[PATCH] Fix undefined __floatdihf in libtvmruntime.so on aarch64. (#4119)

Arm architecture provides optional FP16 floating point support in two alternative formats, IEEE and an an alternative Arm format.

The ACLE (Arm C Language Extension) defined preprocessor symbol __ARM_FP16_FORMAT_IEEE can be used to distinguish between implementations providing IEEE and the Arm alternative format, but cannot, on its own, be used to determined if FP16 HW support is actually present.

Testing this preprocessor symbol can lead to undefined __floatdihf at runtime on an aarch64 target where no FP16 HW is present.

The relevant preprocessor symbol to determine whether FP16 HW support is present in the target is __ARM_FEATURE_FP16_SCALAR_ARITHMETIC, this symbol implies  __ARM_FP16_FORMAT_IEEE.

The relevant preprocessor symbols are defined by the ACLE standard, section 5.5.21 16-bit floating-point data processing operations, https://static.docs.arm.com/101028/0008/Q2-ACLE_2019Q2_release-0008.pdf

diff --git a/src/contrib/sort/sort.cc b/src/contrib/sort/sort.cc
index a87ce07..0ccaee5 100644 (file)
--- a/src/contrib/sort/sort.cc
+++ b/src/contrib/sort/sort.cc
@@ -75,7 +75,7 @@ TVM_REGISTER_GLOBAL("tvm.contrib.sort.argsort_nms")
   // Currently only supports input dtype to be float32.
   CHECK_EQ(dtype.code, 2) << "Currently only supports input dtype "
       "to be float.";
-#if (__ARM_FP16_FORMAT_IEEE != 1)
+#if (__ARM_FEATURE_FP16_SCALAR_ARITHMETIC != 1)
   CHECK_EQ(dtype.bits, 32) << "Currently only supports input dtype "
       "to be float32.";
 #endif
@@ -100,23 +100,23 @@ TVM_REGISTER_GLOBAL("tvm.contrib.sort.argsort_nms")
         sorter.emplace_back(std::make_pair(k, *(data_ptr + full_idx)));
       }
       if (is_ascend) {
-#if (__ARM_FP16_FORMAT_IEEE == 1)
+#if (__ARM_FEATURE_FP16_SCALAR_ARITHMETIC == 1)
         if (dtype.bits == 16) {
           std::stable_sort(sorter.begin(), sorter.end(), CompareAscend<__fp16>);
         } else {
 #endif
         std::stable_sort(sorter.begin(), sorter.end(), CompareAscend<float>);
-#if (__ARM_FP16_FORMAT_IEEE == 1)
+#if (__ARM_FEATURE_FP16_SCALAR_ARITHMETIC == 1)
         }
 #endif
       } else {
-#if (__ARM_FP16_FORMAT_IEEE == 1)
+#if (__ARM_FEATURE_FP16_SCALAR_ARITHMETIC == 1)
         if (dtype.bits == 16) {
           std::stable_sort(sorter.begin(), sorter.end(), CompareDescend<__fp16>);
         } else {
 #endif
         std::stable_sort(sorter.begin(), sorter.end(), CompareDescend<float>);
-#if (__ARM_FP16_FORMAT_IEEE == 1)
+#if (__ARM_FEATURE_FP16_SCALAR_ARITHMETIC == 1)
         }
 #endif
       }
@@ -210,7 +210,7 @@ TVM_REGISTER_GLOBAL("tvm.contrib.sort.argsort")
     } else {
       LOG(FATAL) << "Unsupported output dtype: " << out_dtype;
     }
-#if (__ARM_FP16_FORMAT_IEEE == 1)
+#if (__ARM_FEATURE_FP16_SCALAR_ARITHMETIC == 1)
   } else if (data_dtype == "float16") {
     if (out_dtype == "float16") {
       argsort<__fp16, __fp16>(input, output, axis, is_ascend);