[AMD] Fix nondeterministic atomic tests failure on RDNA (triton-lang#8633)

This PR fixes nondeterministic results in Triton's atomic RMW unit tests
on RDNA. The issue was caused by an unnecessary non-offset LDS store of
the atomic return ("old") that every lane executed to the same LDS
address, creating a race without any functional use.

%16 = atomicrmw add ptr addrspace(1) %dst, i64 %val syncscope("agent")
acq_rel
store i64 %16, ptr addrspace(3) @global_smem ; **<-- non-offset LDS
write**

.
.
.

%8  = (lane_id & 63) << 3
%19 = getelementptr i8, ptr addrspace(3) @global_smem, i32 %8
store <1 x i64> %21, ptr addrspace(3) %19
%24 = load  i64, ptr addrspace(3) %19

There is no read from @global_smem at offset 0 after the first store.
Thus, the zero-offset store is dead, but every lane still writes it,
causing write-write contention. For lane 0, the same location later used
for the real value.

**Mismatched elements: 1 / 64 (1.56%)**
Max absolute difference among violations: 1.87127712
Max relative difference among violations: 1.52464041
**ACTUAL:** array([**-0.643921**, -0.221742, 1.451132, -2.439096,
1.336105, 0.748346,
        0.725985, -0.604978,  2.061625,  1.040246, -0.511721, -0.546404,
-0.163719, 0.279938, -0.730161, -0.21177 , -0.323359, 1.263584,...
**DESIRED:** array([ **1.227356**, -0.221742, 1.451132, -2.439096,
1.336105, 0.748346,
        0.725985, -0.604978,  2.061625,  1.040246, -0.511721, -0.546404,
-0.163719, 0.279938, -0.730161, -0.21177 , -0.323359, 1.263584,...
============================== 1 failed in 3.73s
===============================


A non-offset LDS store (e.g., writing to LDS[0]) only makes sense in a
true scalar scenario when You deliberately perform one atomic per wave
(single output).

**Fix**
We removed/guarded the non-offset LDS write so it is not emitted when it
is unsafe or unnecessary:
No non-offset LDS staging for tensor/multi-output atomics.
No LDS staging when the atomic return is unused.

Author: saeid-rostami

python/test/unit/language/test_core.py

@@ -1404,7 +1404,7 @@ def kernel(X):
 @pytest.mark.interpreter
 @pytest.mark.parametrize("shape, axis, num_ctas, dtype_x_str, check_return_val",
                          [(shape, axis, num_ctas, dtype_x_str, check_return_val)
-                          for shape in [(2, 2), (2, 8), (8, 2), (8, 8), (32, 32), (64, 64)]
+                          for shape in [(2, 2), (2, 8), (8, 2), (8, 8), (32, 32), (64, 64), (128, 128)]
                           for axis in [0, 1]
                           for num_ctas in num_ctas_list
                           for dtype_x_str in ['bfloat16', 'float16', 'float32', 'uint64', 'int64', 'float64']

third_party/amd/lib/TritonAMDGPUToLLVM/LoadStoreOpToLLVM.cpp

@@ -1852,8 +1852,9 @@ struct AtomicRMWOpConversion
         emitRedundantThreadPredicate(freeVarMasks, rewriter, loc, targetInfo);
     auto tid = getThreadId(rewriter, loc);
 
+    bool needLdsStaging = !tensorTy && !opResult.use_empty();
     std::optional<Value> atomicSharedMemBase =
-        op->hasAttr("allocation.offset")
+        op->hasAttr("allocation.offset") && needLdsStaging
             ? std::optional<Value>(getSharedMemoryBase(
                   loc, rewriter, targetInfo, op.getOperation()))
             : std::nullopt;