P0020r1 : Floating Point Atomic

This paper proposes an extension to the atomic operations library [atomics] for atomic addition on an object conforming to the atomic-view-concept (see P0019, Atomic View) instantiated for a floating point type. A class conforming to the atomic-view-concept shall also provide the following operations when T is a floating point (FP) type. This capability is critical for high performance computing (HPC) applications.

template<> struct atomic< FP > {

FP fetch_add( FP operand , memory_order order = memory_order_seq_cst ) volatile noexcept;

FP fetch_add( FP operand , memory_order order = memory_order_seq_cst ) noexcept;

FP fetch_sub( FP operand , memory_order order = memory_order_seq_cst ) volatile noexcept;

FP fetch_sub( FP operand , memory_order order = memory_order_seq_cst ) noexcept;

FP operator+=( FP operand ) volatile noexcept;

FP operator+=( FP operand ) noexcept;

FP operator-=( FP operand ) volatile noexcept;

FP operator-=( FP operand ) noexcept;

};

template<> struct atomic-view-concept < FP > {

FP fetch_add( FP operand , memory_order order = memory_order_seq_cst ) const noexcept;

FP fetch_sub( FP operand , memory_order order = memory_order_seq_cst ) const noexcept;

FP operator+=( FP operand ) const noexcept;

FP operator-=( FP operand ) const noexcept;

};

[Note: The additive operator overloads are consistent with the corresponding overloads specified on std::atomic<integral> [29.5, Atomic types]. However, these arithmatic operator overloads are inconsistent with the corresponding operators for corresponding arithmatic type.

double d(0)
( d += 1.0 ) += 2.0 ; // d+= returns an lvalue

atomic_view<double> ad(d);
( ad += 1.0 ) += 2.0 ; // error: ad+= returns an rvalue

- end note]

Requires: The memory order shall not be memory_order_acquire, memory_order_acq_rel, or memory_order_consume.

Effects: Atomically replaces the value of the floating point object referenced by the atomic view with the sum of the previous value and operand for add or -operand for sub. Memory is affected according to the value of order. This is an atomic read-modify-write operation (1.10).

Returns: fetch_op Atomically, the value of the referenced object immediately before the effects.

Remark: In the event of arithmetic overflow or other floating point exception, the behavior of this operation is implementation defined. It is a quality-of-implementation as to whether arithmetic overflow or other floating point exception is detectable in the resulting value; e.g., a value of NaN.

** FP operator+=(** FP operand )

Effects: Returns fetch_add(operand) + operand .

** FP operator-=(** FP operand )

Effects: Returns fetch_sub(operand) - operand .