[Mlir-commits] [mlir] e652eba - [mlir][acc] Return after NYI for non-NVIDIA per-row barrier (#238116)

[llvmlistbot at llvm.org]

Author: Scott Manley
Date: 2026-11-02T16:09:01-06:00
New Revision: e652eba3de8afe9a5aa73cc41293cae32c9ecbfc

URL: https://github.com/llvm/llvm-project/commit/e652eba3de8afe9a5aa73cc41293cae32c9ecbfc
DIFF: https://github.com/llvm/llvm-project/commit/e652eba3de8afe9a5aa73cc41293cae32c9ecbfc.diff

LOG: [mlir][acc] Return after NYI for non-NVIDIA per-row barrier (#238116)

createPerRowBarrier reports that a per-row barrier is not yet
implemented on non-NVIDIA devices, but it does not return afterwards.
Execution falls through to the NVVM named-barrier code, which asserts
that the device type is NVIDIA, so acc-cg-to-gpu aborts instead of
failing cleanly. This is reachable with the default device type for any
thread-partitioned loop wider than a subgroup. For example,
acc-cg-to-gpu-barrier-after-thread-loop.mlir crashes if its
device-type=nvidia option is removed.

Return after emitting the diagnostic, matching the other NYI sites in
this file, and add a test that checks the diagnostic for this case.

Added: 
    mlir/test/Dialect/OpenACC/acc-cg-to-gpu-per-row-barrier-unsupported.mlir

Modified: 
    mlir/lib/Dialect/OpenACC/Transforms/ACCCGToGPU.cpp

Removed: 
    


################################################################################
diff  --git a/mlir/lib/Dialect/OpenACC/Transforms/ACCCGToGPU.cpp b/mlir/lib/Dialect/OpenACC/Transforms/ACCCGToGPU.cpp
index 0038de3831f562..e5a280f7a9a7f9 100644
--- a/mlir/lib/Dialect/OpenACC/Transforms/ACCCGToGPU.cpp
+++ b/mlir/lib/Dialect/OpenACC/Transforms/ACCCGToGPU.cpp
@@ -1865,6 +1865,7 @@ void ACCCGToGPULowering::createPerRowBarrier(Location loc) {
     (void)accSupport.emitNYI(
         loc,
         "per-row barrier to support worker parallelism on non-NVIDIA device");
+    return;
   }
 
   // Per-row barrier with fully runtime branching.

diff  --git a/mlir/test/Dialect/OpenACC/acc-cg-to-gpu-per-row-barrier-unsupported.mlir b/mlir/test/Dialect/OpenACC/acc-cg-to-gpu-per-row-barrier-unsupported.mlir
new file mode 100644
index 00000000000000..27f1572ab951f6
--- /dev/null
+++ b/mlir/test/Dialect/OpenACC/acc-cg-to-gpu-per-row-barrier-unsupported.mlir
@@ -0,0 +1,35 @@
+// RUN: mlir-opt %s --pass-pipeline="builtin.module(func.func(acc-cg-to-gpu))" -verify-diagnostics
+
+// A row wider than a subgroup is synchronized with an NVVM named barrier,
+// which has no portable equivalent. The pass runs here with the default
+// device type rather than nvidia, so the per-row barrier is reported as not
+// yet implemented instead of being emitted.
+
+func.func @per_row_barrier_non_nvidia(%a: memref<128xf32>, %b: memref<128xf32>) {
+  %c1_pw = arith.constant 1 : index
+  %c128_pw = arith.constant 128 : index
+  %bx = acc.par_width %c1_pw par_dim(#acc.par_dim<block_x>)
+  %tx = acc.par_width %c128_pw par_dim(#acc.par_dim<thread_x>)
+  acc.compute_region launch(%kbx = %bx, %ktx = %tx) ins(%ma = %a, %mb = %b) : (memref<128xf32>, memref<128xf32>) {
+    %c0 = arith.constant 0 : index
+    %c1 = arith.constant 1 : index
+    %c3 = arith.constant 3 : index
+    %c4 = arith.constant 4 : index
+    %c128 = arith.constant 128 : index
+    %one = arith.constant 1.0 : f32
+    scf.parallel (%g) = (%c0) to (%kbx) step (%c1) {
+      scf.for %it = %c0 to %c4 step %c1 {
+        // expected-error @below {{not yet implemented: per-row barrier to support worker parallelism on non-NVIDIA device}}
+        scf.parallel (%i) = (%c0) to (%c128) step (%c1) {
+          memref.store %one, %ma[%i] : memref<128xf32>
+          scf.reduce
+        } {acc.par_dims = #acc<par_dims[thread_x]>}
+        %v = memref.load %ma[%c3] : memref<128xf32>
+        memref.store %v, %mb[%it] : memref<128xf32>
+      }
+      scf.reduce
+    } {acc.par_dims = #acc<par_dims[block_x]>}
+    acc.yield
+  } <{origin = "acc.kernels"}>
+  return
+}