def test_alter_layout_strided_slice():
"""Test rewriting strided_slice during alter_iop_layout"""
def before():
x = relay.var("x", shape=(1, 32, 28, 28))
weight = relay.var('weight', shape=(32, 32, 3, 3))
y = relay.nn.conv2d(x, weight, channels=32, kernel_size=(3, 3), padding=(1, 1))
y = relay.strided_slice(y, begin=[0, 16], end=[None, None])
y = relay.Function(analysis.free_vars(y), y)
return y
def alter_conv2d(attrs, inputs, tinfos):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs['data_layout'] = 'NCHW4c'
return relay.nn.conv2d(data, weight, **new_attrs)
def expected():
x = relay.var("x", shape=(1, 32, 28, 28))
weight = relay.var("weight")
x = relay.layout_transform(x, "NCHW", "NCHW4c")
y = relay.nn.conv2d(x, weight, channels=32, kernel_size=(3, 3), padding=(1, 1),
data_layout="NCHW4c")
y = relay.strided_slice(y, begin=[0, 4], end=[None, 8])
y = relay.layout_transform(y, "NCHW4c", "NCHW")
y = relay.Function(analysis.free_vars(y), y)
return y
with TempOpAttr("nn.conv2d", "FTVMAlterOpLayout", alter_conv2d):
a = before()
a = run_opt_pass(a, [transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = run_opt_pass(expected(), transform.InferType())
assert analysis.alpha_equal(a, b), "Actual = \n" + str(a)
def test_alter_layout_depthwise_conv2d():
"""Test depthwise_conv2d operator"""
def before():
x = relay.var("x", shape=(1, 32, 56, 56))
w = relay.var("w", shape=(32, 1, 3, 3))
y = relay.nn.conv2d(x, w, padding=(1, 1), channels=32, kernel_size=(3, 3), groups=32)
y = relay.Function(analysis.free_vars(y), y)
return y
import topi
@register_alter_op_layout("nn.conv2d", level=110)
def alter_conv2d(attrs, inputs, tinfos):
with tvm.target.create("llvm"):
return topi.nn.conv2d_alter_layout(attrs, inputs, tinfos, relay)
def expected():
x = relay.var("x", shape=(1, 32, 56, 56))
w = relay.var("w", shape=(32, 1, 3, 3))
x = relay.layout_transform(x, "NCHW", "NCHW8c")
w = relay.layout_transform(w, "OIHW", "OIHW1i8o")
y = relay.nn.contrib_depthwise_conv2d_nchwc(x, w, padding=(1, 1), channels=32, kernel_size=(3, 3),
groups=32, data_layout="NCHW8c", kernel_layout="OIHW1i8o",
out_layout="NCHW8c")
y = relay.layout_transform(y, "NCHW8c", "NCHW")
y = relay.Function(analysis.free_vars(y), y)
return y
a = before()
a = run_opt_pass(a, [transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = expected()
b = run_opt_pass(b, transform.InferType())
assert(analysis.alpha_equal(a, b))
def partition_for_dnnl(mod, params=None):
"""Partition the graph greedily offloading supported operators to DNNL.
Parameters
----------
mod : Module
The module to run passes on.
params : Optional[Dict[str, NDArray]]
Constant input parameters.
Returns
-------
mod : Module
Annotated and partitioned module.
"""
if params:
mod["main"] = bind_params_by_name(mod["main"], params)
seq = tvm.transform.Sequential([
transform.CanonicalizeOps(),
transform.InferType(),
transform.SimplifyInference(),
transform.FoldConstant(),
transform.FoldScaleAxis(),
# fold consecutive add ops to simplify pattern `conv2d-bias_add-bn-relu`
transform.SimplifyExpr(),
transform.FoldConstant(),
transform.MergeComposite(pattern_table()),
transform.AnnotateTarget("dnnl"),
transform.MergeCompilerRegions(),
transform.PartitionGraph(),
])
with tvm.transform.PassContext(opt_level=3):
mod = seq(mod)
return mod
def test_alter_layout():
"""Test alternating the layout of a conv2d.
The layout of broadcast operators and the weight should be changed accordingly.
"""
def before():
x = relay.var("x", shape=(1, 64, 56, 56))
bias = relay.var("bias")
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)
# a useless tuple, which will be eliminated
y = relay.Tuple([y])[0]
y = relay.nn.relu(y)
y = relay.nn.max_pool2d(y, pool_size=(2, 2))
y = relay.cast(y, 'int32')
y = relay.nn.batch_flatten(y)
y = relay.Function(analysis.free_vars(y), y)
return y
@register_alter_op_layout("nn.conv2d", level=102)
def alter_conv2d(attrs, inputs, tinfos):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs['data_layout'] = 'NCHW16c'
new_attrs['kernel_layout'] = 'OIHW16i'
return relay.nn.conv2d(data, weight, **new_attrs)
def expected():
x = relay.var("x", shape=(1, 64, 56, 56))
bias = relay.var("bias", shape=(64,))
weight = relay.var("weight", shape=(64, 64, 3, 3))
y = relay.layout_transform(x, "NCHW", "NCHW16c")
w = relay.layout_transform(weight, "OIHW", "OIHW16i")
y = relay.nn.conv2d(y, w,
channels=64,
kernel_size=(3, 3),
padding=(1, 1),
kernel_layout="OIHW16i",
data_layout="NCHW16c")
b = relay.expand_dims(bias, axis=1, num_newaxis=2)
b = relay.layout_transform(b, "CHW", "CHW16c")
y = relay.add(y, b)
y = relay.nn.relu(y)
y = relay.nn.max_pool2d(y, pool_size=(2, 2), layout="NCHW16c")
y = relay.cast(y, 'int32')
y = relay.layout_transform(y, "NCHW16c", "NCHW")
y = relay.nn.batch_flatten(y)
y = relay.Function(analysis.free_vars(y), y)
return y
a = before()
a = run_opt_pass(a, [transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = expected()
b = run_opt_pass(b, transform.InferType())
assert analysis.alpha_equal(a, b), "Actual = \n" + str(a)
def test_alter_layout_depthwise_conv2d():
"""Test depthwise_conv2d operator"""
def before():
x = relay.var("x", shape=(1, 32, 56, 56))
w = relay.var("w", shape=(32, 1, 3, 3))
y = relay.nn.conv2d(x, w, padding=(1, 1), channels=32, kernel_size=(3, 3), groups=32)
y = relay.Function(analysis.free_vars(y), y)
return y
from tvm import topi
def alter_conv2d(attrs, inputs, tinfos, out_type):
with tvm.target.Target("llvm"):
return topi.nn.conv2d_alter_layout(attrs, inputs, tinfos, out_type)
def expected():
x = relay.var("x", shape=(1, 32, 56, 56))
w = relay.var("w", shape=(32, 1, 3, 3))
x = relay.layout_transform(x, "NCHW", "NCHW8c")
w = relay.layout_transform(w, "OIHW", "OIHW1i8o")
y = relay.nn.contrib_depthwise_conv2d_nchwc(x, w, padding=(1, 1, 1, 1), channels=32, kernel_size=(3, 3),
groups=32, data_layout="NCHW8c", kernel_layout="OIHW1i8o",
out_layout="NCHW8c")
y = relay.layout_transform(y, "NCHW8c", "NCHW")
y = relay.Function(analysis.free_vars(y), y)
return y
with TempOpAttr("nn.conv2d", "FTVMAlterOpLayout", alter_conv2d):
a = before()
a = run_opt_pass(a, [transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = run_opt_pass(expected(), transform.InferType())
assert(tvm.ir.structural_equal(a, b))
def test_alter_layout_broadcast_op():
"""Test boradcast operators """
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(analysis.free_vars(y), y)
return y
# Register alter op layout. "level" is used to override the previously registered functions.
@register_alter_op_layout("nn.conv2d", level=105)
def alter_conv2d(attrs, inputs, tinfos):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs['data_layout'] = 'NCHW16c'
return relay.nn.conv2d(data, weight, **new_attrs)
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.expand_dims(bias, 0, 1)
bias = relay.layout_transform(bias, "NCHW", "NCHW16c")
scale = relay.expand_dims(scale, 0, 1)
scale = relay.layout_transform(scale, "NCHW", "NCHW16c")
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(analysis.free_vars(y), y)
return y
a = before()
a = run_opt_pass(a,
[transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = expected()
b = run_opt_pass(b, transform.InferType())
assert analysis.alpha_equal(a, b), "Actual = \n" + str(a)
def test_alter_layout_broadcast_scalar_op():
"""Test alternating the layout of a conv2d.
The layout of broadcast operators and the weight should be changed accordingly.
"""
def before():
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")
y = relay.nn.conv2d(x, kernel, data_layout="NHWC", kernel_layout="HWIO", kernel_size=(3, 3))
y = relay.add(bias, y)
y = relay.nn.relu(y)
y = relay.multiply(multiplier1, y)
y = relay.multiply(y, multiplier2)
y = relay.Function(analysis.free_vars(y), y)
return y
def alter_conv2d(attrs, inputs, tinfos, out_type):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs["data_layout"] = "NCHW16c"
return relay.nn.conv2d(data, weight, **new_attrs)
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
with TempOpAttr("nn.conv2d", "FTVMAlterOpLayout", alter_conv2d):
a = before()
a = run_opt_pass(a, [transform.CanonicalizeOps(), transform.AlterOpLayout()])
b = run_opt_pass(expected(), transform.InferType())
assert tvm.ir.structural_equal(a, b), "Actual = \n" + str(a)
def test_alter_layout_broadcast_op():
"""Test boradcast operators """
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(analysis.free_vars(y), y)
return y
def alter_conv2d(attrs, inputs, tinfos, out_type):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs['data_layout'] = 'NCHW16c'
return relay.nn.conv2d(data, weight, **new_attrs)
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.expand_dims(bias, 0, 1)
bias = relay.layout_transform(bias, "NCHW", "NCHW16c")
scale = relay.expand_dims(scale, 0, 1)
scale = relay.layout_transform(scale, "NCHW", "NCHW16c")
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(analysis.free_vars(y), y)
return y
with TempOpAttr("nn.conv2d", "FTVMAlterOpLayout", alter_conv2d):
a = before()
a = run_opt_pass(
a, [transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = run_opt_pass(expected(), transform.InferType())
assert tvm.ir.structural_equal(a, b), "Actual = \n" + str(a)
def test_alter_layout_lrn():
"""Test alternating the layout of a conv2d.
The layout of broadcast operators and the weight should be changed accordingly.
"""
def before():
x = relay.var("x", shape=(1, 64, 56, 56))
bias = relay.var("bias")
weight = relay.var("weight")
y = relay.nn.conv2d(x,
weight,
channels=64,
kernel_size=(3, 3),
padding=(1, 1))
y = relay.nn.max_pool2d(y, pool_size=(2, 2))
y = relay.nn.lrn(y)
y = relay.Function(analysis.free_vars(y), y)
return y
def alter_conv2d(attrs, inputs, tinfos, out_type):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs['data_layout'] = 'NCHW16c'
new_attrs['kernel_layout'] = 'OIHW16i'
return relay.nn.conv2d(data, weight, **new_attrs)
def expected():
x = relay.var("x", shape=(1, 64, 56, 56))
bias = relay.var("bias", shape=(64, ))
weight = relay.var("weight", shape=(64, 64, 3, 3))
y = relay.layout_transform(x, "NCHW", "NCHW16c")
w = relay.layout_transform(weight, "OIHW", "OIHW16i")
y = relay.nn.conv2d(y,
w,
channels=64,
kernel_size=(3, 3),
padding=(1, 1),
kernel_layout="OIHW16i",
data_layout="NCHW16c")
y = relay.nn.max_pool2d(y, pool_size=(2, 2), layout="NCHW16c")
y = relay.layout_transform(y, "NCHW16c", "NCHW")
y = relay.nn.lrn(y)
y = relay.Function(analysis.free_vars(y), y)
return y
with TempOpAttr("nn.conv2d", "FTVMAlterOpLayout", alter_conv2d):
a = before()
a = run_opt_pass(
a, [transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = run_opt_pass(expected(), transform.InferType())
assert tvm.ir.structural_equal(a, b), "Actual = \n" + str(a)
def test_alter_layout_nchw_upsamping_op():
"""Test upsamping operators """
def before():
x = relay.var("x", shape=(1, 32, 28, 28))
weight = relay.var('weight', shape=(32, 32, 3, 3))
y = relay.nn.conv2d(x,
weight,
channels=32,
kernel_size=(3, 3),
padding=(1, 1))
y = relay.nn.upsampling(y, scale=2)
y = relay.nn.avg_pool2d(y, pool_size=(2, 2), strides=(2, 2))
y = relay.Function(analysis.free_vars(y), y)
return y
# Register alter op layout. "level" is used to override the previously registered functions.
@register_alter_op_layout("nn.conv2d", level=108)
def alter_conv2d(attrs, inputs, tinfos):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs['data_layout'] = 'NCHW16c'
return relay.nn.conv2d(data, weight, **new_attrs)
def expected():
x = relay.var("x", shape=(1, 32, 28, 28))
weight = relay.var("weight")
x = relay.layout_transform(x, "NCHW", "NCHW16c")
y = relay.nn.conv2d(x,
weight,
channels=32,
kernel_size=(3, 3),
padding=(1, 1),
data_layout="NCHW16c")
y = relay.nn.upsampling(y, scale=2, layout="NCHW16c")
y = relay.nn.avg_pool2d(y,
pool_size=(2, 2),
strides=(2, 2),
layout='NCHW16c')
y = relay.layout_transform(y, "NCHW16c", "NCHW")
y = relay.Function(analysis.free_vars(y), y)
return y
a = before()
a = run_opt_pass(a,
[transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = expected()
b = run_opt_pass(b, transform.InferType())
assert analysis.alpha_equal(a, b), "Actual = \n" + str(a)
def quantize_model(model, params, input_dtype, input_shape, qeval='power2'):
skip_conv_layers = [0]
with relay.quantize.qconfig(store_lowbit_output=False,
skip_conv_layers=skip_conv_layers):
from tvm.relay.quantize.quantize import _bind_params
graph = _bind_params(model['main'], params)
mod = relay.Module.from_expr(graph)
optimize = _transform.Sequential([
_transform.SimplifyInference(),
_transform.FoldConstant(),
_transform.FoldScaleAxis(),
_transform.CanonicalizeOps(),
_transform.FoldConstant()
])
with relay.build_config(opt_level=4):
mod = optimize(mod)
mod = relay.quantize.annotate()(mod)
# find scale
cache_file = '%s_%s_scales.pkl' % (VIDEO_FILE, MODEL_NAME)
if os.path.exists(cache_file):
print("Using cached layer statistics...")
with open(cache_file, 'rb') as f:
scales = pickle.load(f)
else:
print("Compute layer statistics...")
scales = calibrate_on_dataset(mod['main'], params, input_dtype,
input_shape)
with open(cache_file, 'wb') as f:
pickle.dump(scales, f)
if qeval == 'power2':
scales = list(
map(
lambda scale: 2**np.math.ceil(np.math.log(scale, 2))
if scale > 0 else 1.0, scales))
weight_scales = 'power2'
elif qeval == 'max':
weight_scales = 'max'
else:
raise ValueError("Invalid quantiziation eval: " + qeval)
mod['main'] = relay.quantize.calibrate(mod['main'],
weight_scales=weight_scales,
scales=scales)
mod = relay.quantize.realize()(mod)
mod = relay.transform.FoldConstant()(mod)
return mod
def test_alter_layout_scalar():
"""Test alternating the layout of a conv2d.
The layout of broadcast operators and the weight should be changed accordingly.
"""
def before():
x = relay.var("x", shape=(1, 64, 56, 56))
weight = relay.var("weight")
y = relay.nn.conv2d(x,
weight,
channels=64,
kernel_size=(3, 3),
padding=(1, 1))
y = relay.add(y, relay.const(1, "float32"))
y = relay.Function(analysis.free_vars(y), y)
return y
# Register alter op layout. "level" is used to override the previously registered functions.
@register_alter_op_layout("nn.conv2d", level=106)
def alter_conv2d(attrs, inputs, tinfos):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs['data_layout'] = 'NCHW16c'
return relay.nn.conv2d(data, weight, **new_attrs)
def expected():
x = relay.var("x", shape=(1, 64, 56, 56))
w = relay.var("weight")
y = relay.layout_transform(x, "NCHW", "NCHW16c")
y = relay.nn.conv2d(y,
w,
channels=64,
kernel_size=(3, 3),
padding=(1, 1),
data_layout="NCHW16c")
y = relay.add(y, relay.const(1.0, "float32"))
y = relay.layout_transform(y, "NCHW16c", "NCHW")
y = relay.Function(analysis.free_vars(y), y)
return y
a = before()
a = run_opt_pass(a,
[transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = expected()
b = run_opt_pass(b, transform.InferType())
assert analysis.alpha_equal(a, b), "Actual = \n" + str(a)
def test_alter_layout_prelu():
"""Test PRelu operator"""
def before():
x = relay.var("x", shape=(1, 64, 56, 56))
weight = relay.var("weight")
alpha = relay.var("alpha", relay.IncompleteType())
y = relay.nn.conv2d(x,
weight,
channels=64,
kernel_size=(3, 3),
padding=(1, 1))
y = relay.nn.prelu(y, alpha)
y = relay.Function(analysis.free_vars(y), y)
return y
# Register alter op layout. "level" is used to override the previously registered functions.
@register_alter_op_layout("nn.conv2d", level=111)
def alter_conv2d(attrs, inputs, tinfos):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs['data_layout'] = 'NCHW16c'
return relay.nn.conv2d(data, weight, **new_attrs)
def expected():
x = relay.var("x", shape=(1, 64, 56, 56))
w = relay.var("weight")
alpha = relay.var("alpha", relay.IncompleteType())
y = relay.layout_transform(x, "NCHW", "NCHW16c")
y = relay.nn.conv2d(y,
w,
channels=64,
kernel_size=(3, 3),
padding=(1, 1),
data_layout="NCHW16c")
y = relay.layout_transform(y, "NCHW16c", "NCHW")
y = relay.nn.prelu(y, alpha)
y = relay.Function(analysis.free_vars(y), y)
return y
a = before()
a = run_opt_pass(a,
[transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = expected()
b = run_opt_pass(b, transform.InferType())
assert (analysis.alpha_equal(a, b))
def test_alter_layout_prelu():
"""Test PRelu operator"""
def before():
x = relay.var("x", shape=(1, 64, 56, 56))
weight = relay.var("weight")
alpha = relay.var("alpha", relay.IncompleteType())
y = relay.nn.conv2d(x,
weight,
channels=64,
kernel_size=(3, 3),
padding=(1, 1))
y = relay.nn.prelu(y, alpha)
y = relay.Function(analysis.free_vars(y), y)
return y
def alter_conv2d(attrs, inputs, tinfos, out_type):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs['data_layout'] = 'NCHW16c'
return relay.nn.conv2d(data, weight, **new_attrs)
def expected():
x = relay.var("x", shape=(1, 64, 56, 56))
w = relay.var("weight")
alpha = relay.var("alpha", relay.IncompleteType())
y = relay.layout_transform(x, "NCHW", "NCHW16c")
y = relay.nn.conv2d(y,
w,
channels=64,
kernel_size=(3, 3),
padding=(1, 1),
data_layout="NCHW16c")
y = relay.layout_transform(y, "NCHW16c", "NCHW")
y = relay.nn.prelu(y, alpha)
y = relay.Function(analysis.free_vars(y), y)
return y
with TempOpAttr("nn.conv2d", "FTVMAlterOpLayout", alter_conv2d):
a = before()
a = run_opt_pass(
a, [transform.CanonicalizeOps(),
transform.AlterOpLayout()])
b = run_opt_pass(expected(), transform.InferType())
assert (analysis.alpha_equal(a, b))
def match_pass_name(name):
if name == 'FoldScaleAxis':
return transform.FoldScaleAxis()
if name == 'BackwardFoldScaleAxis':
return transform.BackwardFoldScaleAxis()
if name == 'ForwardFoldScaleAxis':
return transform.ForwardFoldScaleAxis()
if name == 'FuseOps':
return transform.FuseOps(3)
if name == 'FoldConstant':
return transform.FoldConstant()
if name == 'CombineParallelConv2d':
return transform.CombineParallelConv2D()
if name == 'AlterOpLayout':
return transform.AlterOpLayout()
if name == 'EliminateCommonSubexpr':
return transform.EliminateCommonSubexpr()
if name == 'PartialEvaluate':
return transform.PartialEvaluate()
if name == 'CanonicalizeCast':
return transform.CanonicalizeCast()
if name == 'CanonicalizeOps':
return transform.CanonicalizeOps()
raise Exception('Name {} does not match any pass'.format(name))
def test_alter_layout_scalar_regression():
"""regression test where scalar fails"""
def before():
x = relay.var("x", shape=(1, 56, 56, 64))
weight = relay.var("weight", shape=(3, 3, 64, 16))
bias = relay.var("bias", shape=(1, 1, 1, 16))
y = relay.nn.conv2d(
x,
weight,
channels=16,
kernel_size=(3, 3),
padding=(1, 1),
data_layout="NHWC",
kernel_layout="HWIO",
)
y = relay.add(y, bias)
mean = relay.mean(y, axis=3, exclude=True)
var = relay.variance(y, axis=3, exclude=True)
gamma = relay.var("gamma")
beta = relay.var("beta")
y = relay.nn.batch_norm(y, gamma, beta, mean, var, axis=3)
y = y[0]
y = relay.Function(analysis.free_vars(y), y)
return y
def alter_conv2d(attrs, inputs, tinfos, out_type):
data, weight = inputs
new_attrs = dict(attrs)
new_attrs["data_layout"] = "NCHW16c"
return relay.nn.conv2d(data, weight, **new_attrs)
def expected():
x = relay.var("x", shape=(1, 56, 56, 64))
weight = relay.var("weight", shape=(3, 3, 64, 16))
bias = relay.var("bias", shape=(1, 1, 1, 16))
x = relay.layout_transform(x, src_layout="NHWC", dst_layout="NCHW")
x = relay.layout_transform(x, src_layout="NCHW", dst_layout="NCHW16c")
weight = relay.layout_transform(weight, src_layout="HWIO", dst_layout="OIHW")
y = relay.nn.conv2d(
x, weight, channels=16, kernel_size=(3, 3), padding=(1, 1), data_layout="NCHW16c"
)
bias = relay.layout_transform(bias, src_layout="NHWC", dst_layout="NCHW")
bias = relay.layout_transform(bias, src_layout="NCHW", dst_layout="NCHW16c")
add = relay.add(y, bias)
y = relay.layout_transform(add, src_layout="NCHW16c", dst_layout="NCHW")
y = relay.layout_transform(y, src_layout="NCHW", dst_layout="NHWC")
mean = relay.mean(y, axis=3, exclude=True)
var = relay.variance(y, axis=3, exclude=True)
denom = relay.const(1.0) / relay.sqrt(var + relay.const(1e-05))
gamma = relay.var("gamma", shape=(16,))
denom = denom * gamma
denom_expand1 = relay.expand_dims(denom, axis=1, num_newaxis=2)
denom_expand2 = relay.expand_dims(denom_expand1, axis=0)
denom_nchwc16 = relay.layout_transform(
denom_expand2, src_layout="NCHW", dst_layout="NCHW16c"
)
out = add * denom_nchwc16
beta = relay.var("beta", shape=(16,))
numerator = (-mean) * denom + beta
numerator_expand1 = relay.expand_dims(numerator, axis=1, num_newaxis=2)
numerator_expand2 = relay.expand_dims(numerator_expand1, axis=0)
numerator_nchwc16 = relay.layout_transform(
numerator_expand2, src_layout="NCHW", dst_layout="NCHW16c"
)
out = out + numerator_nchwc16
out = relay.layout_transform(out, src_layout="NCHW16c", dst_layout="NCHW")
y = relay.layout_transform(out, src_layout="NCHW", dst_layout="NHWC")
y = relay.Function(analysis.free_vars(y), y)
return y
with TempOpAttr("nn.conv2d", "FTVMAlterOpLayout", alter_conv2d):
a = before()
desired_layouts = {"nn.conv2d": ["NCHW", "default"], "nn.batch_norm": ["NHWC", "default"]}
a = run_opt_pass(
a,
[
transform.InferType(),
relay.transform.ConvertLayout(desired_layouts),
transform.SimplifyInference(),
transform.CanonicalizeOps(),
transform.AlterOpLayout(),
],
)
b = run_opt_pass(expected(), transform.InferType())
assert tvm.ir.structural_equal(a, b), "Actual = \n" + str(a)
def add_functions(mod, funcs):
"""Workaround for type checker and mutually recursive functions."""
for gv in funcs:
func = funcs[gv]
body = _placeholder_body(func.ret_type)
mod[gv] = relay.Function(func.params, body, func.ret_type)
for gv in funcs:
mod[gv] = funcs[gv]
pass_set = transform.Sequential(
passes=[
transform.SimplifyInference(),
transform.CanonicalizeOps(),
transform.CanonicalizeCast(),
transform.FuseOps(3),
# transform.CombineParallelConv2d(),
transform.AlterOpLayout(),
# transform.RewriteAnnotatedOps(???),
],
opt_level=0)
def optimize(mod):
"""Optimize all the functions in a module.
Modules are the only mutable piece of Relay. We write an
optimization pass over the module which destructively updates each
function while optimizing.
def partition_for_dnnl(mod, params=None, alter_layout=True):
"""Partition the graph greedily offloading supported operators to DNNL.
Parameters
----------
mod : Module
The module to run passes on.
params : Optional[Dict[str, NDArray]]
Constant input parameters.
Returns
-------
mod : Module
Annotated and partitioned module.
"""
if params:
mod["main"] = bind_params_by_name(mod["main"], params)
with TempOpAttr("nn.conv2d", "FTVMLegalize", dnnl.legalize_group_conv):
with TempOpAttr("nn.conv2d_transpose", "FTVMLegalize",
dnnl.legalize_group_conv):
seq = tvm.transform.Sequential([
transform.CanonicalizeOps(),
transform.InferType(),
transform.SimplifyInference(),
transform.FoldConstant(),
transform.FoldScaleAxis(),
# fold consecutive add ops to simplify pattern `conv2d-bias_add-bn-relu`
transform.SimplifyExpr(),
transform.FoldConstant(),
# alter group conv /conv_transpose layout to `GOIHW` / `GIOHW`
transform.Legalize(),
transform.FoldConstant(),
])
with tvm.transform.PassContext(opt_level=3):
mod = seq(mod)
if alter_layout:
with TempOpAttr("nn.conv1d", "FTVMAlterOpLayout", dnnl.alter_conv):
with TempOpAttr("nn.conv2d", "FTVMAlterOpLayout", dnnl.alter_conv):
with TempOpAttr("nn.conv3d", "FTVMAlterOpLayout",
dnnl.alter_conv):
with TempOpAttr("nn.conv2d_transpose", "FTVMAlterOpLayout",
dnnl.alter_conv_transpose):
with TempOpAttr("nn.conv3d_transpose",
"FTVMAlterOpLayout",
dnnl.alter_conv_transpose):
alter_layout_seq = tvm.transform.Sequential([
transform.AlterOpLayout(),
transform.FoldConstant(),
])
with tvm.transform.PassContext(opt_level=3):
mod = alter_layout_seq(mod)
byoc_seq = tvm.transform.Sequential([
transform.MergeComposite(dnnl.pattern_table()),
transform.AnnotateTarget("dnnl"),
transform.MergeCompilerRegions(),
transform.PartitionGraph(),
])
with tvm.transform.PassContext(opt_level=3):
mod = byoc_seq(mod)
return mod
