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 module

program_builder

Program builder: constructs PipelineProgram from a loop body.

Contains:

_construct_mma_blocks, _redistribute_globals
build_double_buffer_program
derive_waits_from_blocks, derive_safe_max_globals, derive_drain_mask
dump_program_blocks
verify* functions, verify_schedule
build_kernel_program
default_kernel_deps_double_buffer
single_buffer_reorder, optimize_within_barriers
mma_block_interleave, mma_block_interleave_list, double_buffer_reorder

Functions

build_double_buffer_program: Build PipelineProgram from LDG ops in the given execution order.
build_kernel_program: Build the finalized PipelineProgram for a double-buffer kernel.
default_kernel_deps_double_buffer: Derive kernel-phase deps for double-buffer by scanning MMA positions.
derive_drain_mask: Derive per-block lgkm drain mask from block content analysis.
derive_safe_max_globals: Derive safe max_globals for uniform global load distribution.
derive_waits_from_blocks: Derive wait counts from the finalized block structure.
double_buffer_reorder: Reorder a double-buffer spec's logical ops into interleaved execution order.
dump_program_blocks: Print the MMA block layout for diagnostic analysis.
mma_block_interleave: Interleave ops across MMA blocks for latency hiding.
mma_block_interleave_list: List-based MMA block interleave (equivalent to mma_block_interleave).
optimize_within_barriers: CSP-optimize op ordering within barrier-delimited segments.
single_buffer_reorder: Single-buffer pipeline reorder: logical iteration → pipelined steady-state.
verify_schedule: Verify structural invariants of a finalized pipeline schedule.

Functions​

Functions