Update ckProfiler for mixed precision cases

profiler/include/profiler/profile_gemm_splitk_impl.hpp

@@ -191,7 +191,20 @@ bool profile_gemm_splitk_impl(int do_verification,
             {
                 c_device_buf.FromDevice(c_m_n_device_result.mData.data());
 
-                pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+                // set softer tolerances for fp8
+                if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
+                             is_same_v<CDataType, f8_t>)
+                {
+                    std::string msg = "Error: Incorrect results!";
+                    double rtol     = 1e-1;
+                    double atol     = 1e-1;
+                    pass            = pass & ck::utils::check_err(
+                                      c_m_n_device_result, c_m_n_host_result, msg, rtol, atol);
+                }
+                else
+                {
+                    pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+                }
 
                 if(do_log)
                 {

profiler/src/profile_gemm_splitk.cpp

@@ -23,6 +23,8 @@ enum struct GemmDataType
     F16_F16_F16,    // 1
     BF16_BF16_BF16, // 2
     INT8_INT8_INT8, // 3
+    F8_F16_F16,     // 4
+    F16_F8_F16,     // 5
 };
 
 #define OP_NAME "gemm_splitk"
@@ -33,7 +35,7 @@ int profile_gemm_splitk(int argc, char* argv[])
     if(argc != 15)
     {
         printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
-        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8)\n");
         printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
         printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
         printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
@@ -65,6 +67,7 @@ int profile_gemm_splitk(int argc, char* argv[])
 
     using F32 = float;
     using F16 = ck::half_t;
+    using F8  = ck::f8_t;
 
     using Row = ck::tensor_layout::gemm::RowMajor;
     using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -143,6 +146,38 @@ int profile_gemm_splitk(int argc, char* argv[])
     {
         return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
     }
+    else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Col{}, Row{});
+    }
     else
     {
         std::cout << "this data_type & layout is not implemented" << std::endl;