31 lines
1.5 KiB
Plaintext
31 lines
1.5 KiB
Plaintext
# SPDX-License-Identifier: AGPL-3.0-or-later
|
|
# Copyright (C) 2026 David (Corey) Herrera
|
|
# matmul_float16b_lofi.cbtable
|
|
# Diagnostic twin of matmul_float16b.cbtable, NOT a shipped configuration. Identical in
|
|
# every field except math_fidelity, so that a program built from this table and one
|
|
# built from matmul_float16b.cbtable differ in EXACTLY one compile-time setting and can
|
|
# be A/B'd on the card without any other variable moving.
|
|
#
|
|
# Why this exists: docs/tt_bringup_log.md §16.3 found that on the six real MiniCPM5-1B
|
|
# matmul shapes (all long-K, Kt >= 48), LoFi measures BETTER PCC and max-abs than HiFi4
|
|
# — backwards from what less arithmetic precision should do — while on `smoke`, the one
|
|
# short-K shape (Kt=2), the ordering is normal (HiFi4 better). That reversal was left
|
|
# INFERRED, not measured: "probably 16-bit DST accumulation dominating rather than
|
|
# product precision." This table is what makes it possible to hold DST_ACCUM_MODE fixed
|
|
# (0, matching the shipped configuration) and vary ONLY fidelity across a K sweep, which
|
|
# is the first of the two orthogonal experiments the inversion needs
|
|
# (foundry_blackhole_backend_dev_notes.md task: "does the gap scale with K").
|
|
#
|
|
# NEVER pointed to by build_tensix_program.sh in Makefile.ops or any shipped program
|
|
# directory. It exists only for foundry/tt/tests/test_fidelity_sweep.c.
|
|
#
|
|
# Format: tools/emit_tensix_cb_formats.py
|
|
|
|
math_fidelity 0
|
|
approx 0
|
|
dst_accum_mode 0
|
|
|
|
cb 0 Float16_b 32 32 16 4
|
|
cb 1 Float16_b 32 32 16 4
|
|
cb 16 Float16_b 32 32 16 4
|