def test_match():
# pair each match keyword with whether it specifies a complete match or not
match_keywords = [("match", True), ("match?", False)]
for (match_keyword, is_complete) in match_keywords:
mod = relay.Module()
list_var = relay.GlobalTypeVar("List")
typ_var = relay.TypeVar("A")
cons_constructor = relay.Constructor(
"Cons", [typ_var, list_var(typ_var)], list_var)
nil_constructor = relay.Constructor("Nil", [], list_var)
list_def = relay.TypeData(list_var, [typ_var],
[cons_constructor, nil_constructor])
mod[list_var] = list_def
length_var = relay.GlobalVar("length")
typ_var = relay.TypeVar("A")
input_type = list_var(typ_var)
input_var = relay.Var("xs", input_type)
rest_var = relay.Var("rest")
cons_case = relay.Let(
_, UNIT,
relay.add(relay.const(1), relay.Call(length_var, [rest_var])))
body = relay.Match(input_var, [
relay.Clause(
relay.PatternConstructor(
cons_constructor,
[relay.PatternWildcard(),
relay.PatternVar(rest_var)]), cons_case),
relay.Clause(relay.PatternConstructor(nil_constructor, []),
relay.const(0))
],
complete=is_complete)
length_func = relay.Function([input_var], body, int32, [typ_var])
mod[length_var] = length_func
assert parses_as(
"""
%s
def @length[A](%%xs: List[A]) -> int32 {
%s (%%xs) {
Cons(_, %%rest) => {
();;
1 + @length(%%rest)
},
Nil => 0,
}
}
""" % (LIST_DEFN, match_keyword), mod)
def test_extern_ccompiler_default_ops():
def expected():
x = relay.var("x", shape=(8, 8))
y = relay.var("y", shape=(8, 8))
x0 = relay.var("x0", shape=(8, 8))
y0 = relay.var("y0", shape=(8, 8))
add = x0 + y0
# Function that uses C compiler
func = relay.Function([x0, y0], add)
func = func.set_attribute("Primitive", tvm.expr.IntImm("int32", 1))
func = func.set_attribute("Compiler",
tvm.expr.StringImm("ccompiler"))
func = func.set_attribute("ExternalSymbol",
tvm.expr.StringImm("ccompiler_0"))
add_call = relay.Call(func, [x, y])
# Function that uses default compiler. Ops are fused in this function.
p0 = relay.var("p0", shape=(8, 8))
log = relay.log(p0)
exp = relay.exp(p0)
concat = relay.concatenate([log, exp], axis=0)
fused_func = relay.Function([p0], concat)
fused_func = fused_func.set_attribute("Primitive",
tvm.expr.IntImm("int32", 1))
fused_call = relay.Call(fused_func, [add_call])
main = relay.Function([x, y], fused_call)
mod = relay.Module()
mod["main"] = main
return mod
x = relay.var("x", shape=(8, 8))
y = relay.var("y", shape=(8, 8))
add = x + y
log = relay.log(add)
exp = relay.exp(add)
concat = relay.concatenate([log, exp], axis=0)
f = relay.Function([x, y], concat)
mod = relay.Module()
mod["main"] = f
mod = WhiteListAnnotator(["add", "subtract", "multiply"], "ccompiler")(mod)
mod = transform.PartitionGraph()(mod)
fused_mod = transform.FuseOps(2)(mod)
expected_mod = expected()
assert relay.alpha_equal(fused_mod, expected_mod)
x_data = np.random.rand(8, 8).astype('float32')
y_data = np.random.rand(8, 8).astype('float32')
np_add = x_data + y_data
res = np.concatenate([np.log(np_add), np.exp(np_add)])
check_result(mod, {"x": x_data, "y": y_data}, (16, 8), res)
def run(dtype):
mod = relay.Module()
p = Prelude(mod)
take = p.get_var('tensor_take', dtype)
tensor2 = p.get_var('tensor2', dtype)
v = relay.var('v')
lower = relay.var('lower')
upper = relay.var('upper')
mod["main"] = relay.Function([v, lower, upper], take(tensor2(v), lower, upper))
v_data = np.random.uniform(size=(10, 10)).astype(dtype)
expected = [np.take(v_data, range(2, 5), axis=0)]
check_tensor_array(mod, expected, *(v_data, 2, 5), dtype=dtype)
expected = [np.take(v_data, range(0, 9), axis=0)]
check_tensor_array(mod, expected, *(v_data, 0, 9), dtype=dtype)
def verify_any_global_pool2d(pool_type, data_shape, layout, static_data_shape,
ref_out_shape):
mod = relay.Module()
dtype = "float32"
pool_func = relay.nn.global_max_pool2d if pool_type == "max" else relay.nn.global_avg_pool2d
data = relay.var('data', shape=data_shape, dtype=dtype)
y = pool_func(data, layout)
mod["main"] = relay.Function([data], y)
data_np = np.random.uniform(size=static_data_shape).astype(dtype)
for kind in ["debug", "vm"]:
ex = relay.create_executor(kind, mod=mod, ctx=tvm.cpu(), target="llvm")
result = ex.evaluate()(data_np)
assert result.asnumpy().shape == ref_out_shape, \
"Shape mismatch: expect %s but got %s." % (str(ref_out_shape), str(result.asnumpy().shape))
def test_tensorrt_not_compatible():
if should_skip():
return
dtype = 'float32'
xshape = (1, 32, 14, 14)
x = relay.var('x', shape=(xshape), dtype=dtype)
y = relay.add(x, x)
z = relay.erf(y)
out = relay.nn.relu(z)
f = relay.Function([x], out)
mod = relay.Module()
mod['main'] = f
mod = relay.tensorrt.EnableTrt(mod)
assert not mod['main'].attrs
def test_any_reshape_like():
mod = relay.Module()
dtype = "float32"
data = relay.var('data', shape=(relay.Any(), 3, 10), dtype=dtype)
shape_like = relay.var('data', shape=(relay.Any(), 5, 6), dtype=dtype)
y = relay.reshape_like(data, shape_like)
mod["main"] = relay.Function([data, shape_like], y)
data_np = np.random.uniform(size=(3, 3, 10)).astype(dtype)
shape_like_np = np.random.uniform(size=(3, 5, 6)).astype(dtype)
for kind in ["debug", "vm"]:
ex = relay.create_executor(kind, mod=mod, ctx=tvm.cpu(), target="llvm")
result = ex.evaluate()(data_np, shape_like_np)
assert result.asnumpy().shape == shape_like_np.shape, \
"Shape mismatch: expect %s but got %s." % (str(shape_like_np.shape), str(result.asnumpy().shape))
def test_ref():
mod = relay.Module()
three_with_ref = relay.GlobalVar('three_with_ref')
i = relay.Var('i')
iv = relay.Var('iv')
u = relay.Var('u')
uv = relay.Var('uv')
body = relay.add(iv, uv)
body = relay.Let(uv, relay.RefRead(i), body)
body = relay.Let(u, relay.RefWrite(i, relay.const(2)), body)
body = relay.Let(iv, relay.RefRead(i), body)
body = relay.Let(i, relay.RefCreate(relay.const(1)), body)
mod[three_with_ref] = relay.Function([], body)
check_eval(three_with_ref, [], 3, mod=mod)
def before():
x = relay.var("x", shape=(1, 64, 56, 56))
weight = relay.var('weight', shape=(64, 64, 3, 3))
y = relay.nn.conv2d(x,
weight,
channels=64,
kernel_size=(3, 3),
padding=(1, 1))
y = relay.nn.relu(y)
mod = relay.Module()
foo = relay.GlobalVar('foo')
mod[foo] = relay.Function([x, weight], y)
mod["main"] = relay.Function([x, weight], foo(x, weight))
return mod
def run(dtype):
x = relay.var('x')
mod = relay.Module()
p = Prelude(mod)
expand_dims_func = p.get_var('tensor_expand_dims', dtype)
tensor1 = p.get_var('tensor1', dtype)
mod["main"] = relay.Function([x], expand_dims_func(tensor1(x)))
for kind in ["debug"]:
ex = relay.create_executor(kind, mod=mod, ctx=tvm.cpu(), target="llvm")
x_np = np.random.uniform(size=(1,)).astype(dtype)
result = ex.evaluate()(x_np)
got = vmobj_to_list(result)
expected = [np.expand_dims(x_np, axis=0)]
tvm.testing.assert_allclose(expected, got)
def run(dtype):
mod = relay.Module()
p = Prelude(mod)
l = relay.var('l')
i = relay.var('i')
read_func = p.get_var('tensor_array_read', dtype)
tensor_array = p.get_var('tensor_array', dtype)
mod["main"] = relay.Function([l, i], read_func(tensor_array(l), i))
for kind in ["debug"]:
ex = relay.create_executor(kind, mod=mod, ctx=tvm.cpu(), target="llvm")
result = ex.evaluate()(10, 5)
got = vmobj_to_list(result)
expected = [0]
tvm.testing.assert_allclose(expected, got)
def test_multiple_cons_defn():
mod = relay.Module()
list_var = relay.GlobalTypeVar("List")
typ_var = relay.TypeVar("A")
prog = relay.TypeData(
list_var,
[typ_var],
[
relay.Constructor("Cons", [typ_var, list_var(typ_var)], list_var),
relay.Constructor("Nil", [], list_var),
])
mod[list_var] = prog
assert parses_as(LIST_DEFN, mod)
def create_vm(f, ctx=tvm.cpu(), target="llvm"):
if isinstance(f, relay.Expr):
mod = relay.Module()
mod["main"] = f
compiler = relay.vm.VMCompiler()
vm = compiler.compile(mod, target)
vm.init(ctx)
return vm
else:
assert isinstance(f, relay.Module), "expected mod as relay.Module"
compiler = relay.vm.VMCompiler()
vm = compiler.compile(f, target)
vm.init(ctx)
return vm
def test_list_tl_empty_list():
mod = relay.Module()
p = Prelude(mod)
nil = p.nil
l = p.l
tl = p.tl
f = relay.Function([], tl(nil()))
mod["main"] = f
result = veval(mod)
print(result)
def veval(f, *args, ctx=tvm.cpu(), target="llvm"):
if isinstance(f, relay.Expr):
mod = relay.Module()
mod["main"] = f
vm = relay.vm.compile(mod, target)
vm.init(tvm.cpu())
return vm.invoke("main", *args)
else:
assert isinstance(f, relay.Module), "expected expression or module"
mod = f
vm = relay.vm.compile(mod, target)
vm.init(tvm.cpu())
ret = vm.invoke("main", *args)
return ret
def test_simple_if():
x = relay.var('x', shape=(10, 10))
y = relay.var('y', shape=(10, 10))
f = relay.Function([x, y], relay.If(any(relay.op.equal(x, y)), x, y))
x_data = np.random.rand(10, 10).astype('float32')
y_data = np.random.rand(10, 10).astype('float32')
mod = relay.Module()
mod["main"] = f
# same
check_result([x_data, x_data], x_data, mod=mod)
# diff
check_result([x_data, y_data], y_data, mod=mod)
def test_match_var():
mod = relay.Module()
box, box_ctor = init_box_adt(mod)
v = relay.Var('v')
w = relay.Var('w')
match = relay.Let(
v, box_ctor(relay.const(1)),
relay.Match(v, [
relay.Clause(relay.PatternVar(w), w)
]))
match_val = run_as_python(match, mod)
assert_constructor_value(match_val, box_ctor, 1)
assert_tensor_value(match_val.fields[0], 1)
def test_add():
mod = relay.Module()
p = Prelude(mod)
nat = p.nat
add = p.add
s = p.s
z = p.z
ctx = tvm.context("llvm", 0)
intrp = create_executor(mod=mod, ctx=ctx, target="llvm")
assert mod[add].checked_type == relay.FuncType([nat(), nat()], nat())
assert count(intrp.evaluate(add(s(z()), s(z())))) == 2
assert count(intrp.evaluate(to_a_normal_form(add(s(z()), s(z())),
mod))) == 2
assert "let" in mod[add].astext()
def verify_any_split(data_shape, indices_or_sections, axis, static_data_shape,
ref_out_shape):
mod = relay.Module()
dtype = "float32"
data = relay.var('data', shape=data_shape, dtype=dtype)
y = relay.split(data, indices_or_sections, axis)
mod["main"] = relay.Function([data], y.astuple())
data_np = np.random.uniform(size=static_data_shape).astype(dtype)
for kind in ["vm"]:
ex = relay.create_executor(kind, mod=mod, ctx=tvm.cpu(), target="llvm")
result = ex.evaluate()(data_np)
for ret, ref_ret in zip(result, ref_out_shape):
assert ret.asnumpy().shape == ref_ret, \
"Shape mismatch: expect %s but got %s." % (str(ref_ret), str(ret.asnumpy().shape))
def test_extern_ccompiler():
x = relay.var('x', shape=(2, 2))
y = relay.var('y', shape=(2, 2))
z = x + x
p = y * y
f = relay.Function([x, y], p - z)
x_data = np.random.rand(2, 2).astype('float32')
y_data = np.random.rand(2, 2).astype('float32')
mod = relay.Module()
mod["main"] = f
mod = WhiteListAnnotator(["add", "subtract", "multiply"], "ccompiler")(mod)
mod = transform.PartitionGraph()(mod)
check_result(mod, {"x": x_data, "y": y_data}, (2, 2), (y_data * y_data) - (x_data + x_data))
def test_tensorrt_simple():
if should_skip():
return
dtype = 'float32'
xshape = (1, 32, 14, 14)
yshape = (1, 32, 1, 1)
zshape = (1, 1, 1, 1)
x = relay.var('x', shape=(xshape), dtype=dtype)
y = relay.var('y', shape=(yshape), dtype=dtype)
z = relay.var('z', shape=(zshape), dtype=dtype)
w = z * (x + y)
out = relay.nn.relu(w)
f = relay.Function([x, y, z], out)
mod = relay.Module()
mod['main'] = f
mod = relay.tensorrt.EnableTrt(mod)
ref_mod = relay.Module()
ref_mod['main'] = f
x_data = np.random.uniform(-1, 1, xshape).astype(dtype)
y_data = np.random.uniform(-1, 1, yshape).astype(dtype)
z_data = np.random.uniform(-1, 1, zshape).astype(dtype)
# Test against reference.
for kind in ["graph"]:
ex = relay.create_executor(kind, mod=mod, ctx=tvm.gpu(0), target='cuda')
# First execution will trigger build of TRT engine(s).
res = ex.evaluate()(x_data, y_data, z_data)
# TRT engine is reused for second execution.
res = ex.evaluate()(x_data, y_data, z_data)
ref_ex = relay.create_executor(kind, mod=ref_mod, ctx=tvm.cpu(0))
ref_res = ref_ex.evaluate()(x_data, y_data, z_data)
tvm.testing.assert_allclose(res.asnumpy(), ref_res.asnumpy(), rtol=1e-5)
def test_nested_match_pattern():
mod = relay.Module()
box, box_ctor = init_box_adt(mod)
v = relay.Var('v')
w = relay.Var('w')
match = relay.Let(
v, box_ctor(box_ctor(relay.const(2))),
relay.Match(v, [
relay.Clause(
relay.PatternConstructor(box_ctor, [
relay.PatternConstructor(box_ctor, [relay.PatternVar(w)])
]), w)
]))
match_val = run_as_python(match, mod)
assert_tensor_value(match_val, 2)
def test_globalvar_as_call_arg():
mod = relay.Module()
p = Prelude(mod)
tensor_array = p.get_var('tensor_array', 'int32')
tensor1 = p.get_var('tensor1', 'int32')
write = p.get_var('tensor_array_write', 'int32')
stack = p.get_var('tensor_array_stack', 'int32')
v = relay.var('v')
init_tensor_array = tensor_array(relay.const(3))
tensor_array1 = write(init_tensor_array, relay.const(0), tensor1(v))
tensor_array2 = stack(tensor_array1)
mod["main"] = relay.Function([v], tensor_array2)
mod = relay.transform.RemoveUnusedFunctions()(mod)
l = set([x[0].name_hint for x in mod.functions.items()])
assert 'tensor_array_int32' in l
def test_constructor():
mod = relay.Module()
box, box_ctor = init_box_adt(mod)
init_box_int = box_ctor(relay.const(1))
box_val_int = run_as_python(init_box_int, mod)
assert_constructor_value(box_val_int, box_ctor, 1)
assert_tensor_value(box_val_int.fields[0], 1)
init_box_tup = box_ctor(relay.Tuple([]))
box_val_tup = run_as_python(init_box_tup, mod)
assert_constructor_value(box_val_tup, box_ctor, 1)
assert_adt_len(box_val_tup.fields[0], 0)
def test_add_op_scalar():
"""
test_add_op_scalar:
fn (x, y) {
return x + y;
}
"""
mod = relay.Module()
x = relay.var('x', shape=())
y = relay.var('y', shape=())
func = relay.Function([x, y], relay.op.add(x, y))
x_data = np.array(10.0, dtype='float32')
y_data = np.array(1.0, dtype='float32')
mod["main"] = func
check_result([x_data, y_data], x_data + y_data, mod=mod)
def test_add_op_broadcast():
"""
test_add_op_broadcast:
fn (x, y) {
return x + y;
}
"""
mod = relay.Module()
x = relay.var('x', shape=(10, 5))
y = relay.var('y', shape=(1, 5))
func = relay.Function([x, y], relay.op.add(x, y))
x_data = np.random.rand(10, 5).astype('float32')
y_data = np.random.rand(1, 5).astype('float32')
mod["main"] = func
check_result([x_data, y_data], x_data + y_data, mod=mod)
def test_let_scalar():
sb = relay.ScopeBuilder()
x = relay.var('x', 'float32')
x1 = sb.let('x1', x)
xplusone = x1 + relay.const(42.0, 'float32')
sb.ret(xplusone)
body = sb.get()
f = relay.Function([x], body)
x_data = np.array(np.random.rand()).astype('float32')
mod = relay.Module()
mod["main"] = f
check_result([x_data], x_data + 42.0, mod=mod)
def run(dtype):
mod = relay.Module()
p = Prelude(mod)
concat = p.get_var('tensor_concatenate', dtype)
tensor1 = p.get_var('tensor1', dtype)
v1 = relay.var('v1')
v2 = relay.var('v2')
mod["main"] = relay.Function([v1, v2], concat(tensor1(v1),
tensor1(v2)))
v1_data = np.random.uniform(low=0.0, high=8.0,
size=(5, )).astype(dtype)
v2_data = np.random.uniform(low=0.0, high=8.0,
size=(5, )).astype(dtype)
expected = [np.concatenate((v1_data, v2_data))]
check_tensor_array(mod, expected, *(v1_data, v2_data), dtype=dtype)
def run(dtype):
mod = relay.Module()
p = Prelude(mod)
v1 = relay.var('v1')
v2 = relay.var('v2')
tensor_array = p.get_var('tensor_array', dtype)
init_tensor_array = tensor_array(relay.const(2))
write_func = p.get_var('tensor_array_write', dtype)
tensor1 = p.get_var('tensor1', dtype)
tensor_array1 = write_func(init_tensor_array, relay.const(0),
tensor1(v1))
tensor_array2 = write_func(tensor_array1, relay.const(1), tensor1(v2))
mod["main"] = relay.Function([v1, v2], tensor_array2)
expected = [3, 7]
check_tensor_array(mod, expected, *(3, 7), dtype=dtype)
def veval(f, *args, ctx=tvm.cpu()):
if isinstance(f, relay.Expr):
ex = relay.create_executor('vm', mod=relay.Module(), ctx=ctx)
if len(args) == 0:
return ex.evaluate(f)
else:
return ex.evaluate(f)(*args)
else:
assert isinstance(f, relay.Module), "expected expression or module"
mod = f
ex = relay.create_executor('vm', mod=mod, ctx=ctx)
if len(args) == 0:
return ex.evaluate()
else:
return ex.evaluate()(*args)
def verify_any_dense(data_shape, weight_shape, units, static_data_shape,
static_weight_shape, ref_out_shape):
mod = relay.Module()
dtype = "float32"
data = relay.var('data', shape=data_shape, dtype=dtype)
weight = relay.var('weight', shape=weight_shape, dtype=dtype)
y = relay.nn.dense(data, weight, units)
mod["main"] = relay.Function([data, weight], y)
data_np = np.random.uniform(size=static_data_shape).astype(dtype)
weight_np = np.random.uniform(size=static_weight_shape).astype(dtype)
for kind in ["debug", "vm"]:
ex = relay.create_executor(kind, mod=mod, ctx=tvm.cpu(), target="llvm")
result = ex.evaluate()(data_np, weight_np)
assert result.asnumpy().shape == ref_out_shape, \
"Shape mismatch: expect %s but got %s." % (str(ref_out_shape), str(result.asnumpy().shape))