[runwangdl]

Adds DEEPLOY_L1_AS_L2, the compile-time switch that turns a tiled deployment into the untiled baseline the tiling ablation needs.

What "untiled" means here

PULP cluster cores cannot dereference HyperRAM addresses, so an untiled L3 deployment is not "no movement". It is a single tile whose L1-annotated buffers are physically allocated in FC L2, with the cluster computing over the L2 fabric. The switch redirects pi_l1_malloc to pi_l2_malloc and replaces the mchan pushes with memcpy.

On PR #21

Its memcpy substitution is correct and I was wrong to call it lazy: the mchan hardware routes loc into L1 banks unconditionally, so with L1-as-L2 there is genuinely no DMA left to program. Two things were missing:

1. Only mchan_transfer_push_1d was substituted, so any tile taking the 2-D path hung. push_2d is substituted too.
2. The define never reached the code it was meant to switch. It was set with target_compile_definitions(${ProjectId} PRIVATE ...), but ProjectId compiles only CycleCounter.c and deeploytraintest.c. Every L1 allocation and every mchan call lives in the generated network code, which is a separate target (${NETWORK_LIB} / optimizer_network). Both arms therefore produced the same binary — DS-CNN measured 3,044,109 cycles untiled against 3,044,050 tiled, a 59-cycle "difference". It is now a directory-scope add_compile_definitions in DeeployTest/CMakeLists.txt; deeploylib is an INTERFACE target so it takes INTERFACE, not PUBLIC.

Measured once the define actually applied

GAP9, 8 cores, gvsoc, --l2 1024000, every run Errors: 0 out of 4:

network	tiled	untiled	slowdown	opt tiled	opt untiled	slowdown
Autoencoder	2,504,870	62,385,885	24.9×	1,241,834	31,799,906	25.6×
DS-CNN	3,044,050	45,917,437	15.1×	130,744	2,978,135	22.8×
CCT-2	256,721,885	2,093,477,099	8.2×	7,335,848	226,170,099	30.8×
MobileNetV1	177,150,391	2,171,487,469	12.3×	9,102,093	233,206,261	25.6×
ResNet-8	184,286,767	1,070,880,302	5.8×	3,193,912	85,103,850	26.6×

Two things worth reading off this:

• The optimizer column is a near-constant 23–31× across all five networks. SGD is elementwise and memory-bound, with no arithmetic to hide the access latency, so that column reads out the TCDM-vs-L2-fabric cost directly.
• The training column varies with arithmetic intensity, 5.8× to 24.9×. ResNet-8 is lowest because convolutions over large feature maps reuse each fetched byte many times; the autoencoder is highest because it is a pure MLP (0.99 MACs per parameter against 115–160 for the convolutional benchmarks). Tiling's benefit is not a fixed multiplier — it is a function of 1/arithmetic intensity.

The Autoencoder and DS-CNN rows predate PR #70's MLperf-compliant fixtures and are being re-measured; the three convolutional/transformer rows are unaffected (those fixtures are byte-identical).

🤖 Generated with Claude Code

https://claude.ai/code/session_012vb2GVs3qbiyLV1KkV2baN