From 8658ca6cf33d791f2ad04618eb5874358f1f2f9e Mon Sep 17 00:00:00 2001
From: Zoltan Lakatos <zoltan.lakatos@streamhpc.com>
Date: Thu, 5 Jun 2025 13:12:07 +0000
Subject: [PATCH] log print added to splitk benchmarks

---
 .../profiler/profile_gemm_add_fastgelu_impl.hpp     | 13 ++++++++++++-
 .../profiler/profile_gemm_add_multiply_impl.hpp     | 13 ++++++++++++-
 test/gemm_add/test_gemm_add_multiply_wmma.cpp       |  2 --
 3 files changed, 24 insertions(+), 4 deletions(-)

diff --git a/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp b/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
index 6f6d881c1e..9e4d30142b 100644
--- a/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
+++ b/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
@@ -33,7 +33,7 @@ template <typename ADataType,
           typename ELayout>
 bool profile_gemm_add_fastgelu_impl(int do_verification,
                                     int init_method,
-                                    bool /*do_log*/,
+                                    bool do_log,
                                     bool time_kernel,
                                     int M,
                                     int N,
@@ -213,6 +213,17 @@ bool profile_gemm_add_fastgelu_impl(int do_verification,
             {
                 e_device_buf.FromDevice(e_m_n_device_result.mData.data());
 
+                if(do_log)
+                {
+                    LogRangeAsType<float>(
+                        std::cout << "e_m_n_device_result: ", e_m_n_device_result.mData, ",")
+                        << std::endl;
+
+                    LogRangeAsType<float>(
+                        std::cout << "e_m_n_host_result: ", e_m_n_host_result.mData, ",")
+                        << std::endl;
+                }
+
                 pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
             }
         }
diff --git a/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp b/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
index 25871dfb2e..fcb546fe96 100644
--- a/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
+++ b/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
@@ -35,7 +35,7 @@ template <typename ADataType,
           typename ELayout>
 bool profile_gemm_add_multiply_impl(int do_verification,
                                     int init_method,
-                                    bool /*do_log*/,
+                                    bool do_log,
                                     bool time_kernel,
                                     int M,
                                     int N,
@@ -223,6 +223,17 @@ bool profile_gemm_add_multiply_impl(int do_verification,
             {
                 e_device_buf.FromDevice(e_m_n_device_result.mData.data());
 
+                if(do_log)
+                {
+                    LogRangeAsType<float>(
+                        std::cout << "e_m_n_device_result: ", e_m_n_device_result.mData, ",")
+                        << std::endl;
+
+                    LogRangeAsType<float>(
+                        std::cout << "e_m_n_host_result: ", e_m_n_host_result.mData, ",")
+                        << std::endl;
+                }
+
                 pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
             }
         }
diff --git a/test/gemm_add/test_gemm_add_multiply_wmma.cpp b/test/gemm_add/test_gemm_add_multiply_wmma.cpp
index 1859a37c00..00c63194aa 100644
--- a/test/gemm_add/test_gemm_add_multiply_wmma.cpp
+++ b/test/gemm_add/test_gemm_add_multiply_wmma.cpp
@@ -11,8 +11,6 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
 using F16  = ck::half_t;
 using F32  = float;
 
-// TODO: inerit TestGemmAddMultiply from TestGemmD0Common after changes are rebased on top of multipleD feature branch.
-// After that clean test...
 template <typename Tuple>
 class TestGemmAddMultiply : public ::testing::Test
 {