Allow int8 type in quantized_conv and im2row (#6049)

Summary:
Pull Request resolved: #6049

As titled.

Reviewed By: zonglinpeng

Differential Revision: D63842548

fbshipit-source-id: 5d535fb75f9ca3374b49126e6653082fa60b2ac1

Author: mcremon-meta

backends/cadence/reference/operators/quantized_conv_out.cpp

@@ -190,34 +190,65 @@ void quantized_conv_out(
   // per-channel
   bool per_tensor_quantized = bias_scale.numel() == 1;
 
-  conv2d_nchw_core_generic<uint8_t, uint8_t, int32_t, uint8_t, true>(
-      input.const_data_ptr<uint8_t>(),
-      weight.const_data_ptr<uint8_t>(),
-      bias.const_data_ptr<int32_t>(),
-      out.mutable_data_ptr<uint8_t>(),
-      n,
-      c,
-      h,
-      w,
-      oc,
-      wc,
-      wh,
-      ww,
-      oh,
-      ow,
-      stride[0],
-      stride[1],
-      padding[0],
-      padding[1],
-      dilation[0],
-      dilation[1],
-      groups,
-      in_zero_point,
-      weight_zero_point.const_data_ptr<int32_t>(),
-      bias_scale.const_data_ptr<float>(),
-      output_scale,
-      (uint8_t)output_zero_point,
-      per_tensor_quantized);
+  if (out.scalar_type() == exec_aten::ScalarType::Byte) {
+    conv2d_nchw_core_generic<uint8_t, uint8_t, int32_t, uint8_t, true>(
+        input.const_data_ptr<uint8_t>(),
+        weight.const_data_ptr<uint8_t>(),
+        bias.const_data_ptr<int32_t>(),
+        out.mutable_data_ptr<uint8_t>(),
+        n,
+        c,
+        h,
+        w,
+        oc,
+        wc,
+        wh,
+        ww,
+        oh,
+        ow,
+        stride[0],
+        stride[1],
+        padding[0],
+        padding[1],
+        dilation[0],
+        dilation[1],
+        groups,
+        in_zero_point,
+        weight_zero_point.const_data_ptr<int32_t>(),
+        bias_scale.const_data_ptr<float>(),
+        output_scale,
+        (uint8_t)output_zero_point,
+        per_tensor_quantized);
+  } else if (out.scalar_type() == exec_aten::ScalarType::Char) {
+    conv2d_nchw_core_generic<int8_t, int8_t, int32_t, int8_t, true>(
+        input.const_data_ptr<int8_t>(),
+        weight.const_data_ptr<int8_t>(),
+        bias.const_data_ptr<int32_t>(),
+        out.mutable_data_ptr<int8_t>(),
+        n,
+        c,
+        h,
+        w,
+        oc,
+        wc,
+        wh,
+        ww,
+        oh,
+        ow,
+        stride[0],
+        stride[1],
+        padding[0],
+        padding[1],
+        dilation[0],
+        dilation[1],
+        groups,
+        in_zero_point,
+        weight_zero_point.const_data_ptr<int32_t>(),
+        bias_scale.const_data_ptr<float>(),
+        output_scale,
+        (int8_t)output_zero_point,
+        per_tensor_quantized);
+  }
 }
 
 }; // namespace native