From 931c8b01479197ba2aa293a609d0e91a221e47bb Mon Sep 17 00:00:00 2001
From: Hariharan Seshadri <shariharan91@gmail.com>
Date: Thu, 1 Sep 2022 18:30:57 -0700
Subject: [PATCH] Resolve GH issue 12706 (#12815)

---
 .../transpose_optimizer/transpose_optimizer.cc       | 12 ++++++++++++
 .../test/optimizer/transpose_optimizer_test.cc       | 12 ++++++++++++
 2 files changed, 24 insertions(+)

diff --git a/onnxruntime/core/optimizer/transpose_optimizer/transpose_optimizer.cc b/onnxruntime/core/optimizer/transpose_optimizer/transpose_optimizer.cc
index 81adcee7cb..2cac02033e 100644
--- a/onnxruntime/core/optimizer/transpose_optimizer/transpose_optimizer.cc
+++ b/onnxruntime/core/optimizer/transpose_optimizer/transpose_optimizer.cc
@@ -1691,7 +1691,19 @@ static const std::unordered_map<std::string_view, const HandlerInfo&> handler_ma
     {"Split", split_handler},
     {"Shape", shape_handler},
     {"Pad", pad_handler},
+// The CUDA Resize kernel assumes that the input is NCHW and
+// Resize can't be supported in ORT builds with CUDA enabled.
+// TODO: Enable this once the CUDA Resize kernel is implemented
+// "generically" (i.e.) aligning with the generic nature of the
+// ONNX spec.
+// See https://github.com/microsoft/onnxruntime/pull/10824 for
+// a similar fix applied to the CPU Resize kernel.
+// Per tests included in #10824, the ROCM EP also generates
+// incorrect results when this handler is used, so the Resize 
+// handler is not enabled even for those builds.
+#if !defined(USE_CUDA) && !defined(USE_ROCM)
     {"Resize", resize_handler},
+#endif
     {"ReduceSum", reduce_sum_handler},
 
     {"ReduceLogSum", reduce_op_handler},
diff --git a/onnxruntime/test/optimizer/transpose_optimizer_test.cc b/onnxruntime/test/optimizer/transpose_optimizer_test.cc
index cfeb9a2202..980ac01b9d 100644
--- a/onnxruntime/test/optimizer/transpose_optimizer_test.cc
+++ b/onnxruntime/test/optimizer/transpose_optimizer_test.cc
@@ -294,6 +294,17 @@ TEST(TransposeOptimizerTests, TestPadNonconst) {
                     /*opset_version*/ 11);
 }
 
+// The CUDA Resize kernel assumes that the input is NCHW and
+// Resize can't be supported in ORT builds with CUDA enabled.
+// TODO: Enable this once the CUDA Resize kernel is implemented
+// "generically" (i.e.) aligning with the generic nature of the
+// ONNX spec.
+// See https://github.com/microsoft/onnxruntime/pull/10824 for
+// a similar fix applied to the CPU Resize kernel.
+// Per tests included in #10824, the ROCM EP also generates
+// incorrect results when this handler is used, so the Resize
+// handler is not enabled even for those builds.
+#if !defined(USE_CUDA) && !defined(USE_ROCM)
 TEST(TransposeOptimizerTests, TestResize) {
   auto build_test_case_1 = [&](ModelTestBuilder& builder) {
     auto* input0_arg = MakeInput<float>(builder, {{4, -1, 2, -1}}, {4, 6, 2, 10}, 0.0, 1.0);
@@ -498,6 +509,7 @@ TEST(TransposeOptimizerTests, TestResizeNonconstOpset13) {
                     /*opset_version*/ 13);
 }
 
+#endif
 TEST(TransposeOptimizerTests, TestAdd) {
   auto build_test_case_1 = [&](ModelTestBuilder& builder) {
     auto* input0_arg = builder.MakeInput<float>({4, 6, 10}, 0.0, 1.0);