def test_dependent_loop_bounds_4():
# https://gitlab.tiker.net/inducer/loopy/issues/23
import loopy as lp
loopy_knl = lp.make_kernel(
[
"{[a]: 0<=a<10}",
"{[b]: b_start<=b<b_end}",
"{[c,idim]: c_start<=c<c_end and 0<=idim<dim}",
],
"""
for a
<> b_start = 1
<> b_end = 2
for b
<> c_start = 1
<> c_end = 2
for c
... nop
end
<>t[idim] = 1
end
end
""",
"...",
seq_dependencies=True)
loopy_knl = lp.fix_parameters(loopy_knl, dim=3)
with lp.CacheMode(False):
lp.generate_code_v2(loopy_knl)
def test_ispc_streaming_stores():
stream_dtype = np.float32
index_dtype = np.int32
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"a[i] = b[i] + scalar * c[i]",
target=lp.ISPCTarget(), index_dtype=index_dtype,
name="stream_triad")
vars = ["a", "b", "c", "scalar"]
knl = lp.assume(knl, "n>0")
knl = lp.split_iname(
knl, "i", 2**18, outer_tag="g.0", slabs=(0, 1))
knl = lp.split_iname(knl, "i_inner", 8, inner_tag="l.0")
knl = lp.tag_instructions(knl, "!streaming_store")
knl = lp.add_and_infer_dtypes(knl, {
var: stream_dtype
for var in vars
})
knl = lp.set_argument_order(knl, vars + ["n"])
knl = lp.preprocess_kernel(knl)
knl = lp.get_one_scheduled_kernel(knl)
lp.generate_code_v2(knl).all_code()
def test_check_bounds_with_caller_assumptions(ctx_factory):
import islpy as isl
from loopy.diagnostic import LoopyIndexError
arange = lp.make_function("{[i]: 0<=i<n}",
"""
y[i] = i
""",
name="arange")
knl = lp.make_kernel(
"{[i]: 0<=i<20}",
"""
[i]: Y[i] = arange(N)
""",
[lp.GlobalArg("Y", shape=(20, )),
lp.ValueArg("N", dtype=np.int32)],
name="epoint")
knl = lp.merge([knl, arange])
with pytest.raises(LoopyIndexError):
lp.generate_code_v2(knl)
knl = knl.with_kernel(
lp.assume(knl.default_entrypoint, isl.BasicSet("[N] -> { : N <= 20}")))
lp.auto_test_vs_ref(knl, ctx_factory(), parameters={"N": 15})
def test_dependent_loop_bounds_3(ctx_factory):
# The point of this test is that it shows a dependency between
# domains that is exclusively mediated by the row_len temporary.
# It also makes sure that row_len gets read before any
# conditionals use it.
dtype = np.dtype(np.float32)
ctx = ctx_factory()
knl = lp.make_kernel([
"{[i]: 0<=i<n}",
"{[jj]: 0<=jj<row_len}",
], [
"<> row_len = a_row_lengths[i]",
"a[i,jj] = 1",
], [
lp.GlobalArg("a_row_lengths", np.int32, shape=lp.auto),
lp.GlobalArg("a", dtype, shape=("n,n"), order="C"),
lp.ValueArg("n", np.int32),
],
target=lp.PyOpenCLTarget(ctx.devices[0]),
name="loopy_kernel")
assert knl["loopy_kernel"].parents_per_domain()[1] == 0
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
print(lp.generate_code_v2(knl).device_code())
knl_bad = lp.split_iname(knl, "jj", 128, outer_tag="g.1", inner_tag="l.1")
with pytest.raises(RuntimeError):
list(lp.generate_code_v2(knl_bad))
def test_ispc_streaming_stores():
stream_dtype = np.float32
index_dtype = np.int32
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"a[i] = b[i] + scalar * c[i]",
target=lp.ISPCTarget(), index_dtype=index_dtype,
name="stream_triad")
vars = ["a", "b", "c", "scalar"]
knl = lp.assume(knl, "n>0")
knl = lp.split_iname(
knl, "i", 2**18, outer_tag="g.0", slabs=(0, 1))
knl = lp.split_iname(knl, "i_inner", 8, inner_tag="l.0")
knl = lp.tag_instructions(knl, "!streaming_store")
knl = lp.add_and_infer_dtypes(knl, {
var: stream_dtype
for var in vars
})
knl = lp.set_argument_order(knl, vars + ["n"])
knl = lp.preprocess_kernel(knl)
knl = lp.get_one_scheduled_kernel(knl)
lp.generate_code_v2(knl).all_code()
def test_dependent_loop_bounds_4():
# https://gitlab.tiker.net/inducer/loopy/issues/23
import loopy as lp
loopy_knl = lp.make_kernel([
"{[a]: 0<=a<10}",
"{[b]: b_start<=b<b_end}",
"{[c,idim]: c_start<=c<c_end and 0<=idim<dim}",
],
"""
for a
<> b_start = 1
<> b_end = 2
for b
<> c_start = 1
<> c_end = 2
for c
... nop
end
<>t[idim] = 1
end
end
""",
"...",
seq_dependencies=True)
loopy_knl = lp.fix_parameters(loopy_knl, dim=3)
with lp.CacheMode(False):
lp.generate_code_v2(loopy_knl)
def test_double_hw_axes_used_in_knl_call(inline):
from loopy.diagnostic import LoopyError
twice = lp.make_function("{[i]: 0<=i<10}",
"""
y[i] = 2*x[i]
""",
name="twice")
knl = lp.make_kernel("{[i]: 0<=i<10}",
"""
y[:, i] = twice(x[:, i])
""", [
lp.GlobalArg("x", shape=(10, 10), dtype=float),
lp.GlobalArg("y", shape=(10, 10))
],
name="outer")
twice = lp.tag_inames(twice, {"i": "l.0"})
knl = lp.tag_inames(knl, {"i": "l.0"})
knl = lp.merge([knl, twice])
if inline:
knl = lp.inline_callable_kernel(knl, "twice")
with pytest.raises(LoopyError):
lp.generate_code_v2(knl)
def test_idempotency(form):
k1 = compile_form(form)[0]
k2 = compile_form(form)[0]
assert k1.ast.gencode() == k2.ast.gencode()
# Test loopy backend
import loopy
k1 = compile_form(form, coffee=False)[0]
k2 = compile_form(form, coffee=False)[0]
assert loopy.generate_code_v2(
k1.ast).device_code() == loopy.generate_code_v2(k2.ast).device_code()
def test_lpy_wide_array_splitter(opts):
from pymbolic.primitives import Subscript, Variable
# create array split
asplit = array_splitter(opts)
# create a test kernel
arg1 = lp.GlobalArg('a1', shape=(10, 10), order=opts.order)
arg2 = lp.GlobalArg('a2', shape=(16, 16), order=opts.order)
k = lp.make_kernel([
'{[i]: 0 <= i < 10}', '{{[j_outer]: 0 <= j_outer < {}}}'.format(
int(np.ceil(10 / VECTOR_WIDTH))),
'{{[j_inner]: 0 <= j_inner < {}}}'.format(VECTOR_WIDTH)
],
"""
for i, j_outer, j_inner
a1[j_outer, i] = 1 {id=a1}
a2[j_outer, i] = 1 {id=a2}
end
""", [arg1, arg2],
silenced_warnings=['no_device_in_pre_codegen_checks'],
target=lp.OpenCLTarget())
a1_hold = k.arg_dict['a1'].copy()
a2_hold = k.arg_dict['a2'].copy()
k = asplit.split_loopy_arrays(k)
k = lp.tag_inames(k, {'j_inner': 'l.0' if not opts.is_simd else 'vec'})
# ensure there's no loopy errors
lp.generate_code_v2(k).device_code()
def __indexer():
if opts.order == 'C':
return (Variable('j_outer'), Variable('i'), Variable('j_inner'))
else:
return (Variable('j_inner'), Variable('j_outer'), Variable('i'))
# check dim
a1 = k.arg_dict['a1']
assert a1.shape == asplit.split_shape(a1_hold)[0]
# and indexing
assign = next(insn.assignee for insn in k.instructions if insn.id == 'a1')
# construct index
assert isinstance(assign, Subscript) and assign.index == __indexer()
# now test with evenly sized
a2 = k.arg_dict['a2']
assert a2.shape == asplit.split_shape(a2_hold)[0]
assign = next(insn.assignee for insn in k.instructions if insn.id == 'a2')
assert isinstance(assign, Subscript) and assign.index == __indexer()
def test_lpy_deep_array_splitter(opts):
from pymbolic.primitives import Subscript, Variable
# create array split
asplit = array_splitter(opts)
# create a test kernel
size = VECTOR_WIDTH * 3
loop_bound = VECTOR_WIDTH * 2
arg1 = lp.GlobalArg('a1', shape=(size, size), order=opts.order)
arg2 = lp.GlobalArg('a2', shape=(16, 16), order=opts.order)
k = lp.make_kernel('{{[i]: 0 <= i < {}}}'.format(loop_bound),
"""
a1[0, i] = 1 {id=a1}
a2[0, i] = 1 {id=a2}
""", [arg1, arg2],
silenced_warnings=['no_device_in_pre_codegen_checks'],
target=lp.OpenCLTarget())
k = lp.split_iname(k,
'i',
VECTOR_WIDTH,
inner_tag='l.0' if not opts.is_simd else 'vec')
a1_hold = k.arg_dict['a1'].copy()
a2_hold = k.arg_dict['a2'].copy()
k = asplit.split_loopy_arrays(k)
# ensure there's no loopy errors
lp.generate_code_v2(k).device_code()
def __indexer():
if opts.order == 'C':
return (0, Variable('i_outer'), Variable('i_inner'))
else:
return (Variable('i_inner'), 0, Variable('i_outer'))
# check dim
a1 = k.arg_dict['a1']
assert a1.shape == asplit.split_shape(a1_hold)[0]
# and indexing
assign = next(insn.assignee for insn in k.instructions if insn.id == 'a1')
# construct index
assert isinstance(assign, Subscript) and assign.index == __indexer()
# now test with evenly sized
a2 = k.arg_dict['a2']
assert a2.shape == asplit.split_shape(a2_hold)[0]
assign = next(insn.assignee for insn in k.instructions if insn.id == 'a2')
assert isinstance(assign, Subscript) and assign.index == __indexer()
def test_global_temporary(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{ [i]: 0<=i<n}",
"""
<> c[i] = a[i + 1]
out[i] = c[i]
""")
knl = lp.add_and_infer_dtypes(knl,
{"a": np.float32, "c": np.float32, "out": np.float32, "n": np.int32})
knl = lp.set_temporary_scope(knl, "c", "global")
ref_knl = knl
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
cgr = lp.generate_code_v2(knl)
assert len(cgr.device_programs) == 2
#print(cgr.device_code())
#print(cgr.host_code())
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=5))
def test_global_temporary(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{ [i]: 0<=i<n}", """
<> c[i] = a[i + 1]
out[i] = c[i]
""")
knl = lp.add_and_infer_dtypes(knl, {
"a": np.float32,
"c": np.float32,
"out": np.float32,
"n": np.int32
})
knl = lp.set_temporary_scope(knl, "c", "global")
ref_knl = knl
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
cgr = lp.generate_code_v2(knl)
assert len(cgr.device_programs) == 2
#print(cgr.device_code())
#print(cgr.host_code())
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=5))
def test_c_execution_with_global_temporaries():
# ensure that the "host" code of a bare ExecutableCTarget with
# global constant temporaries is None
from loopy.target.c import ExecutableCTarget
AS = lp.AddressSpace # noqa
n = 10
knl = lp.make_kernel(
"{[i]: 0 <= i < n}",
"""
a[i] = b[i]
""", [
lp.GlobalArg("a", shape=(n, ), dtype=np.int32),
lp.TemporaryVariable("b",
shape=(n, ),
initializer=np.arange(n, dtype=np.int32),
dtype=np.int32,
read_only=True,
address_space=AS.GLOBAL)
],
target=ExecutableCTarget())
knl = lp.fix_parameters(knl, n=n)
assert ("int b[%d]" % n) not in lp.generate_code_v2(knl).host_code()
assert np.allclose(knl(a=np.zeros(10, dtype=np.int32))[1], np.arange(10))
def test_missing_compilers():
from loopy.target.c import ExecutableCTarget, CTarget
from loopy.target.c.c_execution import CCompiler
from codepy.toolchain import GCCToolchain
def __test(evalfunc, target, **targetargs):
n = 10
knl = lp.make_kernel(
"{[i]: 0 <= i < n}",
"""
a[i] = b[i]
""", [
lp.GlobalArg("a", shape=(n, ), dtype=np.int32),
lp.GlobalArg("b", shape=(n, ), dtype=np.int32)
],
target=target(**targetargs))
knl = lp.fix_parameters(knl, n=n)
return evalfunc(knl)
assert __test(lambda knl: lp.generate_code_v2(knl).device_code(), CTarget)
from pytools.prefork import ExecError
def eval_tester(knl):
return np.allclose(
knl(a=np.zeros(10, dtype=np.int32),
b=np.arange(10, dtype=np.int32))[1], np.arange(10))
import os
path_store = os.environ["PATH"]
ccomp = None
try:
# test with path wiped out such that we can't find gcc
with pytest.raises(ExecError):
os.environ["PATH"] = ""
ccomp = CCompiler()
__test(eval_tester, ExecutableCTarget, compiler=ccomp)
finally:
# make sure we restore the path
os.environ["PATH"] = path_store
# and, with the path restored we should now be able to properly execute with
# the default (non-guessed) toolchain!
__test(eval_tester, ExecutableCTarget, compiler=ccomp)
# and test that we will fail if we remove a required attribute
del ccomp.toolchain.undefines
with pytest.raises(AttributeError):
__test(eval_tester, ExecutableCTarget, compiler=ccomp)
# next test that some made up compiler can be specified
ccomp = CCompiler(cc="foo")
assert isinstance(ccomp.toolchain, GCCToolchain)
assert ccomp.toolchain.cc == "foo"
# and that said made up compiler errors out
with pytest.raises(ExecError):
__test(eval_tester, ExecutableCTarget, compiler=ccomp)
def make_kernels(self, Vf, Vc):
"""
Interpolation and restriction kernels between arbitrary elements.
This is temporary while we wait for structure-preserving tfsc kernels.
"""
self.prolong_kernel = self.prolongation_transfer_kernel_action(
Vf, self.uc)
matrix_kernel = self.prolongation_transfer_kernel_action(
Vf, firedrake.TestFunction(Vc))
# The way we transpose the prolongation kernel is suboptimal.
# A local matrix is generated each time the kernel is executed.
element_kernel = loopy.generate_code_v2(
matrix_kernel.code).device_code()
element_kernel = element_kernel.replace(
"void expression_kernel", "static void expression_kernel")
dimc = Vc.finat_element.space_dimension() * Vc.value_size
dimf = Vf.finat_element.space_dimension() * Vf.value_size
restrict_code = f"""
{element_kernel}
void restriction({ScalarType_c} *restrict Rc, const {ScalarType_c} *restrict Rf, const {ScalarType_c} *restrict w)
{{
{ScalarType_c} Afc[{dimf}*{dimc}] = {{0}};
expression_kernel(Afc);
for ({IntType_c} i = 0; i < {dimf}; i++)
for ({IntType_c} j = 0; j < {dimc}; j++)
Rc[j] += Afc[i*{dimc} + j] * Rf[i] * w[i];
}}
"""
self.restrict_kernel = op2.Kernel(restrict_code, "restriction")
def test_recursive_nested_dependent_reduction(ctx_factory):
dtype = np.dtype(np.int32)
ctx = ctx_factory()
knl = lp.make_kernel([
"{[itgt]: 0 <= itgt < ntgts}", "{[isrc_box]: 0 <= isrc_box < nboxes}",
"{[isrc]: 0 <= isrc < npart}"
],
"""
for itgt
for isrc_box
<> npart = nparticles_per_box[isrc_box]
<> boxsum = sum(isrc, isrc+isrc_box+itgt)
end
a[itgt] = sum(isrc_box, boxsum)
end
""", [
lp.ValueArg("n", np.int32),
lp.GlobalArg("a", dtype, ("n", )),
lp.GlobalArg("nparticles_per_box", np.int32,
("nboxes", )),
lp.ValueArg("ntgts", np.int32),
lp.ValueArg("nboxes", np.int32),
],
assumptions="ntgts>=1",
target=lp.PyOpenCLTarget(ctx.devices[0]))
print(lp.generate_code_v2(knl).device_code())
def test_c_execution_with_global_temporaries():
# ensure that the "host" code of a bare ExecutableCTarget with
# global constant temporaries is None
from loopy.target.c import ExecutableCTarget
from loopy.kernel.data import temp_var_scope as scopes
n = 10
knl = lp.make_kernel(
'{[i]: 0 <= i < n}',
"""
a[i] = b[i]
""", [
lp.GlobalArg('a', shape=(n, ), dtype=np.int32),
lp.TemporaryVariable('b',
shape=(n, ),
initializer=np.arange(n, dtype=np.int32),
dtype=np.int32,
read_only=True,
scope=scopes.GLOBAL)
],
target=ExecutableCTarget())
knl = lp.fix_parameters(knl, n=n)
assert ('int b[%d]' % n) not in lp.generate_code_v2(knl).host_code()
assert np.allclose(knl(a=np.zeros(10, dtype=np.int32))[1], np.arange(10))
def test_automatic_scan_detection():
knl = lp.make_kernel(["[n] -> {[i]: 0<=i<n}", "{[j]: 0<=j<=2*i}"], """
a[i] = sum(j, j**2)
""")
cgr = lp.generate_code_v2(knl)
assert "scan" not in cgr.device_code()
def generate_single_cell_wrapper(iterset, args, forward_args=(), kernel_name=None, wrapper_name=None):
"""Generates wrapper for a single cell. No iteration loop, but cellwise data is extracted.
Cell is expected as an argument to the wrapper. For extruded, the numbering of the cells
is columnwise continuous, bottom to top.
:param iterset: The iteration set
:param args: :class:`Arg`s
:param forward_args: To forward unprocessed arguments to the kernel via the wrapper,
give an iterable of strings describing their C types.
:param kernel_name: Kernel function name
:param wrapper_name: Wrapper function name
:return: string containing the C code for the single-cell wrapper
"""
from pyop2.codegen.builder import WrapperBuilder
from pyop2.codegen.rep2loopy import generate
from loopy.types import OpaqueType
forward_arg_types = [OpaqueType(fa) for fa in forward_args]
builder = WrapperBuilder(iterset=iterset, single_cell=True, forward_arg_types=forward_arg_types)
for arg in args:
builder.add_argument(arg)
builder.set_kernel(Kernel("", kernel_name))
wrapper = generate(builder, wrapper_name)
code = loopy.generate_code_v2(wrapper)
return code.device_code()
def test_kernel_splitting_with_loop(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{ [i,k]: 0<=i<n and 0<=k<3 }",
"""
c[k,i] = a[k, i + 1]
out[k,i] = c[k,i]
""")
knl = lp.add_and_infer_dtypes(knl,
{"a": np.float32, "c": np.float32, "out": np.float32, "n": np.int32})
ref_knl = knl
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
# schedule
from loopy.preprocess import preprocess_kernel
knl = preprocess_kernel(knl)
from loopy.schedule import get_one_scheduled_kernel
knl = get_one_scheduled_kernel(knl)
# map schedule onto host or device
print(knl)
cgr = lp.generate_code_v2(knl)
assert len(cgr.device_programs) == 2
print(cgr.device_code())
print(cgr.host_code())
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=5))
def generate_single_cell_wrapper(iterset,
args,
forward_args=(),
kernel_name=None,
wrapper_name=None):
"""Generates wrapper for a single cell. No iteration loop, but cellwise data is extracted.
Cell is expected as an argument to the wrapper. For extruded, the numbering of the cells
is columnwise continuous, bottom to top.
:param iterset: The iteration set
:param args: :class:`Arg`s
:param forward_args: To forward unprocessed arguments to the kernel via the wrapper,
give an iterable of strings describing their C types.
:param kernel_name: Kernel function name
:param wrapper_name: Wrapper function name
:return: string containing the C code for the single-cell wrapper
"""
from pyop2.codegen.builder import WrapperBuilder
from pyop2.codegen.rep2loopy import generate
from loopy.types import OpaqueType
forward_arg_types = [OpaqueType(fa) for fa in forward_args]
builder = WrapperBuilder(iterset=iterset,
single_cell=True,
forward_arg_types=forward_arg_types)
for arg in args:
builder.add_argument(arg)
builder.set_kernel(Kernel("", kernel_name))
wrapper = generate(builder, wrapper_name)
code = loopy.generate_code_v2(wrapper)
return code.device_code()
def build_loopy_kernel_A_text():
knl_name = "kernel_tensor_A"
knl = lp.make_kernel("{ [i,j,k]: 0<=i,j<n and 0<=k<m }",
"""
A[i,j] = c*sum(k, B[k,i]*B[k,j])
""",
name=knl_name,
assumptions="n >= 1 and m >= 1",
lang_version=lp.MOST_RECENT_LANGUAGE_VERSION,
target=lp.CTarget())
knl = lp.add_and_infer_dtypes(
knl, {
"A": np.dtype(np.double),
"B": np.dtype(np.double),
"c": np.dtype(np.double)
})
knl = lp.fix_parameters(knl, n=3, m=2)
knl = lp.prioritize_loops(knl, "i,j")
#print(knl)
knl_c, knl_h = lp.generate_code_v2(knl).device_code(), str(
lp.generate_header(knl)[0])
replacements = [("__restrict__", "restrict")]
knl_c = utils.replace_strings(knl_c, replacements)
knl_h = utils.replace_strings(knl_h, replacements)
knl_call = "kernel_tensor_A(A, &B[0][0], 1.0/(2.0*Ae));"
return knl_name, knl_call, knl_c, knl_h
def build_loopy_kernel_b_text():
knl_name = "kernel_tensor_b"
knl = lp.make_kernel("{ [i]: 0<=i<n }",
"""
b[i] = c
""",
name="kernel_tensor_b",
lang_version=lp.MOST_RECENT_LANGUAGE_VERSION,
target=lp.CTarget())
knl = lp.add_and_infer_dtypes(knl, {
"b": np.dtype(np.double),
"c": np.dtype(np.double)
})
knl = lp.fix_parameters(knl, n=3)
#print(knl)
knl_c, knl_h = lp.generate_code_v2(knl).device_code(), str(
lp.generate_header(knl)[0])
replacements = [("__restrict__", "restrict")]
knl_c = utils.replace_strings(knl_c, replacements)
knl_h = utils.replace_strings(knl_h, replacements)
knl_call = "kernel_tensor_b(b, Ae / 6.0);"
return knl_name, knl_call, knl_c, knl_h
def test_generate_c_snippet():
from pymbolic import var
I = var("I") # noqa
f = var("f")
df = var("df")
q_v = var("q_v")
eN = var("eN") # noqa
k = var("k")
u = var("u")
from functools import partial
l_sum = partial(lp.Reduction, "sum", allow_simultaneous=True)
Instr = lp.Assignment # noqa
knl = lp.make_kernel("{[I, k]: 0<=I<nSpace and 0<=k<nQuad}", [
Instr(f[I], l_sum(k, q_v[k, I] * u)),
Instr(df[I], l_sum(k, q_v[k, I])),
], [
lp.GlobalArg("q_v", np.float64, shape="nQuad, nSpace"),
lp.GlobalArg("f,df", np.float64, shape="nSpace"),
lp.ValueArg("u", np.float64),
"...",
],
target=CTarget(),
assumptions="nQuad>=1")
if 0: # enable to play with prefetching
# (prefetch currently requires constant sizes)
knl = lp.fix_parameters(knl, nQuad=5, nSpace=3)
knl = lp.add_prefetch(knl, "q_v", "k,I", default_tag=None)
knl = lp.split_iname(knl, "k", 4, inner_tag="unr", slabs=(0, 1))
knl = lp.prioritize_loops(knl, "I,k_outer,k_inner")
print(lp.generate_code_v2(knl))
def test_solve_callable(self, zero_vec, solve_mat, solve_vec):
loopy.set_caching_enabled(False)
k = loopy.make_kernel(
["{[i,j] : 0 <= i,j < 2}"],
"""
x[:] = solve(A[:,:], b[:])
""", [
loopy.GlobalArg('x', dtype=np.float64, shape=(2, )),
loopy.GlobalArg('A', dtype=np.float64, shape=(2, 2)),
loopy.GlobalArg(
'b',
dtype=np.float64,
shape=(2, ),
)
],
target=loopy.CTarget(),
name="callable_kernel2",
lang_version=(2018, 2))
k = loopy.register_function_id_to_in_knl_callable_mapper(
k, solve_fn_lookup)
code = loopy.generate_code_v2(k).device_code()
code.replace('void callable_kernel2', 'static void callable_kernel2')
loopykernel = op2.Kernel(code, k.name, ldargs=["-llapack"])
args = [zero_vec(op2.READ), solve_mat(op2.READ), solve_vec(op2.WRITE)]
op2.par_loop(loopykernel, solve_mat.dataset.set, *args)
expected = np.linalg.solve(solve_mat.data, solve_vec.data)
assert np.allclose(expected, zero_vec.data)
def test_inverse_callable(self, zero_mat, inv_mat):
loopy.set_caching_enabled(False)
k = loopy.make_kernel(
["{[i,j] : 0 <= i,j < 2}"],
"""
B[:,:] = inv(A[:,:])
""", [
loopy.GlobalArg('B', dtype=np.float64, shape=(2, 2)),
loopy.GlobalArg('A', dtype=np.float64, shape=(2, 2))
],
target=loopy.CTarget(),
name="callable_kernel",
lang_version=(2018, 2))
k = loopy.register_function_id_to_in_knl_callable_mapper(
k, inv_fn_lookup)
code = loopy.generate_code_v2(k).device_code()
code.replace('void callable_kernel', 'static void callable_kernel')
loopykernel = op2.Kernel(code, k.name, ldargs=["-llapack"])
op2.par_loop(loopykernel, zero_mat.dataset.set, zero_mat(op2.WRITE),
inv_mat(op2.READ))
expected = np.linalg.inv(inv_mat.data)
assert np.allclose(expected, zero_mat.data)
def loopy_example():
knl = lp.make_kernel("{ [i]: 0<=i<n }",
"out[i] = 2*a[i]",
lang_version=lp.MOST_RECENT_LANGUAGE_VERSION,
target=lp.CTarget())
knl = lp.add_and_infer_dtypes(knl, {"a": np.dtype(np.float32)})
print(lp.generate_code_v2(knl).device_code())
def test_numba_target():
knl = lp.make_kernel("{[i,j,k]: 0<=i,j<M and 0<=k<N}",
"D[i,j] = sqrt(sum(k, (X[i, k]-X[j, k])**2))",
target=lp.NumbaTarget())
knl = lp.add_and_infer_dtypes(knl, {"X": np.float32})
print(lp.generate_code_v2(knl).device_code())
def make_domain(prog):
knl_code = lp.generate_code_v2(prog)
for id, v in knl_code.implemented_domains.items():
# print(v[0].compute_schedule())
print(v[0].to_str())
print(dir(v[0]))
def write_ptx(ctx, knl, filename=None):
cl_program = cl.Program(ctx,
lp.generate_code_v2(knl).device_code()).build(
options=knl.options.cl_build_options)
ptx_src = cl_program.binaries[0]
if not filename:
filename = "ptx_" + knl.name + ".ptx"
ptx_src_file = open(filename, 'w')
ptx_src_file.write(ptx_src.decode('utf-8', 'ignore'))
def test_numba_target():
knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j<M and 0<=k<N}",
"D[i,j] = sqrt(sum(k, (X[i, k]-X[j, k])**2))",
target=lp.NumbaTarget())
knl = lp.add_and_infer_dtypes(knl, {"X": np.float32})
print(lp.generate_code_v2(knl).device_code())
def test_missing_compilers():
from loopy.target.c import ExecutableCTarget, CTarget
from loopy.target.c.c_execution import CCompiler
from codepy.toolchain import GCCToolchain
def __test(evalfunc, target, **targetargs):
n = 10
knl = lp.make_kernel('{[i]: 0 <= i < n}',
"""
a[i] = b[i]
""",
[lp.GlobalArg('a', shape=(n,), dtype=np.int32),
lp.GlobalArg('b', shape=(n,), dtype=np.int32)],
target=target(**targetargs))
knl = lp.fix_parameters(knl, n=n)
return evalfunc(knl)
assert __test(lambda knl: lp.generate_code_v2(knl).device_code(), CTarget)
from pytools.prefork import ExecError
def eval_tester(knl):
return np.allclose(knl(a=np.zeros(10, dtype=np.int32),
b=np.arange(10, dtype=np.int32))[1], np.arange(10))
import os
path_store = os.environ["PATH"]
ccomp = None
try:
# test with path wiped out such that we can't find gcc
with pytest.raises(ExecError):
os.environ["PATH"] = ''
ccomp = CCompiler()
__test(eval_tester, ExecutableCTarget, compiler=ccomp)
finally:
# make sure we restore the path
os.environ["PATH"] = path_store
# and, with the path restored we should now be able to properly execute with
# the default (non-guessed) toolchain!
__test(eval_tester, ExecutableCTarget, compiler=ccomp)
# and test that we will fail if we remove a required attribute
del ccomp.toolchain.undefines
with pytest.raises(AttributeError):
__test(eval_tester, ExecutableCTarget, compiler=ccomp)
# next test that some made up compiler can be specified
ccomp = CCompiler(cc='foo')
assert isinstance(ccomp.toolchain, GCCToolchain)
assert ccomp.toolchain.cc == 'foo'
# and that said made up compiler errors out
with pytest.raises(ExecError):
__test(eval_tester, ExecutableCTarget, compiler=ccomp)
def test_triangle_domain(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel("{[i,j]: 0<=i,j<n and i <= j}",
"a[i,j] = 17",
assumptions="n>=1",
target=lp.PyOpenCLTarget(ctx.devices[0]))
print(knl)
print(lp.generate_code_v2(knl).device_code())
def test_math_function(target, tp):
# Test correct maths functions are generated for C and OpenCL
# backend instead for different data type
data_type = {"f32": np.float32, "f64": np.float64}[tp]
import pymbolic.primitives as p
i = p.Variable("i")
xi = p.Subscript(p.Variable("x"), i)
yi = p.Subscript(p.Variable("y"), i)
zi = p.Subscript(p.Variable("z"), i)
n = 100
domain = "{[i]: 0<=i<%d}" % n
data = [
lp.GlobalArg("x", data_type, shape=(n, )),
lp.GlobalArg("y", data_type, shape=(n, )),
lp.GlobalArg("z", data_type, shape=(n, ))
]
inst = [lp.Assignment(xi, p.Variable("min")(yi, zi))]
knl = lp.make_kernel(domain, inst, data, target=target())
code = lp.generate_code_v2(knl).device_code()
assert "fmin" in code
if tp == "f32" and target == CTarget:
assert "fminf" in code
else:
assert "fminf" not in code
inst = [lp.Assignment(xi, p.Variable("max")(yi, zi))]
knl = lp.make_kernel(domain, inst, data, target=target())
code = lp.generate_code_v2(knl).device_code()
assert "fmax" in code
if tp == "f32" and target == CTarget:
assert "fmaxf" in code
else:
assert "fmaxf" not in code
def test_lpy_iname_presplit(opts):
"""
Tests that inames access to pre-split inames in non-split loopy arrays are
correctly handled
"""
from pymbolic.primitives import Subscript, Variable
# create array split
asplit = array_splitter(opts)
# create a test kernel
arg1 = lp.GlobalArg('a1', shape=(20, 10), order=opts.order)
arg2 = lp.GlobalArg('a2', shape=(16, 16), order=opts.order)
k = lp.make_kernel([
'{[i]: 0 <= i < 10}', '{{[j_outer]: 0 <= j_outer < {}}}'.format(
int(np.ceil(10 / VECTOR_WIDTH))),
'{{[j_inner]: 0 <= j_inner < {}}}'.format(VECTOR_WIDTH)
],
"""
a1[j_outer, i] = 1 {id=a1}
a2[j_outer, i] = 1 {id=a2}
""", [arg1, arg2],
silenced_warnings=['no_device_in_pre_codegen_checks'],
target=lp.OpenCLTarget())
k = asplit.split_loopy_arrays(k, dont_split=['a1', 'a2'])
# ensure there's no loopy errors
lp.generate_code_v2(k).device_code()
def __indexer():
return (Variable('j_outer') * VECTOR_WIDTH + Variable('j_inner'),
Variable('i'))
# check indexing
assign = next(insn.assignee for insn in k.instructions if insn.id == 'a1')
# construct index
assert isinstance(assign, Subscript) and assign.index == __indexer()
# now test with evenly sized
assign = next(insn.assignee for insn in k.instructions if insn.id == 'a2')
assert isinstance(assign, Subscript) and assign.index == __indexer()
def test_math_function(target, tp):
# Test correct maths functions are generated for C and OpenCL
# backend instead for different data type
data_type = {"f32": np.float32,
"f64": np.float64}[tp]
import pymbolic.primitives as p
i = p.Variable("i")
xi = p.Subscript(p.Variable("x"), i)
yi = p.Subscript(p.Variable("y"), i)
zi = p.Subscript(p.Variable("z"), i)
n = 100
domain = "{[i]: 0<=i<%d}" % n
data = [lp.GlobalArg("x", data_type, shape=(n,)),
lp.GlobalArg("y", data_type, shape=(n,)),
lp.GlobalArg("z", data_type, shape=(n,))]
inst = [lp.Assignment(xi, p.Variable("min")(yi, zi))]
knl = lp.make_kernel(domain, inst, data, target=target())
code = lp.generate_code_v2(knl).device_code()
assert "fmin" in code
if tp == "f32" and target == CTarget:
assert "fminf" in code
else:
assert "fminf" not in code
inst = [lp.Assignment(xi, p.Variable("max")(yi, zi))]
knl = lp.make_kernel(domain, inst, data, target=target())
code = lp.generate_code_v2(knl).device_code()
assert "fmax" in code
if tp == "f32" and target == CTarget:
assert "fmaxf" in code
else:
assert "fmaxf" not in code
def test_rob_stroud_bernstein(ctx_factory):
ctx = ctx_factory()
# NOTE: tmp would have to be zero-filled beforehand
knl = lp.make_kernel(
"{[el, i2, alpha1,alpha2]: \
0 <= el < nels and \
0 <= i2 < nqp1d and \
0 <= alpha1 <= deg and 0 <= alpha2 <= deg-alpha1 }",
"""
for el,i2
<> xi = qpts[1, i2]
<> s = 1-xi
<> r = xi/s
<> aind = 0 {id=aind_init}
for alpha1
<> w = s**(deg-alpha1) {id=init_w}
for alpha2
tmp[el,alpha1,i2] = tmp[el,alpha1,i2] + w * coeffs[aind] \
{id=write_tmp,dep=init_w:aind_init}
w = w * r * ( deg - alpha1 - alpha2 ) / (1 + alpha2) \
{id=update_w,dep=init_w:write_tmp}
aind = aind + 1 \
{id=aind_incr,dep=aind_init:write_tmp:update_w}
end
end
end
""",
[
# Must declare coeffs to have "no" shape, to keep loopy
# from trying to figure it out the shape automatically.
lp.GlobalArg("coeffs", None, shape=None),
"..."
],
assumptions="deg>=0 and nels>=1",
target=lp.PyOpenCLTarget(ctx.devices[0])
)
knl = lp.fix_parameters(knl, nqp1d=7, deg=4)
knl = lp.split_iname(knl, "el", 16, inner_tag="l.0")
knl = lp.split_iname(knl, "el_outer", 2, outer_tag="g.0", inner_tag="ilp",
slabs=(0, 1))
knl = lp.tag_inames(knl, dict(i2="l.1", alpha1="unr", alpha2="unr"))
knl = lp.add_dtypes(knl, dict(
qpts=np.float32,
coeffs=np.float32,
tmp=np.float32,
))
print(lp.generate_code_v2(knl))
def create_rand_args(ctx, knl, param_dict):
queue = cl.CommandQueue(ctx)
info = lp.generate_code_v2(knl).implemented_data_info
args, arg_data = lp.auto_test.make_ref_args(knl, info, queue,
param_dict)
args.clear()
del args
rand_args = lp.auto_test.make_args(knl, info, queue, arg_data,
param_dict)
del arg_data[:]
del arg_data
return rand_args
def test_automatic_scan_detection():
knl = lp.make_kernel(
[
"[n] -> {[i]: 0<=i<n}",
"{[j]: 0<=j<=2*i}"
],
"""
a[i] = sum(j, j**2)
"""
)
cgr = lp.generate_code_v2(knl)
assert "scan" not in cgr.device_code()
def test_cuda_short_vector():
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i] = 2*a[i]",
target=lp.CudaTarget())
knl = lp.set_options(knl, write_code=True)
knl = lp.split_iname(knl, "i", 4, slabs=(0, 1), inner_tag="vec")
knl = lp.split_array_axis(knl, "a,out", axis_nr=0, count=4)
knl = lp.tag_array_axes(knl, "a,out", "C,vec")
knl = lp.set_options(knl, write_wrapper=True)
knl = lp.add_and_infer_dtypes(knl, {"a": np.float32})
print(lp.generate_code_v2(knl).device_code())
def test_numba_cuda_target():
knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j<M and 0<=k<N}",
"D[i,j] = sqrt(sum(k, (X[i, k]-X[j, k])**2))",
target=lp.NumbaCudaTarget())
knl = lp.assume(knl, "M>0")
knl = lp.split_iname(knl, "i", 16, outer_tag='g.0')
knl = lp.split_iname(knl, "j", 128, inner_tag='l.0', slabs=(0, 1))
knl = lp.add_prefetch(knl, "X[i,:]", default_tag="l.auto")
knl = lp.fix_parameters(knl, N=3)
knl = lp.prioritize_loops(knl, "i_inner,j_outer")
knl = lp.tag_inames(knl, "k:unr")
knl = lp.tag_array_axes(knl, "X", "N0,N1")
knl = lp.add_and_infer_dtypes(knl, {"X": np.float32})
print(lp.generate_code_v2(knl).all_code())
def code_to_compile(self):
from pyop2.codegen.builder import WrapperBuilder
from pyop2.codegen.rep2loopy import generate
builder = WrapperBuilder(iterset=self._iterset, iteration_region=self._iteration_region, pass_layer_to_kernel=self._pass_layer_arg)
for arg in self._args:
builder.add_argument(arg)
builder.set_kernel(self._kernel)
wrapper = generate(builder)
code = loopy.generate_code_v2(wrapper)
if self._kernel._cpp:
from loopy.codegen.result import process_preambles
preamble = "".join(process_preambles(getattr(code, "device_preambles", [])))
device_code = "\n\n".join(str(dp.ast) for dp in code.device_programs)
return preamble + "\nextern \"C\" {\n" + device_code + "\n}\n"
return code.device_code()
def test_reduction_with_conditional():
# The purpose of the 'l' iname is to force the entire kernel (including the
# predicate) into device code.
knl = lp.make_kernel(
"{ [l,i] : 0<=l,i<42 }",
"""
if l > 0
b[l] = sum(i, l*a[i])
end
""",
[lp.ValueArg("n", dtype=np.int32), "..."])
knl = lp.tag_inames(knl, "l:g.0")
knl = lp.add_and_infer_dtypes(knl, {"a": np.float32})
code = lp.generate_code_v2(knl).device_code()
print(code)
# Check that the if appears before the loop that realizes the reduction.
assert code.index("if") < code.index("for")
def test_ispc_target(occa_mode=False):
from loopy.target.ispc import ISPCTarget
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i] = 2*a[i]",
[
lp.GlobalArg("out,a", np.float32, shape=lp.auto),
"..."
],
target=ISPCTarget(occa_mode=occa_mode))
knl = lp.split_iname(knl, "i", 8, inner_tag="l.0")
knl = lp.split_iname(knl, "i_outer", 4, outer_tag="g.0", inner_tag="ilp")
knl = lp.add_prefetch(knl, "a", ["i_inner", "i_outer_inner"])
codegen_result = lp.generate_code_v2(
lp.get_one_scheduled_kernel(
lp.preprocess_kernel(knl)))
print(codegen_result.device_code())
print(codegen_result.host_code())
def test_c_execution_with_global_temporaries():
# ensure that the "host" code of a bare ExecutableCTarget with
# global constant temporaries is None
from loopy.target.c import ExecutableCTarget
from loopy.kernel.data import temp_var_scope as scopes
n = 10
knl = lp.make_kernel('{[i]: 0 <= i < n}',
"""
a[i] = b[i]
""",
[lp.GlobalArg('a', shape=(n,), dtype=np.int32),
lp.TemporaryVariable('b', shape=(n,),
initializer=np.arange(n, dtype=np.int32),
dtype=np.int32,
read_only=True,
scope=scopes.GLOBAL)],
target=ExecutableCTarget())
knl = lp.fix_parameters(knl, n=n)
assert ('int b[%d]' % n) not in lp.generate_code_v2(knl).host_code()
assert np.allclose(knl(a=np.zeros(10, dtype=np.int32))[1], np.arange(10))
def gen_code(knl):
knl = lp.preprocess_kernel(knl)
knl = lp.get_one_scheduled_kernel(knl)
codegen_result = lp.generate_code_v2(knl)
return codegen_result.device_code() + "\n" + codegen_result.host_code()
import loopy as lp
import pyopencl as cl
import pyopencl.array
from loopy.version import LOOPY_USE_LANGUAGE_VERSION_2018_2
# setup
# -----
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
n = 15 * 10**6
a = cl.array.arange(queue, n, dtype=np.float32)
# create
# ------
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i] = 2*a[i]")
# transform
# ---------
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
# execute
# -------
evt, (out,) = knl(queue, a=a)
# ENDEXAMPLE
knl = lp.add_and_infer_dtypes(knl, {"a": np.dtype(np.float32)})
print(lp.generate_code_v2(knl).device_code())
knl = lp.make_kernel(
"{ [i,k]: 0<=i<n and 0<=k<3 }",
"""
for i, k
... gbarrier
c[k,i] = a[k, i + 1]
... gbarrier
out[k,i] = c[k,i]
end
""", seq_dependencies=True)
# transform
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
knl = lp.add_and_infer_dtypes(knl,
{"a": np.float32, "c": np.float32, "out": np.float32, "n": np.int32})
# schedule
from loopy.preprocess import preprocess_kernel
knl = preprocess_kernel(knl)
from loopy.schedule import get_one_scheduled_kernel
knl = get_one_scheduled_kernel(knl)
# map schedule onto host or device
print(knl)
cgr = lp.generate_code_v2(knl)
print(cgr.device_code())
print(cgr.host_code())