Newer
Older
Andreas Klöckner
committed
print lp.CompiledKernel(ctx, knl).get_highlighted_code()
def test_arg_guessing(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0], [
"{[i,j]: 0<=i,j<n }",
],
"""
a = 1.5 + sum((i,j), i*j)
b[i, j] = i*j
c[i+j, j] = b[j,i]
""",
assumptions="n>=1")
print knl
print lp.CompiledKernel(ctx, knl).get_highlighted_code()
def test_arg_guessing_with_reduction(ctx_factory):
#logging.basicConfig(level=logging.DEBUG)
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0], [
"{[i,j]: 0<=i,j<n }",
],
"""
a = 1.5 + sum((i,j), i*j)
d = 1.5 + sum((i,j), b[i,j])
b[i, j] = i*j
c[i+j, j] = b[j,i]
""",
assumptions="n>=1")
print knl
print lp.CompiledKernel(ctx, knl).get_highlighted_code()
# }}}
def test_nonlinear_index(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0], [
"{[i,j]: 0<=i,j<n }",
],
"""
a[i*i] = 17
""",
[
lp.GlobalArg("a", shape="n"),
lp.ValueArg("n"),
],
assumptions="n>=1")
print knl
print lp.CompiledKernel(ctx, knl).get_highlighted_code()
def test_triangle_domain(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0], [
"{[i,j]: 0<=i,j<n and i <= j}",
],
"a[i,j] = 17",
assumptions="n>=1")
print knl
print lp.CompiledKernel(ctx, knl).get_highlighted_code()
def test_offsets_and_slicing(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl = lp.make_kernel(ctx.devices[0], [
"{[i,j]: 0<=i<n and 0<=j<m }",
],
"""
b[i,j] = 2*a[i,j]
""",
assumptions="n>=1 and m>=1",
default_offset=lp.auto)
knl = lp.tag_data_axes(knl, "a,b", "stride:auto,stride:1")
cknl = lp.CompiledKernel(ctx, knl)
a_full = cl.clrandom.rand(queue, (n, n), np.float64)
a_full_h = a_full.get()
b_full = cl.clrandom.rand(queue, (n, n), np.float64)
b_full_h = b_full.get()
a_sub = (slice(3, 10), slice(5, 10))
a = a_full[a_sub]
b_sub = (slice(3+3, 10+3), slice(5+4, 10+4))
b = b_full[b_sub]
b_full_h[b_sub] = 2*a_full_h[a_sub]
print cknl.get_highlighted_code({"a": a.dtype})
import numpy.linalg as la
assert la.norm(b_full.get() - b_full_h) < 1e-13
Andreas Klöckner
committed
def test_vector_ilp_with_prefetch(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(ctx.devices[0],
"{ [i]: 0<=i<n }",
"out[i] = 2*a[i]",
[
# Tests that comma'd arguments interoperate with
Andreas Klöckner
committed
# argument guessing.
lp.GlobalArg("out,a", np.float32, shape=lp.auto),
"..."
])
knl = lp.split_iname(knl, "i", 128, 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"])
cknl = lp.CompiledKernel(ctx, knl)
cknl.cl_kernel_info()
Andreas Klöckner
committed
import re
Andreas Klöckner
committed
assert len(list(re.finditer("barrier", code))) == 1
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
def test_convolution_like(ctx_factory):
ctx = ctx_factory()
dtype = np.float64
knl = lp.make_kernel(ctx.devices[0],
"{ [im_x, im_y, f_x, f_y]: -f_w <= f_x,f_y <= f_w \
and f_w <= im_x < im_w-f_w and f_w <= im_y < im_h-f_w }",
"""
out[im_x-f_w, im_y-f_w] = sum((f_x, f_y), \
img[im_x-f_x, im_y-f_y] * f[f_w+f_x, f_w+f_y])
""",
[
lp.GlobalArg("f", dtype, shape=lp.auto),
lp.GlobalArg("img", dtype, shape=lp.auto),
lp.GlobalArg("out", dtype, shape=lp.auto),
"..."
],
assumptions="f_w>=1 and im_w, im_h >= 1")
ref_knl = knl
def variant(knl):
knl = lp.split_iname(knl, "im_x", 16, inner_tag="l.0")
return knl
lp.auto_test_vs_ref(ref_knl, ctx, variant(knl),
parameters={"im_w": 1024, "im_h": 1024, "f_w": 7})
if __name__ == "__main__":
import sys
if len(sys.argv) > 1:
exec(sys.argv[1])
else:
from py.test.cmdline import main
main([__file__])
# vim: foldmethod=marker