88 lines
2.9 KiB
Python
88 lines
2.9 KiB
Python
import platform
|
|
import numpy as np
|
|
from numba import types
|
|
import unittest
|
|
from numba import njit
|
|
from numba.core import config
|
|
from numba.tests.support import TestCase
|
|
|
|
_DEBUG = False
|
|
if _DEBUG:
|
|
from llvmlite import binding as llvm
|
|
# Prints debug info from the LLVMs vectorizer
|
|
llvm.set_option("", "--debug-only=loop-vectorize")
|
|
|
|
|
|
_skylake_env = {
|
|
"NUMBA_CPU_NAME": "skylake-avx512",
|
|
"NUMBA_CPU_FEATURES": "",
|
|
}
|
|
|
|
|
|
@unittest.skipIf(platform.machine() != 'x86_64', 'x86_64 only test')
|
|
class TestVectorization(TestCase):
|
|
"""
|
|
Tests to assert that code which should vectorize does indeed vectorize
|
|
"""
|
|
def gen_ir(self, func, args_tuple, fastmath=False):
|
|
self.assertEqual(config.CPU_NAME, "skylake-avx512")
|
|
self.assertEqual(config.CPU_FEATURES, "")
|
|
|
|
jitted = njit(args_tuple, fastmath=fastmath)(func)
|
|
return jitted.inspect_llvm(args_tuple)
|
|
|
|
@TestCase.run_test_in_subprocess(envvars=_skylake_env)
|
|
def test_nditer_loop(self):
|
|
# see https://github.com/numba/numba/issues/5033
|
|
def do_sum(x):
|
|
acc = 0
|
|
for v in np.nditer(x):
|
|
acc += v.item()
|
|
return acc
|
|
|
|
llvm_ir = self.gen_ir(do_sum, (types.float64[::1],), fastmath=True)
|
|
self.assertIn("vector.body", llvm_ir)
|
|
self.assertIn("llvm.loop.isvectorized", llvm_ir)
|
|
|
|
# SLP is off by default due to miscompilations, see #8705. Put this into a
|
|
# subprocess to isolate any potential issues.
|
|
@TestCase.run_test_in_subprocess(
|
|
envvars={'NUMBA_SLP_VECTORIZE': '1', **_skylake_env},
|
|
)
|
|
def test_slp(self):
|
|
# Sample translated from:
|
|
# https://www.llvm.org/docs/Vectorizers.html#the-slp-vectorizer
|
|
|
|
def foo(a1, a2, b1, b2, A):
|
|
A[0] = a1 * (a1 + b1)
|
|
A[1] = a2 * (a2 + b2)
|
|
A[2] = a1 * (a1 + b1)
|
|
A[3] = a2 * (a2 + b2)
|
|
|
|
ty = types.float64
|
|
llvm_ir = self.gen_ir(foo, ((ty,) * 4 + (ty[::1],)), fastmath=True)
|
|
self.assertIn("2 x double", llvm_ir)
|
|
|
|
@TestCase.run_test_in_subprocess(envvars=_skylake_env)
|
|
def test_instcombine_effect(self):
|
|
# Without instcombine running ahead of refprune, the IR has refops that
|
|
# are trivially prunable (same BB) but the arguments are obfuscated
|
|
# through aliases etc. The follow case triggers this situation as the
|
|
# typed.List has a structproxy call for computing `len` and getting the
|
|
# base pointer for use in iteration.
|
|
|
|
def sum_sqrt_list(lst):
|
|
acc = 0.0
|
|
for item in lst:
|
|
acc += np.sqrt(item)
|
|
return acc
|
|
|
|
llvm_ir = self.gen_ir(sum_sqrt_list, (types.ListType(types.float64),),
|
|
fastmath=True)
|
|
self.assertIn("vector.body", llvm_ir)
|
|
self.assertIn("llvm.loop.isvectorized", llvm_ir)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|