ClaudeandClaude Opus 5.5
59921a0181
- GPU test: after hop 2 some INNER_GHOST populations are NaN (pulled from OUTER poisoned on hop 1; 236 slots in pinned CPUREF); assert ACTIVE finite and OUTER NaN, keep the full dense-vs-blocked raw value/NaN-mask comparison. - A 1x1 non-periodic fine layout has zero halo pairs: skip the halo table uploads (CUDA rejects mem_alloc(0)); _fine_halo already skips launch. CPU guard test with driver doubles. - Drop the class-level fine_layout default; the followup fake backend now sets fine_layout = None explicitly. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
157 lines
8.0 KiB
Python
157 lines
8.0 KiB
Python
"""Opt-in blocked fine storage for FixedPatchStepper: CPU tables plus GPU gate.
|
|
|
|
Frozen GPU gate (same stream, 8 coarse steps, nonequilibrium start), in both
|
|
periodic and slot_reservoir modes: dense fine storage vs 1x1, 2x2 and 4x2
|
|
fine blocks give bit-identical masked downloads AND bit-identical raw coarse
|
|
and unpacked owner-core fine arrays, including patch-ghost values/NaN masks.
|
|
"""
|
|
from contextlib import nullcontext
|
|
import unittest
|
|
from unittest.mock import Mock, patch
|
|
|
|
import numpy as np
|
|
import pycuda.driver as cuda
|
|
|
|
from CelerisLab.body.geometry.circle import CircleGeometry
|
|
from CelerisLab.body.objects import SimObject
|
|
from CelerisLab.config import LBMConfig
|
|
from CelerisLab.cuda import compiler_v2 as compiler
|
|
from CelerisLab.lbm.blocks import BlockLayout2D
|
|
from CelerisLab.lbm.descriptors import D2Q9_EX, D2Q9_EY
|
|
from CelerisLab.lbm.refinement import ACTIVE, INNER_GHOST, OUTER_GHOST, FixedPatch2D
|
|
from CelerisLab.lbm.refinement_stepper import FixedPatchStepper, blocked_fine_tables
|
|
|
|
NX, NY, BOUNDS, STEPS = 32, 24, (8, 6, 24, 18), 8
|
|
FINE_BLOCKS = ((32, 40), (16, 20), (16, 10)) # fine window 32x40: 1x1, 2x2, 4x2
|
|
W = np.array([4 / 9] + [1 / 9] * 4 + [1 / 36] * 4)
|
|
CX, CY = np.array(D2Q9_EX, float), np.array(D2Q9_EY, float)
|
|
|
|
|
|
def config(mode):
|
|
if mode == "periodic":
|
|
return LBMConfig(nx=NX, ny=NY, viscosity=.1)
|
|
return LBMConfig(nx=NX, ny=NY, viscosity=.1, velocity=0, wall_normal_velocity=.001,
|
|
wall_slot_width=8, y_wall_bc="on_site_velocity", top_wall_bc="reservoir")
|
|
|
|
|
|
def state(ny, nx, seed):
|
|
"""Positive low-Mach equilibrium plus a seeded nonequilibrium part."""
|
|
y, x = np.mgrid[0:ny, 0:nx]
|
|
rho = 1 + 1e-3 * np.cos(2 * np.pi * x / nx)
|
|
ux, uy = .02 * np.sin(2 * np.pi * y / ny), .01 * np.cos(2 * np.pi * x / nx)
|
|
cu = CX[:, None, None] * ux + CY[:, None, None] * uy
|
|
feq = W[:, None, None] * rho * (1 + 3 * cu + 4.5 * cu * cu - 1.5 * (ux * ux + uy * uy))
|
|
noise = np.random.default_rng(seed).uniform(-1, 1, feq.shape) * 1e-3 * W[:, None, None]
|
|
return (feq + noise).astype(np.float32)
|
|
|
|
|
|
def raw(backend):
|
|
"""Unmasked device arrays; blocked fine storage unpacked from owner cores."""
|
|
coarse, fine = backend._levels
|
|
out = [np.empty(coarse.shape, np.float32)]
|
|
out.append(np.empty(fine.shape if backend.fine_layout is None
|
|
else backend.fine_layout.blocked_shape, np.float32))
|
|
cuda.memcpy_dtoh(out[0], coarse.current)
|
|
cuda.memcpy_dtoh(out[1], fine.current)
|
|
if backend.fine_layout is not None:
|
|
out[1] = backend.fine_layout.to_dense(out[1])
|
|
return out
|
|
|
|
|
|
class BlockedFineCPU(unittest.TestCase):
|
|
def forbidden_context(self):
|
|
return Mock(activate=Mock(side_effect=AssertionError("CUDA reached")), sm_arch="sm_70")
|
|
|
|
def test_rejections_before_cuda(self):
|
|
circle = SimObject(0, geometry=CircleGeometry(16.0, 12.0, 2.0), center=(16.0, 12.0), radius=2.0)
|
|
for kw in ({"fine_block": (7, 7)}, {"fine_block": (16,)},
|
|
{"fine_block": (16, 20), "fine_objects": [circle]}):
|
|
with self.subTest(kw=kw), self.assertRaises(ValueError):
|
|
FixedPatchStepper(config("periodic"), BOUNDS, context=self.forbidden_context(), **kw)
|
|
|
|
def test_single_block_layout_allocates_no_empty_halo_tables(self):
|
|
# A 1x1 non-periodic fine layout has zero halo pairs; CUDA rejects mem_alloc(0).
|
|
context = Mock(activate=nullcontext, sm_arch="sm_70")
|
|
module = Mock()
|
|
module.get_global.return_value = (Mock(), 20)
|
|
sizes = []
|
|
|
|
def alloc(size):
|
|
if size <= 0:
|
|
raise AssertionError("zero-byte device allocation")
|
|
sizes.append(size)
|
|
return Mock()
|
|
with patch.object(cuda, "Stream"), patch.object(compiler, "generate_config"), \
|
|
patch.object(compiler, "compile_kernel"), patch.object(compiler, "cleanup_build_dir"), \
|
|
patch.object(compiler, "load_module", return_value=module), \
|
|
patch.object(cuda, "mem_alloc", side_effect=alloc), patch.object(cuda, "memcpy_htod"):
|
|
backend = FixedPatchStepper(config("periodic"), BOUNDS, context=context,
|
|
fine_block=FINE_BLOCKS[0])
|
|
self.assertEqual(len(backend.fine_layout.halo_dst), 0)
|
|
self.assertIsNone(backend._halo_dst)
|
|
backend._fine_halo(Mock()) # no launch without pairs
|
|
module.get_function.return_value.assert_not_called()
|
|
backend.close()
|
|
self.assertTrue(sizes)
|
|
|
|
def test_tables_map_back_to_dense_slots_in_order(self):
|
|
patch = FixedPatch2D.from_config(config("periodic"), BOUNDS)
|
|
fy, fx = patch.fine_shape
|
|
cells = fy * fx
|
|
order = np.argsort(patch.coalesce_coarse)
|
|
for by, bx in FINE_BLOCKS:
|
|
layout = BlockLayout2D.build(fy, fx, by, bx, periodic=(False, False))
|
|
core = set(layout.owner.ravel().tolist())
|
|
targets, children, kind = blocked_fine_tables(patch, layout, order)
|
|
# Explosion targets: owner cores of exactly the non-ACTIVE fine cells.
|
|
np.testing.assert_array_equal(targets, layout.owner.ravel()[patch.ghost_fine])
|
|
self.assertEqual(sorted(targets.tolist()),
|
|
sorted(layout.owner[patch.fine_kind != ACTIVE].tolist()))
|
|
# Children: same rows/child order, same (q, dense cell), owner cores.
|
|
L = int(np.prod(layout.local_shape))
|
|
b, rest = np.divmod(children, np.uint64(9 * L))
|
|
q, c = np.divmod(rest, np.uint64(L))
|
|
dense = patch.coalesce_fine[order]
|
|
np.testing.assert_array_equal(q, dense // cells)
|
|
np.testing.assert_array_equal(b * np.uint64(L) + c,
|
|
layout.owner.ravel()[(dense % cells).astype(np.intp)])
|
|
self.assertTrue(set((b * np.uint64(L) + c).ravel().tolist()) <= core)
|
|
# Kinds follow owner cores; unused halo entries are OUTER.
|
|
np.testing.assert_array_equal(kind[layout.owner.ravel().astype(np.intp)],
|
|
patch.fine_kind.ravel())
|
|
self.assertEqual(set(np.unique(kind).tolist()) - {ACTIVE, INNER_GHOST, OUTER_GHOST}, set())
|
|
|
|
|
|
class BlockedFineGPU(unittest.TestCase):
|
|
def test_dense_and_blocked_fine_storage_are_bit_identical(self):
|
|
from CelerisLab.cuda.context import CudaContext
|
|
for mode in ("periodic", "slot_reservoir"):
|
|
cfg = config(mode)
|
|
patch = FixedPatch2D.from_config(cfg, BOUNDS)
|
|
g_coarse, g_fine = state(NY, NX, 1), state(*patch.fine_shape, 2)
|
|
results = []
|
|
with CudaContext(0) as context:
|
|
for fine_block in (None, *FINE_BLOCKS):
|
|
with FixedPatchStepper(cfg, BOUNDS, context=context, coarse_boundary=mode,
|
|
fine_block=fine_block) as backend:
|
|
backend.initialize(g_coarse, g_fine)
|
|
backend.step(STEPS)
|
|
self.assertEqual(backend.step_count, STEPS)
|
|
results.append((backend.download(), raw(backend)))
|
|
(dense_public, dense_raw), blocked = results[0], results[1:]
|
|
kind = patch.fine_kind
|
|
self.assertTrue(np.isnan(dense_raw[1][:, kind == OUTER_GHOST]).all(), mode)
|
|
# INNER populations may be NaN after hop 2 (pulled from poisoned OUTER);
|
|
# their exact pattern is pinned by the dense-vs-blocked comparison below.
|
|
self.assertTrue(np.isfinite(dense_raw[1][:, kind == ACTIVE]).all(), mode)
|
|
for fine_block, (public, raw_arrays) in zip(FINE_BLOCKS, blocked):
|
|
msg = f"{mode} fine_block={fine_block}"
|
|
for a, b in zip(public + tuple(raw_arrays), dense_public + tuple(dense_raw)):
|
|
np.testing.assert_array_equal(a, b, err_msg=msg) # values and NaN masks
|
|
print(f"{mode}: dense vs {len(blocked)} fine-block layouts bit-identical "
|
|
f"after {STEPS} coarse steps", flush=True)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|