[GitHub] [tvm] lazycal opened a new issue, #12008: Resize does not reject unsupported layout during AlterOpLayout

[GitBox <gi...@apache.org>]

lazycal opened a new issue, #12008:
URL: https://github.com/apache/tvm/issues/12008

   This model
   ```python
   def @main(%x: Tensor[(1, 1, 1, 1), float32] /* ty=Tensor[(1, 1, 1, 1), float32] */) -> Tensor[(1, 1, 1, 1), float32] {
     %0 = image.resize2d(%x, size=[2, 4], roi=[0f, 0f, 0f, 0f], rounding_method="") /* ty=Tensor[(1, 1, 2, 4), float32] */;
     %1 = transpose(%0, axes=[0, 1, 3, 2]) /* ty=Tensor[(1, 1, 4, 2), float32] */;
     %2 = nn.max_pool2d(%1, pool_size=[1, 1], padding=[0, 0, 0, 0]) /* ty=Tensor[(1, 1, 4, 2), float32] */;
     image.resize2d(%2, size=[1, 1], roi=[0f, 0f, 0f, 0f], rounding_method="") /* ty=Tensor[(1, 1, 1, 1), float32] */
   }
   ```
   fails with error `ValueError: NCWH layout is not supported.`, and this layout is introduced after AlterOpLayout (see below for the IR right after AlterOpLayout):
   ```python
   def @main(%x: Tensor[(1, 1, 1, 1), float32] /* ty=Tensor[(1, 1, 1, 1), float32] */) -> Tensor[(1, 1, 1, 1), float32] {
     %0 = image.resize2d(%x, size=[2, 4], roi=[0f, 0f, 0f, 0f], rounding_method="") /* ty=Tensor[(1, 1, 2, 4), float32] */;
     %1 = transpose(%0, axes=[0, 1, 3, 2]) /* ty=Tensor[(1, 1, 4, 2), float32] */;
     %2 = layout_transform(%1, src_layout="NCHW", dst_layout="NCWH") /* ty=Tensor[(1, 1, 2, 4), float32] */;
     %3 = nn.max_pool2d(%2, pool_size=[1, 1], padding=[0, 0, 0, 0], layout="NCWH") /* ty=Tensor[(1, 1, 2, 4), float32] */;
     %4 = image.resize2d(%3, size=[1, 1], roi=[0f, 0f, 0f, 0f], layout="NCWH", rounding_method="") /* ty=Tensor[(1, 1, 1, 1), float32] */;
     layout_transform(%4, src_layout="NCWH", dst_layout="NCHW") /* ty=Tensor[(1, 1, 1, 1), float32] */
   }
   ```
   where the max_pool2d outputs NCWH layout and resize2d accepts it during AlterOpLayout but complains when lowering in this function:
   https://github.com/apache/tvm/blob/5efe8b0bfdff4c9939185a7581dc77e23cbcb6d5/python/tvm/topi/image/resize.py#L807-L824
   
   Not sure if this is a bug, but I guess ideally resize2d should not accept unsupported layout? The current logic
   https://github.com/apache/tvm/blob/5efe8b0bfdff4c9939185a7581dc77e23cbcb6d5/src/relay/op/image/resize.cc#L49-L53
   does not check this.
   
   ### Expected behavior
   Pass the compilation.
   
   ### Actual behavior
   Pass with opt_level=2 but fail with opt_level=4
   
   ### Environment
   Ubuntu 18.04
   TVM: 1787cca3f90237dff001fba01ffdbaf9a510f886
   ### Steps to reproduce
   
   Run this code:
   ```python
   import tvm
   from tvm import relay
   from tvm.relay.transform import InferType
   
   
   @tvm.instrument.pass_instrument
   class PrintIR:
       """Print the name of the pass, the IR, only before passes execute."""
   
       def __init__(self, print_mod=True, show_meta_data=False) -> None:
           self.pass_cnt = 0
           self.print_mod = print_mod
           self.show_meta_data = show_meta_data
   
       def run_before_pass(self, mod, info):
           with tvm.transform.PassContext(instruments=[]):
               global prev_mod
               if self.print_mod:
                   print(relay.transform.InferType()(mod))
               print('>' * 40, f'Running Pass#{self.pass_cnt}:', info)
   
           self.pass_cnt += 1
   
   
   x = relay.var("x", shape=(1, 1, 1, 1))
   y = relay.image.resize2d(x, (2, 4))
   z = relay.transpose(y, (0, 1, 3, 2))
   a = relay.nn.max_pool2d(z)
   b = relay.image.resize2d(a, (1, 1))
   func = relay.Function((x,), b)
   mod = tvm.IRModule.from_expr(func)
   print(InferType()(mod))
   
   with tvm.transform.PassContext(opt_level=4, instruments=[PrintIR()]):
       relay.create_executor("graph", mod).evaluate()
   ```
   Part of error log:
   ```
     6: _ZZN3tvm5relay11ExprFunctorIFNS_7runtime5ArrayINS_2te6TensorEvEERKNS_
     5: tvm::relay::tec::LowerToTECompute::VisitExpr_(tvm::relay::CallNode const*)
     4: tvm::relay::backend::MemoizedExprTranslator<tvm::runtime::Array<tvm::te::Tensor, void> >::VisitExpr(tvm::RelayExpr const&)
     3: tvm::relay::ExprFunctor<tvm::runtime::Array<tvm::te::Tensor, void> (tvm::RelayExpr const&)>::VisitExpr(tvm::RelayExpr const&)
     2: _ZZN3tvm5relay11ExprFunctorIFNS_7runtime5ArrayINS_2te6TensorEvEERKNS_
     1: tvm::relay::tec::LowerToTECompute::VisitExpr_(tvm::relay::CallNode const*)
     0: tvm::runtime::PackedFuncObj::Extractor<tvm::runtime::PackedFuncSubObj<TVMFuncCreateFromCFunc::{lambda(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)#2}> >::Call(tvm::runtime::PackedFuncObj const*, tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)
     File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
     File "/workspace/workspace/tvm-intact/python/tvm/relay/backend/te_compiler.py", line 317, in lower_call
       best_impl, outputs = select_implementation(op, call.attrs, inputs, ret_type, target)
     File "/workspace/workspace/tvm-intact/python/tvm/relay/backend/te_compiler.py", line 207, in select_implementation
       outs = impl.compute(attrs, inputs, out_type)
     File "/workspace/workspace/tvm-intact/python/tvm/relay/op/op.py", line 126, in compute
       return _OpImplementationCompute(self, attrs, inputs, out_type)
     File "tvm/_ffi/_cython/./packed_func.pxi", line 331, in tvm._ffi._cy3.core.PackedFuncBase.__call__
     File "tvm/_ffi/_cython/./packed_func.pxi", line 276, in tvm._ffi._cy3.core.FuncCall
     File "tvm/_ffi/_cython/./base.pxi", line 181, in tvm._ffi._cy3.core.CHECK_CALL
     3: TVMFuncCall
     2: tvm::runtime::PackedFuncObj::Extractor<tvm::runtime::PackedFuncSubObj<tvm::relay::{lambda(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)#4}> >::Call(tvm::runtime::PackedFuncObj const*, tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)
     1: tvm::relay::OpImplementation::Compute(tvm::Attrs const&, tvm::runtime::Array<tvm::te::Tensor, void> const&, tvm::Type const&)
     0: tvm::runtime::PackedFuncObj::Extractor<tvm::runtime::PackedFuncSubObj<TVMFuncCreateFromCFunc::{lambda(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)#2}> >::Call(tvm::runtime::PackedFuncObj const*, tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)
     File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback
     File "/workspace/workspace/tvm-intact/python/tvm/relay/op/image/_image.py", line 156, in compute_resize2d
       out_dtype,
     File "/workspace/workspace/tvm-intact/python/tvm/topi/image/resize.py", line 824, in resize2d
       raise ValueError("%s layout is not supported." % layout)
   ValueError: NCWH layout is not supported.
   ```


-- 
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.

To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org