Add validity checks for MoE FlatMM scatter and enable bf16 hardware atomic-add (#3236)

* Add validity checks for MoE FlatMM scatter and enable bf16 hardware atomic

* correct clang-format

* removed unused rtol_atol variable from example code

* clang format correction

* remove unused varable max_accumulated_value from example

Author: msaffari-amd

include/ck_tile/core/arch/generic_memory_space_atomic.hpp

@@ -102,6 +102,9 @@ CK_TILE_DEVICE void atomic_add(X* p_dst, const X& x);
 template <>
 CK_TILE_DEVICE void atomic_add<bf16x2_t>(bf16x2_t* p_dst, const bf16x2_t& x)
 {
+#if HAS_GLOBAL_ATOMIC_PK_ADD_BUILTIN
+    __builtin_amdgcn_global_atomic_fadd_v2bf16(c_style_pointer_cast<bf16x2_t*>(p_dst), x);
+#else
     union U32BF162_ADDR
     {
         uint32_t* u32_a;
@@ -128,6 +131,7 @@ CK_TILE_DEVICE void atomic_add<bf16x2_t>(bf16x2_t* p_dst, const bf16x2_t& x)
         new_v      = new_.u32;
         cur_v.u32  = atomicCAS(dword_addr.u32_a, old_v, new_v);
     } while(cur_v.u32 != old_v);
+#endif
 }
 
 template <>

include/ck_tile/ops/flatmm/kernel/moe_flatmm_kernel.hpp

@@ -623,7 +623,7 @@ struct MoeFlatmmKernel
             {
                 return make_naive_tensor_view<address_space_enum::global, DstInMemOp>(
                     e_ptr,
-                    make_tuple(IsInputGemm ? kargs.NumTokens * kargs.TopK : kargs.NumToken,
+                    make_tuple(IsInputGemm ? kargs.NumTokens * kargs.TopK : kargs.NumTokens,
                                IsGateUp ? kargs.N / 2 : kargs.N),
                     make_tuple(1, kargs.stride_C),
                     number<1>{},
@@ -1250,6 +1250,8 @@ struct MoeFlatmmKernel
             constexpr int MPerThread  = TileEncodingPattern::Y2;
             statically_indexed_array<statically_indexed_array<index_t, MPerThread>, NumMEpiTile>
                 c_scatter_offsets;
+            statically_indexed_array<statically_indexed_array<bool, MPerThread>, NumMEpiTile>
+                c_scatter_valids;
             auto c_coord = dram_tile_distribution.calculate_index();
             static_for<0, NumMEpiTile, 1>{}([&](auto mIter) {
                 static_for<0, MPerThread, 1>{}([&](auto m0) {
@@ -1262,6 +1264,7 @@ struct MoeFlatmmKernel
                         scatter_token_id =
                             scatter_token_id * kargs.TopK + (fused_token >> token_id_offset);
                     c_scatter_offsets[mIter][m0] = scatter_token_id * kargs.stride_C;
+                    c_scatter_valids[mIter][m0]  = (scatter_token_id < kargs.NumTokens);
                 });
             });
 
@@ -1302,7 +1305,8 @@ struct MoeFlatmmKernel
                                              c_block_window.get_window_lengths(),
                                              c_block_window.get_window_origin(),
                                              dram_tile_distribution,
-                                             c_scatter_offsets[mIter]);
+                                             c_scatter_offsets[mIter],
+                                             c_scatter_valids[mIter]);
 
                 if constexpr(!IsInputGemm ||
                              EpiloguePipeline::MemoryOperation == memory_operation_enum::atomic_add)