[Mlir-commits] [mlir] [mlir][Target] Teach dense_resource conversion to LLVMIR Target (PR #78958)

Kunwar Grover llvmlistbot at llvm.org
Mon Jan 22 10:27:36 PST 2024


================
@@ -446,6 +447,90 @@ convertDenseElementsAttr(Location loc, DenseElementsAttr denseElementsAttr,
   return buildSequentialConstant(constantsRef, outerShape, llvmType, loc);
 }
 
+/// Convert a dense resource elements attribute to an LLVM IR constant using its
+/// raw data storage if possible. This supports elements attributes of tensor or
+/// vector type and avoids constructing separate objects for individual values
+/// of the innermost dimension. Constants for other dimensions are still
+/// constructed recursively. Returns null if constructing from raw data is not
+/// supported for this type, e.g., element type is not a power-of-two-sized
+/// primitive. Reports other errors at `loc`.
+static llvm::Constant *convertDenseResourceElementsAttr(
+    Location loc, DenseResourceElementsAttr denseResourceAttr,
+    llvm::Type *llvmType, const ModuleTranslation &moduleTranslation) {
+  if (!denseResourceAttr)
+    return nullptr;
+
+  llvm::Type *innermostLLVMType = getInnermostElementType(llvmType);
+  if (!llvm::ConstantDataSequential::isElementTypeCompatible(innermostLLVMType))
+    return nullptr;
+
+  ShapedType type = denseResourceAttr.getType();
+  if (type.getNumElements() == 0)
+    return nullptr;
+
+  ArrayRef<char> rawData =
+      denseResourceAttr.getRawHandle().getBlob()->getData();
+
+  // Check that the raw data size matches what is expected for the scalar size.
+  // TODO: in theory, we could repack the data here to keep constructing from
+  // raw data.
+  // TODO: we may also need to consider endianness when cross-compiling to an
+  // architecture where it is different.
+  int64_t numElements = denseResourceAttr.getType().getNumElements();
+  int64_t elementByteSize = rawData.size() / numElements;
+  if (8 * elementByteSize != innermostLLVMType->getScalarSizeInBits())
+    return nullptr;
+
+  // Compute the shape of all dimensions but the innermost. Note that the
+  // innermost dimension may be that of the vector element type.
+  bool hasVectorElementType = isa<VectorType>(type.getElementType());
+  int64_t numAggregates =
+      numElements / (hasVectorElementType
+                         ? 1
+                         : denseResourceAttr.getType().getShape().back());
----------------
Groverkss wrote:

There is a check for getNumElements == 0 at the start. So that case should never arise here.

https://github.com/llvm/llvm-project/pull/78958


More information about the Mlir-commits mailing list