Skip to content
Snippets Groups Projects

Compare revisions

Changes are shown as if the source revision was being merged into the target revision. Learn more about comparing revisions.

Source

Select target project
No results found

Target

Select target project
No results found
Show changes
Hide whitespace changes
Inline Side-by-side
Showing
with 2177 additions and 972 deletions
pystencils_tests/test_random.py deleted 100644 → 0
View file @ ef496986
import numpy as np
import pystencils as ps
from pystencils.rng import PhiloxFourFloats, PhiloxTwoDoubles, AESNIFourFloats, AESNITwoDoubles
# curand_Philox4x32_10(make_uint4(124, i, j, 0), make_uint2(0, 0))
philox_reference = np.array([[[3576608082, 1252663339, 1987745383, 348040302],
[1032407765, 970978240, 2217005168, 2424826293]],
[[2958765206, 3725192638, 2623672781, 1373196132],
[ 850605163, 1694561295, 3285694973, 2799652583]]])
def test_philox_double():
for target in ('cpu', 'gpu'):
dh = ps.create_data_handling((2, 2), default_ghost_layers=0, default_target=target)
f = dh.add_array("f", values_per_cell=2)
dh.fill('f', 42.0)
philox_node = PhiloxTwoDoubles(dh.dim)
assignments = [philox_node,
ps.Assignment(f(0), philox_node.result_symbols[0]),
ps.Assignment(f(1), philox_node.result_symbols[1])]
kernel = ps.create_kernel(assignments, target=dh.default_target).compile()
dh.all_to_gpu()
dh.run_kernel(kernel, time_step=124)
dh.all_to_cpu()
arr = dh.gather_array('f')
assert np.logical_and(arr <= 1.0, arr >= 0).all()
x = philox_reference[:,:,0::2]
y = philox_reference[:,:,1::2]
z = x ^ y << (53 - 32)
double_reference = z * 2.**-53 + 2.**-54
assert(np.allclose(arr, double_reference, rtol=0, atol=np.finfo(np.float64).eps))
def test_philox_float():
for target in ('cpu', 'gpu'):
dh = ps.create_data_handling((2, 2), default_ghost_layers=0, default_target=target)
f = dh.add_array("f", values_per_cell=4)
dh.fill('f', 42.0)
philox_node = PhiloxFourFloats(dh.dim)
assignments = [philox_node] + [ps.Assignment(f(i), philox_node.result_symbols[i]) for i in range(4)]
kernel = ps.create_kernel(assignments, target=dh.default_target).compile()
dh.all_to_gpu()
dh.run_kernel(kernel, time_step=124)
dh.all_to_cpu()
arr = dh.gather_array('f')
assert np.logical_and(arr <= 1.0, arr >= 0).all()
float_reference = philox_reference * 2.**-32 + 2.**-33
assert(np.allclose(arr, float_reference, rtol=0, atol=np.finfo(np.float32).eps))
def test_aesni_double():
dh = ps.create_data_handling((2, 2), default_ghost_layers=0, default_target="cpu")
f = dh.add_array("f", values_per_cell=2)
dh.fill('f', 42.0)
aesni_node = AESNITwoDoubles(dh.dim)
assignments = [aesni_node,
ps.Assignment(f(0), aesni_node.result_symbols[0]),
ps.Assignment(f(1), aesni_node.result_symbols[1])]
kernel = ps.create_kernel(assignments, target=dh.default_target).compile()
dh.all_to_gpu()
dh.run_kernel(kernel, time_step=124)
dh.all_to_cpu()
arr = dh.gather_array('f')
assert np.logical_and(arr <= 1.0, arr >= 0).all()
def test_aesni_float():
dh = ps.create_data_handling((2, 2), default_ghost_layers=0, default_target="cpu")
f = dh.add_array("f", values_per_cell=4)
dh.fill('f', 42.0)
aesni_node = AESNIFourFloats(dh.dim)
assignments = [aesni_node] + [ps.Assignment(f(i), aesni_node.result_symbols[i]) for i in range(4)]
kernel = ps.create_kernel(assignments, target=dh.default_target).compile()
dh.all_to_gpu()
dh.run_kernel(kernel, time_step=124)
dh.all_to_cpu()
arr = dh.gather_array('f')
assert np.logical_and(arr <= 1.0, arr >= 0).all()
\ No newline at end of file
pystencils_tests/test_size_and_layout_checks_llvm.py deleted 100644 → 0
View file @ ef496986
import numpy as np
from pystencils import Assignment, Field
from pystencils.llvm import create_kernel, make_python_function
def test_size_check():
"""Kernel with two fixed-sized fields creating with same size but calling with wrong size"""
src = np.zeros((20, 21, 9))
dst = np.zeros_like(src)
sym_src = Field.create_from_numpy_array("src", src, index_dimensions=1)
sym_dst = Field.create_from_numpy_array("dst", dst, index_dimensions=1)
update_rule = Assignment(sym_dst(0),
sym_src[-1, 1](1) + sym_src[1, -1](2))
ast = create_kernel([update_rule])
func = make_python_function(ast)
# change size of src field
new_shape = [a - 7 for a in src.shape]
src = np.zeros(new_shape)
dst = np.zeros(new_shape)
try:
func(src=src, dst=dst)
assert False, "Expected ValueError because fields with different sized where passed"
except ValueError:
pass
def test_fixed_size_mismatch_check():
"""Create kernel with two differently sized but constant fields """
src = np.zeros((20, 21, 9))
dst = np.zeros((21, 21, 9))
sym_src = Field.create_from_numpy_array("src", src, index_dimensions=1)
sym_dst = Field.create_from_numpy_array("dst", dst, index_dimensions=1)
update_rule = Assignment(sym_dst(0),
sym_src[-1, 1](1) + sym_src[1, -1](2))
try:
create_kernel([update_rule])
assert False, "Expected ValueError because fields with different sized where passed"
except ValueError:
pass
def test_fixed_and_variable_field_check():
"""Create kernel with two variable sized fields - calling them with different sizes"""
src = np.zeros((20, 21, 9))
sym_src = Field.create_from_numpy_array("src", src, index_dimensions=1)
sym_dst = Field.create_generic("dst", spatial_dimensions=2, index_dimensions=1)
update_rule = Assignment(sym_dst(0),
sym_src[-1, 1](1) + sym_src[1, -1](2))
try:
create_kernel([update_rule])
assert False, "Expected ValueError because fields with different sized where passed"
except ValueError:
pass
def test_two_variable_shaped_fields():
src = np.zeros((20, 21, 9))
dst = np.zeros((22, 21, 9))
sym_src = Field.create_generic("src", spatial_dimensions=2, index_dimensions=1)
sym_dst = Field.create_generic("dst", spatial_dimensions=2, index_dimensions=1)
update_rule = Assignment(sym_dst(0),
sym_src[-1, 1](1) + sym_src[1, -1](2))
ast = create_kernel([update_rule])
func = make_python_function(ast)
try:
func(src=src, dst=dst)
assert False, "Expected ValueError because fields with different sized where passed"
except ValueError:
pass
pystencils_tests/test_sliced_iteration.py deleted 100644 → 0
View file @ ef496986
import numpy as np
import sympy as sp
from pystencils import Assignment, Field, TypedSymbol, create_kernel, make_slice
from pystencils.simp import sympy_cse_on_assignment_list
def test_sliced_iteration():
size = (4, 4)
src_arr = np.ones(size)
dst_arr = np.zeros_like(src_arr)
src_field = Field.create_from_numpy_array('src', src_arr)
dst_field = Field.create_from_numpy_array('dst', dst_arr)
a, b = sp.symbols("a b")
update_rule = Assignment(dst_field[0, 0],
(a * src_field[0, 1] + a * src_field[0, -1] +
b * src_field[1, 0] + b * src_field[-1, 0]) / 4)
x_end = TypedSymbol("x_end", "int")
s = make_slice[1:x_end, 1]
x_end_value = size[1] - 1
kernel = create_kernel(sympy_cse_on_assignment_list([update_rule]), iteration_slice=s).compile()
kernel(src=src_arr, dst=dst_arr, a=1.0, b=1.0, x_end=x_end_value)
expected_result = np.zeros(size)
expected_result[1:x_end_value, 1] = 1
np.testing.assert_almost_equal(expected_result, dst_arr)
def test_sliced_iteration_llvm():
size = (4, 4)
src_arr = np.ones(size)
dst_arr = np.zeros_like(src_arr)
src_field = Field.create_from_numpy_array('src', src_arr)
dst_field = Field.create_from_numpy_array('dst', dst_arr)
a, b = sp.symbols("a b")
update_rule = Assignment(dst_field[0, 0],
(a * src_field[0, 1] + a * src_field[0, -1] +
b * src_field[1, 0] + b * src_field[-1, 0]) / 4)
x_end = TypedSymbol("x_end", "int")
s = make_slice[1:x_end, 1]
x_end_value = size[1] - 1
import pystencils.llvm as llvm_generator
ast = llvm_generator.create_kernel(sympy_cse_on_assignment_list([update_rule]), iteration_slice=s)
kernel = llvm_generator.make_python_function(ast)
kernel(src=src_arr, dst=dst_arr, a=1.0, b=1.0, x_end=x_end_value)
expected_result = np.zeros(size)
expected_result[1:x_end_value, 1] = 1
np.testing.assert_almost_equal(expected_result, dst_arr)
pystencils_tests/test_small_block_benchmark.ipynb deleted 100644 → 0
View file @ ef496986
%% Cell type:code id: tags:
``` python
from pystencils.session import *
from time import perf_counter
from statistics import median
from functools import partial
```
%% Cell type:markdown id: tags:
## Benchmark for Python call overhead
%% Cell type:code id: tags:
``` python
inner_repeats = 100
outer_repeats = 5
sizes = [2**i for i in range(1, 8)]
sizes
```
%% Output
$$\left [ 2, \quad 4, \quad 8, \quad 16, \quad 32, \quad 64, \quad 128\right ]$$
[2, 4, 8, 16, 32, 64, 128]
%% Cell type:code id: tags:
``` python
def benchmark_pure(domain_size, extract_first=False):
src = np.zeros(domain_size)
dst = np.zeros_like(src)
f_src, f_dst = ps.fields("src, dst", src=src, dst=dst)
kernel = ps.create_kernel(ps.Assignment(f_dst.center, f_src.center)).compile()
if extract_first:
kernel = kernel.kernel
start = perf_counter()
for i in range(inner_repeats):
kernel(src=src, dst=dst)
src, dst = dst, src
end = perf_counter()
else:
start = perf_counter()
for i in range(inner_repeats):
kernel(src=src, dst=dst)
src, dst = dst, src
end = perf_counter()
return (end - start) / inner_repeats
def benchmark_datahandling(domain_size, parallel=False):
dh = ps.create_data_handling(domain_size, parallel=parallel)
f_src = dh.add_array('src')
f_dst = dh.add_array('dst')
kernel = ps.create_kernel(ps.Assignment(f_dst.center, f_src.center)).compile()
start = perf_counter()
for i in range(inner_repeats):
dh.run_kernel(kernel)
dh.swap('src', 'dst')
end = perf_counter()
return (end - start) / inner_repeats
name_to_func = {
'pure_extract': partial(benchmark_pure, extract_first=True),
'pure_no_extract': partial(benchmark_pure, extract_first=False),
'dh_serial': partial(benchmark_datahandling, parallel=False),
'dh_parallel': partial(benchmark_datahandling, parallel=True),
}
```
%% Cell type:code id: tags:
``` python
result = {'block_size': [],
'name': [],
'time': []}
for bs in sizes:
print("Computing size ", bs)
for name, func in name_to_func.items():
for i in range(outer_repeats):
time = func((bs, bs))
result['block_size'].append(bs)
result['name'].append(name)
result['time'].append(time)
```
%% Output
Computing size 2
Computing size 4
Computing size 8
Computing size 16
Computing size 32
Computing size 64
Computing size 128
%% Cell type:code id: tags:
``` python
if 'is_test_run' not in globals():
import pandas as pd
import seaborn as sns
data = pd.DataFrame.from_dict(result)
plt.subplot(1,2,1)
sns.barplot(x='block_size', y='time', hue='name', data=data, alpha=0.6)
plt.yscale('log')
plt.subplot(1,2,2)
data = pd.DataFrame.from_dict(result)
sns.barplot(x='block_size', y='time', hue='name', data=data, alpha=0.6)
```
%% Output
pystencils_tests/test_stencils.py deleted 100644 → 0
View file @ ef496986
import pystencils as ps
pystencils_tests/test_sum_prod.py deleted 100644 → 0
View file @ ef496986
# -*- coding: utf-8 -*-
#
# Copyright © 2019 Stephan Seitz <stephan.seitz@fau.de>
#
# Distributed under terms of the GPLv3 license.
"""
"""
import numpy as np
import sympy
from sympy.abc import k
import pystencils
from pystencils.data_types import create_type
def test_sum():
sum = sympy.Sum(k, (k, 1, 100))
expanded_sum = sum.doit()
print(sum)
print(expanded_sum)
x = pystencils.fields('x: float32[1d]')
assignments = pystencils.AssignmentCollection({
x.center(): sum
})
ast = pystencils.create_kernel(assignments)
code = str(pystencils.show_code(ast))
kernel = ast.compile()
print(code)
assert 'double sum' in code
array = np.zeros((10,), np.float32)
kernel(x=array)
assert np.allclose(array, int(expanded_sum) * np.ones_like(array))
def test_sum_use_float():
sum = sympy.Sum(k, (k, 1, 100))
expanded_sum = sum.doit()
print(sum)
print(expanded_sum)
x = pystencils.fields('x: float32[1d]')
assignments = pystencils.AssignmentCollection({
x.center(): sum
})
ast = pystencils.create_kernel(assignments, data_type=create_type('float32'))
code = str(pystencils.show_code(ast))
kernel = ast.compile()
print(code)
print(pystencils.show_code(ast))
assert 'float sum' in code
array = np.zeros((10,), np.float32)
kernel(x=array)
assert np.allclose(array, int(expanded_sum) * np.ones_like(array))
def test_product():
k = pystencils.TypedSymbol('k', create_type('int64'))
sum = sympy.Product(k, (k, 1, 10))
expanded_sum = sum.doit()
print(sum)
print(expanded_sum)
x = pystencils.fields('x: int64[1d]')
assignments = pystencils.AssignmentCollection({
x.center(): sum
})
ast = pystencils.create_kernel(assignments)
code = str(pystencils.show_code(ast))
kernel = ast.compile()
print(code)
assert 'int64_t product' in code
array = np.zeros((10,), np.int64)
kernel(x=array)
assert np.allclose(array, int(expanded_sum) * np.ones_like(array))
def test_prod_var_limit():
k = pystencils.TypedSymbol('k', create_type('int64'))
limit = pystencils.TypedSymbol('limit', create_type('int64'))
sum = sympy.Sum(k, (k, 1, limit))
expanded_sum = sum.replace(limit, 100).doit()
print(sum)
print(expanded_sum)
x = pystencils.fields('x: int64[1d]')
assignments = pystencils.AssignmentCollection({
x.center(): sum
})
ast = pystencils.create_kernel(assignments)
code = str(pystencils.show_code(ast))
kernel = ast.compile()
print(code)
array = np.zeros((10,), np.int64)
kernel(x=array, limit=100)
assert np.allclose(array, int(expanded_sum) * np.ones_like(array))
pystencils_tests/test_types.py deleted 100644 → 0
View file @ ef496986
from pystencils import data_types
from pystencils.data_types import *
import sympy as sp
def test_parsing():
assert str(data_types.create_composite_type_from_string("const double *")) == "double const *"
assert str(data_types.create_composite_type_from_string("double const *")) == "double const *"
t1 = data_types.create_composite_type_from_string("const double * const * const restrict")
t2 = data_types.create_composite_type_from_string(str(t1))
assert t1 == t2
def test_collation():
double_type = create_type("double")
float_type = create_type("float32")
double4_type = VectorType(double_type, 4)
float4_type = VectorType(float_type, 4)
assert collate_types([double_type, float_type]) == double_type
assert collate_types([double4_type, float_type]) == double4_type
assert collate_types([double4_type, float4_type]) == double4_type
def test_dtype_of_constants():
# Some come constants are neither of type Integer,Float,Rational and don't have args
# >>> isinstance(pi, Integer)
# False
# >>> isinstance(pi, Float)
# False
# >>> isinstance(pi, Rational)
# False
# >>> pi.args
# ()
get_type_of_expression(sp.pi)
pystencils_tests/test_vectorization.py deleted 100644 → 0
View file @ ef496986
import numpy as np
import sympy as sp
import pystencils as ps
from pystencils.backends.simd_instruction_sets import get_supported_instruction_sets
from pystencils.cpu.vectorization import vectorize
from pystencils.transformations import replace_inner_stride_with_one
def test_vector_type_propagation():
a, b, c, d, e = sp.symbols("a b c d e")
arr = np.ones((2 ** 2 + 2, 2 ** 3 + 2))
arr *= 10.0
f, g = ps.fields(f=arr, g=arr)
update_rule = [ps.Assignment(a, f[1, 0]),
ps.Assignment(b, a),
ps.Assignment(g[0, 0], b + 3 + f[0, 1])]
ast = ps.create_kernel(update_rule)
vectorize(ast)
func = ast.compile()
dst = np.zeros_like(arr)
func(g=dst, f=arr)
np.testing.assert_equal(dst[1:-1, 1:-1], 2 * 10.0 + 3)
def test_inplace_update():
shape = (9, 9, 3)
arr = np.ones(shape, order='f')
@ps.kernel
def update_rule(s):
f = ps.fields("f(3) : [2D]", f=arr)
s.tmp0 @= f(0)
s.tmp1 @= f(1)
s.tmp2 @= f(2)
f0, f1, f2 = f(0), f(1), f(2)
f0 @= 2 * s.tmp0
f1 @= 2 * s.tmp0
f2 @= 2 * s.tmp0
ast = ps.create_kernel(update_rule, cpu_vectorize_info={'instruction_set': 'avx'})
kernel = ast.compile()
kernel(f=arr)
np.testing.assert_equal(arr, 2)
def test_vectorization_fixed_size():
configurations = []
# Fixed size - multiple of four
arr = np.ones((20 + 2, 24 + 2)) * 5.0
f, g = ps.fields(f=arr, g=arr)
configurations.append((arr, f, g))
# Fixed size - no multiple of four
arr = np.ones((21 + 2, 25 + 2)) * 5.0
f, g = ps.fields(f=arr, g=arr)
configurations.append((arr, f, g))
# Fixed size - different remainder
arr = np.ones((23 + 2, 17 + 2)) * 5.0
f, g = ps.fields(f=arr, g=arr)
configurations.append((arr, f, g))
for arr, f, g in configurations:
update_rule = [ps.Assignment(g[0, 0], f[0, 0] + f[-1, 0] + f[1, 0] + f[0, 1] + f[0, -1] + 42.0)]
ast = ps.create_kernel(update_rule)
vectorize(ast)
func = ast.compile()
dst = np.zeros_like(arr)
func(g=dst, f=arr)
np.testing.assert_equal(dst[1:-1, 1:-1], 5 * 5.0 + 42.0)
def test_vectorization_variable_size():
f, g = ps.fields("f, g : double[2D]")
update_rule = [ps.Assignment(g[0, 0], f[0, 0] + f[-1, 0] + f[1, 0] + f[0, 1] + f[0, -1] + 42.0)]
ast = ps.create_kernel(update_rule)
replace_inner_stride_with_one(ast)
vectorize(ast)
func = ast.compile()
arr = np.ones((23 + 2, 17 + 2)) * 5.0
dst = np.zeros_like(arr)
func(g=dst, f=arr)
np.testing.assert_equal(dst[1:-1, 1:-1], 5 * 5.0 + 42.0)
def test_piecewise1():
a, b, c, d, e = sp.symbols("a b c d e")
arr = np.ones((2 ** 3 + 2, 2 ** 4 + 2)) * 5.0
f, g = ps.fields(f=arr, g=arr)
update_rule = [ps.Assignment(a, f[1, 0]),
ps.Assignment(b, a),
ps.Assignment(c, f[0, 0] > 0.0),
ps.Assignment(g[0, 0], sp.Piecewise((b + 3 + f[0, 1], c), (0.0, True)))]
ast = ps.create_kernel(update_rule)
vectorize(ast)
func = ast.compile()
dst = np.zeros_like(arr)
func(g=dst, f=arr)
np.testing.assert_equal(dst[1:-1, 1:-1], 5 + 3 + 5.0)
def test_piecewise2():
arr = np.zeros((20, 20))
@ps.kernel
def test_kernel(s):
f, g = ps.fields(f=arr, g=arr)
s.condition @= f[0, 0] > 1
s.result @= 0.0 if s.condition else 1.0
g[0, 0] @= s.result
ast = ps.create_kernel(test_kernel)
vectorize(ast)
func = ast.compile()
func(f=arr, g=arr)
np.testing.assert_equal(arr, np.ones_like(arr))
def test_piecewise3():
arr = np.zeros((22, 22))
@ps.kernel
def test_kernel(s):
f, g = ps.fields(f=arr, g=arr)
s.b @= f[0, 1]
g[0, 0] @= 1.0 / (s.b + s.k) if f[0, 0] > 0.0 else 1.0
ast = ps.create_kernel(test_kernel)
vectorize(ast)
ast.compile()
def test_logical_operators():
arr = np.zeros((22, 22))
@ps.kernel
def test_kernel(s):
f, g = ps.fields(f=arr, g=arr)
s.c @= sp.And(f[0, 1] < 0.0, f[1, 0] < 0.0)
g[0, 0] @= sp.Piecewise([1.0 / f[1, 0], s.c], [1.0, True])
ast = ps.create_kernel(test_kernel)
vectorize(ast)
ast.compile()
def test_hardware_query():
instruction_sets = get_supported_instruction_sets()
assert 'sse' in instruction_sets
pytest.ini
View file @ 5600b6b6
[pytest]
testpaths = src tests doc/notebooks
pythonpath = src
python_files = test_*.py *_test.py scenario_*.py
norecursedirs = *.egg-info .git .cache .ipynb_checkpoints htmlcov
addopts = --doctest-modules --durations=20 --cov-config pytest.ini
markers =
longrun: tests only run at night since they have large execution time
notebook: mark for notebooks
# these warnings all come from third party libraries.
filterwarnings =
ignore:an integer is required:DeprecationWarning
ignore:\s*load will be removed, use:PendingDeprecationWarning
ignore:the imp module is deprecated in favour of importlib:DeprecationWarning
ignore:.*is a deprecated alias for the builtin `bool`:DeprecationWarning
ignore:'contextfilter' is renamed to 'pass_context':DeprecationWarning
ignore:Using or importing the ABCs from 'collections' instead of from 'collections.abc':DeprecationWarning
ignore:Animation was deleted without rendering anything:UserWarning
[run]
branch = True
source = pystencils
pystencils_tests
source = src/pystencils
tests
omit = doc/*
pystencils_tests/*
tests/*
setup.py
quicktest.py
conftest.py
pystencils/jupytersetup.py
pystencils/cpu/msvc_detection.py
pystencils/sympy_gmpy_bug_workaround.py
pystencils/cache.py
pystencils/pacxx/benchmark.py
versioneer.py
src/pystencils/jupytersetup.py
src/pystencils/cpu/msvc_detection.py
src/pystencils/sympy_gmpy_bug_workaround.py
src/pystencils/cache.py
src/pystencils/pacxx/benchmark.py
src/pystencils/_version.py
venv/
[report]
exclude_lines =
......@@ -24,10 +42,12 @@ exclude_lines =
pragma: no cover
def __repr__
def _repr_html_
# Don't complain if tests don't hit defensive assertion code:
raise AssertionError
raise NotImplementedError
NotImplementedError()
#raise ValueError
# Don't complain if non-runnable code isn't run:
......@@ -36,7 +56,7 @@ exclude_lines =
if __name__ == .__main__.:
skip_covered = True
fail_under = 74
fail_under = 85
[html]
directory = coverage_report
quicktest.py 0 → 100644
View file @ 5600b6b6
#!/usr/bin/env python3
from contextlib import redirect_stdout
import io
from tests.test_quicktests import (
test_basic_kernel,
test_basic_blocking_staggered,
test_basic_vectorization,
)
quick_tests = [
test_basic_kernel,
test_basic_blocking_staggered,
test_basic_vectorization,
]
if __name__ == "__main__":
print("Running pystencils quicktests")
for qt in quick_tests:
print(f" -> {qt.__name__}")
with redirect_stdout(io.StringIO()):
qt()
release.sh
View file @ 5600b6b6
......@@ -8,6 +8,6 @@ read new_version
git tag -s release/${new_version}
git push origin master release/${new_version}
python setup.py sdist bdist_wheel
rm -rf dist
twine upload dist/*
\ No newline at end of file
python setup.py sdist
twine upload dist/*
setup.py
View file @ 5600b6b6
import distutils
import io
import os
import sys
from contextlib import redirect_stdout
from importlib import import_module
from setuptools import setup, __version__ as setuptools_version
from setuptools import find_packages, setup
if int(setuptools_version.split('.')[0]) < 61:
raise Exception(
"[ERROR] pystencils requires at least setuptools version 61 to install.\n"
"If this error occurs during an installation via pip, it is likely that there is a conflict between "
"versions of setuptools installed by pip and the system package manager. "
"In this case, it is recommended to install pystencils into a virtual environment instead."
)
if '--use-cython' in sys.argv:
USE_CYTHON = True
sys.argv.remove('--use-cython')
else:
USE_CYTHON = False
import versioneer
quick_tests = [
'test_datahandling.test_kernel',
'test_blocking_staggered.test_blocking_staggered',
'test_blocking_staggered.test_blocking_staggered',
'test_vectorization.test_vectorization_variable_size',
]
def get_cmdclass():
return versioneer.get_cmdclass()
class SimpleTestRunner(distutils.cmd.Command):
"""A custom command to run selected tests"""
description = 'run some quick tests'
user_options = []
@staticmethod
def _run_tests_in_module(test):
"""Short test runner function - to work also if py.test is not installed."""
test = 'pystencils_tests.' + test
mod, function_name = test.rsplit('.', 1)
if isinstance(mod, str):
mod = import_module(mod)
func = getattr(mod, function_name)
print(" -> %s in %s" % (function_name, mod.__name__))
with redirect_stdout(io.StringIO()):
func()
def initialize_options(self):
pass
def finalize_options(self):
pass
def run(self):
"""Run command."""
for test in quick_tests:
self._run_tests_in_module(test)
def readme():
with open('README.md') as f:
return f.read()
def cython_extensions(*extensions):
from distutils.extension import Extension
ext = '.pyx' if USE_CYTHON else '.c'
result = [Extension(e, [e.replace('.', '/') + ext]) for e in extensions]
if USE_CYTHON:
from Cython.Build import cythonize
result = cythonize(result, language_level=3)
return result
try:
sys.path.insert(0, os.path.abspath('doc'))
from version_from_git import version_number_from_git
version = version_number_from_git()
with open("RELEASE-VERSION", "w") as f:
f.write(version)
except ImportError:
version = open('RELEASE-VERSION', 'r').read()
setup(name='pystencils',
description='Speeding up stencil computations on CPUs and GPUs',
version=version,
long_description=readme(),
long_description_content_type="text/markdown",
author='Martin Bauer',
license='AGPLv3',
author_email='martin.bauer@fau.de',
url='https://i10git.cs.fau.de/pycodegen/pystencils/',
packages=['pystencils'] + ['pystencils.' + s for s in find_packages('pystencils')],
install_requires=['sympy>=1.1', 'numpy', 'appdirs', 'joblib'],
package_data={'pystencils': ['include/*.h',
'backends/cuda_known_functions.txt',
'backends/opencl1.1_known_functions.txt']},
ext_modules=cython_extensions("pystencils.boundaries.createindexlistcython"),
classifiers=[
'Development Status :: 4 - Beta',
'Framework :: Jupyter',
'Topic :: Software Development :: Code Generators',
'Topic :: Scientific/Engineering :: Physics',
'Intended Audience :: Developers',
'Intended Audience :: Science/Research',
'License :: OSI Approved :: GNU Affero General Public License v3 or later (AGPLv3+)',
],
project_urls={
"Bug Tracker": "https://i10git.cs.fau.de/pycodegen/pystencils/issues",
"Documentation": "http://pycodegen.pages.walberla.net/pystencils/",
"Source Code": "https://i10git.cs.fau.de/pycodegen/pystencils",
},
extras_require={
'gpu': ['pycuda'],
'opencl': ['pyopencl'],
'alltrafos': ['islpy', 'py-cpuinfo'],
'bench_db': ['blitzdb', 'pymongo', 'pandas'],
'interactive': ['matplotlib', 'ipy_table', 'imageio', 'jupyter', 'pyevtk'],
'autodiff': ['pystencils-autodiff'],
'doc': ['sphinx', 'sphinx_rtd_theme', 'nbsphinx',
'sphinxcontrib-bibtex', 'sphinx_autodoc_typehints', 'pandoc'],
},
tests_require=['pytest',
'pytest-cov',
'pytest-html',
'ansi2html',
'pytest-xdist',
'flake8',
'nbformat',
'nbconvert',
'ipython'],
python_requires=">=3.6",
cmdclass={
'quicktest': SimpleTestRunner
},
)
setup(
version=versioneer.get_version(),
cmdclass=get_cmdclass(),
)
pystencils/__init__.py → src/pystencils/__init__.py
View file @ 5600b6b6
"""Module to generate stencil kernels in C or CUDA using sympy expressions and call them as Python functions"""
from . import sympy_gmpy_bug_workaround # NOQA
from .enums import Backend, Target
from . import fd
from . import stencil as stencil
from .assignment import Assignment, assignment_from_stencil
from .data_types import TypedSymbol
from .datahandling import create_data_handling
from .display_utils import show_code, to_dot
from .assignment import Assignment, AddAugmentedAssignment, assignment_from_stencil
from .typing.typed_sympy import TypedSymbol
from .display_utils import get_code_obj, get_code_str, show_code, to_dot
from .field import Field, FieldType, fields
from .kernel_decorator import kernel
from .kernelcreation import create_indexed_kernel, create_kernel, create_staggered_kernel
from .config import CreateKernelConfig
from .cache import clear_cache
from .kernel_decorator import kernel, kernel_config
from .kernelcreation import create_kernel, create_staggered_kernel
from .simp import AssignmentCollection
from .slicing import make_slice
from .spatial_coordinates import x_, x_staggered, x_staggered_vector, x_vector, y_, y_staggered, z_, z_staggered
from .sympyextensions import SymbolCreator
try:
import pystencils_autodiff
autodiff = pystencils_autodiff
except ImportError:
pass
from .datahandling import create_data_handling
__all__ = ['Field', 'FieldType', 'fields',
'TypedSymbol',
'make_slice',
'create_kernel', 'create_indexed_kernel', 'create_staggered_kernel',
'show_code', 'to_dot',
'CreateKernelConfig',
'create_kernel', 'create_staggered_kernel',
'Target', 'Backend',
'show_code', 'to_dot', 'get_code_obj', 'get_code_str',
'AssignmentCollection',
'Assignment',
'Assignment', 'AddAugmentedAssignment',
'assignment_from_stencil',
'SymbolCreator',
'create_data_handling',
'kernel',
'clear_cache',
'kernel', 'kernel_config',
'x_', 'y_', 'z_',
'x_staggered', 'y_staggered', 'z_staggered',
'x_vector', 'x_staggered_vector',
'fd',
'stencil']
from . import _version
__version__ = _version.get_versions()['version']
src/pystencils/_version.py 0 → 100644
View file @ 5600b6b6
# This file helps to compute a version number in source trees obtained from
# git-archive tarball (such as those provided by githubs download-from-tag
# feature). Distribution tarballs (built by setup.py sdist) and build
# directories (produced by setup.py build) will contain a much shorter file
# that just contains the computed version number.
# This file is released into the public domain.
# Generated by versioneer-0.29
# https://github.com/python-versioneer/python-versioneer
"""Git implementation of _version.py."""
import errno
import os
import re
import subprocess
import sys
from typing import Any, Callable, Dict, List, Optional, Tuple
import functools
def get_keywords() -> Dict[str, str]:
"""Get the keywords needed to look up the version information."""
# these strings will be replaced by git during git-archive.
# setup.py/versioneer.py will grep for the variable names, so they must
# each be defined on a line of their own. _version.py will just call
# get_keywords().
git_refnames = "$Format:%d$"
git_full = "$Format:%H$"
git_date = "$Format:%ci$"
keywords = {"refnames": git_refnames, "full": git_full, "date": git_date}
return keywords
class VersioneerConfig:
"""Container for Versioneer configuration parameters."""
VCS: str
style: str
tag_prefix: str
parentdir_prefix: str
versionfile_source: str
verbose: bool
def get_config() -> VersioneerConfig:
"""Create, populate and return the VersioneerConfig() object."""
# these strings are filled in when 'setup.py versioneer' creates
# _version.py
cfg = VersioneerConfig()
cfg.VCS = "git"
cfg.style = "pep440"
cfg.tag_prefix = "release/"
cfg.parentdir_prefix = "pystencils-"
cfg.versionfile_source = "src/pystencils/_version.py"
cfg.verbose = False
return cfg
class NotThisMethod(Exception):
"""Exception raised if a method is not valid for the current scenario."""
LONG_VERSION_PY: Dict[str, str] = {}
HANDLERS: Dict[str, Dict[str, Callable]] = {}
def register_vcs_handler(vcs: str, method: str) -> Callable: # decorator
"""Create decorator to mark a method as the handler of a VCS."""
def decorate(f: Callable) -> Callable:
"""Store f in HANDLERS[vcs][method]."""
if vcs not in HANDLERS:
HANDLERS[vcs] = {}
HANDLERS[vcs][method] = f
return f
return decorate
def run_command(
commands: List[str],
args: List[str],
cwd: Optional[str] = None,
verbose: bool = False,
hide_stderr: bool = False,
env: Optional[Dict[str, str]] = None,
) -> Tuple[Optional[str], Optional[int]]:
"""Call the given command(s)."""
assert isinstance(commands, list)
process = None
popen_kwargs: Dict[str, Any] = {}
if sys.platform == "win32":
# This hides the console window if pythonw.exe is used
startupinfo = subprocess.STARTUPINFO()
startupinfo.dwFlags |= subprocess.STARTF_USESHOWWINDOW
popen_kwargs["startupinfo"] = startupinfo
for command in commands:
try:
dispcmd = str([command] + args)
# remember shell=False, so use git.cmd on windows, not just git
process = subprocess.Popen([command] + args, cwd=cwd, env=env,
stdout=subprocess.PIPE,
stderr=(subprocess.PIPE if hide_stderr
else None), **popen_kwargs)
break
except OSError as e:
if e.errno == errno.ENOENT:
continue
if verbose:
print("unable to run %s" % dispcmd)
print(e)
return None, None
else:
if verbose:
print("unable to find command, tried %s" % (commands,))
return None, None
stdout = process.communicate()[0].strip().decode()
if process.returncode != 0:
if verbose:
print("unable to run %s (error)" % dispcmd)
print("stdout was %s" % stdout)
return None, process.returncode
return stdout, process.returncode
def versions_from_parentdir(
parentdir_prefix: str,
root: str,
verbose: bool,
) -> Dict[str, Any]:
"""Try to determine the version from the parent directory name.
Source tarballs conventionally unpack into a directory that includes both
the project name and a version string. We will also support searching up
two directory levels for an appropriately named parent directory
"""
rootdirs = []
for _ in range(3):
dirname = os.path.basename(root)
if dirname.startswith(parentdir_prefix):
return {"version": dirname[len(parentdir_prefix):],
"full-revisionid": None,
"dirty": False, "error": None, "date": None}
rootdirs.append(root)
root = os.path.dirname(root) # up a level
if verbose:
print("Tried directories %s but none started with prefix %s" %
(str(rootdirs), parentdir_prefix))
raise NotThisMethod("rootdir doesn't start with parentdir_prefix")
@register_vcs_handler("git", "get_keywords")
def git_get_keywords(versionfile_abs: str) -> Dict[str, str]:
"""Extract version information from the given file."""
# the code embedded in _version.py can just fetch the value of these
# keywords. When used from setup.py, we don't want to import _version.py,
# so we do it with a regexp instead. This function is not used from
# _version.py.
keywords: Dict[str, str] = {}
try:
with open(versionfile_abs, "r") as fobj:
for line in fobj:
if line.strip().startswith("git_refnames ="):
mo = re.search(r'=\s*"(.*)"', line)
if mo:
keywords["refnames"] = mo.group(1)
if line.strip().startswith("git_full ="):
mo = re.search(r'=\s*"(.*)"', line)
if mo:
keywords["full"] = mo.group(1)
if line.strip().startswith("git_date ="):
mo = re.search(r'=\s*"(.*)"', line)
if mo:
keywords["date"] = mo.group(1)
except OSError:
pass
return keywords
@register_vcs_handler("git", "keywords")
def git_versions_from_keywords(
keywords: Dict[str, str],
tag_prefix: str,
verbose: bool,
) -> Dict[str, Any]:
"""Get version information from git keywords."""
if "refnames" not in keywords:
raise NotThisMethod("Short version file found")
date = keywords.get("date")
if date is not None:
# Use only the last line. Previous lines may contain GPG signature
# information.
date = date.splitlines()[-1]
# git-2.2.0 added "%cI", which expands to an ISO-8601 -compliant
# datestamp. However we prefer "%ci" (which expands to an "ISO-8601
# -like" string, which we must then edit to make compliant), because
# it's been around since git-1.5.3, and it's too difficult to
# discover which version we're using, or to work around using an
# older one.
date = date.strip().replace(" ", "T", 1).replace(" ", "", 1)
refnames = keywords["refnames"].strip()
if refnames.startswith("$Format"):
if verbose:
print("keywords are unexpanded, not using")
raise NotThisMethod("unexpanded keywords, not a git-archive tarball")
refs = {r.strip() for r in refnames.strip("()").split(",")}
# starting in git-1.8.3, tags are listed as "tag: foo-1.0" instead of
# just "foo-1.0". If we see a "tag: " prefix, prefer those.
TAG = "tag: "
tags = {r[len(TAG):] for r in refs if r.startswith(TAG)}
if not tags:
# Either we're using git < 1.8.3, or there really are no tags. We use
# a heuristic: assume all version tags have a digit. The old git %d
# expansion behaves like git log --decorate=short and strips out the
# refs/heads/ and refs/tags/ prefixes that would let us distinguish
# between branches and tags. By ignoring refnames without digits, we
# filter out many common branch names like "release" and
# "stabilization", as well as "HEAD" and "master".
tags = {r for r in refs if re.search(r'\d', r)}
if verbose:
print("discarding '%s', no digits" % ",".join(refs - tags))
if verbose:
print("likely tags: %s" % ",".join(sorted(tags)))
for ref in sorted(tags):
# sorting will prefer e.g. "2.0" over "2.0rc1"
if ref.startswith(tag_prefix):
r = ref[len(tag_prefix):]
# Filter out refs that exactly match prefix or that don't start
# with a number once the prefix is stripped (mostly a concern
# when prefix is '')
if not re.match(r'\d', r):
continue
if verbose:
print("picking %s" % r)
return {"version": r,
"full-revisionid": keywords["full"].strip(),
"dirty": False, "error": None,
"date": date}
# no suitable tags, so version is "0+unknown", but full hex is still there
if verbose:
print("no suitable tags, using unknown + full revision id")
return {"version": "0+unknown",
"full-revisionid": keywords["full"].strip(),
"dirty": False, "error": "no suitable tags", "date": None}
@register_vcs_handler("git", "pieces_from_vcs")
def git_pieces_from_vcs(
tag_prefix: str,
root: str,
verbose: bool,
runner: Callable = run_command
) -> Dict[str, Any]:
"""Get version from 'git describe' in the root of the source tree.
This only gets called if the git-archive 'subst' keywords were *not*
expanded, and _version.py hasn't already been rewritten with a short
version string, meaning we're inside a checked out source tree.
"""
GITS = ["git"]
if sys.platform == "win32":
GITS = ["git.cmd", "git.exe"]
# GIT_DIR can interfere with correct operation of Versioneer.
# It may be intended to be passed to the Versioneer-versioned project,
# but that should not change where we get our version from.
env = os.environ.copy()
env.pop("GIT_DIR", None)
runner = functools.partial(runner, env=env)
_, rc = runner(GITS, ["rev-parse", "--git-dir"], cwd=root,
hide_stderr=not verbose)
if rc != 0:
if verbose:
print("Directory %s not under git control" % root)
raise NotThisMethod("'git rev-parse --git-dir' returned error")
# if there is a tag matching tag_prefix, this yields TAG-NUM-gHEX[-dirty]
# if there isn't one, this yields HEX[-dirty] (no NUM)
describe_out, rc = runner(GITS, [
"describe", "--tags", "--dirty", "--always", "--long",
"--match", f"{tag_prefix}[[:digit:]]*"
], cwd=root)
# --long was added in git-1.5.5
if describe_out is None:
raise NotThisMethod("'git describe' failed")
describe_out = describe_out.strip()
full_out, rc = runner(GITS, ["rev-parse", "HEAD"], cwd=root)
if full_out is None:
raise NotThisMethod("'git rev-parse' failed")
full_out = full_out.strip()
pieces: Dict[str, Any] = {}
pieces["long"] = full_out
pieces["short"] = full_out[:7] # maybe improved later
pieces["error"] = None
branch_name, rc = runner(GITS, ["rev-parse", "--abbrev-ref", "HEAD"],
cwd=root)
# --abbrev-ref was added in git-1.6.3
if rc != 0 or branch_name is None:
raise NotThisMethod("'git rev-parse --abbrev-ref' returned error")
branch_name = branch_name.strip()
if branch_name == "HEAD":
# If we aren't exactly on a branch, pick a branch which represents
# the current commit. If all else fails, we are on a branchless
# commit.
branches, rc = runner(GITS, ["branch", "--contains"], cwd=root)
# --contains was added in git-1.5.4
if rc != 0 or branches is None:
raise NotThisMethod("'git branch --contains' returned error")
branches = branches.split("\n")
# Remove the first line if we're running detached
if "(" in branches[0]:
branches.pop(0)
# Strip off the leading "* " from the list of branches.
branches = [branch[2:] for branch in branches]
if "master" in branches:
branch_name = "master"
elif not branches:
branch_name = None
else:
# Pick the first branch that is returned. Good or bad.
branch_name = branches[0]
pieces["branch"] = branch_name
# parse describe_out. It will be like TAG-NUM-gHEX[-dirty] or HEX[-dirty]
# TAG might have hyphens.
git_describe = describe_out
# look for -dirty suffix
dirty = git_describe.endswith("-dirty")
pieces["dirty"] = dirty
if dirty:
git_describe = git_describe[:git_describe.rindex("-dirty")]
# now we have TAG-NUM-gHEX or HEX
if "-" in git_describe:
# TAG-NUM-gHEX
mo = re.search(r'^(.+)-(\d+)-g([0-9a-f]+)$', git_describe)
if not mo:
# unparsable. Maybe git-describe is misbehaving?
pieces["error"] = ("unable to parse git-describe output: '%s'"
% describe_out)
return pieces
# tag
full_tag = mo.group(1)
if not full_tag.startswith(tag_prefix):
if verbose:
fmt = "tag '%s' doesn't start with prefix '%s'"
print(fmt % (full_tag, tag_prefix))
pieces["error"] = ("tag '%s' doesn't start with prefix '%s'"
% (full_tag, tag_prefix))
return pieces
pieces["closest-tag"] = full_tag[len(tag_prefix):]
# distance: number of commits since tag
pieces["distance"] = int(mo.group(2))
# commit: short hex revision ID
pieces["short"] = mo.group(3)
else:
# HEX: no tags
pieces["closest-tag"] = None
out, rc = runner(GITS, ["rev-list", "HEAD", "--left-right"], cwd=root)
pieces["distance"] = len(out.split()) # total number of commits
# commit date: see ISO-8601 comment in git_versions_from_keywords()
date = runner(GITS, ["show", "-s", "--format=%ci", "HEAD"], cwd=root)[0].strip()
# Use only the last line. Previous lines may contain GPG signature
# information.
date = date.splitlines()[-1]
pieces["date"] = date.strip().replace(" ", "T", 1).replace(" ", "", 1)
return pieces
def plus_or_dot(pieces: Dict[str, Any]) -> str:
"""Return a + if we don't already have one, else return a ."""
if "+" in pieces.get("closest-tag", ""):
return "."
return "+"
def render_pep440(pieces: Dict[str, Any]) -> str:
"""Build up version string, with post-release "local version identifier".
Our goal: TAG[+DISTANCE.gHEX[.dirty]] . Note that if you
get a tagged build and then dirty it, you'll get TAG+0.gHEX.dirty
Exceptions:
1: no tags. git_describe was just HEX. 0+untagged.DISTANCE.gHEX[.dirty]
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"] or pieces["dirty"]:
rendered += plus_or_dot(pieces)
rendered += "%d.g%s" % (pieces["distance"], pieces["short"])
if pieces["dirty"]:
rendered += ".dirty"
else:
# exception #1
rendered = "0+untagged.%d.g%s" % (pieces["distance"],
pieces["short"])
if pieces["dirty"]:
rendered += ".dirty"
return rendered
def render_pep440_branch(pieces: Dict[str, Any]) -> str:
"""TAG[[.dev0]+DISTANCE.gHEX[.dirty]] .
The ".dev0" means not master branch. Note that .dev0 sorts backwards
(a feature branch will appear "older" than the master branch).
Exceptions:
1: no tags. 0[.dev0]+untagged.DISTANCE.gHEX[.dirty]
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"] or pieces["dirty"]:
if pieces["branch"] != "master":
rendered += ".dev0"
rendered += plus_or_dot(pieces)
rendered += "%d.g%s" % (pieces["distance"], pieces["short"])
if pieces["dirty"]:
rendered += ".dirty"
else:
# exception #1
rendered = "0"
if pieces["branch"] != "master":
rendered += ".dev0"
rendered += "+untagged.%d.g%s" % (pieces["distance"],
pieces["short"])
if pieces["dirty"]:
rendered += ".dirty"
return rendered
def pep440_split_post(ver: str) -> Tuple[str, Optional[int]]:
"""Split pep440 version string at the post-release segment.
Returns the release segments before the post-release and the
post-release version number (or -1 if no post-release segment is present).
"""
vc = str.split(ver, ".post")
return vc[0], int(vc[1] or 0) if len(vc) == 2 else None
def render_pep440_pre(pieces: Dict[str, Any]) -> str:
"""TAG[.postN.devDISTANCE] -- No -dirty.
Exceptions:
1: no tags. 0.post0.devDISTANCE
"""
if pieces["closest-tag"]:
if pieces["distance"]:
# update the post release segment
tag_version, post_version = pep440_split_post(pieces["closest-tag"])
rendered = tag_version
if post_version is not None:
rendered += ".post%d.dev%d" % (post_version + 1, pieces["distance"])
else:
rendered += ".post0.dev%d" % (pieces["distance"])
else:
# no commits, use the tag as the version
rendered = pieces["closest-tag"]
else:
# exception #1
rendered = "0.post0.dev%d" % pieces["distance"]
return rendered
def render_pep440_post(pieces: Dict[str, Any]) -> str:
"""TAG[.postDISTANCE[.dev0]+gHEX] .
The ".dev0" means dirty. Note that .dev0 sorts backwards
(a dirty tree will appear "older" than the corresponding clean one),
but you shouldn't be releasing software with -dirty anyways.
Exceptions:
1: no tags. 0.postDISTANCE[.dev0]
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"] or pieces["dirty"]:
rendered += ".post%d" % pieces["distance"]
if pieces["dirty"]:
rendered += ".dev0"
rendered += plus_or_dot(pieces)
rendered += "g%s" % pieces["short"]
else:
# exception #1
rendered = "0.post%d" % pieces["distance"]
if pieces["dirty"]:
rendered += ".dev0"
rendered += "+g%s" % pieces["short"]
return rendered
def render_pep440_post_branch(pieces: Dict[str, Any]) -> str:
"""TAG[.postDISTANCE[.dev0]+gHEX[.dirty]] .
The ".dev0" means not master branch.
Exceptions:
1: no tags. 0.postDISTANCE[.dev0]+gHEX[.dirty]
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"] or pieces["dirty"]:
rendered += ".post%d" % pieces["distance"]
if pieces["branch"] != "master":
rendered += ".dev0"
rendered += plus_or_dot(pieces)
rendered += "g%s" % pieces["short"]
if pieces["dirty"]:
rendered += ".dirty"
else:
# exception #1
rendered = "0.post%d" % pieces["distance"]
if pieces["branch"] != "master":
rendered += ".dev0"
rendered += "+g%s" % pieces["short"]
if pieces["dirty"]:
rendered += ".dirty"
return rendered
def render_pep440_old(pieces: Dict[str, Any]) -> str:
"""TAG[.postDISTANCE[.dev0]] .
The ".dev0" means dirty.
Exceptions:
1: no tags. 0.postDISTANCE[.dev0]
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"] or pieces["dirty"]:
rendered += ".post%d" % pieces["distance"]
if pieces["dirty"]:
rendered += ".dev0"
else:
# exception #1
rendered = "0.post%d" % pieces["distance"]
if pieces["dirty"]:
rendered += ".dev0"
return rendered
def render_git_describe(pieces: Dict[str, Any]) -> str:
"""TAG[-DISTANCE-gHEX][-dirty].
Like 'git describe --tags --dirty --always'.
Exceptions:
1: no tags. HEX[-dirty] (note: no 'g' prefix)
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
if pieces["distance"]:
rendered += "-%d-g%s" % (pieces["distance"], pieces["short"])
else:
# exception #1
rendered = pieces["short"]
if pieces["dirty"]:
rendered += "-dirty"
return rendered
def render_git_describe_long(pieces: Dict[str, Any]) -> str:
"""TAG-DISTANCE-gHEX[-dirty].
Like 'git describe --tags --dirty --always -long'.
The distance/hash is unconditional.
Exceptions:
1: no tags. HEX[-dirty] (note: no 'g' prefix)
"""
if pieces["closest-tag"]:
rendered = pieces["closest-tag"]
rendered += "-%d-g%s" % (pieces["distance"], pieces["short"])
else:
# exception #1
rendered = pieces["short"]
if pieces["dirty"]:
rendered += "-dirty"
return rendered
def render(pieces: Dict[str, Any], style: str) -> Dict[str, Any]:
"""Render the given version pieces into the requested style."""
if pieces["error"]:
return {"version": "unknown",
"full-revisionid": pieces.get("long"),
"dirty": None,
"error": pieces["error"],
"date": None}
if not style or style == "default":
style = "pep440" # the default
if style == "pep440":
rendered = render_pep440(pieces)
elif style == "pep440-branch":
rendered = render_pep440_branch(pieces)
elif style == "pep440-pre":
rendered = render_pep440_pre(pieces)
elif style == "pep440-post":
rendered = render_pep440_post(pieces)
elif style == "pep440-post-branch":
rendered = render_pep440_post_branch(pieces)
elif style == "pep440-old":
rendered = render_pep440_old(pieces)
elif style == "git-describe":
rendered = render_git_describe(pieces)
elif style == "git-describe-long":
rendered = render_git_describe_long(pieces)
else:
raise ValueError("unknown style '%s'" % style)
return {"version": rendered, "full-revisionid": pieces["long"],
"dirty": pieces["dirty"], "error": None,
"date": pieces.get("date")}
def get_versions() -> Dict[str, Any]:
"""Get version information or return default if unable to do so."""
# I am in _version.py, which lives at ROOT/VERSIONFILE_SOURCE. If we have
# __file__, we can work backwards from there to the root. Some
# py2exe/bbfreeze/non-CPython implementations don't do __file__, in which
# case we can only use expanded keywords.
cfg = get_config()
verbose = cfg.verbose
try:
return git_versions_from_keywords(get_keywords(), cfg.tag_prefix,
verbose)
except NotThisMethod:
pass
try:
root = os.path.realpath(__file__)
# versionfile_source is the relative path from the top of the source
# tree (where the .git directory might live) to this file. Invert
# this to find the root from __file__.
for _ in cfg.versionfile_source.split('/'):
root = os.path.dirname(root)
except NameError:
return {"version": "0+unknown", "full-revisionid": None,
"dirty": None,
"error": "unable to find root of source tree",
"date": None}
try:
pieces = git_pieces_from_vcs(cfg.tag_prefix, root, verbose)
return render(pieces, cfg.style)
except NotThisMethod:
pass
try:
if cfg.parentdir_prefix:
return versions_from_parentdir(cfg.parentdir_prefix, root, verbose)
except NotThisMethod:
pass
return {"version": "0+unknown", "full-revisionid": None,
"dirty": None,
"error": "unable to compute version", "date": None}
pystencils/alignedarray.py → src/pystencils/alignedarray.py
View file @ 5600b6b6
import numpy as np
def aligned_empty(shape, byte_alignment=32, dtype=np.float64, byte_offset=0, order='C', align_inner_coordinate=True):
def aligned_empty(shape, byte_alignment=True, dtype=np.float64, byte_offset=0, order='C', align_inner_coordinate=True):
"""
Creates an aligned empty numpy array
Args:
shape: size of the array
byte_alignment: alignment in bytes, for the start address of the array holds (a % byte_alignment) == 0
By default, use the maximum required by the CPU (or 512 bits if this cannot be detected).
When 'cacheline' is specified, the size of a cache line is used.
dtype: numpy data type
byte_offset: offset in bytes for position that should be aligned i.e. (a+byte_offset) % byte_alignment == 0
typically used to align first inner cell instead of ghost layer
order: storage linearization order
align_inner_coordinate: if True, the start of the innermost coordinate lines are aligned as well
"""
if byte_alignment is True or byte_alignment == 'cacheline':
from pystencils.backends.simd_instruction_sets import (get_supported_instruction_sets, get_cacheline_size,
get_vector_instruction_set)
instruction_sets = get_supported_instruction_sets()
if instruction_sets is None:
byte_alignment = 64
elif byte_alignment == 'cacheline':
cacheline_sizes = [get_cacheline_size(is_name) for is_name in instruction_sets]
if all([s is None for s in cacheline_sizes]) or \
max([s for s in cacheline_sizes if s is not None]) > 0x100000:
widths = [get_vector_instruction_set(dtype, is_name)['width'] * np.dtype(dtype).itemsize
for is_name in instruction_sets
if type(get_vector_instruction_set(dtype, is_name)['width']) is int]
byte_alignment = 64 if all([s is None for s in widths]) else max(widths)
else:
byte_alignment = max([s for s in cacheline_sizes if s is not None])
elif not any([type(get_vector_instruction_set(dtype, is_name)['width']) is int
for is_name in instruction_sets]):
byte_alignment = 64
else:
byte_alignment = max([get_vector_instruction_set(dtype, is_name)['width'] * np.dtype(dtype).itemsize
for is_name in instruction_sets
if type(get_vector_instruction_set(dtype, is_name)['width']) is int])
if (not align_inner_coordinate) or (not hasattr(shape, '__len__')):
size = np.prod(shape)
d = np.dtype(dtype)
......@@ -51,7 +77,7 @@ def aligned_empty(shape, byte_alignment=32, dtype=np.float64, byte_offset=0, ord
return tmp
def aligned_zeros(shape, byte_alignment=16, dtype=float, byte_offset=0, order='C', align_inner_coordinate=True):
def aligned_zeros(shape, byte_alignment=True, dtype=np.float64, byte_offset=0, order='C', align_inner_coordinate=True):
arr = aligned_empty(shape, dtype=dtype, byte_offset=byte_offset,
order=order, byte_alignment=byte_alignment, align_inner_coordinate=align_inner_coordinate)
x = np.zeros((), arr.dtype)
......@@ -59,7 +85,7 @@ def aligned_zeros(shape, byte_alignment=16, dtype=float, byte_offset=0, order='C
return arr
def aligned_ones(shape, byte_alignment=16, dtype=float, byte_offset=0, order='C', align_inner_coordinate=True):
def aligned_ones(shape, byte_alignment=True, dtype=np.float64, byte_offset=0, order='C', align_inner_coordinate=True):
arr = aligned_empty(shape, dtype=dtype, byte_offset=byte_offset,
order=order, byte_alignment=byte_alignment, align_inner_coordinate=align_inner_coordinate)
x = np.ones((), arr.dtype)
......
pystencils/assignment.py → src/pystencils/assignment.py
View file @ 5600b6b6
# -*- coding: utf-8 -*-
import numpy as np
import sympy as sp
from sympy.codegen.ast import Assignment, AugmentedAssignment, AddAugmentedAssignment
from sympy.printing.latex import LatexPrinter
try:
from sympy.codegen.ast import Assignment
except ImportError:
Assignment = None
__all__ = ['Assignment', 'assignment_from_stencil']
__all__ = ['Assignment', 'AugmentedAssignment', 'AddAugmentedAssignment', 'assignment_from_stencil']
def print_assignment_latex(printer, expr):
binop = f"{expr.binop}=" if isinstance(expr, AugmentedAssignment) else ''
"""sympy cannot print Assignments as Latex. Thus, this function is added to the sympy Latex printer"""
printed_lhs = printer.doprint(expr.lhs)
printed_rhs = printer.doprint(expr.rhs)
return r"{printed_lhs} \leftarrow {printed_rhs}".format(printed_lhs=printed_lhs, printed_rhs=printed_rhs)
return fr"{printed_lhs} \leftarrow_{{{binop}}} {printed_rhs}"
def assignment_str(assignment):
return r"{lhs} ← {rhs}".format(lhs=assignment.lhs, rhs=assignment.rhs)
if Assignment:
op = f"{assignment.binop}=" if isinstance(assignment, AugmentedAssignment) else '←'
return fr"{assignment.lhs} {op} {assignment.rhs}"
Assignment.__str__ = assignment_str
LatexPrinter._print_Assignment = print_assignment_latex
else:
# back port for older sympy versions that don't have Assignment yet
_old_new = sp.codegen.ast.Assignment.__new__
class Assignment(sp.Rel): # pragma: no cover
rel_op = ':='
__slots__ = []
# TODO Typing Part2 add default type, defult_float_type, default_int_type and use sane defaults
def _Assignment__new__(cls, lhs, rhs, *args, **kwargs):
if isinstance(lhs, (list, tuple, sp.Matrix)) and isinstance(rhs, (list, tuple, sp.Matrix)):
assert len(lhs) == len(rhs), f'{lhs} and {rhs} must have same length when performing vector assignment!'
return tuple(_old_new(cls, a, b, *args, **kwargs) for a, b in zip(lhs, rhs))
return _old_new(cls, lhs, rhs, *args, **kwargs)
def __new__(cls, lhs, rhs=0, **assumptions):
from sympy.matrices.expressions.matexpr import (
MatrixElement, MatrixSymbol)
lhs = sp.sympify(lhs)
rhs = sp.sympify(rhs)
# Tuple of things that can be on the lhs of an assignment
assignable = (sp.Symbol, MatrixSymbol, MatrixElement, sp.Indexed)
if not isinstance(lhs, assignable):
raise TypeError("Cannot assign to lhs of type %s." % type(lhs))
return sp.Rel.__new__(cls, lhs, rhs, **assumptions)
__str__ = assignment_str
_print_Assignment = print_assignment_latex
Assignment.__str__ = assignment_str
Assignment.__new__ = _Assignment__new__
LatexPrinter._print_Assignment = print_assignment_latex
# Apparently, in SymPy 1.4 Assignment.__hash__ is not implemented. This has been fixed in current master
try:
sympy_version = sp.__version__.split('.')
AugmentedAssignment.__str__ = assignment_str
LatexPrinter._print_AugmentedAssignment = print_assignment_latex
if int(sympy_version[0]) <= 1 and int(sympy_version[1]) <= 4:
def hash_fun(self):
return hash((self.lhs, self.rhs))
Assignment.__hash__ = hash_fun
except Exception:
pass
sp.MutableDenseMatrix.__hash__ = lambda self: hash(tuple(self))
def assignment_from_stencil(stencil_array, input_field, output_field,
......@@ -84,16 +63,19 @@ def assignment_from_stencil(stencil_array, input_field, output_field,
... [0, 6, 0]]
By default 'visual ordering is used - i.e. the stencil is applied as the nested lists are written down
>>> assignment_from_stencil(stencil, f, g, order='visual')
Assignment(g_C, 3*f_W + 6*f_S + 4*f_C + 2*f_N + 5*f_E)
>>> expected_output = Assignment(g[0, 0], 3*f[-1, 0] + 6*f[0, -1] + 4*f[0, 0] + 2*f[0, 1] + 5*f[1, 0])
>>> assignment_from_stencil(stencil, f, g, order='visual') == expected_output
True
'numpy' ordering uses the first coordinate of the stencil array for x offset, second for y offset etc.
>>> assignment_from_stencil(stencil, f, g, order='numpy')
Assignment(g_C, 2*f_W + 3*f_S + 4*f_C + 5*f_N + 6*f_E)
>>> expected_output = Assignment(g[0, 0], 2*f[-1, 0] + 3*f[0, -1] + 4*f[0, 0] + 5*f[0, 1] + 6*f[1, 0])
>>> assignment_from_stencil(stencil, f, g, order='numpy') == expected_output
True
You can also pass field accesses to apply the stencil at an already shifted position:
>>> assignment_from_stencil(stencil, f[1, 0], g[2, 0])
Assignment(g_2E, 3*f_C + 6*f_SE + 4*f_E + 2*f_NE + 5*f_2E)
>>> expected_output = Assignment(g[2, 0], 3*f[0, 0] + 6*f[1, -1] + 4*f[1, 0] + 2*f[1, 1] + 5*f[2, 0])
>>> assignment_from_stencil(stencil, f[1, 0], g[2, 0]) == expected_output
True
"""
from pystencils.field import Field
......
pystencils/astnodes.py → src/pystencils/astnodes.py
View file @ 5600b6b6
import collections.abc
import itertools
import uuid
from typing import Any, List, Optional, Sequence, Set, Union
import sympy as sp
from pystencils.data_types import TypedSymbol, cast_func, create_type
from pystencils.assignment import Assignment
from pystencils.enums import Target, Backend
from pystencils.field import Field
from pystencils.kernelparameters import FieldPointerSymbol, FieldShapeSymbol, FieldStrideSymbol
from pystencils.sympyextensions import fast_subs
from pystencils.typing import (create_type, get_next_parent_of_type,
FieldPointerSymbol, FieldShapeSymbol, FieldStrideSymbol, TypedSymbol, CFunction)
NodeOrExpr = Union['Node', sp.Expr]
......@@ -33,9 +37,13 @@ class Node:
raise NotImplementedError()
def subs(self, subs_dict) -> None:
"""Inplace! substitute, similar to sympy's but modifies the AST inplace."""
for a in self.args:
a.subs(subs_dict)
"""Inplace! Substitute, similar to sympy's but modifies the AST inplace."""
for i, a in enumerate(self.args):
result = a.subs(subs_dict)
if isinstance(a, sp.Expr): # sympy expressions' subs is out-of-place
self.args[i] = result
else: # all other should be in-place
assert result is None
@property
def func(self):
......@@ -102,14 +110,22 @@ class Conditional(Node):
result = self.true_block.undefined_symbols
if self.false_block:
result.update(self.false_block.undefined_symbols)
result.update(self.condition_expr.atoms(sp.Symbol))
if hasattr(self.condition_expr, 'atoms'):
result.update(self.condition_expr.atoms(sp.Symbol))
return result
def __str__(self):
return 'if:({!s}) '.format(self.condition_expr)
return self.__repr__()
def __repr__(self):
return 'if:({!r}) '.format(self.condition_expr)
result = f'if:({self.condition_expr!r}) '
if self.true_block:
result += f'\n\t{self.true_block}) '
if self.false_block:
result = 'else: '
result += f'\n\t{self.false_block} '
return result
def replace_by_true_block(self):
"""Replaces the conditional by its True block"""
......@@ -121,7 +137,6 @@ class Conditional(Node):
class KernelFunction(Node):
class Parameter:
"""Function parameter.
......@@ -161,7 +176,9 @@ class KernelFunction(Node):
def field_name(self):
return self.fields[0].name
def __init__(self, body, target, backend, compile_function, ghost_layers, function_name="kernel"):
def __init__(self, body, target: Target, backend: Backend, compile_function, ghost_layers,
function_name: str = "kernel",
assignments=None):
super(KernelFunction, self).__init__()
self._body = body
body.parent = self
......@@ -175,15 +192,20 @@ class KernelFunction(Node):
self.instruction_set = None # used in `vectorize` function to tell the backend which i.s. (SSE,AVX) to use
# function that compiles the node to a Python callable, is set by the backends
self._compile_function = compile_function
self.assignments = assignments
# If nontemporal stores are activated together with the Neon instruction set it results in cacheline zeroing
# For cacheline zeroing the information of the field size for each field is needed. Thus, in this case
# all field sizes are kernel parameters and not just the common field size used for the loops
self.use_all_written_field_sizes = False
@property
def target(self):
"""Currently either 'cpu' or 'gpu' """
"""See pystencils.Target"""
return self._target
@property
def backend(self):
"""Backend for generating the code e.g. 'llvm', 'c', 'cuda' """
"""Backend for generating the code: `Backend`"""
return self._backend
@property
......@@ -208,13 +230,21 @@ class KernelFunction(Node):
return self._body,
@property
def fields_accessed(self) -> Set['ResolvedFieldAccess']:
def fields_accessed(self) -> Set[Field]:
"""Set of Field instances: fields which are accessed inside this kernel function"""
return set(o.field for o in self.atoms(ResolvedFieldAccess))
return set(o.field for o in itertools.chain(self.atoms(ResolvedFieldAccess)))
def fields_written(self):
assigments = self.atoms(SympyAssignment)
return {a.lhs.field for a in assigments if isinstance(a.lhs, ResolvedFieldAccess)}
@property
def fields_written(self) -> Set[Field]:
assignments = self.atoms(SympyAssignment)
return set().union(itertools.chain.from_iterable([f.field for f in a.lhs.free_symbols if hasattr(f, 'field')]
for a in assignments))
@property
def fields_read(self) -> Set[Field]:
assignments = self.atoms(SympyAssignment)
return set().union(itertools.chain.from_iterable([f.field for f in a.rhs.free_symbols if hasattr(f, 'field')]
for a in assignments))
def get_parameters(self) -> Sequence['KernelFunction.Parameter']:
"""Returns list of parameters for this function.
......@@ -222,6 +252,11 @@ class KernelFunction(Node):
This function is expensive, cache the result where possible!
"""
field_map = {f.name: f for f in self.fields_accessed}
sizes = set()
if self.use_all_written_field_sizes:
sizes = set().union(*(a.shape[:a.spatial_dimensions] for a in self.fields_written))
sizes = filter(lambda s: isinstance(s, FieldShapeSymbol), sizes)
def get_fields(symbol):
if hasattr(symbol, 'field_name'):
......@@ -231,9 +266,13 @@ class KernelFunction(Node):
return ()
argument_symbols = self._body.undefined_symbols - self.global_variables
argument_symbols.update(sizes)
parameters = [self.Parameter(symbol, get_fields(symbol)) for symbol in argument_symbols]
if hasattr(self, 'indexing'):
parameters += [self.Parameter(s, []) for s in self.indexing.symbolic_parameters()]
# Exclude paramters of type CFunction. These parameters will result in a C function call that will be handled
# by including a respective header file in the compute kernel. Hence, it is not a free parameter.
parameters = [p for p in parameters if not isinstance(p.symbol, CFunction)]
parameters.sort(key=lambda p: p.symbol.name)
return parameters
......@@ -244,7 +283,7 @@ class KernelFunction(Node):
def __repr__(self):
params = [p.symbol for p in self.get_parameters()]
return '{0} {1}({2})'.format(type(self).__name__, self.function_name, params)
return f'{type(self).__name__} {self.function_name}({params})'
def compile(self, *args, **kwargs):
if self._compile_function is None:
......@@ -267,12 +306,17 @@ class SkipIteration(Node):
class Block(Node):
def __init__(self, nodes: List[Node]):
def __init__(self, nodes: Union[Node, List[Node]]):
super(Block, self).__init__()
if not isinstance(nodes, list):
nodes = [nodes]
self._nodes = nodes
self.parent = None
for n in self._nodes:
n.parent = self
try:
n.parent = self
except AttributeError:
pass
@property
def args(self):
......@@ -282,23 +326,39 @@ class Block(Node):
for a in self.args:
a.subs(subs_dict)
def insert_front(self, node):
node.parent = self
self._nodes.insert(0, node)
def fast_subs(self, subs_dict, skip=None):
self._nodes = [fast_subs(a, subs_dict, skip) for a in self._nodes]
return self
def insert_front(self, node, if_not_exists=False):
if if_not_exists and len(self._nodes) > 0 and self._nodes[0] == node:
return
if isinstance(node, collections.abc.Iterable):
node = list(node)
for n in node:
n.parent = self
self._nodes = node + self._nodes
else:
node.parent = self
self._nodes.insert(0, node)
def insert_before(self, new_node, insert_before):
def insert_before(self, new_node, insert_before, if_not_exists=False):
new_node.parent = self
assert self._nodes.count(insert_before) == 1
idx = self._nodes.index(insert_before)
# move all assignment (definitions to the top)
if isinstance(new_node, SympyAssignment) and new_node.is_declaration:
while idx > 0:
pn = self._nodes[idx - 1]
if isinstance(pn, LoopOverCoordinate) or isinstance(pn, Conditional):
idx -= 1
else:
break
self._nodes.insert(idx, new_node)
if not if_not_exists or self._nodes[idx] != new_node:
self._nodes.insert(idx, new_node)
def insert_after(self, new_node, insert_after, if_not_exists=False):
new_node.parent = self
assert self._nodes.count(insert_after) == 1
idx = self._nodes.index(insert_after) + 1
if not if_not_exists or not (self._nodes[idx - 1] == new_node
or (idx < len(self._nodes) and self._nodes[idx] == new_node)):
self._nodes.insert(idx, new_node)
def append(self, node):
if isinstance(node, list) or isinstance(node, tuple):
......@@ -315,6 +375,7 @@ class Block(Node):
return tmp
def replace(self, child, replacements):
assert self._nodes.count(child) == 1
idx = self._nodes.index(child)
del self._nodes[idx]
if type(replacements) is list:
......@@ -329,7 +390,10 @@ class Block(Node):
def symbols_defined(self):
result = set()
for a in self.args:
result.update(a.symbols_defined)
if isinstance(a, Assignment):
result.update(a.free_symbols)
else:
result.update(a.symbols_defined)
return result
@property
......@@ -337,8 +401,12 @@ class Block(Node):
result = set()
defined_symbols = set()
for a in self.args:
result.update(a.undefined_symbols)
defined_symbols.update(a.symbols_defined)
if isinstance(a, Assignment):
result.update(a.free_symbols)
defined_symbols.update({a.lhs})
else:
result.update(a.undefined_symbols)
defined_symbols.update(a.symbols_defined)
return result - defined_symbols
def __str__(self):
......@@ -361,9 +429,9 @@ class PragmaBlock(Block):
class LoopOverCoordinate(Node):
LOOP_COUNTER_NAME_PREFIX = "ctr"
BlOCK_LOOP_COUNTER_NAME_PREFIX = "_blockctr"
BLOCK_LOOP_COUNTER_NAME_PREFIX = "_blockctr"
def __init__(self, body, coordinate_to_loop_over, start, stop, step=1, is_block_loop=False):
def __init__(self, body, coordinate_to_loop_over, start, stop, step=1, is_block_loop=False, custom_loop_ctr=None):
super(LoopOverCoordinate, self).__init__(parent=None)
self.body = body
body.parent = self
......@@ -374,11 +442,12 @@ class LoopOverCoordinate(Node):
self.body.parent = self
self.prefix_lines = []
self.is_block_loop = is_block_loop
self.custom_loop_ctr = custom_loop_ctr
def new_loop_with_different_body(self, new_body):
result = LoopOverCoordinate(new_body, self.coordinate_to_loop_over, self.start, self.stop,
self.step, self.is_block_loop)
result.prefix_lines = [l for l in self.prefix_lines]
self.step, self.is_block_loop, self.custom_loop_ctr)
result.prefix_lines = [prefix_line for prefix_line in self.prefix_lines]
return result
def subs(self, subs_dict):
......@@ -390,6 +459,16 @@ class LoopOverCoordinate(Node):
if hasattr(self.step, "subs"):
self.step = self.step.subs(subs_dict)
def fast_subs(self, subs_dict, skip=None):
self.body = fast_subs(self.body, subs_dict, skip)
if isinstance(self.start, sp.Basic):
self.start = fast_subs(self.start, subs_dict, skip)
if isinstance(self.stop, sp.Basic):
self.stop = fast_subs(self.stop, subs_dict, skip)
if isinstance(self.step, sp.Basic):
self.step = fast_subs(self.step, subs_dict, skip)
return self
@property
def args(self):
result = [self.body]
......@@ -422,18 +501,21 @@ class LoopOverCoordinate(Node):
@staticmethod
def get_loop_counter_name(coordinate_to_loop_over):
return "%s_%s" % (LoopOverCoordinate.LOOP_COUNTER_NAME_PREFIX, coordinate_to_loop_over)
return f"{LoopOverCoordinate.LOOP_COUNTER_NAME_PREFIX}_{coordinate_to_loop_over}"
@staticmethod
def get_block_loop_counter_name(coordinate_to_loop_over):
return "%s_%s" % (LoopOverCoordinate.BlOCK_LOOP_COUNTER_NAME_PREFIX, coordinate_to_loop_over)
return f"{LoopOverCoordinate.BLOCK_LOOP_COUNTER_NAME_PREFIX}_{coordinate_to_loop_over}"
@property
def loop_counter_name(self):
if self.is_block_loop:
return LoopOverCoordinate.get_block_loop_counter_name(self.coordinate_to_loop_over)
if self.custom_loop_ctr:
return self.custom_loop_ctr.name
else:
return LoopOverCoordinate.get_loop_counter_name(self.coordinate_to_loop_over)
if self.is_block_loop:
return LoopOverCoordinate.get_block_loop_counter_name(self.coordinate_to_loop_over)
else:
return LoopOverCoordinate.get_loop_counter_name(self.coordinate_to_loop_over)
@staticmethod
def is_loop_counter_symbol(symbol):
......@@ -447,22 +529,26 @@ class LoopOverCoordinate(Node):
@staticmethod
def get_loop_counter_symbol(coordinate_to_loop_over):
return TypedSymbol(LoopOverCoordinate.get_loop_counter_name(coordinate_to_loop_over), 'int')
return TypedSymbol(LoopOverCoordinate.get_loop_counter_name(coordinate_to_loop_over), 'int', nonnegative=True)
@staticmethod
def get_block_loop_counter_symbol(coordinate_to_loop_over):
return TypedSymbol(LoopOverCoordinate.get_block_loop_counter_name(coordinate_to_loop_over), 'int')
return TypedSymbol(LoopOverCoordinate.get_block_loop_counter_name(coordinate_to_loop_over),
'int',
nonnegative=True)
@property
def loop_counter_symbol(self):
if self.is_block_loop:
return self.get_block_loop_counter_symbol(self.coordinate_to_loop_over)
if self.custom_loop_ctr:
return self.custom_loop_ctr
else:
return self.get_loop_counter_symbol(self.coordinate_to_loop_over)
if self.is_block_loop:
return self.get_block_loop_counter_symbol(self.coordinate_to_loop_over)
else:
return self.get_loop_counter_symbol(self.coordinate_to_loop_over)
@property
def is_outermost_loop(self):
from pystencils.transformations import get_next_parent_of_type
return get_next_parent_of_type(self, LoopOverCoordinate) is None
@property
......@@ -482,15 +568,17 @@ class LoopOverCoordinate(Node):
class SympyAssignment(Node):
def __init__(self, lhs_symbol, rhs_expr, is_const=True):
def __init__(self, lhs_symbol, rhs_expr, is_const=True, use_auto=False):
super(SympyAssignment, self).__init__(parent=None)
self._lhs_symbol = lhs_symbol
self.rhs = rhs_expr
self._lhs_symbol = sp.sympify(lhs_symbol)
self._rhs = sp.sympify(rhs_expr)
self._is_const = is_const
self._is_declaration = self.__is_declaration()
self._use_auto = use_auto
def __is_declaration(self):
if isinstance(self._lhs_symbol, cast_func):
from pystencils.typing import CastFunc
if isinstance(self._lhs_symbol, CastFunc):
return False
if any(isinstance(self._lhs_symbol, c) for c in (Field.Access, sp.Indexed, TemporaryMemoryAllocation)):
return False
......@@ -500,15 +588,35 @@ class SympyAssignment(Node):
def lhs(self):
return self._lhs_symbol
@property
def rhs(self):
return self._rhs
@lhs.setter
def lhs(self, new_value):
self._lhs_symbol = new_value
self._is_declaration = self.__is_declaration()
@rhs.setter
def rhs(self, new_rhs_expr):
self._rhs = new_rhs_expr
def subs(self, subs_dict):
self.lhs = fast_subs(self.lhs, subs_dict)
self.rhs = fast_subs(self.rhs, subs_dict)
def fast_subs(self, subs_dict, skip=None):
self.lhs = fast_subs(self.lhs, subs_dict, skip)
self.rhs = fast_subs(self.rhs, subs_dict, skip)
return self
def optimize(self, optimizations):
try:
from sympy.codegen.rewriting import optimize
self.rhs = optimize(self.rhs, optimizations)
except Exception:
pass
@property
def args(self):
return [self._lhs_symbol, self.rhs]
......@@ -529,7 +637,9 @@ class SympyAssignment(Node):
for i in range(len(symbol.offsets)):
loop_counters.add(LoopOverCoordinate.get_loop_counter_symbol(i))
result.update(loop_counters)
result.update(self._lhs_symbol.atoms(sp.Symbol))
return result
@property
......@@ -540,6 +650,10 @@ class SympyAssignment(Node):
def is_const(self):
return self._is_const
@property
def use_auto(self):
return self._use_auto
def replace(self, child, replacement):
if child == self.lhs:
replacement.parent = self
......@@ -548,7 +662,7 @@ class SympyAssignment(Node):
replacement.parent = self
self.rhs = replacement
else:
raise ValueError('%s is not in args of %s' % (replacement, self.__class__))
raise ValueError(f'{replacement} is not in args of {self.__class__}')
def __repr__(self):
return repr(self.lhs) + " ← " + repr(self.rhs)
......@@ -556,13 +670,21 @@ class SympyAssignment(Node):
def _repr_html_(self):
printed_lhs = sp.latex(self.lhs)
printed_rhs = sp.latex(self.rhs)
return "${printed_lhs} \\leftarrow {printed_rhs}$".format(printed_lhs=printed_lhs, printed_rhs=printed_rhs)
return f"${printed_lhs} \\leftarrow {printed_rhs}$"
def __hash__(self):
return hash((self.lhs, self.rhs))
def __eq__(self, other):
return type(self) is type(other) and (self.lhs, self.rhs) == (other.lhs, other.rhs)
class ResolvedFieldAccess(sp.Indexed):
def __new__(cls, base, linearized_index, field, offsets, idx_coordinate_values):
if not isinstance(base, sp.IndexedBase):
assert isinstance(base, TypedSymbol)
base = sp.IndexedBase(base, shape=(1,))
assert isinstance(base.label, TypedSymbol)
obj = super(ResolvedFieldAccess, cls).__new__(cls, base, linearized_index)
obj.field = field
obj.offsets = offsets
......@@ -574,7 +696,7 @@ class ResolvedFieldAccess(sp.Indexed):
self.args[1].subs(old, new),
self.field, self.offsets, self.idx_coordinate_values)
def fast_subs(self, substitutions):
def fast_subs(self, substitutions, skip=None):
if self in substitutions:
return substitutions[self]
return ResolvedFieldAccess(self.args[0].subs(substitutions),
......@@ -591,11 +713,14 @@ class ResolvedFieldAccess(sp.Indexed):
def __str__(self):
top = super(ResolvedFieldAccess, self).__str__()
return "%s (%s)" % (top, self.typed_symbol.dtype)
return f"{top} ({self.typed_symbol.dtype})"
def __getnewargs__(self):
return self.base, self.indices[0], self.field, self.offsets, self.idx_coordinate_values
def __getnewargs_ex__(self):
return (self.base, self.indices[0], self.field, self.offsets, self.idx_coordinate_values), {}
class TemporaryMemoryAllocation(Node):
"""Node for temporary memory buffer allocation.
......@@ -667,59 +792,79 @@ def early_out(condition):
return Conditional(vec_all(condition), Block([SkipIteration()]))
class DestructuringBindingsForFieldClass(Node):
"""
Defines all variables needed for describing a field (shape, pointer, strides)
"""
CLASS_TO_MEMBER_DICT = {
FieldPointerSymbol: "data",
FieldShapeSymbol: "shape[%i]",
FieldStrideSymbol: "stride[%i]"
}
CLASS_NAME_TEMPLATE = "PyStencilsField<{dtype}, {ndim}>"
def get_dummy_symbol(dtype='bool'):
return TypedSymbol(f'dummy{uuid.uuid4().hex}', create_type(dtype))
class SourceCodeComment(Node):
def __init__(self, text):
self.text = text
@property
def fields_accessed(self) -> Set['ResolvedFieldAccess']:
"""Set of Field instances: fields which are accessed inside this kernel function"""
return set(o.field for o in self.atoms(ResolvedFieldAccess))
def args(self):
return []
def __init__(self, body):
super(DestructuringBindingsForFieldClass, self).__init__()
self.headers = ['<PyStencilsField.h>']
self.body = body
@property
def symbols_defined(self):
return set()
@property
def args(self) -> List[NodeOrExpr]:
"""Returns all arguments/children of this node."""
def undefined_symbols(self):
return set()
def __str__(self):
return "/* " + self.text + " */"
def __repr__(self):
return self.__str__()
class EmptyLine(Node):
def __init__(self):
pass
@property
def symbols_defined(self) -> Set[sp.Symbol]:
"""Set of symbols which are defined by this node."""
undefined_field_symbols = {s for s in self.body.undefined_symbols
if isinstance(s, (FieldPointerSymbol, FieldShapeSymbol, FieldStrideSymbol))}
return undefined_field_symbols
def args(self):
return []
@property
def undefined_symbols(self) -> Set[sp.Symbol]:
field_map = {f.name: f for f in self.fields_accessed}
undefined_field_symbols = self.symbols_defined
corresponding_field_names = {s.field_name for s in undefined_field_symbols if hasattr(s, 'field_name')}
corresponding_field_names |= {s.field_names[0] for s in undefined_field_symbols if hasattr(s, 'field_names')}
return {TypedSymbol(f, self.CLASS_NAME_TEMPLATE.format(dtype=field_map[f].dtype, ndim=field_map[f].ndim) + '&')
for f in corresponding_field_names} | (self.body.undefined_symbols - undefined_field_symbols)
def symbols_defined(self):
return set()
def subs(self, subs_dict) -> None:
"""Inplace! substitute, similar to sympy's but modifies the AST inplace."""
self.body.subs(subs_dict)
@property
def undefined_symbols(self):
return set()
def __str__(self):
return ""
def __repr__(self):
return self.__str__()
class ConditionalFieldAccess(sp.Function):
"""
:class:`pystencils.Field.Access` that is only executed if a certain condition is met.
Can be used, for instance, for out-of-bound checks.
"""
def __new__(cls, field_access, outofbounds_condition, outofbounds_value=0):
return sp.Function.__new__(cls, field_access, outofbounds_condition, sp.S(outofbounds_value))
@property
def func(self):
return self.__class__
def access(self):
return self.args[0]
def atoms(self, arg_type) -> Set[Any]:
return self.body.atoms(arg_type) | {s for s in self.symbols_defined if isinstance(s, arg_type)}
@property
def outofbounds_condition(self):
return self.args[1]
@property
def outofbounds_value(self):
return self.args[2]
def get_dummy_symbol(dtype='bool'):
return TypedSymbol('dummy%s' % uuid.uuid4().hex, create_type(dtype))
def __getnewargs__(self):
return self.access, self.outofbounds_condition, self.outofbounds_value
def __getnewargs_ex__(self):
return (self.access, self.outofbounds_condition, self.outofbounds_value), {}
pystencils/backends/__init__.py → src/pystencils/backends/__init__.py
View file @ 5600b6b6
......@@ -6,9 +6,3 @@ try:
__all__.append('print_dot')
except ImportError:
pass
try:
from .llvm import generate_llvm # NOQA
__all__.append('generate_llvm')
except ImportError:
pass
src/pystencils/backends/arm_instruction_sets.py 0 → 100644
View file @ 5600b6b6
from pystencils.typing import CFunction
def get_argument_string(function_shortcut, first=''):
args = function_shortcut[function_shortcut.index('[') + 1: -1]
arg_string = "("
if first:
arg_string += first + ', '
for arg in args.split(","):
arg = arg.strip()
if not arg:
continue
if arg in ('0', '1', '2', '3', '4', '5'):
arg_string += "{" + arg + "},"
else:
arg_string += arg + ","
arg_string = arg_string[:-1] + ")"
return arg_string
def get_vector_instruction_set_arm(data_type='double', instruction_set='neon'):
if instruction_set not in ['neon', 'sme'] and not instruction_set.startswith('sve'):
raise NotImplementedError(instruction_set)
if instruction_set in ['sve', 'sve2', 'sme']:
cmp = 'cmp'
elif instruction_set.startswith('sve2') and instruction_set not in ('sve256', 'sve2048'):
cmp = 'cmp'
bitwidth = int(instruction_set[4:])
elif instruction_set.startswith('sve'):
cmp = 'cmp'
bitwidth = int(instruction_set[3:])
elif instruction_set == 'neon':
cmp = 'c'
bitwidth = 128
base_names = {
'+': 'add[0, 1]',
'-': 'sub[0, 1]',
'*': 'mul[0, 1]',
'/': 'div[0, 1]',
'sqrt': 'sqrt[0]',
'loadU': 'ld1[0]',
'storeU': 'st1[0, 1]',
'abs': 'abs[0]',
'==': f'{cmp}eq[0, 1]',
'!=': f'{cmp}eq[0, 1]',
'<=': f'{cmp}le[0, 1]',
'<': f'{cmp}lt[0, 1]',
'>=': f'{cmp}ge[0, 1]',
'>': f'{cmp}gt[0, 1]',
}
bits = {'double': 64,
'float': 32,
'int': 32}
result = dict()
if instruction_set in ['sve', 'sve2', 'sme']:
width = 'svcntd()' if data_type == 'double' else 'svcntw()'
intwidth = 'svcntw()'
result['bytes'] = 'svcntb()'
else:
width = bitwidth // bits[data_type]
intwidth = bitwidth // bits['int']
result['bytes'] = bitwidth // 8
if instruction_set.startswith('sve') or instruction_set == 'sme':
base_names['stream'] = 'stnt1[0, 1]'
prefix = 'sv'
suffix = f'_f{bits[data_type]}'
elif instruction_set == 'neon':
prefix = 'v'
suffix = f'q_f{bits[data_type]}'
if instruction_set in ['sve', 'sve2', 'sme']:
predicate = f'{prefix}whilelt_b{bits[data_type]}_u64({{loop_counter}}, {{loop_stop}})'
int_predicate = f'{prefix}whilelt_b{bits["int"]}_u64({{loop_counter}}, {{loop_stop}})'
else:
predicate = f'{prefix}whilelt_b{bits[data_type]}(0, {width})'
int_predicate = f'{prefix}whilelt_b{bits["int"]}(0, {intwidth})'
for intrinsic_id, function_shortcut in base_names.items():
function_shortcut = function_shortcut.strip()
name = function_shortcut[:function_shortcut.index('[')]
arg_string = get_argument_string(function_shortcut, first=predicate if prefix == 'sv' else '')
if prefix == 'sv' and not name.startswith('ld') and not name.startswith('st') and not name.startswith(cmp):
undef = '_x'
else:
undef = ''
result[intrinsic_id] = prefix + name + suffix + undef + arg_string
if instruction_set in ['sve', 'sve2', 'sme']:
result['width'] = CFunction(width, "int")
result['intwidth'] = CFunction(intwidth, "int")
else:
result['width'] = width
result['intwidth'] = intwidth
if instruction_set.startswith('sve') or instruction_set == 'sme':
result['makeVecConst'] = f'svdup_f{bits[data_type]}' + '({0})'
result['makeVecConstInt'] = f'svdup_s{bits["int"]}' + '({0})'
result['makeVecIndex'] = f'svindex_s{bits["int"]}' + '({0}, {1})'
if instruction_set != 'sme':
vindex = f'svindex_u{bits[data_type]}(0, {{0}})'
result['storeS'] = f'svst1_scatter_u{bits[data_type]}index_f{bits[data_type]}({predicate}, {{0}}, ' + \
vindex.format("{2}") + ', {1})'
result['loadS'] = f'svld1_gather_u{bits[data_type]}index_f{bits[data_type]}({predicate}, {{0}}, ' + \
vindex.format("{1}") + ')'
if instruction_set.startswith('sve2') and instruction_set not in ('sve256', 'sve2048'):
result['streamS'] = f'svstnt1_scatter_u{bits[data_type]}offset_f{bits[data_type]}({predicate}, {{0}}, ' + \
vindex.format(f"{{2}}*{bits[data_type]//8}") + ', {1})'
result['+int'] = f"svadd_s{bits['int']}_x({int_predicate}, " + "{0}, {1})"
result['float'] = f'svfloat{bits["float"]}_{"s" if instruction_set not in ["sve", "sve2", "sme"] else ""}t'
result['double'] = f'svfloat{bits["double"]}_{"s" if instruction_set not in ["sve", "sve2", "sme"] else ""}t'
result['int'] = f'svint{bits["int"]}_{"s" if instruction_set not in ["sve", "sve2", "sme"] else ""}t'
result['bool'] = f'svbool_{"s" if instruction_set not in ["sve", "sve2", "sme"] else ""}t'
result['headers'] = ['<arm_sve.h>', '<arm_acle.h>', '"arm_neon_helpers.h"']
result['&'] = f'svand_b_z({predicate},' + ' {0}, {1})'
result['|'] = f'svorr_b_z({predicate},' + ' {0}, {1})'
result['blendv'] = f'svsel_f{bits[data_type]}' + '({2}, {1}, {0})'
result['any'] = f'svptest_any({predicate}, {{0}})'
result['all'] = f'svcntp_b{bits[data_type]}({predicate}, {{0}}) == {width}'
result['maskStoreU'] = result['storeU'].replace(predicate, '{2}')
result['maskStream'] = result['stream'].replace(predicate, '{2}')
if instruction_set != 'sme':
result['maskStoreS'] = result['storeS'].replace(predicate, '{3}')
if instruction_set.startswith('sve2') and instruction_set not in ('sve256', 'sve2048'):
result['maskStreamS'] = result['streamS'].replace(predicate, '{3}')
result['streamFence'] = '__dmb(15)'
if instruction_set == 'sme':
result['function_prefix'] = '__attribute__((arm_locally_streaming))'
elif instruction_set not in ['sve', 'sve2', 'sme']:
result['compile_flags'] = [f'-msve-vector-bits={bitwidth}']
else:
result['makeVecConst'] = f'vdupq_n_f{bits[data_type]}' + '({0})'
result['makeVec'] = f'makeVec_f{bits[data_type]}' + '(' + ", ".join(['{' + str(i) + '}' for i in
range(width)]) + ')'
result['makeVecConstInt'] = f'vdupq_n_s{bits["int"]}' + '({0})'
result['makeVecInt'] = f'makeVec_s{bits["int"]}' + '({0}, {1}, {2}, {3})'
result['+int'] = f"vaddq_s{bits['int']}" + "({0}, {1})"
result[data_type] = f'float{bits[data_type]}x{width}_t'
result['int'] = f'int{bits["int"]}x{intwidth}_t'
result['bool'] = f'uint{bits[data_type]}x{width}_t'
result['headers'] = ['<arm_neon.h>', '"arm_neon_helpers.h"']
result['!='] = f'vmvnq_u{bits[data_type]}({result["=="]})'
result['&'] = f'vandq_u{bits[data_type]}' + '({0}, {1})'
result['|'] = f'vorrq_u{bits[data_type]}' + '({0}, {1})'
result['blendv'] = f'vbslq_f{bits[data_type]}' + '({2}, {1}, {0})'
result['any'] = f'vaddlvq_u8(vreinterpretq_u8_u{bits[data_type]}({{0}})) > 0'
result['all'] = f'vaddlvq_u8(vreinterpretq_u8_u{bits[data_type]}({{0}})) == 16*0xff'
# SVE has real nontemporal stores, so we only need to zero cachlines on Neon
result['cachelineZero'] = 'cachelineZero((void*) {0})'
result['cachelineSize'] = 'cachelineSize()'
return result
pystencils/backends/cbackend.py → src/pystencils/backends/cbackend.py
View file @ 5600b6b6
import re
from collections import namedtuple
import hashlib
from typing import Set
import numpy as np
import sympy as sp
from sympy.core import S
from sympy.printing.ccode import C89CodePrinter
from pystencils.astnodes import KernelFunction, Node
from pystencils.cpu.vectorization import vec_all, vec_any
from pystencils.data_types import (
PointerType, VectorType, address_of, cast_func, create_type, get_type_of_expression,
reinterpret_cast_func, vector_memory_access)
from sympy.logic.boolalg import BooleanFalse, BooleanTrue
from sympy.functions.elementary.trigonometric import TrigonometricFunction, InverseTrigonometricFunction
from sympy.functions.elementary.hyperbolic import HyperbolicFunction
from pystencils.astnodes import KernelFunction, LoopOverCoordinate, Node
from pystencils.cpu.vectorization import vec_all, vec_any, CachelineSize
from pystencils.typing import (
PointerType, VectorType, CastFunc, create_type, get_type_of_expression,
ReinterpretCastFunc, VectorMemoryAccess, BasicType, TypedSymbol, CFunction)
from pystencils.enums import Backend
from pystencils.fast_approximation import fast_division, fast_inv_sqrt, fast_sqrt
from pystencils.functions import DivFunc, AddressOf
from pystencils.integer_functions import (
bit_shift_left, bit_shift_right, bitwise_and, bitwise_or, bitwise_xor,
int_div, int_power_of_2, modulo_ceil)
from pystencils.kernelparameters import FieldPointerSymbol
try:
from sympy.printing.ccode import C99CodePrinter as CCodePrinter
from sympy.printing.c import C99CodePrinter as CCodePrinter # for sympy versions > 1.6
except ImportError:
from sympy.printing.ccode import CCodePrinter # for sympy versions < 1.1
from sympy.printing.ccode import C99CodePrinter as CCodePrinter
__all__ = ['generate_c', 'CustomCodeNode', 'PrintNode', 'get_headers', 'CustomSympyPrinter']
KERNCRAFT_NO_TERNARY_MODE = False
HEADER_REGEX = re.compile(r'^[<"].*[">]$')
def generate_c(ast_node: Node, signature_only: bool = False, dialect='c', custom_backend=None) -> str:
def generate_c(ast_node: Node,
signature_only: bool = False,
dialect: Backend = Backend.C,
custom_backend=None,
with_globals=True) -> str:
"""Prints an abstract syntax tree node as C or CUDA code.
This function does not need to distinguish between C, C++ or CUDA code, it just prints 'C-like' code as encoded
in the abstract syntax tree (AST). The AST is built differently for C or CUDA by calling different create_kernel
functions.
This function does not need to distinguish for most AST nodes between C, C++ or CUDA code, it just prints 'C-like'
code as encoded in the abstract syntax tree (AST). The AST is built differently for C or CUDA by calling different
create_kernel functions.
Args:
ast_node:
signature_only:
dialect: 'c' or 'cuda'
ast_node: ast representation of kernel
signature_only: generate signature without function body
dialect: `Backend`: 'C' or 'CUDA'
custom_backend: use own custom printer for code generation
with_globals: enable usage of global variables
Returns:
C-like code for the ast node and its descendants
"""
......@@ -49,26 +61,25 @@ def generate_c(ast_node: Node, signature_only: bool = False, dialect='c', custom
ast_node.global_variables = d.symbols_defined
if custom_backend:
printer = custom_backend
elif dialect == 'c':
elif dialect == Backend.C:
try:
# TODO Vectorization Revamp: instruction_set should not be just slapped on ast
instruction_set = ast_node.instruction_set
except Exception:
instruction_set = None
printer = CBackend(signature_only=signature_only,
vector_instruction_set=instruction_set)
elif dialect == 'cuda':
elif dialect == Backend.CUDA:
from pystencils.backends.cuda_backend import CudaBackend
printer = CudaBackend(signature_only=signature_only)
elif dialect == 'opencl':
from pystencils.backends.opencl_backend import OpenClBackend
printer = OpenClBackend(signature_only=signature_only)
else:
raise ValueError("Unknown dialect: " + str(dialect))
raise ValueError(f'Unknown {dialect=}')
code = printer(ast_node)
if not signature_only and isinstance(ast_node, KernelFunction):
code = "\n" + code
for declaration in global_declarations:
code = printer(declaration) + "\n" + code
if with_globals and global_declarations:
code = "\n" + code
for declaration in global_declarations:
code = printer(declaration) + "\n" + code
return code
......@@ -77,8 +88,8 @@ def get_global_declarations(ast):
global_declarations = []
def visit_node(sub_ast):
nonlocal global_declarations
if hasattr(sub_ast, "required_global_declarations"):
nonlocal global_declarations
global_declarations += sub_ast.required_global_declarations
if hasattr(sub_ast, "args"):
......@@ -87,7 +98,7 @@ def get_global_declarations(ast):
visit_node(ast)
return set(global_declarations)
return sorted(set(global_declarations), key=str)
def get_headers(ast_node: Node) -> Set[str]:
......@@ -100,15 +111,22 @@ def get_headers(ast_node: Node) -> Set[str]:
if hasattr(ast_node, 'headers'):
headers.update(ast_node.headers)
for a in ast_node.args:
if isinstance(a, Node):
if isinstance(a, (sp.Expr, Node)):
headers.update(get_headers(a))
for g in get_global_declarations(ast_node):
if isinstance(g, Node):
headers.update(get_headers(g))
for h in headers:
assert HEADER_REGEX.match(h), f'header /{h}/ does not follow the pattern /"..."/ or /<...>/'
return headers
# --------------------------------------- Backend Specific Nodes -------------------------------------------------------
# TODO future CustomCodeNode should not be backend specific move it elsewhere
class CustomCodeNode(Node):
def __init__(self, code, symbols_read, symbols_defined, parent=None):
super(CustomCodeNode, self).__init__(parent=parent)
......@@ -117,7 +135,7 @@ class CustomCodeNode(Node):
self._symbols_defined = set(symbols_defined)
self.headers = []
def get_code(self, dialect, vector_instruction_set):
def get_code(self, dialect, vector_instruction_set, print_arg):
return self._code
@property
......@@ -132,11 +150,17 @@ class CustomCodeNode(Node):
def undefined_symbols(self):
return self._symbols_read - self._symbols_defined
def __eq__(self, other):
return type(self) is type(other) and self._code == other._code
def __hash__(self):
return hash(self._code)
class PrintNode(CustomCodeNode):
# noinspection SpellCheckingInspection
def __init__(self, symbol_to_print):
code = '\nstd::cout << "%s = " << %s << std::endl; \n' % (symbol_to_print.name, symbol_to_print.name)
code = f'\nstd::cout << "{symbol_to_print.name} = " << {symbol_to_print.name} << std::endl; \n'
super(PrintNode, self).__init__(code, symbols_read=[symbol_to_print], symbols_defined=set())
self.headers.append("<iostream>")
......@@ -147,7 +171,7 @@ class PrintNode(CustomCodeNode):
# noinspection PyPep8Naming
class CBackend:
def __init__(self, sympy_printer=None, signature_only=False, vector_instruction_set=None, dialect='c'):
def __init__(self, sympy_printer=None, signature_only=False, vector_instruction_set=None, dialect=Backend.C):
if sympy_printer is None:
if vector_instruction_set is not None:
self.sympy_printer = VectorizedCustomSympyPrinter(vector_instruction_set)
......@@ -160,6 +184,8 @@ class CBackend:
self._indent = " "
self._dialect = dialect
self._signatureOnly = signature_only
self._kwargs = {}
self.sympy_printer._kwargs = self._kwargs
def __call__(self, node):
prev_is = VectorType.instruction_set
......@@ -172,21 +198,22 @@ class CBackend:
if isinstance(node, str):
return node
for cls in type(node).__mro__:
method_name = "_print_" + cls.__name__
method_name = f"_print_{cls.__name__}"
if hasattr(self, method_name):
return getattr(self, method_name)(node)
raise NotImplementedError(self.__class__.__name__ + " does not support node of type " + node.__class__.__name__)
raise NotImplementedError(f"{self.__class__.__name__} does not support node of type {node.__class__.__name__}")
def _print_Type(self, node):
def _print_AbstractType(self, node):
return str(node)
def _print_KernelFunction(self, node):
function_arguments = ["%s %s" % (self._print(s.symbol.dtype), s.symbol.name) for s in node.get_parameters()]
function_arguments = [f"{self._print(s.symbol.dtype)} {s.symbol.name}" for s in node.get_parameters()
if not type(s.symbol) is CFunction]
launch_bounds = ""
if self._dialect == 'cuda':
if self._dialect == Backend.CUDA:
max_threads = node.indexing.max_threads_per_block()
if max_threads:
launch_bounds = "__launch_bounds__({}) ".format(max_threads)
launch_bounds = f"__launch_bounds__({max_threads}) "
func_declaration = "FUNC_PREFIX %svoid %s(%s)" % (launch_bounds, node.function_name,
", ".join(function_arguments))
if self._signatureOnly:
......@@ -200,54 +227,166 @@ class CBackend:
return "{\n%s\n}" % (self._indent + self._indent.join(block_contents.splitlines(True)))
def _print_PragmaBlock(self, node):
return "%s\n%s" % (node.pragma_line, self._print_Block(node))
return f"{node.pragma_line}\n{self._print_Block(node)}"
def _print_LoopOverCoordinate(self, node):
counter_symbol = node.loop_counter_name
start = "int %s = %s" % (counter_symbol, self.sympy_printer.doprint(node.start))
condition = "%s < %s" % (counter_symbol, self.sympy_printer.doprint(node.stop))
update = "%s += %s" % (counter_symbol, self.sympy_printer.doprint(node.step),)
loop_str = "for (%s; %s; %s)" % (start, condition, update)
counter_name = node.loop_counter_name
counter_dtype = node.loop_counter_symbol.dtype.c_name
start = f"{counter_dtype} {counter_name} = {self.sympy_printer.doprint(node.start)}"
condition = f"{counter_name} < {self.sympy_printer.doprint(node.stop)}"
update = f"{counter_name} += {self.sympy_printer.doprint(node.step)}"
loop_str = f"for ({start}; {condition}; {update})"
self._kwargs['loop_counter'] = counter_name
self._kwargs['loop_stop'] = node.stop
prefix = "\n".join(node.prefix_lines)
if prefix:
prefix += "\n"
return "%s%s\n%s" % (prefix, loop_str, self._print(node.body))
return f"{prefix}{loop_str}\n{self._print(node.body)}"
def _print_SympyAssignment(self, node):
printed_lhs = self.sympy_printer.doprint(node.lhs)
printed_rhs = self.sympy_printer.doprint(node.rhs)
if node.is_declaration:
if node.is_const:
prefix = 'const '
if node.use_auto:
data_type = 'auto'
else:
prefix = ''
data_type = prefix + self._print(node.lhs.dtype) + " "
return "%s%s = %s;" % (data_type, self.sympy_printer.doprint(node.lhs),
self.sympy_printer.doprint(node.rhs))
data_type = self._print(node.lhs.dtype).replace(' const', '')
if node.is_const:
data_type = f'const {data_type}'
return f"{data_type} {printed_lhs} = {printed_rhs};"
else:
lhs_type = get_type_of_expression(node.lhs)
if type(lhs_type) is VectorType and isinstance(node.lhs, cast_func):
arg, data_type, aligned, nontemporal = node.lhs.args
lhs_type = get_type_of_expression(node.lhs) # TOOD: this should have been typed
printed_mask = ""
if type(lhs_type) is VectorType and isinstance(node.lhs, CastFunc):
arg, data_type, aligned, nontemporal, mask, stride = node.lhs.args
instr = 'storeU'
if aligned:
instr = 'stream' if nontemporal else 'storeA'
if nontemporal and 'storeA' not in self._vector_instruction_set and \
'stream' in self._vector_instruction_set:
instr = 'stream'
elif aligned:
instr = 'stream' if nontemporal and 'stream' in self._vector_instruction_set else 'storeA'
if mask != True: # NOQA
instr = 'maskStream' if nontemporal and 'maskStream' in self._vector_instruction_set else \
'maskStoreA' if aligned else 'maskStoreU'
if instr not in self._vector_instruction_set:
if instr == 'maskStream' and 'stream' in self._vector_instruction_set:
store, load = 'stream', 'loadA'
elif (instr in ('maskStream', 'maskStoreA')) and 'storeA' in self._vector_instruction_set:
store, load = 'storeA', 'loadA'
else:
store, load = 'storeU', 'loadU'
load = load if load in self._vector_instruction_set else 'loadU'
self._vector_instruction_set[instr] = self._vector_instruction_set[store].format(
'{0}', self._vector_instruction_set['blendv'].format(
self._vector_instruction_set[load].format('{0}', **self._kwargs),
'{1}', '{2}', **self._kwargs), **self._kwargs)
printed_mask = self.sympy_printer.doprint(mask)
if data_type.base_type.c_name == 'double':
if self._vector_instruction_set['double'] == '__m256d':
printed_mask = f"_mm256_castpd_si256({printed_mask})"
elif self._vector_instruction_set['double'] == '__m128d':
printed_mask = f"_mm_castpd_si128({printed_mask})"
elif data_type.base_type.c_name == 'float':
if self._vector_instruction_set['float'] == '__m256':
printed_mask = f"_mm256_castps_si256({printed_mask})"
elif self._vector_instruction_set['float'] == '__m128':
printed_mask = f"_mm_castps_si128({printed_mask})"
rhs_type = get_type_of_expression(node.rhs)
if type(rhs_type) is not VectorType:
rhs = cast_func(node.rhs, VectorType(rhs_type))
raise ValueError(f'Cannot vectorize {node.rhs} of type {rhs_type} inside of the pretty printer! '
f'This should have happen earlier!')
# rhs = CastFunc(node.rhs, VectorType(rhs_type)) # Unknown width
else:
rhs = node.rhs
return self._vector_instruction_set[instr].format("&" + self.sympy_printer.doprint(node.lhs.args[0]),
self.sympy_printer.doprint(rhs)) + ';'
ptr = "&" + self.sympy_printer.doprint(node.lhs.args[0])
if stride != 1:
instr = ('maskStreamS' if nontemporal and 'maskStreamS' in self._vector_instruction_set else
'maskStoreS') if mask != True else \
('streamS' if nontemporal and 'streamS' in self._vector_instruction_set else 'storeS') # NOQA
return self._vector_instruction_set[instr].format(ptr, self.sympy_printer.doprint(rhs),
stride, printed_mask, **self._kwargs) + ';'
pre_code = ''
if nontemporal and 'cachelineZero' in self._vector_instruction_set and mask == True: # NOQA
first_cond = f"((uintptr_t) {ptr} & {CachelineSize.mask_symbol}) == 0"
offset = sp.Add(*[sp.Symbol(LoopOverCoordinate.get_loop_counter_name(i))
* node.lhs.args[0].field.spatial_strides[i] for i in
range(len(node.lhs.args[0].field.spatial_strides))])
if stride == 1:
offset = offset.subs({node.lhs.args[0].field.spatial_strides[0]: 1})
size = sp.Mul(*node.lhs.args[0].field.spatial_shape)
element_size = 8 if data_type.base_type.c_name == 'double' else 4
size_cond = f"({offset} + {CachelineSize.symbol/element_size}) < {size}"
pre_code = f"if ({first_cond} && {size_cond}) " + "{\n\t" + \
self._vector_instruction_set['cachelineZero'].format(ptr, **self._kwargs) + ';\n}\n'
code = self._vector_instruction_set[instr].format(ptr, self.sympy_printer.doprint(rhs),
printed_mask, **self._kwargs) + ';'
flushcond = f"((uintptr_t) {ptr} & {CachelineSize.mask_symbol}) == {CachelineSize.last_symbol}"
if nontemporal and 'flushCacheline' in self._vector_instruction_set:
code2 = self._vector_instruction_set['flushCacheline'].format(
ptr, self.sympy_printer.doprint(rhs), **self._kwargs) + ';'
code = f"{code}\nif ({flushcond}) {{\n\t{code2}\n}}"
elif aligned and nontemporal and 'storeAAndFlushCacheline' in self._vector_instruction_set:
lhs_hash = hashlib.sha1(self.sympy_printer.doprint(node.lhs).encode('ascii')).hexdigest()[:8]
rhs_hash = hashlib.sha1(self.sympy_printer.doprint(rhs).encode('ascii')).hexdigest()[:8]
tmpvar = f'_tmp_{lhs_hash}_{rhs_hash}'
code = 'const ' + self._print(node.lhs.dtype).replace(' const', '') + ' ' + tmpvar + ' = ' \
+ self.sympy_printer.doprint(rhs) + ';'
code1 = self._vector_instruction_set[instr].format(ptr, tmpvar, printed_mask, **self._kwargs) + ';'
maskStore, store, load = 'maskStoreAAndFlushCacheline', 'storeAAndFlushCacheline', 'loadA'
instr2 = maskStore if mask != True else store # NOQA
if instr2 not in self._vector_instruction_set:
self._vector_instruction_set[maskStore] = self._vector_instruction_set[store].format(
'{0}', self._vector_instruction_set['blendv'].format(
self._vector_instruction_set[load].format('{0}', **self._kwargs),
'{1}', '{2}', **self._kwargs),
**self._kwargs)
code2 = self._vector_instruction_set[instr2].format(ptr, tmpvar, printed_mask, **self._kwargs) + ';'
code += f"\nif ({flushcond}) {{\n\t{code2}\n}} else {{\n\t{code1}\n}}"
return pre_code + code
else:
return "%s = %s;" % (self.sympy_printer.doprint(node.lhs), self.sympy_printer.doprint(node.rhs))
return f"{printed_lhs} = {printed_rhs};"
def _print_NontemporalFence(self, _):
if 'streamFence' in self._vector_instruction_set:
return self._vector_instruction_set['streamFence'] + ';'
else:
return ''
def _print_CachelineSize(self, node):
if 'cachelineSize' in self._vector_instruction_set:
code = f'const size_t {node.symbol} = {self._vector_instruction_set["cachelineSize"]};\n'
code += f'const size_t {node.mask_symbol} = {node.symbol} - 1;\n'
vectorsize = self._vector_instruction_set['bytes']
code += f'const size_t {node.last_symbol} = {node.symbol} - {vectorsize};\n'
return code
else:
return ''
def _print_TemporaryMemoryAllocation(self, node):
align = 64
if self._vector_instruction_set:
align = self._vector_instruction_set['bytes']
else:
align = node.symbol.dtype.base_type.numpy_dtype.itemsize
np_dtype = node.symbol.dtype.base_type.numpy_dtype
required_size = np_dtype.itemsize * node.size + align
size = modulo_ceil(required_size, align)
code = "{dtype} {name}=({dtype})aligned_alloc({align}, {size}) + {offset};"
code = "#if defined(_MSC_VER)\n"
code += "{dtype} {name}=({dtype})_aligned_malloc({size}, {align}) + {offset};\n"
code += "#elif __cplusplus >= 201703L || __STDC_VERSION__ >= 201112L\n"
code += "{dtype} {name}=({dtype})aligned_alloc({align}, {size}) + {offset};\n"
code += "#else\n"
code += "{dtype} {name};\n"
code += "posix_memalign((void**) &{name}, {align}, {size});\n"
code += "{name} += {offset};\n"
code += "#endif"
return code.format(dtype=node.symbol.dtype,
name=self.sympy_printer.doprint(node.symbol.name),
size=self.sympy_printer.doprint(size),
......@@ -255,45 +394,46 @@ class CBackend:
align=align)
def _print_TemporaryMemoryFree(self, node):
align = 64
return "free(%s - %d);" % (self.sympy_printer.doprint(node.symbol.name), node.offset(align))
if self._vector_instruction_set:
align = self._vector_instruction_set['bytes']
else:
align = node.symbol.dtype.base_type.numpy_dtype.itemsize
code = "#if defined(_MSC_VER)\n"
code += "_aligned_free(%s - %d);\n" % (self.sympy_printer.doprint(node.symbol.name), node.offset(align))
code += "#else\n"
code += "free(%s - %d);\n" % (self.sympy_printer.doprint(node.symbol.name), node.offset(align))
code += "#endif"
return code
def _print_SkipIteration(self, _):
return "continue;"
def _print_CustomCodeNode(self, node):
return node.get_code(self._dialect, self._vector_instruction_set)
return node.get_code(self._dialect, self._vector_instruction_set, print_arg=self.sympy_printer._print)
def _print_SourceCodeComment(self, node):
return f"/* {node.text } */"
def _print_EmptyLine(self, node):
return ""
def _print_Conditional(self, node):
if type(node.condition_expr) is BooleanTrue:
return self._print_Block(node.true_block)
elif type(node.condition_expr) is BooleanFalse:
return self._print_Block(node.false_block)
cond_type = get_type_of_expression(node.condition_expr)
if isinstance(cond_type, VectorType):
raise ValueError("Problem with Conditional inside vectorized loop - use vec_any or vec_all")
condition_expr = self.sympy_printer.doprint(node.condition_expr)
true_block = self._print_Block(node.true_block)
result = "if (%s)\n%s " % (condition_expr, true_block)
result = f"if ({condition_expr})\n{true_block} "
if node.false_block:
false_block = self._print_Block(node.false_block)
result += "else " + false_block
result += f"else {false_block}"
return result
def _print_DestructuringBindingsForFieldClass(self, node):
# Define all undefined symbols
undefined_field_symbols = node.symbols_defined
destructuring_bindings = ["%s %s = %s.%s;" %
(u.dtype,
u.name,
u.field_name if hasattr(u, 'field_name') else u.field_names[0],
node.CLASS_TO_MEMBER_DICT[u.__class__] %
(() if type(u) == FieldPointerSymbol else (u.coordinate,)))
for u in undefined_field_symbols
]
destructuring_bindings.sort() # only for code aesthetics
return "{\n" + self._indent + \
("\n" + self._indent).join(destructuring_bindings) + \
"\n" + self._indent + \
("\n" + self._indent).join(self._print(node.body).splitlines()) + \
"\n}"
# ------------------------------------------ Helper function & classes -------------------------------------------------
......@@ -303,27 +443,31 @@ class CustomSympyPrinter(CCodePrinter):
def __init__(self):
super(CustomSympyPrinter, self).__init__()
self._float_type = create_type("float32")
if 'Min' in self.known_functions:
del self.known_functions['Min']
if 'Max' in self.known_functions:
del self.known_functions['Max']
def _print_Pow(self, expr):
"""Don't use std::pow function, for small integer exponents, write as multiplication"""
if not expr.free_symbols:
return self._typed_number(expr.evalf(), get_type_of_expression(expr))
if expr.exp.is_integer and expr.exp.is_number and 0 < expr.exp < 8:
return "(" + self._print(sp.Mul(*[expr.base] * expr.exp, evaluate=False)) + ")"
elif expr.exp.is_integer and expr.exp.is_number and - 8 < expr.exp < 0:
return "1 / ({})".format(self._print(sp.Mul(*[expr.base] * (-expr.exp), evaluate=False)))
# Ideally the printer has as little logic as possible. Therefore,
# powers should be rewritten as `DivFunc`s / unevaluated `Mul`s before
# printing. `NodeCollection` offers a convenience function to do just
# that. However, `cut_loops` rewrites unevaluated multiplications as
# `Pow`s again. Neither `deepcopy` nor `func(*args)` are suited to
# rebuild unevaluated expressions. Therefore, as long as we stick with
# SymPy, this is the only way to avoid printing `pow`s.
exp = expr.exp.expr if isinstance(expr.exp, CastFunc) else expr.exp
one_type = expr.base.dtype if hasattr(expr.base, "dtype") else get_type_of_expression(expr.base)
if exp.is_integer and exp.is_number and (0 < exp <= 8):
return f"({self._print(sp.Mul(*[expr.base] * exp, evaluate=False))})"
elif exp.is_integer and exp.is_number and (-8 <= exp < 0):
return f"{self._typed_number(1, one_type)} / ({self._print(sp.Mul(*([expr.base] * -exp), evaluate=False))})"
else:
return super(CustomSympyPrinter, self)._print_Pow(expr)
# TODO don't print ones in sp.Mul
def _print_Rational(self, expr):
"""Evaluate all rationals i.e. print 0.25 instead of 1.0/4.0"""
res = str(expr.evalf().num)
res = str(expr.evalf(17))
return res
def _print_Equality(self, expr):
......@@ -335,6 +479,15 @@ class CustomSympyPrinter(CCodePrinter):
result = super(CustomSympyPrinter, self)._print_Piecewise(expr)
return result.replace("\n", "")
def _print_Abs(self, expr):
if expr.args[0].is_integer:
return f'abs({self._print(expr.args[0])})'
else:
return f'fabs({self._print(expr.args[0])})'
def _print_AbstractType(self, node):
return str(node)
def _print_Function(self, expr):
infix_functions = {
bitwise_xor: '^',
......@@ -345,34 +498,65 @@ class CustomSympyPrinter(CCodePrinter):
}
if hasattr(expr, 'to_c'):
return expr.to_c(self._print)
if isinstance(expr, reinterpret_cast_func):
if isinstance(expr, ReinterpretCastFunc):
arg, data_type = expr.args
return "*((%s)(& %s))" % (PointerType(data_type, restrict=False), self._print(arg))
elif isinstance(expr, address_of):
if isinstance(data_type, PointerType):
const_str = "const" if data_type.const else ""
return f"(({const_str} {self._print(data_type.base_type)} *)(& {self._print(arg)}))"
else:
return f"*(({self._print(PointerType(data_type, restrict=False))})(& {self._print(arg)}))"
elif isinstance(expr, AddressOf):
assert len(expr.args) == 1, "address_of must only have one argument"
return "&(%s)" % self._print(expr.args[0])
elif isinstance(expr, cast_func):
return f"&({self._print(expr.args[0])})"
elif isinstance(expr, CastFunc):
cast = "(({data_type})({code}))"
arg, data_type = expr.args
if isinstance(arg, sp.Number):
if arg.is_Number and not isinstance(arg, (sp.core.numbers.Infinity, sp.core.numbers.NegativeInfinity)):
return self._typed_number(arg, data_type)
elif isinstance(arg, (InverseTrigonometricFunction, TrigonometricFunction, HyperbolicFunction)) \
and data_type == BasicType('float32'):
known = self.known_functions[arg.__class__.__name__.lower()]
code = self._print(arg)
return code.replace(known, f"{known}f")
elif isinstance(arg, (sp.Pow, sp.exp)) and data_type == BasicType('float32'):
known = ['sqrt', 'cbrt', 'pow', 'exp']
code = self._print(arg)
for k in known:
if k in code:
return code.replace(k, f'{k}f')
# Powers of small integers are printed as divisions/multiplications.
if '/' in code or '*' in code:
return cast.format(data_type=data_type, code=code)
raise ValueError(f"{code} doesn't give {known=} function back.")
else:
return "((%s)(%s))" % (data_type, self._print(arg))
return cast.format(data_type=data_type, code=self._print(arg))
elif isinstance(expr, fast_division):
return "({})".format(self._print(expr.args[0] / expr.args[1]))
raise ValueError("fast_division is only supported for Taget.GPU")
elif isinstance(expr, fast_sqrt):
return "({})".format(self._print(sp.sqrt(expr.args[0])))
raise ValueError("fast_sqrt is only supported for Taget.GPU")
elif isinstance(expr, fast_inv_sqrt):
raise ValueError("fast_inv_sqrt is only supported for Taget.GPU")
elif isinstance(expr, vec_any) or isinstance(expr, vec_all):
return self._print(expr.args[0])
elif isinstance(expr, fast_inv_sqrt):
return "({})".format(self._print(1 / sp.sqrt(expr.args[0])))
elif isinstance(expr, sp.Abs):
return f"abs({self._print(expr.args[0])})"
elif isinstance(expr, sp.Mod):
if expr.args[0].is_integer and expr.args[1].is_integer:
return f"({self._print(expr.args[0])} % {self._print(expr.args[1])})"
else:
return f"fmod({self._print(expr.args[0])}, {self._print(expr.args[1])})"
elif expr.func in infix_functions:
return "(%s %s %s)" % (self._print(expr.args[0]), infix_functions[expr.func], self._print(expr.args[1]))
return f"({self._print(expr.args[0])} {infix_functions[expr.func]} {self._print(expr.args[1])})"
elif expr.func == int_power_of_2:
return "(1 << (%s))" % (self._print(expr.args[0]))
return f"(1 << ({self._print(expr.args[0])}))"
elif expr.func == int_div:
return "((%s) / (%s))" % (self._print(expr.args[0]), self._print(expr.args[1]))
return f"(({self._print(expr.args[0])}) / ({self._print(expr.args[1])}))"
elif expr.func == DivFunc:
return f'(({self._print(expr.divisor)}) / ({self._print(expr.dividend)}))'
else:
return super(CustomSympyPrinter, self)._print_Function(expr)
name = expr.name if hasattr(expr, 'name') else expr.__class__.__name__
arg_str = ', '.join(self._print(a) for a in expr.args)
return f'{name}({arg_str})'
def _typed_number(self, number, dtype):
res = self._print(number)
......@@ -380,56 +564,39 @@ class CustomSympyPrinter(CCodePrinter):
return res + '.0f' if '.' not in res else res + 'f'
elif dtype.numpy_dtype == np.float64:
return res + '.0' if '.' not in res else res
elif dtype.is_int():
tokens = res.split('.')
if len(tokens) == 1:
return res
elif int(tokens[1]) != 0:
raise ValueError(f"Cannot print non-integer number {res} as an integer.")
else:
return tokens[0]
else:
return res
def _print_Sum(self, expr):
template = """[&]() {{
{dtype} sum = ({dtype}) 0;
for ( {iterator_dtype} {var} = {start}; {condition}; {var} += {increment} ) {{
sum += {expr};
}}
return sum;
}}()"""
var = expr.limits[0][0]
start = expr.limits[0][1]
end = expr.limits[0][2]
code = template.format(
dtype=get_type_of_expression(expr.args[0]),
iterator_dtype='int',
var=self._print(var),
start=self._print(start),
end=self._print(end),
expr=self._print(expr.function),
increment=str(1),
condition=self._print(var) + ' <= ' + self._print(end) # if start < end else '>='
)
return code
def _print_Product(self, expr):
template = """[&]() {{
{dtype} product = ({dtype}) 1;
for ( {iterator_dtype} {var} = {start}; {condition}; {var} += {increment} ) {{
product *= {expr};
}}
return product;
}}()"""
var = expr.limits[0][0]
start = expr.limits[0][1]
end = expr.limits[0][2]
code = template.format(
dtype=get_type_of_expression(expr.args[0]),
iterator_dtype='int',
var=self._print(var),
start=self._print(start),
end=self._print(end),
expr=self._print(expr.function),
increment=str(1),
condition=self._print(var) + ' <= ' + self._print(end) # if start < end else '>='
)
return code
_print_Max = C89CodePrinter._print_Max
_print_Min = C89CodePrinter._print_Min
def _print_ConditionalFieldAccess(self, node):
return self._print(sp.Piecewise((node.outofbounds_value, node.outofbounds_condition), (node.access, True)))
def _print_Max(self, expr):
def inner_print_max(args):
if len(args) == 1:
return self._print(args[0])
half = len(args) // 2
a = inner_print_max(args[:half])
b = inner_print_max(args[half:])
return f"(({a} > {b}) ? {a} : {b})"
return inner_print_max(expr.args)
def _print_Min(self, expr):
def inner_print_min(args):
if len(args) == 1:
return self._print(args[0])
half = len(args) // 2
a = inner_print_min(args[:half])
b = inner_print_min(args[half:])
return f"(({a} < {b}) ? {a} : {b})"
return inner_print_min(expr.args)
# noinspection PyPep8Naming
......@@ -448,41 +615,113 @@ class VectorizedCustomSympyPrinter(CustomSympyPrinter):
assert self.instruction_set['width'] == expr_type.width
return None
def _print_Abs(self, expr):
if isinstance(get_type_of_expression(expr), (VectorType, VectorMemoryAccess)):
return self.instruction_set['abs'].format(self._print(expr.args[0]), **self._kwargs)
return super()._print_Abs(expr)
def _typed_vectorized_number(self, expr, data_type):
basic_data_type = data_type.base_type
number = self._typed_number(expr, basic_data_type)
instruction = 'makeVecConst'
if basic_data_type.is_bool():
instruction = 'makeVecConstBool'
# TODO Vectorization Revamp: is int, or sint, or uint (my guess is sint)
elif basic_data_type.is_int():
instruction = 'makeVecConstInt'
return self.instruction_set[instruction].format(number, **self._kwargs)
def _typed_vectorized_symbol(self, expr, data_type):
if not isinstance(expr, TypedSymbol):
raise ValueError(f'{expr} is not a TypeSymbol. It is {expr.type=}')
basic_data_type = data_type.base_type
symbol = self._print(expr)
if basic_data_type != expr.dtype:
symbol = f'(({basic_data_type})({symbol}))'
instruction = 'makeVecConst'
if basic_data_type.is_bool():
instruction = 'makeVecConstBool'
# TODO Vectorization Revamp: is int, or sint, or uint (my guess is sint)
elif basic_data_type.is_int():
instruction = 'makeVecConstInt'
return self.instruction_set[instruction].format(symbol, **self._kwargs)
def _print_CastFunc(self, expr):
arg, data_type = expr.args
if type(data_type) is VectorType:
base_type = data_type.base_type
# vector_memory_access is a cast_func itself so it should't be directly inside a cast_func
assert not isinstance(arg, VectorMemoryAccess)
if isinstance(arg, sp.Tuple):
is_boolean = get_type_of_expression(arg[0]) == create_type("bool")
is_integer = get_type_of_expression(arg[0]) == create_type("int")
printed_args = [self._print(a) for a in arg]
instruction = 'makeVecBool' if is_boolean else 'makeVecInt' if is_integer else 'makeVec'
if instruction == 'makeVecInt' and 'makeVecIndex' in self.instruction_set:
increments = np.array(arg)[1:] - np.array(arg)[:-1]
if len(set(increments)) == 1:
return self.instruction_set['makeVecIndex'].format(printed_args[0], increments[0],
**self._kwargs)
return self.instruction_set[instruction].format(*printed_args, **self._kwargs)
else:
if arg.is_Number and not isinstance(arg, (sp.core.numbers.Infinity, sp.core.numbers.NegativeInfinity)):
return self._typed_vectorized_number(arg, data_type)
elif isinstance(arg, TypedSymbol):
return self._typed_vectorized_symbol(arg, data_type)
elif isinstance(arg, (InverseTrigonometricFunction, TrigonometricFunction, HyperbolicFunction)) \
and base_type == BasicType('float32'):
raise NotImplementedError('Vectorizer is not tested for trigonometric functions yet')
# known = self.known_functions[arg.__class__.__name__.lower()]
# code = self._print(arg)
# return code.replace(known, f"{known}f")
elif isinstance(arg, sp.Pow):
if base_type == BasicType('float32') or base_type == BasicType('float64'):
return self._print_Pow(arg)
else:
raise NotImplementedError('Integer Pow is not implemented')
elif isinstance(arg, sp.UnevaluatedExpr):
return self._print(arg.args[0])
else:
raise NotImplementedError('Vectorizer cannot cast between different datatypes')
# to_type = self.instruction_set['suffix'][data_type.base_type.c_name]
# from_type = self.instruction_set['suffix'][get_type_of_expression(arg).base_type.c_name]
# return self.instruction_set['cast'].format(from_type, to_type, self._print(arg))
else:
return self._scalarFallback('_print_Function', expr)
# raise ValueError(f'Non VectorType cast "{data_type}" in vectorized code.')
def _print_Function(self, expr):
if isinstance(expr, vector_memory_access):
arg, data_type, aligned, _ = expr.args
if isinstance(expr, VectorMemoryAccess):
arg, data_type, aligned, _, mask, stride = expr.args
if stride != 1:
return self.instruction_set['loadS'].format(f"& {self._print(arg)}", stride, **self._kwargs)
instruction = self.instruction_set['loadA'] if aligned else self.instruction_set['loadU']
return instruction.format("& " + self._print(arg))
elif isinstance(expr, cast_func):
arg, data_type = expr.args
if type(data_type) is VectorType:
return self.instruction_set['makeVec'].format(self._print(arg))
elif expr.func == fast_division:
return instruction.format(f"& {self._print(arg)}", **self._kwargs)
elif expr.func == DivFunc:
result = self._scalarFallback('_print_Function', expr)
if not result:
result = self.instruction_set['/'].format(self._print(expr.args[0]), self._print(expr.args[1]))
result = self.instruction_set['/'].format(self._print(expr.divisor), self._print(expr.dividend),
**self._kwargs)
return result
elif expr.func == fast_sqrt:
return "({})".format(self._print(sp.sqrt(expr.args[0])))
elif expr.func == fast_inv_sqrt:
result = self._scalarFallback('_print_Function', expr)
if not result:
if self.instruction_set['rsqrt']:
return self.instruction_set['rsqrt'].format(self._print(expr.args[0]))
else:
return "({})".format(self._print(1 / sp.sqrt(expr.args[0])))
elif isinstance(expr, vec_any):
expr_type = get_type_of_expression(expr.args[0])
if type(expr_type) is not VectorType:
return self._print(expr.args[0])
else:
return self.instruction_set['any'].format(self._print(expr.args[0]))
elif isinstance(expr, vec_all):
elif isinstance(expr, fast_division):
raise ValueError("fast_division is only supported for Taget.GPU")
elif isinstance(expr, fast_sqrt):
raise ValueError("fast_sqrt is only supported for Taget.GPU")
elif isinstance(expr, fast_inv_sqrt):
raise ValueError("fast_inv_sqrt is only supported for Taget.GPU")
elif isinstance(expr, vec_any) or isinstance(expr, vec_all):
instr = 'any' if isinstance(expr, vec_any) else 'all'
expr_type = get_type_of_expression(expr.args[0])
if type(expr_type) is not VectorType:
return self._print(expr.args[0])
else:
return self.instruction_set['all'].format(self._print(expr.args[0]))
if isinstance(expr.args[0], sp.Rel):
op = expr.args[0].rel_op
if (instr, op) in self.instruction_set:
return self.instruction_set[(instr, op)].format(*[self._print(a) for a in expr.args[0].args],
**self._kwargs)
return self.instruction_set[instr].format(self._print(expr.args[0]), **self._kwargs)
return super(VectorizedCustomSympyPrinter, self)._print_Function(expr)
......@@ -495,7 +734,7 @@ class VectorizedCustomSympyPrinter(CustomSympyPrinter):
assert len(arg_strings) > 0
result = arg_strings[0]
for item in arg_strings[1:]:
result = self.instruction_set['&'].format(result, item)
result = self.instruction_set['&'].format(result, item, **self._kwargs)
return result
def _print_Or(self, expr):
......@@ -507,16 +746,31 @@ class VectorizedCustomSympyPrinter(CustomSympyPrinter):
assert len(arg_strings) > 0
result = arg_strings[0]
for item in arg_strings[1:]:
result = self.instruction_set['|'].format(result, item)
result = self.instruction_set['|'].format(result, item, **self._kwargs)
return result
def _print_Add(self, expr, order=None):
result = self._scalarFallback('_print_Add', expr)
try:
result = self._scalarFallback('_print_Add', expr)
except Exception:
result = None
if result:
return result
args = expr.args
# special treatment for all-integer args, for loop index arithmetic until we have proper int vectorization
suffix = ""
if all([(type(e) is CastFunc and str(e.dtype) == self.instruction_set['int']) or isinstance(e, sp.Integer)
or (type(e) is TypedSymbol and isinstance(e.dtype, BasicType) and e.dtype.is_int()) for e in args]):
dtype = set([e.dtype for e in args if type(e) is CastFunc])
assert len(dtype) == 1
dtype = dtype.pop()
args = [CastFunc(e, dtype) if (isinstance(e, sp.Integer) or isinstance(e, TypedSymbol)) else e
for e in args]
suffix = "int"
summands = []
for term in expr.args:
for term in args:
if term.func == sp.Mul:
sign, t = self._print_Mul(term, inside_add=True)
else:
......@@ -532,30 +786,36 @@ class VectorizedCustomSympyPrinter(CustomSympyPrinter):
assert len(summands) >= 2
processed = summands[0].term
for summand in summands[1:]:
func = self.instruction_set['-'] if summand.sign == -1 else self.instruction_set['+']
processed = func.format(processed, summand.term)
func = self.instruction_set['-' + suffix] if summand.sign == -1 else self.instruction_set['+' + suffix]
processed = func.format(processed, summand.term, **self._kwargs)
return processed
def _print_Pow(self, expr):
result = self._scalarFallback('_print_Pow', expr)
# Due to loop cutting sp.Mul is evaluated again.
try:
result = self._scalarFallback('_print_Pow', expr)
except ValueError:
result = None
if result:
return result
one = self.instruction_set['makeVec'].format(1.0)
if expr.exp.is_integer and expr.exp.is_number and 0 < expr.exp < 8:
return "(" + self._print(sp.Mul(*[expr.base] * expr.exp, evaluate=False)) + ")"
elif expr.exp == -1:
one = self.instruction_set['makeVec'].format(1.0)
return self.instruction_set['/'].format(one, self._print(expr.base))
elif expr.exp == 0.5:
return self.instruction_set['sqrt'].format(self._print(expr.base))
elif expr.exp == -0.5:
root = self.instruction_set['sqrt'].format(self._print(expr.base))
return self.instruction_set['/'].format(one, root)
elif expr.exp.is_integer and expr.exp.is_number and - 8 < expr.exp < 0:
return self.instruction_set['/'].format(one,
self._print(sp.Mul(*[expr.base] * (-expr.exp), evaluate=False)))
one = self.instruction_set['makeVecConst'].format(1.0, **self._kwargs)
root = self.instruction_set['sqrt'].format(self._print(expr.base), **self._kwargs)
if isinstance(expr.exp, CastFunc) and expr.exp.args[0].is_number:
exp = expr.exp.args[0]
else:
exp = expr.exp
# TODO the printer should not have any intelligence like this.
# TODO To remove all of these cases the vectoriser needs to be reworked. See loop cutting
if exp.is_integer and exp.is_number and 0 < exp < 8:
return self._print(sp.Mul(*[expr.base] * exp, evaluate=False))
elif exp == 0.5:
return root
elif exp == -0.5:
return self.instruction_set['/'].format(one, root, **self._kwargs)
else:
raise ValueError("Generic exponential not supported: " + str(expr))
......@@ -563,7 +823,10 @@ class VectorizedCustomSympyPrinter(CustomSympyPrinter):
# noinspection PyProtectedMember
from sympy.core.mul import _keep_coeff
result = self._scalarFallback('_print_Mul', expr)
if not inside_add:
result = self._scalarFallback('_print_Mul', expr)
else:
result = None
if result:
return result
......@@ -594,19 +857,19 @@ class VectorizedCustomSympyPrinter(CustomSympyPrinter):
result = a_str[0]
for item in a_str[1:]:
result = self.instruction_set['*'].format(result, item)
result = self.instruction_set['*'].format(result, item, **self._kwargs)
if len(b) > 0:
denominator_str = b_str[0]
for item in b_str[1:]:
denominator_str = self.instruction_set['*'].format(denominator_str, item)
result = self.instruction_set['/'].format(result, denominator_str)
denominator_str = self.instruction_set['*'].format(denominator_str, item, **self._kwargs)
result = self.instruction_set['/'].format(result, denominator_str, **self._kwargs)
if inside_add:
return sign, result
else:
if sign < 0:
return self.instruction_set['*'].format(self._print(S.NegativeOne), result)
return self.instruction_set['*'].format(self._print(S.NegativeOne), result, **self._kwargs)
else:
return result
......@@ -614,13 +877,13 @@ class VectorizedCustomSympyPrinter(CustomSympyPrinter):
result = self._scalarFallback('_print_Relational', expr)
if result:
return result
return self.instruction_set[expr.rel_op].format(self._print(expr.lhs), self._print(expr.rhs))
return self.instruction_set[expr.rel_op].format(self._print(expr.lhs), self._print(expr.rhs), **self._kwargs)
def _print_Equality(self, expr):
result = self._scalarFallback('_print_Equality', expr)
if result:
return result
return self.instruction_set['=='].format(self._print(expr.lhs), self._print(expr.rhs))
return self.instruction_set['=='].format(self._print(expr.lhs), self._print(expr.rhs), **self._kwargs)
def _print_Piecewise(self, expr):
result = self._scalarFallback('_print_Piecewise', expr)
......@@ -638,13 +901,11 @@ class VectorizedCustomSympyPrinter(CustomSympyPrinter):
result = self._print(expr.args[-1][0])
for true_expr, condition in reversed(expr.args[:-1]):
if isinstance(condition, cast_func) and get_type_of_expression(condition.args[0]) == create_type("bool"):
if not KERNCRAFT_NO_TERNARY_MODE:
result = "(({}) ? ({}) : ({}))".format(self._print(condition.args[0]), self._print(true_expr),
result)
else:
print("Warning - skipping ternary op")
if isinstance(condition, CastFunc) and get_type_of_expression(condition.args[0]) == create_type("bool"):
result = "(({}) ? ({}) : ({}))".format(self._print(condition.args[0]), self._print(true_expr),
result, **self._kwargs)
else:
# noinspection SpellCheckingInspection
result = self.instruction_set['blendv'].format(result, self._print(true_expr), self._print(condition))
result = self.instruction_set['blendv'].format(result, self._print(true_expr), self._print(condition),
**self._kwargs)
return result