even faster on pimcomp models

2026-07-31 21:15:28 +02:00
parent 9ca1a0ed9f
commit f4a3b012cc
49 changed files with 1923 additions and 583 deletions
@@ -44,6 +44,19 @@ input size and target linear or sublinear time and space. Avoid repeated full-IR
 walks, nested scans, and per-operation recomputation when indexing, caching, or
 a single traversal can express the same behavior.

+Bufferization cost scales with both the number of operations and the number of
+MLIR values it receives. Upstream lowering and scheduling must therefore keep
+repeated work in compact structured operations, such as statically evaluable
+loops and batches, until after bufferization. Do not compensate for avoidable
+pre-bufferization expansion by weakening, partitioning, or special-casing the
+bufferization analysis; measure both operation and value counts at its input.
+Compactness must also be preserved after bufferization so that liveness,
+verification, memory planning, and other downstream passes do not repeat work
+per logical lane or iteration. Keep structured loops and batches intact until
+PIM ISA code generation is forced to scalarize them into concrete per-core
+instructions; earlier expansion requires an explicit semantic necessity and
+before/after operation and value counts.
+
 When linear-or-better complexity is not possible, use the lowest justified
 complexity and report the actual time and space Big-O, the input variable, and
 why a lower bound is not practical. Include that cost in the final report; do