[Target] Introduce Target Id Registry

[junrushao]

Per RFC TVM Target Specification.

We implemented a target id registry which supports the following syntax:

TVM_REGISTER_TARGET_ID("llvm")
.set_attr<TPreCodegenPass>("TPreCodegenPass", ...)
.add_attr_option<Bool>("system_lib")
.add_attr_option<String>("mtriple")
.add_attr_option<String>("mattr");

[tqchen]

It would be good to have simple testcases around target id

[tqchen]

@junrushao1994 please rebase against the master

[ANSHUMAN87]

@junrushao1994 : Thanks for the PR! Great work 👍

Please find some high level comments. Hope it is helpful. Thanks!

[junrushao]

Thank you @ANSHUMAN87 for the review!

[junrushao]

@tqchen could you take another look?

Changes

• Added schema validation
• Added several tests

[tqchen]

Note about the second level validation:

{
    "id": "cuda",
    "tag": "nvidia/tx2-cudnn",
    "keys": ["cuda", "gpu"],
    "libs": ["cudnn"],
    "target_host": {
        "id": "llvm",
        "system_lib": True,
        "mtriple": "aarch64-linux-gnu",
        "mattr": "+neon"
    }
}

In the above example, the schema of cuda looks like

- keys: Array[str]
- target_host: Target

And we need to be able to validate the information in llvm. For example, if we add an invalid attribute to the target host map(that is not registered by llvm) it will trigger an error.

[junrushao]

@tqchen As far as I could understand, we need validation twice, one for target, one for target_host.

Therefore, the step-by-step procedure should be:

1. Let "target" map be the given map removing "target_host" entry
2. Figure out the "id" entry in "target"
3. Validate "target" using the TargetId indicated by "id"
4. Validate "target_host" in the same way.

Is that correct?

BTW, why the schema of cuda doesn't include "libs" and "tag"?

[tqchen]

Sorrt i omitted the libs and tags, they should be. Would be great if we can have a test example

[junrushao]

@tqchen Just added support for target_host, and right now are can support validating the entire map directly via TargetValidateSchema

[tqchen]

see some additional followups about making Target as part of the schema, to recognize perhaps Array<Target> and Target and do recursive handling.

[junrushao]

Future work may include:

1. Refactor the object TargetNode
2. Allow Target to be a special type supported in schema validation, so that composite/recursive targets can be supported more naturally
3. Support constructing TargetNode from JSON

[tqchen]

Thanks @junrushao1994 @MarisaKirisame @ashutoshparkhi