dequant_per_tensor (#6439)

Summary:

~

Reviewed By: hsharma35

Differential Revision: D64774100

Author: zonglinpeng

backends/cadence/hifi/operators/dequantize_per_tensor.cpp

@@ -15,33 +15,37 @@ namespace impl {
 namespace HiFi {
 namespace native {
 
-using executorch::aten::ScalarType;
-using executorch::aten::Tensor;
-using executorch::runtime::KernelRuntimeContext;
+using ::executorch::aten::ScalarType;
+using ::executorch::aten::Tensor;
+using ::executorch::runtime::KernelRuntimeContext;
+using ::cadence::impl::HiFi::kernels::dequantize;
 
 void dequantize_per_tensor_out(
-    KernelRuntimeContext& context,
+    KernelRuntimeContext& ctx,
     const Tensor& input,
     double scale,
     int64_t zero_point,
-    int64_t quant_min,
-    int64_t quant_max,
+    __ET_UNUSED int64_t quant_min,
+    __ET_UNUSED int64_t quant_max,
     ScalarType dtype,
     Tensor& out) {
   float* out_data = out.mutable_data_ptr<float>();
-  size_t numel = out.numel();
-
+  const size_t numel = out.numel();
   if (input.scalar_type() == ScalarType::Byte) {
     const uint8_t* input_data = input.const_data_ptr<uint8_t>();
-    impl::HiFi::kernels::dequantize<uint8_t>(
+    dequantize<uint8_t>(
         out_data, input_data, scale, zero_point, numel);
   } else if (input.scalar_type() == ScalarType::Char) {
     const int8_t* input_data = input.const_data_ptr<int8_t>();
     xa_nn_elm_dequantize_asym8s_f32(
         out_data, input_data, zero_point, scale, numel);
+  } else if (input.scalar_type() == ScalarType::Short) {
+    const int16_t* input_data = input.const_data_ptr<int16_t>();
+    dequantize<int16_t>(
+        out_data, input_data, scale, zero_point, numel);
   } else if (input.scalar_type() == ScalarType::Int) {
     const int32_t* input_data = input.const_data_ptr<int32_t>();
-    impl::HiFi::kernels::dequantize<int32_t>(
+    dequantize<int32_t>(
         out_data, input_data, scale, zero_point, numel);
   } else {
     ET_CHECK_MSG(false, "Unhandled input dtype %hhd", input.scalar_type());