diff --git a/test/gemm_universal/test_gemm_universal_ut_cases_bf16.inc b/test/gemm_universal/test_gemm_universal_ut_cases_bf16.inc
index 8a6c672a9f..233f86ef43 100644
--- a/test/gemm_universal/test_gemm_universal_ut_cases_bf16.inc
+++ b/test/gemm_universal/test_gemm_universal_ut_cases_bf16.inc
@@ -207,3 +207,35 @@ TYPED_TEST(TestGemmUniversal_BF16_MK_NK, Regular)
     for(int M : Ms)
         this->Run(M, N, K, StrideA, StrideB, StrideC);
 }
+
+TYPED_TEST(TestGemmUniversal_BF16_KM_KN, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 512;
+    constexpr int K = 512;
+
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+    {
+        int StrideA = M;
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+    }
+}
+
+TYPED_TEST(TestGemmUniversal_BF16_KM_NK, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 512;
+    constexpr int K = 512;
+
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+    {
+        int StrideA = M;
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+    }
+}
diff --git a/test/gemm_universal/test_gemm_universal_ut_cases_fp16.inc b/test/gemm_universal/test_gemm_universal_ut_cases_fp16.inc
index b80dc0a325..adc84848f2 100644
--- a/test/gemm_universal/test_gemm_universal_ut_cases_fp16.inc
+++ b/test/gemm_universal/test_gemm_universal_ut_cases_fp16.inc
@@ -207,3 +207,35 @@ TYPED_TEST(TestGemmUniversal_FP16_MK_NK, Regular)
     for(int M : Ms)
         this->Run(M, N, K, StrideA, StrideB, StrideC);
 }
+
+TYPED_TEST(TestGemmUniversal_FP16_KM_KN, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 512;
+    constexpr int K = 512;
+
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+    {
+        int StrideA = M;
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+    }
+}
+
+TYPED_TEST(TestGemmUniversal_FP16_KM_NK, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 512;
+    constexpr int K = 512;
+
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+    {
+        int StrideA = M;
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+    }
+}