bugfix for: 'aisle.GEMM' op operand onnx#1 must be tensor of 32-bit s…

…ignless integer values, but got 'tensor<1x4xi64>
ISOLDE-Project · Jun 20, 2024 · 6574582 · 6574582
1 parent 7e76263
commit 6574582
Showing 1 changed file with 9 additions and 9 deletions.
diff --git a/src/Conversion/ONNXToAISLE/helper.hpp b/src/Conversion/ONNXToAISLE/helper.hpp
@@ -15,7 +15,7 @@
 using namespace mlir;
 namespace onnx_to_aisle {
 
-inline void getVectorShape(Value value, llvm::SmallVector<int64_t>&Vec) {
+inline void getVectorShape(Value value, llvm::SmallVector<int32_t>&Vec) {
   TensorType t_type = value.getType().cast<TensorType>();
   if (!t_type)
     return;
@@ -49,16 +49,16 @@ inline void getAttrValues(::mlir::ArrayAttr array, SmallVector<int> &Vec) {
 template <typename _Class, typename _Func, typename _Operation>
 spade::AISLEQConstantOp create(ConversionPatternRewriter &rewriter,
     _Operation &op, const char *name, _Func mem_func) {
-  auto elementType = rewriter.getI64Type();
+  auto elementType = rewriter.getI32Type();
   auto paramShape = mlir::RankedTensorType::get(
-      llvm::ArrayRef<std::int64_t>{1, 4}, elementType);
+      llvm::ArrayRef<int64_t>{1, 4}, elementType);
   _Class adaptor(op);
   auto get_val = std::mem_fn(mem_func);
   Value X = get_val(adaptor);
-  SmallVector<int64_t> Vec;
+  SmallVector<int32_t> Vec;
   onnx_to_aisle::getVectorShape(X, Vec);
   auto input_shape_dense =
-      DenseElementsAttr::get(paramShape, ArrayRef<int64_t>(Vec));
+      DenseElementsAttr::get(paramShape, ArrayRef<int32_t>(Vec));
   auto iShapeParam = rewriter.create<spade::AISLEQConstantOp>(
       op.getLoc(), name, input_shape_dense);
   return iShapeParam;
@@ -67,10 +67,10 @@ spade::AISLEQConstantOp create(ConversionPatternRewriter &rewriter,
 template <typename _Operation>
 spade::AISLEQConstantOp create(ConversionPatternRewriter &rewriter,
     _Operation &op, const char *name, const SmallVector<int64_t> &shape) {
-  auto elementType = rewriter.getI64Type();
+  auto elementType = rewriter.getI32Type();
   auto paramShape = mlir::RankedTensorType::get(
-      llvm::ArrayRef<std::int64_t>{1, 4}, elementType);
-  SmallVector<int64_t> Vec;
+      llvm::ArrayRef<int64_t>{1, 4}, elementType);
+  SmallVector<int32_t> Vec;
   //make it rank 4
   auto rank = shape.size();
   /*
@@ -86,7 +86,7 @@ spade::AISLEQConstantOp create(ConversionPatternRewriter &rewriter,
   for (size_t i = 0; i < rank; ++i)
     Vec.push_back(shape[i]);
   auto input_shape_dense =
-      DenseElementsAttr::get(paramShape, ArrayRef<int64_t>(Vec));
+      DenseElementsAttr::get(paramShape, ArrayRef<int32_t>(Vec));
   auto iShapeParam = rewriter.create<spade::AISLEQConstantOp>(
       op.getLoc(), name, input_shape_dense);
   return iShapeParam;