| """Drive the simulated multimesh_v8t card through the driver's own seam. |
| |
| python scripts/py/v8t_card_run.py [--build build/vlt_v8t_card] |
| |
| Four things, each a fact about the image rather than about the model: |
| 1. every node's agent answers A_CAPS through its control window (station -> |
| 32->64 converter -> orchestrator); |
| 2. a block written through node 0's memory window reads back through nodes |
| 1..3 (station -> node -> Xache -> the flat 16 GB), and the DRAM behind the |
| Xache holds it (write-through, checked at the backdoor); |
| 3. node 0's mover copies 64 words laid out across all four channels (16 KB |
| stride) to another such region, and node 2 reads the copy back; |
| 4. the station's DECERR counter is still zero. |
| """ |
|
|
| import argparse |
| import pathlib |
| import sys |
| import time |
|
|
| from kohakuaccel.device import mover, rv64load |
| from kohakuaccel.device.registers import A_CAPS |
| from kohakuaccel.transport.verilator import VerilatorTransport |
| from kohakutpu.clock.card import load_board |
| from kohakutpu.host import board_map |
|
|
| ILV_LG, HOME_LSB, NHOME_LG = 14, 32, 2 |
| ROOT = pathlib.Path(__file__).resolve().parents[2] |
|
|
|
|
| def rotate(addr: int) -> tuple[int, int]: |
| """(home, in-channel byte address) of a flat address, the Xache's way: |
| pairs (i, i+2) for i = ILV_LG .. HOME_LSB-1, applied in order.""" |
| a = addr |
| for i in range(ILV_LG, HOME_LSB): |
| j = i + NHOME_LG |
| bi, bj = (a >> i) & 1, (a >> j) & 1 |
| a &= ~((1 << i) | (1 << j)) |
| a |= (bi << j) | (bj << i) |
| return (a >> HOME_LSB) & 3, a & ((1 << HOME_LSB) - 1) |
|
|
|
|
| def main() -> int: |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--build", default=None) |
| ap.add_argument("--native", action="store_true") |
| ap.add_argument("--settle", type=int, default=40000) |
| a = ap.parse_args() |
|
|
| board = load_board("multimesh_v8t") |
| m = board_map(board) |
| ctrl, mem = m["ctrl"], m["mem"] |
| fails = 0 |
|
|
| t0 = time.monotonic() |
| t = VerilatorTransport(build_dir=a.build, wsl=not a.native, settle=a.settle) |
| print(f"model up in {time.monotonic() - t0:.1f}s: {t.ready}") |
|
|
| |
| for i in range(4): |
| caps = t.read64(ctrl[i] + A_CAPS) |
| fw, grid = caps & 0xFFFF, (caps >> 16) & 0xFF |
| ok = fw == 288 |
| fails += not ok |
| print( |
| f" node {i} A_CAPS {caps:#018x} flit_width={fw} grid={grid} {'ok' if ok else 'WRONG'}" |
| ) |
|
|
| |
| addr = 0x0001_0000 |
| data = bytes(((i * 7 + 3) ^ (i >> 5)) & 0xFF for i in range(1024)) |
| t0 = time.monotonic() |
| t.write_block(mem[0] + addr, data) |
| print( |
| f" wrote {len(data)} B through node 0 at flat {addr:#x} in {time.monotonic() - t0:.1f}s" |
| ) |
| for i in range(1, 4): |
| got = t.read_block(mem[i] + addr, len(data)) |
| ok = got == data |
| fails += not ok |
| print( |
| f" node {i} reads it back: {'ok' if ok else 'MISMATCH ' + got[:32].hex()}" |
| ) |
| home, inch = rotate(addr) |
| word = t.backdoor_read(home, inch >> 6) |
| ok = word == data[:64] |
| fails += not ok |
| print( |
| f" DRAM channel {home} word {inch >> 6:#x} (write-through): {'ok' if ok else 'MISMATCH ' + word[:16].hex()}" |
| ) |
|
|
| |
| src, dst = 0x0010_0000, 0x0020_0000 |
| stride, per = 1 << ILV_LG, 16 |
| pattern = {} |
| for c in range(4): |
| blob = bytes(((c * 31 + k) * 13 + 5) & 0xFF for k in range(per * 32)) |
| pattern[c] = blob |
| t.write_block(mem[0] + src + c * stride, blob) |
| print( |
| f" source: 4 x {per} words at {src:#x} + c*{stride:#x} (homes " |
| f"{[rotate(src + c * stride)[0] for c in range(4)]})" |
| ) |
| walk = [(4, stride), (per, 32)] |
| prog = mover.copy( |
| mover.Walker(src, dims=walk), mover.Walker(dst, dims=walk), ewidth=mover.W32 |
| ) |
| stat0 = mover.status(t.read64(ctrl[0] + mover.AUX_STAT)) |
| t0 = time.monotonic() |
| mover.issue(t, prog, ctrl[0]) |
| for _ in range(200): |
| st = mover.status(t.read64(ctrl[0] + mover.AUX_STAT)) |
| if not st["busy"]: |
| break |
| t.run(500) |
| else: |
| print(" mover: still busy after 100k cycles") |
| fails += 1 |
| moved = (st["moves"] - stat0["moves"]) & 0xFF_FFFF |
| print( |
| f" mover on node 0: fault={st['fault']} moves+={moved} reads+={(st['reads'] - stat0['reads']) & 0xFFFF} " |
| f"writes+={(st['writes'] - stat0['writes']) & 0xFFFF} ({time.monotonic() - t0:.1f}s)" |
| ) |
| fails += st["fault_code"] != 0 |
| for c in range(4): |
| got = t.read_block(mem[2] + dst + c * stride, per * 32) |
| ok = got == pattern[c] |
| fails += not ok |
| print( |
| f" node 2 reads the copy, channel-slice {c}: {'ok' if ok else 'MISMATCH ' + got[:32].hex()}" |
| ) |
|
|
| |
| elf = ROOT / "build" / "rv64" / "hello_kohakuaccel.elf" |
| if elf.exists(): |
| win = rv64load.LoadWindow(t, ctrl[0] + rv64load.WINDOW_OFFSET) |
| t0 = time.monotonic() |
| info = win.load_elf(elf) |
| print( |
| f" loaded in {time.monotonic() - t0:.1f}s ({t.calls} host calls so far)", |
| flush=True, |
| ) |
| win.stdin("Kohaku\n") |
| print(f" stdin queued; status {win.status():#x}", flush=True) |
| r = win.run( |
| expect="Nice to meet you, Kohaku", timeout=300, poll=lambda: t.run(2000) |
| ) |
| print( |
| f" node 0 RV64 loaded through +0x8000: text {info['text']} B, spad {info['spad']} B " |
| f"({time.monotonic() - t0:.1f}s incl. run)" |
| ) |
| print(f" console: {r['console'].strip()!r}") |
| print( |
| f" exit {r['exit']:#x} halted={r['halted']} status={r['status']:#x} in {r['seconds']:.1f}s " |
| f"{'ok' if r['ok'] and r['exit'] == 0 else 'WRONG'}" |
| ) |
| fails += not (r["ok"] and r["exit"] == 0) |
| else: |
| print(f" (no {elf}: RV64 program step skipped)") |
| fails += 1 |
|
|
| |
| s = t.status() |
| print( |
| f" station DECERR {s['decerr']:#x}; sys cycles {s['sys']}, ctrl cycles {s['ctrl']}, host calls {t.calls}" |
| ) |
| fails += s["decerr"] != 0 |
| t.close() |
| print("PASS" if not fails else f"FAIL ({fails})") |
| return 0 if not fails else 1 |
|
|
|
|
| if __name__ == "__main__": |
| sys.exit(main()) |
|
|