MAGIA_V3_REBASED_PR - #55
Open
Dequino wants to merge 87 commits into
Open
Conversation
…instead of float16 HP Float support seems to have problems
…gister initialization
…nd tensor as input
- Fix RedMulE register offsets in tile_ctrl.h (X=0x2C, MCFG0=0x20, etc.)
to match gvsoc LightRedmule req_v2 interface; removes computation errors
in test_mm_os
- Restore mcfg_reg1 format bits and MCFG2 write in redmule_isa_utils.h
- Re-add mhartid guard on BSS clear in crt0.S to prevent hart races
- Include base magia tests for magia_v3 in all TEST_MAPPINGS branches
…cate task symbols
Redesign the PULP cluster software interface to align with the new GVSoC
cluster_regs model:
API changes (magia_pulp_utils.h):
- pulp_init(binary) drops the nb_cores argument; broadcasts CLK_EN and
busy-waits until all PULP cores signal READY
- pulp_run_task(task, core_mask) accepts a one-hot core bitmask, sets
nb_cores_to_wait = popcount(mask) internally, writes the mask to PULP_START,
then polls until PULP_START is cleared by the cores' ACK
pulp_crt0.S rewritten with a two-phase completion protocol:
1. ACK phase (before task): each woken hart writes 0 to PULP_START;
cluster_regs counts ACKs and clears the register, unblocking the CV32 poll
2. DONE phase (after task): each hart writes 1 to PULP_DONE after the task
returns; cluster_regs counts DONEs and fires the DONE IRQ to the CV32 event
unit, waking eu_pulp_wait()
Task symbol deduplication:
- Rename hello_task → hello_pulp_task / hello_spatz_task in all test sources
to avoid HELLO_TASK name collision in combined Spatz+PULP tests
- Remove stale FIRST_TASK_NAME overrides from CMakeLists (default now matches
the renamed function names)
- Update extract_pulp_symbols.sh grep pattern to match renamed symbols
Adds tests/spatz_pulp_on_magia/multi_task — a test that exercises two sequential Spatz tasks and four sequential PULP task dispatches on different core pairs within the same run.
Spatz tasks (fp16 RVV):
1. vec_add: vfadd.vv on A+B → C
2. vec_relu: vfmax.vf(0) on X → Y
PULP tasks (int32, one pair of cores each):
1. vec_sum — cores 0+1 (mask=0x03): parallel partial sums
2. vec_dot — cores 2+3 (mask=0x0C): parallel partial dot product
3. vec_scale — cores 4+5 (mask=0x30): element-wise multiply by scalar
4. vec_absmax — cores 6+7 (mask=0xC0): parallel absolute maximum
Each PULP task uses pair_local_id = GET_PULP_LOCAL_ID(hartid) % 2 to split work between the two dispatched cores. CV32 verifies all results before proceeding to the next dispatch.
multi_task used flat L1_BASE for PULP/Spatz buffers, but each tile's L1 is at L1_BASE + tile_id*L1_TILE_OFFSET -> on a mesh all tiles aliased onto tile 0's L1 and clobbered each other (wrong results / hang; ok in 1x1). Use get_l1_base(get_hartid()) instead. Also: build io.c with -fno-tree-loop-distribute-patterns (memset was rewritten into infinite self-recursion at -O3); CV32 sp starts at top of its stack band; add PULP-only multi_task; hello_spatz_pulp HELLO_TASK -> HELLO_SPATZ_TASK.
Heavy-duty application/benchmark for the magia_v3 4x4 mesh: one inference of
a fused NN block, driving RedMulE, Spatz and the 8-core PULP cluster on every
tile.
Z = X*W + bias RedMulE, input-stationary (mm_is dataflow),
iDMA + FractalSync neighbour handover
P = hardswish(Z*s + b) Spatz, fp16 vector
P = layernorm_rowwise(P) Spatz, fp16 vector
Q = softmax_rowwise_q15 PULP cluster, 8 cores, int16 packed SIMD
At M=N=K=512 that is 262 656 parameters and 268.4 MFLOP per inference, the
size of one attention projection in a d_model=512 transformer. Weights are
pseudo-random and the norm/softmax groups follow the mesh partitioning, so it
is a benchmark and not an inference engine; see the header of main.c.
The phases are pipelined rather than run in sequence: block r of Z is picked up
NN_PIPE_LAG rounds after the round that produced it, which the neighbour syncs
alone already order, so no extra barrier is needed. Measured 394 697 cycles per
inference against 461 658 phase-by-phase. RedMulE runs at 97.7% of peak.
Two feed paths: standalone under gvrun with the matrices inside the ELF, or a
payload staged in L2 by a VFIO host (contract in include/nn_is_params.h).
The dataset headers are generated at build time by scripts/gen_nn_is_data.py
and are not committed: nn_is_data.h alone is 6.3 MB of hex literals and is
fully reproducible from -DNN_SIZE / -DNN_TIMESLOTS / -DNN_SEED.
Also: align .pulp_binary to 256 bytes in link.ld with the boundary labels
inside the section. CV32E40P hardwires mtvec[7:2]=0, so the PULP trap handler
must sit on a 256-byte boundary or a dispatch IRQ vectors into the middle of
the startup code.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Realignment with MAGIA_V3, also rebasing the changes in main with chips-it fork.