clean

host/driver_offline/include/device_gemm_xdlops_mk_kn_mn.hpp

@@ -331,37 +331,10 @@ void device_gemm_xdlops_mk_kn_mn(const Tensor<ABType>& a_m_k,
 
     constexpr index_t CThreadTransferDstScalarPerVector = 1;
 #elif 1
-    constexpr index_t BlockSize = 64;
-
-    constexpr index_t MPerBlock = 48;
-    constexpr index_t NPerBlock = 32;
-    constexpr index_t KPerBlock = 4;
-
-    constexpr index_t MPerXDL = 16;
-    constexpr index_t NPerXDL = 16;
-    constexpr index_t K1      = 8;
-
-    constexpr index_t MRepeat = 3;
-    constexpr index_t NRepeat = 2;
-
-    using ABlockTransferThreadSliceLengths_K0_M_K1   = Sequence<4, 1, 8>;
-    using ABlockTransferThreadClusterLengths_K0_M_K1 = Sequence<1, 48, 1>;
-
-    constexpr index_t ABlockTransferSrcScalarPerVector_K1 = 8;
-    constexpr index_t ABlockTransferDstScalarPerVector_K1 = 8;
-
-    using BBlockTransferThreadSliceLengths_K0_N_K1   = Sequence<4, 1, 8>;
-    using BBlockTransferThreadClusterLengths_K0_N_K1 = Sequence<1, 32, 1>;
-
-    constexpr index_t BBlockTransferSrcScalarPerVector_N  = 1;
-    constexpr index_t BBlockTransferDstScalarPerVector_K1 = 8;
-
-    constexpr index_t CThreadTransferDstScalarPerVector = 1;
-#elif 1
-    constexpr index_t BlockSize = 64;
+    constexpr index_t BlockSize = 256;
 
-    constexpr index_t MPerBlock = 48;
-    constexpr index_t NPerBlock = 16;
+    constexpr index_t MPerBlock = 96;
+    constexpr index_t NPerBlock = 128;
     constexpr index_t KPerBlock = 4;
 
     constexpr index_t MPerXDL = 16;
@@ -369,18 +342,18 @@ void device_gemm_xdlops_mk_kn_mn(const Tensor<ABType>& a_m_k,
     constexpr index_t K1      = 8;
 
     constexpr index_t MRepeat = 3;
-    constexpr index_t NRepeat = 1;
+    constexpr index_t NRepeat = 4;
 
-    using ABlockTransferThreadSliceLengths_K0_M_K1   = Sequence<4, 1, 8>;
-    using ABlockTransferThreadClusterLengths_K0_M_K1 = Sequence<1, 48, 1>;
+    using ABlockTransferThreadSliceLengths_K0_M_K1   = Sequence<1, 3, 8>;
+    using ABlockTransferThreadClusterLengths_K0_M_K1 = Sequence<4, 32, 1>;
 
     constexpr index_t ABlockTransferSrcScalarPerVector_K1 = 8;
     constexpr index_t ABlockTransferDstScalarPerVector_K1 = 8;
 
-    using BBlockTransferThreadSliceLengths_K0_N_K1   = Sequence<4, 1, 8>;
-    using BBlockTransferThreadClusterLengths_K0_N_K1 = Sequence<1, 16, 1>;
+    using BBlockTransferThreadSliceLengths_K0_N_K1   = Sequence<1, 2, 8>;
+    using BBlockTransferThreadClusterLengths_K0_N_K1 = Sequence<4, 64, 1>;
 
-    constexpr index_t BBlockTransferSrcScalarPerVector_N  = 1;
+    constexpr index_t BBlockTransferSrcScalarPerVector_N  = 2;
     constexpr index_t BBlockTransferDstScalarPerVector_K1 = 8;
 
     constexpr index_t CThreadTransferDstScalarPerVector = 1;