-
Michael Kuron authored
[FIX] Vector Size for SVE instruction set is in free kernel parameters See merge request !412
Michael Kuron authored[FIX] Vector Size for SVE instruction set is in free kernel parameters See merge request !412
test_vectorization_specific.py 14.35 KiB
import pytest
import numpy as np
import pystencils.config
import sympy as sp
import pystencils as ps
from pystencils.backends.simd_instruction_sets import (get_cacheline_size, get_supported_instruction_sets,
get_vector_instruction_set)
from pystencils.enums import Target
from pystencils.typing import CFunction
from . import test_vectorization
supported_instruction_sets = get_supported_instruction_sets() if get_supported_instruction_sets() else []
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
def test_vectorisation_varying_arch(instruction_set):
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
config = pystencils.config.CreateKernelConfig(cpu_vectorize_info={'instruction_set': instruction_set})
ast = ps.create_kernel(update_rule, config=config)
kernel = ast.compile()
kernel(f=arr)
np.testing.assert_equal(arr, 2)
@pytest.mark.parametrize('dtype', ('float32', 'float64'))
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
def test_vectorized_abs_field(instruction_set, dtype):
"""Some instructions sets have abs, some don't.
Furthermore, the special treatment of unary minus makes this data type-sensitive too.
"""
arr = np.ones((2 ** 2 + 2, 2 ** 3 + 2), dtype=dtype)
arr[-3:, :] = -1
f, g = ps.fields(f=arr, g=arr)
update_rule = [ps.Assignment(g.center(), sp.Abs(f.center()))]
config = pystencils.config.CreateKernelConfig(cpu_vectorize_info={'instruction_set': instruction_set})
ast = ps.create_kernel(update_rule, config=config)
func = ast.compile()
dst = np.zeros_like(arr)
func(g=dst, f=arr)
np.testing.assert_equal(np.sum(dst[1:-1, 1:-1]), 2 ** 2 * 2 ** 3)
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
def test_vectorized_abs_scalar(instruction_set):
"""Some instructions sets have abs, some don't.
Furthermore, the special treatment of unary minus makes this data type-sensitive too.
"""
arr = np.zeros((2 ** 2 + 2, 2 ** 3 + 2), dtype="float64")
f = ps.fields(f=arr)
update_rule = [ps.Assignment(f.center(), sp.Abs(sp.Symbol("a")))]
config = pystencils.config.CreateKernelConfig(cpu_vectorize_info={'instruction_set': instruction_set})
ast = ps.create_kernel(update_rule, config=config)
func = ast.compile()
func(f=arr, a=-1)
np.testing.assert_equal(np.sum(arr[1:-1, 1:-1]), 2 ** 2 * 2 ** 3)
@pytest.mark.parametrize('dtype', ('float32', 'float64'))
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
@pytest.mark.parametrize('nontemporal', [False, True])
def test_strided(instruction_set, dtype, nontemporal):
f, g = ps.fields(f"f, g : {dtype}[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)]
config = pystencils.config.CreateKernelConfig(cpu_vectorize_info={'instruction_set': instruction_set,
'nontemporal': nontemporal},
default_number_float=dtype)
if 'storeS' not in get_vector_instruction_set(dtype, instruction_set) \
and instruction_set not in ['avx512', 'avx512vl', 'rvv'] and not instruction_set.startswith('sve'):
with pytest.warns(UserWarning) as warn:
ast = ps.create_kernel(update_rule, config=config)
assert 'Could not vectorize loop' in warn[0].message.args[0]
else:
with pytest.warns(None) as warn:
ast = ps.create_kernel(update_rule, config=config)
assert len(warn) == 0
instruction = 'streamS' if nontemporal and 'streamS' in ast.instruction_set else 'storeS'
assert ast.instruction_set[instruction].split('{')[0] in ps.get_code_str(ast)
instruction = 'cachelineZero'
if instruction in ast.instruction_set:
assert ast.instruction_set[instruction] not in ps.get_code_str(ast)
# ps.show_code(ast)
func = ast.compile()
ref_config = pystencils.config.CreateKernelConfig(default_number_float=dtype)
ref_func = ps.create_kernel(update_rule, config=ref_config).compile()
# For some reason other array creations fail on the emulated ppc pipeline
size = (25, 19)
arr = np.zeros(size).astype(dtype)
for i in range(size[0]):
for j in range(size[1]):
arr[i, j] = i * j
dst = np.zeros_like(arr, dtype=dtype)
ref = np.zeros_like(arr, dtype=dtype)
func(g=dst, f=arr)
ref_func(g=ref, f=arr)
# print("dst: ", dst)
# print("np array: ", arr)
np.testing.assert_almost_equal(dst[1:-1, 1:-1], ref[1:-1, 1:-1], 13 if dtype == 'float64' else 5)
@pytest.mark.parametrize('dtype', ('float32', 'float64'))
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
@pytest.mark.parametrize('gl_field, gl_kernel', [(1, 0), (0, 1), (1, 1)])
def test_alignment_and_correct_ghost_layers(gl_field, gl_kernel, instruction_set, dtype):
domain_size = (128, 128)
dh = ps.create_data_handling(domain_size, periodicity=(True, True), default_target=Target.CPU)
src = dh.add_array("src", values_per_cell=1, dtype=dtype, ghost_layers=gl_field, alignment=True)
dh.fill(src.name, 1.0, ghost_layers=True)
dst = dh.add_array("dst", values_per_cell=1, dtype=dtype, ghost_layers=gl_field, alignment=True)
dh.fill(dst.name, 1.0, ghost_layers=True)
update_rule = ps.Assignment(dst[0, 0], src[0, 0])
opt = {'instruction_set': instruction_set, 'assume_aligned': True,
'nontemporal': True, 'assume_inner_stride_one': True}
config = pystencils.config.CreateKernelConfig(target=dh.default_target,
cpu_vectorize_info=opt, ghost_layers=gl_kernel)
ast = ps.create_kernel(update_rule, config=config)
kernel = ast.compile()
if ('loadA' in ast.instruction_set or 'storeA' in ast.instruction_set) and gl_kernel != gl_field:
with pytest.raises(ValueError):
dh.run_kernel(kernel)
else:
dh.run_kernel(kernel)
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
def test_cacheline_size(instruction_set):
cacheline_size = get_cacheline_size(instruction_set)
if cacheline_size is None and instruction_set in ['sse', 'avx', 'avx512', 'avx512vl', 'rvv']:
pytest.skip()
instruction_set = get_vector_instruction_set('double', instruction_set)
vector_size = instruction_set['bytes']
assert 8 < cacheline_size < 0x100000, "Cache line size is implausible"
if type(vector_size) is int:
assert cacheline_size % vector_size == 0, "Cache line size should be multiple of vector size"
assert cacheline_size & (cacheline_size - 1) == 0, "Cache line size is not a power of 2"
# TODO move to vectorise
@pytest.mark.parametrize('instruction_set',
sorted(set(supported_instruction_sets) - {test_vectorization.instruction_set}))
@pytest.mark.parametrize('function',
[f for f in test_vectorization.__dict__ if f.startswith('test_') and f not in ['test_hardware_query', 'test_aligned_and_nt_stores']])
def test_vectorization_other(instruction_set, function):
test_vectorization.__dict__[function](instruction_set)
@pytest.mark.parametrize('dtype', ('float32', 'float64'))
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
@pytest.mark.parametrize('field_layout', ('fzyx', 'zyxf'))
def test_square_root(dtype, instruction_set, field_layout):
config = pystencils.config.CreateKernelConfig(data_type=dtype,
default_number_float=dtype,
cpu_vectorize_info={'instruction_set': instruction_set,
'assume_inner_stride_one': True,
'assume_aligned': False,
'nontemporal': False})
src_field = ps.Field.create_generic('pdfs', 2, dtype, index_dimensions=1, layout=field_layout, index_shape=(9,))
eq = [ps.Assignment(sp.Symbol("xi"), sum(src_field.center_vector)),
ps.Assignment(sp.Symbol("xi_2"), sp.Symbol("xi") * sp.sqrt(src_field.center))]
ast = ps.create_kernel(eq, config=config)
ast.compile()
code = ps.get_code_str(ast)
print(code)
@pytest.mark.parametrize('dtype', ('float32', 'float64'))
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
@pytest.mark.parametrize('padding', (True, False))
def test_square_root_2(dtype, instruction_set, padding):
x, y = sp.symbols("x y")
src = ps.fields(f"src: {dtype}[2D]", layout='fzyx')
up = ps.Assignment(src[0, 0], 1 / x + sp.sqrt(y * 0.52 + x ** 2))
cpu_vec = {'instruction_set': instruction_set, 'assume_inner_stride_one': True,
'assume_sufficient_line_padding': padding,
'assume_aligned': True}
config = ps.CreateKernelConfig(data_type=dtype, default_number_float=dtype, cpu_vectorize_info=cpu_vec)
ast = ps.create_kernel(up, config=config)
ast.compile()
code = ps.get_code_str(ast)
print(code)
@pytest.mark.parametrize('dtype', ('float32', 'float64'))
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
@pytest.mark.parametrize('padding', (True, False))
def test_pow(dtype, instruction_set, padding):
config = pystencils.config.CreateKernelConfig(data_type=dtype,
default_number_float=dtype,
cpu_vectorize_info={'instruction_set': instruction_set,
'assume_inner_stride_one': True,
'assume_sufficient_line_padding': padding,
'assume_aligned': False, 'nontemporal': False})
src_field = ps.Field.create_generic('pdfs', 2, dtype, index_dimensions=1, layout='fzyx', index_shape=(9,))
eq = [ps.Assignment(sp.Symbol("xi"), sum(src_field.center_vector)),
ps.Assignment(sp.Symbol("xi_2"), sp.Symbol("xi") * sp.Pow(src_field.center, 0.5))]
ast = ps.create_kernel(eq, config=config)
ast.compile()
code = ps.get_code_str(ast)
print(code)
@pytest.mark.parametrize('dtype', ('float32', 'float64'))
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
@pytest.mark.parametrize('padding', (True, False))
def test_issue62(dtype, instruction_set, padding):
opt = {'instruction_set': instruction_set, 'assume_aligned': True,
'assume_inner_stride_one': True,
'assume_sufficient_line_padding': padding}
dx = sp.Symbol("dx")
dy = sp.Symbol("dy")
src, dst, rhs = ps.fields(f"src, src_tmp, rhs: {dtype}[2D]", layout='fzyx')
up = ps.Assignment(dst[0, 0], ((dy ** 2 * (src[1, 0] + src[-1, 0]))
+ (dx ** 2 * (src[0, 1] + src[0, -1]))
- (rhs[0, 0] * dx ** 2 * dy ** 2)) / (2 * (dx ** 2 + dy ** 2)))
config = ps.CreateKernelConfig(data_type=dtype,
default_number_float=dtype,
cpu_vectorize_info=opt)
ast = ps.create_kernel(up, config=config)
ast.compile()
code = ps.get_code_str(ast)
print(code)
assert 'pow' not in code
@pytest.mark.parametrize('dtype', ('float32', 'float64'))
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
def test_div_and_unevaluated_expr(dtype, instruction_set):
opt = {'instruction_set': instruction_set, 'assume_aligned': True, 'assume_inner_stride_one': True,
'assume_sufficient_line_padding': False}
x, y, z = sp.symbols("x y z")
rhs = (-4 * x ** 4 * y ** 2 * z ** 2 + (x ** 4 * y ** 2 / 3) + (x ** 4 * z ** 2 / 3)) / x ** 3
src = ps.fields(f"src: {dtype}[2D]", layout='fzyx')
up = ps.Assignment(src[0, 0], rhs)
config = ps.CreateKernelConfig(data_type=dtype,
default_number_float=dtype,
cpu_vectorize_info=opt)
ast = ps.create_kernel(up, config=config)
code = ps.get_code_str(ast)
# print(code)
ast.compile()
assert 'pow' not in code
@pytest.mark.parametrize('dtype', ('float32', 'float64'))
@pytest.mark.parametrize('instruction_set', ('sve', 'sve2', 'sme', 'rvv'))
def test_check_ast_parameters_sizeless(dtype, instruction_set):
f, g = ps.fields(f"f, g: {dtype}[3D]", layout='fzyx')
update_rule = [ps.Assignment(g.center(), 2 * f.center())]
config = pystencils.config.CreateKernelConfig(data_type=dtype,
cpu_vectorize_info={'instruction_set': instruction_set})
ast = ps.create_kernel(update_rule, config=config)
ast_symbols = [p.symbol for p in ast.get_parameters()]
assert ast.instruction_set['width'] not in ast_symbols
assert ast.instruction_set['intwidth'] not in ast_symbols
# TODO this test case needs a complete rework of the vectoriser. The reason is that the vectoriser does not
# TODO vectorise symbols at the moment because they could be strides or field sizes, thus involved in pointer arithmetic
# TODO This means that the vectoriser only works if fields are involved on the rhs.
# @pytest.mark.parametrize('dtype', ('float32', 'float64'))
# @pytest.mark.parametrize('instruction_set', supported_instruction_sets)
# def test_vectorised_symbols(dtype, instruction_set):
# opt = {'instruction_set': instruction_set, 'assume_aligned': True, 'assume_inner_stride_one': True,
# 'assume_sufficient_line_padding': False}
#
# x, y, z = sp.symbols("x y z")
# rhs = 1 / x ** 2 * (x + y)
#
# src = ps.fields(f"src: {dtype}[2D]", layout='fzyx')
#
# up = ps.Assignment(src[0, 0], rhs)
#
# config = ps.CreateKernelConfig(data_type=dtype,
# default_number_float=dtype,
# cpu_vectorize_info=opt)
#
# ast = ps.create_kernel(up, config=config)
# code = ps.get_code_str(ast)
# print(code)
#
# ast.compile()
#
# assert 'pow' not in code