def test_reduce_map(target, dev, ref_data, in_shape, axis, keepdims,
reduce_type, dtype):
target = tvm.target.Target(target)
if target.kind.name == "vulkan" and reduce_type in ["sum", "any", "all"]:
pytest.xfail(f"Vulkan backend has known errors on {reduce_type}")
in_npy, in_npy_map, out_npy = ref_data
# Build the logic and compile the function
A = te.placeholder(shape=in_shape, name="A", dtype=dtype)
A1 = topi.sqrt(topi.exp(A))
out_dtype = dtype
if reduce_type == "sum":
B = topi.sum(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "all":
B = topi.all(A, axis=axis, keepdims=keepdims)
elif reduce_type == "any":
B = topi.any(A, axis=axis, keepdims=keepdims)
elif reduce_type == "max":
B = topi.max(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "min":
B = topi.min(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "argmax":
B = topi.argmax(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
elif reduce_type == "argmin":
B = topi.argmin(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
else:
raise NotImplementedError
with tvm.target.Target(target):
s = tvm.topi.testing.get_reduce_schedule(target)(B)
foo = tvm.build(s, [A, B], target, name=reduce_type)
data_tvm = tvm.nd.array(in_npy, device=dev)
out_tvm = tvm.nd.empty(shape=out_npy.shape, device=dev, dtype=out_dtype)
foo(data_tvm, out_tvm)
if reduce_type == "argmax" or reduce_type == "argmin":
out_tvm_indices = out_tvm.numpy()
if keepdims:
out_tvm_indices = np.take(out_tvm_indices, indices=0, axis=axis)
if axis is None:
out_tvm_val = in_npy_map.ravel()[out_tvm_indices]
else:
other_indices = tuple(
np.indices(in_shape[0:axis] + in_shape[(axis + 1):]))
sel_indices = other_indices[0:axis] + (
out_tvm_indices, ) + other_indices[axis:]
out_tvm_val = in_npy_map[sel_indices]
if reduce_type == "argmax":
tvm.testing.assert_allclose(out_tvm_val, in_npy_map.max(axis=axis),
1e-3, 1e-3)
elif reduce_type == "argmin":
tvm.testing.assert_allclose(out_tvm_val, in_npy_map.min(axis=axis),
1e-3, 1e-3)
else:
tvm.testing.assert_allclose(out_tvm.numpy(), out_npy, 1e-3, 1e-3)
def test_reduce_map(in_shape, axis, keepdims, type="sum", test_id=0):
global TASK
# Build the logic and compile the function
A = te.placeholder(shape=in_shape, name="A")
if type == "sum":
TASK = "sum_map_id%d" % test_id
B = topi.sum(A, axis=axis, keepdims=keepdims)
elif type == "max":
TASK = "max_map_id%d" % test_id
B = topi.max(A, axis=axis, keepdims=keepdims)
elif type == "min":
TASK = "min_map_id%d" % test_id
B = topi.min(A, axis=axis, keepdims=keepdims)
else:
raise NotImplementedError
s = topi.cuda.schedule_reduce(B)
with tvm.transform.PassContext(
config={"tir.UnrollLoop": {
"auto_max_step": 16,
}}):
fcuda = tvm.build(s, [A, B], "cuda", name="sum")
# Test
in_npy = np.random.normal(size=in_shape).astype(np.float32)
if type == "sum":
out_npy = in_npy.sum(axis=axis, keepdims=keepdims)
elif type == "max":
out_npy = in_npy.max(axis=axis, keepdims=keepdims)
elif type == "min":
out_npy = in_npy.min(axis=axis, keepdims=keepdims)
else:
raise NotImplementedError
data_tvm = tvm.nd.array(in_npy, device=tvm.cuda())
out_tvm = tvm.nd.empty(shape=out_npy.shape, device=tvm.cuda())
for _ in range(2):
fcuda(data_tvm, out_tvm)
tvm.testing.assert_allclose(out_tvm.asnumpy(),
out_npy,
rtol=4e-4,
atol=4e-4)
def verify_reduce_map_ele(in_shape,
axis,
keepdims,
type="sum",
dtype="float32"):
# Build the logic and compile the function
A = te.placeholder(shape=in_shape, name="A", dtype=dtype)
A1 = topi.sqrt(topi.exp(A))
out_dtype = dtype
if type == "sum":
B = topi.sum(A1, axis=axis, keepdims=keepdims)
elif type == "all":
B = topi.all(A, axis=axis, keepdims=keepdims)
elif type == "any":
B = topi.any(A, axis=axis, keepdims=keepdims)
elif type == "max":
B = topi.max(A1, axis=axis, keepdims=keepdims)
elif type == "min":
B = topi.min(A1, axis=axis, keepdims=keepdims)
elif type == "argmax":
B = topi.argmax(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
elif type == "argmin":
B = topi.argmin(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
else:
raise NotImplementedError
def check_device(device, ctx):
print("Running on target: %s" % device)
with tvm.target.Target(device):
s = tvm.topi.testing.get_reduce_schedule(device)(B)
foo = tvm.build(s, [A, B], device, name=type)
# Test
if dtype == "bool":
in_npy_map = in_npy = np.random.choice([True, False],
size=in_shape)
else:
in_npy = np.random.uniform(-1, 1, size=in_shape).astype(dtype)
in_npy_map = np.sqrt(np.exp(in_npy)).astype(dtype)
if type == "sum":
out_npy = in_npy_map.sum(axis=axis, keepdims=keepdims)
elif type == "all" and dtype == "bool":
out_npy = in_npy_map.all(axis=axis, keepdims=keepdims)
elif type == "any" and dtype == "bool":
out_npy = in_npy_map.any(axis=axis, keepdims=keepdims)
elif type == "max":
out_npy = in_npy_map.max(axis=axis, keepdims=keepdims)
elif type == "min":
out_npy = in_npy_map.min(axis=axis, keepdims=keepdims)
elif type == "argmax":
out_npy = _my_npy_argmax(in_npy_map, axis=axis, keepdims=keepdims)
elif type == "argmin":
out_npy = _my_npy_argmin(in_npy_map, axis=axis, keepdims=keepdims)
else:
raise NotImplementedError
data_tvm = tvm.nd.array(in_npy, ctx=ctx)
out_tvm = tvm.nd.empty(shape=out_npy.shape, ctx=ctx, dtype=out_dtype)
for _ in range(1):
foo(data_tvm, out_tvm)
if type == "argmax" or type == "argmin":
out_tvm_indices = out_tvm.asnumpy()
if keepdims:
out_tvm_indices = np.take(out_tvm_indices,
indices=0,
axis=axis)
if axis is None:
out_tvm_val = in_npy_map.ravel()[out_tvm_indices]
else:
other_indices = tuple(
np.indices(in_shape[0:axis] + in_shape[(axis + 1):]))
sel_indices = other_indices[0:axis] + (
out_tvm_indices, ) + other_indices[axis:]
out_tvm_val = in_npy_map[sel_indices]
if type == "argmax":
tvm.testing.assert_allclose(out_tvm_val,
in_npy_map.max(axis=axis), 1e-3,
1e-3)
elif type == "argmin":
tvm.testing.assert_allclose(out_tvm_val,
in_npy_map.min(axis=axis), 1e-3,
1e-3)
else:
tvm.testing.assert_allclose(out_tvm.asnumpy(), out_npy, 1e-3, 1e-3)
for device, ctx in tvm.testing.enabled_targets():
check_device(device, ctx)
def test_reduce_map(hexagon_session: Session, ref_data, in_shape, axis,
keepdims, reduce_type, dtype):
in_npy, in_npy_map, out_npy = ref_data
# Build the logic and compile the function
A = te.placeholder(shape=in_shape, name="A", dtype=dtype)
A1 = topi.sqrt(topi.exp(A))
out_dtype = dtype
if reduce_type == "sum":
B = topi.sum(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "all":
B = topi.all(A, axis=axis, keepdims=keepdims)
elif reduce_type == "any":
B = topi.any(A, axis=axis, keepdims=keepdims)
elif reduce_type == "max":
B = topi.max(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "min":
B = topi.min(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "argmax":
B = topi.argmax(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
elif reduce_type == "argmin":
B = topi.argmin(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
else:
raise NotImplementedError
target_hexagon = tvm.target.hexagon("v68")
with tvm.target.Target(target_hexagon):
fschedule = topi.hexagon.schedule_reduce
s = fschedule(B)
func = tvm.build(s, [A, B],
tvm.target.Target(target_hexagon, host=target_hexagon),
name=reduce_type)
mod = hexagon_session.load_module(func)
dev = hexagon_session.device
data_tvm = tvm.nd.array(in_npy, device=dev)
out_tvm = tvm.nd.empty(shape=out_npy.shape, device=dev, dtype=out_dtype)
mod[reduce_type](data_tvm, out_tvm)
if reduce_type == "argmax" or reduce_type == "argmin":
out_tvm_indices = out_tvm.numpy()
if keepdims:
out_tvm_indices = np.take(out_tvm_indices, indices=0, axis=axis)
if axis is None:
out_tvm_val = in_npy_map.ravel()[out_tvm_indices]
else:
other_indices = tuple(
np.indices(in_shape[0:axis] + in_shape[(axis + 1):]))
sel_indices = other_indices[0:axis] + (
out_tvm_indices, ) + other_indices[axis:]
out_tvm_val = in_npy_map[sel_indices]
if reduce_type == "argmax":
tvm.testing.assert_allclose(out_tvm_val, in_npy_map.max(axis=axis),
1e-3, 1e-3)
elif reduce_type == "argmin":
tvm.testing.assert_allclose(out_tvm_val, in_npy_map.min(axis=axis),
1e-3, 1e-3)
else:
tvm.testing.assert_allclose(out_tvm.numpy(), out_npy, 1e-3, 1e-3)