For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo struct
ContextParams
struct ContextParams
Comptime half of Context. Bundled into one template parameter so callers write Context[params] instead of threading eight separate comptime args.
Tier discriminator (GPU): _tier (a ReduceTier) picks exactly one
of warp / serial / split-K, or Block (also every CPU tier).
emit_tile_width > 1 further marks the tiled (SIMD-on-outputs) tier,
independent of _tier. The same fields exist on CPU but only
emit_tile_width and simd_width matter; _tier is always Block.
_tier is a _-prefixed scaffolder internal (set by the tier
picker) — a body never reads it.
target (comptime) is the dispatcher's switch — "cpu" selects
the CPU backend, anything else the GPU backend.
Fields
- axis (
Int): Axis being reduced. - emit_tile_width (
Int): Rows-per-thread.1for warp/block/serial/split-K tiers,> 1for the tiled tier (SIMD-on-outputs). - BLOCK_SIZE (
Int): Threads per block (kernel-launch shape). GPU-only;1on CPU. - simd_width (
Int): SIMD width. On GPU, the block tier's axis-direction load width (ignored by tiled/warp/serial). On CPU, the width the scaffolder walks the axis with. - target (
StaticString):"cpu"or"gpu"— picks which backend the unifiedrowwise.{reduce, pjoin, once, launch}dispatches to. - name (
StaticString): Operation name propagated from the graph op (_trace_name) so the tier kernels'@__nameidentifies the op in profiler traces. Defaults to"rowwise"when the caller doesn't provide one. Scaffolder-internal — a body never reads it. - pdl_level (
PDLLevel): Programmatic dependent launch level of the GPU tier kernels (NVIDIA SM90+ only).ONwaits on the predecessor grid at entry and releases the dependent grids at exit;OVERLAP_AT_BEGINNINGreleases them right after the wait.NO_WAIT_OVERLAP_AT_ENDis unsupported, and the per-element split-K tier always releases at exit. Scaffolder-internal — a body never reads it.
Implemented traits
AnyType,
Copyable,
Deinitable,
ImplicitlyCopyable,
Movable,
RegisterPassable,
TrivialRegisterPassable
Methods
__init__
def __init__(axis: Int, emit_tile_width: Int, BLOCK_SIZE: Int, simd_width: Int, target: StaticString, tier: ReduceTier = ReduceTier.Block, num_phases: Int = Int(0), name: StaticString = "rowwise", pdl_level: PDLLevel = PDLLevel.OVERLAP_AT_BEGINNING) -> Self
Initializes a ContextParams from per-tier comptime values.
Args:
- axis (
Int): Axis being reduced. - emit_tile_width (
Int): Rows-per-thread (1for cooperative tiers). - BLOCK_SIZE (
Int): Threads per block. - simd_width (
Int): SIMD width for the block tier's axis load. - target (
StaticString):"cpu"or"gpu". - tier (
ReduceTier): Tier discriminator (ReduceTier.Blockby default — every CPU tier and the GPU tiled/cooperative tiers). - num_phases (
Int): Per-element-output split-K total phase count (<= 1disables the tier). - name (
StaticString): Operation name for the tier kernels'@__name(defaults to"rowwise"). - pdl_level (
PDLLevel): PDL level of the GPU tier kernels.