tweak

host/driver_offline/include/device_gemm_xdlops_mk_nk_mn.hpp

@@ -33,7 +33,7 @@ void device_gemm_xdlops_mk_nk_mn(const ADesc& a_m_k_grid_desc,
     c_m_n_device_buf.ToDevice(c_m_n.mData.data());
 
 #if 0
-    // [M, N, K0, K1] = [256, 128, 4, 4] for fp32
+    // [M, N, K0, K1] = [128, 256, 4, 4] for fp32
     constexpr index_t BlockSize = 256;
 
     constexpr index_t MPerBlock = 128;
@@ -61,6 +61,34 @@ void device_gemm_xdlops_mk_nk_mn(const ADesc& a_m_k_grid_desc,
 
     constexpr index_t CThreadTransferDstScalarPerVector = 1;
 #elif 1
+    // [M, N, K0, K1] = [256, 128, 4, 8] for fp16
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t MPerBlock = 256;
+    constexpr index_t NPerBlock = 128;
+    constexpr index_t KPerBlock = 4;
+
+    constexpr index_t MPerXDL = 32;
+    constexpr index_t NPerXDL = 32;
+    constexpr index_t K1      = 8;
+
+    constexpr index_t MRepeat = 4;
+    constexpr index_t NRepeat = 2;
+
+    using ABlockTransferThreadSliceLengths_K0_M_K1   = Sequence<1, 4, 8>;
+    using ABlockTransferThreadClusterLengths_K0_M_K1 = Sequence<4, 64, 1>;
+
+    constexpr index_t ABlockTransferSrcScalarPerVector_K1 = 8;
+    constexpr index_t ABlockTransferDstScalarPerVector_K1 = 8;
+
+    using BBlockTransferThreadSliceLengths_K0_N_K1   = Sequence<1, 2, 8>;
+    using BBlockTransferThreadClusterLengths_K0_N_K1 = Sequence<4, 64, 1>;
+
+    constexpr index_t BBlockTransferSrcScalarPerVector_K1 = 8;
+    constexpr index_t BBlockTransferDstScalarPerVector_K1 = 8;
+
+    constexpr index_t CThreadTransferDstScalarPerVector = 1;
+#elif 0
     // [M, N, K0, K1] = [128, 256, 4, 8] for fp16
     constexpr index_t BlockSize = 256;
 
@@ -87,6 +115,34 @@ void device_gemm_xdlops_mk_nk_mn(const ADesc& a_m_k_grid_desc,
     constexpr index_t BBlockTransferSrcScalarPerVector_K1 = 8;
     constexpr index_t BBlockTransferDstScalarPerVector_K1 = 8;
 
+    constexpr index_t CThreadTransferDstScalarPerVector = 1;
+#elif 1
+    // [M, N, K0, K1] = [128, 128, 4, 8] for fp16
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t MPerBlock = 128;
+    constexpr index_t NPerBlock = 128;
+    constexpr index_t KPerBlock = 4;
+
+    constexpr index_t MPerXDL = 32;
+    constexpr index_t NPerXDL = 32;
+    constexpr index_t K1      = 8;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 2;
+
+    using ABlockTransferThreadSliceLengths_K0_M_K1   = Sequence<1, 2, 8>;
+    using ABlockTransferThreadClusterLengths_K0_M_K1 = Sequence<4, 64, 1>;
+
+    constexpr index_t ABlockTransferSrcScalarPerVector_K1 = 8;
+    constexpr index_t ABlockTransferDstScalarPerVector_K1 = 8;
+
+    using BBlockTransferThreadSliceLengths_K0_N_K1   = Sequence<1, 2, 8>;
+    using BBlockTransferThreadClusterLengths_K0_N_K1 = Sequence<4, 64, 1>;
+
+    constexpr index_t BBlockTransferSrcScalarPerVector_K1 = 8;
+    constexpr index_t BBlockTransferDstScalarPerVector_K1 = 8;
+
     constexpr index_t CThreadTransferDstScalarPerVector = 1;
 #endif
 

script/run.sh

@@ -12,13 +12,16 @@
 #export OLC_DEBUG_HIP_DUMP=1
 #export OLC_DEBUG_SAVE_TEMP_DIR=1
 
- make -j conv_fwd_driver_offline
- make -j conv_bwd_driver_offline
- make -j conv_fwd_driver_online
-
 #rm -rf /root/_hip_binary_kernels_/
 #rm -rf /tmp/olCompile*
 
+#make -j conv_fwd_driver_offline
+#make -j conv_bwd_driver_offline
+#make -j conv_wrw_driver_offline
+#make -j conv_fwd_driver_online
+
+ make -j gemm_driver_offline
+
 LAYOUT=$1
 ALGO=$2
 VERIFY=$3
@@ -30,7 +33,7 @@ REPEAT=$6
 #./host/driver_offline/conv_fwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  128  128  192 3 3  71   71     2 2       1 1      1 1       1 1
 #./host/driver_offline/conv_fwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  128  256  192 3 3  71   71     2 2       1 1      1 1       1 1
 #./host/driver_offline/conv_fwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  128  256 1024 1 7  17   17     1 1       1 1      0 3       0 3
- ./host/driver_offline/conv_fwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  256  256  256 3 3  14   14     1 1       1 1      1 1       1 1
+#./host/driver_offline/conv_fwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  256  256  256 3 3  14   14     1 1       1 1      1 1       1 1
 #./host/driver_offline/conv_fwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  128  128  128 3 3  14   14     1 1       1 1      1 1       1 1
 #./host/driver_offline/conv_fwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  256  512  512 3 3   7    7     1 1       1 1      1 1       1 1
 
@@ -44,4 +47,12 @@ REPEAT=$6
 
 #./host/driver_offline/conv_bwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  256  256  256 3 3  14   14     1 1       1 1      1 1       1 1
 
-#./host/driver_online/conv_fwd_driver_online    $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  128  256  192 3 3  71   71     2 2       1 1      1 1       1 1
+#./host/driver_offline/conv_wrw_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  128  256  128 3 3  14   14     1 1       1 1      1 1       1 1
+
+#./host/driver_online/conv_fwd_driver_online    $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  128  128  192 3 3  71   71     2 2       1 1      1 1       1 1
+
+################################################ layout  algo  verify  init  log  repeat  M___ N___ K___
+#./host/driver_offline/gemm_driver_offline      $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT   960 1024 1024
+#./host/driver_offline/gemm_driver_offline      $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  1920 2048 2048
+ ./host/driver_offline/gemm_driver_offline      $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  3840 4096 4096
+#./host/driver_offline/gemm_driver_offline      $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  7680 8192 8192