Lower tensor programs to efficient execution

Connect high-level tensor semantics to layouts, buffers, schedules, vector instructions, and accelerator kernels with verifiable lowering steps.

10 published lessons1 module262 min of guided study
  1. 01
    Lower tensor programs to efficient executionadvanced · 10 lessons

Follow the path or begin with a question. Every lesson is open.

Your notebook0 of 10 lessons completed

Your place, at your pace. Progress is optional and saved only in this browser when available.

Inside this module

Each lesson includes an interactive lab

Lower tensor programs to efficient execution

Connect high-level tensor semantics to layouts, buffers, schedules, vector instructions, and accelerator kernels with verifiable lowering steps.

  1. 01

    Affine accesses and legal execution order

    Separate analyzable index expressions from permission to reorder dependent operations.

    26 min
  2. 02

    Logical tensors and physical layouts

    Track shapes, strides, and layout constraints without mistaking a view for a required data copy.

    25 min
  3. 03

    Bufferization preserves tensor versions

    Convert value semantics to storage while respecting aliases, old reads, ownership, and asynchronous users.

    26 min
  4. 04

    Tile coverage, remainders, and legal interchange

    Partition iteration domains without dropping tails or reversing dependencies in the pursuit of reuse.

    26 min
  5. 05

    Fusion preserves effects and numeric contracts

    Remove intermediate traffic only when consumers, side effects, and floating-point semantics remain valid.

    27 min
  6. 06

    Vector lanes and safe tails

    Map independent iterations to lanes while preserving bounds, masks, aliasing, and floating-point order.

    26 min
  7. 07

    Promote data within a complete resource budget

    Stage reused values only when memory-space semantics, synchronization, and occupancy constraints remain valid.

    26 min
  8. 08

    Outlining kernels and mapping launch coordinates

    Preserve captured arguments, execution scope, and asynchronous dependencies when separating host and device work.

    26 min
  9. 09

    Autotuning pays for every candidate

    Search legal schedules with measured costs while accounting for compilation expense, noise, and workload overfit.

    27 min
  10. 10

    Differential testing needs a valid reference domain

    Compare lowering stages under the same semantics, with explicit floating-point criteria and memory-safety checks.

    27 min
All engineering modules