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 2021/03/02 17:09:48 UTC

[GitHub] [tvm] kazimuth opened a new issue #7565: Bring-your-own-codegen c demo doesn't seem to handle tuples

kazimuth opened a new issue #7565:
URL: https://github.com/apache/tvm/issues/7565


   The bring-your-own-codegen c compiler demo fails to handle the following relay IR:
   ```
   def @main(%x: Tensor[(10, 10), float32], %y: Tensor[(10, 10), float32]) {
     %0 = (%x, %y);
     %5 = fn (%xy: (Tensor[(10, 10), float32], Tensor[(10, 10), float32]), Primitive=1, Compiler="ccompiler", global_symbol="tuple_thingamabob") {
       %1 = %xy.0;
       %2 = %xy.1;
       %3 = add(%1, %2);
       %4 = multiply(%1, %2);
       (%3, %4)
     };
     %6 = %5(%0);
     %6.0
   }
   ```
   Trying to compile this with the Relay VM throws an exception: `Check failed: res.size() > static_cast<size_t>(op->index) (1 vs. 1)`, [here](https://github.com/apache/tvm/blob/b52267e/src/relay/backend/contrib/codegen_c/codegen.cc#L72). However, if the external codegen attributes (`Compiler` and `global_symbol`) are removed, then it works fine.
   
   Repro python script (based on `tests/python/relay/test_external_codegen.py`):
   
   ```python
   import os
   import sys
   import numpy as np
   
   import tvm
   import tvm.relay.testing
   import tvm.relay.transform
   from tvm import relay
   from tvm import runtime
   from tvm.contrib import utils
   
   def set_external_func_attr(func, compiler, ext_symbol):
       func = func.with_attr("Primitive", tvm.tir.IntImm("int32", 1))
       func = func.with_attr("Compiler", compiler)
       func = func.with_attr("global_symbol", ext_symbol)
       return func
   
   
   def check_result(mod, map_inputs, out_shape, result, tol=1e-5, target="llvm", ctx=tvm.cpu()):
       if sys.platform == "win32":
           print("Skip test on Windows for now")
           return
   
       def update_lib(lib):
           test_dir = os.path.dirname(os.path.realpath(os.path.expanduser(__file__)))
           source_dir = os.path.join(test_dir, "..", "..", "..")
           contrib_path = os.path.join(source_dir, "src", "runtime", "contrib")
   
           for m in lib.imported_modules:
               if m.type_key == 'c':
                   print(m.get_source())
   
           kwargs = {}
           kwargs["options"] = ["-O2", "-std=c++14", "-I" + contrib_path]
           tmp_path = utils.tempdir()
           lib_name = "lib.so"
           lib_path = tmp_path.relpath(lib_name)
           lib.export_library(lib_path, fcompile=False, **kwargs)
           lib = tvm.runtime.load_module(lib_path)
   
           return lib
   
       def check_vm_result():
           with tvm.transform.PassContext(opt_level=3, disabled_pass=["AlterOpLayout"]):
               exe = relay.vm.compile(mod, target=target)
           code, lib = exe.save()
           lib = update_lib(lib)
           exe = runtime.vm.Executable.load_exec(code, lib)
           vm = runtime.vm.VirtualMachine(exe, ctx)
           out = vm.run(**map_inputs)
           tvm.testing.assert_allclose(out.asnumpy(), result, rtol=tol, atol=tol)
   
       def check_graph_runtime_result():
           with tvm.transform.PassContext(opt_level=3, disabled_pass=["AlterOpLayout"]):
               json, lib, _ = relay.build(mod, target=target)
           lib = update_lib(lib)
           rt_mod = tvm.contrib.graph_runtime.create(json, lib, ctx)
   
           for name, data in map_inputs.items():
               rt_mod.set_input(name, data)
           rt_mod.run()
           out = tvm.nd.empty(out_shape, ctx=ctx)
           out = rt_mod.get_output(0, out)
   
           tvm.testing.assert_allclose(out.asnumpy(), result, rtol=tol, atol=tol)
   
       check_vm_result()
       #check_graph_runtime_result() # tuples don't work in graph runtime anyway afaik
   
   def test_tuple():
       x = relay.var("x", shape=(10, 10))
       y = relay.var("y", shape=(10, 10))
       tuple = relay.Tuple([x, y])
   
       xy = relay.var("xy", relay.TupleType([
           relay.TensorType((10, 10), dtype='float32'),
           relay.TensorType((10, 10), dtype='float32'),
       ]))
       x_ = relay.TupleGetItem(xy, 0)
       y_ = relay.TupleGetItem(xy, 1)
       sum = x_ + y_
       prod = x_ * y_
       result = relay.Tuple([sum, prod])
       subgraph = relay.Function([xy], result)
       subgraph = set_external_func_attr(subgraph, "ccompiler", "tuple_thingamabob")
   
       call = relay.Call(subgraph, [tuple])
       result = relay.TupleGetItem(call, 0)
       f = relay.Function([x, y], result)
       mod = tvm.IRModule()
       mod["main"] = f
   
       print(mod)
   
       x_data = np.random.rand(10, 10).astype("float32")
       y_data = np.random.rand(10, 10).astype("float32")
   
       check_result(
           mod,
           {
               'x': x_data,
               'y': y_data
           },
           (10, 10),
           (x_data + y_data)
       )
   
   if __name__ == '__main__':
       test_tuple()
   ```


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



[GitHub] [tvm] comaniac commented on issue #7565: Bring-your-own-codegen c demo doesn't seem to handle tuples

Posted by GitBox <gi...@apache.org>.
comaniac commented on issue #7565:
URL: https://github.com/apache/tvm/issues/7565#issuecomment-789081636


   Thanks for your interest. As its name indicates, the C codegen in BYOC is only for demo purpose so its functionality is very limited. You are welcome to send a PR to add the tuple support to the BYOC C codegen.


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



[GitHub] [tvm] comaniac closed issue #7565: Bring-your-own-codegen c demo doesn't seem to handle tuples

Posted by GitBox <gi...@apache.org>.
comaniac closed issue #7565:
URL: https://github.com/apache/tvm/issues/7565


   


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