add timer

host/driver_offline/include/device_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw.hpp

@@ -25,7 +25,7 @@ void device_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw(
     const Tensor<TInWei>& in_n_c_hi_wi,
     const Tensor<TInWei>& wei_k_c_y_x,
     Tensor<TOut>& out_n_k_ho_wo,
-    ck::index_t /* nrepeat */)
+    ck::index_t nrepeat)
 {
     using namespace ck;
 
@@ -102,15 +102,15 @@ void device_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw(
     constexpr index_t BlockSize = 64;
 
     constexpr index_t KPerBlock  = 16;
-    constexpr index_t HoPerBlock = 8;
+    constexpr index_t HoPerBlock = 32;
     constexpr index_t WoPerBlock = 8;
 
     constexpr index_t E1        = 16;
     constexpr index_t EPerBlock = 8;
 
     constexpr index_t KPerThread  = KPerBlock;
-    constexpr index_t HoPerThread = 1;
-    constexpr index_t WoPerThread = 1;
+    constexpr index_t HoPerThread = 2;
+    constexpr index_t WoPerThread = 2;
     constexpr index_t EPerThread  = EPerBlock;
 
     using ABlockTransferThreadSliceLengths_E0_E1_K   = Sequence<1, 4, 1>;
@@ -146,16 +146,18 @@ void device_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw(
             BThreadTransferSrcScalarPerVector_E,
             CThreadTransferDstScalarPerVector_K>{};
 
-    conv_driver.Run(wei_k_c0_y_x_c1_desc,
-                    in_n_c0_hi_wi_c1_desc,
-                    out_n_k0_ho_wo_k1_desc,
-                    conv_strides,
-                    conv_dilations,
-                    in_left_pads,
-                    in_right_pads,
-                    static_cast<TInWei*>(wei_k_c0_y_x_c1_device_buf.GetDeviceBuffer()),
-                    static_cast<TInWei*>(in_n_c0_hi_wi_c1_device_buf.GetDeviceBuffer()),
-                    static_cast<TOut*>(out_n_k0_ho_wo_k1_device_buf.GetDeviceBuffer()));
+    const auto ave_time =
+        conv_driver.Run(wei_k_c0_y_x_c1_desc,
+                        in_n_c0_hi_wi_c1_desc,
+                        out_n_k0_ho_wo_k1_desc,
+                        conv_strides,
+                        conv_dilations,
+                        in_left_pads,
+                        in_right_pads,
+                        static_cast<TInWei*>(wei_k_c0_y_x_c1_device_buf.GetDeviceBuffer()),
+                        static_cast<TInWei*>(in_n_c0_hi_wi_c1_device_buf.GetDeviceBuffer()),
+                        static_cast<TOut*>(out_n_k0_ho_wo_k1_device_buf.GetDeviceBuffer()),
+                        nrepeat);
 
     out_n_k0_ho_wo_k1_device_buf.FromDevice(out_n_k0_ho_wo_k1.mData.data());
 
@@ -164,4 +166,11 @@ void device_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw(
     };
 
     make_ParallelTensorFunctor(f_nk0hwk1_to_nkhw, N, K, Ho, Wo)();
+
+    {
+        float perf = static_cast<float>(std::size_t(2) * N * K * Ho * Wo * C * Y * X) /
+                     (std::size_t(1000) * 1000 * 1000) / ave_time;
+
+        std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
+    }
 }

host/driver_offline/include/driver_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw_outpad.hpp

@@ -43,12 +43,11 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
                        const InRightPads& in_right_pads,
                        const FloatAB* __restrict__ p_a_grid,
                        const FloatAB* __restrict__ p_b_grid,
-                       FloatC* __restrict__ p_c_grid) const
+                       FloatC* __restrict__ p_c_grid,
+                       const int nrepeat) const
     {
         using namespace ck;
 
-        const int nrepeat = 100;
-
         constexpr auto I0 = Number<0>{};
         constexpr auto I1 = Number<1>{};
         constexpr auto I2 = Number<2>{};
@@ -179,28 +178,28 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
         constexpr auto a_e0_e1_k_global_move_slice_window_step_hack = Sequence<0, 0, 0, 0, 0>{};
 
         constexpr auto b_e0_e1_n_ho_wo_global_step_hacks = make_tuple(
-            make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
-                       Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
+            make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0>{},
+                       Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0>{},
                        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
                        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
                        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),
-            make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
-                       Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
+            make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0>{},
+                       Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0>{},
                        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
                        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
                        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}));
 
         constexpr auto b_e0_e1_n_ho_wo_global_move_slice_window_step_hack =
-            Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{};
+            Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0>{};
 
         // hack to control index calculation when iterating over c_m0_m1_n0_n1_global tensor
         // hack for NKHW format
         constexpr auto c_k_n_ho_wo_global_tensor_step_hacks =
-            make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
+            make_tuple(make_tuple(Sequence<0, 1, 0, 0, 0>{},
                                   Sequence<0, 0, 0, 0, 0>{},
                                   Sequence<0, 0, 0, 0, 0>{},
                                   Sequence<0, 0, 0, 0, 0>{}),
-                       make_tuple(Sequence<0, 0, 0, 0, 0>{},
+                       make_tuple(Sequence<0, 2, 0, 0, 0>{},
                                   Sequence<0, 0, 0, 0, 0>{},
                                   Sequence<0, 0, 0, 0, 0>{},
                                   Sequence<0, 0, 0, 0, 0>{}));

script/run.sh

@@ -51,8 +51,7 @@ REPEAT=$6
 
 #./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
 
-#./host/driver_offline/conv_fwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  1  16  16 3 3  8   8     1 1       1 1      1 1       1 1
-./host/driver_offline/conv_fwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  1  16  16 3 3  8   8     1 1       1 1      0 0       0 0
+./host/driver_offline/conv_fwd_driver_offline  $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT  1  16  16 3 3  1080   1920     1 1       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