Merge branch 'release/0.5' into my_fork

Author: SS-JIA

backends/vulkan/test/test_vulkan_delegate.py

@@ -117,6 +117,7 @@ def run_test():
 
             edge_program = to_edge_transform_and_lower(
                 program,
+                compile_config=self._edge_compile_config,
                 transform_passes=[
                     I64toI32(self._edge_compile_config._skip_dim_order),
                 ],

exir/capture/_config.py

@@ -45,8 +45,7 @@ class EdgeCompileConfig:
     )
     _skip_type_promotion: bool = False
     # TODO(gasoonjia): remove this
-    # TODO(T192537614): reenanle dim order as default
-    _skip_dim_order: bool = True
+    _skip_dim_order: bool = False
 
 
 @compatibility(is_backward_compatible=False)

extension/export_util/utils.py

@@ -60,7 +60,6 @@ def _core_aten_to_edge(
     if not edge_compile_config:
         edge_compile_config = exir.EdgeCompileConfig(
             _check_ir_validity=False,  # quant ops currently break ir verification
-            _skip_dim_order=True,  # TODO(T182928844): dim order ops can not delegate to backend
         )
     edge_manager: EdgeProgramManager = to_edge(
         core_aten_exir_ep,

kernels/portable/cpu/op__to_dim_order_copy.cpp

@@ -96,13 +96,17 @@ Tensor& _to_dim_order_copy_out(
       InvalidArgument,
       out);
 
-  ET_SWITCH_REALHB_TYPES(
+  if (self.numel() == 0) {
+    return out;
+  }
+
+  ET_SWITCH_REALHBBF16_TYPES(
       self.scalar_type(),
       ctx,
       "dim_order_ops::_to_dim_order_copy.out",
       CTYPE_IN,
       [&] {
-        ET_SWITCH_REALHB_TYPES(
+        ET_SWITCH_REALHBBF16_TYPES(
             out.scalar_type(),
             ctx,
             "dim_order_ops::_to_dim_order_copy.out",

kernels/test/op__to_dim_order_copy_test.cpp

@@ -36,7 +36,9 @@ typedef std::map<
           std::type_index,
           std::variant<
             std::vector<float>,
-            std::vector<double>>>
+            std::vector<double>,
+            std::vector<exec_aten::Half>,
+            std::vector<exec_aten::BFloat16>>>
         FloatingTypeToDataMap;
 
 typedef std::map<
@@ -381,9 +383,9 @@ TEST_F(OpToDimOrderCopyTest, NanInfSupported) {
       ScalarType::OUTPUT_DTYPE>(test_cases);
 
 #define TEST_ENTRY(INPUT_CTYPE, INPUT_DTYPE) \
-  ET_FORALL_FLOAT_TYPES_WITH2(INPUT_CTYPE, INPUT_DTYPE, TEST_KERNEL);
+  ET_FORALL_FLOATHBF16_TYPES_WITH2(INPUT_CTYPE, INPUT_DTYPE, TEST_KERNEL);
 
-  ET_FORALL_FLOAT_TYPES(TEST_ENTRY);
+  ET_FORALL_FLOATHBF16_TYPES(TEST_ENTRY);
 
 #undef TEST_ENTRY
 #undef TEST_KERNEL
@@ -413,6 +415,13 @@ TEST_F(OpToDimOrderCopyTest, HardcodeFloatConvertInt) {
       -0.30919688936285893988};
   // clang-format on
 
+  std::vector<exec_aten::Half> half_data;
+  std::vector<exec_aten::BFloat16> bf16_data;
+  for (auto d : double_data) {
+    half_data.emplace_back(d);
+    bf16_data.emplace_back(d);
+  }
+
   std::vector<int64_t> int64_data = {
       -1, -4, 2, -2, 3, 3, -3, -4, 3, 3, 0, 2, 0, -1, 0};
   std::vector<int32_t> int32_data = {
@@ -426,6 +435,8 @@ TEST_F(OpToDimOrderCopyTest, HardcodeFloatConvertInt) {
   FloatingTypeToDataMap floating_point_data;
   floating_point_data[typeid(float)] = float_data;
   floating_point_data[typeid(double)] = double_data;
+  floating_point_data[typeid(exec_aten::Half)] = half_data;
+  floating_point_data[typeid(exec_aten::BFloat16)] = bf16_data;
 
   // Gathering all int data together for better traversial
   IntTypeToDataMap int_data;
@@ -444,7 +455,7 @@ TEST_F(OpToDimOrderCopyTest, HardcodeFloatConvertInt) {
 #define TEST_ENTRY(INPUT_CTYPE, INPUT_DTYPE) \
   ET_FORALL_INT_TYPES_WITH2(INPUT_CTYPE, INPUT_DTYPE, TEST_KERNEL);
 
-  ET_FORALL_FLOAT_TYPES(TEST_ENTRY);
+  ET_FORALL_FLOATHBF16_TYPES(TEST_ENTRY);
 }
 
 TEST_F(OpToDimOrderCopyTest, MismatchedSizesDie) {