Support Half/BFloat16 in sum (pytorch#7897)

Partial fix for pytorch#7748.

Author: swolchok

kernels/portable/cpu/op_sum.cpp

@@ -43,10 +43,10 @@ Tensor& sum_dim_out(
 
   ET_KERNEL_CHECK(ctx, tensor_is_default_dim_order(in), InvalidArgument, out);
 
-  ET_SWITCH_REAL_TYPES_AND(
-      Bool, in.scalar_type(), ctx, "sum.IntList_out", CTYPE_IN, [&] {
-        ET_SWITCH_REAL_TYPES_AND(
-            Bool, out.scalar_type(), ctx, "sum.IntList_out", CTYPE_OUT, [&] {
+  ET_SWITCH_REALHBBF16_TYPES(
+      in.scalar_type(), ctx, "sum.IntList_out", CTYPE_IN, [&] {
+        ET_SWITCH_REALHBBF16_TYPES(
+            out.scalar_type(), ctx, "sum.IntList_out", CTYPE_OUT, [&] {
               CTYPE_OUT* out_data = out.mutable_data_ptr<CTYPE_OUT>();
               for (size_t out_ix = 0; out_ix < out.numel(); ++out_ix) {
                 CTYPE_OUT sum = 0;

kernels/test/op_sum_test.cpp

@@ -302,9 +302,9 @@ TEST_F(OpSumOutTest, AllRealInputRealOutputPasses) {
   test_sum_dim_out_dtype<ScalarType::INPUT_DTYPE, ScalarType::OUTPUT_DTYPE>();
 
 #define TEST_ENTRY(INPUT_CTYPE, INPUT_DTYPE) \
-  ET_FORALL_REAL_TYPES_WITH2(INPUT_CTYPE, INPUT_DTYPE, TEST_KERNEL);
+  ET_FORALL_REALHBF16_TYPES_WITH2(INPUT_CTYPE, INPUT_DTYPE, TEST_KERNEL);
 
-  ET_FORALL_REAL_TYPES(TEST_ENTRY);
+  ET_FORALL_REALHBF16_TYPES(TEST_ENTRY);
 #undef TEST_ENTRY
 #undef TEST_KERNEL
 }