def build_ispc(code):
with open("tasksys.cpp", "r") as ts_file:
tasksys_source = ts_file.read()
tmpdir = tempfile.mkdtemp()
build_ispc_shared_lib(
tmpdir,
[("my.ispc", code)],
[("tasksys.cpp", tasksys_source)],
cxx_options=["-g", "-fopenmp", "-DISPC_USE_OMP"],
ispc_options=([
"-g", "--no-omit-frame-pointer",
"--target=" + ISPC_TARGET,
"--opt=force-aligned-memory",
"--opt=disable-loop-unroll",
#"--math-lib=fast",
#"--opt=fast-math",
#"--opt=disable-fma",
]
#+ (["--addressing=64"] if INDEX_DTYPE == np.int64 else [])
),
#ispc_bin="/home/andreask/pack/ispc-v1.9.0-linux/ispc",
quiet=False,
)
shared_obj = os.path.join(tmpdir, "shared.so")
return shared_obj
def main():
with open("tasksys.cpp") as ts_file:
tasksys_source = ts_file.read()
def make_knl(name, insn, vars):
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
insn,
target=lp.ISPCTarget(), index_dtype=INDEX_DTYPE,
name="stream_"+name+"_tasks")
knl = transform(knl, vars, STREAM_DTYPE)
return knl
init_knl = make_knl("init", """
a[i] = 1
b[i] = 2
c[i] = 0
""", "a,b,c")
triad_knl = make_knl("triad", """
a[i] = b[i] + scalar * c[i]
""", "a,b,c,scalar")
with TemporaryDirectory() as tmpdir:
ispc_code = gen_code(init_knl) + gen_code(triad_knl)
print(ispc_code)
build_ispc_shared_lib(
tmpdir,
[("stream.ispc", ispc_code)],
[("tasksys.cpp", tasksys_source)],
cxx_options=["-g", "-fopenmp", "-DISPC_USE_OMP"],
ispc_options=([
#"-g", "--no-omit-frame-pointer",
"--target=avx2-i32x8",
"--opt=force-aligned-memory",
"--opt=disable-loop-unroll",
#"--opt=fast-math",
#"--opt=disable-fma",
]
+ (["--addressing=64"] if INDEX_DTYPE == np.int64 else [])
),
#ispc_bin="/home/andreask/pack/ispc-v1.9.0-linux/ispc",
quiet=False,
)
knl_lib = ctypes.cdll.LoadLibrary(os.path.join(tmpdir, "shared.so"))
scalar = 5
a = empty_aligned(ARRAY_SIZE, dtype=STREAM_DTYPE, n=ALIGN_TO)
b = empty_aligned(ARRAY_SIZE, dtype=STREAM_DTYPE, n=ALIGN_TO)
c = empty_aligned(ARRAY_SIZE, dtype=STREAM_DTYPE, n=ALIGN_TO)
print(
hex(address_from_numpy(a)),
hex(address_from_numpy(b)),
hex(address_from_numpy(c)))
assert address_from_numpy(a) % ALIGN_TO == 0
assert address_from_numpy(b) % ALIGN_TO == 0
assert address_from_numpy(c) % ALIGN_TO == 0
knl_lib.stream_init_tasks(
cptr_from_numpy(a),
cptr_from_numpy(b),
cptr_from_numpy(c),
INDEX_CTYPE(ARRAY_SIZE),
)
def call_kernel():
knl_lib.stream_triad_tasks(
cptr_from_numpy(a),
cptr_from_numpy(b),
cptr_from_numpy(c),
STREAM_CTYPE(scalar),
INDEX_CTYPE(ARRAY_SIZE),
)
call_kernel()
call_kernel()
start_time = time()
for irun in range(NRUNS):
call_kernel()
elapsed = time() - start_time
print(elapsed/NRUNS)
print(1e-9*3*a.nbytes*NRUNS/elapsed, "GB/s")
assert la.norm(a-b+scalar*c, np.inf) < np.finfo(STREAM_DTYPE).eps * 10
def run(approx, code, size, NRUNS):
ALIGN_TO = 4096
if approx.dtype_name == "float":
dt = np.float32
STREAM_DTYPE = np.float32
STREAM_CTYPE = ctypes.c_float
INDEX_DTYPE = np.int32
INDEX_CTYPE = ctypes.c_int
elif approx.dtype_name == "double":
dt = np.float64
STREAM_DTYPE = np.float64
STREAM_CTYPE = ctypes.c_doublev
INDEX_DTYPE = np.int64
INDEX_CTYPE = ctypes.c_longlong
with open("tests/tasksys.cpp", "r") as ts_file:
tasksys_source = ts_file.read()
with TemporaryDirectory() as tmpdir:
print(code)
build_ispc_shared_lib(
tmpdir,
[("stream.ispc", code)],
[("tasksys.cpp", tasksys_source)],
cxx_options=["-g", "-fopenmp", "-DISPC_USE_OMP"],
ispc_options=([
"-g",
"--no-omit-frame-pointer",
#"--target=avx2-i32x8",
"--arch=x86-64",
"--target=avx2",
#"--opt=force-aligned-memory",
"--opt=disable-loop-unroll",
#"--opt=fast-math",
"--opt=disable-fma",
"--woff"
] + (["--addressing=64"] if INDEX_DTYPE == np.int64 else [])),
ispc_bin="/home/ubuntu-boot/Desktop/ispc-v1.9.1-linux/ispc",
quiet=True,
)
x = np.linspace(approx.lower_bound, approx.upper_bound, size, dtype=dt)
y = np.zeros(size, dtype=dt)
approx.tree_1d = np.array(approx.tree_1d, dtype=dt)
approx.map = np.array(approx.map, dtype=dt)
approx.mid = np.array(approx.mid, dtype=dt)
approx.left = np.array(approx.left, dtype=dt)
approx.right = np.array(approx.right, dtype=dt)
approx.interval_a = np.array(approx.interval_a, dtype=dt)
approx.interval_b = np.array(approx.interval_b, dtype=dt)
approx.coeff = np.array(approx.coeff, dtype=dt)
knl_lib = ctypes.cdll.LoadLibrary(os.path.join(tmpdir, "shared.so"))
def call_kernel():
if 'map' in approx.optimizations:
knl_lib.eval(
cptr_from_numpy(approx.mid),
cptr_from_numpy(approx.left),
cptr_from_numpy(approx.right),
cptr_from_numpy(approx.interval_a),
cptr_from_numpy(approx.interval_b),
cptr_from_numpy(approx.coeff),
cptr_from_numpy(approx.map),
cptr_from_numpy(x),
cptr_from_numpy(y),
INDEX_CTYPE(size),
)
else:
knl_lib.eval(
cptr_from_numpy(approx.tree_1d),
cptr_from_numpy(x),
cptr_from_numpy(y),
INDEX_CTYPE(size),
)
# clear the kernel
for i in range(10):
call_kernel()
#print(y)
#print(np.max(y))
plt.figure()
plt.plot(x[::2048], y[::2048])
plt.show()
start_time = time.time()
for irun in range(NRUNS):
call_kernel()
elapsed = time.time() - start_time
FLOPS = (4 + 2 + 2 * (approx.max_order - 2))
print("Average Runtime:", elapsed / NRUNS)
# times size*4 because thats the number of bytes in x
GFLOPS = FLOPS * size / (2**30)
print(GFLOPS / elapsed, "GFLOPS/s")
return y
def main():
with open("tasksys.cpp", "r") as ts_file:
tasksys_source = ts_file.read()
def make_knl(name, insn, vars):
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
insn,
target=lp.ISPCTarget(), index_dtype=INDEX_DTYPE,
name="stream_"+name+"_tasks")
knl = transform(knl, vars, STREAM_DTYPE)
return knl
init_knl = make_knl("init", """
a[i] = 1
b[i] = 2
c[i] = 0
""", "a,b,c")
triad_knl = make_knl("triad", """
a[i] = b[i] + scalar * c[i]
""", "a,b,c,scalar")
with TemporaryDirectory() as tmpdir:
ispc_code = gen_code(init_knl) + gen_code(triad_knl)
print(ispc_code)
build_ispc_shared_lib(
tmpdir,
[("stream.ispc", ispc_code)],
[("tasksys.cpp", tasksys_source)],
cxx_options=["-g", "-fopenmp", "-DISPC_USE_OMP"],
ispc_options=([
#"-g", "--no-omit-frame-pointer",
"--target=avx2-i32x8",
"--opt=force-aligned-memory",
"--opt=disable-loop-unroll",
#"--opt=fast-math",
#"--opt=disable-fma",
]
+ (["--addressing=64"] if INDEX_DTYPE == np.int64 else [])
),
#ispc_bin="/home/andreask/pack/ispc-v1.9.0-linux/ispc",
quiet=False,
)
knl_lib = ctypes.cdll.LoadLibrary(os.path.join(tmpdir, "shared.so"))
scalar = 5
a = empty_aligned(ARRAY_SIZE, dtype=STREAM_DTYPE, n=ALIGN_TO)
b = empty_aligned(ARRAY_SIZE, dtype=STREAM_DTYPE, n=ALIGN_TO)
c = empty_aligned(ARRAY_SIZE, dtype=STREAM_DTYPE, n=ALIGN_TO)
print(
hex(address_from_numpy(a)),
hex(address_from_numpy(b)),
hex(address_from_numpy(c)))
assert address_from_numpy(a) % ALIGN_TO == 0
assert address_from_numpy(b) % ALIGN_TO == 0
assert address_from_numpy(c) % ALIGN_TO == 0
knl_lib.stream_init_tasks(
cptr_from_numpy(a),
cptr_from_numpy(b),
cptr_from_numpy(c),
INDEX_CTYPE(ARRAY_SIZE),
)
def call_kernel():
knl_lib.stream_triad_tasks(
cptr_from_numpy(a),
cptr_from_numpy(b),
cptr_from_numpy(c),
STREAM_CTYPE(scalar),
INDEX_CTYPE(ARRAY_SIZE),
)
call_kernel()
call_kernel()
start_time = time()
for irun in range(NRUNS):
call_kernel()
elapsed = time() - start_time
print(elapsed/NRUNS)
print(1e-9*3*a.nbytes*NRUNS/elapsed, "GB/s")
assert la.norm(a-b+scalar*c, np.inf) < np.finfo(STREAM_DTYPE).eps * 10
def main(experiment):
print()
print("Task: ", experiment)
with open("tests/tasksys.cpp", "r") as ts_file:
tasksys_source = ts_file.read()
def make_knl(name, insn, vars):
knl = lp.make_kernel("{[i]: 0<=i<n}",
insn,
target=lp.ISPCTarget(),
index_dtype=INDEX_DTYPE,
name="stream_" + name + "_tasks")
knl = transform(knl, vars, STREAM_DTYPE)
return knl
if experiment == "triad":
init_knl = make_knl(
"init", """
a[i] = 1
b[i] = 3
c[i] = 0
scalar = 7
""", "a,b,c,scalar")
triad_knl = make_knl(
"triad", """
a[i] = b[i] + scalar * c[i]
""", "a,b,c,scalar")
else:
init_knl = make_knl(
"init", """
a[i] = 1
b[i] = 9
""", "a,b")
copy_knl = make_knl("copy", """
a[i] = b[i]
""", "a,b")
with TemporaryDirectory() as tmpdir:
if experiment == "copy":
ispc_code = gen_code(init_knl) + gen_code(copy_knl)
else:
ispc_code = gen_code(init_knl) + gen_code(triad_knl)
print(ispc_code)
build_ispc_shared_lib(
tmpdir,
[("stream.ispc", ispc_code)],
[("tasksys.cpp", tasksys_source)],
cxx_options=["-g", "-fopenmp", "-DISPC_USE_OMP"],
ispc_options=([
"-g",
"-O0",
"--no-omit-frame-pointer",
"--target=avx2-i32x8",
#"--opt=force-aligned-memory",
"--opt=disable-loop-unroll",
#"--opt=fast-math",
"--opt=disable-fma",
"--addressing=32",
]),
ispc_bin="/home/ubuntu-boot/Desktop/ispc-v1.9.1-linux/ispc",
quiet=True,
)
knl_lib = ctypes.cdll.LoadLibrary(os.path.join(tmpdir, "shared.so"))
scalar = 5
a = empty_aligned(ARRAY_SIZE, dtype=STREAM_DTYPE, n=ALIGN_TO)
b = empty_aligned(ARRAY_SIZE, dtype=STREAM_DTYPE, n=ALIGN_TO)
c = empty_aligned(ARRAY_SIZE, dtype=STREAM_DTYPE, n=ALIGN_TO)
#print( hex(address_from_numpy(a)),
# hex(address_from_numpy(b)),
# hex(address_from_numpy(c)))
assert address_from_numpy(a) % ALIGN_TO == 0
assert address_from_numpy(b) % ALIGN_TO == 0
assert address_from_numpy(c) % ALIGN_TO == 0
if experiment == "copy":
g = knl_lib.stream_copy_tasks
x = [
cptr_from_numpy(a),
cptr_from_numpy(b),
INDEX_CTYPE(ARRAY_SIZE),
]
else:
g = knl_lib.stream_triad_tasks
x = [
cptr_from_numpy(a),
cptr_from_numpy(b),
cptr_from_numpy(c),
STREAM_CTYPE(scalar),
INDEX_CTYPE(ARRAY_SIZE),
]
for i in range(4):
g(*x)
def call_kernel():
g(*x)
for i in range(3):
call_kernel()
start_time = time()
for irun in range(NRUNS):
call_kernel()
elapsed = time() - start_time
print("Avg Time: ", elapsed / NRUNS)
by = 3 if experiment == "triad" else 2
print("MB: ", 1e-9 * by * a.nbytes * NRUNS / elapsed, "GB/s")