[llvm-branch-commits] [mlir] [mlir][AMDGPU] Update gather_to_lds with explicit-async support (PR #181082)

[Krzysztof Drewniak via llvm-branch-commits]

================
@@ -1099,14 +1100,21 @@ def AMDGPU_GatherToLDSOp :
     * `$transferType`: type of the data to be transferred by each thread. This is used to determine
       the size of the data to be transferred and the number of threads in the subgroup.
       The transfer type must be a scalar type or a vector type with a single element type.
+    * If `$async` is set, the compiler will not attempt to infer the
+      memory waits needed to ensure that the DMA operation has succeeded
+      before a load that might access the stored-to LDS is performed.
+      Instead, the `rocdl.asyncmark` and `rocdl.wait.asyncmark N`
+      operations must be used to explicitly indicate the desired completion
+      behavior. This enables more precise calculation of these waits at the
+      cost of requiring user management of asynchrony.
 
     The `$dst`, along with its indices, points to the memory location the subgroup of this thread
     will write to.
 
     Note: only supported on gfx9 and gfx10.
   }];
   let assemblyFormat = [{
-    $src `[` $srcIndices `]` `,` $dst `[` $dstIndices `]` attr-dict `:` $transferType `,` type($src) `,` type($dst)
+    (`async` $async^)? $src `[` $srcIndices `]` `,` $dst `[` $dstIndices `]` attr-dict `:` $transferType `,` type($src) `,` type($dst)
----------------
krzysz00 wrote:

`amdgpu.sched_barrier` is allowed because the enum there is adding ergonomic value - it makes the IR clearer than the integer constants you get in ROCDL 

I wrote up some documentation on this, but to roughly summarize, amdgpu-level wrappers come in to interact with the MLIR type system in a richer way, for meaningful ergonomic improvements (representing enums, say), and/or if you need to create an extended version of an intrinsic that you're planning to unroll/bitcast/... .

That is, `amdgpu.sched_barrier` has a reason to exist, `amdgpu.s.sleep` would *not* because it's exactly 1:1 with its rocdl counterpart.

(Or, to put it another way, there's no point is wasting bytes on a rewriter that just changes the op name)

https://github.com/llvm/llvm-project/pull/181082