masahi commented on a change in pull request #5272: [BYOC] Add example of
Composite + Annotate for DNNL fused op
URL: https://github.com/apache/incubator-tvm/pull/5272#discussion_r405765549
##########
File path: tests/python/relay/test_pass_partition_graph.py
##########
@@ -856,6 +857,111 @@ def expected():
partitioned = transform.PartitionGraph()(mod)
assert tvm.ir.structural_equal(partitioned, ref_mod, map_free_vars=True)
+
+def test_partition_conv_bias_relu():
+ def make_pattern():
+ data = relay.var("data", relay.TensorType((1, 3, 224, 224), "float32"))
+ weight = relay.var("weight")
+ bias = relay.var("bias")
+ conv = relay.nn.conv2d(data=data, weight=weight, kernel_size=(3, 3),
+ channels=8, padding=(1, 1))
+ add = relay.add(conv, bias)
+ return relay.nn.relu(add)
+
+ def get_blocks(prefix, data, in_channel, out_channel,
+ include_bn=True, include_sigmoid=False):
+ weight = relay.var(prefix + "weight")
+ bn_gamma = relay.var(prefix + "bn_gamma")
+ bn_beta = relay.var(prefix + "bn_beta")
+ bn_mmean = relay.var(prefix + "bn_mean")
+ bn_mvar = relay.var(prefix + "bn_var")
+
+ layer = relay.nn.conv2d(data=data, weight=weight, kernel_size=(3, 3),
+ channels=out_channel, padding=(1, 1))
+ if include_bn:
+ bn_output = relay.nn.batch_norm(layer, bn_gamma, bn_beta,
+ bn_mmean, bn_mvar)
+ layer = bn_output[0]
+ if include_sigmoid:
+ # dummy layer to prevent pattern detection
+ layer = relay.sigmoid(layer)
+ layer = relay.nn.relu(layer)
+ return layer
+
+ def get_net(include_bn=True, include_sigmoid=False):
+ data = relay.var("data", relay.TensorType((1, 3, 224, 224), "float32"))
+ layer1 = get_blocks("layer1_", data, 3, 8, include_bn, include_sigmoid)
+ layer2 = get_blocks("layer2_", layer1, 8, 8, include_bn,
include_sigmoid)
+ return relay.Function(relay.analysis.free_vars(layer2), layer2)
+
+ def get_partitoned_mod(mod, params):
+ # This is required for constant folding
+ mod["main"] = bind_params_by_name(mod["main"], params)
+ pattern_table = [
+ ("dnnl.conv_bias_relu", make_pattern())
+ ]
+ remove_bn_pass = transform.Sequential([
+ transform.InferType(),
+ transform.SimplifyInference(),
+ transform.FoldConstant(),
+ transform.FoldScaleAxis(),
+ ])
+ composite_partition = transform.Sequential([
+ remove_bn_pass,
+ transform.MergeComposite(pattern_table),
+ transform.AnnotateTarget("dnnl"),
Review comment:
I'm not familiar with what MergeCompilerRegions does, but assuming that it
merges consecutive ops that are supported by the same compiler, yes it is
possible in this case because all ops involved in the composite above, conv2d,
add, and relu, are supported by the dnnl backend
https://github.com/apache/incubator-tvm/blob/master/python/tvm/relay/op/contrib/dnnl.py#L59-L64.
But I think composite is supposed to be more general, it works even if
`_register_external_op_helper("add")` is missing in dnnl.py above, for example.
The purpose of composite is to target a "complex" op whose ops that constitute
it are not necessary supported by the compiler by itself.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
