[Mlir-commits] [mlir] [mlir][amdgpu] Add gfx12+ split barrier wrappers (PR #175579)

[Krzysztof Drewniak]

================
@@ -977,6 +977,69 @@ def AMDGPU_LDSBarrierOp : AMDGPU_Op<"lds_barrier"> {
   let hasCanonicalizer = 1;
 }
 
+def AMDGPU_LDSBarrierSignalOp : AMDGPU_Op<"lds_barrier_signal"> {
+  let summary = "Signal phase of split LDS barrier.";
+  let description = [{
+    `amdgpu.lds_barrier_signal` is the signal phase of a split barrier that
+    initiates a wait for all operations that affect the Local Data Store (LDS)
+    issued from the workgroup. This operation allows computation to be overlapped
+    with LDS memory operations by separating the barrier signal from the wait.
----------------
krzysz00 wrote:

Yeah, especially since this is the release side of a fence - I think the shape of the wording is "all LDS writes before this will be visible after the corresponding wait, and no writes from after that wait will be visible before this signal"

And I'll concur on the naming being weird but also a function of unfortunate naming from back before these async LDS barriers were a thing. (Long-term, we'll want to deprecate `amdgpu.lds_barrier`, but that's pending a bunch of improvements to `gpu.barrier`)

https://github.com/llvm/llvm-project/pull/175579