use array

example/15_grouped_gemm/grouped_gemm_bias_xdl_fp16.cpp

@@ -82,7 +82,7 @@ int main(int argc, char* argv[])
     // GEMM shape
     std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
     std::vector<const void*> p_a, p_b;
-    std::vector<std::vector<const void*>> p_ds;
+    std::vector<std::array<const void*, 1>> p_ds;
     std::vector<void*> p_c;
 
     gemm_descs.reserve(group_count);

example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp

@@ -200,7 +200,7 @@ int main(int argc, char* argv[])
     auto gemm    = DeviceGemmInstance{};
     auto invoker = gemm.MakeInvoker();
 
-    std::vector<std::vector<const void*>> p_Ds = {};
+    std::vector<std::array<const void*, 0>> p_Ds = {};
 
     // do GEMM
     auto argument = gemm.MakeArgument(

include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp

@@ -28,10 +28,12 @@ template <typename ALayout,
           typename CElementwiseOperation>
 struct DeviceGroupedGemm : public BaseOperator
 {
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
     virtual std::unique_ptr<BaseArgument>
     MakeArgumentPointer(std::vector<const void*>& p_a,
                         std::vector<const void*>& p_b,
-                        std::vector<std::vector<const void*>>& p_ds,
+                        std::vector<std::array<const void*, NumDTensor>>& p_ds,
                         std::vector<void*>& p_e,
                         std::vector<GemmDesc>& gemm_desc,
                         AElementwiseOperation a_element_op,

include/ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp

@@ -532,7 +532,7 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
     {
         Argument(std::vector<const void*>& p_As,
                  std::vector<const void*>& p_Bs,
-                 std::vector<std::vector<const void*>>& p_Ds,
+                 std::vector<std::array<const void*, NumDTensor>>& p_Ds,
                  std::vector<void*>& p_Es,
                  std::vector<GemmDesc>& gemm_descs,
                  AElementwiseOperation a_element_op,
@@ -755,7 +755,7 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
 
     static auto MakeArgument(std::vector<const void*>& p_As,
                              std::vector<const void*>& p_Bs,
-                             std::vector<std::vector<const void*>>& p_Ds,
+                             std::vector<std::array<const void*, NumDTensor>>& p_Ds,
                              std::vector<void*>& p_Es,
                              std::vector<GemmDesc> gemm_descs,
                              AElementwiseOperation a_element_op,
@@ -769,14 +769,15 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
     static auto MakeInvoker() { return Invoker{}; }
 
     // polymorphic
-    std::unique_ptr<BaseArgument> MakeArgumentPointer(std::vector<const void*>& p_As,
-                                                      std::vector<const void*>& p_Bs,
-                                                      std::vector<std::vector<const void*>>& p_Ds,
-                                                      std::vector<void*>& p_Es,
-                                                      std::vector<GemmDesc>& gemm_descs,
-                                                      AElementwiseOperation a_element_op,
-                                                      BElementwiseOperation b_element_op,
-                                                      CDEElementwiseOperation c_element_op) override
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_As,
+                        std::vector<const void*>& p_Bs,
+                        std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                        std::vector<void*>& p_Es,
+                        std::vector<GemmDesc>& gemm_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation c_element_op) override
     {
         return std::make_unique<Argument>(
             p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op);

profiler/include/profile_grouped_gemm_impl.hpp

@@ -175,7 +175,7 @@ bool profile_grouped_gemm_impl(int do_verification,
     float best_tflops     = 0;
     float best_gb_per_sec = 0;
 
-    auto p_ds = std::vector<std::vector<const void*>>{};
+    auto p_ds = std::vector<std::array<const void*, 0>>{};
 
     // profile device GEMM instances
     for(auto& gemm_ptr : op_ptrs)