[Mlir-commits] [mlir] 5829652 - [MLIR][NVVM] Add clusterlaunchcontrol Ops (#156585)

llvmlistbot at llvm.org llvmlistbot at llvm.org
Wed Sep 17 00:19:16 PDT 2025


Author: Srinivasa Ravi
Date: 2025-09-17T12:49:12+05:30
New Revision: 5829652a6def4031853104b43ee6dfd331162352

URL: https://github.com/llvm/llvm-project/commit/5829652a6def4031853104b43ee6dfd331162352
DIFF: https://github.com/llvm/llvm-project/commit/5829652a6def4031853104b43ee6dfd331162352.diff

LOG: [MLIR][NVVM] Add clusterlaunchcontrol Ops (#156585)

This change adds the `clusterlaunchcontrol.try.cancel` and
`clusterlaunchcontrol.query.cancel` Ops to the NVVM dialect.

Tests are added in `clusterlaunchcontrol.mlir`.

PTX Reference:
https://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions-clusterlaunchcontrol-try-cancel

Added: 
    mlir/test/Target/LLVMIR/nvvm/clusterlaunchcontrol.mlir

Modified: 
    mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
    mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
    mlir/test/Target/LLVMIR/nvvmir-invalid.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td b/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
index 70ab1df876d35..797f8ada9f238 100644
--- a/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
+++ b/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
@@ -4431,6 +4431,116 @@ def NVVM_DotAccumulate2WayOp : NVVM_Op<"dot.accumulate.2way"> {
   }];
 }
 
+//===----------------------------------------------------------------------===//
+// NVVM clusterlaunchcontrol Ops.
+//===----------------------------------------------------------------------===//
+
+def NVVM_ClusterLaunchControlTryCancelOp
+    : NVVM_Op<"clusterlaunchcontrol.try.cancel", [NVVMRequiresSM<100>]> {
+  let summary = "Request atomically canceling the launch of a cluster that has not started running yet";
+  let description = [{
+    `clusterlaunchcontrol.try.cancel` requests atomically canceling the launch 
+    of a cluster that has not started running yet. It asynchronously writes an 
+    opaque response to shared memory indicating whether the operation succeeded 
+    or failed.
+
+    Operand `smemAddress` specifies the naturally aligned address of the 
+    16-byte wide shared memory location where the request's response is written.
+
+    Operand `mbarrier` specifies the mbarrier object used to track the 
+    completion of the asynchronous operation.
+
+    If `multicast` is specified, the response is asynchronously written to the 
+    corresponding local shared memory location (specifed by `addr`) of each CTA 
+    in the requesting cluster.
+
+    [For more information, see PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions-clusterlaunchcontrol-try-cancel)
+  }];
+
+  let arguments = (ins UnitAttr:$multicast, 
+                       LLVM_PointerShared: $smemAddress,
+                       LLVM_PointerShared: $mbarrier);
+
+  let assemblyFormat = "(`multicast` $multicast^ `,`)? $smemAddress `,` $mbarrier attr-dict";
+
+  let extraClassDeclaration = [{
+    static mlir::NVVM::IDArgPair
+    getIntrinsicIDAndArgs(Operation &op, LLVM::ModuleTranslation &mt,
+                          llvm::IRBuilderBase &builder);
+  }];
+  
+  string llvmBuilder = [{
+    auto [id, args] = 
+    NVVM::ClusterLaunchControlTryCancelOp::getIntrinsicIDAndArgs(
+                        *op, moduleTranslation, builder);
+    createIntrinsicCall(builder, id, args);
+  }];
+}
+
+def ClusterLaunchControlIsCanceled
+  : I32EnumCase<"IS_CANCELED", 0, "is_canceled">;
+def ClusterLaunchControlGetFirstCTAIDX
+  : I32EnumCase<"GET_FIRST_CTA_ID_X", 1, "get_first_cta_id_x">;
+def ClusterLaunchControlGetFirstCTAIDY
+  : I32EnumCase<"GET_FIRST_CTA_ID_Y", 2, "get_first_cta_id_y">;
+def ClusterLaunchControlGetFirstCTAIDZ
+  : I32EnumCase<"GET_FIRST_CTA_ID_Z", 3, "get_first_cta_id_z">;
+
+def ClusterLaunchControlQueryType
+  : I32Enum<"ClusterLaunchControlQueryType",
+      "NVVM ClusterLaunchControlQueryType", 
+      [ClusterLaunchControlIsCanceled, ClusterLaunchControlGetFirstCTAIDX,
+      ClusterLaunchControlGetFirstCTAIDY, ClusterLaunchControlGetFirstCTAIDZ]> {
+  let cppNamespace = "::mlir::NVVM";
+}
+
+def ClusterLaunchControlQueryTypeAttr
+  : EnumAttr<NVVM_Dialect,
+      ClusterLaunchControlQueryType, "cluster_launch_control_query_type"> {
+  let assemblyFormat = "$value";
+}
+
+def NVVM_ClusterLaunchControlQueryCancelOp
+    : NVVM_Op<"clusterlaunchcontrol.query.cancel", [NVVMRequiresSM<100>]> {
+  let summary = "Query the response of a clusterlaunchcontrol.try.cancel operation";
+  let description = [{
+    `clusterlaunchcontrol.query.cancel` queries the response of a 
+    `clusterlaunchcontrol.try.cancel` operation specified by operand 
+    `try_cancel_response`.
+
+    Operand `query_type` specifies the type of query to perform and can be one 
+    of the following:
+    - `is_canceled` : Returns true if the try cancel request succeeded, 
+    and false otherwise.
+    - `get_first_cta_id_{x/y/z}` : Returns the x, y, or z coordinate of the 
+    first CTA in the canceled cluster. Behaviour is defined only if the try 
+    cancel request succeeded. 
+
+    [For more information, see PTX ISA](https://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions-clusterlaunchcontrol-query-cancel)
+  }];
+
+  let arguments = (ins ClusterLaunchControlQueryTypeAttr:$query_type, 
+                       I128:$try_cancel_response);
+  let results = (outs AnyTypeOf<[I1, I32]>:$res);
+                                 
+  let assemblyFormat = "`query` `=` $query_type `,` $try_cancel_response attr-dict `:` type($res)";
+  
+  let hasVerifier = 1;
+
+  let extraClassDeclaration = [{
+    static mlir::NVVM::IDArgPair
+    getIntrinsicIDAndArgs(Operation &op, LLVM::ModuleTranslation &mt,
+                          llvm::IRBuilderBase &builder);
+  }];
+  
+  string llvmBuilder = [{
+    auto [id, args] = 
+    NVVM::ClusterLaunchControlQueryCancelOp::getIntrinsicIDAndArgs(
+                        *op, moduleTranslation, builder);
+    $res = createIntrinsicCall(builder, id, args);
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // NVVM target attribute.
 //===----------------------------------------------------------------------===//

diff  --git a/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp b/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
index 9dba0848eba94..13f1dd9a664e5 100644
--- a/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
+++ b/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
@@ -1401,6 +1401,24 @@ LogicalResult NVVM::PrefetchOp::verify() {
   return success();
 }
 
+LogicalResult NVVM::ClusterLaunchControlQueryCancelOp::verify() {
+  switch (getQueryType()) {
+  case NVVM::ClusterLaunchControlQueryType::IS_CANCELED:
+    if (!getType().isInteger(1))
+      return emitOpError("is_canceled query type returns an i1");
+    break;
+  case NVVM::ClusterLaunchControlQueryType::GET_FIRST_CTA_ID_X:
+  case NVVM::ClusterLaunchControlQueryType::GET_FIRST_CTA_ID_Y:
+  case NVVM::ClusterLaunchControlQueryType::GET_FIRST_CTA_ID_Z:
+    if (!getType().isInteger(32)) {
+      return emitOpError("get_first_cta_id_x, get_first_cta_id_y, "
+                         "get_first_cta_id_z query types return an i32");
+    }
+    break;
+  }
+  return success();
+}
+
 /// Packs the given `field` into the `result`.
 /// The `result` is 64-bits and each `field` can be 32-bits or narrower.
 static llvm::Value *
@@ -2087,6 +2105,51 @@ bool NVVM::InlinePtxOp::getAsmValues(
   return false; // No manual mapping needed
 }
 
+NVVM::IDArgPair ClusterLaunchControlTryCancelOp::getIntrinsicIDAndArgs(
+    Operation &op, LLVM::ModuleTranslation &mt, llvm::IRBuilderBase &builder) {
+  auto curOp = cast<NVVM::ClusterLaunchControlTryCancelOp>(op);
+  llvm::SmallVector<llvm::Value *> args;
+  args.push_back(mt.lookupValue(curOp.getSmemAddress()));
+  args.push_back(mt.lookupValue(curOp.getMbarrier()));
+
+  llvm::Intrinsic::ID intrinsicID =
+      curOp.getMulticast()
+          ? llvm::Intrinsic::
+                nvvm_clusterlaunchcontrol_try_cancel_async_multicast_shared
+          : llvm::Intrinsic::nvvm_clusterlaunchcontrol_try_cancel_async_shared;
+
+  return {intrinsicID, args};
+}
+
+NVVM::IDArgPair ClusterLaunchControlQueryCancelOp::getIntrinsicIDAndArgs(
+    Operation &op, LLVM::ModuleTranslation &mt, llvm::IRBuilderBase &builder) {
+  auto curOp = cast<NVVM::ClusterLaunchControlQueryCancelOp>(op);
+  llvm::SmallVector<llvm::Value *> args;
+  args.push_back(mt.lookupValue(curOp.getTryCancelResponse()));
+
+  llvm::Intrinsic::ID intrinsicID;
+
+  switch (curOp.getQueryType()) {
+  case NVVM::ClusterLaunchControlQueryType::IS_CANCELED:
+    intrinsicID =
+        llvm::Intrinsic::nvvm_clusterlaunchcontrol_query_cancel_is_canceled;
+    break;
+  case NVVM::ClusterLaunchControlQueryType::GET_FIRST_CTA_ID_X:
+    intrinsicID = llvm::Intrinsic::
+        nvvm_clusterlaunchcontrol_query_cancel_get_first_ctaid_x;
+    break;
+  case NVVM::ClusterLaunchControlQueryType::GET_FIRST_CTA_ID_Y:
+    intrinsicID = llvm::Intrinsic::
+        nvvm_clusterlaunchcontrol_query_cancel_get_first_ctaid_y;
+    break;
+  case NVVM::ClusterLaunchControlQueryType::GET_FIRST_CTA_ID_Z:
+    intrinsicID = llvm::Intrinsic::
+        nvvm_clusterlaunchcontrol_query_cancel_get_first_ctaid_z;
+    break;
+  }
+  return {intrinsicID, args};
+}
+
 //===----------------------------------------------------------------------===//
 // NVVMDialect initialization, type parsing, and registration.
 //===----------------------------------------------------------------------===//

diff  --git a/mlir/test/Target/LLVMIR/nvvm/clusterlaunchcontrol.mlir b/mlir/test/Target/LLVMIR/nvvm/clusterlaunchcontrol.mlir
new file mode 100644
index 0000000000000..6ff85fa38a0a0
--- /dev/null
+++ b/mlir/test/Target/LLVMIR/nvvm/clusterlaunchcontrol.mlir
@@ -0,0 +1,55 @@
+// RUN: mlir-translate -mlir-to-llvmir %s | FileCheck %s
+
+llvm.func @clusterlaunchcontrol_try_cancel(%addr: !llvm.ptr<3>, %mbar: !llvm.ptr<3>) {
+  // CHECK-LABEL: define void @clusterlaunchcontrol_try_cancel(ptr addrspace(3) %0, ptr addrspace(3) %1) {
+  // CHECK-NEXT: call void @llvm.nvvm.clusterlaunchcontrol.try_cancel.async.shared(ptr addrspace(3) %0, ptr addrspace(3) %1)
+  // CHECK-NEXT: ret void
+  // CHECK-NEXT: }
+  nvvm.clusterlaunchcontrol.try.cancel %addr, %mbar
+  llvm.return
+}
+
+llvm.func @clusterlaunchcontrol_try_cancel_multicast(%addr: !llvm.ptr<3>, %mbar: !llvm.ptr<3>) {
+  // CHECK-LABEL: define void @clusterlaunchcontrol_try_cancel_multicast(ptr addrspace(3) %0, ptr addrspace(3) %1) {
+  // CHECK-NEXT: call void @llvm.nvvm.clusterlaunchcontrol.try_cancel.async.multicast.shared(ptr addrspace(3) %0, ptr addrspace(3) %1)
+  // CHECK-NEXT: ret void
+  // CHECK-NEXT: }
+  nvvm.clusterlaunchcontrol.try.cancel multicast, %addr, %mbar
+  llvm.return
+}
+
+llvm.func @clusterlaunchcontrol_query_cancel_is_canceled(%try_cancel_response: i128) {
+  // CHECK-LABEL: define void @clusterlaunchcontrol_query_cancel_is_canceled(i128 %0) {
+  // CHECK-NEXT: %2 = call i1 @llvm.nvvm.clusterlaunchcontrol.query_cancel.is_canceled(i128 %0)
+  // CHECK-NEXT: ret void
+  // CHECK-NEXT: }
+  %res = nvvm.clusterlaunchcontrol.query.cancel query = is_canceled, %try_cancel_response : i1
+  llvm.return
+}
+
+llvm.func @clusterlaunchcontrol_query_cancel_get_first_cta_id_x(%try_cancel_response: i128) {
+  // CHECK-LABEL: define void @clusterlaunchcontrol_query_cancel_get_first_cta_id_x(i128 %0) {
+  // CHECK-NEXT: %2 = call i32 @llvm.nvvm.clusterlaunchcontrol.query_cancel.get_first_ctaid.x(i128 %0)
+  // CHECK-NEXT: ret void
+  // CHECK-NEXT: }
+  %res = nvvm.clusterlaunchcontrol.query.cancel query = get_first_cta_id_x, %try_cancel_response : i32
+  llvm.return
+}
+
+llvm.func @clusterlaunchcontrol_query_cancel_get_first_cta_id_y(%try_cancel_response: i128) {
+  // CHECK-LABEL: define void @clusterlaunchcontrol_query_cancel_get_first_cta_id_y(i128 %0) {
+  // CHECK-NEXT: %2 = call i32 @llvm.nvvm.clusterlaunchcontrol.query_cancel.get_first_ctaid.y(i128 %0)
+  // CHECK-NEXT: ret void
+  // CHECK-NEXT: }
+  %res = nvvm.clusterlaunchcontrol.query.cancel query = get_first_cta_id_y, %try_cancel_response : i32
+  llvm.return
+}
+
+llvm.func @clusterlaunchcontrol_query_cancel_get_first_cta_id_z(%try_cancel_response: i128) {
+  // CHECK-LABEL: define void @clusterlaunchcontrol_query_cancel_get_first_cta_id_z(i128 %0) {
+  // CHECK-NEXT: %2 = call i32 @llvm.nvvm.clusterlaunchcontrol.query_cancel.get_first_ctaid.z(i128 %0)
+  // CHECK-NEXT: ret void
+  // CHECK-NEXT: }
+  %res = nvvm.clusterlaunchcontrol.query.cancel query = get_first_cta_id_z, %try_cancel_response : i32
+  llvm.return
+}

diff  --git a/mlir/test/Target/LLVMIR/nvvmir-invalid.mlir b/mlir/test/Target/LLVMIR/nvvmir-invalid.mlir
index b35a6dbcca286..383f4829f3287 100644
--- a/mlir/test/Target/LLVMIR/nvvmir-invalid.mlir
+++ b/mlir/test/Target/LLVMIR/nvvmir-invalid.mlir
@@ -535,3 +535,19 @@ llvm.func @nanosleep() {
   nvvm.nanosleep 100000000000000
   llvm.return
 }
+
+// -----
+
+llvm.func @clusterlaunchcontrol_query_cancel_is_canceled_invalid_return_type(%try_cancel_response: i128) {
+  // expected-error at +1 {{'nvvm.clusterlaunchcontrol.query.cancel' op is_canceled query type returns an i1}}
+  %res = nvvm.clusterlaunchcontrol.query.cancel query = is_canceled, %try_cancel_response : i32
+  llvm.return
+}
+
+// -----
+
+llvm.func @clusterlaunchcontrol_query_cancel_get_first_cta_id_invalid_return_type(%try_cancel_response: i128) {
+  // expected-error at +1 {{'nvvm.clusterlaunchcontrol.query.cancel' op get_first_cta_id_x, get_first_cta_id_y, get_first_cta_id_z query types return an i32}}
+  %res = nvvm.clusterlaunchcontrol.query.cancel query = get_first_cta_id_x, %try_cancel_response : i1
+  llvm.return
+}


        


More information about the Mlir-commits mailing list