[RELAY] Filter PlaceholderOp from schedule.

[srkreddy1238]

Thanks for contributing to TVM! Please refer to guideline https://docs.tvm.ai/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from Reviewers.

[tqchen]

What is the motivation on this? The compile engine should have included all the inputs including placeholder

[srkreddy1238]

this is to handle a situation like below

--------------------------
fn (%p0: Tensor[(1, 10000), float32],
    %p1: Tensor[(1, 2, 1, 200), float32],
    %p2: Tensor[(1, 2, 1, 200), float32])
    -> Tuple[Tensor[(1, 10000), float32], Tensor[(2, 2, 1, 200), float32]] {
  %0 = (%p1, %p2)
  %1 = concatenate(%0) # ty=Tensor[(2, 2, 1, 200), float32]
  %2 = (%p0, %1)
  %2
}
--------------------------

where p0 is a bypass. Ref. #2216 (comment)

[tqchen]

OK, this seems to be fine, can you please document the intent of the code as a comment block? Please also add a test case as per https://docs.tvm.ai/contribute/code_review.html#ensure-test-coverage

[srkreddy1238]

Test case added, but enabling fuse pass has another issue as given below.

Traceback (most recent call last):
  File "./tests/python/relay/test_backend_compile_engine.py", line 48, in <module>
    test_compile_placeholder_bypass()
  File "./tests/python/relay/test_backend_compile_engine.py", line 44, in test_compile_placeholder_bypass
    graph, lib, params = relay.build(func, 'llvm')
  File "/home/srk/.local/lib/python3.6/site-packages/tvm-0.5.dev0-py3.6-linux-x86_64.egg/tvm/relay/build_module.py", line 276, in build
    lowered_funcs, target=target, target_host=target_host)
  File "/home/srk/.local/lib/python3.6/site-packages/tvm-0.5.dev0-py3.6-linux-x86_64.egg/tvm/build_module.py", line 592, in build
    mhost = codegen.build_module(fhost_all, str(target_host))
  File "/home/srk/.local/lib/python3.6/site-packages/tvm-0.5.dev0-py3.6-linux-x86_64.egg/tvm/codegen.py", line 20, in build_module
    return _Build(lowered_func, target)
  File "/home/srk/.local/lib/python3.6/site-packages/tvm-0.5.dev0-py3.6-linux-x86_64.egg/tvm/_ffi/_ctypes/function.py", line 185, in __call__
    ctypes.byref(ret_val), ctypes.byref(ret_tcode)))
  File "/home/srk/.local/lib/python3.6/site-packages/tvm-0.5.dev0-py3.6-linux-x86_64.egg/tvm/_ffi/base.py", line 72, in check_call
    raise TVMError(py_str(_LIB.TVMGetLastError()))
tvm._ffi.base.TVMError: [11:53:12] /home/srk/work/DMLC/tvm/src/codegen/llvm/llvm_module.cc:173: LLVM module verification failed with the following errors: 
Invalid operand types for FCmp instruction
  %243 = fcmp oeq i8* %34, %242

I may need some help here.

[jroesch]

This makes sense to me, we need someone with more TVM experience to explain the LLVM failure. It looks like the compiler is producing invalid IR.

[tqchen]

@srkreddy1238 can you act on @jroesch 's comment, and we can merge this. The new error seems was due to wrong type info

[srkreddy1238]

cc @tqchen

[tqchen]

Thanks to @srkreddy1238 @jroesch , this is now merged