triton.language¶

Programming Model¶

`tensor`	Represents an N-dimensional array of values or pointers.
`program_id`	Returns the id of the current program instance along the given `axis`.
`num_programs`	Returns the number of program instances launched along the given `axis`.

`arange`	Returns contiguous values within the half-open interval `[start, end)`.
`cat`	Concatenate the given blocks
`full`	Returns a tensor filled with the scalar value for the given `shape` and `dtype`.
`zeros`	Returns a tensor filled with the scalar value 0 for the given `shape` and `dtype`.
`zeros_like`	Returns a tensor of zeros with the same shape and type as a given tensor.
`cast`	Casts a tensor to the given `dtype`.

`broadcast`	Tries to broadcast the two given blocks to a common compatible shape.
`broadcast_to`	Tries to broadcast the given tensor to a new `shape`.
`expand_dims`	Expand the shape of a tensor, by inserting new length-1 dimensions.
`interleave`	Interleaves the values of two tensors along their last dimension.
`join`	Join the given tensors in a new, minor dimension.
`permute`	Permutes the dimensions of a tensor.
`ravel`	Returns a contiguous flattened view of `x`.
`reshape`	Returns a tensor with the same number of elements as input but with the provided shape.
`split`	Split a tensor in two along its last dim, which must have size 2.
`trans`	Permutes the dimensions of a tensor.
`view`	Returns a tensor with the same elements as input but a different shape.

`dot`	Returns the matrix product of two blocks.
`dot_scaled`	Returns the matrix product of two blocks in microscaling format.

`load`	Return a tensor of data whose values are loaded from memory at location defined by pointer:
`store`	Store a tensor of data into memory locations defined by pointer.
`make_block_ptr`	Returns a pointer to a block in a parent tensor
`advance`	Advance a block pointer

`flip`	Flips a tensor x along the dimension dim.
`where`	Returns a tensor of elements from either `x` or `y`, depending on `condition`.
`swizzle2d`	Transforms the indices of a row-major size_i * size_j matrix into the indices of a column-major matrix for each group of size_g rows.

`abs`	Computes the element-wise absolute value of `x`.
`cdiv`	Computes the ceiling division of `x` by `div`
`ceil`	Computes the element-wise ceil of `x`.
`clamp`	Clamps the input tensor `x` within the range [min, max].
`cos`	Computes the element-wise cosine of `x`.
`div_rn`	Computes the element-wise precise division (rounding to nearest wrt the IEEE standard) of `x` and `y`.
`erf`	Computes the element-wise error function of `x`.
`exp`	Computes the element-wise exponential of `x`.
`exp2`	Computes the element-wise exponential (base 2) of `x`.
`fdiv`	Computes the element-wise fast division of `x` and `y`.
`floor`	Computes the element-wise floor of `x`.
`fma`	Computes the element-wise fused multiply-add of `x`, `y`, and `z`.
`log`	Computes the element-wise natural logarithm of `x`.
`log2`	Computes the element-wise logarithm (base 2) of `x`.
`maximum`	Computes the element-wise maximum of `x` and `y`.
`minimum`	Computes the element-wise minimum of `x` and `y`.
`rsqrt`	Computes the element-wise inverse square root of `x`.
`sigmoid`	Computes the element-wise sigmoid of `x`.
`sin`	Computes the element-wise sine of `x`.
`softmax`	Computes the element-wise softmax of `x`.
`sqrt`	Computes the element-wise fast square root of `x`.
`sqrt_rn`	Computes the element-wise precise square root (rounding to nearest wrt the IEEE standard) of `x`.
`umulhi`	Computes the element-wise most significant N bits of the 2N-bit product of `x` and `y`.

`argmax`	Returns the maximum index of all elements in the `input` tensor along the provided `axis`
`argmin`	Returns the minimum index of all elements in the `input` tensor along the provided `axis`
`max`	Returns the maximum of all elements in the `input` tensor along the provided `axis`
`min`	Returns the minimum of all elements in the `input` tensor along the provided `axis`
`reduce`	Applies the combine_fn to all elements in `input` tensors along the provided `axis`
`sum`	Returns the sum of all elements in the `input` tensor along the provided `axis`
`xor_sum`	Returns the xor sum of all elements in the `input` tensor along the provided `axis`

`associative_scan`	Applies the combine_fn to each elements with a carry in `input` tensors along the provided `axis` and update the carry
`cumprod`	Returns the cumprod of all elements in the `input` tensor along the provided `axis`
`cumsum`	Returns the cumsum of all elements in the `input` tensor along the provided `axis`
`histogram`	computes an histogram based on input tensor with num_bins bins, the bins have a width of 1 and start at 0.
`sort`	Sorts a tensor along a specified dimension.

`atomic_add`	Performs an atomic add at the memory location specified by `pointer`.
`atomic_and`	Performs an atomic logical and at the memory location specified by `pointer`.
`atomic_cas`	Performs an atomic compare-and-swap at the memory location specified by `pointer`.
`atomic_max`	Performs an atomic max at the memory location specified by `pointer`.
`atomic_min`	Performs an atomic min at the memory location specified by `pointer`.
`atomic_or`	Performs an atomic logical or at the memory location specified by `pointer`.
`atomic_xchg`	Performs an atomic exchange at the memory location specified by `pointer`.
`atomic_xor`	Performs an atomic logical xor at the memory location specified by `pointer`.

`randint4x`	Given a `seed` scalar and an `offset` block, returns four blocks of random `int32`.
`randint`	Given a `seed` scalar and an `offset` block, returns a single block of random `int32`.
`rand`	Given a `seed` scalar and an `offset` block, returns a block of random `float32` in \(U(0, 1)\).
`randn`	Given a `seed` scalar and an `offset` block, returns a block of random `float32` in \(\mathcal{N}(0, 1)\).

`range`	Iterator that counts upward forever.
`static_range`	Iterator that counts upward forever.

Execute inline assembly over a tensor.

`debug_barrier`	Insert a barrier to synchronize all threads in a block.
`max_constancy`	Let the compiler know that the value first values in `input` are constant.
`max_contiguous`	Let the compiler know that the value first values in `input` are contiguous.
`multiple_of`	Let the compiler know that the values in `input` are all multiples of `value`.

`static_print`	Print the values at compile time.
`static_assert`	Assert the condition at compile time.
`device_print`	Print the values at runtime from the device.
`device_assert`	Assert the condition at runtime from the device.