1. change blockwise gemm loopover direction from kmn to mnk ( ~1% improvement)

2. change kernel timing mode to 50 warmup + 50 timed repeat

example/01_gemm/gemm_wmma_fp16.cpp

@@ -37,14 +37,14 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
            GemmDefault, 
            2,           // Prefetch stage
            128,         // BlockSize
-           64,          // MPerBlock
-           128,         // NPerBlock
+           128,          // MPerBlock
+           64,         // NPerBlock
            64,          // KPerBlock
            8,           // K1
            16,          // MPerWmma
            16,          // NPerWmma
-           2,           // M-Repeat // M-PerWmma / M-Repeat = M-Wave
-           4,           // N-Repeat // N-PerWmma / N-Repeat = N-Wave
+           4,           // M-Repeat // M-PerWmma / M-Repeat = M-Wave
+           2,           // N-Repeat // N-PerWmma / N-Repeat = N-Wave
            S<4, 32, 1>,     
            S<1, 0, 2>,     
            S<1, 0, 2>,              

include/ck/host_utility/kernel_launch.hpp

@@ -32,10 +32,13 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
 
         printf("Warm up 1 time\n");
 #endif
-        // warm up
-        // kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+        const int nrepeat = 50;
+
+        for(int i = 0; i < nrepeat; ++i)    
+        {
+            kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+        }
 
-        const int nrepeat = 100;
 #if DEBUG_LOG
         printf("Start running %d times...\n", nrepeat);
 #endif

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

@@ -312,9 +312,13 @@ struct BlockwiseGemmWMMA
         // basic intrinsic to determine loopover direction
         if constexpr(MRepeat < NRepeat)
         {
+            
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                       
+
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
                             static_for<0, KPerBlock / WmmaK, 1>{}(
                 [&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
-                    static_for<0, MRepeat, 1>{}([&](auto m0) {
                              // read A
                         a_thread_copy_.Run(
                             a_block_desc_k0_m0_m1_m2_k1,
@@ -327,8 +331,6 @@ struct BlockwiseGemmWMMA
                             a_thread_desc_,
                             make_tuple(I0, m0, I0, I0, I0),
                             a_thread_buf);
-
-                        static_for<0, NRepeat, 1>{}([&](auto n0) {
                             // read B
                             b_thread_copy_.Run(
                                 b_block_desc_k0_n0_n1_n2_k1,
@@ -370,9 +372,13 @@ struct BlockwiseGemmWMMA
         }
         else
         {
+           
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        
+
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
                              static_for<0, KPerBlock / WmmaK, 1>{}(
                 [&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
-                    static_for<0, NRepeat, 1>{}([&](auto n0) {
                 // read B
                         b_thread_copy_.Run(
                             b_block_desc_k0_n0_n1_n2_k1,
@@ -385,8 +391,6 @@ struct BlockwiseGemmWMMA
                             b_thread_desc_,
                             make_tuple(I0, n0, I0, I0, I0),
                             b_thread_buf);
-
-                        static_for<0, MRepeat, 1>{}([&](auto m0) {
                             // read A
                             a_thread_copy_.Run(
                                 a_block_desc_k0_m0_m1_m2_k1,