[mlir][gpu] Introduce gpu.dynamic_shared_memory Op

mlir/include/mlir/Dialect/GPU/IR/GPUBase.td

@@ -52,6 +52,14 @@ def GPU_Dialect : Dialect {
     /// Returns the numeric value used to identify the private memory address
     /// space.
     static AddressSpace getPrivateAddressSpace() { return AddressSpace::Private; }
+
+    /// Return true if the given MemRefType has an address space that matches 
+    /// with the gpu::AddressSpaceAttr attribute with value 'workgroup`.
+    static bool hasWorkgroupMemoryAddressSpace(MemRefType type);
+
+    /// Return true if the given Attribute is an gpu::AddressSpaceAttr 
+    /// attribute with value 'workgroup`.
+    static bool isWorkgroupMemoryAddressSpace(Attribute memorySpace);  
   }];
 
   let dependentDialects = ["arith::ArithDialect"];

mlir/include/mlir/Dialect/GPU/IR/GPUOps.td

@@ -433,6 +433,32 @@ def GPU_GPUFuncOp : GPU_Op<"func", [
   let hasVerifier = 1;
 }
 
+def GPU_DynamicSharedMemoryOp : GPU_Op<"dynamic_shared_memory", [Pure]>
+{
+  let summary = "Get the memref for dynamic shared memory";
+
+  let description = [{
+    This operation provides a memref pointer to the start of dynamic shared 
+    memory, often referred to as workgroup memory. It's important to note that
+    this dynamic shared memory needs to be allocated at kernel launch. One can 
+    conveniently utilize `the dynamic_shared_memory_size` parameter of 
+    `gpu.launch` for this purpose.
+
+    Examples: 
+    ```mlir        
+    %0 = gpu.dynamic.shared.memory : memref<?xi8, #gpu.address_space<workgroup>>
+    %1 = memref.view %0[%c8192][] : memref<?xi8, #gpu.address_space<workgroup>> 
+                            to memref<32x64xf32, #gpu.address_space<workgroup>>
+    %2 = memref.view %0[%c16384][] : memref<?xi8, #gpu.address_space<workgroup>> 
+                            to memref<32x64xf32, #gpu.address_space<workgroup>>
+    ```
+  }];  
+  let arguments = (ins);
+  let results = (outs Arg<MemRefRankOf<[I8], [1]>>:$resultMemref);
+  let assemblyFormat = [{ attr-dict `:` type($resultMemref) }];
+  let hasVerifier = 1;
+}
+
 def LaunchIndx : AnyTypeOf<[Index, I32, I64]>;
 
 def GPU_LaunchFuncOp :GPU_Op<"launch_func", [

mlir/include/mlir/Dialect/LLVMIR/NVVMDialect.h

@@ -27,6 +27,9 @@
 namespace mlir {
 namespace NVVM {
 
+// Shared memory has 128-bit alignment
+constexpr int kSharedMemoryAlignmentBit = 128;
+
 /// NVVM memory space identifiers.
 enum NVVMMemorySpace {
   /// Global memory space identifier.

mlir/include/mlir/IR/SymbolTable.h

@@ -103,6 +103,24 @@ class SymbolTable {
     Nested,
   };
 
+  /// Generate a unique symbol name. Iteratively increase uniquingCounter
+  /// and use it as a suffix for symbol names until uniqueChecker does not
+  /// detect any conflict.
+  template <unsigned N, typename UniqueChecker>
+  static SmallString<N> generateSymbolName(StringRef name,
+                                           UniqueChecker uniqueChecker,
+                                           unsigned &uniquingCounter) {
+    SmallString<N> nameBuffer(name);
+    unsigned originalLength = nameBuffer.size();
+    do {
+      nameBuffer.resize(originalLength);
+      nameBuffer += '_';
+      nameBuffer += std::to_string(uniquingCounter++);
+    } while (uniqueChecker(nameBuffer));
+
+    return nameBuffer;
+  }
+
   /// Returns the name of the given symbol operation, aborting if no symbol is
   /// present.
   static StringAttr getSymbolName(Operation *symbol);

mlir/lib/Conversion/GPUCommon/GPUOpsLowering.cpp

@@ -14,6 +14,7 @@
 #include "mlir/IR/Builders.h"
 #include "mlir/IR/BuiltinTypes.h"
 #include "llvm/ADT/SmallVectorExtras.h"
+#include "llvm/ADT/StringSet.h"
 #include "llvm/Support/FormatVariadic.h"
 
 using namespace mlir;
@@ -554,6 +555,104 @@ static IntegerAttr wrapNumericMemorySpace(MLIRContext *ctx, unsigned space) {
   return IntegerAttr::get(IntegerType::get(ctx, 64), space);
 }
 
+/// Generates a symbol with 0-sized array type for dynamic shared memory usage,
+/// or uses existing symbol.
+LLVM::GlobalOp
+getDynamicSharedMemorySymbol(ConversionPatternRewriter &rewriter,
+                             Operation *moduleOp, gpu::DynamicSharedMemoryOp op,
+                             const LLVMTypeConverter *typeConverter,
+                             MemRefType memrefType, unsigned alignmentBit) {
+  uint64_t alignmentByte = alignmentBit / memrefType.getElementTypeBitWidth();
+
+  FailureOr<unsigned> addressSpace =
+      typeConverter->getMemRefAddressSpace(memrefType);
+  if (failed(addressSpace)) {
+    op->emitError() << "conversion of memref memory space "
+                    << memrefType.getMemorySpace()
+                    << " to integer address space "
+                       "failed. Consider adding memory space conversions.";
+  }
+
+  // Step 1. Collect symbol names of LLVM::GlobalOp Ops. Also if any of
+  // LLVM::GlobalOp is suitable for shared memory, return it.
+  llvm::StringSet<> existingGlobalNames;
+  for (auto globalOp :
+       moduleOp->getRegion(0).front().getOps<LLVM::GlobalOp>()) {
+    existingGlobalNames.insert(globalOp.getSymName());
+    if (auto arrayType = dyn_cast<LLVM::LLVMArrayType>(globalOp.getType())) {
+      if (globalOp.getAddrSpace() == addressSpace.value() &&
+          arrayType.getNumElements() == 0 &&
+          globalOp.getAlignment().value_or(0) == alignmentByte) {
+        return globalOp;
+      }
+    }
+  }
+
+  // Step 2. Find a unique symbol name
+  unsigned uniquingCounter = 0;
+  SmallString<128> symName = SymbolTable::generateSymbolName<128>(
+      "__dynamic_shmem_",
+      [&](StringRef candidate) {
+        return existingGlobalNames.contains(candidate);
+      },
+      uniquingCounter);
+
+  // Step 3. Generate a global op
+  OpBuilder::InsertionGuard guard(rewriter);
+  rewriter.setInsertionPoint(&moduleOp->getRegion(0).front().front());
+
+  auto zeroSizedArrayType = LLVM::LLVMArrayType::get(
+      typeConverter->convertType(memrefType.getElementType()), 0);
+
+  return rewriter.create<LLVM::GlobalOp>(
+      op->getLoc(), zeroSizedArrayType, /*isConstant=*/false,
+      LLVM::Linkage::Internal, symName, /*value=*/Attribute(), alignmentByte,
+      addressSpace.value());
+}
+
+LogicalResult GPUDynamicSharedMemoryOpLowering::matchAndRewrite(
+    gpu::DynamicSharedMemoryOp op, OpAdaptor adaptor,
+    ConversionPatternRewriter &rewriter) const {
+  Location loc = op.getLoc();
+  MemRefType memrefType = op.getResultMemref().getType();
+  Type elementType = typeConverter->convertType(memrefType.getElementType());
+
+  // Step 1: Generate a memref<0xi8> type
+  MemRefLayoutAttrInterface layout = {};
+  auto memrefType0sz =
+      MemRefType::get({0}, elementType, layout, memrefType.getMemorySpace());
+
+  // Step 2: Generate a global symbol or existing for the dynamic shared
+  // memory with memref<0xi8> type
+  LLVM::LLVMFuncOp funcOp = op->getParentOfType<LLVM::LLVMFuncOp>();
+  LLVM::GlobalOp shmemOp = {};
+  Operation *moduleOp = funcOp->getParentWithTrait<OpTrait::SymbolTable>();
+  shmemOp = getDynamicSharedMemorySymbol(
+      rewriter, moduleOp, op, getTypeConverter(), memrefType0sz, alignmentBit);
+
+  // Step 3. Get address of the global symbol
+  OpBuilder::InsertionGuard guard(rewriter);
+  rewriter.setInsertionPoint(op);
+  auto basePtr = rewriter.create<LLVM::AddressOfOp>(loc, shmemOp);
+  Type baseType = basePtr->getResultTypes().front();
+
+  // Step 4. Generate GEP using offsets
+  SmallVector<LLVM::GEPArg> gepArgs = {0};
+  Value shmemPtr = rewriter.create<LLVM::GEPOp>(loc, baseType, elementType,
+                                                basePtr, gepArgs);
+  // Step 5. Create a memref descriptor
+  SmallVector<Value> shape, strides;
+  Value sizeBytes;
+  getMemRefDescriptorSizes(loc, memrefType0sz, {}, rewriter, shape, strides,
+                           sizeBytes);
+  auto memRefDescriptor = this->createMemRefDescriptor(
+      loc, memrefType0sz, shmemPtr, shmemPtr, shape, strides, rewriter);
+
+  // Step 5. Replace the op with memref descriptor
+  rewriter.replaceOp(op, {memRefDescriptor});
+  return success();
+}
+
 void mlir::populateGpuMemorySpaceAttributeConversions(
     TypeConverter &typeConverter, const MemorySpaceMapping &mapping) {
   typeConverter.addTypeAttributeConversion(

mlir/lib/Conversion/GPUCommon/GPUOpsLowering.h

@@ -14,6 +14,27 @@
 
 namespace mlir {
 
+/// Lowering for gpu.dynamic.shared.memory to LLVM dialect. The pattern first
+/// create a 0-sized global array symbol similar as LLVM expects. It constructs
+/// a memref descriptor with these values and return it.
+struct GPUDynamicSharedMemoryOpLowering
+    : public ConvertOpToLLVMPattern<gpu::DynamicSharedMemoryOp> {
+  using ConvertOpToLLVMPattern<
+      gpu::DynamicSharedMemoryOp>::ConvertOpToLLVMPattern;
+  GPUDynamicSharedMemoryOpLowering(const LLVMTypeConverter &converter,
+                                   unsigned alignmentBit = 0)
+      : ConvertOpToLLVMPattern<gpu::DynamicSharedMemoryOp>(converter),
+        alignmentBit(alignmentBit) {}
+
+  LogicalResult
+  matchAndRewrite(gpu::DynamicSharedMemoryOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override;
+
+private:
+  // Alignment bit
+  unsigned alignmentBit;
+};
+
 struct GPUFuncOpLowering : ConvertOpToLLVMPattern<gpu::GPUFuncOp> {
   GPUFuncOpLowering(const LLVMTypeConverter &converter,
                     unsigned allocaAddrSpace, unsigned workgroupAddrSpace,

mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp

@@ -325,6 +325,9 @@ void mlir::populateGpuToNVVMConversionPatterns(LLVMTypeConverter &converter,
            GPULaneIdOpToNVVM, GPUShuffleOpLowering, GPUReturnOpLowering>(
           converter);
 
+  patterns.add<GPUDynamicSharedMemoryOpLowering>(
+      converter, NVVM::kSharedMemoryAlignmentBit);
+
   // Explicitly drop memory space when lowering private memory
   // attributions since NVVM models it as `alloca`s in the default
   // memory space and does not support `alloca`s with addrspace(5).

mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

@@ -23,6 +23,7 @@
 #include "mlir/IR/Matchers.h"
 #include "mlir/IR/OpImplementation.h"
 #include "mlir/IR/PatternMatch.h"
+#include "mlir/IR/SymbolTable.h"
 #include "mlir/IR/TypeUtilities.h"
 #include "mlir/Interfaces/FunctionImplementation.h"
 #include "mlir/Interfaces/SideEffectInterfaces.h"
@@ -164,17 +165,18 @@ MMAMatrixType::verify(function_ref<InFlightDiagnostic()> emitError,
 // GPUDialect
 //===----------------------------------------------------------------------===//
 
-/// GPU memory space identifiers.
-enum GPUMemorySpace {
-  /// Generic memory space identifier.
-  kGenericMemorySpace = 0,
-
-  /// Global memory space identifier.
-  kGlobalMemorySpace = 1,
+bool GPUDialect::isWorkgroupMemoryAddressSpace(Attribute memorySpace) {
+  if (!memorySpace)
+    return false;
+  if (auto gpuAttr = llvm::dyn_cast<gpu::AddressSpaceAttr>(memorySpace))
+    return gpuAttr.getValue() == getWorkgroupAddressSpace();
+  return false;
+}
 
-  /// Shared memory space identifier.
-  kSharedMemorySpace = 3
-};
+bool GPUDialect::hasWorkgroupMemoryAddressSpace(MemRefType type) {
+  Attribute memorySpace = type.getMemorySpace();
+  return isWorkgroupMemoryAddressSpace(memorySpace);
+}
 
 bool GPUDialect::isKernel(Operation *op) {
   UnitAttr isKernelAttr = op->getAttrOfType<UnitAttr>(getKernelFuncAttrName());
@@ -2024,6 +2026,28 @@ gpu::SelectObjectAttr::verify(function_ref<InFlightDiagnostic()> emitError,
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// DynamicSharedMemoryOp
+//===----------------------------------------------------------------------===//
+
+LogicalResult gpu::DynamicSharedMemoryOp::verify() {
+  if (!getOperation()->getParentWithTrait<OpTrait::SymbolTable>())
+    return emitOpError() << "must be inside an op with symbol table";
+
+  MemRefType memrefType = getResultMemref().getType();
+  // Check address space
+  if (!GPUDialect::hasWorkgroupMemoryAddressSpace(memrefType)) {
+    return emitOpError() << "address space must be "
+                         << gpu::AddressSpaceAttr::getMnemonic() << "<"
+                         << stringifyEnum(gpu::AddressSpace::Workgroup) << ">";
+  }
+  if (memrefType.hasStaticShape()) {
+    return emitOpError() << "result memref type must be memref<?xi8, "
+                            "#gpu.address_space<workgroup>>";
+  }
+  return success();
+}
+
 //===----------------------------------------------------------------------===//
 // GPU target options
 //===----------------------------------------------------------------------===//

mlir/lib/IR/SymbolTable.cpp

@@ -200,20 +200,16 @@ StringAttr SymbolTable::insert(Operation *symbol, Block::iterator insertPt) {
   // If the symbol was already in the table, also return.
   if (symbolTable.lookup(name) == symbol)
     return name;
-  // If a conflict was detected, then the symbol will not have been added to
-  // the symbol table. Try suffixes until we get to a unique name that works.
-  SmallString<128> nameBuffer(name.getValue());
-  unsigned originalLength = nameBuffer.size();
 
   MLIRContext *context = symbol->getContext();
-
-  // Iteratively try suffixes until we find one that isn't used.
-  do {
-    nameBuffer.resize(originalLength);
-    nameBuffer += '_';
-    nameBuffer += std::to_string(uniquingCounter++);
-  } while (!symbolTable.insert({StringAttr::get(context, nameBuffer), symbol})
-                .second);
+  SmallString<128> nameBuffer = generateSymbolName<128>(
+      name.getValue(),
+      [&](StringRef candidate) {
+        return !symbolTable
+                    .insert({StringAttr::get(context, candidate), symbol})
+                    .second;
+      },
+      uniquingCounter);
   setSymbolName(symbol, nameBuffer);
   return getSymbolName(symbol);
 }