create seperate fusion fun

host/driver_offline/include/driver_convolution_add_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp

@@ -336,12 +336,7 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0
             decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks),
             decltype(d_k0_k1_n_h0_h1_h2x2_w0_w1_w2x2_global_tensor_step_hacks),
             decltype(a_e0_e1_k_e2_global_move_slice_window_step_hack),
-            decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_move_slice_window_step_hack),
-            activ_type,
-            1, // bias_type
-            0, // out_type
-            1  // add_type
-            >;
+            decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_move_slice_window_step_hack)>;
 
         const auto a_e0_e1_k0_k1_e2_grid_desc =
             GridwiseGemm::MakeAE0E1K0K1E2GridDescriptor(a_e0_e1_k_e2_grid_desc);
@@ -350,7 +345,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0
         const auto c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc =
             GridwiseGemm::MakeCK0K1NH0H1H2W0W1W2GridDescriptor(c_k_n_hop_wop_grid_desc);
         const auto d_k0_k1_n_h0_h1_h2x2_w0_w1_w2x2_grid_desc =
-            GridwiseGemm::MakeDK0K1NH0H1HxW0W1WxGridDescriptor(d_k_n_hopx2_wopx2_grid_desc);
+            GridwiseGemm::MakeDK0K1NH0H1HxW0W1WxGridDescriptorResizeAdd(
+                d_k_n_hopx2_wopx2_grid_desc);
 
         using AGridDesc_E0_E1_K0_K1_E2 = decltype(a_e0_e1_k0_k1_e2_grid_desc);
         using BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2 =

host/driver_offline/include/driver_convolution_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp

@@ -301,12 +301,7 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0
             decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks),
             decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks),
             decltype(a_e0_e1_k_e2_global_move_slice_window_step_hack),
-            decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_move_slice_window_step_hack),
-            activ_type,
-            1, // bias_type
-            1, // out_type
-            0  // add_type
-            >;
+            decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_move_slice_window_step_hack)>;
 
         const auto a_e0_e1_k0_k1_e2_grid_desc =
             GridwiseGemm::MakeAE0E1K0K1E2GridDescriptor(a_e0_e1_k_e2_grid_desc);

host/driver_offline/include/driver_convolution_maxpool_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp

@@ -340,12 +340,7 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0
             decltype(c_k0_k1_n_h0_h1_h2_w0_w1_w2_global_tensor_step_hacks),
             decltype(d_k0_k1_n_h0_h1_hx_w0_w1_wx_global_tensor_step_hacks),
             decltype(a_e0_e1_k_e2_global_move_slice_window_step_hack),
-            decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_move_slice_window_step_hack),
-            activ_type,
-            1, // bias_type
-            1, // out_type
-            2  // add_type
-            >;
+            decltype(b_e0_e1_n_h0_h1_h2_w0_w1_w2_e2_global_move_slice_window_step_hack)>;
 
         const auto a_e0_e1_k0_k1_e2_grid_desc =
             GridwiseGemm::MakeAE0E1K0K1E2GridDescriptor(a_e0_e1_k_e2_grid_desc);
@@ -354,7 +349,7 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0
         const auto c_k0_k1_n_h0_h1_h2_w0_w1_w2_grid_desc =
             GridwiseGemm::MakeCK0K1NH0H1H2W0W1W2GridDescriptor(c_k_n_hop_wop_grid_desc);
         const auto d_k0_k1_n_h0_h1_hx_w0_w1_wx_grid_desc =
-            GridwiseGemm::MakeDK0K1NH0H1HxW0W1WxGridDescriptor(d_k_n_hx_wx_grid_desc);
+            GridwiseGemm::MakeDK0K1NH0H1HxW0W1WxGridDescriptorMaxPool(d_k_n_hx_wx_grid_desc);
 
         using AGridDesc_E0_E1_K0_K1_E2 = decltype(a_e0_e1_k0_k1_e2_grid_desc);
         using BGridDesc_E0_E1_N_H0_H1_H2_W0_W1_W2_E2 =

host/driver_offline/src/conv_fwd_driver_offline_nchwc.cpp

@@ -92,8 +92,8 @@ int main(int argc, char* argv[])
     const bool do_log          = std::stoi(argv[4]);
     const int nrepeat          = std::stoi(argv[5]);
 
-    constexpr ck::ActivTypeEnum_t activ_type = ActivTypeEnum_t::Sigmoid;
-    // constexpr ck::ActivTypeEnum_t activ_type = ActivTypeEnum_t::LeakyRelu;
+    // constexpr ck::ActivTypeEnum_t activ_type = ActivTypeEnum_t::Sigmoid;
+    constexpr ck::ActivTypeEnum_t activ_type = ActivTypeEnum_t::LeakyRelu;
 
 #if 0
     constexpr auto N              = Number<1>{};