[mlir][AMDGPU] Add gfx941 to buffer atomics emulation

[fmorac]

Overall looks good to me, can you add a description & tests for MaxSIOp & MinUIOp?

[krzysz00]

Ping

mlir/lib/Dialect/AMDGPU/Transforms/EmulateAtomics.cpp

void mlir::amdgpu::populateAmdgpuEmulateAtomicsPatterns(
     ConversionTarget &target, RewritePatternSet &patterns, Chipset chipset) {
   // gfx10 has no atomic adds.
   if (chipset.majorVersion == 10 || chipset.majorVersion < 9 ||
       (chipset.majorVersion == 9 && chipset.minorVersion < 0x08)) {
     target.addIllegalOp<RawBufferAtomicFaddOp>();
   }
   // gfx9 has no to a very limited support for floating-point min and max.
   if (chipset.majorVersion == 9) {
-    if (chipset.minorVersion >= 0x0a) {
+    if (chipset.minorVersion >= 0x0a && chipset.minorVersion != 0x41) {
       // gfx90a supports f64 max (and min, but we don't have a min wrapper right
       // now) but all other types need to be emulated.
       target.addDynamicallyLegalOp<RawBufferAtomicFmaxOp>(
           [](RawBufferAtomicFmaxOp op) -> bool {
             return op.getValue().getType().isF64();
           });
     } else {
       target.addIllegalOp<RawBufferAtomicFmaxOp>();
     }
+    if (chipset.minorVersion == 0x41) {
+      // gfx941 requires non-CAS atomics to be implemented with CAS loops.
+      // The workaround here mirrors HIP and OpenMP.
+      target.addIllegalOp<RawBufferAtomicFaddOp, RawBufferAtomicFmaxOp,
+                          RawBufferAtomicSmaxOp, RawBufferAtomicUminOp>();
+    }
   }
   patterns.add<
       RawBufferAtomicByCasPattern<RawBufferAtomicFaddOp, arith::AddFOp>,
-      RawBufferAtomicByCasPattern<RawBufferAtomicFmaxOp, arith::MaximumFOp>>(
+      RawBufferAtomicByCasPattern<RawBufferAtomicFmaxOp, arith::MaximumFOp>,
+      RawBufferAtomicByCasPattern<RawBufferAtomicSmaxOp, arith::MaxSIOp>,
+      RawBufferAtomicByCasPattern<RawBufferAtomicUminOp, arith::MinUIOp>>(
       patterns.getContext());
 }
 
 void AmdgpuEmulateAtomicsPass::runOnOperation() {
   Operation *op = getOperation();
   FailureOr<Chipset> maybeChipset = Chipset::parse(chipset);
   if (failed(maybeChipset)) {
     emitError(op->getLoc(), "Invalid chipset name: " + chipset);