Use TmpTensor for MatMul op.

Summary: This diff introduces the use of temporary tensors to reduce memory usage in the width packed versions of the matmul op shader.

Differential Revision: D68924743

Author: trivedivivek

backends/vulkan/runtime/graph/ops/impl/MatMul.cpp

@@ -166,7 +166,8 @@ void add_matmul_optimized_node(
       /*passthrough = */ true);
 
   // Ensure mat1 is width packed
-  ValueRef mat1_W_packed = graph.add_tensor_like(mat1, utils::kWidthPacked);
+  TmpTensor mat1_tmp(&graph, graph.sizes_of(mat1), graph.dtype_of(mat1), utils::kWidthPacked);
+  ValueRef mat1_W_packed = mat1_tmp;
   auto viewFn = VK_GET_OP_FN("aten.view_copy.default");
   viewFn(graph, {mat1, graph.add_none(), mat1_W_packed});
 
@@ -176,8 +177,9 @@ void add_matmul_optimized_node(
   ValueRef mat2_packed = mat2;
   const utils::GPUMemoryLayout mat2_layout =
       mat2_is_transposed_val ? utils::kWidthPacked : utils::kHeightPacked;
+  TmpTensor mat2_tmp(&graph, graph.sizes_of(mat2), graph.dtype_of(mat2), mat2_layout);
   if (graph.estimate_memory_layout_of(mat2) != mat2_layout) {
-    mat2_packed = graph.add_tensor_like(mat2, mat2_layout);
+    mat2_packed = mat2_tmp;
     viewFn(graph, {mat2, graph.add_none(), mat2_packed});
   }