optimize main loop

example/01_gemm/gemm_xdl_skip_lds_fp16.cpp

@@ -51,7 +51,7 @@ using BDataType   = ck::half_t;
 using CDataType   = ck::half_t;
 using AccDataType = float;
 #else  
-                   <   F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,     64,    16,   16,     4,  1,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,               7,               1>;
+                   <   F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,     64,    16,   16,     4,  4,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>;
 using ADataType   = float;
 using BDataType   = float;
 using CDataType   = float;
@@ -85,14 +85,14 @@ int main(int argc, char* argv[])
     int nrepeat          = 5;
 
     // GEMM shape
-#if 0
+#if 1
     ck::index_t M = 16;
-    ck::index_t N = 4096;
-    ck::index_t K = 64;
+    ck::index_t N = 1152;
+    ck::index_t K = 5120;
 
-    ck::index_t StrideA = 64;
+    ck::index_t StrideA = 5120;
     ck::index_t StrideB = 5120;
-    ck::index_t StrideC = 4096;
+    ck::index_t StrideC = 1152;
 #else
     ck::index_t M = 16;
     ck::index_t N = 16;

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_lds_v2r3.hpp

@@ -687,13 +687,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_skip_lds_v2r3
                 do
                 {
                     a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf);
-                    // block_sync_lds();
                     b_threadwise_copy.Run(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
                                           b_grid_buf,
                                           b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3,
                                           make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
                                           b_thread_odd_buf);
                     blockwise_gemm.ResetABlockStartWindow();
+                    block_sync_lds();
                     blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
 
                     // only move b windows
@@ -705,7 +705,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_skip_lds_v2r3
                                           b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3,
                                           make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
                                           b_thread_even_buf);
-                    // block_sync_lds();
+
                     blockwise_gemm.MoveABlockSliceWindow();
                     blockwise_gemm.Run(a_block_buf, b_thread_odd_buf, c_thread_buf);