Fix scalar (single element tensor) binary ops on HiFi

Differential Revision: D71667998

Pull Request resolved: #9523

Author: mcremon-meta

backends/cadence/hifi/operators/op_add.cpp

@@ -142,13 +142,13 @@ Tensor& add_out(
       (a_type == ScalarType::Float) && (b_type == ScalarType::Float);
 
   if ((a_dim == 0) && float_types) {
-    for (int i = 0; i < max_dim; i++)
+    for (int i = 0; i < b.numel(); i++)
       out.mutable_data_ptr<float>()[i] =
           a.const_data_ptr<float>()[0] + b.const_data_ptr<float>()[i];
     return out;
   }
   if ((b_dim == 0) && float_types) {
-    for (int i = 0; i < max_dim; i++)
+    for (int i = 0; i < a.numel(); i++)
       out.mutable_data_ptr<float>()[i] =
           a.const_data_ptr<float>()[i] + b.const_data_ptr<float>()[0];
     return out;

backends/cadence/hifi/operators/op_div.cpp

@@ -90,13 +90,13 @@ div_out(RuntimeContext& ctx, const Tensor& a, const Tensor& b, Tensor& out) {
       (a_type == ScalarType::Float) && (b_type == ScalarType::Float);
 
   if ((a_dim == 0) && float_types) {
-    for (int i = 0; i < max_dim; i++)
+    for (int i = 0; i < b.numel(); i++)
       out.mutable_data_ptr<float>()[i] =
           a.const_data_ptr<float>()[0] / b.const_data_ptr<float>()[i];
     return out;
   }
   if ((b_dim == 0) && float_types) {
-    for (int i = 0; i < max_dim; i++)
+    for (int i = 0; i < a.numel(); i++)
       out.mutable_data_ptr<float>()[i] =
           a.const_data_ptr<float>()[i] / b.const_data_ptr<float>()[0];
     return out;
@@ -218,13 +218,13 @@ Tensor& div_out_mode(
       (a_type == ScalarType::Float) && (b_type == ScalarType::Float);
 
   if ((a_dim == 0) && float_types) {
-    for (int i = 0; i < max_dim; i++)
+    for (int i = 0; i < b.numel(); i++)
       out.mutable_data_ptr<float>()[i] =
           a.const_data_ptr<float>()[0] / b.const_data_ptr<float>()[i];
     return out;
   }
   if ((b_dim == 0) && float_types) {
-    for (int i = 0; i < max_dim; i++)
+    for (int i = 0; i < a.numel(); i++)
       out.mutable_data_ptr<float>()[i] =
           a.const_data_ptr<float>()[i] / b.const_data_ptr<float>()[0];
     return out;

backends/cadence/hifi/operators/op_mul.cpp

@@ -108,13 +108,13 @@ mul_out(RuntimeContext& ctx, const Tensor& a, const Tensor& b, Tensor& out) {
       (a_type == ScalarType::Float) && (b_type == ScalarType::Float);
 
   if ((a_dim == 0) && float_types) {
-    for (int i = 0; i < max_dim; i++)
+    for (int i = 0; i < b.numel(); i++)
       out.mutable_data_ptr<float>()[i] =
           a.const_data_ptr<float>()[0] * b.const_data_ptr<float>()[i];
     return out;
   }
   if ((b_dim == 0) && float_types) {
-    for (int i = 0; i < max_dim; i++)
+    for (int i = 0; i < a.numel(); i++)
       out.mutable_data_ptr<float>()[i] =
           a.const_data_ptr<float>()[i] * b.const_data_ptr<float>()[0];
     return out;

backends/cadence/hifi/operators/op_sub.cpp

@@ -137,13 +137,13 @@ Tensor& sub_out(
       (a_type == ScalarType::Float) && (b_type == ScalarType::Float);
 
   if ((a_dim == 0) && float_types) {
-    for (int i = 0; i < max_dim; i++)
+    for (int i = 0; i < b.numel(); i++)
       out.mutable_data_ptr<float>()[i] =
           a.const_data_ptr<float>()[0] - b.const_data_ptr<float>()[i];
     return out;
   }
   if ((b_dim == 0) && float_types) {
-    for (int i = 0; i < max_dim; i++)
+    for (int i = 0; i < a.numel(); i++)
       out.mutable_data_ptr<float>()[i] =
           a.const_data_ptr<float>()[i] - b.const_data_ptr<float>()[0];
     return out;