Dont error check unused tensors in convolution_backward.out

Summary:
In ATen if this tensor is unused for the functional op it returns an empty tensor.

In Executorch we essentially just serialize a place holder to tensor to be passed along, but since the tensor should be unused we dont bother trying to give it robust meta data. In this diff I'm skipping the meta data checks on the tensors if the gradients are unused.

Reviewed By: manuelcandales

Differential Revision: D69211399

Author: JacobSzwejbka

kernels/portable/cpu/op_convolution_backward.cpp

@@ -34,7 +34,7 @@ bool check_convolution_backward_args(
     bool transposed,
     IntArrayRef output_padding,
     int64_t groups,
-    ET_UNUSED executorch::aten::ArrayRef<bool> output_mask,
+    executorch::aten::ArrayRef<bool> output_mask,
     Tensor& grad_input,
     Tensor& grad_weight,
     Tensor& grad_bias) {
@@ -45,9 +45,18 @@ bool check_convolution_backward_args(
 
   ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(weight, input));
   ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(grad_output, input));
-  ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(grad_input, input));
-  ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(grad_weight, input));
-  ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(grad_bias, input));
+
+  if (output_mask[0]) {
+    ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(grad_input, input));
+  }
+
+  if (output_mask[1]) {
+    ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(grad_weight, input));
+  }
+
+  if (output_mask[2]) {
+    ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(grad_bias, input));
+  }
 
   ET_LOG_MSG_AND_RETURN_IF_FALSE(
       check_convolution_args(
@@ -267,19 +276,23 @@ std::tuple<Tensor&, Tensor&, Tensor&> convolution_backward_out(
       InvalidArgument,
       ret_val);
 
-  ET_KERNEL_CHECK(
+  if (output_mask[0]) {
+    ET_KERNEL_CHECK(
       ctx,
       resize_tensor(grad_input, input.sizes()) == Error::Ok,
       InvalidArgument,
       ret_val);
+  }
 
-  ET_KERNEL_CHECK(
-      ctx,
-      resize_tensor(grad_weight, weight.sizes()) == Error::Ok,
-      InvalidArgument,
-      ret_val);
+  if (output_mask[1]) {
+    ET_KERNEL_CHECK(
+        ctx,
+        resize_tensor(grad_weight, weight.sizes()) == Error::Ok,
+        InvalidArgument,
+        ret_val);
+  }
 
-  if (bias_sizes_opt.has_value()) {
+  if (bias_sizes_opt.has_value() && output_mask[2]) {
     ET_KERNEL_CHECK(
         ctx,
         resize_tensor(grad_bias, bias_sizes_opt.value()) == Error::Ok,