format

profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp

@@ -187,8 +187,21 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
     // profile device GEMM instances
     for(auto& op_ptr : op_ptrs)
     {
-        auto argument_ptr =
-            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+
+        std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 19, 20, 32, 38};
+
+        if(KBatch > 0)
+        {
+            kbatch_list = {KBatch};
+        }
+
+        for(std::size_t i = 0; i < kbatch_list.size(); i++)
+        {
+
+            auto kbatch_curr = kbatch_list[i];
+
+            auto argument_ptr = op_ptr->MakeArgumentPointer(
+                static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
                 static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
                 std::array<const void*, 2>{d0_device_buf.GetDeviceBuffer(),
                                            d1_device_buf.GetDeviceBuffer()},
@@ -213,7 +226,8 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
                 // re-init C to zero before profiling next kernel
                 c_device_buf.SetZero();
 
-            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false, 0, n_warmup, n_iter});
+                invoker_ptr->Run(argument_ptr.get(),
+                                 StreamConfig{nullptr, false, 0, n_warmup, n_iter});
 
                 if(do_verification)
                 {
@@ -225,31 +239,37 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
                     {
                         LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
                         LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
-                    LogRangeAsType<float>(std::cout << "c_host  : ", e_m_n_host_result.mData, ",")
+                        LogRangeAsType<float>(
+                            std::cout << "c_host  : ", e_m_n_host_result.mData, ",")
                             << std::endl;
-                    LogRangeAsType<float>(std::cout << "c_device: ", e_m_n_device_result.mData, ",")
+                        LogRangeAsType<float>(
+                            std::cout << "c_device: ", e_m_n_device_result.mData, ",")
                             << std::endl;
                     }
                 }
 
                 std::string op_name = op_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(
-                argument_ptr.get(),
-                StreamConfig{
-                    nullptr, time_kernel, 0, n_warmup, n_iter, rotating_count > 1, rotating_count});
+                float ave_time = invoker_ptr->Run(argument_ptr.get(),
+                                                  StreamConfig{nullptr,
+                                                               time_kernel,
+                                                               0,
+                                                               n_warmup,
+                                                               n_iter,
+                                                               rotating_count > 1,
+                                                               rotating_count});
 
                 std::size_t flop = std::size_t(2) * M * N * K;
 
-            std::size_t num_btype =
-                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+                std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                        sizeof(EDataType) * M * N;
 
                 float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
 
                 float gb_per_sec = num_btype / 1.E6 / ave_time;
 
-            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
-                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
+                          << " TFlops, " << gb_per_sec << " GB/s, " << op_name << std::endl;
 
 #if defined CK_ENABLE_FP8
                 // set softer tolerances for fp8
@@ -280,7 +300,9 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
             }
             else
             {
-            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+                std::cout << op_ptr->GetTypeString() << " does not support this problem"
+                          << std::endl;
+            }
         }
     }