Cast long inputs to float32 before feeding to xnnpack graph runtime (#553)

Summary:
Pull Request resolved: #553

In DeepLab v3, we saw that some inputs are int64, (this is largely from interpolation decomposition). XNNPACK can not handle int64, which means that the int64 inputs would wrongly be interpreted as float32.

In order to correctly feed the right inputs to XNNPACK, we must cast the int64 data to float32. We do something similar when casting int32 data to int64 when we are returning indicies from argmax pooling.

Reviewed By: digantdesai

Differential Revision: D49797819

fbshipit-source-id: 1f4f6ab38c4c8e36ba0a7017300f5e3a3ccbf810

Author: mcr229

backends/xnnpack/runtime/XNNExecutor.cpp

@@ -68,6 +68,17 @@ Error XNNExecutor::set_external_input(uint32_t id, Tensor* input) {
     return Error::NotSupported;
 #endif
   } else {
+    // TODO(T165403530): Test insure accuracy for int64 --> float32 conversion
+    if (input->scalar_type() == ScalarType::Long) {
+      // Input data type is int64. However, XNNPACK doesn't support
+      // int64. This means that the data needs to be casted to float
+      // In order for XNNPACK to properly use it.
+      const int64_t* data_64 = input->const_data_ptr<int64_t>();
+      float* data_f32 = input->mutable_data_ptr<float>();
+      for (int j = 0; j < input->numel(); j++) {
+        data_f32[j] = data_64[j];
+      }
+    }
     externals_.emplace_back(xnn_external_value{id, input->mutable_data_ptr()});
   }
   return Error::Ok;