CUTLASS 3.0.0 (#786)

* CUTLASS 3.0.0

include/cutlass/functional.h

@@ -56,6 +56,12 @@ struct absolute_value_op {
   }
 };
 
+template <>
+struct absolute_value_op<float> {
+  CUTLASS_HOST_DEVICE
+  float operator()(float lhs) const { return fabs(lhs); }
+};
+
 template <typename T>
 struct plus {
   CUTLASS_HOST_DEVICE
@@ -83,6 +89,30 @@ struct multiplies {
   }
 };
 
+// Maximum with nan propogation
+// To propgate the NANs, the "max" of a two element that contains NaNs should also return a NaN 
+template <typename T>
+struct maximum_with_nan_propogation {
+  CUTLASS_HOST_DEVICE
+  T operator()(T const &lhs, T const &rhs) const {
+    return lhs > rhs or std::isnan(lhs) ? lhs : rhs;
+  }
+};
+
+template <>
+struct maximum_with_nan_propogation<float> {
+  CUTLASS_HOST_DEVICE
+  float operator()(float const lhs, float const rhs) const {
+    float res;
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 800)
+    asm volatile("max.NaN.f32 %0, %1, %2;\n" : "=f"(res) : "f"(lhs), "f"(rhs));
+#else
+    res = lhs > rhs or std::isnan(lhs) ? lhs : rhs;
+#endif
+    return res;
+  }
+};
+
 /// Squares with optional conversion
 template <typename T, typename Output = T>
 struct square {