xref: /llvm-project/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp (revision 66e41a1a20f2190a800669028a0e80bd86e735ce)
1 //===- NVVMDialect.cpp - NVVM IR Ops and Dialect registration -------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the types and operation details for the NVVM IR dialect in
10 // MLIR, and the LLVM IR dialect.  It also registers the dialect.
11 //
12 // The NVVM dialect only contains GPU specific additions on top of the general
13 // LLVM dialect.
14 //
15 //===----------------------------------------------------------------------===//
16 
17 #include "mlir/Dialect/LLVMIR/NVVMDialect.h"
18 
19 #include "mlir/Conversion/ConvertToLLVM/ToLLVMInterface.h"
20 #include "mlir/Dialect/GPU/IR/CompilationInterfaces.h"
21 #include "mlir/Dialect/Utils/StaticValueUtils.h"
22 #include "mlir/IR/Builders.h"
23 #include "mlir/IR/BuiltinAttributes.h"
24 #include "mlir/IR/BuiltinTypes.h"
25 #include "mlir/IR/Diagnostics.h"
26 #include "mlir/IR/DialectImplementation.h"
27 #include "mlir/IR/MLIRContext.h"
28 #include "mlir/IR/Operation.h"
29 #include "mlir/IR/OperationSupport.h"
30 #include "mlir/IR/Types.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/TypeSwitch.h"
33 #include "llvm/AsmParser/Parser.h"
34 #include "llvm/IR/Attributes.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/Type.h"
37 #include "llvm/Support/Casting.h"
38 #include "llvm/Support/SourceMgr.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include <cassert>
41 #include <optional>
42 #include <string>
43 
44 using namespace mlir;
45 using namespace NVVM;
46 
47 #include "mlir/Dialect/LLVMIR/NVVMOpsDialect.cpp.inc"
48 #include "mlir/Dialect/LLVMIR/NVVMOpsEnums.cpp.inc"
49 
50 //===----------------------------------------------------------------------===//
51 // Printing/parsing for NVVM ops
52 //===----------------------------------------------------------------------===//
53 
54 static void printNVVMIntrinsicOp(OpAsmPrinter &p, Operation *op) {
55   p << " " << op->getOperands();
56   if (op->getNumResults() > 0)
57     p << " : " << op->getResultTypes();
58 }
59 
60 // <operation> ::= `llvm.nvvm.vote.ballot.sync %mask, %pred` : result_type
61 ParseResult VoteBallotOp::parse(OpAsmParser &parser, OperationState &result) {
62   MLIRContext *context = parser.getContext();
63   auto int32Ty = IntegerType::get(context, 32);
64   auto int1Ty = IntegerType::get(context, 1);
65 
66   SmallVector<OpAsmParser::UnresolvedOperand, 8> ops;
67   Type type;
68   return failure(parser.parseOperandList(ops) ||
69                  parser.parseOptionalAttrDict(result.attributes) ||
70                  parser.parseColonType(type) ||
71                  parser.addTypeToList(type, result.types) ||
72                  parser.resolveOperands(ops, {int32Ty, int1Ty},
73                                         parser.getNameLoc(), result.operands));
74 }
75 
76 void VoteBallotOp::print(OpAsmPrinter &p) { printNVVMIntrinsicOp(p, *this); }
77 
78 // This verifier is shared among the following Ops:
79 // CpAsyncBulkTensorGlobalToSharedClusterOp (TMA Load)
80 // CpAsyncBulkTensorPrefetchOp (TMA Prefetch)
81 // CpAsyncBulkTensorReduceOp (TMA Store-Reduce)
82 static LogicalResult CpAsyncBulkTensorCommonVerifier(size_t tensorDims,
83                                                      bool isIm2Col,
84                                                      size_t numIm2ColOffsets,
85                                                      Location loc) {
86   if (tensorDims < 1 || tensorDims > 5)
87     return emitError(loc, "expects coordinates between 1 to 5 dimension");
88 
89   // For Im2Col mode, there are two constraints:
90   if (isIm2Col) {
91     // 1. Tensor must always be at least 3-d.
92     if (tensorDims < 3)
93       return emitError(
94           loc,
95           "to use im2col mode, the tensor has to be at least 3-dimensional");
96     // 2. When there are Im2ColOffsets, they must be (Dims - 2) in number.
97     if (numIm2ColOffsets && (tensorDims != (numIm2ColOffsets + 2)))
98       return emitError(
99           loc, "im2col offsets must be 2 less than number of coordinates");
100   }
101   return success();
102 }
103 
104 LogicalResult CpAsyncBulkTensorGlobalToSharedClusterOp::verify() {
105   size_t numIm2ColOffsets = getIm2colOffsets().size();
106   bool isIm2Col = numIm2ColOffsets > 0;
107   return CpAsyncBulkTensorCommonVerifier(getCoordinates().size(), isIm2Col,
108                                          numIm2ColOffsets, getLoc());
109 }
110 
111 LogicalResult CpAsyncBulkTensorSharedCTAToGlobalOp::verify() {
112   if (getCoordinates().size() > 5)
113     return emitError("Maximum 5 coordinates and dimension is supported.");
114   return success();
115 }
116 
117 LogicalResult CpAsyncOp::verify() {
118   if (getModifier() != LoadCacheModifierKind::CG &&
119       getModifier() != LoadCacheModifierKind::CA)
120     return emitError("Only CG and CA cache modifiers are supported.");
121   if (getSize() != 4 && getSize() != 8 && getSize() != 16)
122     return emitError("expected byte size to be either 4, 8 or 16.");
123   if (getModifier() == LoadCacheModifierKind::CG && getSize() != 16)
124     return emitError("CG cache modifier is only support for 16 bytes copy.");
125   return success();
126 }
127 
128 LogicalResult CpAsyncBulkTensorPrefetchOp::verify() {
129   size_t numIm2ColOffsets = getIm2colOffsets().size();
130   bool isIm2Col = numIm2ColOffsets > 0;
131   return CpAsyncBulkTensorCommonVerifier(getCoordinates().size(), isIm2Col,
132                                          numIm2ColOffsets, getLoc());
133 }
134 
135 LogicalResult CpAsyncBulkTensorReduceOp::verify() {
136   bool isIm2Col = (getMode() == TMAStoreMode::IM2COL);
137   return CpAsyncBulkTensorCommonVerifier(getCoordinates().size(), isIm2Col, 0,
138                                          getLoc());
139 }
140 
141 // Given the element type of an operand and whether or not it is an accumulator,
142 // this function returns the PTX type (`NVVM::MMATypes`) that corresponds to the
143 // operand's element type.
144 std::optional<mlir::NVVM::MMATypes>
145 MmaOp::inferOperandMMAType(Type operandElType, bool isAccumulator) {
146   auto half2Type =
147       LLVM::getFixedVectorType(Float16Type::get(operandElType.getContext()), 2);
148   if (operandElType.isF64())
149     return NVVM::MMATypes::f64;
150   if (operandElType.isF16() || operandElType == half2Type)
151     return NVVM::MMATypes::f16;
152   if (operandElType.isF32() && isAccumulator)
153     return NVVM::MMATypes::f32;
154   if (operandElType.isF32() && !isAccumulator)
155     return NVVM::MMATypes::tf32;
156   if (llvm::isa<IntegerType>(operandElType)) {
157     if (isAccumulator)
158       return NVVM::MMATypes::s32;
159     return std::nullopt;
160   }
161 
162   if (auto structType = llvm::dyn_cast<LLVM::LLVMStructType>(operandElType)) {
163     if (structType.getBody().empty())
164       return std::nullopt;
165     return inferOperandMMAType(structType.getBody()[0], isAccumulator);
166   }
167 
168   return std::nullopt;
169 }
170 
171 static bool isInt4PtxType(MMATypes type) {
172   return (type == MMATypes::u4 || type == MMATypes::s4);
173 }
174 
175 static bool isInt8PtxType(MMATypes type) {
176   return (type == MMATypes::u8 || type == MMATypes::s8);
177 }
178 
179 static bool isIntegerPtxType(MMATypes type) {
180   return isInt4PtxType(type) || isInt8PtxType(type) || type == MMATypes::b1 ||
181          type == MMATypes::s32;
182 }
183 
184 MMATypes MmaOp::accumPtxType() {
185   std::optional<mlir::NVVM::MMATypes> val = inferOperandMMAType(
186       getODSOperands(2).getTypes().front(), /*isAccum=*/true);
187   assert(val.has_value() && "accumulator PTX type should always be inferrable");
188   return val.value();
189 }
190 
191 MMATypes MmaOp::resultPtxType() {
192   std::optional<mlir::NVVM::MMATypes> val =
193       inferOperandMMAType(getResult().getType(), /*isAccum=*/true);
194   assert(val.has_value() && "result PTX type should always be inferrable");
195   return val.value();
196 }
197 
198 void MmaOp::print(OpAsmPrinter &p) {
199   SmallVector<Type, 4> regTypes;
200   struct OperandFragment {
201     StringRef operandName;
202     StringRef ptxTypeAttr;
203     SmallVector<Value, 4> regs;
204     explicit OperandFragment(StringRef name, StringRef ptxTypeName)
205         : operandName(name), ptxTypeAttr(ptxTypeName) {}
206   };
207 
208   std::array<OperandFragment, 3> frags{
209       OperandFragment("A", getMultiplicandAPtxTypeAttrName()),
210       OperandFragment("B", getMultiplicandBPtxTypeAttrName()),
211       OperandFragment("C", "")};
212   SmallVector<StringRef, 4> ignoreAttrNames{
213       mlir::NVVM::MmaOp::getOperandSegmentSizeAttr()};
214 
215   for (unsigned fragIdx = 0; fragIdx < frags.size(); fragIdx++) {
216     auto &frag = frags[fragIdx];
217     auto varOperandSpec = getODSOperandIndexAndLength(fragIdx);
218     for (auto operandIdx = varOperandSpec.first;
219          operandIdx < varOperandSpec.first + varOperandSpec.second;
220          operandIdx++) {
221       frag.regs.push_back(this->getOperand(operandIdx));
222       if (operandIdx == 0) {
223         regTypes.push_back(this->getOperand(operandIdx).getType());
224       }
225     }
226     std::optional<MMATypes> inferredType =
227         inferOperandMMAType(regTypes.back(), /*isAccum=*/fragIdx >= 2);
228     if (inferredType)
229       ignoreAttrNames.push_back(frag.ptxTypeAttr);
230   }
231 
232   auto printMmaOperand = [&](const OperandFragment &frag) -> void {
233     p << " " << frag.operandName;
234     p << "[";
235     p.printOperands(frag.regs);
236     p << "] ";
237   };
238 
239   for (const auto &frag : frags) {
240     printMmaOperand(frag);
241   }
242 
243   p.printOptionalAttrDict(this->getOperation()->getAttrs(), ignoreAttrNames);
244 
245   // Print the types of the operands and result.
246   p << " : " << "(";
247   llvm::interleaveComma(SmallVector<Type, 3>{frags[0].regs[0].getType(),
248                                              frags[1].regs[0].getType(),
249                                              frags[2].regs[0].getType()},
250                         p);
251   p << ")";
252   p.printArrowTypeList(TypeRange{this->getRes().getType()});
253 }
254 
255 void MmaOp::build(OpBuilder &builder, OperationState &result, Type resultType,
256                   ValueRange operandA, ValueRange operandB, ValueRange operandC,
257                   ArrayRef<int64_t> shape, std::optional<MMAB1Op> b1Op,
258                   std::optional<MMAIntOverflow> intOverflow,
259                   std::optional<std::array<MMATypes, 2>> multiplicandPtxTypes,
260                   std::optional<std::array<MMALayout, 2>> multiplicandLayouts) {
261 
262   assert(shape.size() == 3 && "expected shape to have size 3 (m, n, k)");
263   MLIRContext *ctx = builder.getContext();
264   result.addAttribute(
265       "shape", builder.getAttr<MMAShapeAttr>(shape[0], shape[1], shape[2]));
266 
267   result.addOperands(operandA);
268   result.addOperands(operandB);
269   result.addOperands(operandC);
270 
271   if (multiplicandPtxTypes) {
272     result.addAttribute("multiplicandAPtxType",
273                         MMATypesAttr::get(ctx, (*multiplicandPtxTypes)[0]));
274     result.addAttribute("multiplicandBPtxType",
275                         MMATypesAttr::get(ctx, (*multiplicandPtxTypes)[1]));
276   } else {
277     if (auto res = inferOperandMMAType(operandA[0].getType(), false))
278       result.addAttribute("multiplicandAPtxType", MMATypesAttr::get(ctx, *res));
279     if (auto res = inferOperandMMAType(operandB[0].getType(), false))
280       result.addAttribute("multiplicandBPtxType", MMATypesAttr::get(ctx, *res));
281   }
282 
283   if (multiplicandLayouts) {
284     result.addAttribute("layoutA",
285                         MMALayoutAttr::get(ctx, (*multiplicandLayouts)[0]));
286     result.addAttribute("layoutB",
287                         MMALayoutAttr::get(ctx, (*multiplicandLayouts)[1]));
288   } else {
289     result.addAttribute("layoutA", MMALayoutAttr::get(ctx, MMALayout::row));
290     result.addAttribute("layoutB", MMALayoutAttr::get(ctx, MMALayout::col));
291   }
292 
293   if (intOverflow.has_value())
294     result.addAttribute("intOverflowBehavior",
295                         MMAIntOverflowAttr::get(ctx, *intOverflow));
296   if (b1Op.has_value())
297     result.addAttribute("b1Op", MMAB1OpAttr::get(ctx, *b1Op));
298 
299   result.addTypes(resultType);
300   result.addAttribute(
301       MmaOp::getOperandSegmentSizeAttr(),
302       builder.getDenseI32ArrayAttr({static_cast<int32_t>(operandA.size()),
303                                     static_cast<int32_t>(operandB.size()),
304                                     static_cast<int32_t>(operandC.size())}));
305 }
306 
307 // <operation> :=
308 //   A `[` $operandA `]` B `[` $operandB `]` C `[` $operandC `]`
309 //   attr-dict : (type($operandA[0]), type($operandB[0]), type($operandC[0]))
310 //     `->` type($res)
311 ParseResult MmaOp::parse(OpAsmParser &parser, OperationState &result) {
312   struct OperandFragment {
313     std::optional<MMATypes> elemtype;
314     SmallVector<OpAsmParser::UnresolvedOperand, 4> regs;
315     SmallVector<Type> regTypes;
316   };
317 
318   Builder &builder = parser.getBuilder();
319   std::array<OperandFragment, 4> frags;
320 
321   NamedAttrList namedAttributes;
322 
323   // A helper to parse the operand segments.
324   auto parseMmaOperand = [&](StringRef operandName,
325                              OperandFragment &frag) -> LogicalResult {
326     if (parser.parseKeyword(operandName).failed())
327       return failure();
328     if (parser
329             .parseOperandList(frag.regs, OpAsmParser::Delimiter::OptionalSquare)
330             .failed())
331       return failure();
332     return success();
333   };
334 
335   // Parse the operand segments.
336   if (parseMmaOperand("A", frags[0]).failed())
337     return failure();
338   if (parseMmaOperand("B", frags[1]).failed())
339     return failure();
340   if (parseMmaOperand("C", frags[2]).failed())
341     return failure();
342 
343   if (parser.parseOptionalAttrDict(namedAttributes).failed())
344     return failure();
345 
346   // Parse the type specification and resolve operands.
347   SmallVector<Type, 3> operandTypes;
348   if (failed(parser.parseColon()))
349     return failure();
350   if (failed(parser.parseLParen()))
351     return failure();
352   if (failed(parser.parseTypeList(operandTypes)))
353     return failure();
354   if (failed(parser.parseRParen()))
355     if (operandTypes.size() != 3)
356       return parser.emitError(
357           parser.getNameLoc(),
358           "expected one type for each operand segment but got " +
359               Twine(operandTypes.size()) + " types");
360   for (const auto &iter : llvm::enumerate(operandTypes)) {
361     auto &frag = frags[iter.index()];
362     frag.regTypes.resize(frag.regs.size(), iter.value());
363     if (failed(parser.resolveOperands(frag.regs, frag.regTypes,
364                                       parser.getNameLoc(), result.operands)))
365       return failure();
366     frag.elemtype =
367         inferOperandMMAType(frag.regTypes[0], /*isAccum=*/iter.index() < 2);
368   }
369 
370   Type resultType;
371   if (parser.parseArrow() || parser.parseType(resultType))
372     return failure();
373   frags[3].elemtype = inferOperandMMAType(resultType, /*isAccum=*/true);
374 
375   std::array<StringRef, 2> names{"multiplicandAPtxType",
376                                  "multiplicandBPtxType"};
377   for (unsigned idx = 0; idx < names.size(); idx++) {
378     const auto &frag = frags[idx];
379     std::optional<NamedAttribute> attr = namedAttributes.getNamed(names[idx]);
380     if (!frag.elemtype.has_value() && !attr.has_value()) {
381       return parser.emitError(
382           parser.getNameLoc(),
383           "attribute " + names[idx] +
384               " is not provided explicitly and cannot be inferred");
385     }
386     if (!attr.has_value())
387       result.addAttribute(
388           names[idx], MMATypesAttr::get(parser.getContext(), *frag.elemtype));
389   }
390 
391   result.addTypes(resultType);
392   if (!namedAttributes.empty())
393     result.addAttributes(namedAttributes);
394   result.addAttribute(MmaOp::getOperandSegmentSizeAttr(),
395                       builder.getDenseI32ArrayAttr({
396                           static_cast<int32_t>(frags[0].regs.size()),
397                           static_cast<int32_t>(frags[1].regs.size()),
398                           static_cast<int32_t>(frags[2].regs.size()),
399                       }));
400   return success();
401 }
402 
403 LogicalResult MmaOp::verify() {
404   MLIRContext *context = getContext();
405   auto f16Ty = Float16Type::get(context);
406   auto i32Ty = IntegerType::get(context, 32);
407   auto f16x2Ty = LLVM::getFixedVectorType(f16Ty, 2);
408   auto f32Ty = Float32Type::get(context);
409   auto f16x2x4StructTy = LLVM::LLVMStructType::getLiteral(
410       context, {f16x2Ty, f16x2Ty, f16x2Ty, f16x2Ty});
411 
412   auto s32x4StructTy =
413       LLVM::LLVMStructType::getLiteral(context, {i32Ty, i32Ty, i32Ty, i32Ty});
414   auto f32x8StructTy =
415       LLVM::LLVMStructType::getLiteral(context, SmallVector<Type>(8, f32Ty));
416   auto f16x2x2StructTy =
417       LLVM::LLVMStructType::getLiteral(context, {f16x2Ty, f16x2Ty});
418   auto f32x4StructTy =
419       LLVM::LLVMStructType::getLiteral(context, {f32Ty, f32Ty, f32Ty, f32Ty});
420   auto s32x2StructTy =
421       LLVM::LLVMStructType::getLiteral(context, {i32Ty, i32Ty});
422 
423   std::array<int64_t, 3> mmaShape{getShapeAttr().getM(), getShapeAttr().getN(),
424                                   getShapeAttr().getK()};
425 
426   // These variables define the set of allowed data types for matrices A, B, C,
427   // and result.
428   using AllowedShapes = SmallVector<std::array<int64_t, 3>, 2>;
429   using AllowedTypes = SmallVector<SmallVector<Type, 4>, 2>;
430   AllowedShapes allowedShapes;
431   AllowedTypes expectedA;
432   AllowedTypes expectedB;
433   AllowedTypes expectedC;
434   SmallVector<Type> expectedResult;
435 
436   // When M = 16, we just need to calculate the number of 8xk tiles, where
437   // k is a factor that depends on the data type.
438   if (mmaShape[0] == 16) {
439     int64_t kFactor;
440     Type multiplicandFragType;
441     switch (*getMultiplicandAPtxType()) {
442     case MMATypes::tf32:
443       kFactor = 4;
444       multiplicandFragType = i32Ty;
445       expectedResult.push_back(LLVM::LLVMStructType::getLiteral(
446           context, {f32Ty, f32Ty, f32Ty, f32Ty}));
447       break;
448     case MMATypes::f16:
449     case MMATypes::bf16:
450       kFactor = 8;
451       multiplicandFragType = f16x2Ty;
452       expectedResult.push_back(f16x2x2StructTy);
453       expectedResult.push_back(f32x4StructTy);
454       break;
455     case MMATypes::s4:
456     case MMATypes::u4:
457       kFactor = 32;
458       break;
459     case MMATypes::b1:
460       kFactor = 128;
461       break;
462     case MMATypes::s8:
463     case MMATypes::u8:
464       kFactor = 16;
465       break;
466     default:
467       return emitError("invalid shape or multiplicand type: " +
468                        stringifyEnum(getMultiplicandAPtxType().value()));
469     }
470 
471     if (isIntegerPtxType(getMultiplicandAPtxType().value())) {
472       expectedResult.push_back(s32x4StructTy);
473       expectedC.emplace_back(4, i32Ty);
474       multiplicandFragType = i32Ty;
475     } else {
476       expectedC.emplace_back(2, f16x2Ty);
477       expectedC.emplace_back(4, f32Ty);
478     }
479 
480     int64_t unitA = (mmaShape[0] / 8) * (mmaShape[2] / kFactor);
481     int64_t unitB = (mmaShape[1] / 8) * (mmaShape[2] / kFactor);
482     expectedA.emplace_back(unitA, multiplicandFragType);
483     expectedB.emplace_back(unitB, multiplicandFragType);
484     allowedShapes.push_back({16, 8, kFactor});
485     allowedShapes.push_back({16, 8, kFactor * 2});
486   }
487 
488   // In the M=8 case, there is only 1 possible case per data type.
489   if (mmaShape[0] == 8) {
490     if (*getMultiplicandAPtxType() == MMATypes::f16) {
491       expectedA.emplace_back(2, f16x2Ty);
492       expectedB.emplace_back(2, f16x2Ty);
493       expectedResult.push_back(f16x2x4StructTy);
494       expectedResult.push_back(f32x8StructTy);
495       expectedC.emplace_back(4, f16x2Ty);
496       expectedC.emplace_back(8, f32Ty);
497       allowedShapes.push_back({8, 8, 4});
498     }
499     if (*getMultiplicandAPtxType() == MMATypes::f64) {
500       Type f64Ty = Float64Type::get(context);
501       expectedA.emplace_back(1, f64Ty);
502       expectedB.emplace_back(1, f64Ty);
503       expectedC.emplace_back(2, f64Ty);
504       // expectedC.emplace_back(1, LLVM::getFixedVectorType(f64Ty, 2));
505       expectedResult.emplace_back(LLVM::LLVMStructType::getLiteral(
506           context, SmallVector<Type>(2, f64Ty)));
507       allowedShapes.push_back({8, 8, 4});
508     }
509     if (isIntegerPtxType(getMultiplicandAPtxType().value())) {
510       expectedA.push_back({i32Ty});
511       expectedB.push_back({i32Ty});
512       expectedC.push_back({i32Ty, i32Ty});
513       expectedResult.push_back(s32x2StructTy);
514       if (isInt4PtxType(getMultiplicandAPtxType().value()))
515         allowedShapes.push_back({8, 8, 32});
516       if (isInt8PtxType(getMultiplicandAPtxType().value()))
517         allowedShapes.push_back({8, 8, 16});
518       if (getMultiplicandAPtxType().value() == MMATypes::b1)
519         allowedShapes.push_back({8, 8, 128});
520     }
521   }
522 
523   std::string errorMessage;
524   llvm::raw_string_ostream errorStream(errorMessage);
525 
526   // Check that we matched an existing shape/dtype combination.
527   if (expectedA.empty() || expectedB.empty() || expectedC.empty() ||
528       !llvm::is_contained(allowedShapes, mmaShape)) {
529     errorStream << "unimplemented variant for MMA shape <";
530     llvm::interleaveComma(mmaShape, errorStream);
531     errorStream << ">";
532     return emitOpError(errorMessage);
533   }
534 
535   // Verify the operand types for segments of A, B, and C operands.
536   std::array<StringRef, 3> operandNames{"A", "B", "C"};
537   for (const auto &iter : llvm::enumerate(
538            SmallVector<AllowedTypes, 3>{expectedA, expectedB, expectedC})) {
539     auto spec = this->getODSOperandIndexAndLength(iter.index());
540     SmallVector<Type, 4> operandTySeg(operand_type_begin() + spec.first,
541                                       operand_type_begin() + spec.first +
542                                           spec.second);
543     bool match = llvm::is_contained(iter.value(), operandTySeg);
544 
545     if (!match) {
546       errorStream << "Could not match types for the "
547                   << operandNames[iter.index()]
548                   << " operands; expected one of ";
549       for (const auto &x : iter.value()) {
550         errorStream << x.size() << "x" << x[0] << " ";
551       }
552       errorStream << "but got ";
553       llvm::interleaveComma(operandTySeg, errorStream);
554       return emitOpError(errorMessage);
555     }
556   }
557 
558   // Check the result type
559   if (!llvm::any_of(expectedResult, [&](Type expectedResultType) {
560         return expectedResultType == getResult().getType();
561       })) {
562     errorStream
563         << "Could not match allowed types for the result; expected one of ";
564     llvm::interleaveComma(expectedResult, errorStream);
565     errorStream << " but got " << getResult().getType();
566     return emitOpError(errorMessage);
567   }
568 
569   // Ensure that binary MMA variants have a b1 MMA operation defined.
570   if (getMultiplicandAPtxType() == MMATypes::b1 && !getB1Op()) {
571     return emitOpError("op requires " + getB1OpAttrName().strref() +
572                        " attribute");
573   }
574 
575   // Ensure int4/int8 MMA variants specify the accum overflow behavior
576   // attribute.
577   if (isInt4PtxType(*getMultiplicandAPtxType()) ||
578       isInt8PtxType(*getMultiplicandAPtxType())) {
579     if (!getIntOverflowBehavior())
580       return emitOpError("op requires " +
581                          getIntOverflowBehaviorAttrName().strref() +
582                          " attribute");
583   }
584 
585   return success();
586 }
587 
588 LogicalResult ShflOp::verify() {
589   if (!(*this)->getAttrOfType<UnitAttr>("return_value_and_is_valid"))
590     return success();
591   auto type = llvm::dyn_cast<LLVM::LLVMStructType>(getType());
592   auto elementType = (type && type.getBody().size() == 2)
593                          ? llvm::dyn_cast<IntegerType>(type.getBody()[1])
594                          : nullptr;
595   if (!elementType || elementType.getWidth() != 1)
596     return emitError("expected return type to be a two-element struct with "
597                      "i1 as the second element");
598   return success();
599 }
600 
601 std::pair<mlir::Type, unsigned> NVVM::inferMMAType(NVVM::MMATypes type,
602                                                    NVVM::MMAFrag frag, int nRow,
603                                                    int nCol,
604                                                    MLIRContext *context) {
605   unsigned numberElements = 0;
606   Type elementType;
607   OpBuilder builder(context);
608   Type f16x2 = VectorType::get(2, builder.getF16Type());
609   if (type == NVVM::MMATypes::f16) {
610     elementType = f16x2;
611     if (frag == NVVM::MMAFrag::a || frag == NVVM::MMAFrag::b)
612       numberElements = 8;
613     else
614       numberElements = 4;
615   } else if (type == NVVM::MMATypes::f32) {
616     elementType = builder.getF32Type();
617     numberElements = 8;
618   } else if (type == NVVM::MMATypes::tf32) {
619     elementType = builder.getI32Type();
620     numberElements = 4;
621   } else if (type == NVVM::MMATypes::s8 || type == NVVM::MMATypes::u8) {
622     elementType = builder.getI32Type();
623     int parallelSize = 0;
624     if (frag == NVVM::MMAFrag::a)
625       parallelSize = nRow;
626     if (frag == NVVM::MMAFrag::b)
627       parallelSize = nCol;
628 
629     // m == 16 && n == 16 && k == 16
630     if (parallelSize == 16)
631       numberElements = 2;
632     // m == 8 && n == 32 && k == 16 or m == 32 && n == 8 && k == 16
633     else if (parallelSize == 8)
634       numberElements = 1;
635     else if (parallelSize == 32)
636       numberElements = 4;
637   } else if (type == NVVM::MMATypes::s32) {
638     elementType = builder.getI32Type();
639     numberElements = 8;
640   }
641   assert(numberElements != 0 && elementType != nullptr);
642   return std::make_pair(elementType, numberElements);
643 }
644 
645 static std::pair<mlir::Type, unsigned>
646 inferMMATypeFromMNK(NVVM::MMATypes type, NVVM::MMAFrag frag, int m, int n,
647                     int k, MLIRContext *context) {
648   int nRow, nCol;
649   if (frag == NVVM::MMAFrag::a) {
650     nRow = m;
651     nCol = k;
652   } else if (frag == NVVM::MMAFrag::b) {
653     nRow = k;
654     nCol = n;
655   } else {
656     nRow = m;
657     nCol = n;
658   }
659   assert(nRow && nCol);
660   return inferMMAType(type, frag, nRow, nCol, context);
661 }
662 
663 LogicalResult NVVM::WMMALoadOp::verify() {
664   unsigned addressSpace =
665       llvm::cast<LLVM::LLVMPointerType>(getPtr().getType()).getAddressSpace();
666   if (addressSpace != 0 && addressSpace != NVVM::kGlobalMemorySpace &&
667       addressSpace != NVVM::kSharedMemorySpace)
668     return emitOpError("expected source pointer in memory "
669                        "space 0, 1, 3");
670 
671   if (NVVM::WMMALoadOp::getIntrinsicID(getM(), getN(), getK(), getLayout(),
672                                        getEltype(), getFrag()) == 0)
673     return emitOpError() << "invalid attribute combination";
674   std::pair<Type, unsigned> typeInfo = inferMMATypeFromMNK(
675       getEltype(), getFrag(), getM(), getN(), getK(), getContext());
676   Type dstType = LLVM::LLVMStructType::getLiteral(
677       getContext(), SmallVector<Type, 8>(typeInfo.second, typeInfo.first));
678   if (getType() != dstType)
679     return emitOpError("expected destination type is a structure of ")
680            << typeInfo.second << " elements of type " << typeInfo.first;
681   return success();
682 }
683 
684 LogicalResult NVVM::WMMAStoreOp::verify() {
685   unsigned addressSpace =
686       llvm::cast<LLVM::LLVMPointerType>(getPtr().getType()).getAddressSpace();
687   if (addressSpace != 0 && addressSpace != NVVM::kGlobalMemorySpace &&
688       addressSpace != NVVM::kSharedMemorySpace)
689     return emitOpError("expected operands to be a source pointer in memory "
690                        "space 0, 1, 3");
691 
692   if (NVVM::WMMAStoreOp::getIntrinsicID(getM(), getN(), getK(), getLayout(),
693                                         getEltype()) == 0)
694     return emitOpError() << "invalid attribute combination";
695   std::pair<Type, unsigned> typeInfo = inferMMATypeFromMNK(
696       getEltype(), NVVM::MMAFrag::c, getM(), getN(), getK(), getContext());
697   if (getArgs().size() != typeInfo.second)
698     return emitOpError() << "expected " << typeInfo.second << " data operands";
699   if (llvm::any_of(getArgs(), [&typeInfo](Value operands) {
700         return operands.getType() != typeInfo.first;
701       }))
702     return emitOpError() << "expected data operands of type " << typeInfo.first;
703   return success();
704 }
705 
706 LogicalResult NVVM::WMMAMmaOp::verify() {
707   if (NVVM::WMMAMmaOp::getIntrinsicID(getM(), getN(), getK(), getLayoutA(),
708                                       getLayoutB(), getEltypeA(),
709                                       getEltypeB()) == 0)
710     return emitOpError() << "invalid attribute combination";
711   std::pair<Type, unsigned> typeInfoA = inferMMATypeFromMNK(
712       getEltypeA(), NVVM::MMAFrag::a, getM(), getN(), getK(), getContext());
713   std::pair<Type, unsigned> typeInfoB = inferMMATypeFromMNK(
714       getEltypeA(), NVVM::MMAFrag::b, getM(), getN(), getK(), getContext());
715   std::pair<Type, unsigned> typeInfoC = inferMMATypeFromMNK(
716       getEltypeB(), NVVM::MMAFrag::c, getM(), getN(), getK(), getContext());
717   SmallVector<Type, 32> arguments;
718   arguments.append(typeInfoA.second, typeInfoA.first);
719   arguments.append(typeInfoB.second, typeInfoB.first);
720   arguments.append(typeInfoC.second, typeInfoC.first);
721   unsigned numArgs = arguments.size();
722   if (getArgs().size() != numArgs)
723     return emitOpError() << "expected " << numArgs << " arguments";
724   for (unsigned i = 0; i < numArgs; i++) {
725     if (getArgs()[i].getType() != arguments[i])
726       return emitOpError() << "expected argument " << i << " to be of type "
727                            << arguments[i];
728   }
729   Type dstType = LLVM::LLVMStructType::getLiteral(
730       getContext(), SmallVector<Type, 8>(typeInfoC.second, typeInfoC.first));
731   if (getType() != dstType)
732     return emitOpError("expected destination type is a structure of ")
733            << typeInfoC.second << " elements of type " << typeInfoC.first;
734   return success();
735 }
736 
737 LogicalResult NVVM::LdMatrixOp::verify() {
738   unsigned addressSpace =
739       llvm::cast<LLVM::LLVMPointerType>(getPtr().getType()).getAddressSpace();
740   if (addressSpace != NVVM::kSharedMemorySpace)
741     return emitOpError("expected source pointer in memory space 3");
742 
743   if (getNum() != 1 && getNum() != 2 && getNum() != 4)
744     return emitOpError("expected num attribute to be 1, 2 or 4");
745 
746   Type i32 = IntegerType::get(getContext(), 32);
747   if (getNum() == 1 && getType() != i32)
748     return emitOpError("expected destination type is i32");
749   if (getNum() == 2 || getNum() == 4) {
750     Type dstType = LLVM::LLVMStructType::getLiteral(
751         getContext(), SmallVector<Type>(getNum(), i32));
752     if (getType() != dstType)
753       return emitOpError("expected destination type is a structure of ")
754              << getNum() << " elements of type i32";
755   }
756   return success();
757 }
758 
759 LogicalResult NVVM::StMatrixOp::verify() {
760   unsigned addressSpace =
761       llvm::cast<LLVM::LLVMPointerType>(getPtr().getType()).getAddressSpace();
762   if (addressSpace != NVVM::kSharedMemorySpace)
763     return emitOpError("expected source pointer in memory space 3");
764 
765   int numMatrix = getSources().size();
766   if (numMatrix != 1 && numMatrix != 2 && numMatrix != 4)
767     return emitOpError("expected num attribute to be 1, 2 or 4");
768 
769   return success();
770 }
771 
772 FailureOr<int> getAllowedSizeK(NVVM::WGMMATypes typeA) {
773   if (typeA == NVVM::WGMMATypes::tf32)
774     return 8;
775   if (typeA == NVVM::WGMMATypes::f16 || typeA == NVVM::WGMMATypes::bf16)
776     return 16;
777   if (typeA == NVVM::WGMMATypes::s8 || typeA == NVVM::WGMMATypes::u8)
778     return 32;
779   if (typeA == NVVM::WGMMATypes::e4m3 || typeA == NVVM::WGMMATypes::e5m2)
780     return 32;
781   if (typeA == NVVM::WGMMATypes::b1)
782     return 256;
783   return failure();
784 }
785 
786 LogicalResult isAllowedWGMMADataType(NVVM::WGMMATypes typeD,
787                                      NVVM::WGMMATypes typeA,
788                                      NVVM::WGMMATypes typeB) {
789   switch (typeA) {
790   case NVVM::WGMMATypes::f16:
791     if ((typeD == NVVM::WGMMATypes::f32 || typeD == NVVM::WGMMATypes::f16) &&
792         typeB == NVVM::WGMMATypes::f16)
793       return success();
794     break;
795   case NVVM::WGMMATypes::tf32:
796     if (typeD == NVVM::WGMMATypes::f32 && typeB == NVVM::WGMMATypes::tf32)
797       return success();
798     break;
799   case NVVM::WGMMATypes::u8:
800   case NVVM::WGMMATypes::s8:
801     if (typeD == NVVM::WGMMATypes::s32 &&
802         (typeB == NVVM::WGMMATypes::u8 || typeB == NVVM::WGMMATypes::s8))
803       return success();
804     break;
805   case NVVM::WGMMATypes::b1:
806     if (typeD == NVVM::WGMMATypes::s32 && typeB == NVVM::WGMMATypes::b1)
807       return success();
808     break;
809   case NVVM::WGMMATypes::bf16:
810     if ((typeD == NVVM::WGMMATypes::f32 || typeD == NVVM::WGMMATypes::f16) &&
811         typeB == NVVM::WGMMATypes::bf16)
812       return success();
813     break;
814   case NVVM::WGMMATypes::e4m3:
815   case NVVM::WGMMATypes::e5m2:
816     if ((typeD == NVVM::WGMMATypes::f32 || typeD == NVVM::WGMMATypes::f16) &&
817         (typeB == NVVM::WGMMATypes::e5m2 || typeB == NVVM::WGMMATypes::e4m3))
818       return success();
819     break;
820   case WGMMATypes::f32:
821   case WGMMATypes::s32:
822     llvm_unreachable("unsupported input types");
823     break;
824   }
825   return failure();
826 }
827 
828 LogicalResult isAllowedSizeN(int sizeN, NVVM::WGMMATypes typeA) {
829   SmallVector<int> allowedN = {8,   16,  24,  32,  40,  48,  56,  64,
830                                72,  80,  88,  96,  104, 112, 120, 128,
831                                136, 144, 152, 160, 168, 176, 184, 192,
832                                200, 208, 216, 224, 232, 240, 248, 256};
833   SmallVector<int> allowedNshort = {8,   16,  24,  32,  48,  64,
834                                     80,  96,  112, 128, 144, 160,
835                                     176, 192, 208, 224, 240, 256};
836   switch (typeA) {
837   case WGMMATypes::f16:
838   case WGMMATypes::tf32:
839   case WGMMATypes::bf16:
840   case WGMMATypes::e4m3:
841   case WGMMATypes::e5m2:
842     if (llvm::is_contained(allowedN, sizeN))
843       return success();
844     break;
845   case WGMMATypes::u8:
846   case WGMMATypes::s8:
847   case WGMMATypes::b1:
848     if (llvm::is_contained(allowedNshort, sizeN))
849       return success();
850     break;
851   case WGMMATypes::f32:
852   case WGMMATypes::s32:
853     llvm_unreachable("unsupported input types");
854     break;
855   }
856   return failure();
857 }
858 
859 LogicalResult NVVM::WgmmaMmaAsyncOp::verify() {
860   Value outValue = getResults();
861   auto stype = dyn_cast<LLVM::LLVMStructType>(outValue.getType());
862   if (!stype)
863     return emitOpError() << "expected results to be struct";
864   int outputSize = stype.getBody().size();
865   WGMMATypes typeD = getTypeD();
866   WGMMATypes typeA = getTypeA();
867   WGMMATypes typeB = getTypeB();
868 
869   for (Type t : stype.getBody()) {
870     if (t != stype.getBody().front())
871       return emitOpError()
872              << "all elements in struct must be same type but there is " << t;
873   }
874 
875   if (typeD != WGMMATypes::f32 && typeD != WGMMATypes::f16 &&
876       typeD != WGMMATypes::s32) {
877     return emitOpError() << "does not support the given output type "
878                          << NVVM::stringifyWGMMATypes(typeD);
879   }
880   if (typeD == WGMMATypes::s32 &&
881       (getScaleA() == WGMMAScaleIn::neg || getScaleB() == WGMMAScaleIn::neg)) {
882     return emitOpError() << "has s32 output, scaleA and scaleB cannot be neg";
883   }
884 
885   if (failed(isAllowedWGMMADataType(typeD, typeA, typeB))) {
886     return emitOpError() << NVVM::stringifyWGMMATypes(typeD)
887                          << " += " << NVVM::stringifyWGMMATypes(typeA) << " * "
888                          << NVVM::stringifyWGMMATypes(typeB)
889                          << ", it is not supported.";
890   }
891 
892   // Check M
893   if (getShape().getM() != 64)
894     return emitOpError() << "shape 'm' must be 64";
895 
896   // Check K
897   FailureOr<int> allowedK = getAllowedSizeK(typeA);
898   if (failed(allowedK) || allowedK.value() != getShape().getK())
899     return emitOpError() << "shape 'k' must be " << allowedK.value()
900                          << " for input type "
901                          << NVVM::stringifyWGMMATypes(typeA);
902 
903   // Check N
904   if (failed(isAllowedSizeN(getShape().getN(), typeA))) {
905     return emitOpError() << "has input type "
906                          << NVVM::stringifyWGMMATypes(typeA) << " n is set to "
907                          << getShape().getN() << ", it is not supported.";
908   }
909 
910   // Check transpose (only available for f16/bf16)
911   // Matrices A should be stored in row-major and B in column-major.
912   // Only f16/bf16 matrices can be stored in either column-major or row-major
913   // by setting the tranpose value(imm-trans-a,imm-trans-b) in PTX code.
914   if ((typeA != WGMMATypes::f16 && typeA != WGMMATypes::bf16) &&
915       (getLayoutA() == mlir::NVVM::MMALayout::col ||
916        getLayoutB() == mlir::NVVM::MMALayout::row)) {
917     return emitOpError()
918            << "given layouts layout_a = " << stringifyMMALayout(getLayoutA())
919            << " and layout_b = " << stringifyMMALayout(getLayoutB())
920            << " for input types " << stringifyWGMMATypes(typeA) << " and "
921            << stringifyWGMMATypes(typeB)
922            << " requires transpose. However, this is only supported for: "
923            << stringifyMMATypes(MMATypes::f16) << " and "
924            << stringifyMMATypes(MMATypes::bf16);
925   }
926 
927   // Check result registers
928   int expectedOutput = 0;
929   if (typeD == WGMMATypes::f32 || typeD == WGMMATypes::s32)
930     expectedOutput = getShape().getN() / 2;
931   if (typeD == WGMMATypes::f16)
932     expectedOutput = getShape().getN() / 4;
933   if (outputSize != expectedOutput) {
934     return emitOpError() << "results " << expectedOutput
935                          << ", however output struct has " << outputSize
936                          << " elements";
937   }
938   // Check satfinite (only available for s32 accumulator)
939   if (typeD != WGMMATypes::s32 &&
940       getSatfinite().value_or(NVVM::MMAIntOverflow::wrapped) ==
941           NVVM::MMAIntOverflow::satfinite) {
942     return emitOpError()
943            << " `satfinite` can be only used with s32 accumulator, however "
944               "the current accumulator is "
945            << NVVM::stringifyWGMMATypes(typeD);
946   }
947 
948   return success();
949 }
950 
951 std::string NVVM::WgmmaMmaAsyncOp::getPtx() {
952 
953   int m = getShape().getM(), n = getShape().getN(), k = getShape().getK();
954   bool isF16 = getTypeA() == WGMMATypes::f16 || getTypeA() == WGMMATypes::bf16;
955 
956   StringRef outputTypeName = stringifyWGMMATypes(getTypeD());
957 
958   int expectedOutputRegisters = 0;
959   if (getTypeD() == WGMMATypes::f16)
960     expectedOutputRegisters = getShape().getN() / 4;
961   else
962     expectedOutputRegisters = getShape().getN() / 2;
963 
964   std::string ptx;
965   llvm::raw_string_ostream ss(ptx);
966 
967   ss << "{\n"
968         ".reg .pred p;\n"
969         "setp.ne.b32 p, $"
970      << ((expectedOutputRegisters * 2) + 2)
971      << ", 0;\n"
972         "wgmma.mma_async.sync.aligned.m"
973      << m << "n" << n << "k" << k << "." << outputTypeName << "."
974      << stringifyWGMMATypes(getTypeA()) << "."
975      << stringifyWGMMATypes(getTypeB());
976   if (getSatfinite().value_or(NVVM::MMAIntOverflow::wrapped) ==
977       NVVM::MMAIntOverflow::satfinite)
978     ss << ".satfinite";
979   ss << " {";
980   int regCnt = 0;
981   for (; regCnt < expectedOutputRegisters; ++regCnt) {
982     ss << "$" << regCnt;
983     if (regCnt != expectedOutputRegisters - 1)
984       ss << ", ";
985   }
986 
987   ss << "},";
988   // Need to map read/write registers correctly.
989   regCnt = (regCnt * 2);
990   ss << " $" << (regCnt) << "," << " $" << (regCnt + 1) << "," << " p";
991   if (getTypeD() != WGMMATypes::s32) {
992     ss << ", $" << (regCnt + 3) << ",  $" << (regCnt + 4);
993   }
994   // Don't add transpose parameters unless needed.
995   if (isF16) {
996     ss << ", $" << (regCnt + 5) << ",  $" << (regCnt + 6);
997   }
998   ss << ";\n"
999      << "}\n";
1000   return ptx;
1001 }
1002 
1003 void NVVM::WgmmaMmaAsyncOp::getAsmValues(
1004     RewriterBase &rewriter,
1005     llvm::SmallVectorImpl<std::pair<mlir::Value, mlir::NVVM::PTXRegisterMod>>
1006         &asmValues) {
1007   bool isF16 = getTypeA() == WGMMATypes::f16 || getTypeA() == WGMMATypes::bf16;
1008   if (getResults())
1009     asmValues.push_back({getResults(), mlir::NVVM::PTXRegisterMod::Write});
1010   if (getInouts())
1011     asmValues.push_back({getInouts(), mlir::NVVM::PTXRegisterMod::ReadWrite});
1012   asmValues.push_back({getDescriptorA(), mlir::NVVM::PTXRegisterMod::Read});
1013   asmValues.push_back({getDescriptorB(), mlir::NVVM::PTXRegisterMod::Read});
1014   asmValues.push_back({makeConstantI32(rewriter, static_cast<int>(getScaleD())),
1015                        mlir::NVVM::PTXRegisterMod::Read});
1016   if (getTypeD() != WGMMATypes::s32) {
1017     asmValues.push_back(
1018         {makeConstantI32(rewriter,
1019                          getScaleA() == NVVM::WGMMAScaleIn::neg ? -1 : 1),
1020          mlir::NVVM::PTXRegisterMod::Read});
1021     asmValues.push_back(
1022         {makeConstantI32(rewriter,
1023                          getScaleB() == NVVM::WGMMAScaleIn::neg ? -1 : 1),
1024          mlir::NVVM::PTXRegisterMod::Read});
1025   }
1026   if (isF16) {
1027     asmValues.push_back(
1028         {makeConstantI32(rewriter, static_cast<int>(getLayoutA())),
1029          mlir::NVVM::PTXRegisterMod::Read});
1030     asmValues.push_back(
1031         {makeConstantI32(rewriter, 1 - static_cast<int>(getLayoutB())),
1032          mlir::NVVM::PTXRegisterMod::Read});
1033   }
1034 }
1035 LogicalResult NVVM::FenceProxyOp::verify() {
1036   if (getKind() == NVVM::ProxyKind::TENSORMAP)
1037     return emitOpError() << "tensormap proxy is not a supported proxy kind";
1038   if (getKind() == NVVM::ProxyKind::GENERIC)
1039     return emitOpError() << "generic proxy not a supported proxy kind";
1040   if (getKind() == NVVM::ProxyKind::async_shared && !getSpace().has_value()) {
1041     return emitOpError() << "async_shared fence requires space attribute";
1042   }
1043   if (getKind() != NVVM::ProxyKind::async_shared && getSpace().has_value()) {
1044     return emitOpError() << "only async_shared fence can have space attribute";
1045   }
1046   return success();
1047 }
1048 
1049 LogicalResult NVVM::FenceProxyAcquireOp::verify() {
1050   if (getFromProxy() != NVVM::ProxyKind::GENERIC)
1051     return emitOpError("uni-directional proxies only support generic for "
1052                        "from_proxy attribute");
1053 
1054   if (getToProxy() != NVVM::ProxyKind::TENSORMAP)
1055     return emitOpError("uni-directional proxies only support tensormap "
1056                        "for to_proxy attribute");
1057 
1058   return success();
1059 }
1060 
1061 LogicalResult NVVM::FenceProxyReleaseOp::verify() {
1062   if (getFromProxy() != NVVM::ProxyKind::GENERIC)
1063     return emitOpError("uni-directional proxies only support generic for "
1064                        "from_proxy attribute");
1065 
1066   if (getToProxy() != NVVM::ProxyKind::TENSORMAP)
1067     return emitOpError("uni-directional proxies only support tensormap "
1068                        "for to_proxy attribute");
1069 
1070   return success();
1071 }
1072 
1073 LogicalResult NVVM::SetMaxRegisterOp::verify() {
1074   if (getRegCount() % 8)
1075     return emitOpError("new register size must be multiple of 8");
1076   if (getRegCount() < 24 || getRegCount() > 256)
1077     return emitOpError("new register size must be in between 24 to 256");
1078   return success();
1079 }
1080 
1081 LogicalResult NVVM::BarrierOp::verify() {
1082   if (getNumberOfThreads() && !getBarrierId())
1083     return emitOpError(
1084         "barrier id is missing, it should be set between 0 to 15");
1085   return success();
1086 }
1087 
1088 llvm::Intrinsic::ID CpAsyncBulkTensorPrefetchOp::getIntrinsicID(int tensorDims,
1089                                                                 bool isIm2Col) {
1090   switch (tensorDims) {
1091   case 1:
1092     return llvm::Intrinsic::nvvm_cp_async_bulk_tensor_prefetch_tile_1d;
1093   case 2:
1094     return llvm::Intrinsic::nvvm_cp_async_bulk_tensor_prefetch_tile_2d;
1095   case 3:
1096     return isIm2Col
1097                ? llvm::Intrinsic::nvvm_cp_async_bulk_tensor_prefetch_im2col_3d
1098                : llvm::Intrinsic::nvvm_cp_async_bulk_tensor_prefetch_tile_3d;
1099   case 4:
1100     return isIm2Col
1101                ? llvm::Intrinsic::nvvm_cp_async_bulk_tensor_prefetch_im2col_4d
1102                : llvm::Intrinsic::nvvm_cp_async_bulk_tensor_prefetch_tile_4d;
1103   case 5:
1104     return isIm2Col
1105                ? llvm::Intrinsic::nvvm_cp_async_bulk_tensor_prefetch_im2col_5d
1106                : llvm::Intrinsic::nvvm_cp_async_bulk_tensor_prefetch_tile_5d;
1107   default:
1108     llvm_unreachable("Invalid TensorDim in CpAsyncBulkTensorPrefetchOp.");
1109   }
1110 }
1111 
1112 #define CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, dim, mode)                        \
1113   llvm::Intrinsic::nvvm_cp_async_bulk_tensor_##op##_##mode##_##dim##d
1114 
1115 #define CP_ASYNC_BULK_TENSOR_REDUCE(op, dim, is_im2col)                        \
1116   is_im2col ? CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, dim, im2col)                \
1117             : CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, dim, tile)
1118 
1119 #define GET_CP_ASYNC_BULK_TENSOR_ID(op, dims, is_im2col)                       \
1120   [&]() -> auto {                                                              \
1121     switch (dims) {                                                            \
1122     case 1:                                                                    \
1123       return CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, 1, tile);                    \
1124     case 2:                                                                    \
1125       return CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, 2, tile);                    \
1126     case 3:                                                                    \
1127       return CP_ASYNC_BULK_TENSOR_REDUCE(op, 3, is_im2col);                    \
1128     case 4:                                                                    \
1129       return CP_ASYNC_BULK_TENSOR_REDUCE(op, 4, is_im2col);                    \
1130     case 5:                                                                    \
1131       return CP_ASYNC_BULK_TENSOR_REDUCE(op, 5, is_im2col);                    \
1132     default:                                                                   \
1133       llvm_unreachable("Invalid TensorDim in CpAsyncBulkTensorReduceOp.");     \
1134     }                                                                          \
1135   }()
1136 
1137 llvm::Intrinsic::ID CpAsyncBulkTensorReduceOp::getIntrinsicID(
1138     int tensorDims, NVVM::TMAReduxKind kind, bool isIm2Col) {
1139   using RedTy = NVVM::TMAReduxKind;
1140   switch (kind) {
1141   case RedTy::ADD:
1142     return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_add, tensorDims, isIm2Col);
1143   case RedTy::MIN:
1144     return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_min, tensorDims, isIm2Col);
1145   case RedTy::MAX:
1146     return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_max, tensorDims, isIm2Col);
1147   case RedTy::INC:
1148     return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_inc, tensorDims, isIm2Col);
1149   case RedTy::DEC:
1150     return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_dec, tensorDims, isIm2Col);
1151   case RedTy::AND:
1152     return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_and, tensorDims, isIm2Col);
1153   case RedTy::OR:
1154     return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_or, tensorDims, isIm2Col);
1155   case RedTy::XOR:
1156     return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_xor, tensorDims, isIm2Col);
1157   }
1158   llvm_unreachable("Invalid Reduction Op for CpAsyncBulkTensorReduceOp");
1159 }
1160 
1161 /// Infer the result ranges for the NVVM SpecialRangeableRegisterOp that might
1162 /// have ConstantRangeAttr.
1163 static void nvvmInferResultRanges(Operation *op, Value result,
1164                                   ArrayRef<::mlir::ConstantIntRanges> argRanges,
1165                                   SetIntRangeFn setResultRanges) {
1166   if (auto rangeAttr = op->getAttrOfType<LLVM::ConstantRangeAttr>("range")) {
1167     setResultRanges(result, {rangeAttr.getLower(), rangeAttr.getUpper(),
1168                              rangeAttr.getLower(), rangeAttr.getUpper()});
1169   }
1170 }
1171 
1172 //===----------------------------------------------------------------------===//
1173 // NVVMDialect initialization, type parsing, and registration.
1174 //===----------------------------------------------------------------------===//
1175 
1176 // TODO: This should be the llvm.nvvm dialect once this is supported.
1177 void NVVMDialect::initialize() {
1178   addOperations<
1179 #define GET_OP_LIST
1180 #include "mlir/Dialect/LLVMIR/NVVMOps.cpp.inc"
1181       >();
1182   addAttributes<
1183 #define GET_ATTRDEF_LIST
1184 #include "mlir/Dialect/LLVMIR/NVVMOpsAttributes.cpp.inc"
1185       >();
1186 
1187   // Support unknown operations because not all NVVM operations are
1188   // registered.
1189   allowUnknownOperations();
1190   declarePromisedInterface<ConvertToLLVMPatternInterface, NVVMDialect>();
1191   declarePromisedInterface<gpu::TargetAttrInterface, NVVMTargetAttr>();
1192 }
1193 
1194 LogicalResult NVVMDialect::verifyOperationAttribute(Operation *op,
1195                                                     NamedAttribute attr) {
1196   StringAttr attrName = attr.getName();
1197   // Kernel function attribute should be attached to functions.
1198   if (attrName == NVVMDialect::getKernelFuncAttrName()) {
1199     if (!isa<LLVM::LLVMFuncOp>(op)) {
1200       return op->emitError() << "'" << NVVMDialect::getKernelFuncAttrName()
1201                              << "' attribute attached to unexpected op";
1202     }
1203   }
1204   // If maxntid / reqntid / cluster_dim exist, it must be an array with max 3
1205   // dim
1206   if (attrName == NVVMDialect::getMaxntidAttrName() ||
1207       attrName == NVVMDialect::getReqntidAttrName() ||
1208       attrName == NVVMDialect::getClusterDimAttrName()) {
1209     auto values = llvm::dyn_cast<DenseI32ArrayAttr>(attr.getValue());
1210     if (!values || values.empty() || values.size() > 3)
1211       return op->emitError()
1212              << "'" << attrName
1213              << "' attribute must be integer array with maximum 3 index";
1214   }
1215   // If minctasm / maxnreg / cluster_max_blocks exist, it must be an integer
1216   // attribute
1217   if (attrName == NVVMDialect::getMinctasmAttrName() ||
1218       attrName == NVVMDialect::getMaxnregAttrName() ||
1219       attrName == NVVMDialect::getClusterMaxBlocksAttrName()) {
1220     if (!llvm::dyn_cast<IntegerAttr>(attr.getValue()))
1221       return op->emitError()
1222              << "'" << attrName << "' attribute must be integer constant";
1223   }
1224 
1225   return success();
1226 }
1227 
1228 LogicalResult NVVMDialect::verifyRegionArgAttribute(Operation *op,
1229                                                     unsigned regionIndex,
1230                                                     unsigned argIndex,
1231                                                     NamedAttribute argAttr) {
1232   auto funcOp = dyn_cast<FunctionOpInterface>(op);
1233   if (!funcOp)
1234     return success();
1235 
1236   bool isKernel = op->hasAttr(NVVMDialect::getKernelFuncAttrName());
1237   StringAttr attrName = argAttr.getName();
1238   if (attrName == NVVM::NVVMDialect::getGridConstantAttrName()) {
1239     if (!isKernel) {
1240       return op->emitError()
1241              << "'" << attrName
1242              << "' attribute must be present only on kernel arguments";
1243     }
1244     if (!isa<UnitAttr>(argAttr.getValue()))
1245       return op->emitError() << "'" << attrName << "' must be a unit attribute";
1246     if (!funcOp.getArgAttr(argIndex, LLVM::LLVMDialect::getByValAttrName())) {
1247       return op->emitError()
1248              << "'" << attrName
1249              << "' attribute requires the argument to also have attribute '"
1250              << LLVM::LLVMDialect::getByValAttrName() << "'";
1251     }
1252   }
1253 
1254   return success();
1255 }
1256 
1257 //===----------------------------------------------------------------------===//
1258 // NVVM target attribute.
1259 //===----------------------------------------------------------------------===//
1260 LogicalResult
1261 NVVMTargetAttr::verify(function_ref<InFlightDiagnostic()> emitError,
1262                        int optLevel, StringRef triple, StringRef chip,
1263                        StringRef features, DictionaryAttr flags,
1264                        ArrayAttr files) {
1265   if (optLevel < 0 || optLevel > 3) {
1266     emitError() << "The optimization level must be a number between 0 and 3.";
1267     return failure();
1268   }
1269   if (triple.empty()) {
1270     emitError() << "The target triple cannot be empty.";
1271     return failure();
1272   }
1273   if (chip.empty()) {
1274     emitError() << "The target chip cannot be empty.";
1275     return failure();
1276   }
1277   if (files && !llvm::all_of(files, [](::mlir::Attribute attr) {
1278         return attr && mlir::isa<StringAttr>(attr);
1279       })) {
1280     emitError() << "All the elements in the `link` array must be strings.";
1281     return failure();
1282   }
1283   return success();
1284 }
1285 
1286 #define GET_OP_CLASSES
1287 #include "mlir/Dialect/LLVMIR/NVVMOps.cpp.inc"
1288 
1289 #define GET_ATTRDEF_CLASSES
1290 #include "mlir/Dialect/LLVMIR/NVVMOpsAttributes.cpp.inc"
1291