Lines Matching defs:barrier

210 /// async load for which there is not yet a barrier.
868   /// Build a tma load from global memory to shared memory using `barrier` to
873                     TypedValue<nvgpu::MBarrierGroupType> barrier,
875   void buildBarrierArriveTx(TypedValue<nvgpu::MBarrierGroupType> barrier,
884       TypedValue<nvgpu::MBarrierGroupType> barrier);
886   void buildTryWaitParity(TypedValue<nvgpu::MBarrierGroupType> barrier);
895     TypedValue<nvgpu::MBarrierGroupType> barrier) {
911         OpFoldResult sz = buildTmaAsyncLoad(desc, shmem, barrier, loadOps);
914       // TODO: Note that cutlass predeclares the barrier arrive tx before the tma.async.load.
916       buildBarrierArriveTx(barrier, sizes);
923       buildBarrierArriveTx(barrier, getAsIndexOpFoldResult(rewriter.getContext(), 0));
939   Value barrier = rewriter.create<nvgpu::MBarrierCreateOp>(
944       loc, barrier, getValueOrCreateConstantIndexOp(rewriter, loc, numThreads),
947   return cast<TypedValue<nvgpu::MBarrierGroupType>>(barrier);
982     TypedValue<nvgpu::MBarrierGroupType> barrier,
987       loc, sharedMemref, barrier, globalDesc, ValueRange{zero, zero}, zero,
1002     TypedValue<nvgpu::MBarrierGroupType> barrier,
1013   rewriter.create<nvgpu::MBarrierArriveExpectTxOp>(loc, barrier, sizeVal, zero,
1018     TypedValue<nvgpu::MBarrierGroupType> barrier) {
1027   rewriter.create<nvgpu::MBarrierTryWaitParityOp>(loc, barrier, parity,
1051   // 1. Init a barrier object in shared memory.
1062   TypedValue<nvgpu::MBarrierGroupType> barrier =
1089       buildPredicateLoadsOnThread0(globalDescs, shmems, barrier);
1092   buildTryWaitParity(barrier);