[SYCL] Fix bfloat16::to_float() host implementation. (#6675)

* [SYCL] Fix bfloat16::to_float() host implementation.

Signed-off-by: Konstantin S Bobrovsky <[email protected]>

Author: kbobrovs

sycl/include/sycl/ext/oneapi/experimental/bfloat16.hpp

@@ -63,10 +63,9 @@ class bfloat16 {
     return __spirv_ConvertBF16ToFINTEL(a);
 #endif
 #else
-    // Shift temporary variable to silence the warning
     uint32_t bits = a;
     bits <<= 16;
-    return static_cast<float>(bits);
+    return sycl::bit_cast<float>(bits);
 #endif
   }
 

sycl/test/extensions/bfloat16_host.cpp

@@ -73,15 +73,17 @@ int main() {
   Success &= check_bf16_from_float(std::numeric_limits<float>::quiet_NaN(),
                                    std::stoi("1111111111000001", nullptr, 2));
 
+  // see https://float.exposed/b0xffff
   Success &= check_bf16_to_float(
       0, bitsToFloatConv(std::string("00000000000000000000000000000000")));
   Success &= check_bf16_to_float(
-      1, bitsToFloatConv(std::string("01000111100000000000000000000000")));
+      1, bitsToFloatConv(std::string("00000000000000010000000000000000")));
   Success &= check_bf16_to_float(
-      42, bitsToFloatConv(std::string("01001010001010000000000000000000")));
+      42, bitsToFloatConv(std::string("00000000001010100000000000000000")));
   Success &= check_bf16_to_float(
-      std::numeric_limits<uint16_t>::max(),
-      bitsToFloatConv(std::string("01001111011111111111111100000000")));
+      // std::numeric_limits<uint16_t>::max() - 0xffff is bfloat16 -Nan and
+      // -Nan == -Nan check in check_bf16_to_float would fail, so use not Nan:
+      65407, bitsToFloatConv(std::string("11111111011111110000000000000000")));
   if (!Success)
     return -1;
   return 0;