some gridwise gemm write to C matrix

composable_kernel/include/tensor_operation/blockwise_gemm_xdlops.hpp

@@ -131,6 +131,17 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
         return make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, I1, M0, M1, M2, N));
     }
 
+    __host__ __device__ static constexpr auto GetCGM0N0M1N1M2M3M4N2ThreadDescriptor()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, I1, I1, I1, I1, M0, M1, M2, N));
+    }
     __host__ __device__ static constexpr auto GetCM0N0M1N1M2M3M4N2BlockDescriptor()
     {
         constexpr auto c_m0_n0_m1_n1_m2_n2_block_desc =
@@ -144,6 +155,20 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
         return xdlops_gemm.MakeCM0N0M1N1M2M3M4N2Descriptor(c_m0_n0_m1_n1_m2_n2_block_desc);
     }
 
+    __host__ __device__ static constexpr auto GetCGM0N0M1N1M2M3M4N2BlockDescriptor()
+    {
+        constexpr auto c_m0_n0_m1_n1_m2_n2_block_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<1>{},
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCM0N0M1N1M2M3M4N2Descriptor(c_m0_n0_m1_n1_m2_n2_block_desc);
+    }
+
     template <typename CMNGridDesc>
     __host__ __device__ static constexpr auto
     MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CMNGridDesc& c_m_n_grid_desc)

host/driver_offline/include/device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwgc_gkyxc_nhwgk.hpp

@@ -255,24 +255,24 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwgc_gkyxc_nhwgk(
                               Sequence<0, 0, 0, 0, 0, 0, 0>{})); // 3-: GemmK1
 
     constexpr auto out_gemmg_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
-        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0+: M0
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1+: M0
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2+: N0
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3+: M1
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4+: N1
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5+: M2
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6+: M3
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 7+: M4
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),  // 8+: N2
-                   make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0-: M0
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1-: M0
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2-: N0
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3-: M1
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4-: N1
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5-: M2
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6-: M3
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 7-: M4
-                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 8-: N2
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0+: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1+: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2+: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3+: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4+: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5+: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6+: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 7+: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),  // 8+: N2
+                   make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0-: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1-: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2-: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3-: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4-: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5-: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6-: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 7-: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 8-: N2
 
     constexpr auto in_gemmg_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
         Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{};