Merge branch 'develop' into gemm_bf16_sk_muozturk

include/ck_tile/ref/README.md

+# reference
+
+this folder contains reference implementation of a specific op. Note by including a specific header, you are including the implementation(expecially the gpu implementation) into your source code, and compile that kernel into the fatbin, hence may increase your kernel obj code length. Usually the header starts with `reference_` is a cpu reference implementation. The header starts with `naive_` contains a gpu implementation with a small launcher.
+
+TODO: move `host/reference` under this folder

include/ck_tile/remod.py

@@ -7,6 +7,7 @@ import copy
 
 NS = 'ck_tile'
 OPS = 'ops'
+REF = 'ref'
 OPS_COMMON = 'common' # common header will be duplicated into ops/* other module
 
 HEADER_COMMON = f"""// SPDX-License-Identifier: MIT
@@ -29,6 +30,9 @@ class submodule_t:
     def push(self, f):
         if len(f.parents) != 1: # ignore ./xxx.hpp
             mod = get_module(f)
+            # ref is supposed to include one header on demand
+            if mod == REF:
+                return
             if mod == OPS:
                 if mod not in self.m.keys():
                     self.m[mod] = dict()

library/src/tensor_operation_instance/gpu/gemm_universal_batched/device_batched_gemm_xdl_universal_bf16_bf16_bf16/device_batched_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp

@@ -52,6 +52,9 @@ using device_batched_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_instances =
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   224,   256,    64,   8,   8,  16,   16,    7,    8,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           2,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   224,    64,   8,   8,  16,   16,    8,    7,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          2,           1,                   S<1, 32, 1, 8>,          S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   160,    64,   8,   8,  16,   16,    8,    5,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          2,           1,                   S<1, 32, 1, 8>,          S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   160,    64,   8,   8,  32,   32,    1,    5,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 64, 1, 4>,          S<8>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   160,   128,    64,   8,   8,  32,   32,    5,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>

library/src/tensor_operation_instance/gpu/gemm_universal_batched/device_batched_gemm_xdl_universal_f8_f8_bf16/device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn.hpp

@@ -42,6 +42,7 @@ using device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_instances = std
         //##################################|        |        |         |        | Type|  Type|         Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
         //##################################|        |        |         |        |     |      |             |      |        |         |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
         //##################################|        |        |         |        |     |      |             |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
+        
 #ifdef __gfx94__
         // Compute friendly
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,    Row,     F8,     F8, DsDataType, BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    64,  16,  16,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,               S<8>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
@@ -72,6 +73,7 @@ using device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_instances = std:
         //##################################|        |        |         |        | Type|  Type|         Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|          Pipeline|                     Pipeline|
         //##################################|        |        |         |        |     |      |             |      |        |         |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|         Scheduler|                     Verision|
         //##################################|        |        |         |        |     |      |             |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                  |                             |
+
 #if defined(__gfx94__) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,    Row,     F8,     F8, DsDataType,   BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    128, 16,  16,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               S<2>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,    Row,     F8,     F8, DsDataType,   BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    128, 16,  16,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 4>,               S<4>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,

library/src/tensor_operation_instance/gpu/mha/CMakeLists.txt

@@ -6,7 +6,7 @@ set(CK_TILE_SRC_FOLDER ${CMAKE_SOURCE_DIR}/include/ck_tile/)
 # CK Codegen requires dataclass which is added in Python 3.7
 # Python version 3.8 is required for general good practice as it is default for Ubuntu 20.04
 if(NOT CK_USE_ALTERNATIVE_PYTHON)
-   find_package(PythonInterp 3 REQUIRED)
+   find_package(Python3 COMPONENTS Interpreter Development)
 else()
    message("Using alternative python version")
    set(EXTRA_PYTHON_PATH)
@@ -33,7 +33,7 @@ set(FMHA_KNOWN_APIS "fwd,fwd_splitkv,fwd_appendkv,bwd")
 # Note: The receipt 3 arg filters the generated backwards instances to reduce compilation time.
 # With receipt 3 set, we are generating instances for datatype == {fp16 || bfp16}, bias == {no || alibi}, deterministic == off, and dpad == dvpad.
 execute_process(
-  COMMAND ${PYTHON_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
+  COMMAND ${Python3_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
   --list_blobs ${FMHA_CPP_FOLDER}/blob_list.txt
   --api ${FMHA_KNOWN_APIS}
   --receipt 3
@@ -50,7 +50,7 @@ endif()
 # With receipt 3 set, we are generating instances for datatype == {fp16 || bfp16}, bias == {no || alibi}, deterministic == off, and dpad == dvpad.
 add_custom_command(
   OUTPUT ${FMHA_GEN_BLOBS}
-  COMMAND ${PYTHON_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
+  COMMAND ${Python3_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
   --output_dir ${FMHA_CPP_FOLDER}
   --api ${FMHA_KNOWN_APIS}
   --receipt 3

modified_files.txt

-example/01_gemm/gemm_xdl_fp8_streamk_v3.cpp
-example/01_gemm/run_gemm_example_streamk_v2.inc
-include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp
-include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_streamk_v3.hpp
-library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f8_f16/device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
-library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f8_f16/device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_comp_mnpadding_instance.cpp
-library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f8_f16_f16/device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
-library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f8_f16_f16/device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_comp_mnpadding_instance.cpp
-profiler/src/profile_gemm_universal_streamk.cpp
-modified_files.txt

profiler/include/profiler/profile_gemm_universal_batched_impl.hpp

@@ -48,6 +48,7 @@ bool profile_gemm_universal_batched_impl(int do_verification,
                                          int StrideB,
                                          int StrideC,
                                          int BatchCount,
+                                         int KBatch,
                                          int n_warmup,
                                          int n_iter,
                                          uint64_t rotating = 0)
@@ -147,89 +148,100 @@ bool profile_gemm_universal_batched_impl(int do_verification,
     float best_ave_time   = 0;
     float best_tflops     = 0;
     float best_gb_per_sec = 0;
+    float best_kbatch     = 0;
 
     // profile device op instances
     for(auto& op_ptr : op_ptrs)
     {
-        std::unique_ptr<tensor_operation::device::BaseArgument> argument_ptr;
-        // false branch for multi d dl kernel
-
-        argument_ptr =
-            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
-                                        static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
-                                        {},
-                                        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
-                                        M,
-                                        N,
-                                        K,
-                                        BatchCount,
-                                        StrideA,
-                                        StrideB,
-                                        {},
-                                        StrideC,
-                                        BatchStrideA,
-                                        BatchStrideB,
-                                        {},
-                                        BatchStrideC,
-                                        ck::tensor_operation::element_wise::PassThrough{},
-                                        ck::tensor_operation::element_wise::PassThrough{},
-                                        ck::tensor_operation::element_wise::PassThrough{});
-
-        auto invoker_ptr = op_ptr->MakeInvokerPointer();
-
-        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
-        {
-            // re-init C to zero before profiling next kernel
-            c_device_buf.SetZero();
-
-            std::string op_name = op_ptr->GetTypeString();
+        std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
 
-            float ave_time = invoker_ptr->Run(
-                argument_ptr.get(),
-                StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter, true, rotating_count});
+        if(KBatch > 0)
+        {
+            kbatch_list = {KBatch};
+        }
 
-            std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
+        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()),
+                                            {},
+                                            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                            M,
+                                            N,
+                                            K,
+                                            BatchCount,
+                                            StrideA,
+                                            StrideB,
+                                            {},
+                                            StrideC,
+                                            BatchStrideA,
+                                            BatchStrideB,
+                                            {},
+                                            BatchStrideC,
+                                            ck::tensor_operation::element_wise::PassThrough{},
+                                            ck::tensor_operation::element_wise::PassThrough{},
+                                            ck::tensor_operation::element_wise::PassThrough{},
+                                            kbatch_curr);
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                std::string op_name = op_ptr->GetTypeString();
 
-            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
-                                     sizeof(CDataType) * M * N) *
-                                    BatchCount;
+                float ave_time = invoker_ptr->Run(
+                    argument_ptr.get(),
+                    StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter, true, rotating_count});
 
-            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+                std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
 
-            float gb_per_sec = num_btype / 1.E6 / ave_time;
+                std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                         sizeof(CDataType) * M * N) *
+                                        BatchCount;
 
-            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
-                      << " GB/s, " << op_name << std::endl;
+                float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
 
-            if(tflops > best_tflops)
-            {
-                best_op_name    = op_name;
-                best_tflops     = tflops;
-                best_ave_time   = ave_time;
-                best_gb_per_sec = gb_per_sec;
-            }
+                float gb_per_sec = num_btype / 1.E6 / ave_time;
 
-            if(do_verification)
-            {
-                c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
+                std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                          << " GB/s, " << op_name << ", KBatch " << kbatch_curr << std::endl;
 
-                pass = pass & ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
+                if(tflops > best_tflops)
+                {
+                    best_op_name    = op_name;
+                    best_tflops     = tflops;
+                    best_ave_time   = ave_time;
+                    best_gb_per_sec = gb_per_sec;
+                    best_kbatch     = kbatch_curr;
+                }
 
-                if(do_log)
+                if(do_verification)
                 {
-                    LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
-                    LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
-                    LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
-                        << std::endl;
-                    LogRangeAsType<float>(
-                        std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
-                        << std::endl;
+                    c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
+
+                    pass = pass & ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
+
+                    if(do_log)
+                    {
+                        LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
+                        LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
+                            << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
+                            << std::endl;
+                    }
                 }
             }
-        }
-        else
-        {
-            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+            else
+            {
+                std::cout << op_ptr->GetTypeString() << " does not support this problem"
+                          << std::endl;
+            }
         }
     }
 
@@ -270,8 +282,8 @@ bool profile_gemm_universal_batched_impl(int do_verification,
 
     std::cout << " B = " << BatchCount << " M = " << M << " N = " << N << " K = " << K
               << " StrideA = " << StrideA << " StrideB = " << StrideB << " StrideC = " << StrideC
-              << ": " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
-              << " GB/s, " << best_op_name << std::endl;
+              << " KBatch = " << best_kbatch << ": " << best_ave_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
 
     return pass;
 }

script/process_perf_data.py

@@ -82,7 +82,7 @@ def parse_logfile(logfile):
     StrideA=[]
     StrideB=[]
     StrideC=[]
-    if 'perf_gemm.log' in logfile:
+    if 'perf_gemm' in logfile and 'gemm_bilinear' not in logfile:
         for line in open(logfile):
             if 'Best Perf' in line:
                 lst=line.split()
@@ -260,7 +260,7 @@ def main():
         conn = sqlEngine.connect()
 
         #save gemm performance tests:
-        if 'perf_gemm.log' in filename:
+        if 'perf_gemm' in filename and 'gemm_bilinear' not in filename:
             #write the ck_gemm_test_params table only needed once the test set changes
             #post_test_params(test_list,conn)
             for i in range(1,len(results)+1):
@@ -332,7 +332,7 @@ def main():
             table_name="ck_fmha_bwd_tflops"
 
         tflops_base = get_baseline(table_name,conn)
-        store_new_test_result(table_name, results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, conn)
+        store_new_test_result(table_name, results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, sqlEngine)
         conn.close()
 
     #compare the results to the baseline if baseline exists