set up inter-wave configuration

include/ck/ck.hpp

@@ -126,7 +126,7 @@
 #define CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST 1
 
 // experimental feature: optimize for inter-wave scheduling policy
-#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING 0
+#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING 1
 #define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS 1
 
 // hack: have underlying assumption that need to be satsified, otherwise it's a bug

include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp

@@ -557,6 +557,10 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
     AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
     BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
 
+#else
+    static_assert(false,
+                  "Configuration error: using BlockwiseGemmXdlopsInterwave without defining "
+                  "macro CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING=1");
 #endif // #if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
 };
 

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp

@@ -426,9 +426,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
         //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
         //       register
         // sanity check
-
-        auto blockwise_gemm =
-            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
             FloatAB,
             FloatAcc,
             decltype(a_block_desc_k0_m_k1),
@@ -437,7 +436,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
             NPerXDL,
             MXdlPerWave,
             NXdlPerWave,
-                                                                K1>{};
+            K1,
+            LoopSched>();
 
         auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();