[Codegen][Tuner] default tuning spec available per-SKU #19748
Conversation
bangtianliu
requested review from
kuhar,
antiagainst,
qedawkins,
hanhanW,
MaheshRavishankar and
ScottTodd
as code owners
bangtianliu
force-pushed
the
add_default_td
branch
from
72347e3 to
c1b7532
Compare
Signed-off-by: Bangtian Liu <[email protected]>
bangtianliu
force-pushed
the
add_default_td
branch
from
c1b7532 to
fc8c3e1
Compare
Signed-off-by: Bangtian Liu <[email protected]>
| @@ -34,6 +34,8 @@ | |||
| #define DBGS() (llvm::dbgs() << "[" DEBUG_TYPE "]: ") | |||
| #define LDBG(X) LLVM_DEBUG(DBGS() << X << "\n") | |||
|
|
|||
| extern llvm::cl::opt<std::string> clTestTarget; | |||
There was a problem hiding this comment.
In general, we should no reference global variables like this. This makes the code hard to maintain and LLVM converged on external storage for flags meant to be external, and having these flag storage variables declared in headers.
Here specifically, we should not rely on test flags in this code. All the information we use must come from the gpu target attr. If the information we need there is not available, we should work on adding it.
| // Entry point | ||
| //===----------------------------------------------------------------------===// | ||
|
|
||
| transform.named_sequence @__kernel_config(%variant_op: !transform.any_op {transform.consumed}) -> !transform.any_op |
There was a problem hiding this comment.
We can stage this PR such that we first commit a simple tuning spec to make sure everything works, and then separately work on adding configs we care about.
These configs should be tested to make sure they apply on the intended code, and we should quantify the improvements making sure we don't populate these specs with configs that give us only marginal improvements. This needs to be backed by data.
|
This already landed as a few smaller PRs |
This PR addresses the tasks outlined in the following issues: #19720 and #19721.