triton 总览

triton op 支持度总览

	Triton Op	int8	int16	int32	uint32	int64	fp16	fp32	bf16	bool
Creation Ops	arange	×	×	✓	×	×	×	×	×	×
	cat	✓	✓	✓	×	✓	✓	✓	✓	✓
	full	✓	✓	✓	×	✓	✓	✓	✓	✓
	zeros	✓	✓	✓	×	✓	✓	✓	✓	✓
	zeros_like	✓	✓	✓	×	✓	✓	✓	✓	✓
	cast	✓	✓	✓	×	✓	✓	✓	✓	✓
Shape Manipulation Ops	broadcast	✓	✓	✓	×	✓	✓	✓	✓	✓
	broadcast_to	✓	✓	✓	×	✓	✓	✓	✓	✓
	expand_dims	✓	✓	✓	×	✓	✓	✓	✓	✓
	interleave	✓	✓	✓	×	✓	✓	✓	✓	✓
	join	✓	✓	✓	×	✓	✓	✓	✓	✓
	permute	✓	✓	✓	×	×	✓	✓	✓	✓
	ravel	✓	✓	✓	×	✓	✓	✓	✓	✓
	reshape	✓	✓	✓	×	✓	✓	✓	✓	✓
	split	✓	✓	✓	×	✓	✓	✓	✓	✓
	trans	✓	✓	✓	×	×	✓	✓	✓	✓
	view	✓	✓	✓	×	✓	✓	✓	✓	✓
Linear Algebra Ops	dot	✓	×	×	×	×	✓	✓	✓	×
	dot_scaled	×	×	×	×	×	×	×	×	×
Memory/Pointer Ops	load	✓	✓	✓	×	✓	✓	✓	✓	✓
	store	✓	✓	✓	×	✓	✓	✓	✓	✓
	make_block_ptr	✓	✓	✓	×	✓	✓	✓	✓	×
	make_tensor_descriptor	✓	✓	✓	×	✓	✓	✓	✓	×
	load_tensor_descriptor	✓	✓	✓	×	✓	✓	✓	✓	×
	store_tensor_descriptor	✓	✓	✓	×	✓	✓	✓	✓	×
	advance	✓	✓	✓	×	✓	✓	✓	✓	×
Indexing Ops	flip	✓	✓	✓	×	✓	✓	✓	✓	✓
	where	✓	✓	✓	×	×	✓	✓	✓	✓*
	swizzle2d	✓	✓	✓	×	✓	×	×	×	×
Math Ops	add	✓	✓	✓	×	✓	✓	✓	✓	✓*
	sub	✓	✓	✓	×	✓	✓	✓	✓	✓*
	mul	✓	✓	✓	×	✓	✓	✓	✓	✓*
	div	✓	✓	✓	×	✓	✓	✓	✓	✓*
	floordiv(//)	✓	✓	✓	×	✓	×	×	×	✓*
	mod	✓	✓	✓	×	✓	✓	✓	✓	✓*
	neg	✓	✓	✓	×	✓	✓	✓	✓	×
	invert(~)	✓	✓	✓	×	✓	×	×	×	✓
	and(&)	✓	✓	✓	×	✓	×	×	×	✓
	or(|)	✓	✓	✓	×	✓	×	×	×	✓
	xor(^)	✓	✓	✓	×	✓	×	×	×	✓
	not(!)	✓	✓	✓	×	✓	×	×	×	✓
	lshift(<<)	✓	✓	✓	×	✓	×	×	×	×
	rshift(>>)	✓	✓	✓	×	✓	×	×	×	×
	gt	✓	✓	✓	×	✓	✓	✓	✓	✓*
	ge	✓	✓	✓	×	✓	✓	✓	✓	✓*
	lt	✓	✓	✓	×	✓	✓	✓	✓	✓*
	le	✓	✓	✓	×	✓	✓	✓	✓	✓*
	eq	✓	✓	✓	×	✓	✓	✓	✓	✓*
	ne	✓	✓	✓	×	✓	✓	✓	✓	✓*
	logical and	✓	✓	✓	✓	✓	✓	✓	✓	✓
	logical or	✓	✓	✓	✓	✓	✓	✓	✓	✓
	abs	✓	✓	✓	×	✓	✓	✓	✓	✓*
	cdiv	✓	✓	✓	×	✓	✓	✓	✓	×
	ceil	×	×	×	×	×	✓	✓	✓	×
	clamp	×	×	×	×	×	✓	✓	✓	×
	cos	×	×	×	×	×	✓	✓	✓	×
	div_rn	×	×	×	×	×	✓	✓	✓	×
	erf	×	×	×	×	×	✓	✓	✓	×
	exp	×	×	×	×	×	✓	✓	✓	×
	exp2	×	×	×	×	×	✓	✓	✓	×
	fdiv	×	×	×	×	×	✓	✓	✓	×
	floor	×	×	×	×	×	✓	✓	✓	×
	fma	×	×	×	×	×	✓	✓	✓	×
	log	×	×	×	×	×	✓	✓	✓	×
	log2	×	×	×	×	×	✓	✓	✓	×
	maximum	✓	✓	✓	×	✓	✓	✓	✓	✓*
	minimum	✓	✓	✓	×	✓	✓	✓	✓	✓*
	round	×	×	×	×	×	×	✓	×	×
	rsqrt	×	×	×	×	×	✓	✓	✓	×
	sigmoid	×	×	×	×	×	✓	✓	✓	×
	sin	×	×	×	×	×	✓	✓	✓	×
	softmax	×	×	×	×	×	✓	✓	✓	×
	sqrt	×	×	×	×	×	✓	✓	✓	×
	sqrt_rn	×	×	×	×	×	✓	✓	✓	×
	umulhi	×	×	✓	×	×	×	×	×	×
Reduction Ops	argmax	✓	✓	✓	×	✓	✓	✓	✓	×
	argmin	✓	✓	✓	×	✓	✓	✓	✓	×
	max	✓	✓	✓	×	✓	✓	✓	✓	✓*
	min	✓	✓	✓	×	✓	✓	✓	✓	✓*
	reduce	✓	✓	✓	×	✓	✓	✓	✓	✓*
	sum	✓	✓	✓	×	✓	✓	✓	✓	✓*
	xor_sum	✓	✓	✓	×	✓	×	×	×	✓*
Scan/Sort Ops	associative_scan	✓	✓	✓	✓	✓	✓	✓	×	✓
	cumprod	✓	✓	✓	×	✓	✓	✓	✓	✓
	cumsum	✓	✓	✓	×	✓	✓	✓	✓	✓
	histogram	×	×	✓	✓	✓	×	×	×	×
	sort	×	×	×	×	×	×	×	×	×
	gather	×	×	×	×	×	✓	✓	✓	×
Atomic Ops	atomic_add	✓	✓	✓	✓	×	✓	✓	✓	×
	atomic_and	✓	✓	✓	✓	✓	×	×	×	×
	atomic_cas	×	✓	✓	✓	✓	✓	✓	×	×
	atomic_max	✓	✓	✓	×	×	✓	✓	✓	×
	atomic_min	✓	✓	✓	×	×	✓	✓	✓	×
	atomic_or	✓	✓	✓	✓	✓	×	×	×	×
	atomic_xchg	✓	✓	✓	✓	✓	✓	✓	×	×
	atomic_xor	✓	✓	✓	✓	✓	×	×	×	×
Random Number Generation	randint4x	✓	✓	✓	✓	×	×	×	×	✓
	randint	✓	✓	✓	✓	×	×	×	×	✓
	rand	×	×	×	×	×	✓	✓	✓	✓
	randn	×	×	×	×	×	✓	✓	✓	✓
Iterators	range	✓	✓	✓	×	✓	×	×	×	×
	static_range	✓	✓	✓	×	✓	×	×	×	×
Inline Assembly	inline_asm_elementwise	×	×	×	×	×	×	×	×	×
Compiler Hint Ops	assume	×	×	×	×	×	×	×	×	✓
	debug_barrier	✓	✓	✓	×	✓	✓	✓	✓	✓
	max_constancy	×	×	×	×	×	×	×	×	×
	max_contiguous	×	×	×	×	×	×	×	×	×
	multiple_of	×	×	×	×	×	×	×	×	×
Debug Ops	static_print	✓	✓	✓	×	✓	✓	✓	✓	✓
	static_assert	✓	✓	✓	×	✓	✓	✓	✓	✓
	device_print	✓	✓	✓	×	✓	✓	✓	×	✓
	device_assert	×	×	×	×	×	×	×	×	✓

约束说明

• dot: 两个输入A[batch(optional), M, K], B[batch(optional), K, N]。

• gather: triton.gather(x, index, axis)，假设x的shape为n维度，目前只支持axis=n-1。

• permute: triton.permute(x, dims)，不支持dims=[2, 1, 0]。

• trans: triton.trans(x, dims)，不支持dims=[2, 1 , 0]。

• device_print: 需要增加1个环境变量，TRITON_DEVICE_PRINT=1。

• device_assert: 生效需要设置两个环境变量，TRITON_DEBUG=1，TRITON_DEVICE_PRINT=1。

• atomic_add: 昇腾不支持atomic_add实现多核add+保存中间结果，需要修改成普通add来保存中间结果

• atomic类op: 对于昇腾后端，sem只支持默认值"acq_rel"模式，其他值均按默认值处理；scope只支持默认值"gpu"，其他值均按默认值处理

• atomic_or/atomic_xor/atomic_and/atomic_xchg/atomic_cas: 昇腾暂不支持在loop中使用

• permute: 不支持不相邻轴转置，如(0, 1, 2) -> (2, 1, 0)

• trans: 不支持不相邻轴转置，如(0, 1, 2) -> (2, 1, 0)

• umulhi: 不支持负数输入

• mod: int64仅支持处理 -2^24 ~ 2^24 范围内的数值

• rand类op: 所支持的数据类型仅针对算子的输出。

• tensor_descriptor类op: 当前仅支持绑定使用，即 make/load/store_tensor_descriptor 需配套使用

• ALL: int8类型由于特殊处理，会占用更大的片上空间，编译时容易造成ub overflow报错，通常调整tiling即可解决

• ALL: triton kernel中同时存在所有tensor总和不能超过96KB，若关闭double buffer，则不能超过192KB

• ALL: 所有tensor不允许某个shape的size小于1

• ALL: ✓*表示triton内部将bool类型转为int8类型进行运算，并能够执行得到结果的OP

• ALL: 不支持使用shape为"[[]]"的标量tensor进行计算