Add host API (#220)

* Add host API

* manually rebase on develop

* clean

* manually rebase on develop

* exclude tests from all target

* address review comments

* update client app name

* fix missing lib name

* clang-format update

* refactor

* refactor

* refactor

* refactor

* refactor

* fix test issue

* refactor

* refactor

* refactor

* upate cmake and readme
Co-authored-by: Chao Liu <chao.liu2@amd.com>

profiler/include/profile_gemm_bias_relu_add_impl.hpp

@@ -48,7 +48,7 @@ template <typename ADataType,
 void profile_gemm_bias_relu_add_impl(int do_verification,
                                      int init_method,
                                      bool do_log,
-                                     int nrepeat,
+                                     bool time_kernel,
                                      int M,
                                      int N,
                                      int K,
@@ -232,7 +232,8 @@ void profile_gemm_bias_relu_add_impl(int do_verification,
         {
             std::string gemm_name = gemm_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * M * N * K;
 

profiler/include/profile_gemm_bias_relu_impl.hpp

@@ -48,7 +48,7 @@ template <typename ADataType,
 void profile_gemm_bias_relu_impl(int do_verification,
                                  int init_method,
                                  bool do_log,
-                                 int nrepeat,
+                                 bool time_kernel,
                                  int M,
                                  int N,
                                  int K,
@@ -212,7 +212,8 @@ void profile_gemm_bias_relu_impl(int do_verification,
         {
             std::string gemm_name = gemm_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * M * N * K;
 

profiler/include/profile_gemm_impl.hpp

@@ -91,7 +91,7 @@ template <typename ADataType,
 void profile_gemm_impl(int do_verification,
                        int init_method,
                        bool do_log,
-                       int nrepeat,
+                       bool time_kernel,
                        int M,
                        int N,
                        int K,
@@ -416,7 +416,8 @@ void profile_gemm_impl(int do_verification,
 
             std::string gemm_name = gemm_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * M * N * K;
 

profiler/include/profile_gemm_reduce_impl.hpp

@@ -52,7 +52,7 @@ template <typename ADataType,
 bool profile_gemm_reduce_impl(int do_verification,
                               int init_method,
                               bool do_log,
-                              int nrepeat,
+                              bool time_kernel,
                               int M,
                               int N,
                               int K,
@@ -243,36 +243,18 @@ bool profile_gemm_reduce_impl(int do_verification,
 
         if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
         {
-            // warm up
-            invoker_ptr->Run(argument_ptr.get());
+            // init DO, D1 to 0
+            d0_device_buf.SetZero();
+            d1_device_buf.SetZero();
 
-            // timing
-            float total_time = 0;
-
-            for(int i = 0; i < nrepeat; ++i)
-            {
-                // init DO, D1 to 0
-                d0_device_buf.SetZero();
-                d1_device_buf.SetZero();
-
-                KernelTimer timer;
-
-                timer.Start();
-
-                invoker_ptr->Run(argument_ptr.get());
-
-                timer.End();
-
-                total_time += timer.GetElapsedTime();
-            }
-
-            float ave_time = total_time / nrepeat;
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::string gemm_name = gemm_ptr->GetTypeString();
 
             std::size_t flop = std::size_t(2) * M * N * K;
 
-            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
                                     sizeof(CDataType) * M * N + sizeof(CDataType) * N;
 
             float tflops = static_cast<float>(flop) / 1.E9 / ave_time;

profiler/include/profile_grouped_gemm_impl.hpp

@@ -49,7 +49,7 @@ template <typename ADataType,
 void profile_grouped_gemm_impl(int do_verification,
                                int init_method,
                                bool do_log,
-                               int nrepeat,
+                               bool time_kernel,
                                const std::vector<int>& Ms,
                                const std::vector<int>& Ns,
                                const std::vector<int>& Ks,
@@ -231,7 +231,8 @@ void profile_grouped_gemm_impl(int do_verification,
         {
             std::string gemm_name = gemm_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = 0, num_btype = 0;
             for(std::size_t i = 0; i < gemm_shapes.size(); i++)

profiler/include/profile_reduce_impl.hpp

@@ -157,7 +157,7 @@ void profile_reduce_impl_impl(bool do_verification,
                               int init_method,
                               bool do_log,
                               bool do_dumpout,
-                              int nrepeat,
+                              bool time_kernel,
                               const std::vector<size_t>& inLengths,
                               const std::vector<int>& reduceDims,
                               float alpha,
@@ -430,7 +430,8 @@ void profile_reduce_impl_impl(bool do_verification,
 
             auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
 
-            float avg_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t num_bytes =
                 invariant_total_length * reduce_total_length * sizeof(InDataType) +
@@ -516,7 +517,8 @@ void profile_reduce_impl_impl(bool do_verification,
 
             auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
 
-            float avg_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t num_bytes =
                 invariant_total_length * reduce_total_length * sizeof(InDataType) +
@@ -554,7 +556,8 @@ void profile_reduce_impl_impl(bool do_verification,
 
                 auto invoker2_ptr = reduce2_ptr->MakeInvokerPointer();
 
-                float avg_time_2 = invoker2_ptr->Run(argument2_ptr.get(), nrepeat);
+                float avg_time_2 =
+                    invoker2_ptr->Run(argument2_ptr.get(), StreamConfig{nullptr, time_kernel});
 
                 std::size_t num_bytes_2 =
                     static_cast<size_t>(inLengths2[0]) * inLengths2[1] * sizeof(AccDataType);
@@ -625,7 +628,7 @@ void profile_reduce_impl(bool do_verification,
                          int init_method,
                          bool do_log,
                          bool do_dumpout,
-                         int nrepeat,
+                         bool time_kernel,
                          const std::vector<size_t>& inLengths,
                          const std::vector<int>& reduceDims,
                          ReduceTensorOp ReduceOpId,
@@ -663,7 +666,7 @@ void profile_reduce_impl(bool do_verification,
             init_method,
             do_log,
             do_dumpout,
-            nrepeat,
+            time_kernel,
             inLengths,
             reduceDims,
             alpha,

profiler/src/profile_batched_gemm.cpp

@@ -48,8 +48,8 @@ int profile_batched_gemm(int argc, char* argv[])
         printf("                     3: A[g, k, m] * B[g, n, k] = C[g, m, n])\n");
         printf("arg4: verification (0: no; 1: yes)\n");
         printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
-        printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg7: run kernel # of times (>1)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
         printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
         exit(1);
     }
@@ -59,7 +59,7 @@ int profile_batched_gemm(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[4]);
     const int init_method      = std::stoi(argv[5]);
     const bool do_log          = std::stoi(argv[6]);
-    const int nrepeat          = std::stoi(argv[7]);
+    const bool time_kernel     = std::stoi(argv[7]);
 
     const int M = std::stoi(argv[8]);
     const int N = std::stoi(argv[9]);
@@ -82,7 +82,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -102,7 +102,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -122,7 +122,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -142,7 +142,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -162,7 +162,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -182,7 +182,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -202,7 +202,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -222,7 +222,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -242,7 +242,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -262,7 +262,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -282,7 +282,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -302,7 +302,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -322,7 +322,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -342,7 +342,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -362,7 +362,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -382,7 +382,7 @@ int profile_batched_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,

profiler/src/profile_batched_gemm_reduce.cpp

@@ -33,8 +33,8 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
         printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
         printf("arg4: verification (0: no; 1: yes)\n");
         printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
-        printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg7: run kernel # of times (>1)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
         printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
         printf("arg15: split k into  mulitiple batch\n");
         exit(1);
@@ -45,7 +45,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[4]);
     const int init_method      = std::stoi(argv[5]);
     const bool do_log          = std::stoi(argv[6]);
-    const int nrepeat          = std::stoi(argv[7]);
+    const bool time_kernel     = std::stoi(argv[7]);
 
     const int M = std::stoi(argv[8]);
     const int N = std::stoi(argv[9]);
@@ -69,7 +69,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -91,7 +91,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -113,7 +113,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -135,7 +135,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,

profiler/src/profile_conv_bwd_data.cpp

@@ -44,7 +44,7 @@ int profile_conv_bwd_data(int argc, char* argv[])
         printf("arg6: verification (0: no; 1: yes)\n");
         printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
         printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg9: run kernel # of times (>1)\n");
+        printf("arg9: time kernel (0=n0, 1=yes)\n");
         printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
                "RightPx\n");
         exit(1);
@@ -57,7 +57,7 @@ int profile_conv_bwd_data(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[6]);
     const int init_method      = std::stoi(argv[7]);
     const bool do_log          = std::stoi(argv[8]);
-    const int nrepeat          = std::stoi(argv[9]);
+    const bool time_kernel     = std::stoi(argv[9]);
 
     const ck::index_t N  = std::stoi(argv[10]);
     const ck::index_t K  = std::stoi(argv[11]);
@@ -96,7 +96,7 @@ int profile_conv_bwd_data(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            StreamControl{nullptr, time_kernel},
             N,
             K,
             C,
@@ -122,7 +122,7 @@ int profile_conv_bwd_data(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            StreamControl{nullptr, time_kernel},
             N,
             K,
             C,
@@ -148,7 +148,7 @@ int profile_conv_bwd_data(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            StreamControl{nullptr, time_kernel},
             N,
             K,
             C,
@@ -174,7 +174,7 @@ int profile_conv_bwd_data(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            StreamControl{nullptr, time_kernel},
             N,
             K,
             C,

profiler/src/profile_conv_bwd_weight.cpp

@@ -58,7 +58,7 @@ int profile_conv_bwd_weight(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[6]);
     const int init_method      = std::stoi(argv[7]);
     const bool do_log          = std::stoi(argv[8]);
-    const int nrepeat          = std::stoi(argv[9]);
+    const bool time_kernel     = std::stoi(argv[9]);
 
     const ck::index_t N  = std::stoi(argv[10]);
     const ck::index_t K  = std::stoi(argv[11]);
@@ -98,7 +98,7 @@ int profile_conv_bwd_weight(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             N,
             K,
             C,
@@ -124,7 +124,7 @@ int profile_conv_bwd_weight(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             N,
             K,
             C,

profiler/src/profile_conv_fwd_bias_relu.cpp

@@ -42,7 +42,7 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
         printf("arg6: verification (0: no; 1: yes)\n");
         printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
         printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg9: run kernel # of times (>1)\n");
+        printf("arg9: time kernel (0=n0, 1=yes)\n");
         printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
                "RightPx\n");
         exit(1);
@@ -55,7 +55,7 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[6]);
     const int init_method      = std::stoi(argv[7]);
     const bool do_log          = std::stoi(argv[8]);
-    const int nrepeat          = std::stoi(argv[9]);
+    const bool time_kernel     = std::stoi(argv[9]);
 
     const ck::index_t N  = std::stoi(argv[10]);
     const ck::index_t K  = std::stoi(argv[11]);
@@ -93,7 +93,7 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             N,
             K,
             C,

profiler/src/profile_conv_fwd_bias_relu_add.cpp

@@ -43,7 +43,7 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
         printf("arg6: verification (0: no; 1: yes)\n");
         printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
         printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg9: run kernel # of times (>1)\n");
+        printf("arg9: time kernel (0=n0, 1=yes)\n");
         printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
                "RightPx\n");
         exit(1);
@@ -56,7 +56,7 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[6]);
     const int init_method      = std::stoi(argv[7]);
     const bool do_log          = std::stoi(argv[8]);
-    const int nrepeat          = std::stoi(argv[9]);
+    const bool time_kernel     = std::stoi(argv[9]);
 
     const ck::index_t N  = std::stoi(argv[10]);
     const ck::index_t K  = std::stoi(argv[11]);
@@ -94,7 +94,7 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             N,
             K,
             C,

profiler/src/profile_conv_fwd_bias_relu_atomic_add.cpp

@@ -43,7 +43,7 @@ int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[])
         printf("arg6: verification (0: no; 1: yes)\n");
         printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
         printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg9: run kernel # of times (>1)\n");
+        printf("arg9: time kernel (0=n0, 1=yes)\n");
         printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
                "RightPx\n");
         exit(1);
@@ -56,7 +56,7 @@ int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[6]);
     const int init_method      = std::stoi(argv[7]);
     const bool do_log          = std::stoi(argv[8]);
-    const int nrepeat          = std::stoi(argv[9]);
+    const bool time_kernel     = std::stoi(argv[9]);
 
     const ck::index_t N  = std::stoi(argv[10]);
     const ck::index_t K  = std::stoi(argv[11]);
@@ -95,7 +95,7 @@ int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             N,
             K,
             C,

profiler/src/profile_convnd_bwd_data.cpp

@@ -95,7 +95,7 @@ int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
         printf("arg6: verification (0: no; 1: yes)\n");
         printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
         printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg9: run kernel # of times (>1)\n");
+        printf("arg9: time kernel (0=n0, 1=yes)\n");
         printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
                "RightPx\n");
         return 1;
@@ -108,7 +108,7 @@ int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
     const bool do_verification = std::stoi(argv[6]);
     const int init_method      = std::stoi(argv[7]);
     const bool do_log          = std::stoi(argv[8]);
-    const int nrepeat          = std::stoi(argv[9]);
+    const bool time_kernel     = std::stoi(argv[9]);
 
     ck::utils::conv::ConvParams params = parse_conv_params(num_dim_spatial, argv, preParams);
 
@@ -132,7 +132,7 @@ int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
                 do_verification,
                 init_method,
                 do_log,
-                nrepeat,
+                time_kernel,
                 params.N_,
                 params.K_,
                 params.C_,
@@ -157,7 +157,7 @@ int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
                 do_verification,
                 init_method,
                 do_log,
-                nrepeat,
+                time_kernel,
                 params.N_,
                 params.K_,
                 params.C_,
@@ -182,7 +182,7 @@ int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
                 do_verification,
                 init_method,
                 do_log,
-                nrepeat,
+                time_kernel,
                 params.N_,
                 params.K_,
                 params.C_,

profiler/src/profile_convnd_fwd.cpp

@@ -119,7 +119,7 @@ template <int NDim,
 void profile_convnd_instances_impl(const ck::utils::conv::ConvParams& params,
                                    bool do_verification,
                                    bool do_log,
-                                   int nrepeat,
+                                   bool time_kernel,
                                    int init_method,
                                    ConvLayouts)
 {
@@ -185,7 +185,7 @@ void profile_convnd_instances_impl(const ck::utils::conv::ConvParams& params,
                                                                          reference_conv_fwd_fun);
     auto best_conf = run_engine.Profile(
         conv::ConvolutionFwdInstances<InDataType, WeiDataType, OutDataType>::template Get<NDim>(),
-        nrepeat,
+        time_kernel,
         do_verification,
         do_log);
 
@@ -201,7 +201,7 @@ void profile_convnd_instances(ConvDataType data_type,
                               const ck::utils::conv::ConvParams& params,
                               bool do_verification,
                               bool do_log,
-                              int nrepeat,
+                              bool time_kernel,
                               int init_method)
 {
     switch(data_layout)
@@ -214,7 +214,7 @@ void profile_convnd_instances(ConvDataType data_type,
                 params,
                 do_verification,
                 do_log,
-                nrepeat,
+                time_kernel,
                 init_method,
                 ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
             break;
@@ -223,7 +223,7 @@ void profile_convnd_instances(ConvDataType data_type,
                 params,
                 do_verification,
                 do_log,
-                nrepeat,
+                time_kernel,
                 init_method,
                 ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
             break;
@@ -232,7 +232,7 @@ void profile_convnd_instances(ConvDataType data_type,
                 params,
                 do_verification,
                 do_log,
-                nrepeat,
+                time_kernel,
                 init_method,
                 ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
             break;
@@ -241,7 +241,7 @@ void profile_convnd_instances(ConvDataType data_type,
                 params,
                 do_verification,
                 do_log,
-                nrepeat,
+                time_kernel,
                 init_method,
                 ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
             break;
@@ -256,7 +256,7 @@ void profile_convnd_instances(ConvDataType data_type,
                 params,
                 do_verification,
                 do_log,
-                nrepeat,
+                time_kernel,
                 init_method,
                 ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
             break;
@@ -265,7 +265,7 @@ void profile_convnd_instances(ConvDataType data_type,
                 params,
                 do_verification,
                 do_log,
-                nrepeat,
+                time_kernel,
                 init_method,
                 ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
             break;
@@ -274,7 +274,7 @@ void profile_convnd_instances(ConvDataType data_type,
                 params,
                 do_verification,
                 do_log,
-                nrepeat,
+                time_kernel,
                 init_method,
                 ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
             break;
@@ -283,7 +283,7 @@ void profile_convnd_instances(ConvDataType data_type,
                 params,
                 do_verification,
                 do_log,
-                nrepeat,
+                time_kernel,
                 init_method,
                 ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
             break;
@@ -304,7 +304,7 @@ int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
     bool do_verification{true};
     int init_method{2};
     bool do_log{false};
-    int nrepeat{100};
+    bool time_kernel{false};
     int num_dim_spatial{2};
     ConvParams params;
 
@@ -318,7 +318,7 @@ int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
         do_verification = std::stoi(argv[4]);
         init_method     = std::stoi(argv[5]);
         do_log          = std::stoi(argv[6]);
-        nrepeat         = std::stoi(argv[7]);
+        time_kernel     = std::stoi(argv[7]);
         num_dim_spatial = std::stoi(argv[8]);
     }
     if(argc >= 10)
@@ -332,15 +332,15 @@ int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
     {
     case 1:
         profile_convnd_instances<1>(
-            data_type, data_layout, params, do_verification, do_log, nrepeat, init_method);
+            data_type, data_layout, params, do_verification, do_log, time_kernel, init_method);
         break;
     case 2:
         profile_convnd_instances<2>(
-            data_type, data_layout, params, do_verification, do_log, nrepeat, init_method);
+            data_type, data_layout, params, do_verification, do_log, time_kernel, init_method);
         break;
     case 3:
         profile_convnd_instances<3>(
-            data_type, data_layout, params, do_verification, do_log, nrepeat, init_method);
+            data_type, data_layout, params, do_verification, do_log, time_kernel, init_method);
         break;
     default:
         throw std::runtime_error("profile_conv_fwd: unsupported num_dim_spatial value: " +

profiler/src/profile_gemm.cpp

@@ -38,8 +38,8 @@ int profile_gemm(int argc, char* argv[])
         printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
         printf("arg4: verification (0: no; 1: yes)\n");
         printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
-        printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg7: run kernel # of times (>1)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
         printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
         printf("arg14: split k into  mulitiple batch\n");
         exit(1);
@@ -50,7 +50,7 @@ int profile_gemm(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[4]);
     const int init_method      = std::stoi(argv[5]);
     const bool do_log          = std::stoi(argv[6]);
-    const int nrepeat          = std::stoi(argv[7]);
+    const bool time_kernel     = std::stoi(argv[7]);
 
     const int M = std::stoi(argv[8]);
     const int N = std::stoi(argv[9]);
@@ -74,7 +74,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -94,7 +94,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -114,7 +114,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -134,7 +134,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -154,7 +154,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -174,7 +174,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -194,7 +194,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -214,7 +214,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -234,7 +234,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -254,7 +254,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -274,7 +274,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -294,7 +294,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -314,7 +314,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -334,7 +334,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -354,7 +354,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -374,7 +374,7 @@ int profile_gemm(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,

profiler/src/profile_gemm_bias_2d.cpp

@@ -36,8 +36,8 @@ int profile_gemm_bias_2d(int argc, char* argv[])
         printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
         printf("arg4: verification (0: no; 1: yes)\n");
         printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
-        printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg7: run kernel # of times (>1)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
         printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
         printf("arg14: alpha\n");
         printf("arg15: beta\n");
@@ -50,7 +50,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[4]);
     const int init_method      = std::stoi(argv[5]);
     const bool do_log          = std::stoi(argv[6]);
-    const int nrepeat          = std::stoi(argv[7]);
+    const bool time_kernel     = std::stoi(argv[7]);
 
     const int M = std::stoi(argv[8]);
     const int N = std::stoi(argv[9]);
@@ -76,7 +76,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -99,7 +99,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -122,7 +122,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -145,7 +145,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -168,7 +168,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -191,7 +191,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -214,7 +214,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -237,7 +237,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,

profiler/src/profile_gemm_bias_relu.cpp

@@ -36,8 +36,8 @@ int profile_gemm_bias_relu(int argc, char* argv[])
         printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
         printf("arg4: verification (0: no; 1: yes)\n");
         printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
-        printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg7: run kernel # of times (>1)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
         printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
         printf("arg14: split k into  mulitiple batch\n");
         exit(1);
@@ -48,7 +48,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[4]);
     const int init_method      = std::stoi(argv[5]);
     const bool do_log          = std::stoi(argv[6]);
-    const int nrepeat          = std::stoi(argv[7]);
+    const bool time_kernel     = std::stoi(argv[7]);
 
     const int M = std::stoi(argv[8]);
     const int N = std::stoi(argv[9]);
@@ -69,7 +69,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -88,7 +88,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -107,7 +107,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -126,7 +126,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,

profiler/src/profile_gemm_bias_relu_add.cpp

@@ -36,8 +36,8 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
         printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
         printf("arg4: verification (0: no; 1: yes)\n");
         printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
-        printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg7: run kernel # of times (>1)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
         printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, StrideC1\n");
         printf("arg15: split k into  mulitiple batch\n");
         exit(1);
@@ -48,7 +48,7 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[4]);
     const int init_method      = std::stoi(argv[5]);
     const bool do_log          = std::stoi(argv[6]);
-    const int nrepeat          = std::stoi(argv[7]);
+    const bool time_kernel     = std::stoi(argv[7]);
 
     const int M = std::stoi(argv[8]);
     const int N = std::stoi(argv[9]);
@@ -70,7 +70,7 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -90,7 +90,7 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -110,7 +110,7 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -130,7 +130,7 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,

profiler/src/profile_gemm_reduce.cpp

@@ -32,8 +32,8 @@ int profile_gemm_reduce(int argc, char* argv[])
         printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
         printf("arg4: verification (0: no; 1: yes)\n");
         printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
-        printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg7: run kernel # of times (>1)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
         printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
         printf("arg14: split k into  mulitiple batch\n");
         exit(1);
@@ -44,7 +44,7 @@ int profile_gemm_reduce(int argc, char* argv[])
     const bool do_verification = std::stoi(argv[4]);
     const int init_method      = std::stoi(argv[5]);
     const bool do_log          = std::stoi(argv[6]);
-    const int nrepeat          = std::stoi(argv[7]);
+    const bool time_kernel     = std::stoi(argv[7]);
 
     const int M = std::stoi(argv[8]);
     const int N = std::stoi(argv[9]);
@@ -66,7 +66,7 @@ int profile_gemm_reduce(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -87,7 +87,7 @@ int profile_gemm_reduce(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -108,7 +108,7 @@ int profile_gemm_reduce(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,
@@ -129,7 +129,7 @@ int profile_gemm_reduce(int argc, char* argv[])
             do_verification,
             init_method,
             do_log,
-            nrepeat,
+            time_kernel,
             M,
             N,
             K,