[ROCDL] Added asynccnt and tensorcnt wait ops (gfx1250) (#163533)

This patch introduces some missing `s.wait.asynccnt` and `tensorcnt`
instructions in the ROCDL dialect.

The op is lowered to `@llvm.amdgcn.s.wait.asynccnt(i16 0)` and
`@llvm.amdgcn.s.wait.tensorcnt(i16 0)` intrinsic calls

Author: ravil-mobile

mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td

@@ -412,6 +412,32 @@ def ROCDL_WaitExpcntOp: ROCDL_ConcreteNonMemIntrOp<"s.wait.expcnt", [], 0, [0],
   let assemblyFormat = "$count attr-dict";
 }
 
+def ROCDL_WaitAsynccntOp: ROCDL_ConcreteNonMemIntrOp<"s.wait.asynccnt", [], 0, [0], ["count"]>,
+  Arguments<(ins I16Attr:$count)> {
+  let summary = "Wait until ASYNCCNT is less than or equal to `count`";
+  let description = [{
+      Wait for the counter specified to be less-than or equal-to the `count`
+      before continuing.
+
+      Available on gfx1250+.
+  }];
+  let results = (outs);
+  let assemblyFormat = "$count attr-dict";
+}
+
+def ROCDL_WaitTensorcntOp: ROCDL_ConcreteNonMemIntrOp<"s.wait.tensorcnt", [], 0, [0], ["count"]>,
+  Arguments<(ins I16Attr:$count)> {
+  let summary = "Wait until TENSORCNT is less than or equal to `count`";
+  let description = [{
+      Wait for the counter specified to be less-than or equal-to the `count`
+      before continuing.
+
+      Available on gfx1250+.
+  }];
+  let results = (outs);
+  let assemblyFormat = "$count attr-dict";
+}
+
 def ROCDL_SetPrioOp : ROCDL_ConcreteNonMemIntrOp<"s.setprio", [], 0, [0], ["priority"]>,
   Arguments<(ins I16Attr:$priority)> {
   let assemblyFormat = "$priority attr-dict";

mlir/test/Dialect/LLVMIR/rocdl.mlir

@@ -1035,6 +1035,20 @@ llvm.func @rocdl.s.wait.expcnt() {
   llvm.return
 }
 
+llvm.func @rocdl.s.wait.asynccnt() {
+  // CHECK-LABEL: rocdl.s.wait.asynccnt
+  // CHECK: rocdl.s.wait.asynccnt 0
+  rocdl.s.wait.asynccnt 0
+  llvm.return
+}
+
+llvm.func @rocdl.s.wait.tensorcnt() {
+  // CHECK-LABEL: rocdl.s.wait.tensorcnt
+  // CHECK: rocdl.s.wait.tensorcnt 0
+  rocdl.s.wait.tensorcnt 0
+  llvm.return
+}
+
 // -----
 
 llvm.func @rocdl.readfirstlane(%src : f32) -> f32 {

mlir/test/Target/LLVMIR/rocdl.mlir

@@ -276,6 +276,20 @@ llvm.func @rocdl.s.wait.expcnt() {
   llvm.return
 }
 
+llvm.func @rocdl.s.wait.asynccnt() {
+  // CHECK-LABEL: rocdl.s.wait.asynccnt
+  // CHECK-NEXT: call void @llvm.amdgcn.s.wait.asynccnt(i16 0)
+  rocdl.s.wait.asynccnt 0
+  llvm.return
+}
+
+llvm.func @rocdl.s.wait.tensorcnt() {
+  // CHECK-LABEL: rocdl.s.wait.tensorcnt
+  // CHECK-NEXT: call void @llvm.amdgcn.s.wait.tensorcnt(i16 0)
+  rocdl.s.wait.tensorcnt 0
+  llvm.return
+}
+
 llvm.func @rocdl.setprio() {
   // CHECK: call void @llvm.amdgcn.s.setprio(i16 0)
   rocdl.s.setprio 0