From ac6c376163e98c13b72fa4ac3e374ee98b082765 Mon Sep 17 00:00:00 2001
From: Tai Ly <tai.ly@arm.com>
Date: Wed, 11 Oct 2023 19:15:57 +0000
Subject: [tosa_mlir_translator] Fix pad deserialization

This patch fixes deserialization of Pad Op's padding input
to have shape {rank(shape_1)/2, 2}

Signed-off-by: Tai Ly <tai.ly@arm.com>
Change-Id: Ia9cee740b75985a5c605503ee4c97b9df220de0a
---
 src/TosaDeserialize.cpp | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/src/TosaDeserialize.cpp b/src/TosaDeserialize.cpp
index 031c57f..612e8aa 100644
--- a/src/TosaDeserialize.cpp
+++ b/src/TosaDeserialize.cpp
@@ -282,13 +282,12 @@ BuildDenseI16ElementsAttr(mlir::OpBuilder *op_builder,
 template <class T>
 mlir::DenseElementsAttr
 BuildDenseI32ElementsAttr(mlir::OpBuilder *op_builder,
+                          mlir::RankedTensorType &type,
                           const std::vector<T> &values) {
   llvm::SmallVector<int32_t> vec;
   for (auto val : values) {
     vec.push_back(val);
   }
-  auto type = mlir::RankedTensorType::get({static_cast<int64_t>(vec.size())},
-                                          op_builder->getI32Type());
   return mlir::DenseElementsAttr::get(type, llvm::ArrayRef(vec));
 }
 
@@ -932,6 +931,8 @@ template <>
 std::vector<mlir::Value>
 TosaMlirOperatorBuilder::build<Op_PAD>(TosaSerializationOperator *op) const {
   mlir::Value input_val = tensor_map->at(op->GetInputTensorNames()[0]);
+  mlir::RankedTensorType input_type =
+      tensor_type_map->at(op->GetInputTensorNames()[0]);
   mlir::RankedTensorType output_type =
       tensor_type_map->at(op->GetOutputTensorNames()[0]);
 
@@ -939,9 +940,11 @@ TosaMlirOperatorBuilder::build<Op_PAD>(TosaSerializationOperator *op) const {
          Attribute_PadAttribute); // double check attribute type
   TosaPadAttribute *attr = static_cast<TosaPadAttribute *>(op->GetAttribute());
 
-  auto padding_attr = BuildDenseI32ElementsAttr(op_builder, attr->padding());
-  auto padding_type = mlir::RankedTensorType::get(
-      {static_cast<int64_t>(attr->padding().size())}, op_builder->getI32Type());
+  // padding has shape {rank(input_type), 2}
+  auto padding_type = mlir::RankedTensorType::get({input_type.getRank(), 2},
+                                                  op_builder->getI32Type());
+  auto padding_attr =
+      BuildDenseI32ElementsAttr(op_builder, padding_type, attr->padding());
   mlir::Operation *mlir_const_op =
       op_builder->create<mlir::tosa::ConstOp>(loc, padding_type, padding_attr);
   block->push_back(mlir_const_op);
@@ -1024,7 +1027,7 @@ std::vector<mlir::Value> TosaMlirOperatorBuilder::build<Op_TRANSPOSE>(
   auto const_type = mlir::RankedTensorType::get(
       {static_cast<int64_t>(perms_values.size())}, op_builder->getI32Type());
   mlir::DenseElementsAttr const_attr =
-      BuildDenseI32ElementsAttr(op_builder, perms_values);
+      BuildDenseI32ElementsAttr(op_builder, const_type, perms_values);
   mlir::Operation *mlir_const_op =
       op_builder->create<mlir::tosa::ConstOp>(loc, const_type, const_attr);
   auto perms_val = mlir_const_op->getResult(0);
-- 
cgit v1.2.1