Expand description
Original PTX specification:
§// Atomic operation with scalar type: atom{.sem}{.scope}{.space}.op{.level::cache_hint}.type d, [a], b{, cache-policy}; atom{.sem}{.scope}{.space}.op.type d, [a], b, c; atom{.sem}{.scope}{.space}.cas.b16 d, [a], b, c; atom{.sem}{.scope}{.space}.cas.b128 d, [a], b, c; atom{.sem}{.scope}{.space}.exch{.level::cache_hint}.b128 d, [a], b {, cache-policy}; atom{.sem}{.scope}{.space}.add.noftz{.level::cache_hint}.f16 d, [a], b{, cache-policy}; atom{.sem}{.scope}{.space}.add.noftz{.level::cache_hint}.f16x2 d, [a], b{, cache-policy}; atom{.sem}{.scope}{.space}.add.noftz{.level::cache_hint}.bf16 d, [a], b{, cache-policy}; atom{.sem}{.scope}{.space}.add.noftz{.level::cache_hint}.bf16x2 d, [a], b{, cache-policy}; .space = { .global, .shared, .shared::cta, .shared::cluster}; .sem = { .relaxed, .acquire, .release, .acq_rel }; .scope = { .cta, .cluster, .gpu, .sys }; .op = { .and, .or, .xor, .cas, .exch, .add, .inc, .dec, .min, .max }; .level::cache_hint = { .L2::cache_hint }; .type = { .b32, .b64, .u32, .u64, .s32, .s64, .f32, .f64 };
// Atomic operation with vector type: atom{.sem}{.scope}{.global}.add{.level::cache_hint}.vec_32_bit.f32 d, [a], b{, cache-policy}; atom{.sem}{.scope}{.global}.op.noftz{.level::cache_hint}.vec_16_bit.half_word_type d, [a], b{, cache-policy}; atom{.sem}{.scope}{.global}.op.noftz{.level::cache_hint}.vec_32_bit.packed_type d, [a], b{, cache-policy}; .sem = { .relaxed, .acquire, .release, .acq_rel }; .scope = { .cta, .cluster, .gpu, .sys }; .op = { .add, .min, .max }; .half_word_type = { .f16, .bf16 }; .packed_type = { .f16x2, .bf16x2 }; .vec_16_bit = { .v2, .v4, .v8 }; .vec_32_bit = { .v2, .v4 }; .level::cache_hint = { .L2::cache_hint };