trevor-m commented on issue #5261: [RELAY][BYOC] Add support for composite 
functions in BYOC
URL: https://github.com/apache/incubator-tvm/pull/5261#issuecomment-610663673
 
 
   > Using MergeComposite, AnnotateTarget and PartitionGraph, I get the 
following graph for conv + bias + relu pattern:
   > 
   > ```
   > def @dnnl_0(%dnnl_0_i0: Tensor[(1, 3, 224, 224), float32], Inline=1, 
Compiler="dnnl", global_symbol=runtime.String(0x55a8d9cddbd0), Primitive=1) -> 
Tensor[(1, 1, 224, 224), float32] {
   >   %2 = fn (%data: Tensor[(1, 3, 224, 224), float32], %weight: Tensor[(1, 
3, 3, 3), float32], %bias: Tensor[(1, 1, 1), float32], 
Composite="dnnl.conv_bias_relu") -> Tensor[(1, 1, 224, 224), float32] {
   >     %0 = nn.conv2d(%data, %weight, padding=[1, 1, 1, 1], channels=1, 
kernel_size=[3, 3]) /* ty=Tensor[(1, 1, 224, 224), float32] */;
   >     %1 = add(%0, %bias) /* ty=Tensor[(1, 1, 224, 224), float32] */;
   >     nn.relu(%1) /* ty=Tensor[(1, 1, 224, 224), float32] */
   >   };
   >   %2(%dnnl_0_i0, meta[relay.Constant][0] /* ty=Tensor[(1, 3, 3, 3), 
float32] */ /* ty=Tensor[(1, 3, 3, 3), float32] */, meta[relay.Constant][1] /* 
ty=Tensor[(1, 1, 1), float32] */ /* ty=Tensor[(1, 1, 1), float32] */) /* 
ty=Tensor[(1, 1, 224, 224), float32] */
   > }
   > 
   > def @main(%data1: Tensor[(1, 3, 224, 224), float32]) -> Tensor[(1, 1, 224, 
224), float32] {
   >   @dnnl_0(%data1) /* ty=Tensor[(1, 1, 224, 224), float32] */
   > }
   > ```
   > 
   > Is it possible to inline composite function `%2` there into `dnnl_0`? What 
I want is this:
   > 
   > ```
   > def @dnnl_0(%dnnl_0_i0: Tensor[(1, 3, 224, 224), float32], Inline=1, 
Compiler="dnnl", global_symbol=runtime.String(0x5599b307c370), Primitive=1) -> 
Tensor[(1, 1, 224, 224), float32] {
   >   %0 = nn.conv2d(%dnnl_0_i0, meta[relay.Constant][0] /* ty=Tensor[(1, 3, 
3, 3), float32] */ /* ty=Tensor[(1, 3, 3, 3), float32] */, padding=[1, 1, 1, 
1], channels=1, kernel_size=[3, 3]) /* ty=Tensor[(1, 1, 224, 224), float32] */;
   >   %1 = add(%0, meta[relay.Constant][1] /* ty=Tensor[(1, 1, 1), float32] */ 
/* ty=Tensor[(1, 1, 1), float32] */) /* ty=Tensor[(1, 1, 224, 224), float32] */;
   >   nn.relu(%1) /* ty=Tensor[(1, 1, 224, 224), float32] */
   > }
   > 
   > def @main(%data: Tensor[(1, 3, 224, 224), float32]) -> Tensor[(1, 1, 224, 
224), float32] {
   >   @dnnl_0(%data) /* ty=Tensor[(1, 1, 224, 224), float32] */
   > }
   > ```
   > 
   > Otherwise I have to support function call in DNNL codegen. @zhiics @mbaret
   > 
   > UPDATE: hmm if I inline the composite function, I lose the composite 
attribute and hence cannot detect fused call. Is supporting function call in 
the DNNL codegen a better option?
   
   Hi @masahi, I think it is best to leave this to the codegen. That way we 
have the option to handle a composite all at once at call node or invididually 
by just doing VisitExpr(func->body) when we encounter one.

----------------------------------------------------------------
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

Reply via email to