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_target_nv50.h"
27 Target
*getTargetNV50(unsigned int chipset
)
29 return new TargetNV50(chipset
);
32 TargetNV50::TargetNV50(unsigned int card
) : Target(true, true, false)
37 for (unsigned int i
= 0; i
<= SV_LAST
; ++i
)
38 sysvalLocation
[i
] = ~0;
44 // BULTINS / LIBRARY FUNCTIONS:
47 static const uint32_t nvc0_builtin_code
[] =
51 static const uint16_t nvc0_builtin_offsets
[NV50_BUILTIN_COUNT
] =
57 TargetNV50::getBuiltinCode(const uint32_t **code
, uint32_t *size
) const
64 TargetNV50::getBuiltinOffset(int builtin
) const
72 unsigned int mNeg
: 4;
73 unsigned int mAbs
: 4;
74 unsigned int mNot
: 4;
75 unsigned int mSat
: 4;
76 unsigned int fConst
: 3;
77 unsigned int fShared
: 3;
78 unsigned int fAttrib
: 3;
79 unsigned int fImm
: 3;
82 static const struct opProperties _initProps
[] =
84 // neg abs not sat c[] s[], a[], imm
85 { OP_ADD
, 0x3, 0x0, 0x0, 0x8, 0x2, 0x1, 0x1, 0x2 },
86 { OP_SUB
, 0x3, 0x0, 0x0, 0x8, 0x2, 0x1, 0x1, 0x2 },
87 { OP_MUL
, 0x3, 0x0, 0x0, 0x0, 0x2, 0x1, 0x1, 0x2 },
88 { OP_MAX
, 0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 },
89 { OP_MIN
, 0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 },
90 { OP_MAD
, 0x7, 0x0, 0x0, 0x8, 0x6, 0x1, 0x1, 0x0 }, // special constraint
91 { OP_ABS
, 0x0, 0x0, 0x0, 0x0, 0x0, 0x1, 0x1, 0x0 },
92 { OP_NEG
, 0x0, 0x1, 0x0, 0x0, 0x0, 0x1, 0x1, 0x0 },
93 { OP_CVT
, 0x1, 0x1, 0x0, 0x8, 0x0, 0x1, 0x1, 0x0 },
94 { OP_AND
, 0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 },
95 { OP_OR
, 0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 },
96 { OP_XOR
, 0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 },
97 { OP_SHL
, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x2 },
98 { OP_SHR
, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x2 },
99 { OP_SET
, 0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 },
100 { OP_PREEX2
, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
101 { OP_PRESIN
, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
102 { OP_EX2
, 0x0, 0x0, 0x0, 0x8, 0x0, 0x0, 0x0, 0x0 },
103 { OP_LG2
, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
104 { OP_RCP
, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
105 { OP_RSQ
, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
106 { OP_DFDX
, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
107 { OP_DFDY
, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
110 void TargetNV50::initOpInfo()
114 static const uint32_t commutative
[(OP_LAST
+ 31) / 32] =
116 // ADD, MUL, MAD, FMA, AND, OR, XOR, MAX, MIN, SET_AND, SET_OR, SET_XOR,
118 0x0ce0ca00, 0x0000007e, 0x00000000, 0x00000000
120 static const uint32_t shortForm
[(OP_LAST
+ 31) / 32] =
122 // MOV, ADD, SUB, MUL, MAD, SAD, RCP, L/PINTERP, TEX, TXF
123 0x00014e40, 0x00000080, 0x00001260, 0x00000000
125 static const operation noDestList
[] =
127 OP_STORE
, OP_WRSV
, OP_EXPORT
, OP_BRA
, OP_CALL
, OP_RET
, OP_EXIT
,
128 OP_DISCARD
, OP_CONT
, OP_BREAK
, OP_PRECONT
, OP_PREBREAK
, OP_PRERET
,
129 OP_JOIN
, OP_JOINAT
, OP_BRKPT
, OP_MEMBAR
, OP_EMIT
, OP_RESTART
,
130 OP_QUADON
, OP_QUADPOP
, OP_TEXBAR
, OP_SUSTB
, OP_SUSTP
, OP_SUREDP
,
133 static const operation noPredList
[] =
135 OP_CALL
, OP_PREBREAK
, OP_PRERET
, OP_QUADON
, OP_QUADPOP
, OP_JOINAT
,
139 for (i
= 0; i
< DATA_FILE_COUNT
; ++i
)
140 nativeFileMap
[i
] = (DataFile
)i
;
141 nativeFileMap
[FILE_PREDICATE
] = FILE_FLAGS
;
143 for (i
= 0; i
< OP_LAST
; ++i
) {
144 opInfo
[i
].variants
= NULL
;
145 opInfo
[i
].op
= (operation
)i
;
146 opInfo
[i
].srcTypes
= 1 << (int)TYPE_F32
;
147 opInfo
[i
].dstTypes
= 1 << (int)TYPE_F32
;
148 opInfo
[i
].immdBits
= 0xffffffff;
149 opInfo
[i
].srcNr
= operationSrcNr
[i
];
151 for (j
= 0; j
< opInfo
[i
].srcNr
; ++j
) {
152 opInfo
[i
].srcMods
[j
] = 0;
153 opInfo
[i
].srcFiles
[j
] = 1 << (int)FILE_GPR
;
155 opInfo
[i
].dstMods
= 0;
156 opInfo
[i
].dstFiles
= 1 << (int)FILE_GPR
;
158 opInfo
[i
].hasDest
= 1;
159 opInfo
[i
].vector
= (i
>= OP_TEX
&& i
<= OP_TEXCSAA
);
160 opInfo
[i
].commutative
= (commutative
[i
/ 32] >> (i
% 32)) & 1;
161 opInfo
[i
].pseudo
= (i
< OP_MOV
);
162 opInfo
[i
].predicate
= !opInfo
[i
].pseudo
;
163 opInfo
[i
].flow
= (i
>= OP_BRA
&& i
<= OP_JOIN
);
164 opInfo
[i
].minEncSize
= (shortForm
[i
/ 32] & (1 << (i
% 32))) ? 4 : 8;
166 for (i
= 0; i
< sizeof(noDestList
) / sizeof(noDestList
[0]); ++i
)
167 opInfo
[noDestList
[i
]].hasDest
= 0;
168 for (i
= 0; i
< sizeof(noPredList
) / sizeof(noPredList
[0]); ++i
)
169 opInfo
[noPredList
[i
]].predicate
= 0;
171 for (i
= 0; i
< sizeof(_initProps
) / sizeof(_initProps
[0]); ++i
) {
172 const struct opProperties
*prop
= &_initProps
[i
];
174 for (int s
= 0; s
< 3; ++s
) {
175 if (prop
->mNeg
& (1 << s
))
176 opInfo
[prop
->op
].srcMods
[s
] |= NV50_IR_MOD_NEG
;
177 if (prop
->mAbs
& (1 << s
))
178 opInfo
[prop
->op
].srcMods
[s
] |= NV50_IR_MOD_ABS
;
179 if (prop
->mNot
& (1 << s
))
180 opInfo
[prop
->op
].srcMods
[s
] |= NV50_IR_MOD_NOT
;
181 if (prop
->fConst
& (1 << s
))
182 opInfo
[prop
->op
].srcFiles
[s
] |= 1 << (int)FILE_MEMORY_CONST
;
183 if (prop
->fShared
& (1 << s
))
184 opInfo
[prop
->op
].srcFiles
[s
] |= 1 << (int)FILE_MEMORY_SHARED
;
185 if (prop
->fAttrib
& (1 << s
))
186 opInfo
[prop
->op
].srcFiles
[s
] |= 1 << (int)FILE_SHADER_INPUT
;
187 if (prop
->fImm
& (1 << s
))
188 opInfo
[prop
->op
].srcFiles
[s
] |= 1 << (int)FILE_IMMEDIATE
;
191 opInfo
[prop
->op
].dstMods
= NV50_IR_MOD_SAT
;
195 opInfo
[OP_MUL
].dstMods
= NV50_IR_MOD_SAT
;
199 TargetNV50::getFileSize(DataFile file
) const
202 case FILE_NULL
: return 0;
203 case FILE_GPR
: return 256; // in 16-bit units **
204 case FILE_PREDICATE
: return 0;
205 case FILE_FLAGS
: return 4;
206 case FILE_ADDRESS
: return 4;
207 case FILE_IMMEDIATE
: return 0;
208 case FILE_MEMORY_CONST
: return 65536;
209 case FILE_SHADER_INPUT
: return 0x200;
210 case FILE_SHADER_OUTPUT
: return 0x200;
211 case FILE_MEMORY_BUFFER
: return 0xffffffff;
212 case FILE_MEMORY_GLOBAL
: return 0xffffffff;
213 case FILE_MEMORY_SHARED
: return 16 << 10;
214 case FILE_MEMORY_LOCAL
: return 48 << 10;
215 case FILE_SYSTEM_VALUE
: return 16;
217 assert(!"invalid file");
220 // ** only first 128 units encodable for 16-bit regs
224 TargetNV50::getFileUnit(DataFile file
) const
226 if (file
== FILE_GPR
|| file
== FILE_ADDRESS
)
228 if (file
== FILE_SYSTEM_VALUE
)
234 TargetNV50::getSVAddress(DataFile shaderFile
, const Symbol
*sym
) const
236 switch (sym
->reg
.data
.sv
.sv
) {
241 uint32_t addr
= sysvalLocation
[sym
->reg
.data
.sv
.sv
];
242 for (int c
= 0; c
< sym
->reg
.data
.sv
.index
; ++c
)
243 if (wposMask
& (1 << c
))
247 case SV_PRIMITIVE_ID
:
248 return shaderFile
== FILE_SHADER_INPUT
? 0x18 :
249 sysvalLocation
[sym
->reg
.data
.sv
.sv
];
251 return 0x8 + 2 * sym
->reg
.data
.sv
.index
;
253 return 0xc + 2 * sym
->reg
.data
.sv
.index
;
255 return 0x2 + 2 * sym
->reg
.data
.sv
.index
;
259 return 0; /* sample position is handled differently */
261 return sysvalLocation
[sym
->reg
.data
.sv
.sv
];
265 // long: rrr, arr, rcr, acr, rrc, arc, gcr, grr
266 // short: rr, ar, rc, gr
269 TargetNV50::insnCanLoad(const Instruction
*i
, int s
,
270 const Instruction
*ld
) const
272 DataFile sf
= ld
->src(0).getFile();
274 // immediate 0 can be represented by GPR $r63/$r127
275 if (sf
== FILE_IMMEDIATE
&& ld
->getSrc(0)->reg
.data
.u64
== 0)
276 return (!i
->isPseudo() &&
278 i
->op
!= OP_EXPORT
&& i
->op
!= OP_STORE
);
280 if (sf
== FILE_IMMEDIATE
&& (i
->predSrc
>= 0 || i
->flagsDef
>= 0))
282 if (s
>= opInfo
[i
->op
].srcNr
)
284 if (!(opInfo
[i
->op
].srcFiles
[s
] & (1 << (int)sf
)))
286 if (s
== 2 && i
->src(1).getFile() != FILE_GPR
)
289 // NOTE: don't rely on flagsDef
290 if (sf
== FILE_IMMEDIATE
)
291 for (int d
= 0; i
->defExists(d
); ++d
)
292 if (i
->def(d
).getFile() == FILE_FLAGS
)
297 for (int z
= 0; z
< Target::operationSrcNr
[i
->op
]; ++z
) {
298 DataFile zf
= (z
== s
) ? sf
: i
->src(z
).getFile();
302 case FILE_MEMORY_SHARED
:
303 case FILE_SHADER_INPUT
:
304 mode
|= 1 << (z
* 2);
306 case FILE_MEMORY_CONST
:
307 mode
|= 2 << (z
* 2);
310 mode
|= 3 << (z
* 2);
326 // Shader inputs get transformed to p[] in geometry shaders, and those
327 // aren't allowed to be used at the same time as c[].
328 if (ld
->bb
->getProgram()->getType() == Program::TYPE_GEOMETRY
)
332 if (ld
->bb
->getProgram()->getType() != Program::TYPE_GEOMETRY
)
341 if ((i
->op
== OP_MUL
|| i
->op
== OP_MAD
) && !isFloatType(i
->dType
)) {
342 // 32-bit MUL will be split into 16-bit MULs
343 if (ld
->src(0).isIndirect(0))
345 if (sf
== FILE_IMMEDIATE
)
347 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
&& sf
== FILE_MEMORY_CONST
)
351 ldSize
= typeSizeof(ld
->dType
);
354 if (sf
== FILE_IMMEDIATE
)
358 // Check if memory access is encodable:
360 if (ldSize
< 4 && sf
== FILE_SHADER_INPUT
) // no < 4-byte aligned a[] access
362 if (ld
->getSrc(0)->reg
.data
.offset
> (int32_t)(127 * ldSize
))
365 if (ld
->src(0).isIndirect(0)) {
366 for (int z
= 0; i
->srcExists(z
); ++z
)
367 if (i
->src(z
).isIndirect(0))
370 // s[] access only possible in CP, $aX always applies
371 if (sf
== FILE_MEMORY_SHARED
)
373 if (!ld
->bb
) // can't check type ...
375 Program::Type pt
= ld
->bb
->getProgram()->getType();
377 // $aX applies to c[] only in VP, FP, GP if p[] is not accessed
378 if (pt
== Program::TYPE_COMPUTE
)
380 if (pt
== Program::TYPE_GEOMETRY
) {
381 if (sf
== FILE_MEMORY_CONST
)
382 return i
->src(s
).getFile() != FILE_SHADER_INPUT
;
383 return sf
== FILE_SHADER_INPUT
;
385 return sf
== FILE_MEMORY_CONST
;
391 TargetNV50::insnCanLoadOffset(const Instruction
*i
, int s
, int offset
) const
393 if (!i
->src(s
).isIndirect(0))
395 offset
+= i
->src(s
).get()->reg
.data
.offset
;
396 if (i
->op
== OP_LOAD
|| i
->op
== OP_STORE
) {
397 // There are some restrictions in theory, but in practice they're never
398 // going to be hit. When we enable shared/global memory, this will
399 // become more important.
402 return offset
>= 0 && offset
<= (int32_t)(127 * i
->src(s
).get()->reg
.size
);
406 TargetNV50::isAccessSupported(DataFile file
, DataType ty
) const
408 if (ty
== TYPE_B96
|| ty
== TYPE_NONE
)
410 if (typeSizeof(ty
) > 4)
411 return (file
== FILE_MEMORY_LOCAL
) || (file
== FILE_MEMORY_GLOBAL
) ||
412 (file
== FILE_MEMORY_BUFFER
);
417 TargetNV50::isOpSupported(operation op
, DataType ty
) const
419 if (ty
== TYPE_F64
&& chipset
< 0xa0)
424 return chipset
>= 0xa0;
426 return chipset
>= 0xa3 && chipset
!= 0xaa && chipset
!= 0xac;
439 case OP_EXIT
: // want exit modifier instead (on NOP if required)
444 return ty
== TYPE_S32
;
446 return !isFloatType(ty
);
453 TargetNV50::isModSupported(const Instruction
*insn
, int s
, Modifier mod
) const
455 if (!isFloatType(insn
->dType
)) {
468 if (insn
->src(s
? 0 : 1).mod
.neg())
473 return insn
->src(1).mod
.neg() ? false : true;
476 if (insn
->sType
!= TYPE_F32
)
483 if (s
>= opInfo
[insn
->op
].srcNr
|| s
>= 3)
485 return (mod
& Modifier(opInfo
[insn
->op
].srcMods
[s
])) == mod
;
489 TargetNV50::mayPredicate(const Instruction
*insn
, const Value
*pred
) const
491 if (insn
->getPredicate() || insn
->flagsSrc
>= 0)
493 for (int s
= 0; insn
->srcExists(s
); ++s
)
494 if (insn
->src(s
).getFile() == FILE_IMMEDIATE
)
496 return opInfo
[insn
->op
].predicate
;
500 TargetNV50::isSatSupported(const Instruction
*insn
) const
502 if (insn
->op
== OP_CVT
)
504 if (insn
->dType
!= TYPE_F32
)
506 return opInfo
[insn
->op
].dstMods
& NV50_IR_MOD_SAT
;
509 int TargetNV50::getLatency(const Instruction
*i
) const
511 // TODO: tune these values
512 if (i
->op
== OP_LOAD
) {
513 switch (i
->src(0).getFile()) {
514 case FILE_MEMORY_LOCAL
:
515 case FILE_MEMORY_GLOBAL
:
516 case FILE_MEMORY_BUFFER
:
517 return 100; // really 400 to 800
525 // These are "inverse" throughput values, i.e. the number of cycles required
526 // to issue a specific instruction for a full warp (32 threads).
528 // Assuming we have more than 1 warp in flight, a higher issue latency results
529 // in a lower result latency since the MP will have spent more time with other
531 // This also helps to determine the number of cycles between instructions in
534 int TargetNV50::getThroughput(const Instruction
*i
) const
536 // TODO: tune these values
537 if (i
->dType
== TYPE_F32
) {
551 if (i
->dType
== TYPE_U32
|| i
->dType
== TYPE_S32
) {
554 if (i
->dType
== TYPE_F64
) {
562 recordLocation(uint16_t *locs
, uint8_t *masks
,
563 const struct nv50_ir_varying
*var
)
565 uint16_t addr
= var
->slot
[0] * 4;
568 case TGSI_SEMANTIC_POSITION
: locs
[SV_POSITION
] = addr
; break;
569 case TGSI_SEMANTIC_INSTANCEID
: locs
[SV_INSTANCE_ID
] = addr
; break;
570 case TGSI_SEMANTIC_VERTEXID
: locs
[SV_VERTEX_ID
] = addr
; break;
571 case TGSI_SEMANTIC_PRIMID
: locs
[SV_PRIMITIVE_ID
] = addr
; break;
572 case TGSI_SEMANTIC_LAYER
: locs
[SV_LAYER
] = addr
; break;
573 case TGSI_SEMANTIC_VIEWPORT_INDEX
: locs
[SV_VIEWPORT_INDEX
] = addr
; break;
577 if (var
->sn
== TGSI_SEMANTIC_POSITION
&& masks
)
578 masks
[0] = var
->mask
;
582 TargetNV50::parseDriverInfo(const struct nv50_ir_prog_info
*info
)
585 for (i
= 0; i
< info
->numOutputs
; ++i
)
586 recordLocation(sysvalLocation
, NULL
, &info
->out
[i
]);
587 for (i
= 0; i
< info
->numInputs
; ++i
)
588 recordLocation(sysvalLocation
, &wposMask
, &info
->in
[i
]);
589 for (i
= 0; i
< info
->numSysVals
; ++i
)
590 recordLocation(sysvalLocation
, NULL
, &info
->sv
[i
]);
592 if (sysvalLocation
[SV_POSITION
] >= 0x200) {
593 // not assigned by driver, but we need it internally
595 sysvalLocation
[SV_POSITION
] = 0;
598 Target::parseDriverInfo(info
);
601 } // namespace nv50_ir