P0020r00 : Floating Point Atomic View

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. We leave open the question of whether the existing class atomic<T> should be similarly extended when T is a floating point type.

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;

void add( fp operand , memory_order order = memory_order_seq_cst ) volatile noexcept;

void add( fp operand , memory_order order = memory_order_seq_cst ) noexcept;

void sub( fp operand , memory_order order = memory_order_seq_cst ) volatile noexcept;

void sub( fp operand , memory_order order = memory_order_seq_cst ) noexcept;

void operator+=( FP operand ) volatile noexcept;

void operator+=( FP operand ) noexcept;

void operator-=( FP operand ) volatile noexcept;

void 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;

void add( fp operand , memory_order order = memory_order_seq_cst ) const noexcept;

void sub( fp operand , memory_order order = memory_order_seq_cst ) const noexcept;

void operator+=( FP operand ) const noexcept;

void operator-=( FP operand ) const noexcept;

};

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: The add and sub functions omit returning the value before the effects. Omission of a return value may allow a better performing implementation.

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.

Effects: += performs add(operand) and -= performs sub(operand)

Remark: These operators omit a return value to avoid conflicting with either the corresponding arithmetic operators for FP or the arithmetic operators for atomic<integral>. Note that the arithmetic operators for atomic<integral> are inconsistent with the corresponding arithmetic operators for integral.

int i(0);
( i += 1 ) += 2 ; // i+= returns an lvalue

atomic<int> ai(0);
( ai += 1 ) += 2 ; // error: i+= returns an rvalue