gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[FRONTEND] Added support for element-wise function defined in external LLVM bitcode (e.g., libdevice) (#562)

Browse Source
This commit is contained in:
Keren Zhou
2022-07-13 15:52:21 -07:00
committed by GitHub
parent 971f5782b4
commit 4912916c11
24 changed files with 2634 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
python/test/unit/language/test_core.py
View File

@@ -1300,3 +1300,49 @@ def test_num_warps_pow2():
_kernel[(1,)](dst=dst, num_warps=1)
_kernel[(1,)](dst=dst, num_warps=2)
_kernel[(1,)](dst=dst, num_warps=4)
# -------------
# test extern
# -------------
@pytest.mark.parametrize("dtype_str, expr, lib_path",
[('int32', 'libdevice.ffs', ''),
('float32', 'libdevice.pow', '/usr/local/cuda/nvvm/libdevice/libdevice.10.bc'),
('float64', 'libdevice.norm4d', '')])
def test_libdevice(dtype_str, expr, lib_path):
@triton.jit
def kernel(X, Y, BLOCK: tl.constexpr):
x = tl.load(X + tl.arange(0, BLOCK))
y = GENERATE_TEST_HERE
tl.store(Y + tl.arange(0, BLOCK), y)
shape = (128, )
rs = RandomState(17)
# limit the range of integers so that the sum does not overflow
x = numpy_random(shape, dtype_str=dtype_str, rs=rs)
if expr == 'libdevice.ffs':
kernel = patch_kernel(kernel, {'GENERATE_TEST_HERE': 'tl.libdevice.ffs(x)'})
y_ref = np.zeros(shape, dtype=x.dtype)
for i in range(shape[0]):
y_ref[i] = (int(x[i]) & int(-x[i])).bit_length()
elif expr == 'libdevice.pow':
# numpy does not allow negative factors in power, so we use abs()
x = np.abs(x)
kernel = patch_kernel(kernel, {'GENERATE_TEST_HERE': 'tl.libdevice.pow(x, x)'})
y_ref = np.power(x, x)
elif expr == 'libdevice.norm4d':
kernel = patch_kernel(kernel, {'GENERATE_TEST_HERE': 'tl.libdevice.norm4d(x, x, x, x)'})
y_ref = np.sqrt(4 * np.power(x, 2))
x_tri = to_triton(x)
# triton result
y_tri = to_triton(numpy_random((shape[0],), dtype_str=dtype_str, rs=rs), device='cuda')
kernel[(1,)](x_tri, y_tri, BLOCK=shape[0], extern_libs={'libdevice': lib_path})
# compare
if expr == 'libdevice.ffs':
np.testing.assert_equal(y_ref, to_numpy(y_tri))
else:
np.testing.assert_allclose(y_ref, to_numpy(y_tri), rtol=0.01)