Release Notes

CUDA 13.5

Support for return statements inside for, while, do, and switch clauses in tile code.
New ct::scaled_mma() function for performing block scaled matrix multiplication.
New ct::tile constructor for building a tile from a list of compile-time constant scalars (for example ct::tile t({1, 2, 3, 4})).
New ct::insert() function to update a portion of a tile with the elements of another tile.
New ct::repack() function for reinterpreting data across multiple elements of a tile.
New masked atomic read-modify-write member functions on ct::partition_view and ct::strided_view which have well-defined behavior for partially out-of-bounds chunks.
New ct::powi() function for elementwise float-to-int exponentiation.
Support for __nv_fp8_e8m0, __nv_fp8_ue5m3, and __nv_fp4_e2m1 as tile elements.
Support for rounding and saturation modes in ct::element_cast().
Support for view padding when loading restricted floating point types from ct::partition_view, ct::strided_view, and ct::gather_scatter_view.
Support for optimization hints on variable declaration and return statements.
The num_worker_warps_per_cta optimization hint now accepts any power of two in the inclusive range \([1, 32]\). Previously only the values \(4\) and \(8\) were supported.
Support for integer exponents in ct::pow().

New ct::strided_view type for performing structured memory accesses over chunks of data separated by a stride.
New ct::gather_scatter_view type for performing non-adjacent memory access to chunks of data along a gather-scatter dimension.
New ct::partition_view type member functions: atomic_and(), atomic_or(), atomic_xor(), atomic_max(), atomic_min(), atomic_add(), atomic_sub().
New ct::max_element() and ct::min_element() functions.
Tile argument constraint for ct::isinf() and ct::isnan() changed from to basic_floating_point_tile to floating_point_tile (more floating point types now supported).
ct::atomic_sub() and ct::atomic_add() now support tile with element type __nv_bfloat16, when targeting architecture >= sm_90.
Added explicit rounding mode template parameter to ct::exp() and ct::tanh(). Supported modes are round full and round approximate.
New ct::strides type for representing stride values in layout mappings and ct::strided_view. Associated APIs include ct::strides_like, ct::dynamic_stride, and ct::strides_equal().
ct::layout_strided_mapping now accepts ct::strides in addition to ct::extents when specifying stride values.
The bitwise complement, left shift, and right shift APIs now support tiles of bool. For details, see ct::operator~(), ct::operator<<(), and ct::operator>>().
Switch statements are now supported in tile code. See the CUDA Programming Guide for limitations.

The addition of the rounding mode template parameter changes the template signature of ct::exp() and ct::tanh(). The first template parameter is now rounding_mode rather than the tile element type. Code that explicitly specified the tile element type as a template argument in CUDA 13.3, such as ct::exp<double>(0.0), will no longer compile. Users should rely on template argument deduction instead, e.g. ct::exp(tile).