added moe interleaving pipeline

include/ck_tile/ops/flatmm.hpp

@@ -5,6 +5,7 @@
 
 #include "ck_tile/ops/flatmm/block/flatmm_32x512x128_1x4x1_16x16x32.hpp"
 #include "ck_tile/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp"
+#include "ck_tile/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32_itl.hpp"
 #include "ck_tile/ops/flatmm/block/flatmm_uk_config.hpp"
 #include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp

@@ -810,21 +810,46 @@ struct FusedMoeGemmPipelineFlatmmPolicy
     CK_TILE_HOST_DEVICE static constexpr auto GetUK_1()
     {
         using S_ = typename Problem::BlockShape;
+        using T_ = typename Problem::Traits;
         if constexpr(std::is_same_v<typename Problem::YDataType, ck_tile::bf16_t> &&
                      std::is_same_v<typename Problem::DDataType, ck_tile::bf16_t> &&
                      std::is_same_v<typename Problem::TopkWeightDataType, float> &&
                      S_::Block_M1 == 32 && S_::Block_N1 == 128 && S_::Block_K1 == 512 &&
-                     S_::Warp_M0 == 16 && S_::Warp_N0 == 16 && S_::Warp_K0 == 32)
+                     S_::Warp_M0 == 16 && S_::Warp_N0 == 16 && S_::Warp_K0 == 32 &&
+                     T_::PipeInterleave == false)
         {
             return FlatmmSn_32x128x512_1x4x1_16x16x32_BF16{};
+            // return FlatmmSn_32x128x512_1x4x1_16x16x32_BF16_itl{};
         }
         else if constexpr(std::is_same_v<typename Problem::YDataType, ck_tile::fp16_t> &&
                           std::is_same_v<typename Problem::DDataType, ck_tile::fp16_t> &&
                           std::is_same_v<typename Problem::TopkWeightDataType, float> &&
                           S_::Block_M1 == 32 && S_::Block_N1 == 128 && S_::Block_K1 == 512 &&
-                          S_::Warp_M0 == 16 && S_::Warp_N0 == 16 && S_::Warp_K0 == 32)
+                          S_::Warp_M0 == 16 && S_::Warp_N0 == 16 && S_::Warp_K0 == 32 &&
+                     T_::PipeInterleave == false)
         {
             return FlatmmSn_32x128x512_1x4x1_16x16x32_FP16{};
+            // return FlatmmSn_32x128x512_1x4x1_16x16x32_FP16_itl{};
+        }
+        else if constexpr(std::is_same_v<typename Problem::YDataType, ck_tile::bf16_t> &&
+                          std::is_same_v<typename Problem::DDataType, ck_tile::bf16_t> &&
+                          std::is_same_v<typename Problem::TopkWeightDataType, float> &&
+                          S_::Block_M1 == 32 && S_::Block_N1 == 128 && S_::Block_K1 == 512 &&
+                          S_::Warp_M0 == 16 && S_::Warp_N0 == 16 && S_::Warp_K0 == 32 &&
+                     T_::PipeInterleave == true)
+        {
+            // return FlatmmSn_32x128x512_1x4x1_16x16x32_FP16{};
+            return FlatmmSn_32x128x512_1x4x1_16x16x32_BF16_itl{};
+        }
+         else if constexpr(std::is_same_v<typename Problem::YDataType, ck_tile::fp16_t> &&
+                          std::is_same_v<typename Problem::DDataType, ck_tile::fp16_t> &&
+                          std::is_same_v<typename Problem::TopkWeightDataType, float> &&
+                          S_::Block_M1 == 32 && S_::Block_N1 == 128 && S_::Block_K1 == 512 &&
+                          S_::Warp_M0 == 16 && S_::Warp_N0 == 16 && S_::Warp_K0 == 32 &&
+                     T_::PipeInterleave == true)
+        {
+            // return FlatmmSn_32x128x512_1x4x1_16x16x32_FP16{};
+            return FlatmmSn_32x128x512_1x4x1_16x16x32_FP16_itl{};
         }
     }
 };

include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_traits.hpp

@@ -22,7 +22,8 @@ template <bool IsGateOnly_,
           FusedMoeGemmWeightPermuteEnum PermuteEnum_ =
               FusedMoeGemmWeightPermuteEnum::b_nr_kr_waveflatten,
           bool PadHiddenSize_       = false,
-          bool PadIntermediateSize_ = false>
+          bool PadIntermediateSize_ = false,
+          bool PipeInterleave_ = true>
 struct FusedMoeGemmTraits
 {
     // Gate+Up or Gate only
@@ -32,6 +33,7 @@ struct FusedMoeGemmTraits
     static constexpr FusedMoeGemmWeightPermuteEnum PermuteEnum = PermuteEnum_;
     static constexpr bool PadHiddenSize                        = PadHiddenSize_;
     static constexpr bool PadIntermediateSize                  = PadIntermediateSize_;
+    static constexpr bool PipeInterleave                       = PipeInterleave_;
 };
 
 // Note: this need to be a bit mask