Newer
Older
from __future__ import division
__copyright__ = "Copyright (C) 2012 Andreas Kloeckner"
__license__ = """
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
"""
import numpy as np
import loopy as lp
import pyopencl as cl
from pyopencl.tools import pytest_generate_tests_for_pyopencl \
as pytest_generate_tests
__all__ = ["pytest_generate_tests",
"cl" # 'cl.create_some_context'
]
Andreas Klöckner
committed
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
def test_complicated_subst(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[i]: 0<=i<n}",
"""
f(x) := x*a[x]
g(x) := 12 + f(x)
h(x) := 1 + g(x) + 20*g$two(x)
a[i] = h$one(i) * h$two(i)
""",
[
lp.GlobalArg("a", np.float32, shape=("n",)),
lp.ValueArg("n", np.int32),
])
from loopy.subst import expand_subst
knl = expand_subst(knl, "g$two < h$two")
print knl
sr_keys = knl.substitutions.keys()
for letter, how_many in [
("f", 1),
("g", 1),
("h", 2)
]:
substs_with_letter = sum(1 for k in sr_keys if k.startswith(letter))
assert substs_with_letter == how_many
def test_type_inference_no_artificial_doubles(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[i]: 0<=i<n}",
"""
<> bb = a[i] - b[i]
c[i] = bb
""",
[
lp.GlobalArg("a", np.float32, shape=("n",)),
lp.GlobalArg("b", np.float32, shape=("n",)),
lp.GlobalArg("c", np.float32, shape=("n",)),
lp.ValueArg("n", np.int32),
],
assumptions="n>=1")
for k in lp.generate_loop_schedules(knl):
code = lp.generate_code(k)
assert "double" not in code
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
def test_sized_and_complex_literals(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[i]: 0<=i<n}",
"""
<> aa = 5jf
<> bb = 5j
a[i] = imag(aa)
b[i] = imag(bb)
c[i] = 5f
""",
[
lp.GlobalArg("a", np.float32, shape=("n",)),
lp.GlobalArg("b", np.float32, shape=("n",)),
lp.GlobalArg("c", np.float32, shape=("n",)),
lp.ValueArg("n", np.int32),
],
assumptions="n>=1")
lp.auto_test_vs_ref(knl, ctx, lp.generate_loop_schedules(knl),
parameters=dict(n=5))
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
def test_simple_side_effect(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[i,j]: 0<=i,j<100}",
"""
a[i] = a[i] + 1
""",
[lp.GlobalArg("a", np.float32, shape=(100,))]
)
kernel_gen = lp.generate_loop_schedules(knl)
kernel_gen = lp.check_kernels(kernel_gen)
for gen_knl in kernel_gen:
print gen_knl
compiled = lp.CompiledKernel(ctx, gen_knl)
print compiled.code
def test_nonsense_reduction(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[i]: 0<=i<100}",
"""
a[i] = sum(i, 2)
""",
[lp.GlobalArg("a", np.float32, shape=(100,))]
)
import pytest
with pytest.raises(RuntimeError):
list(lp.generate_loop_schedules(knl))
def test_owed_barriers(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[i]: 0<=i<100}",
[
Andreas Klöckner
committed
"<float32> z[i] = a[i]"
],
[lp.GlobalArg("a", np.float32, shape=(100,))]
Andreas Klöckner
committed
knl = lp.tag_inames(knl, dict(i="l.0"))
kernel_gen = lp.generate_loop_schedules(knl)
kernel_gen = lp.check_kernels(kernel_gen)
for gen_knl in kernel_gen:
compiled = lp.CompiledKernel(ctx, gen_knl)
print compiled.code
Andreas Klöckner
committed
def test_wg_too_small(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[i]: 0<=i<100}",
[
Andreas Klöckner
committed
"<float32> z[i] = a[i] {id=copy}"
Andreas Klöckner
committed
],
[lp.GlobalArg("a", np.float32, shape=(100,))],
Andreas Klöckner
committed
local_sizes={0: 16})
Andreas Klöckner
committed
knl = lp.tag_inames(knl, dict(i="l.0"))
Andreas Klöckner
committed
kernel_gen = lp.generate_loop_schedules(knl)
kernel_gen = lp.check_kernels(kernel_gen)
for gen_knl in kernel_gen:
try:
lp.CompiledKernel(ctx, gen_knl).get_code()
Andreas Klöckner
committed
except RuntimeError, e:
assert "implemented and desired" in str(e)
pass # expected!
else:
assert False # expecting an error
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
def test_join_inames(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[i,j]: 0<=i,j<16}",
[
"b[i,j] = 2*a[i,j]"
],
[
lp.GlobalArg("a", np.float32, shape=(16, 16,)),
lp.GlobalArg("b", np.float32, shape=(16, 16,))
],
)
ref_knl = knl
knl = lp.add_prefetch(knl, "a", sweep_inames=["i", "j"])
knl = lp.join_inames(knl, ["a_dim_0", "a_dim_1"])
kernel_gen = lp.generate_loop_schedules(knl)
kernel_gen = lp.check_kernels(kernel_gen)
lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen)
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
def test_divisibility_assumption(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"[n] -> {[i]: 0<=i<n}",
[
"b[i] = 2*a[i]"
],
[
lp.GlobalArg("a", np.float32, shape=("n",)),
lp.GlobalArg("b", np.float32, shape=("n",)),
lp.ValueArg("n", np.int32),
],
assumptions="n>=1 and (exists zz: n = 16*zz)")
ref_knl = knl
knl = lp.split_iname(knl, "i", 16)
for k in lp.generate_loop_schedules(knl):
code = lp.generate_code(k)
assert "if" not in code
kernel_gen = lp.generate_loop_schedules(knl)
kernel_gen = lp.check_kernels(kernel_gen)
lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen,
parameters={"n": 16**3})
def test_multi_cse(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[i]: 0<=i<100}",
[
Andreas Klöckner
committed
"<float32> z[i] = a[i] + a[i]**2"
],
[lp.GlobalArg("a", np.float32, shape=(100,))],
local_sizes={0: 16})
knl = lp.split_iname(knl, "i", 16, inner_tag="l.0")
knl = lp.add_prefetch(knl, "a", [])
kernel_gen = lp.generate_loop_schedules(knl)
kernel_gen = lp.check_kernels(kernel_gen)
for gen_knl in kernel_gen:
compiled = lp.CompiledKernel(ctx, gen_knl)
print compiled.code
def test_stencil(ctx_factory):
ctx = ctx_factory()
# n=32 causes corner case behavior in size calculations for temprorary (a
# non-unifiable, two-constant-segments PwAff as the base index)
n = 256
knl = lp.make_kernel(ctx.devices[0],
"{[i,j]: 0<= i,j < %d}" % n,
"a_offset(ii, jj) := a[ii+1, jj+1]",
"z[i,j] = -2*a_offset(i,j)"
" + a_offset(i,j-1)"
" + a_offset(i,j+1)"
" + a_offset(i-1,j)"
" + a_offset(i+1,j)"
lp.GlobalArg("a", np.float32, shape=(n+2,n+2,)),
lp.GlobalArg("z", np.float32, shape=(n+2,n+2,))
def variant_1(knl):
knl = lp.split_iname(knl, "i", 16, outer_tag="g.1", inner_tag="l.1")
knl = lp.split_iname(knl, "j", 16, outer_tag="g.0", inner_tag="l.0")
knl = lp.add_prefetch(knl, "a", ["i_inner", "j_inner"])
for variant in [variant_1]:
kernel_gen = lp.generate_loop_schedules(variant(knl),
loop_priority=["i_outer", "i_inner_0", "j_0"])
kernel_gen = lp.check_kernels(kernel_gen)
lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen,
fills_entire_output=False, print_ref_code=True,
op_count=[n*n], op_label=["cells"])
def test_eq_constraint(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[i,j]: 0<= i,j < 32}",
[
"a[i] = b[i]"
],
[
lp.GlobalArg("a", np.float32, shape=(1000,)),
lp.GlobalArg("b", np.float32, shape=(1000,))
knl = lp.split_iname(knl, "i", 16, outer_tag="g.0")
knl = lp.split_iname(knl, "i_inner", 16, outer_tag=None, inner_tag="l.0")
kernel_gen = lp.generate_loop_schedules(knl)
kernel_gen = lp.check_kernels(kernel_gen)
for knl in kernel_gen:
print lp.generate_code(knl)
def test_argmax(ctx_factory):
dtype = np.dtype(np.float32)
ctx = ctx_factory()
order = "C"
n = 10000
knl = lp.make_kernel(ctx.devices[0],
"{[i]: 0<=i<%d}" % n,
[
"<> result = argmax(i, fabs(a[i]))",
"max_idx = result.index",
"max_val = result.value",
],
[
lp.GlobalArg("a", dtype, shape=(n,), order=order),
lp.GlobalArg("max_idx", np.int32, shape=(), order=order),
lp.GlobalArg("max_val", dtype, shape=(), order=order),
])
a = np.random.randn(10000).astype(dtype)
cknl = lp.CompiledKernel(ctx, knl)
evt, (max_idx, max_val) = cknl(queue, a=a, out_host=True)
assert max_val == np.max(np.abs(a))
assert max_idx == np.where(np.abs(a)==max_val)[-1]
def make_random_value():
from random import randrange, uniform
v = randrange(3)
if v == 0:
while True:
z = randrange(-1000, 1000)
if z:
return z
elif v == 1:
return uniform(-10, 10)
else:
cval = uniform(-10, 10) + 1j*uniform(-10, 10)
if randrange(0, 2) == 0:
return np.complex128(cval)
else:
return np.complex128(cval)
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
def make_random_expression(var_values, size):
from random import randrange
import pymbolic.primitives as p
v = randrange(1500)
size[0] += 1
if v < 500 and size[0] < 40:
term_count = randrange(2, 5)
if randrange(2) < 1:
cls = p.Sum
else:
cls = p.Product
return cls(tuple(
make_random_expression(var_values, size)
for i in range(term_count)))
elif v < 750:
return make_random_value()
elif v < 1000:
var_name = "var_%d" % len(var_values)
assert var_name not in var_values
var_values[var_name] = make_random_value()
return p.Variable(var_name)
elif v < 1250:
# Cannot use '-' because that destroys numpy constants.
return p.Sum((
make_random_expression(var_values, size),
- make_random_expression(var_values, size)))
elif v < 1500:
# Cannot use '/' because that destroys numpy constants.
return p.Quotient(
make_random_expression(var_values, size),
make_random_expression(var_values, size))
def generate_random_fuzz_examples(count):
for i in xrange(count):
size = [0]
var_values = {}
expr = make_random_expression(var_values, size)
yield expr, var_values
def test_fuzz_code_generator(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
#from expr_fuzz import get_fuzz_examples
for expr, var_values in generate_random_fuzz_examples(50):
#for expr, var_values in get_fuzz_examples():
from pymbolic import evaluate
try:
true_value = evaluate(expr, var_values)
except ZeroDivisionError:
continue
def get_dtype(x):
if isinstance(x, (complex, np.complexfloating)):
return np.complex128
else:
return np.float64
knl = lp.make_kernel(ctx.devices[0], "{ : }",
[lp.Instruction(None, "value", expr)],
[lp.GlobalArg("value", np.complex128, shape=())]
+ [
for name, val in var_values.iteritems()
])
ck = lp.CompiledKernel(ctx, knl)
evt, (lp_value,) = ck(queue, out_host=True, **var_values)
err = abs(true_value-lp_value)/abs(true_value)
if abs(err) > 1e-10:
print "---------------------------------------------------------------------"
print "WRONG: rel error=%g" % err
print "true=%r" % true_value
print "loopy=%r" % lp_value
print "---------------------------------------------------------------------"
print ck.code
print "---------------------------------------------------------------------"
print var_values
print "---------------------------------------------------------------------"
print repr(expr)
print "---------------------------------------------------------------------"
print expr
print "---------------------------------------------------------------------"
1/0
Andreas Klöckner
committed
def test_empty_reduction(ctx_factory):
dtype = np.dtype(np.float32)
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl = lp.make_kernel(ctx.devices[0],
[
"{[i]: 0<=i<20}",
Andreas Klöckner
committed
],
[
"a[i] = sum(j, j)",
],
[
lp.GlobalArg("a", dtype, (20,)),
])
cknl = lp.CompiledKernel(ctx, knl)
evt, (a,) = cknl(queue)
assert (a.get() == 0).all()
Andreas Klöckner
committed
def test_nested_dependent_reduction(ctx_factory):
Andreas Klöckner
committed
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl = lp.make_kernel(ctx.devices[0],
[
Andreas Klöckner
committed
"{[j]: 0<=j<i+sumlen}"
],
[
"<> sumlen = l[i]",
"a[i] = sum(j, j)",
],
[
lp.GlobalArg("a", dtype, ("n",)),
lp.GlobalArg("l", np.int32, ("n",)),
Andreas Klöckner
committed
])
cknl = lp.CompiledKernel(ctx, knl)
n = 330
l = np.arange(n, dtype=np.int32)
evt, (a,) = cknl(queue, l=l, n=n, out_host=True)
tgt_result = (2*l-1)*2*l/2
assert (a == tgt_result).all()
Andreas Klöckner
committed
def test_multi_nested_dependent_reduction(ctx_factory):
dtype = np.dtype(np.int32)
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
[
"{[itgt]: 0 <= itgt < ntgts}",
"{[isrc_box]: 0 <= isrc_box < nboxes}",
"{[isrc]: 0 <= isrc < npart}"
],
[
"<> npart = nparticles_per_box[isrc_box]",
"a[itgt] = sum((isrc_box, isrc), 1)",
],
[
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")
cknl = lp.CompiledKernel(ctx, knl)
print cknl.code
# FIXME: Actually test functionality.
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
def test_recursive_nested_dependent_reduction(ctx_factory):
dtype = np.dtype(np.int32)
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
[
"{[itgt]: 0 <= itgt < ntgts}",
"{[isrc_box]: 0 <= isrc_box < nboxes}",
"{[isrc]: 0 <= isrc < npart}"
],
[
"<> npart = nparticles_per_box[isrc_box]",
"<> boxsum = sum(isrc, isrc+isrc_box+itgt)",
"a[itgt] = sum(isrc_box, boxsum)",
],
[
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")
cknl = lp.CompiledKernel(ctx, knl)
print cknl.code
# FIXME: Actually test functionality.
Andreas Klöckner
committed
def test_dependent_loop_bounds(ctx_factory):
dtype = np.dtype(np.float32)
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
[
"{[i]: 0<=i<n}",
"{[jj]: 0<=jj<row_len}",
],
[
"<> row_len = a_rowstarts[i+1] - a_rowstarts[i]",
"ax[i] = sum(jj, a_values[a_rowstarts[i]+jj])",
],
[
lp.GlobalArg("a_rowstarts", np.int32),
lp.GlobalArg("a_indices", np.int32),
lp.GlobalArg("a_values", dtype),
lp.GlobalArg("x", dtype),
lp.GlobalArg("ax", dtype),
Andreas Klöckner
committed
],
assumptions="n>=1 and row_len>=1")
cknl = lp.CompiledKernel(ctx, knl)
print "---------------------------------------------------"
print cknl.get_highlighted_code()
Andreas Klöckner
committed
print "---------------------------------------------------"
def test_dependent_loop_bounds_2(ctx_factory):
dtype = np.dtype(np.float32)
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
[
"{[i]: 0<=i<n}",
"{[jj]: 0<=jj<row_len}",
],
Andreas Klöckner
committed
[
"<> row_start = a_rowstarts[i]",
"<> row_len = a_rowstarts[i+1] - row_start",
"ax[i] = sum(jj, a_values[row_start+jj])",
],
[
lp.GlobalArg("a_rowstarts", np.int32),
lp.GlobalArg("a_indices", np.int32),
lp.GlobalArg("a_values", dtype),
lp.GlobalArg("x", dtype),
lp.GlobalArg("ax", dtype),
Andreas Klöckner
committed
],
assumptions="n>=1 and row_len>=1")
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0",
Andreas Klöckner
committed
inner_tag="l.0")
cknl = lp.CompiledKernel(ctx, knl)
print "---------------------------------------------------"
print cknl.get_highlighted_code()
Andreas Klöckner
committed
print "---------------------------------------------------"
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.
Andreas Klöckner
committed
dtype = np.dtype(np.float32)
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
[
"{[i]: 0<=i<n}",
"{[jj]: 0<=jj<row_len}",
],
Andreas Klöckner
committed
[
"<> row_len = a_row_lengths[i]",
Andreas Klöckner
committed
],
[
lp.GlobalArg("a_row_lengths", np.int32),
lp.GlobalArg("a", dtype, shape=("n,n"), order="C"),
Andreas Klöckner
committed
])
assert knl.parents_per_domain()[1] == 0
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0",
Andreas Klöckner
committed
inner_tag="l.0")
Andreas Klöckner
committed
cknl = lp.CompiledKernel(ctx, knl)
print "---------------------------------------------------"
print cknl.get_highlighted_code()
Andreas Klöckner
committed
print "---------------------------------------------------"
knl_bad = lp.split_iname(knl, "jj", 128, outer_tag="g.1",
inner_tag="l.1")
import pytest
with pytest.raises(RuntimeError):
list(lp.generate_loop_schedules(knl_bad))
def test_independent_multi_domain(ctx_factory):
dtype = np.dtype(np.float32)
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl = lp.make_kernel(ctx.devices[0],
[
"{[i]: 0<=i<n}",
"{[j]: 0<=j<n}",
],
[
lp.GlobalArg("a", dtype, shape=("n"), order="C"),
lp.GlobalArg("b", dtype, shape=("n"), order="C"),
knl = lp.split_iname(knl, "i", 16, outer_tag="g.0",
knl = lp.split_iname(knl, "j", 16, outer_tag="g.0",
assert knl.parents_per_domain() == 2*[None]
n = 50
cknl = lp.CompiledKernel(ctx, knl)
evt, (a, b) = cknl(queue, n=n, out_host=True)
assert a.shape == (50,)
assert b.shape == (50,)
def test_bare_data_dependency(ctx_factory):
dtype = np.dtype(np.float32)
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl = lp.make_kernel(ctx.devices[0],
[
"[znirp] -> {[i]: 0<=i<znirp}",
],
[
"<> znirp = n",
"a[i] = 1",
],
[
lp.GlobalArg("a", dtype, shape=("n"), order="C"),
])
cknl = lp.CompiledKernel(ctx, knl)
n = 20000
evt, (a,) = cknl(queue, n=n, out_host=True)
assert a.shape == (n,)
assert (a == 1).all()
Andreas Klöckner
committed
def test_equality_constraints(ctx_factory):
dtype = np.float32
ctx = ctx_factory()
order = "C"
knl = lp.make_kernel(ctx.devices[0], [
"[n] -> {[i,j]: 0<=i,j<n }",
"{[k]: k =i+5 and k < n}",
],
"a[i,j] = 5 {id=set_all}",
"a[i,k] = 22 {dep=set_all}",
lp.GlobalArg("a", dtype, shape="n, n", order=order),
lp.ValueArg("n", np.int32, approximately=1000),
name="equality_constraints", assumptions="n>=1")
seq_knl = knl
knl = lp.split_iname(knl, "i", 16, outer_tag="g.0", inner_tag="l.0")
knl = lp.split_iname(knl, "j", 16, outer_tag="g.1", inner_tag="l.1")
#print knl
#print knl.domains[0].detect_equalities()
kernel_gen = lp.generate_loop_schedules(knl)
kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
parameters=dict(n=n), print_ref_code=True)
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
def test_stride(ctx_factory):
dtype = np.float32
ctx = ctx_factory()
order = "C"
n = 10
knl = lp.make_kernel(ctx.devices[0], [
"{[i]: 0<=i<n and (exists l: i = 2*l)}",
],
[
"a[i] = 5",
],
[
lp.GlobalArg("a", dtype, shape="n", order=order),
lp.ValueArg("n", np.int32, approximately=1000),
],
assumptions="n>=1")
seq_knl = knl
kernel_gen = lp.generate_loop_schedules(knl)
kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
parameters=dict(n=n), fills_entire_output=False)
def test_domain_dependency_via_existentially_quantified_variable(ctx_factory):
dtype = np.float32
ctx = ctx_factory()
order = "C"
n = 10
knl = lp.make_kernel(ctx.devices[0], [
"{[i]: 0<=i<n }",
"{[k]: k=i and (exists l: k = 2*l) }",
],
[
"a[i] = 5 {id=set}",
"a[k] = 6 {dep=set}",
],
[
lp.GlobalArg("a", dtype, shape="n", order=order),
lp.ValueArg("n", np.int32, approximately=1000),
],
assumptions="n>=1")
seq_knl = knl
kernel_gen = lp.generate_loop_schedules(knl)
kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
parameters=dict(n=n), )
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
def test_double_sum(ctx_factory):
dtype = np.float32
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
n = 20
knl = lp.make_kernel(ctx.devices[0], [
"{[i,j]: 0<=i,j<n }",
],
[
"a = sum((i,j), i*j)",
"b = sum(i, sum(j, i*j))",
],
[
lp.GlobalArg("a", dtype, shape=()),
lp.GlobalArg("b", dtype, shape=()),
lp.ValueArg("n", np.int32, approximately=1000),
],
assumptions="n>=1")
cknl = lp.CompiledKernel(ctx, knl)
evt, (a, b) = cknl(queue, n=n)
ref = sum(i*j for i in xrange(n) for j in xrange(n))
assert a.get() == ref
assert b.get() == ref
# {{{ test race detection
def test_ilp_write_race_detection_global(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0], [
"[n] -> {[i,j]: 0<=i,j<n }",
],
[
Andreas Klöckner
committed
"a[i] = 5+i+j",
],
[
lp.GlobalArg("a", np.float32),
lp.ValueArg("n", np.int32, approximately=1000),
],
assumptions="n>=1")
Andreas Klöckner
committed
knl = lp.tag_inames(knl, dict(j="ilp"))
from loopy.check import WriteRaceConditionError
import pytest
with pytest.raises(WriteRaceConditionError):
list(lp.generate_loop_schedules(knl))
def test_ilp_write_race_avoidance_local(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
Andreas Klöckner
committed
"<> a[i] = 5+i+j",
Andreas Klöckner
committed
knl = lp.tag_inames(knl, dict(i="l.0", j="ilp"))
for k in lp.generate_loop_schedules(knl):
assert k.temporary_variables["a"].shape == (16,17)
def test_ilp_write_race_avoidance_private(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{[j]: 0<=j<16 }",
[
Andreas Klöckner
committed
"<> a = 5+j",