For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Mojo struct
OpDesc
struct OpDesc
Describes a single operation in the pipeline schedule.
Fields: tag: The type of operation (kernel-specific, used by _emit dispatch). stage: Buffer stage index (0 or 1 for double-buffering). subtile: Subtile index within the stage (0 or 1). k_offset: How to compute the K dimension offset for loads. vm_cost: Number of vmcnt (global load) ops this produces. lgkm_cost: Number of lgkmcnt (LDS) ops this produces. wait_value: For WAIT_VM/WAIT_LGKM ops, the count to wait for. resource: Hardware execution unit (GLOBAL_MEM, LDS, MMA_UNIT, SCALAR). latency: Estimated execution latency in cycles. role: Pipeline data-flow role (GLOBAL_LOAD, FRAGMENT_LOAD, etc.). channel: Data path identifier for edge derivation. Ops on the same channel share a buffer (e.g., 0=A matrix, 1=B matrix). -1 = none. vgpr_def: VGPRs this op brings into scope (new live register values). vgpr_kill: VGPRs this op releases (last use of some register buffer).
Fields
- tag (
Int): - stage (
Int): - subtile (
Int): - k_offset (
KOffsetKind): - vm_cost (
Int): - lgkm_cost (
Int): - wait_value (
Int): - resource (
ResourceKind): - latency (
Int): - role (
OpRole): - channel (
Int): - vgpr_def (
Int): VGPRs this op brings into scope (new live register values). - vgpr_kill (
Int): VGPRs this op releases (last use of some register buffer).
Implemented traits
AnyType,
Copyable,
ImplicitlyCopyable,
ImplicitlyDestructible,
Movable
Methods
__init__
__init__(out self, *, tag: Int, stage: Int = 0, subtile: Int = 0, k_offset: KOffsetKind = KOffsetKind.NONE, vm_cost: Int = 0, lgkm_cost: Int = 0, wait_value: Int = 0, resource: ResourceKind = ResourceKind.NONE, latency: Int = 0, role: OpRole = OpRole.NONE, channel: Int = -1, vgpr_def: Int = 0, vgpr_kill: Int = 0)
is_present
op
static op(tag: Int, resource: ResourceKind, latency: Int, role: OpRole, *, channel: Int = -1, stage: Int = 0, subtile: Int = 0, k_offset: KOffsetKind = KOffsetKind.NONE, vm_cost: Int = 0, lgkm_cost: Int = 0, wait_value: Int = 0, vgpr_def: Int = 0, vgpr_kill: Int = 0) -> Self
Construct an OpDesc with all metadata inline.
All scheduling metadata (tag, resource, latency, role, channel) is specified directly. Per-instance fields (stage, subtile, k_offset) are keyword arguments.
logical
static logical(tag: Int, *, channel: Int = -1, stage: Int = 0, subtile: Int = 0, k_offset: KOffsetKind = KOffsetKind.NONE) -> Self
Declare a logical op — buffer metadata only, no hardware costs.
Resource, latency, and role are left as sentinel values (NONE/0). Call annotate_ops() with a TargetCostModel to fill them in before scheduling. This separates the algorithm (what ops exist) from the target (how expensive they are).
barrier
static barrier() -> Self
wait_vm
static wait_vm[count: Int]() -> Self
wait_vm_n
static wait_vm_n(count: Int) -> Self
Runtime-parameterized wait_vm (for ScheduleConfig-driven waits).
wait_lgkm
static wait_lgkm[count: Int]() -> Self
wait_lgkm_n
static wait_lgkm_n(count: Int) -> Self
Runtime-parameterized wait_lgkm (for ScheduleConfig-driven waits).
set_prio
static set_prio[priority: Int]() -> Self
Priority hint: s_setpriopriority.
schedule_barrier
static schedule_barrier() -> Self
Compiler scheduling fence.
none
static none() -> Self
No-op sentinel (used for optional MMA block fields).
sched_group_barrier
static sched_group_barrier[mask: Int, count: Int]() -> Self
Schedule_group_barrier hint. Mask encoded in subtile, count in wait_value.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!