Fix nightly operator test failures across nxp_rt600 and DLA_V130 backends

Differential Revision: D100873709

Pull Request resolved: #18951

Author: ethansfng

backends/cadence/hifi/operators/op_permute_copy.cpp

@@ -73,8 +73,7 @@ Tensor& permute_copy_out(
 
   bool optimized = false;
 
-  if (out.scalar_type() == ScalarType::Float ||
-      out.scalar_type() == ScalarType::Char ||
+  if (out.scalar_type() == ScalarType::Char ||
       out.scalar_type() == ScalarType::Byte)
     optimized = true;
 
@@ -101,22 +100,7 @@ Tensor& permute_copy_out(
       p_permute_vec[i] = dims[i];
     }
 
-    if (in_type == ScalarType::Float) {
-      WORD32* p_inp = (WORD32*)in.const_data_ptr<float>();
-      WORD32* p_out = (WORD32*)out.mutable_data_ptr<float>();
-
-      WORD32 ret_val = xa_nn_transpose_32_32(
-          p_out,
-          p_out_shape,
-          p_inp,
-          p_inp_shape,
-          p_permute_vec,
-          num_out_dims,
-          num_inp_dims);
-
-      ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
-
-    } else if (in_type == ScalarType::Char) {
+    if (in_type == ScalarType::Char) {
       WORD8* p_inp = (WORD8*)in.const_data_ptr<char>();
       WORD8* p_out = (WORD8*)out.mutable_data_ptr<char>();
 

backends/cadence/hifi/operators/op_softmax.cpp

@@ -68,6 +68,9 @@ Tensor& _softmax_out(
   if (in.dim() > kNnlibMaxDim)
     optimized = false;
 
+  if (dim < in.dim() - 1)
+    optimized = false;
+
   if (optimized) {
     int* p_inp = (int*)in.const_data_ptr<float>();
     int* out_data = (int*)out.mutable_data_ptr<float>();

backends/cadence/hifi/operators/op_where.cpp

@@ -81,6 +81,9 @@ Tensor& where_self_out(
   if ((broadcast == 1) && (max_dim > kNnlibMaxDim))
     optimized = 0;
 
+  if (cond_is_broadcasted)
+    optimized = 0;
+
   if (optimized) {
     const float* a_data = a.const_data_ptr<float>();
     const float* b_data = b.const_data_ptr<float>();

backends/cadence/utils/facto_util.py

@@ -249,7 +249,7 @@ def random_size_constraint(deps: object, r: int, d: int) -> int:
         case "permute_copy.default":
             tensor_constraints.extend(
                 [
-                    cp.Dtype.In(lambda deps: [torch.float32, torch.int8, torch.uint8]),
+                    cp.Dtype.In(lambda deps: [torch.float32, torch.int32]),
                     cp.Rank.Le(
                         lambda deps: 5
                     ),  # xa_nn_transpose only supports up to 5D
@@ -391,12 +391,13 @@ def random_size_constraint(deps: object, r: int, d: int) -> int:
             tensor_constraints.extend(
                 [
                     cp.Dtype.In(lambda deps: [torch.float32, torch.int32]),
+                    cp.Value.Ge(lambda deps, dtype, struct: 0),
                 ]
             )
         case "div.Tensor_mode" | "minimum.default":
             if index == 0:
                 tensor_constraints = [
-                    cp.Dtype.In(lambda deps: [torch.int64, torch.int32, torch.float32]),
+                    cp.Dtype.In(lambda deps: [torch.int32, torch.float32]),
                     cp.Value.Ge(lambda deps, dtype, struct: -(2**4)),
                     cp.Value.Le(lambda deps, dtype, struct: 2**4),
                     cp.Rank.Ge(lambda deps: 1),
@@ -405,7 +406,7 @@ def random_size_constraint(deps: object, r: int, d: int) -> int:
                 ]
             else:
                 tensor_constraints = [
-                    cp.Dtype.In(lambda deps: [torch.int64, torch.int32, torch.float32]),
+                    cp.Dtype.In(lambda deps: [torch.int32, torch.float32]),
                     cp.Value.Ge(lambda deps, dtype, struct: -(2**4)),
                     cp.Value.Le(lambda deps, dtype, struct: 2**4),
                     cp.Value.Ne(