"""
Tests for CellMatch-based assemble dependency establishment.
Covers three dependency types:
- WAW (Write After Write): both ops are NORMAL_WRITE → same type + NORMAL_WRITE → mutex
- RAW (Read After Write): consumer=READ, producer=NORMAL_WRITE → different types → mutex
- WAR (Write After Read): consumer=NORMAL_WRITE, producer=READ → different types → mutex
"""
import os
import pypto
import pytest
import numpy as np
import torch
from numpy.testing import assert_allclose
import torch_npu
DTYPE = pypto.DT_FP32
@pytest.mark.skip(reason="temporarily skipped")
def test_stitch_all_dependencies():
"""
Unified test covering WAW, RAW, and WAR in a single kernel.
All phases converge on the SAME region out[8:16] (offset=8):
out[0:8] — WAW: 5 writes fully overlap
out[8:16] — COMMON: RAW producer write → RAW consumers read →
WAR init write → WAR readers read → WAR final write
out[16:24] — RAW consumer 0 result
out[24:32] — RAW consumer 1 result
out[32:40] — WAR reader 0 result
out[40:48] — WAR reader 1 result
out[48:56] — WAR reader 2 result
=== WAW (out[0:8]) ===
5 iterations, each views a different x slice (idx*8) as input.
All assemble to out[0:8] — NORMAL_WRITE + NORMAL_WRITE → mutex (WAW).
Last writer (iter 4, x[32:40]) determines golden.
=== RAW (out[8:16] written, read offset == WAR write offset) ===
Producer: x[8:16] + 100 → out[8:16] (NORMAL_WRITE).
2 consumers: view out[8:16] (READ) + x[16:24]/x[24:32] → out[16:24], out[24:32].
READ vs NORMAL_WRITE → RAW mutex automatically.
=== WAR (out[8:16] read by 3 readers, WRITE offset == RAW READ offset) ===
Init overwrite: x[32:40] + 400 → out[8:16].
3 readers: view out[8:16] (READ) + x[0:8]/x[8:16]/x[16:24] → out[32:40]/[40:48]/[48:56].
Final writer: x[32:40] + 800 → out[8:16] (NORMAL_WRITE).
READ vs NORMAL_WRITE → WAR mutex — final write waits for all readers.
"""
device_id = int(os.environ.get('TILE_FWK_DEVICE_ID', 0))
torch.npu.set_device(device_id)
pypto.runtime._device_init()
x = pypto.tensor([40], DTYPE)
out = pypto.tensor([56], DTYPE)
with pypto.function("main_all_deps", x, out):
pypto.set_vec_tile_shapes(8)
for idx in pypto.loop(5, name="waw", idx_name="idx"):
tmp = pypto.view(x, [8], [idx * 8])
t1 = pypto.add(tmp, 10.0)
t2 = pypto.mul(t1, 2.0)
t3 = pypto.add(t2, 1.0)
pypto.assemble(t3, [0], out)
raw_src = pypto.view(x, [8], [8])
raw_prod = pypto.add(raw_src, 100.0)
pypto.assemble(raw_prod, [8], out)
for idx in pypto.loop(2, name="raw_consumer", idx_name="idx"):
raw_tmp = pypto.view(out, [8], [8])
add_t = pypto.view(x, [8], [16 + idx * 8])
raw_res = pypto.add(raw_tmp, add_t)
pypto.assemble(raw_res, [16 + idx * 8], out)
war_init_src = pypto.view(x, [8], [32])
war_init = pypto.add(war_init_src, 400.0)
pypto.assemble(war_init, [8], out)
for idx in pypto.loop(3, name="war_reader", idx_name="idx"):
war_tmp = pypto.view(out, [8], [8])
add_t = pypto.view(x, [8], [idx * 8])
war_res = pypto.add(war_tmp, add_t)
pypto.assemble(war_res, [32 + idx * 8], out)
war_final_src = pypto.view(x, [8], [32])
war_final = pypto.add(war_final_src, 800.0)
pypto.assemble(war_final, [8], out)
torch_x = torch.arange(40, dtype=torch.float32)
golden = torch.zeros(56, dtype=torch.float32)
golden[0:8] = torch_x[32:40] * 2.0 + 21.0
golden[8:16] = torch_x[32:40] + 800.0
golden[16:24] = torch_x[8:16] + torch_x[16:24] + 100.0
golden[24:32] = torch_x[8:16] + torch_x[24:32] + 100.0
golden[32:40] = torch_x[32:40] + torch_x[0:8] + 400.0
golden[40:48] = torch_x[32:40] + torch_x[8:16] + 400.0
golden[48:56] = torch_x[32:40] + torch_x[16:24] + 400.0
pto_x = pypto.from_torch(torch_x, "x")
torch_out = torch.zeros(56)
pto_out = pypto.from_torch(torch_out, "out")
pypto.runtime._device_run_once_data_from_host(pto_x, pto_out)
assert_allclose(torch_out, golden, atol=1e-5, verbose=True)
pypto.runtime._device_fini()