Visualizations#
Explore how cuda.coop.cutlass primitives move, combine, compare, count, decode,
order, and choose values. Change the settings, step through the stages, and
select a value to follow its ownership. These diagrams show data movement;
their timing and geometry do not predict GPU performance.
Using the kernel fragments#
The fragments use import cuda.coop.cutlass as coop with CuTe kernels.
They assume one-dimensional blocks and an items_per_thread kernel
argument declared cutlass.Constexpr on both kernel and launcher. Use
cutlass.range_constexpr for loops that index payload slots.
Inside @cute.kernel, obtain block_index from
cute.arch.block_idx()[0] and thread_rank from
cute.arch.thread_idx()[0]. Pass a supported CuTe tensor or pointer-backed
view to Load/Store. For scalar indexing, construct a tensor with the required
layout. The Programming Guide and
examples provide complete launch and memory examples.
The diagrams describe the supported contracts. They simplify thread counts and intermediate stages for teaching; they do not predict instruction timing, transaction counts, or GPU performance.