def before(x, conv_weight, out_bias, out_scale, channels):
args = [x, conv_weight, out_bias]
y0 = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y0 = relay.multiply(y0, out_scale)
y0 = relay.nn.relu(y0)
y1 = relay.nn.conv2d(y0, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y1 = relay.multiply(y1, out_scale)
y1 = relay.nn.relu(y1)
y2 = relay.nn.conv2d(y0, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y2 = relay.multiply(y2, out_scale)
y2 = relay.nn.relu(y2)
y = relay.add(y1, y2)
return relay.Function(args, y)
def before(x, w1, w2, scale1, scale2):
args = [x, w1, w2, scale1, scale2]
y1 = relay.nn.conv2d(x, w1)
y1 = relay.multiply(y1, scale1)
y2 = relay.nn.conv2d(x, w2)
y2 = relay.multiply(y2, scale2)
y = relay.Tuple((y1, y2))
return relay.Function(args, y)
def expected(x, w1, w2, scale1, scale2, channels1, channels2):
args = [x, w1, w2, scale1, scale2]
w = relay.concatenate((w1, w2), axis=0)
y = relay.nn.conv2d(x, w, channels=channels1 + channels2)
y1 = relay.strided_slice(y, [0, 0], [None, channels1])
y2 = relay.strided_slice(y, [0, channels1], [None, channels1 + channels2])
y1 = relay.multiply(y1, scale1)
y2 = relay.multiply(y2, scale2)
y = relay.Tuple((y1, y2))
return relay.Function(args, y)
def simple_bn(x, gamma, beta, moving_mean, moving_var,
axis=1, epsilon=1e-5, shape=None):
# expect = (x - moving_mean) / sqrt(moving_var + eps) * gamma + beta
scale = rly.multiply(rly.const(1, 'float32') /
rly.sqrt(moving_var + rly.const(epsilon, 'float32')), gamma)
shift = rly.add(
rly.multiply(rly.negative(moving_mean), scale), beta)
num_newaxis = len(shape) - (axis + 1)
if num_newaxis:
scale = rly.expand_dims(scale, axis=1, num_newaxis=num_newaxis)
shift = rly.expand_dims(shift, axis=1, num_newaxis=num_newaxis)
return x * scale + shift
def expected():
add = relay.add(a, b)
mul = relay.multiply(c, d)
copy_mul_sub = relay.device_copy(mul, cpu_ctx, dev_ctx)
sub = relay.subtract(add, copy_mul_sub)
func = relay.Function([a, b, c, d], sub)
return func
def test_mul_param():
x = relay.var('x', shape=(10, 10))
y = relay.var('y', shape=(1, 10))
func = relay.Function([x, y], relay.multiply(x, y))
x_data = np.random.rand(10, 10).astype('float32')
y_data = np.random.rand(1, 10).astype('float32')
check_eval(func, [x_data, y_data], x_data * y_data)
def expected(x, conv_weight, out_bias, out_scale, channels):
# use a fixed order of args so alpha equal check can pass
args = [x, conv_weight, out_bias]
out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
squeezed_scale = relay.squeeze(out_scale, axis=[1,2])
conv_weight = relay.multiply(
conv_weight , relay.expand_dims(squeezed_scale, axis=1, num_newaxis=3))
y = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
out_bias = relay.multiply(out_bias,
relay.expand_dims(squeezed_scale, axis=1, num_newaxis=2))
y = relay.add(y, out_bias)
y = relay.nn.relu(y)
return relay.Function(args, y)
def before(x, conv_weight, out_scale, channels):
args = [x, conv_weight]
y = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y = relay.multiply(y, out_scale)
return relay.Function(args, y)
def before():
c = relay.const(c_data)
x = relay.var("x")
y = relay.add(c, c)
y = relay.multiply(y, relay.const(2, "float32"))
y = relay.add(x, y)
z = relay.add(y, c)
return relay.Function([x], z)
def before(x, conv_weight, out_scale, channels):
y = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
data_layout="NCHW",
padding=(1, 1))
y = relay.nn.relu(y)
y = relay.multiply(x, out_scale)
return relay.Function(relay.ir_pass.free_vars(y), y)
def expected():
x = relay.var("x", shape=(1, 64, 56, 56))
weight = relay.var('weight', shape=(64, 64, 3, 3))
y = relay.nn.conv2d(x, relay.multiply(weight, relay.const(2.0, "float32")),
channels=64,
kernel_size=(3, 3),
padding=(1, 1))
y = relay.nn.relu(y)
y = relay.Function([x, weight], y)
return y
def before():
x = relay.var("x", shape=(1, 64, 56, 56))
bias = relay.var("bias", shape=(64,))
scale = relay.var("scale", shape=(64, 1, 1))
weight = relay.var("weight")
y = relay.nn.conv2d(x, weight, channels=64, kernel_size=(3, 3), padding=(1, 1))
y = relay.nn.bias_add(y, bias) # test broadcasting to lhs
y = relay.multiply(scale, y) # test broadcasting to rhs
y = relay.Function(free_vars(y), y)
return y
def before(x, conv_weight, in_bias, in_scale, channels):
x = relay.multiply(x, in_scale)
xx = relay.nn.leaky_relu(x, alpha=0.1)
y1 = relay.nn.conv2d(xx, conv_weight,
channels=channels,
kernel_size=(3, 3),
data_layout="NHWC",
padding=(1, 1))
z = relay.add(y1, x)
return relay.Function(relay.ir_pass.free_vars(z), z)
def before(x, conv_weight, out_bias, out_scale, channels):
args = [x, conv_weight, out_bias]
out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
y = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y = relay.add(y, out_bias)
y = relay.nn.relu(y)
y = relay.multiply(y, out_scale)
return relay.Function(args, y)
def fail2(x, conv_weight, out_bias, out_scale, channels):
args = [x, conv_weight, out_bias]
out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
y1 = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y2 = relay.nn.relu(y1)
# fold will fail because y1 is referred also by y2
y1 = relay.multiply(y1, out_scale)
y = relay.add(y1, y2)
return relay.Function(args, y)
def before(x, conv_weight, in_bias, in_scale, channels):
args = [x, conv_weight, in_bias, in_scale]
in_scale = relay.expand_dims(in_scale, axis=1, num_newaxis=2)
in_bias = relay.expand_dims(in_bias, axis=1, num_newaxis=2)
x = relay.multiply(x, in_scale)
x = relay.nn.relu(x)
x = relay.add(x, in_bias)
y = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
return relay.Function(args, y)
def expected(x, conv_weight, in_bias, in_scale, channels):
args = [x, conv_weight, in_bias, in_scale]
x = relay.nn.relu(x)
in_bias = relay.divide(in_bias, in_scale)
x = relay.subtract(x, in_bias)
y1 = relay.nn.conv2d(x,
relay.multiply(conv_weight, in_scale),
channels=channels,
kernel_size=(3, 3),
data_layout="NHWC",
weight_layout="HWIO",
groups=channels,
padding=(1, 1))
y2 = relay.nn.conv2d(x,
relay.multiply(conv_weight, in_scale),
channels=channels,
kernel_size=(3, 3),
data_layout="NHWC",
weight_layout="HWIO",
groups=channels,
padding=(1, 1))
z = relay.add(y1, y2)
return relay.Function(args, z)
def expected():
x = relay.var("x", shape=(1, 64, 56, 56))
weight = relay.var('weight', shape=(64, 64, 3, 3))
y = relay.nn.conv2d(x,
relay.multiply(weight, relay.const(2.0,
"float32")),
channels=64,
kernel_size=(3, 3),
padding=(1, 1))
y = relay.nn.relu(y)
mod = tvm.IRModule()
foo = relay.GlobalVar('foo')
mod[foo] = relay.Function([x, weight], y)
mod["main"] = relay.Function([x, weight], foo(x, weight))
return mod
def expected():
x = relay.var("x", shape=(1, 500, 500, 64))
kernel = relay.var('kernel', shape=(3, 3, 64, 64), dtype='float32')
bias = relay.var("bias", shape=(64,))
multiplier1 = relay.var('multiplier1', shape=(1, ), dtype='float32')
multiplier2 = relay.var('multiplier2', shape=(1, 1), dtype='float32')
b = relay.expand_dims(bias, axis=0, num_newaxis=3)
b = relay.layout_transform(b, "NHWC", "NCHW16c")
y = relay.layout_transform(x, "NHWC", "NCHW16c")
y = relay.nn.conv2d(y, kernel,
data_layout='NCHW16c',
kernel_layout="HWIO",
kernel_size=(3, 3))
y = relay.add(b, y)
y = relay.nn.relu(y)
y = relay.multiply(multiplier1, y)
y = relay.multiply(y, multiplier2)
y = relay.layout_transform(y, "NCHW16c", "NHWC")
y = relay.Function(analysis.free_vars(y), y)
return y
def test_qemu_make_fail(temp_dir, board, west_cmd, tvm_debug):
"""Testing QEMU make fail."""
if board not in ["qemu_x86", "mps2_an521", "mps3_an547"]:
pytest.skip(msg="Only for QEMU targets.")
model = test_utils.ZEPHYR_BOARDS[board]
build_config = {"debug": tvm_debug}
shape = (10, )
dtype = "float32"
# Construct Relay program.
x = relay.var("x", relay.TensorType(shape=shape, dtype=dtype))
xx = relay.multiply(x, x)
z = relay.add(xx, relay.const(np.ones(shape=shape, dtype=dtype)))
func = relay.Function([x], z)
ir_mod = tvm.IRModule.from_expr(func)
target = tvm.target.target.micro(model)
executor = Executor("aot")
runtime = Runtime("crt")
with tvm.transform.PassContext(opt_level=3,
config={"tir.disable_vectorize": True}):
lowered = relay.build(ir_mod,
target,
executor=executor,
runtime=runtime)
sample = np.zeros(shape=shape, dtype=dtype)
project, project_dir = test_utils.generate_project(temp_dir,
board,
west_cmd,
lowered,
build_config,
sample,
shape,
dtype,
load_cmsis=False)
file_path = (pathlib.Path(project_dir) / "build" / "zephyr" /
"CMakeFiles" / "run.dir" / "build.make")
assert file_path.is_file(), f"[{file_path}] does not exist."
# Remove a file to create make failure.
os.remove(file_path)
project.flash()
with pytest.raises(server.JSONRPCError) as excinfo:
project.transport().open()
assert "QEMU setup failed" in str(excinfo.value)
def before(x, conv_weight, in_bias, in_scale, channels, blocking):
args = [x, conv_weight, in_bias]
x = relay.multiply(x, in_scale)
x = relay.nn.relu(x)
x = relay.add(x, in_bias)
y = relay.nn.conv2d(
x,
conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1),
data_layout="NCHW{}c".format(blocking[0]) if blocking else "NCHW",
kernel_layout="OIHW2i{}o".format(blocking[1])
if blocking else "OIHW")
return relay.Function(args, y)
def expected(x, conv_weight, in_bias, in_scale, channels):
# use a fixed order of args so alpha equal check can pass
args = [x, conv_weight, in_bias, in_scale]
in_scale = relay.expand_dims(in_scale, axis=1, num_newaxis=2)
in_bias = relay.expand_dims(in_bias, axis=1, num_newaxis=2)
squeezed_scale = relay.squeeze(in_scale, axis=[1,2])
x = relay.nn.relu(x)
in_bias = relay.divide(in_bias, relay.expand_dims(squeezed_scale, axis=1, num_newaxis=2))
x = relay.add(x, in_bias)
conv_weight = relay.multiply(
conv_weight , relay.expand_dims(squeezed_scale, axis=1, num_newaxis=2))
y = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
return relay.Function(args, y)
def expected(x, w1, w2, scale1, scale2, bias, channels1, channels2):
args = [x, w1, w2, scale1, scale2, bias]
w = relay.concatenate((w1, w2), axis=0)
scale = relay.concatenate((scale1, scale2), axis=0)
y = relay.nn.conv2d(x, w, channels=channels1 + channels2)
y = relay.multiply(y, scale)
y = relay.nn.relu(y)
y1 = relay.strided_slice(
y, begin=[0, 0], end=[-1, channels1], strides=[1, 1], slice_mode="size"
)
y2 = relay.strided_slice(
y, begin=[0, channels1], end=[-1, channels2], strides=[1, 1], slice_mode="size"
)
y2 = relay.add(y2, bias)
y = relay.Tuple((y1, y2))
return relay.Function(args, y)
def annotated():
add = relay.add(a, b)
_add = relay.annotation.on_device(add, dev_ctx)
mul = relay.multiply(c, d)
_mul = relay.annotation.on_device(mul, cpu_ctx)
sub = relay.subtract(add, mul)
_sub = relay.annotation.on_device(sub, dev_ctx)
func = relay.Function([a, b, c, d],
relay.Tuple(tvm.convert([_add, _mul,
_sub, sub])))
func = relay.ir_pass.infer_type(func)
func = relay.ir_pass.rewrite_annotated_ops(func,
dev_ctx.device_type)
func = relay.ir_pass.infer_type(func)
return relay.Function(relay.ir_pass.free_vars(func.body[3]),
func.body[3])
def test_mul_param(interface_api, use_unpacked_api, test_runner):
x = relay.var("x", shape=(10, 10))
y = relay.var("y", shape=(1, 10))
func = relay.Function([x, y], relay.multiply(x, y))
x_data = np.random.rand(10, 10).astype("float32")
y_data = np.random.rand(1, 10).astype("float32")
inputs = OrderedDict([("x", x_data), ("y", y_data)])
output_list = generate_ref_data(func, inputs)
compile_and_run(
AOTTestModel(module=IRModule.from_expr(func), inputs=inputs, outputs=output_list),
test_runner,
interface_api,
use_unpacked_api,
)
def fail2(x, conv_weight, out_bias, out_scale, in_channels, channels, blocking):
args = [x, conv_weight, out_bias]
y1 = relay.nn.conv2d(
x,
conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1),
data_layout="NCHW{}c".format(blocking[0]) if blocking else "NCHW",
kernel_layout="OIHW1i{}o".format(blocking[1]) if blocking else "OIHW",
)
y2 = relay.nn.relu(y1)
# fold will fail because y1 is referred also by y2
y1 = relay.multiply(y1, out_scale)
y = relay.add(y1, y2)
return relay.Function(args, y)
def expected():
x = relay.var("x", shape=(1, 64, 56, 56))
bias = relay.var("bias", shape=(64,))
scale = relay.var("scale", shape=(64, 1, 1))
weight = relay.var("weight")
x = relay.layout_transform(x, "NCHW", "NCHW16c")
bias = relay.expand_dims(bias, 1, 2)
bias = relay.layout_transform(bias, "CHW", "CHW16c")
scale = relay.layout_transform(scale, "CHW", "CHW16c")
y = relay.nn.conv2d(x, weight, channels=64, kernel_size=(3, 3), padding=(1, 1),
data_layout="NCHW16c")
y = relay.add(y, bias) # test broadcasting to lhs
y = relay.multiply(scale, y) # test broadcasting to rhs
y = relay.layout_transform(y, "NCHW16c", "NCHW")
y = relay.Function(free_vars(y), y)
return y
def make_qgraph(data, weight):
out = data * relay.const(32.0)
out = relay.round(out)
out = relay.clip(out, a_min=-127, a_max=127)
out = out.astype('int8')
out = relay.nn.conv2d(out,
weight,
kernel_size=(3, 3),
padding=(1, 1),
channels=c,
out_dtype='int32')
out = out.astype('float32')
out = relay.multiply(out, relay.const(0.00024414062))
out = relay.Function(relay.ir_pass.free_vars(out), out)
return out
def expected():
x = relay.var("x", shape=(1, 64, 56, 56))
bias = relay.var("bias", shape=(64,))
scale = relay.var("scale", shape=(64, 1, 1))
weight = relay.var("weight")
x = relay.layout_transform(x, "NCHW", "NCHW16c")
bias = relay.expand_dims(bias, 1, 2)
bias = relay.layout_transform(bias, "CHW", "CHW16c")
scale = relay.layout_transform(scale, "CHW", "CHW16c")
y = relay.nn.conv2d(x, weight, channels=64, kernel_size=(3, 3), padding=(1, 1),
data_layout="NCHW16c")
y = relay.add(y, bias) # test broadcasting to lhs
y = relay.multiply(scale, y) # test broadcasting to rhs
y = relay.layout_transform(y, "NCHW16c", "NCHW")
y = relay.Function(free_vars(y), y)
return y
def after_B():
inputs = [relay.var('input_' + str(i), shape=(10, 10)) for i in range(8)]
add_relu_calls = []
for i in range(4):
x = relay.var('x' + str(i))
y = relay.var('x' + str(i))
add_relu = relay.add(x, y)
add_relu = relay.nn.relu(add_relu)
add_relu = relay.Function([x, y], add_relu)
add_relu = add_relu.set_attribute('Composite', tir.StringImm('add_relu'))
add_relu_call = relay.Call(add_relu, [inputs[i*2], inputs[i*2+1]])
add_relu_calls.append(add_relu_call)
add = relay.add(add_relu_calls[0], add_relu_calls[1])
sub = relay.subtract(add_relu_calls[2], add_relu_calls[3])
out = relay.multiply(add, sub)
return relay.Function(inputs, out)
def test_compile_nhwc_pack():
data = relay.var("data", shape=(1, 1, 1, 1024), dtype="uint8")
weight = relay.var("weight", shape=(1, 1, 1024, 1001), dtype="int8")
p2 = relay.var("p2", shape=(1, 1, 1, 1), dtype="int32")
conv = relay.nn.conv2d(data,
weight,
kernel_size=(1, 1),
data_layout="NHWC",
kernel_layout="HWIO",
out_dtype="int32")
multiply = relay.multiply(relay.const(-22, dtype='int32'), p2)
tile = relay.tile(multiply, reps=(1, 1, 1, 1001))
subtract = relay.subtract(conv, tile)
func = subtract
mod = relay.Function(relay.analysis.free_vars(func), func)
relay.build(mod, target="llvm")
def create_external_func1(mod_, compiler_name, symbol_name):
x_int = relay.var("x_int", shape=(10, 10))
w0_int = relay.var("w0_int", shape=(10, 10))
w1_int = relay.var("w1_int", shape=(10, 10))
w2_int = relay.var("w2_int", shape=(10, 10))
z0 = relay.add(x_int, w0_int)
p0 = relay.subtract(z0, w1_int)
q0 = relay.multiply(z0, w2_int)
f1_o_tuple = relay.Tuple([p0, q0])
f1 = relay.Function([x_int, w0_int, w1_int, w2_int], f1_o_tuple)
f1 = set_func_attr(f1, compiler_name, symbol_name)
glb_f1 = relay.GlobalVar(symbol_name)
mod_[glb_f1] = f1
mod_ = relay.transform.InferType()(mod_)
return glb_f1, mod_
def before(x, conv_weight, out_bias, out_scale, channels):
args = [x, conv_weight, out_bias]
y1 = relay.nn.conv2d(x,
conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y1 = relay.nn.relu(y1)
y2 = relay.nn.conv2d(x,
conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y2 = relay.nn.relu(y2)
y = relay.add(y1, y2)
y = relay.multiply(y, out_scale)
return relay.Function(args, y)
def __impl_philox_2x_round(self, ctr, key):
"""Compute a round in Philox2x32.
:param ctr: uint64 vector
:param key: uint32 scalar
:return:
"""
ctr_0 = relay.right_shift(ctr, RELAY_UINT64_32)
ctr_1 = relay.bitwise_and(ctr, RELAY_UINT64_CLEAR_HIGH)
# mul_hi_lo
product = relay.multiply(RELAY_PHILOX_M2x32_0, ctr_0)
key_64 = relay.cast(key, "uint64")
ctr_1_xor_key = relay.bitwise_xor(ctr_1, key_64)
ctr_1_xor_key_up = relay.left_shift(ctr_1_xor_key, RELAY_UINT64_32)
return relay.bitwise_xor(product, ctr_1_xor_key_up)
def before(x, conv_weight, out_bias, out_scale, channels):
args = [x, conv_weight, out_bias, out_scale]
out_scale = relay.expand_dims(out_scale, axis=1, num_newaxis=2)
out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
y1 = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y1 = relay.nn.relu(y1)
y2 = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y2 = relay.nn.relu(y2)
y = relay.add(y1, y2)
y = relay.multiply(y, out_scale)
return relay.Function(args, y)
def test_checkpoint():
dtype = "float32"
xs = [relay.var("x{}".format(i), dtype) for i in range(4)]
f = relay.multiply(relay.add(xs[0], xs[1]), relay.add(xs[2], xs[3]))
f_checkpoint = relay.annotation.checkpoint(f)
func, func_checkpoint = relay.Function(xs, f), relay.Function(xs, f_checkpoint)
f, f_checkpoint = run_infer_type(func), run_infer_type(func_checkpoint)
assert f.checked_type == f_checkpoint.checked_type
inputs = [np.random.uniform() for _ in range(len(xs))]
for target, dev in tvm.testing.enabled_targets():
for kind in ["graph", "debug"]:
intrp = relay.create_executor(kind, device=dev, target=target)
f_res = intrp.evaluate(f)(*inputs)
f_checkpoint_res = intrp.evaluate(f_checkpoint)(*inputs)
tvm.testing.assert_allclose(f_res.numpy(), f_checkpoint_res.numpy(), 0, 0)
def test_qemu_make_fail(platform, west_cmd, skip_build, tvm_debug):
if platform not in ["host", "mps2_an521"]:
pytest.skip(msg="Only for QEMU targets.")
"""Testing QEMU make fail."""
model, zephyr_board = PLATFORMS[platform]
build_config = {"skip_build": skip_build, "debug": tvm_debug}
shape = (10, )
dtype = "float32"
this_dir = pathlib.Path(__file__).parent
tvm_source_dir = this_dir / ".." / ".." / ".."
runtime_path = tvm_source_dir / "apps" / "microtvm" / "zephyr" / "aot_demo"
# Construct Relay program.
x = relay.var("x", relay.TensorType(shape=shape, dtype=dtype))
xx = relay.multiply(x, x)
z = relay.add(xx, relay.const(np.ones(shape=shape, dtype=dtype)))
func = relay.Function([x], z)
target = tvm.target.target.micro(
model, options=["-link-params=1", "--executor=aot"])
with tvm.transform.PassContext(opt_level=3,
config={"tir.disable_vectorize": True}):
lowered = relay.build(func, target)
# Generate input/output header files
model_files_path = os.path.join(runtime_path, "include")
_create_header_file((f"input_data"), np.zeros(shape=shape, dtype=dtype),
model_files_path)
_create_header_file("output_data", np.zeros(shape=shape, dtype=dtype),
model_files_path)
session_kw = _build_session_kw(model, target, zephyr_board, west_cmd,
lowered.lib, runtime_path, build_config)
file_path = os.path.join(session_kw["binary"].base_dir,
"zephyr/CMakeFiles/run.dir/build.make")
assert os.path.isfile(file_path), f"[{file_path}] does not exist."
# Remove a file to create make failure.
os.remove(file_path)
transport = session_kw["flasher"].flash(session_kw["binary"])
with pytest.raises(RuntimeError) as excinfo:
transport.open()
assert "QEMU setup failed" in str(excinfo.value)
def expected():
a = relay.var('a', shape=(10, 10))
b = relay.var('b', shape=(10, 10))
c = relay.var('c', shape=(10, 10))
# add_sub_mul function
in_1 = relay.var('in_1', shape=(10, 10))
in_2 = relay.var('in_2', shape=(10, 10))
add_node = relay.add(in_1, in_2)
sub_node = relay.subtract(in_1, in_2)
mul_node = relay.multiply(add_node, sub_node)
add_sub_mul = relay.Function([in_1, in_2], mul_node)
# merged function
add_sub_mul_1 = relay.Call(add_sub_mul, [a, b])
add_sub_mul_2 = relay.Call(add_sub_mul, [c, add_sub_mul_1])
r = relay.nn.relu(add_sub_mul_2)
return relay.Function([a, b, c], r)
def before(x, conv_weight, in_bias, in_scale, channels):
args = [x, conv_weight, in_bias]
x = relay.multiply(x, in_scale)
x = relay.nn.relu(x)
x = relay.add(x, in_bias)
x_var = relay.Var("x_var")
y1 = relay.nn.conv2d(
x_var,
conv_weight,
channels=channels,
kernel_size=(3, 3),
data_layout="NHWC",
kernel_layout="HWIO",
padding=(1, 1),
)
z = relay.add(y1, x)
let = relay.Let(x_var, x, z)
return relay.Function(args, let)
def test_multiply_left_constant(self):
right = relay.var("right", relay.TensorType((-1, 4, 2, 2), "float32"))
left = relay.expr.const(np.zeros((2, 2), dtype=np.float32))
net = relay.multiply(left, right)
net = relay.Function([right], net)
mod = tvm.IRModule.from_expr(net)
mod = relay.transform.InferType()(mod)
xgraph = xf_relay.from_relay(mod, {})
layers = xgraph.get_layers()
assert layers[0].type[0] == 'Input'
assert 'relay_id' in layers[0].attrs
assert layers[1].type[0] == 'Scale'
assert layers[1].shapes == [-1, 4, 2, 2]
assert 'relay_id' in layers[1].attrs
def test_mul_param(interface_api, use_unpacked_api, use_calculated_workspaces):
x = relay.var("x", shape=(10, 10))
y = relay.var("y", shape=(1, 10))
func = relay.Function([x, y], relay.multiply(x, y))
x_data = np.random.rand(10, 10).astype("float32")
y_data = np.random.rand(1, 10).astype("float32")
inputs = OrderedDict([("x", x_data), ("y", y_data)])
output_list = generate_ref_data(func, inputs)
compile_and_run(
func,
inputs,
output_list,
interface_api,
use_unpacked_api,
use_calculated_workspaces,
)
def relay_take_grad_inp(c, _nb_indices, _indices, _values):
assert _nb_indices.is_constant(int)
values = c.ref(_values)
r_indices = relay.reshape(c.ref(_indices),
tuple(_indices.abstract.xshape()) + (1, ))
n_rows = _nb_indices.value
n_cols = _values.abstract.xshape()[-1]
outputs = []
indices_dtype = type_to_np_dtype(_indices.abstract.element.xtype())
out_dtype = type_to_np_dtype(_values.abstract.element.xtype())
for i in range(n_rows):
select_entries = relay.equal(r_indices, relay.const(i, indices_dtype))
casted_select = relay.cast(select_entries, out_dtype)
select_dout = relay.multiply(casted_select, values)
reshape_out = relay.reshape(select_dout, (-1, n_cols))
vector = relay.sum(reshape_out, 0)
outputs.append(relay.reshape(vector, (1, n_cols)))
return relay.concatenate(outputs, 0)
def before():
data = relay.var('data', shape=(1, 512, 28, 28))
kernel = relay.var('kernel', shape=(256, 512, 1, 1))
bias = relay.var('bias', shape=(256,))
a = relay.var('a', shape=(1, 256, 28, 28))
b = relay.var('b', shape=(1, 256, 28, 28))
conv_node = relay.nn.conv2d(data,
kernel,
kernel_size=(1, 1),
padding=(0, 0),
strides=(1, 1))
bias_node = relay.nn.bias_add(conv_node, bias)
relu_node = relay.nn.relu(bias_node)
add_node = relay.add(relu_node, a)
relu_node_2 = relay.nn.relu(add_node)
r = relay.multiply(relu_node_2, b)
return relay.Function([data, kernel, bias, a, b], r)
def fail1(x, conv_weight, out_bias, out_scale, channels):
args = [x, conv_weight, out_bias]
out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
y1 = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y1 = relay.nn.relu(y1)
y2 = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1),
out_layout="CNHW")
# fold will fail because the axis from two path
# differs from each other.
y2 = relay.nn.relu(y2)
y = relay.add(y1, y2)
y = relay.multiply(y, out_scale)
return relay.Function(args, y)
def fail1(x, conv_weight, out_bias, out_scale, channels):
args = [x, conv_weight, out_bias]
out_bias = relay.expand_dims(out_bias, axis=1, num_newaxis=2)
y1 = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1))
y1 = relay.nn.relu(y1)
y2 = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
padding=(1, 1),
out_layout="CNHW")
# fold will fail because the axis from two path
# differs from each other.
y2 = relay.nn.relu(y2)
y = relay.add(y1, y2)
y = relay.multiply(y, out_scale)
return relay.Function(args, y)
def expected():
data = relay.var('data', shape=(1, 512, 28, 28))
kernel = relay.var('kernel', shape=(256, 512, 1, 1))
bias = relay.var('bias', shape=(256, ))
a = relay.var('a', shape=(1, 256, 28, 28))
b = relay.var('b', shape=(1, 256, 28, 28))
# conv_bias_relu function
in_1 = relay.var('in_1', shape=(1, 512, 28, 28))
in_2 = relay.var('in_2', shape=(256, 512, 1, 1))
in_3 = relay.var('in_3', shape=(256, ))
conv_node = relay.nn.conv2d(in_1,
in_2,
kernel_size=(1, 1),
padding=(0, 0),
strides=(1, 1))
bias_node = relay.nn.bias_add(conv_node, in_3)
r = relay.nn.relu(bias_node)
conv_bias_add_relu = relay.Function([in_1, in_2, in_3], r)
conv_bias_add_relu = conv_bias_add_relu.set_attribute(
"Primitive", tir.IntImm("int32", 1))
conv_bias_add_relu = conv_bias_add_relu.set_attribute(
"Composite", tir.StringImm("conv2d_bias_relu"))
# add_relu function
in_4 = relay.var('in_4', shape=(1, 256, 28, 28))
in_5 = relay.var('in_5', shape=(1, 256, 28, 28))
add_node = relay.add(in_4, in_5)
r = relay.nn.relu(add_node)
add_relu = relay.Function([in_4, in_5], r)
add_relu = add_relu.set_attribute("Primitive", tir.IntImm("int32", 1))
add_relu = add_relu.set_attribute("Composite",
tir.StringImm("add_relu"))
# merged function
conv_bias_add_relu_1 = relay.Call(conv_bias_add_relu,
[data, kernel, bias])
add_relu_1 = relay.Call(add_relu, [conv_bias_add_relu_1, a])
r = relay.multiply(add_relu_1, b)
return relay.Function([data, kernel, bias, a, b], r)
def graph2relay(graph: Graph, batch_size):
node2var = {}
params = {}
for node in graph.nodes():
if node is graph.enter_node:
node2var[node] = relay.var(name_hint=node.name, shape=(batch_size, *node.output_shape))
else:
term_vars = []
for term in node.inputs:
value_vars = []
for value in term:
if value.begin == 0 and value.end == value.node.output_shape[0]:
var = node2var[value.node]
else:
var = relay.strided_slice(node2var[value.node], begin=[0, value.begin, 0, 0], end=[batch_size, value.end, *value.node.output_shape[1:]])
value_vars.append(var)
term_var = value_vars[0]
for value_var in value_vars[1:]:
if isinstance(node, Element):
if node.op_type == 'mul':
term_var = relay.multiply(term_var, value_var)
elif node.op_type == 'add':
term_var = term_var + value_var
else:
raise ValueError
else:
term_var = term_var + value_var
term_vars.append(term_var)
if len(term_vars) > 1:
x = relay.concatenate(term_vars, axis=1)
else:
x = term_vars[0]
node2var[node] = do_layer(x, node, params)
x = node2var[graph.exit_node]
fn = relay.Function(relay.analysis.free_vars(x), x)
if tvm_minor_version() <= 6:
return relay.Module.from_expr(fn), params
else:
return tvm.ir.IRModule.from_expr(fn), params
def expected(x, w1, w2, b1, b2, scale1, scale2, newshape):
args = [x, w1, w2, b1, b2, scale1, scale2]
x_stacked = relay.stack((x, x), axis=0)
w = relay.stack((w1, w2), axis=0)
y = relay.nn.batch_matmul(x_stacked, w)
b1 = relay.expand_dims(b1, 0)
b2 = relay.expand_dims(b2, 0)
b = relay.stack((b1, b2), axis=0)
y = relay.add(y, b)
scale1 = relay.expand_dims(scale1, 0)
scale2 = relay.expand_dims(scale2, 0)
scale = relay.stack((scale1, scale2), axis=0)
y = relay.multiply(y, scale)
(y1, y2) = relay.split(y, 2)
y1 = relay.squeeze(y1, [0])
y2 = relay.squeeze(y2, [0])
y1 = relay.reshape(y1, newshape=newshape)
y2 = relay.reshape(y2, newshape=newshape)
y = relay.Tuple((y1, y2))
return relay.Function(args, y)
def test_lower_to_tir():
data = relay.var("data", shape=(1, 1, 1, 1024), dtype="uint8")
weight = relay.var("weight", shape=(1, 1, 1024, 1001), dtype="int8")
p2 = relay.var("p2", shape=(1, 1, 1, 1), dtype="int32")
conv = relay.nn.conv2d(
data,
weight,
kernel_size=(1, 1),
data_layout="NHWC",
kernel_layout="HWIO",
out_dtype="int32",
)
multiply = relay.multiply(relay.const(-22, dtype="int32"), p2)
tile = relay.tile(multiply, reps=(1, 1, 1, 1001))
subtract = relay.subtract(conv, tile)
func = subtract
expr = relay.Function(relay.analysis.free_vars(func), func)
mod = tvm.IRModule.from_expr(expr)
mod = relay.transform.InferType()(mod)
lower_to_tir(mod["main"])
def expected():
data = relay.var("data", shape=(1, 512, 28, 28))
kernel = relay.var("kernel", shape=(256, 512, 1, 1))
bias = relay.var("bias", shape=(256, ))
a = relay.var("a", shape=(1, 256, 28, 28))
b = relay.var("b", shape=(1, 256, 28, 28))
# conv_bias_relu function
in_1 = relay.var("in_1", shape=(1, 512, 28, 28))
in_2 = relay.var("in_2", shape=(256, 512, 1, 1))
in_3 = relay.var("in_3", shape=(256, ))
conv_node = relay.nn.conv2d(in_1,
in_2,
kernel_size=(1, 1),
padding=(0, 0),
strides=(1, 1))
bias_node = relay.nn.bias_add(conv_node, in_3)
r = relay.nn.relu(bias_node)
conv_bias_add_relu = relay.Function([in_1, in_2, in_3], r)
conv_bias_add_relu = conv_bias_add_relu.with_attr(
"Composite", "conv2d_bias_relu")
conv_bias_add_relu = conv_bias_add_relu.with_attr(
"PartitionedFromPattern", "nn.conv2d_nn.bias_add_nn.relu_")
# add_relu function
in_4 = relay.var("in_4", shape=(1, 256, 28, 28))
in_5 = relay.var("in_5", shape=(1, 256, 28, 28))
add_node = relay.add(in_4, in_5)
r = relay.nn.relu(add_node)
add_relu = relay.Function([in_4, in_5], r)
add_relu = add_relu.with_attr("Composite", "add_relu")
add_relu = add_relu.with_attr("PartitionedFromPattern", "add_nn.relu_")
# merged function
conv_bias_add_relu_1 = relay.Call(conv_bias_add_relu,
[data, kernel, bias])
add_relu_1 = relay.Call(add_relu, [conv_bias_add_relu_1, a])
r = relay.multiply(add_relu_1, b)
return relay.Function([data, kernel, bias, a, b], r)
def before(x, conv_weight, in_bias, in_scale, channels):
args = [x, conv_weight, in_bias]
x = relay.multiply(in_scale, x)
x = relay.nn.relu(x)
x = relay.subtract(x, in_bias)
y1 = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
data_layout="NHWC",
kernel_layout="HWIO",
groups=channels,
padding=(1, 1))
y2 = relay.nn.conv2d(x, conv_weight,
channels=channels,
kernel_size=(3, 3),
data_layout="NHWC",
kernel_layout="HWIO",
groups=channels,
padding=(1, 1))
z = relay.add(y1, y2)
return relay.Function(args, z)
def alter_conv2d(attrs, inputs, tinfos):
data, weight = inputs
weight = relay.multiply(weight, relay.const(2.0, "float32"))
return relay.nn.conv2d(data, weight, **attrs)
def fold_conv_weight():
squeezed_scale = relay.squeeze(out_scale, axis=[1,2])
return relay.multiply(
conv_weight ,
relay.expand_dims(squeezed_scale, axis=1, num_newaxis=3))
def lstm_cell(num_hidden, batch_size=1, dtype="float32", name=""):
"""Long-Short Term Memory (LSTM) network cell.
Parameters
----------
num_hidden : int
Number of units in output symbol.
batch_size : int
Batch size (length of states).
Returns
-------
result : tvm.relay.Function
A Relay function that evaluates an LSTM cell.
The function takes in a tensor of input data, a tuple of two
states, and weights and biases for dense operations on the
inputs and on the state. It returns a tuple with two members,
an output tensor and a tuple of two new states.
"""
builder = relay.ScopeBuilder()
input_type = relay.TensorType((batch_size, num_hidden), dtype)
weight_type = relay.TensorType((4*num_hidden, num_hidden), dtype)
bias_type = relay.TensorType((4*num_hidden,), dtype)
dense_type = relay.TensorType((batch_size, 4*num_hidden), dtype)
slice_type = relay.TupleType([input_type, input_type,
input_type, input_type])
ret_type = relay.TupleType([input_type,
relay.TupleType([input_type, input_type])])
inputs = relay.Var("inputs", input_type)
states = relay.Var("states",
relay.TupleType([input_type, input_type]))
i2h_weight = relay.Var("i2h_weight", weight_type)
i2h_bias = relay.Var("i2h_bias", bias_type)
h2h_weight = relay.Var("h2h_weight", weight_type)
h2h_bias = relay.Var("h2h_bias", bias_type)
i2h = builder.let(("i2h", dense_type),
layers.dense_add_bias(
data=inputs,
units=num_hidden * 4,
weight=i2h_weight, bias=i2h_bias,
name="%si2h" % name))
h2h = builder.let(("h2h", dense_type),
layers.dense_add_bias(
data=relay.TupleGetItem(states, 0),
units=num_hidden * 4,
weight=h2h_weight, bias=h2h_bias,
name="%sh2h" % name))
gates = builder.let(("gates", dense_type), relay.add(i2h, h2h))
slice_gates = builder.let(("slice_gates", slice_type),
relay.split(gates,
indices_or_sections=4,
axis=1).astuple())
in_gate = builder.let(("in_gate", input_type),
relay.sigmoid(relay.TupleGetItem(slice_gates, 0)))
forget_gate = builder.let(("forget_gate", input_type),
relay.sigmoid(relay.TupleGetItem(slice_gates, 1)))
in_transform = builder.let(("in_transform", input_type),
relay.tanh(relay.TupleGetItem(slice_gates, 2)))
out_gate = builder.let(("out_gate", input_type),
relay.sigmoid(relay.TupleGetItem(slice_gates, 3)))
next_c = builder.let(("next_c", input_type),
relay.add(relay.multiply(forget_gate,
relay.TupleGetItem(states, 1)),
relay.multiply(in_gate, in_transform)))
next_h = builder.let(("next_h", input_type),
relay.multiply(out_gate, relay.tanh(next_c)))
ret = builder.let(("ret", ret_type),
relay.Tuple([next_h, relay.Tuple([next_h, next_c])]))
builder.ret(ret)
body = builder.get()
return relay.Function([inputs, states, i2h_weight,
i2h_bias, h2h_weight, h2h_bias],
body, ret_type)
def fold_conv_weight():
return relay.multiply(
conv_weight ,
relay.expand_dims(squeezed_scale, axis=1, num_newaxis=3))
def test_call():
# select right function to call: simple ident case
id_func = relay.Var("id")
assert parses_as(
"""
let %id = fn (%x) { %x };
10 * %id(10)
""",
relay.Let(
id_func,
relay.Function([X], X, None, []),
relay.multiply(relay.const(10), relay.Call(id_func, [relay.const(10)]))
)
)
# 0 args
constant = relay.Var("constant")
assert parses_as(
"""
let %constant = fn () { 0 };
%constant()
""",
relay.Let(
constant,
relay.Function([], relay.const(0), None, []),
relay.Call(constant, [], None, None)
)
)
# 1 arg
id_var = relay.Var("id")
assert parses_as(
"""
let %id = fn (%x) { %x };
%id(1)
""",
relay.Let(
id_var,
relay.Function([X], X, None, []),
relay.Call(id_var, [relay.const(1)], None, None)
)
)
# 2 args
multiply = relay.Var("multiply")
assert parses_as(
"""
let %multiply = fn (%x, %y) { %x * %y };
%multiply(0, 0)
""",
relay.Let(
multiply,
relay.Function(
[X, Y],
relay.multiply(X, Y),
None,
[]
),
relay.Call(multiply, [relay.const(0), relay.const(0)], None, None)
)
)
# anonymous function
assert parses_as(
"""
(fn (%x) { %x })(0)
""",
relay.Call(
relay.Function(
[X],
X,
None,
[]
),
[relay.const(0)],
None,
None
)
)
# TODO(@jmp): re-enable after sequence parsing improvements
# curried function
# curried_mult = relay.Var("curried_mult")
# assert parses_as(
# """
# let %curried_mult =
# fn (%x) {
# fn (%y) {
# %x * %y
# }
# };
# %curried_mult(0);
# %curried_mult(0)(0)
# """,
# relay.Let(
# curried_mult,
# relay.Function(
# [X],
# relay.Function(
# [Y],
# relay.multiply(X, Y),
# None,
# []
# ),
# None,
# []
# ),
# relay.Let(
# _,
# relay.Call(curried_mult, [relay.const(0)], None, None),
# relay.Call(relay.Call(curried_mult, [relay.const(0)], None, None), [relay.const(0)], None, None)
# )
# )
# )
# op
assert parses_as(
"abs(1)",
relay.Call(relay.op.get("abs"), [relay.const(1)], None, None)
)
