1//===- GPUOpsLowering.cpp - GPU FuncOp / ReturnOp lowering ----------------===//
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#include "GPUOpsLowering.h"
10
11#include "mlir/Conversion/GPUCommon/GPUCommonPass.h"
12#include "mlir/Conversion/LLVMCommon/VectorPattern.h"
13#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
14#include "mlir/IR/Attributes.h"
15#include "mlir/IR/Builders.h"
16#include "mlir/IR/BuiltinTypes.h"
17#include "llvm/ADT/SmallVectorExtras.h"
18#include "llvm/ADT/StringSet.h"
19#include "llvm/Support/FormatVariadic.h"
20
21using namespace mlir;
22
23LLVM::LLVMFuncOp mlir::getOrDefineFunction(gpu::GPUModuleOp moduleOp,
24 Location loc, OpBuilder &b,
25 StringRef name,
26 LLVM::LLVMFunctionType type) {
27 LLVM::LLVMFuncOp ret;
28 if (!(ret = moduleOp.template lookupSymbol<LLVM::LLVMFuncOp>(name))) {
29 OpBuilder::InsertionGuard guard(b);
30 b.setInsertionPointToStart(moduleOp.getBody());
31 ret = b.create<LLVM::LLVMFuncOp>(loc, name, type, LLVM::Linkage::External);
32 }
33 return ret;
34}
35
36static SmallString<16> getUniqueSymbolName(gpu::GPUModuleOp moduleOp,
37 StringRef prefix) {
38 // Get a unique global name.
39 unsigned stringNumber = 0;
40 SmallString<16> stringConstName;
41 do {
42 stringConstName.clear();
43 (prefix + Twine(stringNumber++)).toStringRef(Out&: stringConstName);
44 } while (moduleOp.lookupSymbol(stringConstName));
45 return stringConstName;
46}
47
48LLVM::GlobalOp
49mlir::getOrCreateStringConstant(OpBuilder &b, Location loc,
50 gpu::GPUModuleOp moduleOp, Type llvmI8,
51 StringRef namePrefix, StringRef str,
52 uint64_t alignment, unsigned addrSpace) {
53 llvm::SmallString<20> nullTermStr(str);
54 nullTermStr.push_back(Elt: '\0'); // Null terminate for C
55 auto globalType =
56 LLVM::LLVMArrayType::get(llvmI8, nullTermStr.size_in_bytes());
57 StringAttr attr = b.getStringAttr(nullTermStr);
58
59 // Try to find existing global.
60 for (auto globalOp : moduleOp.getOps<LLVM::GlobalOp>())
61 if (globalOp.getGlobalType() == globalType && globalOp.getConstant() &&
62 globalOp.getValueAttr() == attr &&
63 globalOp.getAlignment().value_or(0) == alignment &&
64 globalOp.getAddrSpace() == addrSpace)
65 return globalOp;
66
67 // Not found: create new global.
68 OpBuilder::InsertionGuard guard(b);
69 b.setInsertionPointToStart(moduleOp.getBody());
70 SmallString<16> name = getUniqueSymbolName(moduleOp, namePrefix);
71 return b.create<LLVM::GlobalOp>(loc, globalType,
72 /*isConstant=*/true, LLVM::Linkage::Internal,
73 name, attr, alignment, addrSpace);
74}
75
76LogicalResult
77GPUFuncOpLowering::matchAndRewrite(gpu::GPUFuncOp gpuFuncOp, OpAdaptor adaptor,
78 ConversionPatternRewriter &rewriter) const {
79 Location loc = gpuFuncOp.getLoc();
80
81 SmallVector<LLVM::GlobalOp, 3> workgroupBuffers;
82 if (encodeWorkgroupAttributionsAsArguments) {
83 // Append an `llvm.ptr` argument to the function signature to encode
84 // workgroup attributions.
85
86 ArrayRef<BlockArgument> workgroupAttributions =
87 gpuFuncOp.getWorkgroupAttributions();
88 size_t numAttributions = workgroupAttributions.size();
89
90 // Insert all arguments at the end.
91 unsigned index = gpuFuncOp.getNumArguments();
92 SmallVector<unsigned> argIndices(numAttributions, index);
93
94 // New arguments will simply be `llvm.ptr` with the correct address space
95 Type workgroupPtrType =
96 rewriter.getType<LLVM::LLVMPointerType>(workgroupAddrSpace);
97 SmallVector<Type> argTypes(numAttributions, workgroupPtrType);
98
99 // Attributes: noalias, llvm.mlir.workgroup_attribution(<size>, <type>)
100 std::array attrs{
101 rewriter.getNamedAttr(LLVM::LLVMDialect::getNoAliasAttrName(),
102 rewriter.getUnitAttr()),
103 rewriter.getNamedAttr(
104 getDialect().getWorkgroupAttributionAttrHelper().getName(),
105 rewriter.getUnitAttr()),
106 };
107 SmallVector<DictionaryAttr> argAttrs;
108 for (BlockArgument attribution : workgroupAttributions) {
109 auto attributionType = cast<MemRefType>(attribution.getType());
110 IntegerAttr numElements =
111 rewriter.getI64IntegerAttr(attributionType.getNumElements());
112 Type llvmElementType =
113 getTypeConverter()->convertType(attributionType.getElementType());
114 if (!llvmElementType)
115 return failure();
116 TypeAttr type = TypeAttr::get(llvmElementType);
117 attrs.back().setValue(
118 rewriter.getAttr<LLVM::WorkgroupAttributionAttr>(numElements, type));
119 argAttrs.push_back(rewriter.getDictionaryAttr(attrs));
120 }
121
122 // Location match function location
123 SmallVector<Location> argLocs(numAttributions, gpuFuncOp.getLoc());
124
125 // Perform signature modification
126 rewriter.modifyOpInPlace(
127 gpuFuncOp, [gpuFuncOp, &argIndices, &argTypes, &argAttrs, &argLocs]() {
128 LogicalResult inserted =
129 static_cast<FunctionOpInterface>(gpuFuncOp).insertArguments(
130 argIndices, argTypes, argAttrs, argLocs);
131 (void)inserted;
132 assert(succeeded(inserted) &&
133 "expected GPU funcs to support inserting any argument");
134 });
135 } else {
136 workgroupBuffers.reserve(gpuFuncOp.getNumWorkgroupAttributions());
137 for (auto [idx, attribution] :
138 llvm::enumerate(gpuFuncOp.getWorkgroupAttributions())) {
139 auto type = dyn_cast<MemRefType>(attribution.getType());
140 assert(type && type.hasStaticShape() && "unexpected type in attribution");
141
142 uint64_t numElements = type.getNumElements();
143
144 auto elementType =
145 cast<Type>(typeConverter->convertType(type.getElementType()));
146 auto arrayType = LLVM::LLVMArrayType::get(elementType, numElements);
147 std::string name =
148 std::string(llvm::formatv("__wg_{0}_{1}", gpuFuncOp.getName(), idx));
149 uint64_t alignment = 0;
150 if (auto alignAttr = dyn_cast_or_null<IntegerAttr>(
151 gpuFuncOp.getWorkgroupAttributionAttr(
152 idx, LLVM::LLVMDialect::getAlignAttrName())))
153 alignment = alignAttr.getInt();
154 auto globalOp = rewriter.create<LLVM::GlobalOp>(
155 gpuFuncOp.getLoc(), arrayType, /*isConstant=*/false,
156 LLVM::Linkage::Internal, name, /*value=*/Attribute(), alignment,
157 workgroupAddrSpace);
158 workgroupBuffers.push_back(globalOp);
159 }
160 }
161
162 // Remap proper input types.
163 TypeConverter::SignatureConversion signatureConversion(
164 gpuFuncOp.front().getNumArguments());
165
166 Type funcType = getTypeConverter()->convertFunctionSignature(
167 gpuFuncOp.getFunctionType(), /*isVariadic=*/false,
168 getTypeConverter()->getOptions().useBarePtrCallConv, signatureConversion);
169 if (!funcType) {
170 return rewriter.notifyMatchFailure(gpuFuncOp, [&](Diagnostic &diag) {
171 diag << "failed to convert function signature type for: "
172 << gpuFuncOp.getFunctionType();
173 });
174 }
175
176 // Create the new function operation. Only copy those attributes that are
177 // not specific to function modeling.
178 SmallVector<NamedAttribute, 4> attributes;
179 ArrayAttr argAttrs;
180 for (const auto &attr : gpuFuncOp->getAttrs()) {
181 if (attr.getName() == SymbolTable::getSymbolAttrName() ||
182 attr.getName() == gpuFuncOp.getFunctionTypeAttrName() ||
183 attr.getName() ==
184 gpu::GPUFuncOp::getNumWorkgroupAttributionsAttrName() ||
185 attr.getName() == gpuFuncOp.getWorkgroupAttribAttrsAttrName() ||
186 attr.getName() == gpuFuncOp.getPrivateAttribAttrsAttrName() ||
187 attr.getName() == gpuFuncOp.getKnownBlockSizeAttrName() ||
188 attr.getName() == gpuFuncOp.getKnownGridSizeAttrName())
189 continue;
190 if (attr.getName() == gpuFuncOp.getArgAttrsAttrName()) {
191 argAttrs = gpuFuncOp.getArgAttrsAttr();
192 continue;
193 }
194 attributes.push_back(attr);
195 }
196
197 DenseI32ArrayAttr knownBlockSize = gpuFuncOp.getKnownBlockSizeAttr();
198 DenseI32ArrayAttr knownGridSize = gpuFuncOp.getKnownGridSizeAttr();
199 // Ensure we don't lose information if the function is lowered before its
200 // surrounding context.
201 auto *gpuDialect = cast<gpu::GPUDialect>(gpuFuncOp->getDialect());
202 if (knownBlockSize)
203 attributes.emplace_back(gpuDialect->getKnownBlockSizeAttrHelper().getName(),
204 knownBlockSize);
205 if (knownGridSize)
206 attributes.emplace_back(gpuDialect->getKnownGridSizeAttrHelper().getName(),
207 knownGridSize);
208
209 // Add a dialect specific kernel attribute in addition to GPU kernel
210 // attribute. The former is necessary for further translation while the
211 // latter is expected by gpu.launch_func.
212 if (gpuFuncOp.isKernel()) {
213 if (kernelAttributeName)
214 attributes.emplace_back(kernelAttributeName, rewriter.getUnitAttr());
215 // Set the dialect-specific block size attribute if there is one.
216 if (kernelBlockSizeAttributeName && knownBlockSize) {
217 attributes.emplace_back(kernelBlockSizeAttributeName, knownBlockSize);
218 }
219 }
220 LLVM::CConv callingConvention = gpuFuncOp.isKernel()
221 ? kernelCallingConvention
222 : nonKernelCallingConvention;
223 auto llvmFuncOp = rewriter.create<LLVM::LLVMFuncOp>(
224 gpuFuncOp.getLoc(), gpuFuncOp.getName(), funcType,
225 LLVM::Linkage::External, /*dsoLocal=*/false, callingConvention,
226 /*comdat=*/nullptr, attributes);
227
228 {
229 // Insert operations that correspond to converted workgroup and private
230 // memory attributions to the body of the function. This must operate on
231 // the original function, before the body region is inlined in the new
232 // function to maintain the relation between block arguments and the
233 // parent operation that assigns their semantics.
234 OpBuilder::InsertionGuard guard(rewriter);
235
236 // Rewrite workgroup memory attributions to addresses of global buffers.
237 rewriter.setInsertionPointToStart(&gpuFuncOp.front());
238 unsigned numProperArguments = gpuFuncOp.getNumArguments();
239
240 if (encodeWorkgroupAttributionsAsArguments) {
241 // Build a MemRefDescriptor with each of the arguments added above.
242
243 unsigned numAttributions = gpuFuncOp.getNumWorkgroupAttributions();
244 assert(numProperArguments >= numAttributions &&
245 "Expecting attributions to be encoded as arguments already");
246
247 // Arguments encoding workgroup attributions will be in positions
248 // [numProperArguments, numProperArguments+numAttributions)
249 ArrayRef<BlockArgument> attributionArguments =
250 gpuFuncOp.getArguments().slice(numProperArguments - numAttributions,
251 numAttributions);
252 for (auto [idx, vals] : llvm::enumerate(llvm::zip_equal(
253 gpuFuncOp.getWorkgroupAttributions(), attributionArguments))) {
254 auto [attribution, arg] = vals;
255 auto type = cast<MemRefType>(attribution.getType());
256
257 // Arguments are of llvm.ptr type and attributions are of memref type:
258 // we need to wrap them in memref descriptors.
259 Value descr = MemRefDescriptor::fromStaticShape(
260 rewriter, loc, *getTypeConverter(), type, arg);
261
262 // And remap the arguments
263 signatureConversion.remapInput(numProperArguments + idx, descr);
264 }
265 } else {
266 for (const auto [idx, global] : llvm::enumerate(workgroupBuffers)) {
267 auto ptrType = LLVM::LLVMPointerType::get(rewriter.getContext(),
268 global.getAddrSpace());
269 Value address = rewriter.create<LLVM::AddressOfOp>(
270 loc, ptrType, global.getSymNameAttr());
271 Value memory =
272 rewriter.create<LLVM::GEPOp>(loc, ptrType, global.getType(),
273 address, ArrayRef<LLVM::GEPArg>{0, 0});
274
275 // Build a memref descriptor pointing to the buffer to plug with the
276 // existing memref infrastructure. This may use more registers than
277 // otherwise necessary given that memref sizes are fixed, but we can try
278 // and canonicalize that away later.
279 Value attribution = gpuFuncOp.getWorkgroupAttributions()[idx];
280 auto type = cast<MemRefType>(attribution.getType());
281 Value descr = MemRefDescriptor::fromStaticShape(
282 rewriter, loc, *getTypeConverter(), type, memory);
283 signatureConversion.remapInput(numProperArguments + idx, descr);
284 }
285 }
286
287 // Rewrite private memory attributions to alloca'ed buffers.
288 unsigned numWorkgroupAttributions = gpuFuncOp.getNumWorkgroupAttributions();
289 auto int64Ty = IntegerType::get(rewriter.getContext(), 64);
290 for (const auto [idx, attribution] :
291 llvm::enumerate(gpuFuncOp.getPrivateAttributions())) {
292 auto type = cast<MemRefType>(attribution.getType());
293 assert(type && type.hasStaticShape() && "unexpected type in attribution");
294
295 // Explicitly drop memory space when lowering private memory
296 // attributions since NVVM models it as `alloca`s in the default
297 // memory space and does not support `alloca`s with addrspace(5).
298 Type elementType = typeConverter->convertType(type.getElementType());
299 auto ptrType =
300 LLVM::LLVMPointerType::get(rewriter.getContext(), allocaAddrSpace);
301 Value numElements = rewriter.create<LLVM::ConstantOp>(
302 gpuFuncOp.getLoc(), int64Ty, type.getNumElements());
303 uint64_t alignment = 0;
304 if (auto alignAttr =
305 dyn_cast_or_null<IntegerAttr>(gpuFuncOp.getPrivateAttributionAttr(
306 idx, LLVM::LLVMDialect::getAlignAttrName())))
307 alignment = alignAttr.getInt();
308 Value allocated = rewriter.create<LLVM::AllocaOp>(
309 gpuFuncOp.getLoc(), ptrType, elementType, numElements, alignment);
310 Value descr = MemRefDescriptor::fromStaticShape(
311 rewriter, loc, *getTypeConverter(), type, allocated);
312 signatureConversion.remapInput(
313 numProperArguments + numWorkgroupAttributions + idx, descr);
314 }
315 }
316
317 // Move the region to the new function, update the entry block signature.
318 rewriter.inlineRegionBefore(gpuFuncOp.getBody(), llvmFuncOp.getBody(),
319 llvmFuncOp.end());
320 if (failed(rewriter.convertRegionTypes(&llvmFuncOp.getBody(), *typeConverter,
321 &signatureConversion)))
322 return failure();
323
324 // Get memref type from function arguments and set the noalias to
325 // pointer arguments.
326 for (const auto [idx, argTy] :
327 llvm::enumerate(gpuFuncOp.getArgumentTypes())) {
328 auto remapping = signatureConversion.getInputMapping(idx);
329 NamedAttrList argAttr =
330 argAttrs ? cast<DictionaryAttr>(argAttrs[idx]) : NamedAttrList();
331 auto copyAttribute = [&](StringRef attrName) {
332 Attribute attr = argAttr.erase(attrName);
333 if (!attr)
334 return;
335 for (size_t i = 0, e = remapping->size; i < e; ++i)
336 llvmFuncOp.setArgAttr(remapping->inputNo + i, attrName, attr);
337 };
338 auto copyPointerAttribute = [&](StringRef attrName) {
339 Attribute attr = argAttr.erase(attrName);
340
341 if (!attr)
342 return;
343 if (remapping->size > 1 &&
344 attrName == LLVM::LLVMDialect::getNoAliasAttrName()) {
345 emitWarning(llvmFuncOp.getLoc(),
346 "Cannot copy noalias with non-bare pointers.\n");
347 return;
348 }
349 for (size_t i = 0, e = remapping->size; i < e; ++i) {
350 if (isa<LLVM::LLVMPointerType>(
351 llvmFuncOp.getArgument(remapping->inputNo + i).getType())) {
352 llvmFuncOp.setArgAttr(remapping->inputNo + i, attrName, attr);
353 }
354 }
355 };
356
357 if (argAttr.empty())
358 continue;
359
360 copyAttribute(LLVM::LLVMDialect::getReturnedAttrName());
361 copyAttribute(LLVM::LLVMDialect::getNoUndefAttrName());
362 copyAttribute(LLVM::LLVMDialect::getInRegAttrName());
363 bool lowersToPointer = false;
364 for (size_t i = 0, e = remapping->size; i < e; ++i) {
365 lowersToPointer |= isa<LLVM::LLVMPointerType>(
366 llvmFuncOp.getArgument(remapping->inputNo + i).getType());
367 }
368
369 if (lowersToPointer) {
370 copyPointerAttribute(LLVM::LLVMDialect::getNoAliasAttrName());
371 copyPointerAttribute(LLVM::LLVMDialect::getNoCaptureAttrName());
372 copyPointerAttribute(LLVM::LLVMDialect::getNoFreeAttrName());
373 copyPointerAttribute(LLVM::LLVMDialect::getAlignAttrName());
374 copyPointerAttribute(LLVM::LLVMDialect::getReadonlyAttrName());
375 copyPointerAttribute(LLVM::LLVMDialect::getWriteOnlyAttrName());
376 copyPointerAttribute(LLVM::LLVMDialect::getReadnoneAttrName());
377 copyPointerAttribute(LLVM::LLVMDialect::getNonNullAttrName());
378 copyPointerAttribute(LLVM::LLVMDialect::getDereferenceableAttrName());
379 copyPointerAttribute(
380 LLVM::LLVMDialect::getDereferenceableOrNullAttrName());
381 copyPointerAttribute(
382 LLVM::LLVMDialect::WorkgroupAttributionAttrHelper::getNameStr());
383 }
384 }
385 rewriter.eraseOp(op: gpuFuncOp);
386 return success();
387}
388
389LogicalResult GPUPrintfOpToHIPLowering::matchAndRewrite(
390 gpu::PrintfOp gpuPrintfOp, gpu::PrintfOpAdaptor adaptor,
391 ConversionPatternRewriter &rewriter) const {
392 Location loc = gpuPrintfOp->getLoc();
393
394 mlir::Type llvmI8 = typeConverter->convertType(rewriter.getI8Type());
395 auto ptrType = LLVM::LLVMPointerType::get(rewriter.getContext());
396 mlir::Type llvmI32 = typeConverter->convertType(rewriter.getI32Type());
397 mlir::Type llvmI64 = typeConverter->convertType(rewriter.getI64Type());
398 // Note: this is the GPUModule op, not the ModuleOp that surrounds it
399 // This ensures that global constants and declarations are placed within
400 // the device code, not the host code
401 auto moduleOp = gpuPrintfOp->getParentOfType<gpu::GPUModuleOp>();
402
403 auto ocklBegin =
404 getOrDefineFunction(moduleOp, loc, rewriter, "__ockl_printf_begin",
405 LLVM::LLVMFunctionType::get(llvmI64, {llvmI64}));
406 LLVM::LLVMFuncOp ocklAppendArgs;
407 if (!adaptor.getArgs().empty()) {
408 ocklAppendArgs = getOrDefineFunction(
409 moduleOp, loc, rewriter, "__ockl_printf_append_args",
410 LLVM::LLVMFunctionType::get(
411 llvmI64, {llvmI64, /*numArgs*/ llvmI32, llvmI64, llvmI64, llvmI64,
412 llvmI64, llvmI64, llvmI64, llvmI64, /*isLast*/ llvmI32}));
413 }
414 auto ocklAppendStringN = getOrDefineFunction(
415 moduleOp, loc, rewriter, "__ockl_printf_append_string_n",
416 LLVM::LLVMFunctionType::get(
417 llvmI64,
418 {llvmI64, ptrType, /*length (bytes)*/ llvmI64, /*isLast*/ llvmI32}));
419
420 /// Start the printf hostcall
421 Value zeroI64 = rewriter.create<LLVM::ConstantOp>(loc, llvmI64, 0);
422 auto printfBeginCall = rewriter.create<LLVM::CallOp>(loc, ocklBegin, zeroI64);
423 Value printfDesc = printfBeginCall.getResult();
424
425 // Create the global op or find an existing one.
426 LLVM::GlobalOp global = getOrCreateStringConstant(
427 rewriter, loc, moduleOp, llvmI8, "printfFormat_", adaptor.getFormat());
428
429 // Get a pointer to the format string's first element and pass it to printf()
430 Value globalPtr = rewriter.create<LLVM::AddressOfOp>(
431 loc,
432 LLVM::LLVMPointerType::get(rewriter.getContext(), global.getAddrSpace()),
433 global.getSymNameAttr());
434 Value stringStart =
435 rewriter.create<LLVM::GEPOp>(loc, ptrType, global.getGlobalType(),
436 globalPtr, ArrayRef<LLVM::GEPArg>{0, 0});
437 Value stringLen = rewriter.create<LLVM::ConstantOp>(
438 loc, llvmI64, cast<StringAttr>(global.getValueAttr()).size());
439
440 Value oneI32 = rewriter.create<LLVM::ConstantOp>(loc, llvmI32, 1);
441 Value zeroI32 = rewriter.create<LLVM::ConstantOp>(loc, llvmI32, 0);
442
443 auto appendFormatCall = rewriter.create<LLVM::CallOp>(
444 loc, ocklAppendStringN,
445 ValueRange{printfDesc, stringStart, stringLen,
446 adaptor.getArgs().empty() ? oneI32 : zeroI32});
447 printfDesc = appendFormatCall.getResult();
448
449 // __ockl_printf_append_args takes 7 values per append call
450 constexpr size_t argsPerAppend = 7;
451 size_t nArgs = adaptor.getArgs().size();
452 for (size_t group = 0; group < nArgs; group += argsPerAppend) {
453 size_t bound = std::min(a: group + argsPerAppend, b: nArgs);
454 size_t numArgsThisCall = bound - group;
455
456 SmallVector<mlir::Value, 2 + argsPerAppend + 1> arguments;
457 arguments.push_back(Elt: printfDesc);
458 arguments.push_back(
459 rewriter.create<LLVM::ConstantOp>(loc, llvmI32, numArgsThisCall));
460 for (size_t i = group; i < bound; ++i) {
461 Value arg = adaptor.getArgs()[i];
462 if (auto floatType = dyn_cast<FloatType>(arg.getType())) {
463 if (!floatType.isF64())
464 arg = rewriter.create<LLVM::FPExtOp>(
465 loc, typeConverter->convertType(rewriter.getF64Type()), arg);
466 arg = rewriter.create<LLVM::BitcastOp>(loc, llvmI64, arg);
467 }
468 if (arg.getType().getIntOrFloatBitWidth() != 64)
469 arg = rewriter.create<LLVM::ZExtOp>(loc, llvmI64, arg);
470
471 arguments.push_back(Elt: arg);
472 }
473 // Pad out to 7 arguments since the hostcall always needs 7
474 for (size_t extra = numArgsThisCall; extra < argsPerAppend; ++extra) {
475 arguments.push_back(Elt: zeroI64);
476 }
477
478 auto isLast = (bound == nArgs) ? oneI32 : zeroI32;
479 arguments.push_back(Elt: isLast);
480 auto call = rewriter.create<LLVM::CallOp>(loc, ocklAppendArgs, arguments);
481 printfDesc = call.getResult();
482 }
483 rewriter.eraseOp(op: gpuPrintfOp);
484 return success();
485}
486
487LogicalResult GPUPrintfOpToLLVMCallLowering::matchAndRewrite(
488 gpu::PrintfOp gpuPrintfOp, gpu::PrintfOpAdaptor adaptor,
489 ConversionPatternRewriter &rewriter) const {
490 Location loc = gpuPrintfOp->getLoc();
491
492 mlir::Type llvmI8 = typeConverter->convertType(rewriter.getIntegerType(8));
493 mlir::Type ptrType =
494 LLVM::LLVMPointerType::get(rewriter.getContext(), addressSpace);
495
496 // Note: this is the GPUModule op, not the ModuleOp that surrounds it
497 // This ensures that global constants and declarations are placed within
498 // the device code, not the host code
499 auto moduleOp = gpuPrintfOp->getParentOfType<gpu::GPUModuleOp>();
500
501 auto printfType =
502 LLVM::LLVMFunctionType::get(rewriter.getI32Type(), {ptrType},
503 /*isVarArg=*/true);
504 LLVM::LLVMFuncOp printfDecl =
505 getOrDefineFunction(moduleOp, loc, rewriter, "printf", printfType);
506
507 // Create the global op or find an existing one.
508 LLVM::GlobalOp global = getOrCreateStringConstant(
509 rewriter, loc, moduleOp, llvmI8, "printfFormat_", adaptor.getFormat(),
510 /*alignment=*/0, addressSpace);
511
512 // Get a pointer to the format string's first element
513 Value globalPtr = rewriter.create<LLVM::AddressOfOp>(
514 loc,
515 LLVM::LLVMPointerType::get(rewriter.getContext(), global.getAddrSpace()),
516 global.getSymNameAttr());
517 Value stringStart =
518 rewriter.create<LLVM::GEPOp>(loc, ptrType, global.getGlobalType(),
519 globalPtr, ArrayRef<LLVM::GEPArg>{0, 0});
520
521 // Construct arguments and function call
522 auto argsRange = adaptor.getArgs();
523 SmallVector<Value, 4> printfArgs;
524 printfArgs.reserve(N: argsRange.size() + 1);
525 printfArgs.push_back(Elt: stringStart);
526 printfArgs.append(argsRange.begin(), argsRange.end());
527
528 rewriter.create<LLVM::CallOp>(loc, printfDecl, printfArgs);
529 rewriter.eraseOp(op: gpuPrintfOp);
530 return success();
531}
532
533LogicalResult GPUPrintfOpToVPrintfLowering::matchAndRewrite(
534 gpu::PrintfOp gpuPrintfOp, gpu::PrintfOpAdaptor adaptor,
535 ConversionPatternRewriter &rewriter) const {
536 Location loc = gpuPrintfOp->getLoc();
537
538 mlir::Type llvmI8 = typeConverter->convertType(rewriter.getIntegerType(8));
539 mlir::Type ptrType = LLVM::LLVMPointerType::get(rewriter.getContext());
540
541 // Note: this is the GPUModule op, not the ModuleOp that surrounds it
542 // This ensures that global constants and declarations are placed within
543 // the device code, not the host code
544 auto moduleOp = gpuPrintfOp->getParentOfType<gpu::GPUModuleOp>();
545
546 // Create a valid global location removing any metadata attached to the
547 // location as debug info metadata inside of a function cannot be used outside
548 // of that function.
549 Location globalLoc = loc->findInstanceOfOrUnknown<FileLineColLoc>();
550
551 auto vprintfType =
552 LLVM::LLVMFunctionType::get(rewriter.getI32Type(), {ptrType, ptrType});
553 LLVM::LLVMFuncOp vprintfDecl = getOrDefineFunction(
554 moduleOp, globalLoc, rewriter, "vprintf", vprintfType);
555
556 // Create the global op or find an existing one.
557 LLVM::GlobalOp global =
558 getOrCreateStringConstant(rewriter, globalLoc, moduleOp, llvmI8,
559 "printfFormat_", adaptor.getFormat());
560
561 // Get a pointer to the format string's first element
562 Value globalPtr = rewriter.create<LLVM::AddressOfOp>(loc, global);
563 Value stringStart =
564 rewriter.create<LLVM::GEPOp>(loc, ptrType, global.getGlobalType(),
565 globalPtr, ArrayRef<LLVM::GEPArg>{0, 0});
566 SmallVector<Type> types;
567 SmallVector<Value> args;
568 // Promote and pack the arguments into a stack allocation.
569 for (Value arg : adaptor.getArgs()) {
570 Type type = arg.getType();
571 Value promotedArg = arg;
572 assert(type.isIntOrFloat());
573 if (isa<FloatType>(type)) {
574 type = rewriter.getF64Type();
575 promotedArg = rewriter.create<LLVM::FPExtOp>(loc, type, arg);
576 }
577 types.push_back(type);
578 args.push_back(promotedArg);
579 }
580 Type structType =
581 LLVM::LLVMStructType::getLiteral(gpuPrintfOp.getContext(), types);
582 Value one = rewriter.create<LLVM::ConstantOp>(loc, rewriter.getI64Type(),
583 rewriter.getIndexAttr(1));
584 Value tempAlloc =
585 rewriter.create<LLVM::AllocaOp>(loc, ptrType, structType, one,
586 /*alignment=*/0);
587 for (auto [index, arg] : llvm::enumerate(First&: args)) {
588 Value ptr = rewriter.create<LLVM::GEPOp>(
589 loc, ptrType, structType, tempAlloc,
590 ArrayRef<LLVM::GEPArg>{0, static_cast<int32_t>(index)});
591 rewriter.create<LLVM::StoreOp>(loc, arg, ptr);
592 }
593 std::array<Value, 2> printfArgs = {stringStart, tempAlloc};
594
595 rewriter.create<LLVM::CallOp>(loc, vprintfDecl, printfArgs);
596 rewriter.eraseOp(op: gpuPrintfOp);
597 return success();
598}
599
600/// Helper for impl::scalarizeVectorOp. Scalarizes vectors to elements.
601/// Used either directly (for ops on 1D vectors) or as the callback passed to
602/// detail::handleMultidimensionalVectors (for ops on higher-rank vectors).
603static Value scalarizeVectorOpHelper(Operation *op, ValueRange operands,
604 Type llvm1DVectorTy,
605 ConversionPatternRewriter &rewriter,
606 const LLVMTypeConverter &converter) {
607 TypeRange operandTypes(operands);
608 VectorType vectorType = cast<VectorType>(llvm1DVectorTy);
609 Location loc = op->getLoc();
610 Value result = rewriter.create<LLVM::PoisonOp>(loc, vectorType);
611 Type indexType = converter.convertType(rewriter.getIndexType());
612 StringAttr name = op->getName().getIdentifier();
613 Type elementType = vectorType.getElementType();
614
615 for (int64_t i = 0; i < vectorType.getNumElements(); ++i) {
616 Value index = rewriter.create<LLVM::ConstantOp>(loc, indexType, i);
617 auto extractElement = [&](Value operand) -> Value {
618 if (!isa<VectorType>(Val: operand.getType()))
619 return operand;
620 return rewriter.create<LLVM::ExtractElementOp>(loc, operand, index);
621 };
622 auto scalarOperands = llvm::map_to_vector(C&: operands, F&: extractElement);
623 Operation *scalarOp =
624 rewriter.create(loc, name, scalarOperands, elementType, op->getAttrs());
625 result = rewriter.create<LLVM::InsertElementOp>(
626 loc, result, scalarOp->getResult(0), index);
627 }
628 return result;
629}
630
631/// Unrolls op to array/vector elements.
632LogicalResult impl::scalarizeVectorOp(Operation *op, ValueRange operands,
633 ConversionPatternRewriter &rewriter,
634 const LLVMTypeConverter &converter) {
635 TypeRange operandTypes(operands);
636 if (llvm::any_of(Range&: operandTypes, P: llvm::IsaPred<VectorType>)) {
637 VectorType vectorType =
638 cast<VectorType>(converter.convertType(op->getResultTypes()[0]));
639 rewriter.replaceOp(op, scalarizeVectorOpHelper(op, operands, vectorType,
640 rewriter, converter));
641 return success();
642 }
643
644 if (llvm::any_of(operandTypes, llvm::IsaPred<LLVM::LLVMArrayType>)) {
645 return LLVM::detail::handleMultidimensionalVectors(
646 op, operands, typeConverter: converter,
647 createOperand: [&](Type llvm1DVectorTy, ValueRange operands) -> Value {
648 return scalarizeVectorOpHelper(op, operands, llvm1DVectorTy, rewriter,
649 converter);
650 },
651 rewriter);
652 }
653
654 return rewriter.notifyMatchFailure(arg&: op, msg: "no llvm.array or vector to unroll");
655}
656
657static IntegerAttr wrapNumericMemorySpace(MLIRContext *ctx, unsigned space) {
658 return IntegerAttr::get(IntegerType::get(ctx, 64), space);
659}
660
661/// Generates a symbol with 0-sized array type for dynamic shared memory usage,
662/// or uses existing symbol.
663LLVM::GlobalOp getDynamicSharedMemorySymbol(
664 ConversionPatternRewriter &rewriter, gpu::GPUModuleOp moduleOp,
665 gpu::DynamicSharedMemoryOp op, const LLVMTypeConverter *typeConverter,
666 MemRefType memrefType, unsigned alignmentBit) {
667 uint64_t alignmentByte = alignmentBit / memrefType.getElementTypeBitWidth();
668
669 FailureOr<unsigned> addressSpace =
670 typeConverter->getMemRefAddressSpace(type: memrefType);
671 if (failed(Result: addressSpace)) {
672 op->emitError() << "conversion of memref memory space "
673 << memrefType.getMemorySpace()
674 << " to integer address space "
675 "failed. Consider adding memory space conversions.";
676 }
677
678 // Step 1. Collect symbol names of LLVM::GlobalOp Ops. Also if any of
679 // LLVM::GlobalOp is suitable for shared memory, return it.
680 llvm::StringSet<> existingGlobalNames;
681 for (auto globalOp : moduleOp.getBody()->getOps<LLVM::GlobalOp>()) {
682 existingGlobalNames.insert(globalOp.getSymName());
683 if (auto arrayType = dyn_cast<LLVM::LLVMArrayType>(globalOp.getType())) {
684 if (globalOp.getAddrSpace() == addressSpace.value() &&
685 arrayType.getNumElements() == 0 &&
686 globalOp.getAlignment().value_or(0) == alignmentByte) {
687 return globalOp;
688 }
689 }
690 }
691
692 // Step 2. Find a unique symbol name
693 unsigned uniquingCounter = 0;
694 SmallString<128> symName = SymbolTable::generateSymbolName<128>(
695 name: "__dynamic_shmem_",
696 uniqueChecker: [&](StringRef candidate) {
697 return existingGlobalNames.contains(key: candidate);
698 },
699 uniquingCounter);
700
701 // Step 3. Generate a global op
702 OpBuilder::InsertionGuard guard(rewriter);
703 rewriter.setInsertionPointToStart(moduleOp.getBody());
704
705 auto zeroSizedArrayType = LLVM::LLVMArrayType::get(
706 typeConverter->convertType(memrefType.getElementType()), 0);
707
708 return rewriter.create<LLVM::GlobalOp>(
709 op->getLoc(), zeroSizedArrayType, /*isConstant=*/false,
710 LLVM::Linkage::Internal, symName, /*value=*/Attribute(), alignmentByte,
711 addressSpace.value());
712}
713
714LogicalResult GPUDynamicSharedMemoryOpLowering::matchAndRewrite(
715 gpu::DynamicSharedMemoryOp op, OpAdaptor adaptor,
716 ConversionPatternRewriter &rewriter) const {
717 Location loc = op.getLoc();
718 MemRefType memrefType = op.getResultMemref().getType();
719 Type elementType = typeConverter->convertType(memrefType.getElementType());
720
721 // Step 1: Generate a memref<0xi8> type
722 MemRefLayoutAttrInterface layout = {};
723 auto memrefType0sz =
724 MemRefType::get({0}, elementType, layout, memrefType.getMemorySpace());
725
726 // Step 2: Generate a global symbol or existing for the dynamic shared
727 // memory with memref<0xi8> type
728 auto moduleOp = op->getParentOfType<gpu::GPUModuleOp>();
729 LLVM::GlobalOp shmemOp = getDynamicSharedMemorySymbol(
730 rewriter, moduleOp, op, getTypeConverter(), memrefType0sz, alignmentBit);
731
732 // Step 3. Get address of the global symbol
733 OpBuilder::InsertionGuard guard(rewriter);
734 rewriter.setInsertionPoint(op);
735 auto basePtr = rewriter.create<LLVM::AddressOfOp>(loc, shmemOp);
736 Type baseType = basePtr->getResultTypes().front();
737
738 // Step 4. Generate GEP using offsets
739 SmallVector<LLVM::GEPArg> gepArgs = {0};
740 Value shmemPtr = rewriter.create<LLVM::GEPOp>(loc, baseType, elementType,
741 basePtr, gepArgs);
742 // Step 5. Create a memref descriptor
743 SmallVector<Value> shape, strides;
744 Value sizeBytes;
745 getMemRefDescriptorSizes(loc, memrefType0sz, {}, rewriter, shape, strides,
746 sizeBytes);
747 auto memRefDescriptor = this->createMemRefDescriptor(
748 loc, memrefType0sz, shmemPtr, shmemPtr, shape, strides, rewriter);
749
750 // Step 5. Replace the op with memref descriptor
751 rewriter.replaceOp(op, {memRefDescriptor});
752 return success();
753}
754
755LogicalResult GPUReturnOpLowering::matchAndRewrite(
756 gpu::ReturnOp op, OpAdaptor adaptor,
757 ConversionPatternRewriter &rewriter) const {
758 Location loc = op.getLoc();
759 unsigned numArguments = op.getNumOperands();
760 SmallVector<Value, 4> updatedOperands;
761
762 bool useBarePtrCallConv = getTypeConverter()->getOptions().useBarePtrCallConv;
763 if (useBarePtrCallConv) {
764 // For the bare-ptr calling convention, extract the aligned pointer to
765 // be returned from the memref descriptor.
766 for (auto it : llvm::zip(op->getOperands(), adaptor.getOperands())) {
767 Type oldTy = std::get<0>(it).getType();
768 Value newOperand = std::get<1>(it);
769 if (isa<MemRefType>(oldTy) && getTypeConverter()->canConvertToBarePtr(
770 cast<BaseMemRefType>(oldTy))) {
771 MemRefDescriptor memrefDesc(newOperand);
772 newOperand = memrefDesc.allocatedPtr(rewriter, loc);
773 } else if (isa<UnrankedMemRefType>(oldTy)) {
774 // Unranked memref is not supported in the bare pointer calling
775 // convention.
776 return failure();
777 }
778 updatedOperands.push_back(newOperand);
779 }
780 } else {
781 updatedOperands = llvm::to_vector<4>(adaptor.getOperands());
782 (void)copyUnrankedDescriptors(rewriter, loc, op.getOperands().getTypes(),
783 updatedOperands,
784 /*toDynamic=*/true);
785 }
786
787 // If ReturnOp has 0 or 1 operand, create it and return immediately.
788 if (numArguments <= 1) {
789 rewriter.replaceOpWithNewOp<LLVM::ReturnOp>(
790 op, TypeRange(), updatedOperands, op->getAttrs());
791 return success();
792 }
793
794 // Otherwise, we need to pack the arguments into an LLVM struct type before
795 // returning.
796 auto packedType = getTypeConverter()->packFunctionResults(
797 op.getOperandTypes(), useBarePtrCallConv);
798 if (!packedType) {
799 return rewriter.notifyMatchFailure(op, "could not convert result types");
800 }
801
802 Value packed = rewriter.create<LLVM::PoisonOp>(loc, packedType);
803 for (auto [idx, operand] : llvm::enumerate(First&: updatedOperands)) {
804 packed = rewriter.create<LLVM::InsertValueOp>(loc, packed, operand, idx);
805 }
806 rewriter.replaceOpWithNewOp<LLVM::ReturnOp>(op, TypeRange(), packed,
807 op->getAttrs());
808 return success();
809}
810
811void mlir::populateGpuMemorySpaceAttributeConversions(
812 TypeConverter &typeConverter, const MemorySpaceMapping &mapping) {
813 typeConverter.addTypeAttributeConversion(
814 callback: [mapping](BaseMemRefType type, gpu::AddressSpaceAttr memorySpaceAttr) {
815 gpu::AddressSpace memorySpace = memorySpaceAttr.getValue();
816 unsigned addressSpace = mapping(memorySpace);
817 return wrapNumericMemorySpace(memorySpaceAttr.getContext(),
818 addressSpace);
819 });
820}
821

source code of mlir/lib/Conversion/GPUCommon/GPUOpsLowering.cpp