From 68f909eed553bc9fe7923d024371f4945a87b740 Mon Sep 17 00:00:00 2001
From: Michael Kuron <m.kuron@gmx.de>
Date: Thu, 27 May 2021 13:41:10 +0200
Subject: [PATCH 1/2] Check if fluctuating LB compiles when vectorized
---
lbmpy_tests/test_fluctuating_lb.py | 49 ++++++++++++++++++++++++++++--
1 file changed, 46 insertions(+), 3 deletions(-)
diff --git a/lbmpy_tests/test_fluctuating_lb.py b/lbmpy_tests/test_fluctuating_lb.py
index 4055b0e2..76f4bdf9 100644
--- a/lbmpy_tests/test_fluctuating_lb.py
+++ b/lbmpy_tests/test_fluctuating_lb.py
@@ -6,6 +6,10 @@ from lbmpy.creationfunctions import *
from lbmpy.macroscopic_value_kernels import macroscopic_values_setter
import numpy as np
from lbmpy.moments import is_bulk_moment, is_shear_moment, get_order
+from pystencils.rng import PhiloxTwoDoubles
+
+import pytest
+from pystencils.backends.simd_instruction_sets import get_supported_instruction_sets
def single_component_maxwell(x1, x2, kT, mass):
@@ -54,7 +58,7 @@ def add_pressure_output_to_collision_rule(collision_rule, pressure_field):
collision_rule.main_assignments = collision_rule.main_assignments + pressure_ouput
-def get_fluctuating_lb(size=None, kT=None, omega_shear=None, omega_bulk=None, omega_odd=None, omega_even=None, rho_0=None, target=None):
+def get_fluctuating_lb(size=None, kT=None, omega_shear=None, omega_bulk=None, omega_odd=None, omega_even=None, rho_0=None, target=None, cpu_vectorize_info=None):
# Parameters
stencil = get_stencil('D3Q19')
@@ -81,7 +85,8 @@ def get_fluctuating_lb(size=None, kT=None, omega_shear=None, omega_bulk=None, om
collision_rule = create_lb_collision_rule(
method,
fluctuating={
- 'temperature': kT
+ 'temperature': kT,
+ 'rng_node': PhiloxTwoDoubles,
},
optimization={'cse_global': True}
)
@@ -98,7 +103,7 @@ def get_fluctuating_lb(size=None, kT=None, omega_shear=None, omega_bulk=None, om
{'density': rho, 'velocity': u})
opts = {'cpu_openmp': True,
- 'cpu_vectorize_info': None,
+ 'cpu_vectorize_info': cpu_vectorize_info,
'target': dh.default_target}
# Compile kernels
@@ -242,3 +247,41 @@ def test_point_force(target="cpu"):
momentum, introduced_momentum + 0.5 * point_force, atol=1E-10)
dh.cpu_arrays["force"][force_pos[0],
force_pos[1], force_pos[2]] = np.zeros(3)
+
+@pytest.mark.skipif(not get_supported_instruction_sets(), reason="No vector instruction sets supported")
+@pytest.mark.parametrize('assume_aligned', (True, False))
+@pytest.mark.parametrize('assume_inner_stride_one', (True, False))
+@pytest.mark.parametrize('assume_sufficient_line_padding', (True, False))
+def test_vectorization(assume_aligned, assume_inner_stride_one, assume_sufficient_line_padding):
+ method = create_mrt_orthogonal(
+ stencil=get_stencil('D2Q9'),
+ compressible=True,
+ weighted=True,
+ relaxation_rate_getter=rr_getter)
+ collision_rule = create_lb_collision_rule(
+ method,
+ fluctuating={
+ 'temperature': sp.Symbol("kT"),
+ 'rng_node': PhiloxTwoDoubles,
+ 'block_offsets': (0, 0),
+ },
+ optimization={'cse_global': True}
+ )
+
+ collision = create_lb_update_rule(collision_rule=collision_rule,
+ stencil=method.stencil,
+ method=method,
+ compressible=True,
+ kernel_type='collide_only')
+
+ opts = {'cpu_openmp': False,
+ 'cpu_vectorize_info': {
+ 'instruction_set': get_supported_instruction_sets()[0],
+ 'assume_aligned': assume_aligned,
+ 'assume_inner_stride_one': assume_inner_stride_one,
+ 'assume_sufficient_line_padding': assume_sufficient_line_padding,
+ },
+ 'target': 'cpu'}
+
+ code = ps.create_kernel(collision, **opts)
+ code.compile()
--
GitLab
From 982feddfc73f0df56a574bb54be8d37f85798976 Mon Sep 17 00:00:00 2001
From: Michael Kuron <m.kuron@gmx.de>
Date: Thu, 27 May 2021 15:29:32 +0200
Subject: [PATCH 2/2] check for warning when vectorizing without stride-one
---
lbmpy_tests/test_fluctuating_lb.py | 32 +++++++++++++++++++++++-------
1 file changed, 25 insertions(+), 7 deletions(-)
diff --git a/lbmpy_tests/test_fluctuating_lb.py b/lbmpy_tests/test_fluctuating_lb.py
index 76f4bdf9..3ad3958e 100644
--- a/lbmpy_tests/test_fluctuating_lb.py
+++ b/lbmpy_tests/test_fluctuating_lb.py
@@ -9,7 +9,8 @@ from lbmpy.moments import is_bulk_moment, is_shear_moment, get_order
from pystencils.rng import PhiloxTwoDoubles
import pytest
-from pystencils.backends.simd_instruction_sets import get_supported_instruction_sets
+from pystencils.backends.simd_instruction_sets import get_supported_instruction_sets, get_vector_instruction_set
+from pystencils.cpu.cpujit import get_compiler_config
def single_component_maxwell(x1, x2, kT, mass):
@@ -58,7 +59,7 @@ def add_pressure_output_to_collision_rule(collision_rule, pressure_field):
collision_rule.main_assignments = collision_rule.main_assignments + pressure_ouput
-def get_fluctuating_lb(size=None, kT=None, omega_shear=None, omega_bulk=None, omega_odd=None, omega_even=None, rho_0=None, target=None, cpu_vectorize_info=None):
+def get_fluctuating_lb(size=None, kT=None, omega_shear=None, omega_bulk=None, omega_odd=None, omega_even=None, rho_0=None, target=None):
# Parameters
stencil = get_stencil('D3Q19')
@@ -85,8 +86,7 @@ def get_fluctuating_lb(size=None, kT=None, omega_shear=None, omega_bulk=None, om
collision_rule = create_lb_collision_rule(
method,
fluctuating={
- 'temperature': kT,
- 'rng_node': PhiloxTwoDoubles,
+ 'temperature': kT
},
optimization={'cse_global': True}
)
@@ -103,7 +103,7 @@ def get_fluctuating_lb(size=None, kT=None, omega_shear=None, omega_bulk=None, om
{'density': rho, 'velocity': u})
opts = {'cpu_openmp': True,
- 'cpu_vectorize_info': cpu_vectorize_info,
+ 'cpu_vectorize_info': None,
'target': dh.default_target}
# Compile kernels
@@ -248,6 +248,7 @@ def test_point_force(target="cpu"):
dh.cpu_arrays["force"][force_pos[0],
force_pos[1], force_pos[2]] = np.zeros(3)
+
@pytest.mark.skipif(not get_supported_instruction_sets(), reason="No vector instruction sets supported")
@pytest.mark.parametrize('assume_aligned', (True, False))
@pytest.mark.parametrize('assume_inner_stride_one', (True, False))
@@ -274,14 +275,31 @@ def test_vectorization(assume_aligned, assume_inner_stride_one, assume_sufficien
compressible=True,
kernel_type='collide_only')
+ instruction_sets = get_supported_instruction_sets()
+ if get_compiler_config()['os'] == 'windows':
+ # skip instruction sets supported by the CPU but not by the compiler
+ if 'avx' in instruction_sets and ('/arch:avx2' not in get_compiler_config()['flags'].lower()
+ and '/arch:avx512' not in get_compiler_config()['flags'].lower()):
+ instruction_sets.remove('avx')
+ if 'avx512' in instruction_sets and '/arch:avx512' not in get_compiler_config()['flags'].lower():
+ instruction_sets.remove('avx512')
+ instruction_set = instruction_sets[-1]
+
opts = {'cpu_openmp': False,
'cpu_vectorize_info': {
- 'instruction_set': get_supported_instruction_sets()[0],
+ 'instruction_set': instruction_set,
'assume_aligned': assume_aligned,
'assume_inner_stride_one': assume_inner_stride_one,
'assume_sufficient_line_padding': assume_sufficient_line_padding,
},
'target': 'cpu'}
- code = ps.create_kernel(collision, **opts)
+ if not assume_inner_stride_one and 'storeS' not in get_vector_instruction_set('double', instruction_set):
+ with pytest.warns(UserWarning) as warn:
+ code = ps.create_kernel(collision, **opts)
+ assert 'Could not vectorize loop' in warn[0].message.args[0]
+ else:
+ with pytest.warns(None) as warn:
+ code = ps.create_kernel(collision, **opts)
+ assert len(warn) == 0
code.compile()
--
GitLab