perf(PULPOpen): let Gemm take a one-dimensional bias instead of widening it

[runwangdl]

GEMMLayer.computeShapes widens the C operand to [M, N] unconditionally, because PULP_Gemm_fp32_fp32_fp32_fp32 walked the bias as a full matrix (pDstC[i * O + j]). A Linear's bias is O values shared by every output row, so the widened operand stores the same vector M times — on CCT-2 at 64 tokens, 32 KB to carry 512 B.

What changes

The kernel takes a bias row stride. O reproduces the previous walk exactly; 0 makes every row re-read the same vector, so one kernel serves both layouts and the existing behaviour is bit-identical when the stride is O.

file	change
TargetLibraries/PULPOpen/src/Gemm.c	biasStride parameter, 4 indexing sites
TargetLibraries/{PULPOpen,GAP9}/inc/...	declarations
Templates/FloatGemmTemplate.py	stride chosen from the C operand's rank
PULPOpen/Layers.py	PULPGEMMLayer skips the widening for a 1-D bias
{PULPOpen,GAP9}/Platform.py	register it
TileConstraints/GEMMTileConstraint.py	tie only the trailing dimension

Targets/Generic/Layers.py is untouched, so other targets keep the [M, N] bias their kernels expect.

One thing worth knowing

GAP9 keeps its own operator mapping in Deeploy/Targets/GAP9/Platform.py. Registering a layer only in the PULPOpen mapping has no effect on a GAP9 deployment, and an unregistered change is indistinguishable from an ineffective one when you measure it. Both mappings are updated here.

Tiling

The kernel side is only half of it. serializeTilingSolution built the C cube from
the output cube, so a [O] bias was still transferred as if it were [M, O] -- on
CCT at 64 tokens, 32768 B moved out of a 512 B buffer. A broadcast bias now gets a
cube of its own rank, with the trailing extent tracking the output tile and the
leading dimensions held at 1. Measured on the generated code for cct_train at
--l1 128000 --defaultMemLevel L3, node_84:

	C_transfer_size	total
before	{28672, 4096}	32768 B
after	{448, 64}	512 B

512 B is the entire bias, O = 128 floats across the two O tiles of 112 and 16.

This is invisible whenever the operand is already resident, which is why
siracusa-training-tiled-l2 passed while the four L3 and promote jobs each failed
with 4 errors out of 4, and why the GAP9 jobs, whose smaller L1 budget produces a
different tiling solution, did not catch it either.

Verification

GAP9 gvsoc, CCT linear probing, --l1 122000 --defaultMemLevel L3: Errors: 0, 37.69M train cycles.

Scope

This reaches graphs whose Linear layers are exported as Gemm. The single-block CCT variants measured alongside it are exported as MatMul + Add and contain no Gemm nodes, so their 252 KB of widened bias is untouched here — that path is addressed separately.

🤖 Generated with Claude Code

[runwangdl]

Folded into #54 so this ships as a single merge. The branch is kept, and the content is in #54 with its description.