Newer
Older
knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32))
ref_knl = knl
knl = lp.link_inames(knl, "i,j", "i0")
from loopy.diagnostic import AutomaticTestFailure
with pytest.raises(AutomaticTestFailure):
lp.auto_test_vs_ref(
ref_knl, ctx, knl,
parameters=dict(n=123))
def test_generate_c_snippet():
from loopy.target.c import CTarget
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)
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
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.set_loop_priority(knl, "I,k_outer,k_inner")
knl = lp.preprocess_kernel(knl)
knl = lp.get_one_scheduled_kernel(knl)
print(lp.generate_body(knl))
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
def test_precompute_with_preexisting_inames(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[e,i,j,k]: 0<=e<E and 0<=i,j,k<n}",
"""
result[e,i] = sum(j, D1[i,j]*u[e,j])
result2[e,i] = sum(k, D2[i,k]*u[e,k])
""")
knl = lp.add_and_infer_dtypes(knl, {
"u": np.float32,
"D1": np.float32,
"D2": np.float32,
})
knl = lp.fix_parameters(knl, n=13)
ref_knl = knl
knl = lp.extract_subst(knl, "D1_subst", "D1[ii,jj]", parameters="ii,jj")
knl = lp.extract_subst(knl, "D2_subst", "D2[ii,jj]", parameters="ii,jj")
knl = lp.precompute(knl, "D1_subst", "i,j", default_tag="for",
precompute_inames="ii,jj")
knl = lp.precompute(knl, "D2_subst", "i,k", default_tag="for",
precompute_inames="ii,jj")
knl = lp.set_loop_priority(knl, "ii,jj,e,j,k")
lp.auto_test_vs_ref(
ref_knl, ctx, knl,
parameters=dict(E=200))
def test_precompute_with_preexisting_inames_fail():
knl = lp.make_kernel(
"{[e,i,j,k]: 0<=e<E and 0<=i,j<n and 0<=k<2*n}",
"""
result[e,i] = sum(j, D1[i,j]*u[e,j])
result2[e,i] = sum(k, D2[i,k]*u[e,k])
""")
knl = lp.add_and_infer_dtypes(knl, {
"u": np.float32,
"D1": np.float32,
"D2": np.float32,
})
knl = lp.fix_parameters(knl, n=13)
knl = lp.extract_subst(knl, "D1_subst", "D1[ii,jj]", parameters="ii,jj")
knl = lp.extract_subst(knl, "D2_subst", "D2[ii,jj]", parameters="ii,jj")
knl = lp.precompute(knl, "D1_subst", "i,j", default_tag="for",
precompute_inames="ii,jj")
with pytest.raises(lp.LoopyError):
lp.precompute(knl, "D2_subst", "i,k", default_tag="for",
precompute_inames="ii,jj")
def test_vectorize(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"""
<> temp = 2*b[i]
a[i] = temp
""")
knl = lp.add_and_infer_dtypes(knl, dict(b=np.float32))
knl = lp.set_array_dim_names(knl, "a,b", "i")
knl = lp.split_array_dim(knl, [("a", 0), ("b", 0)], 4,
split_kwargs=dict(slabs=(0, 1)))
knl = lp.tag_data_axes(knl, "a,b", "c,vec")
ref_knl = knl
ref_knl = lp.tag_inames(ref_knl, {"i_inner": "unr"})
knl = lp.tag_inames(knl, {"i_inner": "vec"})
knl = lp.preprocess_kernel(knl)
knl = lp.get_one_scheduled_kernel(knl)
code, inf = lp.generate_code(knl)
lp.auto_test_vs_ref(
ref_knl, ctx, knl,
parameters=dict(n=30))
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
def test_alias_temporaries(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"""
times2(i) := 2*a[i]
times3(i) := 3*a[i]
times4(i) := 4*a[i]
x[i] = times2(i)
y[i] = times3(i)
z[i] = times4(i)
""")
knl = lp.add_and_infer_dtypes(knl, {"a": np.float32})
ref_knl = knl
knl = lp.split_iname(knl, "i", 16, outer_tag="g.0", inner_tag="l.0")
knl = lp.precompute(knl, "times2", "i_inner")
knl = lp.precompute(knl, "times3", "i_inner")
knl = lp.precompute(knl, "times4", "i_inner")
knl = lp.alias_temporaries(knl, ["times2_0", "times3_0", "times4_0"])
lp.auto_test_vs_ref(
ref_knl, ctx, knl,
parameters=dict(n=30))
def test_fusion():
exp_kernel = lp.make_kernel(
''' { [i]: 0<=i<n } ''',
''' exp[i] = pow(E, z[i])''',
assumptions="n>0")
sum_kernel = lp.make_kernel(
'{ [j]: 0<=j<n }',
'out2 = sum(j, exp[j])',
assumptions='n>0')
knl = lp.fuse_kernels([exp_kernel, sum_kernel])
print(knl)
def test_sci_notation_literal(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
set_kernel = lp.make_kernel(
''' { [i]: 0<=i<12 } ''',
''' out[i] = 1e-12''')
set_kernel = lp.set_options(set_kernel, write_cl=True)
evt, (out,) = set_kernel(queue)
assert (np.abs(out.get() - 1e-12) < 1e-20).all()
def test_rename_argument(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
kernel = lp.make_kernel(
'''{ [i]: 0<=i<n }''',
'''out[i] = a + 2''')
kernel = lp.rename_argument(kernel, "a", "b")
evt, (out,) = kernel(queue, b=np.float32(12), n=20)
assert (np.abs(out.get() - 14) < 1e-8).all()
def test_to_batched(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl = lp.make_kernel(
''' { [i,j]: 0<=i,j<n } ''',
''' out[i] = sum(j, a[i,j]*x[j])''')
bknl = lp.to_batched(knl, "nbatches", "out,x")
a = np.random.randn(5, 5)
x = np.random.randn(7, 5)
bknl(queue, a=a, x=x)
def test_variable_size_temporary():
knl = lp.make_kernel(
''' { [i,j]: 0<=i,j<n } ''',
''' out[i] = sum(j, a[i,j])''')
knl = lp.add_and_infer_dtypes(knl, {"a": np.float32})
knl = lp.add_prefetch(
knl, "a[:,:]", default_tag=None)
# Make sure that code generation succeeds even if
# there are variable-length arrays.
knl = lp.preprocess_kernel(knl)
for k in lp.generate_loop_schedules(knl):
lp.generate_code(k)
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
def test_indexof(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl = lp.make_kernel(
''' { [i,j]: 0<=i,j<5 } ''',
''' out[i,j] = indexof(out[i,j])''')
knl = lp.set_options(knl, write_cl=True)
(evt, (out,)) = knl(queue)
out = out.get()
assert np.array_equal(out.ravel(order="C"), np.arange(25))
def test_indexof_vec(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
if ctx.devices[0].platform.name.startswith("Portable"):
# Accurate as of 2015-10-08
pytest.skip("POCL miscompiles vector code")
knl = lp.make_kernel(
''' { [i,j,k]: 0<=i,j,k<4 } ''',
''' out[i,j,k] = indexof_vec(out[i,j,k])''')
knl = lp.tag_inames(knl, {"i": "vec"})
knl = lp.tag_data_axes(knl, "out", "vec,c,c")
knl = lp.set_options(knl, write_cl=True)
(evt, (out,)) = knl(queue)
#out = out.get()
#assert np.array_equal(out.ravel(order="C"), np.arange(25))
Andreas Klöckner
committed
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
def test_finite_difference_expr_subst(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
grid = np.linspace(0, 2*np.pi, 2048, endpoint=False)
h = grid[1] - grid[0]
u = cl.clmath.sin(cl.array.to_device(queue, grid))
fin_diff_knl = lp.make_kernel(
"{[i]: 1<=i<=n}",
"out[i] = -(f[i+1] - f[i-1])/h",
[lp.GlobalArg("out", shape="n+2"), "..."])
flux_knl = lp.make_kernel(
"{[j]: 1<=j<=n}",
"f[j] = u[j]**2/2",
[
lp.GlobalArg("f", shape="n+2"),
lp.GlobalArg("u", shape="n+2"),
])
fused_knl = lp.fuse_kernels([fin_diff_knl, flux_knl],
data_flow=[
("f", 1, 0)
])
Andreas Klöckner
committed
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
fused_knl = lp.set_options(fused_knl, write_cl=True)
evt, _ = fused_knl(queue, u=u, h=np.float32(1e-1))
fused_knl = lp.assignment_to_subst(fused_knl, "f")
fused_knl = lp.set_options(fused_knl, write_cl=True)
# This is the real test here: The automatically generated
# shape expressions are '2+n' and the ones above are 'n+2'.
# Is loopy smart enough to understand that these are equal?
evt, _ = fused_knl(queue, u=u, h=np.float32(1e-1))
fused0_knl = lp.affine_map_inames(fused_knl, "i", "inew", "inew+1=i")
gpu_knl = lp.split_iname(
fused0_knl, "inew", 128, outer_tag="g.0", inner_tag="l.0")
precomp_knl = lp.precompute(
gpu_knl, "f_subst", "inew_inner", fetch_bounding_box=True)
precomp_knl = lp.tag_inames(precomp_knl, {"j_0_outer": "unr"})
precomp_knl = lp.set_options(precomp_knl, return_dict=True)
evt, _ = precomp_knl(queue, u=u, h=h)
def test_is_expression_equal():
from loopy.symbolic import is_expression_equal
from pymbolic import var
x = var("x")
y = var("y")
assert is_expression_equal(x+2, 2+x)
assert is_expression_equal((x+2)**2, x**2 + 4*x + 4)
assert is_expression_equal((x+y)**2, x**2 + 2*x*y + y**2)
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
def test_collect_common_factors(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j<n}",
"""
<float32> out_tmp = 0 {id=out_init,inames=i}
out_tmp = out_tmp + alpha[i]*a[i,j]*b1[j] {id=out_up1,dep=out_init}
out_tmp = out_tmp + alpha[i]*a[j,i]*b2[j] {id=out_up2,dep=out_init}
out[i] = out_tmp {dep=out_up1:out_up2}
""")
knl = lp.add_and_infer_dtypes(knl,
dict(a=np.float32, alpha=np.float32, b1=np.float32, b2=np.float32))
ref_knl = knl
knl = lp.split_iname(knl, "i", 256, outer_tag="g.0", inner_tag="l.0")
knl = lp.collect_common_factors_on_increment(knl, "out_tmp")
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=13))
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.preprocess_kernel(knl)))
print(codegen_result.device_code())
print(codegen_result.host_code())
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
def test_cuda_target():
from loopy.target.cuda import CudaTarget
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i] = 2*a[i]",
[
lp.GlobalArg("out,a", np.float32, shape=lp.auto),
"..."
],
target=CudaTarget())
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"])
print(
lp.generate_code(
lp.get_one_scheduled_kernel(
lp.preprocess_kernel(knl)))[0])
def test_chunk_iname(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i] = 2*a[i]",
[
lp.GlobalArg("out,a", np.float32, shape=lp.auto),
"..."
],
assumptions="n>0")
ref_knl = knl
knl = lp.chunk_iname(knl, "i", 3, inner_tag="l.0")
knl = lp.set_loop_priority(knl, "i_outer, i_inner")
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=130))
@pytest.mark.parametrize("dtype", [np.int32, np.int64, np.float32, np.float64])
def test_atomic(ctx_factory, dtype):
ctx = ctx_factory()
if (
np.dtype(dtype).itemsize == 8
and "cl_khr_int64_base_atomics" not in ctx.devices[0].extensions):
pytest.skip("64-bit atomics not supported on device")
import pyopencl.version # noqa
if (
cl.version.VERSION < (2015, 2)
and dtype == np.int64):
pytest.skip("int64 RNG not supported in PyOpenCL < 2015.2")
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i%20] = out[i%20] + 2*a[i] {atomic}",
[
lp.GlobalArg("out", dtype, shape=lp.auto, for_atomic=True),
lp.GlobalArg("a", dtype, shape=lp.auto),
"..."
],
assumptions="n>0")
ref_knl = knl
knl = lp.split_iname(knl, "i", 512)
knl = lp.split_iname(knl, "i_inner", 128, outer_tag="unr", inner_tag="g.0")
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=10000))
def test_clamp(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
n = 15 * 10**6
x = cl.clrandom.rand(queue, n, dtype=np.float32)
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i] = clamp(x[i], a, b)")
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
knl = lp.set_options(knl, write_cl=True)
evt, (out,) = knl(queue, x=x, a=np.float32(12), b=np.float32(15))
def test_forced_iname_deps_and_reduction():
# See https://github.com/inducer/loopy/issues/24
# This is (purposefully) somewhat un-idiomatic, to replicate the conditions
# under which the above bug was found. If assignees were phi[i], then the
# iname propagation heuristic would not assume that dependent instructions
# need to run inside of 'i', and hence the forced_iname_* bits below would not
# be needed.
i1 = lp.CInstruction("i",
"doSomethingToGetPhi();",
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
from pymbolic.primitives import Subscript, Variable
i2 = lp.Assignment("a",
lp.Reduction("sum", "j", Subscript(Variable("phi"), Variable("j"))),
forced_iname_deps=frozenset(),
forced_iname_deps_is_final=True)
k = lp.make_kernel("{[i,j] : 0<=i,j<n}",
[i1, i2],
[
lp.GlobalArg("a", dtype=np.float32, shape=()),
lp.ValueArg("n", dtype=np.int32),
lp.TemporaryVariable("phi", dtype=np.float32, shape=("n",)),
],
target=lp.CTarget(),
)
k = lp.preprocess_kernel(k)
assert 'i' not in k.insn_inames("insn_0_j_update")
print(k.stringify(with_dependencies=True))
@pytest.mark.parametrize("tp", ["f32", "f64"])
def test_random123(ctx_factory, tp):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
import pyopencl.version # noqa
if cl.version.VERSION < (2016, 2):
pytest.skip("Random123 RNG not supported in PyOpenCL < 2016.2")
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
n = 150000
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"""
<> key2 = make_uint2(i, 324830944) {inames=i}
<> key4 = make_uint4(i, 324830944, 234181, 2233) {inames=i}
<> ctr = make_uint4(0, 1, 2, 3) {inames=i}
<> real, ctr = philox4x32_TYPE(ctr, key2)
<> imag, ctr = threefry4x32_TYPE(ctr, key4)
out[i, 0] = real.s0 + 1j * imag.s0
out[i, 1] = real.s1 + 1j * imag.s1
out[i, 2] = real.s2 + 1j * imag.s2
out[i, 3] = real.s3 + 1j * imag.s3
""".replace("TYPE", tp))
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
knl = lp.set_options(knl, write_cl=True)
evt, (out,) = knl(queue, n=n)
out = out.get()
assert (out < 1).all()
assert (0 <= out).all()
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
def test_kernel_splitting(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})
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 test_kernel_splitting_with_loop(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})
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 test_kernel_splitting_with_loop_and_private_temporary(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{ [i,k]: 0<=i<n and 0<=k<3 }",
"""
<> t_private = a[k,i+1]
c[k,i] = a[k,i+1]
out[k,i] = c[k,i] + t_private
""")
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
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=5))
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
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))
if __name__ == "__main__":
if len(sys.argv) > 1:
exec(sys.argv[1])
else:
from py.test.cmdline import main
main([__file__])
# vim: foldmethod=marker