You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2020/01/21 17:57:53 UTC
[GitHub] [incubator-tvm] abergeron opened a new issue #4758: [relay] ADT
match becomes not well formed during VM optimization
abergeron opened a new issue #4758: [relay] ADT match becomes not well formed during VM optimization
URL: https://github.com/apache/incubator-tvm/issues/4758
I have this sample program:
```python
import tvm
from tvm import relay
from tvm.relay import adt
ctx = tvm.ndarray.context('cpu', 0)
mod = relay.Module({})
union_type = relay.GlobalTypeVar("u")
c0_type = relay.ty.TupleType([relay.ty.scalar_type('int32'), union_type()])
c0 = adt.Constructor("c0", [c0_type], union_type)
c1 = adt.Constructor("c1", [relay.ty.TupleType([])], union_type)
mod[union_type] = adt.TypeData(union_type, [], [c0, c1])
gv = relay.GlobalVar("fn")
p = relay.var('p', union_type())
v = relay.Var('v')
cond = adt.Match(p, [adt.Clause(adt.PatternConstructor(c0, [adt.PatternWildcard\
()]), relay.const(True)),
adt.Clause(adt.PatternWildcard(), relay.const(False))])
mm = adt.Match(p, [adt.Clause(adt.PatternConstructor(c0, [adt.PatternVar(v)]), \
v)], complete=False)
fn = relay.Function(
[p],
relay.If(
cond,
relay.const(0),
relay.add(relay.TupleGetItem(mm, 0),
relay.Call(gv, [relay.TupleGetItem(mm, 1)]))
),
ret_type=relay.ty.scalar_type('int32')
)
mod[gv] = fn
q = relay.var("q", union_type())
mod["main"] = relay.Function([q], relay.Call(gv, [q]))
print(str(mod))
vm = relay.create_executor(mod=mod, ctx=ctx, target='llvm', kind="vm")
```
The module looks like this before compiling:
```
v0.0.4
type u {
c0((int32, u[])),
c1(()),
}
def @main(%q: u[]) -> int32 {
@fn(%q) /* ty=int32 */
}
def @fn(%p: u[]) -> int32 {
%0 = match (%p) {
c0(_) => True /* ty=bool */,
_ => False /* ty=bool */,
};
if (%0) {
0 /* ty=int32 */
} else {
%1 = match? (%p) {
c0(%v: (int32, u[])) => %v,
};
%2 = %1.0;
%3 = %1.1;
%4 = @fn(%3) /* ty=int32 */;
add(%2, %4) /* ty=int32 */
}
}
```
And I get this error during compilation with kind="vm", but it works correctly for kind="debug":
```
Traceback (most recent call last):
File "tst.py", line 42, in <module>
vm = relay.create_executor(mod=mod, ctx=ctx, target='llvm', kind="vm")
File "/home/anakha/ext/tvm/python/tvm/relay/build_module.py", line 411, in create_executor
return VMExecutor(mod, ctx, target)
File "/home/anakha/ext/tvm/python/tvm/relay/backend/vm.py", line 540, in __init__
self.executable = compile(mod, target)
File "/home/anakha/ext/tvm/python/tvm/relay/backend/vm.py", line 399, in compile
compiler.lower(mod, target, target_host)
File "/home/anakha/ext/tvm/python/tvm/relay/backend/vm.py", line 455, in lower
self._lower(mod, target, target_host)
File "tvm/_ffi/_cython/./function.pxi", line 304, in tvm._ffi._cy3.core.FunctionBase.__call__
File "tvm/_ffi/_cython/./function.pxi", line 239, in tvm._ffi._cy3.core.FuncCall
File "tvm/_ffi/_cython/./function.pxi", line 228, in tvm._ffi._cy3.core.FuncCall3
File "tvm/_ffi/_cython/./base.pxi", line 157, in tvm._ffi._cy3.core.CALL
tvm._ffi.base.TVMError: Traceback (most recent call last):
[bt] (8) /home/anakha/ext/tvm/build/libtvm.so(+0x2097b7d) [0x7f0bebd40b7d]
[bt] (7) /home/anakha/ext/tvm/build/libtvm.so(+0x2097d9c) [0x7f0bebd40d9c]
[bt] (6) /home/anakha/ext/tvm/build/libtvm.so(+0x2097f08) [0x7f0bebd40f08]
[bt] (5) /home/anakha/ext/tvm/build/libtvm.so(+0x2097579) [0x7f0bebd40579]
[bt] (4) /home/anakha/ext/tvm/build/libtvm.so(tvm::relay::vm::PrimitiveInliner::Inline()+0x276) [0x7f0bebd41a1c]
[bt] (3) /home/anakha/ext/tvm/build/libtvm.so(tvm::IRModuleNode::Add(tvm::GlobalVar const&, tvm::BaseFunc const&, bool)+0xcb) [0x7f0beb3daca3]
[bt] (2) /home/anakha/ext/tvm/build/libtvm.so(tvm::RunTypeCheck(tvm::IRModule const&, tvm::GlobalVar const&, tvm::relay::Function)+0x5b) [0x7f0beb3da662]
[bt] (1) /home/anakha/ext/tvm/build/libtvm.so(tvm::relay::DeDup(tvm::RelayExpr const&)+0x106) [0x7f0bebb460cd]
[bt] (0) /home/anakha/ext/tvm/build/libtvm.so(dmlc::LogMessageFatal::~LogMessageFatal()+0x4a) [0x7f0beb36f884]
File "/home/anakha/ext/tvm/src/relay/pass/de_duplicate.cc", line 109
TVMError: Check failed: WellFormed(e): v0.0.4
fn (%p: u[]) -> int32 {
%0 = match (%p) {
c0(_) => True /* ty=bool */,
_ => False /* ty=bool */,
};
if (%0) {
0 /* ty=int32 */
} else {
%1 = match? (%p) {
c0(%v: (int32, u[])) => %v,
};
%2 = %1.0;
%3 = match? (%p) {
c0(%v: (int32, u[])) => %v,
};
%4 = %3.1;
%5 = @fn(%4);
add(%2, %5)
}
}
```
The only thing I can notice is that something duplicated the match in the else branch and this probably makes the code not well formed. I traced the WellFormed Error and it comes from %v. I may be wrong about this.
----------------------------------------------------------------
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:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] abergeron commented on issue #4758: [relay]
PrimitiveInliner doesn't correctly handle match expressions that have more
than one use
Posted by GitBox <gi...@apache.org>.
abergeron commented on issue #4758: [relay] PrimitiveInliner doesn't correctly handle match expressions that have more than one use
URL: https://github.com/apache/incubator-tvm/issues/4758#issuecomment-580367317
Fixed by #4783
----------------------------------------------------------------
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:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] abergeron closed issue #4758: [relay]
PrimitiveInliner doesn't correctly handle match expressions that have more
than one use
Posted by GitBox <gi...@apache.org>.
abergeron closed issue #4758: [relay] PrimitiveInliner doesn't correctly handle match expressions that have more than one use
URL: https://github.com/apache/incubator-tvm/issues/4758
----------------------------------------------------------------
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:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] abergeron commented on issue #4758: [relay] ADT
match becomes not well formed during VM optimization
Posted by GitBox <gi...@apache.org>.
abergeron commented on issue #4758: [relay] ADT match becomes not well formed during VM optimization
URL: https://github.com/apache/incubator-tvm/issues/4758#issuecomment-578918024
After some experimentation, if I comment out InlinePrimitives at line 921 in compiler.cc (https://github.com/apache/incubator-tvm/blob/master/src/relay/backend/vm/compiler.cc#L921), then the compilation works. It seems this is the line that introduces the weird change.
If I add an override like this:
```
Expr VisitExpr_(const MatchNode* m) {
std::vector<Clause> clauses;
for (const Clause& p : m->clauses) {
clauses.push_back(VisitClause(p));
}
return GetRef<Expr>(m);
}
```
to the PrimitiveInliner class (in https://github.com/apache/incubator-tvm/blob/master/src/relay/backend/vm/inline_primitives.cc#L54), it appears to make the compilation work (at least in my limited example), but this may have some other consequences that I am not aware of (I suspect this means that no transformations will be applied to the body of match clauses, which might be bad).
I'll keep working on that for a bit trying to find an acceptable solution, but I will gladly take any help/hints that I can get.
----------------------------------------------------------------
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:
users@infra.apache.org
With regards,
Apache Git Services