Prior to this commit, if a TIR variable was required to compute the output of BlockBuilder.call_te, but that TIR variable could not be inferred from the shape of any tensor arguments, it would be provided in an optional tir_vars argument to R.call_tir. In C++, this would be then be accessed as an optional
call->args[2].as<ShapeExprNode>().
This extra argument can cause unexpected bugs. For example, the bug that was fixed in https://github.com/apache/tvm/pull/17086 was caused by RewriteDataflowReshape identifying the output buffer using prim_func->buffer_map.Get(prim_func->params.back()), which is only correct if tir_vars is empty. Rather than fixing these issues as they come up, it would be better to make the general Relax guarantees stronger by removing the tir_vars argument altogether.
Use of extra R.shape parameter to specify additional tir_vars predates the existence of relax::PrimValue, and is no longer required. This commit updates BlockBuilder.call_te to use additional relax.PrimValue arguments to handle symbolic values that cannot be inferred from tensor shapes, rather than tir_vars.
masahi
commented
3 weeks ago
Prior to this commit, if a TIR variable was required to compute the output of
BlockBuilder.call_te, but that TIR variable could not be inferred from the shape of any tensor arguments, it would be provided in an optionaltir_varsargument toR.call_tir. In C++, this would be then be accessed as an optionalcall->args[2].as<ShapeExprNode>().This extra argument can cause unexpected bugs. For example, the bug that was fixed in https://github.com/apache/tvm/pull/17086 was caused by
RewriteDataflowReshapeidentifying the output buffer usingprim_func->buffer_map.Get(prim_func->params.back()), which is only correct iftir_varsis empty. Rather than fixing these issues as they come up, it would be better to make the general Relax guarantees stronger by removing thetir_varsargument altogether.Use of extra
R.shapeparameter to specify additionaltir_varspredates the existence ofrelax::PrimValue, and is no longer required. This commit updatesBlockBuilder.call_teto use additionalrelax.PrimValuearguments to handle symbolic values that cannot be inferred from tensor shapes, rather thantir_vars.