterpret_cast<Graph::Node
*>(bi
->get());
394 BasicBlock
*bb
= BasicBlock::get(n
);
395 const int bbId
= bb
->getId();
396 for (Graph::EdgeIterator ei
= n
->incident(); !ei
.end(); ei
.next()) {
397 BasicBlock
*in
= BasicBlock::get(ei
.getNode());
398 const int inId
= in
->getId();
399 limitT
[bbId
].min
= MAX2(limitT
[bbId
].min
, limitB
[inId
].min
);
400 limitT
[bbId
].max
= MAX2(limitT
[bbId
].max
, limitB
[inId
].max
);
402 // I just hope this is correct ...
403 if (limitS
[bbId
].max
== std::numeric_limits
<int>::max()) {
405 limitB
[bbId
].min
= limitT
[bbId
].min
+ limitS
[bbId
].min
;
406 limitB
[bbId
].max
= limitT
[bbId
].max
+ limitS
[bbId
].min
;
408 // block contained a barrier
409 limitB
[bbId
].min
= MIN2(limitS
[bbId
].max
,
410 limitT
[bbId
].min
+ limitS
[bbId
].min
);
411 limitB
[bbId
].max
= MIN2(limitS
[bbId
].max
,
412 limitT
[bbId
].max
+ limitS
[bbId
].min
);
416 // finally delete unnecessary barriers
417 for (bi
->reset(); !bi
->end(); bi
->next()) {
418 Graph::Node
*n
= reinterpret_cast<Graph::Node
*>(bi
->get());
419 BasicBlock
*bb
= BasicBlock::get(n
);
420 Instruction
*prev
= NULL
;
422 int max
= limitT
[bb
->getId()].max
;
423 for (Instruction
*i
= bb
->getFirst(); i
; i
= next
) {
425 if (i
->op
== OP_TEXBAR
) {
426 if (i
->subOp
>= max
) {
427 delete_Instruction(prog
, i
);
431 if (prev
&& prev
->op
== OP_TEXBAR
&& prev
->subOp
>= max
) {
432 delete_Instruction(prog
, prev
);
437 if (isTextureOp(i
->op
)) {
440 if (i
&& !i
->isNop())
448 NVC0LegalizePostRA::visit(Function
*fn
)
451 insertTextureBarriers(fn
);
453 rZero
= new_LValue(fn
, FILE_GPR
);
454 carry
= new_LValue(fn
, FILE_FLAGS
);
456 rZero
->reg
.data
.id
= prog
->getTarget()->getFileSize(FILE_GPR
);
457 carry
->reg
.data
.id
= 0;
463 NVC0LegalizePostRA::replaceZero(Instruction
*i
)
465 for (int s
= 0; i
->srcExists(s
); ++s
) {
466 if (s
== 2 && i
->op
== OP_SUCLAMP
)
468 ImmediateValue
*imm
= i
->getSrc(s
)->asImm();
469 if (imm
&& imm
->reg
.data
.u64
== 0)
474 // replace CONT with BRA for single unconditional continue
476 NVC0LegalizePostRA::tryReplaceContWithBra(BasicBlock
*bb
)
478 if (bb
->cfg
.incidentCount() != 2 || bb
->getEntry()->op
!= OP_PRECONT
)
480 Graph::EdgeIterator ei
= bb
->cfg
.incident();
481 if (ei
.getType() != Graph::Edge::BACK
)
483 if (ei
.getType() != Graph::Edge::BACK
)
485 BasicBlock
*contBB
= BasicBlock::get(ei
.getNode());
487 if (!contBB
->getExit() || contBB
->getExit()->op
!= OP_CONT
||
488 contBB
->getExit()->getPredicate())
490 contBB
->getExit()->op
= OP_BRA
;
491 bb
->remove(bb
->getEntry()); // delete PRECONT
494 assert(ei
.end() || ei
.getType() != Graph::Edge::BACK
);
498 // replace branches to join blocks with join ops
500 NVC0LegalizePostRA::propagateJoin(BasicBlock
*bb
)
502 if (bb
->getEntry()->op
!= OP_JOIN
|| bb
->getEntry()->asFlow()->limit
)
504 for (Graph::EdgeIterator ei
= bb
->cfg
.incident(); !ei
.end(); ei
.next()) {
505 BasicBlock
*in
= BasicBlock::get(ei
.getNode());
506 Instruction
*exit
= in
->getExit();
508 in
->insertTail(new FlowInstruction(func
, OP_JOIN
, bb
));
509 // there should always be a terminator instruction
510 WARN("inserted missing terminator in BB:%i\n", in
->getId());
512 if (exit
->op
== OP_BRA
) {
514 exit
->asFlow()->limit
= 1; // must-not-propagate marker
517 bb
->remove(bb
->getEntry());
521 NVC0LegalizePostRA::visit(BasicBlock
*bb
)
523 Instruction
*i
, *next
;
525 // remove pseudo operations and non-fixed no-ops, split 64 bit operations
526 for (i
= bb
->getFirst(); i
; i
= next
) {
528 if (i
->op
== OP_EMIT
|| i
->op
== OP_RESTART
) {
529 if (!i
->getDef(0)->refCount())
531 if (i
->src(0).getFile() == FILE_IMMEDIATE
)
532 i
->setSrc(0, rZero
); // initial value must be 0
538 if (i
->op
== OP_BAR
&& i
->subOp
== NV50_IR_SUBOP_BAR_SYNC
&&
539 prog
->getType() != Program::TYPE_COMPUTE
) {
540 // It seems like barriers are never required for tessellation since
541 // the warp size is 32, and there are always at most 32 tcs threads.
544 if (i
->op
== OP_LOAD
&& i
->subOp
== NV50_IR_SUBOP_LDC_IS
) {
545 int offset
= i
->src(0).get()->reg
.data
.offset
;
546 if (abs(offset
) > 0x10000)
547 i
->src(0).get()->reg
.fileIndex
+= offset
>> 16;
548 i
->src(0).get()->reg
.data
.offset
= (int)(short)offset
;
550 // TODO: Move this to before register allocation for operations that
551 // need the $c register !
552 if (typeSizeof(i
->dType
) == 8) {
554 hi
= BuildUtil::split64BitOpPostRA(func
, i
, rZero
, carry
);
559 if (i
->op
!= OP_MOV
&& i
->op
!= OP_PFETCH
)
566 if (!tryReplaceContWithBra(bb
))
572 NVC0LoweringPass::NVC0LoweringPass(Program
*prog
) : targ(prog
->getTarget())
574 bld
.setProgram(prog
);
579 NVC0LoweringPass::visit(Function
*fn
)
581 if (prog
->getType() == Program::TYPE_GEOMETRY
) {
582 assert(!strncmp(fn
->getName(), "MAIN", 4));
583 // TODO: when we generate actual functions pass this value along somehow
584 bld
.setPosition(BasicBlock::get(fn
->cfg
.getRoot()), false);
585 gpEmitAddress
= bld
.loadImm(NULL
, 0)->asLValue();
587 bld
.setPosition(BasicBlock::get(fn
->cfgExit
)->getExit(), false);
588 bld
.mkMovToReg(0, gpEmitAddress
);
595 NVC0LoweringPass::visit(BasicBlock
*bb
)
601 NVC0LoweringPass::loadTexHandle(Value
*ptr
, unsigned int slot
)
603 uint8_t b
= prog
->driver
->io
.resInfoCBSlot
;
604 uint32_t off
= prog
->driver
->io
.texBindBase
+ slot
* 4;
606 mkLoadv(TYPE_U32
, bld
.mkSymbol(FILE_MEMORY_CONST
, b
, TYPE_U32
, off
), ptr
);
609 // move array source to first slot, convert to u16, add indirections
611 NVC0LoweringPass::handleTEX(TexInstruction
*i
)
613 const int dim
= i
->tex
.target
.getDim() + i
->tex
.target
.isCube();
614 const int arg
= i
->tex
.target
.getArgCount();
615 const int lyr
= arg
- (i
->tex
.target
.isMS() ? 2 : 1);
616 const int chipset
= prog
->getTarget()->getChipset();
618 // Arguments to the TEX instruction are a little insane. Even though the
619 // encoding is identical between SM20 and SM30, the arguments mean
620 // different things between Fermi and Kepler+. A lot of arguments are
621 // optional based on flags passed to the instruction. This summarizes the
631 // - tg4: 8 bits each, either 2 (1 offset reg) or 8 (2 offset reg)
632 // - other: 4 bits each, single reg
636 // array (+ offsets for txd in upper 16 bits)
641 // offsets (same as fermi, except txd which takes it with array)
658 if (chipset
>= NVISA_GK104_CHIPSET
) {
659 if (i
->tex
.rIndirectSrc
>= 0 || i
->tex
.sIndirectSrc
>= 0) {
660 // XXX this ignores tsc, and assumes a 1:1 mapping
661 assert(i
->tex
.rIndirectSrc
>= 0);
662 Value
*hnd
= loadTexHandle(
663 bld
.mkOp2v(OP_SHL
, TYPE_U32
, bld
.getSSA(),
664 i
->getIndirectR(), bld
.mkImm(2)),
668 i
->setIndirectR(hnd
);
669 i
->setIndirectS(NULL
);
670 } else if (i
->tex
.r
== i
->tex
.s
|| i
->op
== OP_TXF
) {
671 i
->tex
.r
+= prog
->driver
->io
.texBindBase
/ 4;
672 i
->tex
.s
= 0; // only a single cX[] value possible here
674 Value
*hnd
= bld
.getScratch();
675 Value
*rHnd
= loadTexHandle(NULL
, i
->tex
.r
);
676 Value
*sHnd
= loadTexHandle(NULL
, i
->tex
.s
);
678 bld
.mkOp3(OP_INSBF
, TYPE_U32
, hnd
, rHnd
, bld
.mkImm(0x1400), sHnd
);
680 i
->tex
.r
= 0; // not used for indirect tex
682 i
->setIndirectR(hnd
);
684 if (i
->tex
.target
.isArray()) {
685 LValue
*layer
= new_LValue(func
, FILE_GPR
);
686 Value
*src
= i
->getSrc(lyr
);
687 const int sat
= (i
->op
== OP_TXF
) ? 1 : 0;
688 DataType sTy
= (i
->op
== OP_TXF
) ? TYPE_U32
: TYPE_F32
;
689 bld
.mkCvt(OP_CVT
, TYPE_U16
, layer
, sTy
, src
)->saturate
= sat
;
690 if (i
->op
!= OP_TXD
|| chipset
< NVISA_GM107_CHIPSET
) {
691 for (int s
= dim
; s
>= 1; --s
)
692 i
->setSrc(s
, i
->getSrc(s
- 1));
695 i
->setSrc(dim
, layer
);
698 // Move the indirect reference to the first place
699 if (i
->tex
.rIndirectSrc
>= 0 && (
700 i
->op
== OP_TXD
|| chipset
< NVISA_GM107_CHIPSET
)) {
701 Value
*hnd
= i
->getIndirectR();
703 i
->setIndirectR(NULL
);
704 i
->moveSources(0, 1);
706 i
->tex
.rIndirectSrc
= 0;
707 i
->tex
.sIndirectSrc
= -1;
710 // (nvc0) generate and move the tsc/tic/array source to the front
711 if (i
->tex
.target
.isArray() || i
->tex
.rIndirectSrc
>= 0 || i
->tex
.sIndirectSrc
>= 0) {
712 LValue
*src
= new_LValue(func
, FILE_GPR
); // 0xttxsaaaa
714 Value
*ticRel
= i
->getIndirectR();
715 Value
*tscRel
= i
->getIndirectS();
718 i
->setSrc(i
->tex
.rIndirectSrc
, NULL
);
720 ticRel
= bld
.mkOp2v(OP_ADD
, TYPE_U32
, bld
.getScratch(),
721 ticRel
, bld
.mkImm(i
->tex
.r
));
724 i
->setSrc(i
->tex
.sIndirectSrc
, NULL
);
726 tscRel
= bld
.mkOp2v(OP_ADD
, TYPE_U32
, bld
.getScratch(),
727 tscRel
, bld
.mkImm(i
->tex
.s
));
730 Value
*arrayIndex
= i
->tex
.target
.isArray() ? i
->getSrc(lyr
) : NULL
;
731 for (int s
= dim
; s
>= 1; --s
)
732 i
->setSrc(s
, i
->getSrc(s
- 1));
733 i
->setSrc(0, arrayIndex
);
736 int sat
= (i
->op
== OP_TXF
) ? 1 : 0;
737 DataType sTy
= (i
->op
== OP_TXF
) ? TYPE_U32
: TYPE_F32
;
738 bld
.mkCvt(OP_CVT
, TYPE_U16
, src
, sTy
, arrayIndex
)->saturate
= sat
;
744 bld
.mkOp3(OP_INSBF
, TYPE_U32
, src
, ticRel
, bld
.mkImm(0x0917), src
);
746 bld
.mkOp3(OP_INSBF
, TYPE_U32
, src
, tscRel
, bld
.mkImm(0x0710), src
);
751 // For nvc0, the sample id has to be in the second operand, as the offset
752 // does. Right now we don't know how to pass both in, and this case can't
753 // happen with OpenGL. On nve0, the sample id is part of the texture
754 // coordinate argument.
755 assert(chipset
>= NVISA_GK104_CHIPSET
||
756 !i
->tex
.useOffsets
|| !i
->tex
.target
.isMS());
758 // offset is between lod and dc
759 if (i
->tex
.useOffsets
) {
761 int s
= i
->srcCount(0xff, true);
762 if (i
->op
!= OP_TXD
|| chipset
< NVISA_GK104_CHIPSET
) {
763 if (i
->tex
.target
.isShadow())
765 if (i
->srcExists(s
)) // move potential predicate out of the way
766 i
->moveSources(s
, 1);
767 if (i
->tex
.useOffsets
== 4 && i
->srcExists(s
+ 1))
768 i
->moveSources(s
+ 1, 1);
770 if (i
->op
== OP_TXG
) {
771 // Either there is 1 offset, which goes into the 2 low bytes of the
772 // first source, or there are 4 offsets, which go into 2 sources (8
773 // values, 1 byte each).
774 Value
*offs
[2] = {NULL
, NULL
};
775 for (n
= 0; n
< i
->tex
.useOffsets
; n
++) {
776 for (c
= 0; c
< 2; ++c
) {
777 if ((n
% 2) == 0 && c
== 0)
778 offs
[n
/ 2] = i
->offset
[n
][c
].get();
780 bld
.mkOp3(OP_INSBF
, TYPE_U32
,
782 i
->offset
[n
][c
].get(),
783 bld
.mkImm(0x800 | ((n
* 16 + c
* 8) % 32)),
787 i
->setSrc(s
, offs
[0]);
789 i
->setSrc(s
+ 1, offs
[1]);
792 assert(i
->tex
.useOffsets
== 1);
793 for (c
= 0; c
< 3; ++c
) {
795 if (!i
->offset
[0][c
].getImmediate(val
))
796 assert(!"non-immediate offset passed to non-TXG");
797 imm
|= (val
.reg
.data
.u32
& 0xf) << (c
* 4);
799 if (i
->op
== OP_TXD
&& chipset
>= NVISA_GK104_CHIPSET
) {
800 // The offset goes into the upper 16 bits of the array index. So
801 // create it if it's not already there, and INSBF it if it already
803 s
= (i
->tex
.rIndirectSrc
>= 0) ? 1 : 0;
804 if (chipset
>= NVISA_GM107_CHIPSET
)
806 if (i
->tex
.target
.isArray()) {
807 bld
.mkOp3(OP_INSBF
, TYPE_U32
, i
->getSrc(s
),
808 bld
.loadImm(NULL
, imm
), bld
.mkImm(0xc10),
811 i
->moveSources(s
, 1);
812 i
->setSrc(s
, bld
.loadImm(NULL
, imm
<< 16));
815 i
->setSrc(s
, bld
.loadImm(NULL
, imm
));
820 if (chipset
>= NVISA_GK104_CHIPSET
) {
822 // If TEX requires more than 4 sources, the 2nd register tuple must be
823 // aligned to 4, even if it consists of just a single 4-byte register.
825 // XXX HACK: We insert 0 sources to avoid the 5 or 6 regs case.
827 int s
= i
->srcCount(0xff, true);
828 if (s
> 4 && s
< 7) {
829 if (i
->srcExists(s
)) // move potential predicate out of the way
830 i
->moveSources(s
, 7 - s
);
832 i
->setSrc(s
++, bld
.loadImm(NULL
, 0));
840 NVC0LoweringPass::handleManualTXD(TexInstruction
*i
)
842 static const uint8_t qOps
[4][2] =
844 { QUADOP(MOV2
, ADD
, MOV2
, ADD
), QUADOP(MOV2
, MOV2
, ADD
, ADD
) }, // l0
845 { QUADOP(SUBR
, MOV2
, SUBR
, MOV2
), QUADOP(MOV2
, MOV2
, ADD
, ADD
) }, // l1
846 { QUADOP(MOV2
, ADD
, MOV2
, ADD
), QUADOP(SUBR
, SUBR
, MOV2
, MOV2
) }, // l2
847 { QUADOP(SUBR
, MOV2
, SUBR
, MOV2
), QUADOP(SUBR
, SUBR
, MOV2
, MOV2
) }, // l3
852 Value
*zero
= bld
.loadImm(bld
.getSSA(), 0);
854 const int dim
= i
->tex
.target
.getDim() + i
->tex
.target
.isCube();
855 const int array
= i
->tex
.target
.isArray();
857 i
->op
= OP_TEX
; // no need to clone dPdx/dPdy later
859 for (c
= 0; c
< dim
; ++c
)
860 crd
[c
] = bld
.getScratch();
862 bld
.mkOp(OP_QUADON
, TYPE_NONE
, NULL
);
863 for (l
= 0; l
< 4; ++l
) {
864 // mov coordinates from lane l to all lanes
865 for (c
= 0; c
< dim
; ++c
)
866 bld
.mkQuadop(0x00, crd
[c
], l
, i
->getSrc(c
+ array
), zero
);
867 // add dPdx from lane l to lanes dx
868 for (c
= 0; c
< dim
; ++c
)
869 bld
.mkQuadop(qOps
[l
][0], crd
[c
], l
, i
->dPdx
[c
].get(), crd
[c
]);
870 // add dPdy from lane l to lanes dy
871 for (c
= 0; c
< dim
; ++c
)
872 bld
.mkQuadop(qOps
[l
][1], crd
[c
], l
, i
->dPdy
[c
].get(), crd
[c
]);
874 bld
.insert(tex
= cloneForward(func
, i
));
875 for (c
= 0; c
< dim
; ++c
)
876 tex
->setSrc(c
+ array
, crd
[c
]);
878 for (c
= 0; i
->defExists(c
); ++c
) {
880 def
[c
][l
] = bld
.getSSA();
881 mov
= bld
.mkMov(def
[c
][l
], tex
->getDef(c
));
886 bld
.mkOp(OP_QUADPOP
, TYPE_NONE
, NULL
);
888 for (c
= 0; i
->defExists(c
); ++c
) {
889 Instruction
*u
= bld
.mkOp(OP_UNION
, TYPE_U32
, i
->getDef(c
));
890 for (l
= 0; l
< 4; ++l
)
891 u
->setSrc(l
, def
[c
][l
]);
899 NVC0LoweringPass::handleTXD(TexInstruction
*txd
)
901 int dim
= txd
->tex
.target
.getDim() + txd
->tex
.target
.isCube();
902 unsigned arg
= txd
->tex
.target
.getArgCount();
903 unsigned expected_args
= arg
;
904 const int chipset
= prog
->getTarget()->getChipset();
906 if (chipset
>= NVISA_GK104_CHIPSET
) {
907 if (!txd
->tex
.target
.isArray() && txd
->tex
.useOffsets
)
909 if (txd
->tex
.rIndirectSrc
>= 0 || txd
->tex
.sIndirectSrc
>= 0)
912 if (txd
->tex
.useOffsets
)
914 if (!txd
->tex
.target
.isArray() && (
915 txd
->tex
.rIndirectSrc
>= 0 || txd
->tex
.sIndirectSrc
>= 0))
919 if (expected_args
> 4 ||
921 txd
->tex
.target
.isShadow())
925 while (txd
->srcExists(arg
))
928 txd
->tex
.derivAll
= true;
929 if (txd
->op
== OP_TEX
)
930 return handleManualTXD(txd
);
932 assert(arg
== expected_args
);
933 for (int c
= 0; c
< dim
; ++c
) {
934 txd
->setSrc(arg
+ c
* 2 + 0, txd
->dPdx
[c
]);
935 txd
->setSrc(arg
+ c
* 2 + 1, txd
->dPdy
[c
]);
936 txd
->dPdx
[c
].set(NULL
);
937 txd
->dPdy
[c
].set(NULL
);
943 NVC0LoweringPass::handleTXQ(TexInstruction
*txq
)
945 const int chipset
= prog
->getTarget()->getChipset();
946 if (chipset
>= NVISA_GK104_CHIPSET
&& txq
->tex
.rIndirectSrc
< 0)
947 txq
->tex
.r
+= prog
->driver
->io
.texBindBase
/ 4;
949 if (txq
->tex
.rIndirectSrc
< 0)
952 Value
*ticRel
= txq
->getIndirectR();
954 txq
->setIndirectS(NULL
);
955 txq
->tex
.sIndirectSrc
= -1;
959 if (chipset
< NVISA_GK104_CHIPSET
) {
960 LValue
*src
= new_LValue(func
, FILE_GPR
); // 0xttxsaaaa
962 txq
->setSrc(txq
->tex
.rIndirectSrc
, NULL
);
964 ticRel
= bld
.mkOp2v(OP_ADD
, TYPE_U32
, bld
.getScratch(),
965 ticRel
, bld
.mkImm(txq
->tex
.r
));
967 bld
.mkOp2(OP_SHL
, TYPE_U32
, src
, ticRel
, bld
.mkImm(0x17));
969 txq
->moveSources(0, 1);
972 Value
*hnd
= loadTexHandle(
973 bld
.mkOp2v(OP_SHL
, TYPE_U32
, bld
.getSSA(),
974 txq
->getIndirectR(), bld
.mkImm(2)),
979 txq
->setIndirectR(NULL
);
980 txq
->moveSources(0, 1);
982 txq
->tex
.rIndirectSrc
= 0;
989 NVC0LoweringPass::handleTXLQ(TexInstruction
*i
)
991 /* The outputs are inverted compared to what the TGSI instruction
992 * expects. Take that into account in the mask.
994 assert((i
->tex
.mask
& ~3) == 0);
995 if (i
->tex
.mask
== 1)
997 else if (i
->tex
.mask
== 2)
1000 bld
.setPosition(i
, true);
1002 /* The returned values are not quite what we want:
1003 * (a) convert from s16/u16 to f32
1004 * (b) multiply by 1/256
1006 for (int def
= 0; def
< 2; ++def
) {
1007 if (!i
->defExists(def
))
1009 enum DataType type
= TYPE_S16
;
1010 if (i
->tex
.mask
== 2 || def
> 0)
1012 bld
.mkCvt(OP_CVT
, TYPE_F32
, i
->getDef(def
), type
, i
->getDef(def
));
1013 bld
.mkOp2(OP_MUL
, TYPE_F32
, i
->getDef(def
),
1014 i
->getDef(def
), bld
.loadImm(NULL
, 1.0f
/ 256));
1016 if (i
->tex
.mask
== 3) {
1017 LValue
*t
= new_LValue(func
, FILE_GPR
);
1018 bld
.mkMov(t
, i
->getDef(0));
1019 bld
.mkMov(i
->getDef(0), i
->getDef(1));
1020 bld
.mkMov(i
->getDef(1), t
);
1027 NVC0LoweringPass::handleATOM(Instruction
*atom
)
1031 switch (atom
->src(0).getFile()) {
1032 case FILE_MEMORY_LOCAL
:
1035 case FILE_MEMORY_SHARED
:
1039 assert(atom
->src(0).getFile() == FILE_MEMORY_GLOBAL
);
1043 bld
.mkOp1v(OP_RDSV
, TYPE_U32
, bld
.getScratch(), bld
.mkSysVal(sv
, 0));
1044 Value
*ptr
= atom
->getIndirect(0, 0);
1046 atom
->setSrc(0, cloneShallow(func
, atom
->getSrc(0)));
1047 atom
->getSrc(0)->reg
.file
= FILE_MEMORY_GLOBAL
;
1049 base
= bld
.mkOp2v(OP_ADD
, TYPE_U32
, base
, base
, ptr
);
1050 atom
->setIndirect(0, 0, base
);
1056 NVC0LoweringPass::handleCasExch(Instruction
*cas
, bool needCctl
)
1058 if (cas
->subOp
!= NV50_IR_SUBOP_ATOM_CAS
&&
1059 cas
->subOp
!= NV50_IR_SUBOP_ATOM_EXCH
)
1061 bld
.setPosition(cas
, true);
1064 Instruction
*cctl
= bld
.mkOp1(OP_CCTL
, TYPE_NONE
, NULL
, cas
->getSrc(0));
1065 cctl
->setIndirect(0, 0, cas
->getIndirect(0, 0));
1067 cctl
->subOp
= NV50_IR_SUBOP_CCTL_IV
;
1068 if (cas
->isPredicated())
1069 cctl
->setPredicate(cas
->cc
, cas
->getPredicate());
1072 if (cas
->defExists(0) && cas
->subOp
== NV50_IR_SUBOP_ATOM_CAS
) {
1073 // CAS is crazy. It's 2nd source is a double reg, and the 3rd source
1074 // should be set to the high part of the double reg or bad things will
1075 // happen elsewhere in the universe.
1076 // Also, it sometimes returns the new value instead of the old one
1077 // under mysterious circumstances.
1078 Value
*dreg
= bld
.getSSA(8);
1079 bld
.setPosition(cas
, false);
1080 bld
.mkOp2(OP_MERGE
, TYPE_U64
, dreg
, cas
->getSrc(1), cas
->getSrc(2));
1081 cas
->setSrc(1, dreg
);
1088 NVC0LoweringPass::loadResInfo32(Value
*ptr
, uint32_t off
)
1090 uint8_t b
= prog
->driver
->io
.resInfoCBSlot
;
1091 off
+= prog
->driver
->io
.suInfoBase
;
1093 mkLoadv(TYPE_U32
, bld
.mkSymbol(FILE_MEMORY_CONST
, b
, TYPE_U32
, off
), ptr
);
1097 NVC0LoweringPass::loadMsInfo32(Value
*ptr
, uint32_t off
)
1099 uint8_t b
= prog
->driver
->io
.msInfoCBSlot
;
1100 off
+= prog
->driver
->io
.msInfoBase
;
1102 mkLoadv(TYPE_U32
, bld
.mkSymbol(FILE_MEMORY_CONST
, b
, TYPE_U32
, off
), ptr
);
1105 /* On nvc0, surface info is obtained via the surface binding points passed
1106 * to the SULD/SUST instructions.
1107 * On nve4, surface info is stored in c[] and is used by various special
1108 * instructions, e.g. for clamping coordiantes or generating an address.
1109 * They couldn't just have added an equivalent to TIC now, couldn't they ?
1111 #define NVE4_SU_INFO_ADDR 0x00
1112 #define NVE4_SU_INFO_FMT 0x04
1113 #define NVE4_SU_INFO_DIM_X 0x08
1114 #define NVE4_SU_INFO_PITCH 0x0c
1115 #define NVE4_SU_INFO_DIM_Y 0x10
1116 #define NVE4_SU_INFO_ARRAY 0x14
1117 #define NVE4_SU_INFO_DIM_Z 0x18
1118 #define NVE4_SU_INFO_UNK1C 0x1c
1119 #define NVE4_SU_INFO_WIDTH 0x20
1120 #define NVE4_SU_INFO_HEIGHT 0x24
1121 #define NVE4_SU_INFO_DEPTH 0x28
1122 #define NVE4_SU_INFO_TARGET 0x2c
1123 #define NVE4_SU_INFO_CALL 0x30
1124 #define NVE4_SU_INFO_RAW_X 0x34
1125 #define NVE4_SU_INFO_MS_X 0x38
1126 #define NVE4_SU_INFO_MS_Y 0x3c
1128 #define NVE4_SU_INFO__STRIDE 0x40
1130 #define NVE4_SU_INFO_DIM(i) (0x08 + (i) * 8)
1131 #define NVE4_SU_INFO_SIZE(i) (0x20 + (i) * 4)
1132 #define NVE4_SU_INFO_MS(i) (0x38 + (i) * 4)
1134 static inline uint16_t getSuClampSubOp(const TexInstruction
*su
, int c
)
1136 switch (su
->tex
.target
.getEnum()) {
1137 case TEX_TARGET_BUFFER
: return NV50_IR_SUBOP_SUCLAMP_PL(0, 1);
1138 case TEX_TARGET_RECT
: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1139 case TEX_TARGET_1D
: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1140 case TEX_TARGET_1D_ARRAY
: return (c
== 1) ?
1141 NV50_IR_SUBOP_SUCLAMP_PL(0, 2) :
1142 NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1143 case TEX_TARGET_2D
: return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
1144 case TEX_TARGET_2D_MS
: return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
1145 case TEX_TARGET_2D_ARRAY
: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1146 case TEX_TARGET_2D_MS_ARRAY
: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1147 case TEX_TARGET_3D
: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1148 case TEX_TARGET_CUBE
: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1149 case TEX_TARGET_CUBE_ARRAY
: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1157 NVC0LoweringPass::adjustCoordinatesMS(TexInstruction
*tex
)
1159 const uint16_t base
= tex
->tex
.r
* NVE4_SU_INFO__STRIDE
;
1160 const int arg
= tex
->tex
.target
.getArgCount();
1162 if (tex
->tex
.target
== TEX_TARGET_2D_MS
)
1163 tex
->tex
.target
= TEX_TARGET_2D
;
1165 if (tex
->tex
.target
== TEX_TARGET_2D_MS_ARRAY
)
1166 tex
->tex
.target
= TEX_TARGET_2D_ARRAY
;
1170 Value
*x
= tex
->getSrc(0);
1171 Value
*y
= tex
->getSrc(1);
1172 Value
*s
= tex
->getSrc(arg
- 1);
1174 Value
*tx
= bld
.getSSA(), *ty
= bld
.getSSA(), *ts
= bld
.getSSA();
1176 Value
*ms_x
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_MS(0));
1177 Value
*ms_y
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_MS(1));
1179 bld
.mkOp2(OP_SHL
, TYPE_U32
, tx
, x
, ms_x
);
1180 bld
.mkOp2(OP_SHL
, TYPE_U32
, ty
, y
, ms_y
);
1182 s
= bld
.mkOp2v(OP_AND
, TYPE_U32
, ts
, s
, bld
.loadImm(NULL
, 0x7));
1183 s
= bld
.mkOp2v(OP_SHL
, TYPE_U32
, ts
, ts
, bld
.mkImm(3));
1185 Value
*dx
= loadMsInfo32(ts
, 0x0);
1186 Value
*dy
= loadMsInfo32(ts
, 0x4);
1188 bld
.mkOp2(OP_ADD
, TYPE_U32
, tx
, tx
, dx
);
1189 bld
.mkOp2(OP_ADD
, TYPE_U32
, ty
, ty
, dy
);
1193 tex
->moveSources(arg
, -1);
1196 // Sets 64-bit "generic address", predicate and format sources for SULD/SUST.
1197 // They're computed from the coordinates using the surface info in c[] space.
1199 NVC0LoweringPass::processSurfaceCoordsNVE4(TexInstruction
*su
)
1202 const bool atom
= su
->op
== OP_SUREDB
|| su
->op
== OP_SUREDP
;
1204 su
->op
== OP_SULDB
|| su
->op
== OP_SUSTB
|| su
->op
== OP_SUREDB
;
1205 const int idx
= su
->tex
.r
;
1206 const int dim
= su
->tex
.target
.getDim();
1207 const int arg
= dim
+ (su
->tex
.target
.isArray() ? 1 : 0);
1208 const uint16_t base
= idx
* NVE4_SU_INFO__STRIDE
;
1210 Value
*zero
= bld
.mkImm(0);
1214 Value
*bf
, *eau
, *off
;
1217 off
= bld
.getScratch(4);
1218 bf
= bld
.getScratch(4);
1219 addr
= bld
.getSSA(8);
1220 pred
= bld
.getScratch(1, FILE_PREDICATE
);
1222 bld
.setPosition(su
, false);
1224 adjustCoordinatesMS(su
);
1226 // calculate clamped coordinates
1227 for (c
= 0; c
< arg
; ++c
) {
1228 src
[c
] = bld
.getScratch();
1230 v
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_RAW_X
);
1232 v
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_DIM(c
));
1233 bld
.mkOp3(OP_SUCLAMP
, TYPE_S32
, src
[c
], su
->getSrc(c
), v
, zero
)
1234 ->subOp
= getSuClampSubOp(su
, c
);
1239 // set predicate output
1240 if (su
->tex
.target
== TEX_TARGET_BUFFER
) {
1241 src
[0]->getInsn()->setFlagsDef(1, pred
);
1243 if (su
->tex
.target
.isArray()) {
1244 p1
= bld
.getSSA(1, FILE_PREDICATE
);
1245 src
[dim
]->getInsn()->setFlagsDef(1, p1
);
1248 // calculate pixel offset
1250 if (su
->tex
.target
!= TEX_TARGET_BUFFER
)
1251 bld
.mkOp2(OP_AND
, TYPE_U32
, off
, src
[0], bld
.loadImm(NULL
, 0xffff));
1254 v
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_UNK1C
);
1255 bld
.mkOp3(OP_MADSP
, TYPE_U32
, off
, src
[2], v
, src
[1])
1256 ->subOp
= NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
1258 v
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_PITCH
);
1259 bld
.mkOp3(OP_MADSP
, TYPE_U32
, off
, off
, v
, src
[0])
1260 ->subOp
= NV50_IR_SUBOP_MADSP(0,2,8); // u32 u16l u16l
1263 v
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_PITCH
);
1264 bld
.mkOp3(OP_MADSP
, TYPE_U32
, off
, src
[1], v
, src
[0])
1265 ->subOp
= su
->tex
.target
.isArray() ?
1266 NV50_IR_SUBOP_MADSP_SD
: NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
1269 // calculate effective address part 1
1270 if (su
->tex
.target
== TEX_TARGET_BUFFER
) {
1274 v
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_FMT
);
1275 bld
.mkOp3(OP_VSHL
, TYPE_U32
, bf
, src
[0], v
, zero
)
1276 ->subOp
= NV50_IR_SUBOP_V1(7,6,8|2);
1290 if (!su
->tex
.target
.isArray()) {
1291 z
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_UNK1C
);
1292 subOp
= NV50_IR_SUBOP_SUBFM_3D
;
1296 subOp
= NV50_IR_SUBOP_SUBFM_3D
;
1300 insn
= bld
.mkOp3(OP_SUBFM
, TYPE_U32
, bf
, src
[0], y
, z
);
1301 insn
->subOp
= subOp
;
1302 insn
->setFlagsDef(1, pred
);
1306 v
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_ADDR
);
1308 if (su
->tex
.target
== TEX_TARGET_BUFFER
) {
1311 eau
= bld
.mkOp3v(OP_SUEAU
, TYPE_U32
, bld
.getScratch(4), off
, bf
, v
);
1313 // add array layer offset
1314 if (su
->tex
.target
.isArray()) {
1315 v
= loadResInfo32(NULL
, base
+ NVE4_SU_INFO_ARRAY
);
1317 bld
.mkOp3(OP_MADSP
, TYPE_U32
, eau
, src
[1], v
, eau
)
1318 ->subOp
= NV50_IR_SUBOP_MADSP(4,0,0); // u16 u24 u32
1320 bld
.mkOp3(OP_MADSP
, TYPE_U32
, eau
, v
, src
[2], eau
)
1321 ->subOp
= NV50_IR_SUBOP_MADSP(0,0,0); // u32 u24 u32
1322 // combine predicates
1324 bld
.mkOp2(OP_OR
, TYPE_U8
, pred
, pred
, p1
);
1329 if (su
->tex
.target
== TEX_TARGET_BUFFER
) {
1333 // bf == g[] address & 0xff
1334 // eau == g[] address >> 8
1335 bld
.mkOp3(OP_PERMT
, TYPE_U32
, bf
, lo
, bld
.loadImm(NULL
, 0x6540), eau
);
1336 bld
.mkOp3(OP_PERMT
, TYPE_U32
, eau
, zero
, bld
.loadImm(NULL
, 0x0007), eau
);
1338 if (su
->op
== OP_SULDP
&& su
->tex
.target
== TEX_TARGET_BUFFER
) {
1339 // Convert from u32 to u8 address format, which is what the library code
1340 // doing SULDP currently uses.
1341 // XXX: can SUEAU do this ?
1342 // XXX: does it matter that we don't mask high bytes in bf ?
1344 bld
.mkOp2(OP_SHR
, TYPE_U32
, off
, bf
, bld
.mkImm(8));
1345 bld
.mkOp2(OP_ADD
, TYPE_U32
, eau
, eau
, off
);
1348 bld
.mkOp2(OP_MERGE
, TYPE_U64
, addr
, bf
, eau
);
1350 if (atom
&& su
->tex
.target
== TEX_TARGET_BUFFER
)
1351 bld
.mkOp2(OP_ADD
, TYPE_U64
, addr
, addr
, off
);
1353 // let's just set it 0 for raw access and hope it works
1355 bld
.mkImm(0) : loadResInfo32(NULL
, base
+ NVE4_SU_INFO_FMT
);
1357 // get rid of old coordinate sources, make space for fmt info and predicate
1358 su
->moveSources(arg
, 3 - arg
);
1359 // set 64 bit address and 32-bit format sources
1360 su
->setSrc(0, addr
);
1362 su
->setSrc(2, pred
);
1366 NVC0LoweringPass::handleSurfaceOpNVE4(TexInstruction
*su
)
1368 processSurfaceCoordsNVE4(su
);
1370 // Who do we hate more ? The person who decided that nvc0's SULD doesn't
1371 // have to support conversion or the person who decided that, in OpenCL,
1372 // you don't have to specify the format here like you do in OpenGL ?
1374 if (su
->op
== OP_SULDP
) {
1375 // We don't patch shaders. Ever.
1376 // You get an indirect call to our library blob here.
1377 // But at least it's uniform.
1378 FlowInstruction
*call
;
1381 uint16_t base
= su
->tex
.r
* NVE4_SU_INFO__STRIDE
+ NVE4_SU_INFO_CALL
;
1383 for (int i
= 0; i
< 4; ++i
)
1384 (r
[i
] = bld
.getScratch(4, FILE_GPR
))->reg
.data
.id
= i
;
1385 for (int i
= 0; i
< 3; ++i
)
1386 (p
[i
] = bld
.getScratch(1, FILE_PREDICATE
))->reg
.data
.id
= i
;
1387 (r
[4] = bld
.getScratch(8, FILE_GPR
))->reg
.data
.id
= 4;
1389 bld
.mkMov(p
[1], bld
.mkImm((su
->cache
== CACHE_CA
) ? 1 : 0), TYPE_U8
);
1390 bld
.mkMov(p
[2], bld
.mkImm((su
->cache
== CACHE_CG
) ? 1 : 0), TYPE_U8
);
1391 bld
.mkMov(p
[0], su
->getSrc(2), TYPE_U8
);
1392 bld
.mkMov(r
[4], su
->getSrc(0), TYPE_U64
);
1393 bld
.mkMov(r
[2], su
->getSrc(1), TYPE_U32
);
1395 call
= bld
.mkFlow(OP_CALL
, NULL
, su
->cc
, su
->getPredicate());
1399 call
->setSrc(0, bld
.mkSymbol(FILE_MEMORY_CONST
,
1400 prog
->driver
->io
.resInfoCBSlot
, TYPE_U32
,
1401 prog
->driver
->io
.suInfoBase
+ base
));
1402 call
->setSrc(1, r
[2]);
1403 call
->setSrc(2, r
[4]);
1404 for (int i
= 0; i
< 3; ++i
)
1405 call
->setSrc(3 + i
, p
[i
]);
1406 for (int i
= 0; i
< 4; ++i
) {
1407 call
->setDef(i
, r
[i
]);
1408 bld
.mkMov(su
->getDef(i
), r
[i
]);
1410 call
->setDef(4, p
[1]);
1411 delete_Instruction(bld
.getProgram(), su
);
1414 if (su
->op
== OP_SUREDB
|| su
->op
== OP_SUREDP
) {
1415 // FIXME: for out of bounds access, destination value will be undefined !
1416 Value
*pred
= su
->getSrc(2);
1417 CondCode cc
= CC_NOT_P
;
1418 if (su
->getPredicate()) {
1419 pred
= bld
.getScratch(1, FILE_PREDICATE
);
1421 if (cc
== CC_NOT_P
) {
1422 bld
.mkOp2(OP_OR
, TYPE_U8
, pred
, su
->getPredicate(), su
->getSrc(2));
1424 bld
.mkOp2(OP_AND
, TYPE_U8
, pred
, su
->getPredicate(), su
->getSrc(2));
1425 pred
->getInsn()->src(1).mod
= Modifier(NV50_IR_MOD_NOT
);
1428 Instruction
*red
= bld
.mkOp(OP_ATOM
, su
->dType
, su
->getDef(0));
1429 red
->subOp
= su
->subOp
;
1431 gMemBase
= bld
.mkSymbol(FILE_MEMORY_GLOBAL
, 0, TYPE_U32
, 0);
1432 red
->setSrc(0, gMemBase
);
1433 red
->setSrc(1, su
->getSrc(3));
1434 if (su
->subOp
== NV50_IR_SUBOP_ATOM_CAS
)
1435 red
->setSrc(2, su
->getSrc(4));
1436 red
->setIndirect(0, 0, su
->getSrc(0));
1437 red
->setPredicate(cc
, pred
);
1438 delete_Instruction(bld
.getProgram(), su
);
1439 handleCasExch(red
, true);
1441 su
->sType
= (su
->tex
.target
== TEX_TARGET_BUFFER
) ? TYPE_U32
: TYPE_U8
;
1446 NVC0LoweringPass::handleWRSV(Instruction
*i
)
1452 // must replace, $sreg are not writeable
1453 addr
= targ
->getSVAddress(FILE_SHADER_OUTPUT
, i
->getSrc(0)->asSym());
1456 sym
= bld
.mkSymbol(FILE_SHADER_OUTPUT
, 0, i
->sType
, addr
);
1458 st
= bld
.mkStore(OP_EXPORT
, i
->dType
, sym
, i
->getIndirect(0, 0),
1460 st
->perPatch
= i
->perPatch
;
1462 bld
.getBB()->remove(i
);
1467 NVC0LoweringPass::readTessCoord(LValue
*dst
, int c
)
1469 Value
*laneid
= bld
.getSSA();
1472 bld
.mkOp1(OP_RDSV
, TYPE_U32
, laneid
, bld
.mkSysVal(SV_LANEID
, 0));
1487 bld
.mkFetch(x
, TYPE_F32
, FILE_SHADER_OUTPUT
, 0x2f0, NULL
, laneid
);
1489 bld
.mkFetch(y
, TYPE_F32
, FILE_SHADER_OUTPUT
, 0x2f4, NULL
, laneid
);
1492 bld
.mkOp2(OP_ADD
, TYPE_F32
, dst
, x
, y
);
1493 bld
.mkOp2(OP_SUB
, TYPE_F32
, dst
, bld
.loadImm(NULL
, 1.0f
), dst
);
1498 NVC0LoweringPass::handleRDSV(Instruction
*i
)
1500 Symbol
*sym
= i
->getSrc(0)->asSym();
1501 const SVSemantic sv
= sym
->reg
.data
.sv
.sv
;
1504 uint32_t addr
= targ
->getSVAddress(FILE_SHADER_INPUT
, sym
);
1506 if (addr
>= 0x400) {
1508 if (sym
->reg
.data
.sv
.index
== 3) {
1509 // TGSI backend may use 4th component of TID,NTID,CTAID,NCTAID
1511 i
->setSrc(0, bld
.mkImm((sv
== SV_NTID
|| sv
== SV_NCTAID
) ? 1 : 0));
1513 if (sv
== SV_VERTEX_COUNT
) {
1514 bld
.setPosition(i
, true);
1515 bld
.mkOp2(OP_EXTBF
, TYPE_U32
, i
->getDef(0), i
->getDef(0), bld
.mkImm(0x808));
1522 assert(prog
->getType() == Program::TYPE_FRAGMENT
);
1523 if (i
->srcExists(1)) {
1524 // Pass offset through to the interpolation logic
1525 ld
= bld
.mkInterp(NV50_IR_INTERP_LINEAR
| NV50_IR_INTERP_OFFSET
,
1526 i
->getDef(0), addr
, NULL
);
1527 ld
->setSrc(1, i
->getSrc(1));
1529 bld
.mkInterp(NV50_IR_INTERP_LINEAR
, i
->getDef(0), addr
, NULL
);
1534 Value
*face
= i
->getDef(0);
1535 bld
.mkInterp(NV50_IR_INTERP_FLAT
, face
, addr
, NULL
);
1536 if (i
->dType
== TYPE_F32
) {
1537 bld
.mkOp2(OP_OR
, TYPE_U32
, face
, face
, bld
.mkImm(0x00000001));
1538 bld
.mkOp1(OP_NEG
, TYPE_S32
, face
, face
);
1539 bld
.mkCvt(OP_CVT
, TYPE_F32
, face
, TYPE_S32
, face
);
1544 assert(prog
->getType() == Program::TYPE_TESSELLATION_EVAL
);
1545 readTessCoord(i
->getDef(0)->asLValue(), i
->getSrc(0)->reg
.data
.sv
.index
);
1550 assert(targ
->getChipset() >= NVISA_GK104_CHIPSET
); // mov $sreg otherwise
1551 if (sym
->reg
.data
.sv
.index
== 3) {
1553 i
->setSrc(0, bld
.mkImm(sv
== SV_GRIDID
? 0 : 1));
1556 addr
+= prog
->driver
->prop
.cp
.gridInfoBase
;
1557 bld
.mkLoad(TYPE_U32
, i
->getDef(0),
1558 bld
.mkSymbol(FILE_MEMORY_CONST
, 0, TYPE_U32
, addr
), NULL
);
1560 case SV_SAMPLE_INDEX
:
1561 // TODO: Properly pass source as an address in the PIX address space
1562 // (which can be of the form [r0+offset]). But this is currently
1564 ld
= bld
.mkOp1(OP_PIXLD
, TYPE_U32
, i
->getDef(0), bld
.mkImm(0));
1565 ld
->subOp
= NV50_IR_SUBOP_PIXLD_SAMPLEID
;
1567 case SV_SAMPLE_POS
: {
1568 Value
*off
= new_LValue(func
, FILE_GPR
);
1569 ld
= bld
.mkOp1(OP_PIXLD
, TYPE_U32
, i
->getDef(0), bld
.mkImm(0));
1570 ld
->subOp
= NV50_IR_SUBOP_PIXLD_SAMPLEID
;
1571 bld
.mkOp2(OP_SHL
, TYPE_U32
, off
, i
->getDef(0), bld
.mkImm(3));
1572 bld
.mkLoad(TYPE_F32
,
1575 FILE_MEMORY_CONST
, prog
->driver
->io
.resInfoCBSlot
,
1576 TYPE_U32
, prog
->driver
->io
.sampleInfoBase
+
1577 4 * sym
->reg
.data
.sv
.index
),
1581 case SV_SAMPLE_MASK
:
1582 ld
= bld
.mkOp1(OP_PIXLD
, TYPE_U32
, i
->getDef(0), bld
.mkImm(0));
1583 ld
->subOp
= NV50_IR_SUBOP_PIXLD_COVMASK
;
1586 case SV_BASEINSTANCE
:
1588 ld
= bld
.mkLoad(TYPE_U32
, i
->getDef(0),
1589 bld
.mkSymbol(FILE_MEMORY_CONST
,
1590 prog
->driver
->io
.auxCBSlot
,
1592 prog
->driver
->io
.drawInfoBase
+
1593 4 * (sv
- SV_BASEVERTEX
)),
1597 if (prog
->getType() == Program::TYPE_TESSELLATION_EVAL
&& !i
->perPatch
)
1598 vtx
= bld
.mkOp1v(OP_PFETCH
, TYPE_U32
, bld
.getSSA(), bld
.mkImm(0));
1599 ld
= bld
.mkFetch(i
->getDef(0), i
->dType
,
1600 FILE_SHADER_INPUT
, addr
, i
->getIndirect(0, 0), vtx
);
1601 ld
->perPatch
= i
->perPatch
;
1604 bld
.getBB()->remove(i
);
1609 NVC0LoweringPass::handleDIV(Instruction
*i
)
1611 if (!isFloatType(i
->dType
))
1613 bld
.setPosition(i
, false);
1614 Instruction
*rcp
= bld
.mkOp1(OP_RCP
, i
->dType
, bld
.getSSA(typeSizeof(i
->dType
)), i
->getSrc(1));
1616 i
->setSrc(1, rcp
->getDef(0));
1621 NVC0LoweringPass::handleMOD(Instruction
*i
)
1623 if (!isFloatType(i
->dType
))
1625 LValue
*value
= bld
.getScratch(typeSizeof(i
->dType
));
1626 bld
.mkOp1(OP_RCP
, i
->dType
, value
, i
->getSrc(1));
1627 bld
.mkOp2(OP_MUL
, i
->dType
, value
, i
->getSrc(0), value
);
1628 bld
.mkOp1(OP_TRUNC
, i
->dType
, value
, value
);
1629 bld
.mkOp2(OP_MUL
, i
->dType
, value
, i
->getSrc(1), value
);
1631 i
->setSrc(1, value
);
1636 NVC0LoweringPass::handleSQRT(Instruction
*i
)
1638 Value
*pred
= bld
.getSSA(1, FILE_PREDICATE
);
1639 Value
*zero
= bld
.getSSA();
1642 bld
.mkOp1(OP_MOV
, TYPE_U32
, zero
, bld
.mkImm(0));
1643 if (i
->dType
== TYPE_F64
)
1644 zero
= bld
.mkOp2v(OP_MERGE
, TYPE_U64
, bld
.getSSA(8), zero
, zero
);
1645 bld
.mkCmp(OP_SET
, CC_LE
, i
->dType
, pred
, i
->dType
, i
->getSrc(0), zero
);
1646 bld
.mkOp1(OP_MOV
, i
->dType
, i
->getDef(0), zero
)->setPredicate(CC_P
, pred
);
1647 rsq
= bld
.mkOp1(OP_RSQ
, i
->dType
,
1648 bld
.getSSA(typeSizeof(i
->dType
)), i
->getSrc(0));
1649 rsq
->setPredicate(CC_NOT_P
, pred
);
1651 i
->setSrc(1, rsq
->getDef(0));
1652 i
->setPredicate(CC_NOT_P
, pred
);
1659 NVC0LoweringPass::handlePOW(Instruction
*i
)
1661 LValue
*val
= bld
.getScratch();
1663 bld
.mkOp1(OP_LG2
, TYPE_F32
, val
, i
->getSrc(0));
1664 bld
.mkOp2(OP_MUL
, TYPE_F32
, val
, i
->getSrc(1), val
)->dnz
= 1;
1665 bld
.mkOp1(OP_PREEX2
, TYPE_F32
, val
, val
);
1675 NVC0LoweringPass::handleEXPORT(Instruction
*i
)
1677 if (prog
->getType() == Program::TYPE_FRAGMENT
) {
1678 int id
= i
->getSrc(0)->reg
.data
.offset
/ 4;
1680 if (i
->src(0).isIndirect(0)) // TODO, ugly
1683 i
->subOp
= NV50_IR_SUBOP_MOV_FINAL
;
1684 i
->src(0).set(i
->src(1));
1686 i
->setDef(0, new_LValue(func
, FILE_GPR
));
1687 i
->getDef(0)->reg
.data
.id
= id
;
1689 prog
->maxGPR
= MAX2(prog
->maxGPR
, id
);
1691 if (prog
->getType() == Program::TYPE_GEOMETRY
) {
1692 i
->setIndirect(0, 1, gpEmitAddress
);
1698 NVC0LoweringPass::handleOUT(Instruction
*i
)
1700 Instruction
*prev
= i
->prev
;
1701 ImmediateValue stream
, prevStream
;
1703 // Only merge if the stream ids match. Also, note that the previous
1704 // instruction would have already been lowered, so we take arg1 from it.
1705 if (i
->op
== OP_RESTART
&& prev
&& prev
->op
== OP_EMIT
&&
1706 i
->src(0).getImmediate(stream
) &&
1707 prev
->src(1).getImmediate(prevStream
) &&
1708 stream
.reg
.data
.u32
== prevStream
.reg
.data
.u32
) {
1709 i
->prev
->subOp
= NV50_IR_SUBOP_EMIT_RESTART
;
1710 delete_Instruction(prog
, i
);
1712 assert(gpEmitAddress
);
1713 i
->setDef(0, gpEmitAddress
);
1714 i
->setSrc(1, i
->getSrc(0));
1715 i
->setSrc(0, gpEmitAddress
);
1720 // Generate a binary predicate if an instruction is predicated by
1721 // e.g. an f32 value.
1723 NVC0LoweringPass::checkPredicate(Instruction
*insn
)
1725 Value
*pred
= insn
->getPredicate();
1728 if (!pred
|| pred
->reg
.file
== FILE_PREDICATE
)
1730 pdst
= new_LValue(func
, FILE_PREDICATE
);
1732 // CAUTION: don't use pdst->getInsn, the definition might not be unique,
1733 // delay turning PSET(FSET(x,y),0) into PSET(x,y) to a later pass
1735 bld
.mkCmp(OP_SET
, CC_NEU
, insn
->dType
, pdst
, insn
->dType
, bld
.mkImm(0), pred
);
1737 insn
->setPredicate(insn
->cc
, pdst
);
1741 // - add quadop dance for texturing
1742 // - put FP outputs in GPRs
1743 // - convert instruction sequences
1746 NVC0LoweringPass::visit(Instruction
*i
)
1749 bld
.setPosition(i
, false);
1751 if (i
->cc
!= CC_ALWAYS
)
1760 return handleTEX(i
->asTex());
1762 return handleTXD(i
->asTex());
1764 return handleTXLQ(i
->asTex());
1766 return handleTXQ(i
->asTex());
1768 bld
.mkOp1(OP_PREEX2
, TYPE_F32
, i
->getDef(0), i
->getSrc(0));
1769 i
->setSrc(0, i
->getDef(0));
1772 return handlePOW(i
);
1774 return handleDIV(i
);
1776 return handleMOD(i
);
1778 return handleSQRT(i
);
1780 ret
= handleEXPORT(i
);
1784 return handleOUT(i
);
1786 return handleRDSV(i
);
1788 return handleWRSV(i
);
1790 if (i
->src(0).getFile() == FILE_SHADER_INPUT
) {
1791 if (prog
->getType() == Program::TYPE_COMPUTE
) {
1792 i
->getSrc(0)->reg
.file
= FILE_MEMORY_CONST
;
1793 i
->getSrc(0)->reg
.fileIndex
= 0;
1795 if (prog
->getType() == Program::TYPE_GEOMETRY
&&
1796 i
->src(0).isIndirect(0)) {
1797 // XXX: this assumes vec4 units
1798 Value
*ptr
= bld
.mkOp2v(OP_SHL
, TYPE_U32
, bld
.getSSA(),
1799 i
->getIndirect(0, 0), bld
.mkImm(4));
1800 i
->setIndirect(0, 0, ptr
);
1804 assert(prog
->getType() != Program::TYPE_FRAGMENT
); // INTERP
1806 } else if (i
->src(0).getFile() == FILE_MEMORY_CONST
) {
1807 if (i
->src(0).isIndirect(1)) {
1809 if (i
->src(0).isIndirect(0))
1810 ptr
= bld
.mkOp3v(OP_INSBF
, TYPE_U32
, bld
.getSSA(),
1811 i
->getIndirect(0, 1), bld
.mkImm(0x1010),
1812 i
->getIndirect(0, 0));
1814 ptr
= bld
.mkOp2v(OP_SHL
, TYPE_U32
, bld
.getSSA(),
1815 i
->getIndirect(0, 1), bld
.mkImm(16));
1816 i
->setIndirect(0, 1, NULL
);
1817 i
->setIndirect(0, 0, ptr
);
1818 i
->subOp
= NV50_IR_SUBOP_LDC_IS
;
1820 } else if (i
->src(0).getFile() == FILE_SHADER_OUTPUT
) {
1821 assert(prog
->getType() == Program::TYPE_TESSELLATION_CONTROL
);
1827 const bool cctl
= i
->src(0).getFile() == FILE_MEMORY_GLOBAL
;
1829 handleCasExch(i
, cctl
);
1838 if (targ
->getChipset() >= NVISA_GK104_CHIPSET
)
1839 handleSurfaceOpNVE4(i
->asTex());
1845 /* Kepler+ has a special opcode to compute a new base address to be used
1846 * for indirect loads.
1848 if (targ
->getChipset() >= NVISA_GK104_CHIPSET
&& !i
->perPatch
&&
1849 (i
->op
== OP_VFETCH
|| i
->op
== OP_EXPORT
) && i
->src(0).isIndirect(0)) {
1850 Instruction
*afetch
= bld
.mkOp1(OP_AFETCH
, TYPE_U32
, bld
.getSSA(),
1851 cloneShallow(func
, i
->getSrc(0)));
1852 afetch
->setIndirect(0, 0, i
->getIndirect(0, 0));
1853 i
->src(0).get()->reg
.data
.offset
= 0;
1854 i
->setIndirect(0, 0, afetch
->getDef(0));
1861 TargetNVC0::runLegalizePass(Program
*prog
, CGStage stage
) const
1863 if (stage
== CG_STAGE_PRE_SSA
) {
1864 NVC0LoweringPass
pass(prog
);
1865 return pass
.run(prog
, false, true);
1867 if (stage
== CG_STAGE_POST_RA
) {
1868 NVC0LegalizePostRA
pass(prog
);
1869 return pass
.run(prog
, false, true);
1871 if (stage
== CG_STAGE_SSA
) {
1872 NVC0LegalizeSSA pass
;
1873 return pass
.run(prog
, false, true);
1878 } // namespace nv50_ir