diff --git a/src/blas/backends/cublas/cublas_batch.cpp b/src/blas/backends/cublas/cublas_batch.cpp
index bea86b596..85eadbe42 100644
--- a/src/blas/backends/cublas/cublas_batch.cpp
+++ b/src/blas/backends/cublas/cublas_batch.cpp
@@ -162,10 +162,10 @@ inline void gemm_batch(const char *func_name, Func func, sycl::queue &queue, tra
             auto b_ = sc.get_mem<cuDataType *>(b_acc);
             auto c_ = sc.get_mem<cuDataType *>(c_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_operation(transa),
-                                get_cublas_operation(transb), m, n, k, (cuDataType *)&alpha, a_,
-                                lda, stride_a, b_, ldb, stride_b, (cuDataType *)&beta, c_, ldc,
-                                stride_c, batch_size);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_operation(transa),
+                                     get_cublas_operation(transb), m, n, k, (cuDataType *)&alpha,
+                                     a_, lda, stride_a, b_, ldb, stride_b, (cuDataType *)&beta, c_,
+                                     ldc, stride_c, batch_size);
         });
     });
 }
@@ -495,10 +495,10 @@ inline sycl::event gemm_batch(const char *func_name, Func func, sycl::queue &que
             auto b_ = reinterpret_cast<const cuDataType *>(b);
             auto c_ = reinterpret_cast<cuDataType *>(c);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_operation(transa),
-                                get_cublas_operation(transb), m, n, k, (cuDataType *)&alpha, a_,
-                                lda, stride_a, b_, ldb, stride_b, (cuDataType *)&beta, c_, ldc,
-                                stride_c, batch_size);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_operation(transa),
+                                     get_cublas_operation(transb), m, n, k, (cuDataType *)&alpha,
+                                     a_, lda, stride_a, b_, ldb, stride_b, (cuDataType *)&beta, c_,
+                                     ldc, stride_c, batch_size);
         });
     });
     return done;
@@ -550,11 +550,11 @@ inline sycl::event gemm_batch(const char *func_name, Func func, sycl::queue &que
                 auto **a_ = reinterpret_cast<const cuDataType **>(a);
                 auto **b_ = reinterpret_cast<const cuDataType **>(b);
                 auto **c_ = reinterpret_cast<cuDataType **>(c);
-                CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_operation(transa[i]),
-                                    get_cublas_operation(transb[i]), (int)m[i], (int)n[i],
-                                    (int)k[i], (cuDataType *)&alpha[i], a_ + offset, (int)lda[i],
-                                    b_ + offset, (int)ldb[i], (cuDataType *)&beta[i], c_ + offset,
-                                    (int)ldc[i], (int)group_size[i]);
+                CUBLAS_ERROR_FUNC_T_SYNC(
+                    func_name, func, err, handle, get_cublas_operation(transa[i]),
+                    get_cublas_operation(transb[i]), (int)m[i], (int)n[i], (int)k[i],
+                    (cuDataType *)&alpha[i], a_ + offset, (int)lda[i], b_ + offset, (int)ldb[i],
+                    (cuDataType *)&beta[i], c_ + offset, (int)ldc[i], (int)group_size[i]);
                 offset += group_size[i];
             }
         });
@@ -632,7 +632,7 @@ inline sycl::event trsm_batch(const char *func_name, Func func, sycl::queue &que
             for (int64_t i = 0; i < group_count; i++) {
                 auto **a_ = reinterpret_cast<const cuDataType **>(a);
                 auto **b_ = reinterpret_cast<cuDataType **>(b);
-                CUBLAS_ERROR_FUNC_T(
+                CUBLAS_ERROR_FUNC_T_SYNC(
                     func_name, func, err, handle, get_cublas_side_mode(left_right[i]),
                     get_cublas_fill_mode(upper_lower[i]), get_cublas_operation(trans[i]),
                     get_cublas_diag_type(unit_diag[i]), (int)m[i], (int)n[i],
diff --git a/src/blas/backends/cublas/cublas_helper.hpp b/src/blas/backends/cublas/cublas_helper.hpp
index bf803a7c7..0ee9930e3 100644
--- a/src/blas/backends/cublas/cublas_helper.hpp
+++ b/src/blas/backends/cublas/cublas_helper.hpp
@@ -180,11 +180,23 @@ class cuda_error : virtual public std::runtime_error {
         throw cublas_error(std::string(#name) + std::string(" : "), err); \
     }
 
-#define CUBLAS_ERROR_FUNC_T(name, func, err, ...)                        \
-    err = func(__VA_ARGS__);                                             \
+#define CUBLAS_ERROR_FUNC_SYNC(name, err, handle, ...)                    \
+    err = name(handle, __VA_ARGS__);                                      \
+    if (err != CUBLAS_STATUS_SUCCESS) {                                   \
+        throw cublas_error(std::string(#name) + std::string(" : "), err); \
+    }                                                                     \
+    cudaStream_t currentStreamId;                                         \
+    CUBLAS_ERROR_FUNC(cublasGetStream, err, handle, &currentStreamId);    \
+    cuStreamSynchronize(currentStreamId);
+
+#define CUBLAS_ERROR_FUNC_T_SYNC(name, func, err, handle, ...)           \
+    err = func(handle, __VA_ARGS__);                                     \
     if (err != CUBLAS_STATUS_SUCCESS) {                                  \
         throw cublas_error(std::string(name) + std::string(" : "), err); \
-    }
+    }                                                                    \
+    cudaStream_t currentStreamId;                                        \
+    CUBLAS_ERROR_FUNC(cublasGetStream, err, handle, &currentStreamId);   \
+    cuStreamSynchronize(currentStreamId);
 
 inline cublasOperation_t get_cublas_operation(oneapi::mkl::transpose trn) {
     switch (trn) {
diff --git a/src/blas/backends/cublas/cublas_level1.cpp b/src/blas/backends/cublas/cublas_level1.cpp
index 7dee2657a..24f18f096 100644
--- a/src/blas/backends/cublas/cublas_level1.cpp
+++ b/src/blas/backends/cublas/cublas_level1.cpp
@@ -53,7 +53,7 @@ inline void asum(const char *func_name, Func func, sycl::queue &queue, int64_t n
             auto res_ = sc.get_mem<cuDataType2 *>(res_acc);
             cublasStatus_t err;
             // ASUM does not support negative index
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, std::abs(incx), res_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, std::abs(incx), res_);
             // Higher level BLAS functions expect CUBLAS_POINTER_MODE_HOST
             // to be set, therfore we need to reset this to the default value
             // in order to avoid CUDA_ERROR_ILLEGAL_ADRESS errors
@@ -86,8 +86,8 @@ inline void scal(const char *func_name, Func func, sycl::queue &queue, int64_t n
             auto x_ = sc.get_mem<cuDataType2 *>(x_acc);
             cublasStatus_t err;
             // SCAL does not support negative incx
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, (cuDataType1 *)&a, x_,
-                                std::abs(incx));
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, (cuDataType1 *)&a, x_,
+                                     std::abs(incx));
         });
     });
 }
@@ -117,8 +117,8 @@ inline void axpy(const char *func_name, Func func, sycl::queue &queue, int64_t n
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, (cuDataType *)&alpha, x_, incx, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy);
         });
     });
 }
@@ -180,7 +180,7 @@ inline void rotg(const char *func_name, Func func, sycl::queue &queue, sycl::buf
             auto c_ = sc.get_mem<cuDataType2 *>(c_acc);
             auto s_ = sc.get_mem<cuDataType1 *>(s_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, a_, b_, c_, s_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, a_, b_, c_, s_);
             // Higher level BLAS functions expect CUBLAS_POINTER_MODE_HOST
             // to be set, therfore we need to reset this to the default value
             // in order to avoid CUDA_ERROR_ILLEGAL_ADRESS errors
@@ -223,7 +223,7 @@ inline void rotm(const char *func_name, Func func, sycl::queue &queue, int64_t n
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             auto param_ = sc.get_mem<cuDataType *>(param_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, y_, incy, param_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, y_, incy, param_);
             // Higher level BLAS functions expect CUBLAS_POINTER_MODE_HOST
             // to be set, therfore we need to reset this to the default value
             // in order to avoid CUDA_ERROR_ILLEGAL_ADRESS errors
@@ -255,7 +255,7 @@ inline void copy(const char *func_name, Func func, sycl::queue &queue, int64_t n
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, y_, incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, y_, incy);
         });
     });
 }
@@ -294,7 +294,7 @@ inline void dot(const char *func_name, Func func, sycl::queue &queue, int64_t n,
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             auto res_ = sc.get_mem<cuDataType *>(res_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, y_, incy, res_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, y_, incy, res_);
             // Higher level BLAS functions expect CUBLAS_POINTER_MODE_HOST
             // to be set, therfore we need to reset this to the default value
             // in order to avoid CUDA_ERROR_ILLEGAL_ADRESS errors
@@ -338,8 +338,8 @@ inline void rot(const char *func_name, Func func, sycl::queue &queue, int64_t n,
             auto x_ = sc.get_mem<cuDataType1 *>(x_acc);
             auto y_ = sc.get_mem<cuDataType1 *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, y_, incy,
-                                (cuDataType2 *)&c, (cuDataType3 *)&s);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, y_, incy,
+                                     (cuDataType2 *)&c, (cuDataType3 *)&s);
         });
     });
 }
@@ -376,7 +376,7 @@ void sdsdot(sycl::queue &queue, int64_t n, float sb, sycl::buffer<float, 1> &x,
             auto y_ = sc.get_mem<float *>(y_acc);
             auto res_ = sc.get_mem<float *>(res_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC(cublasSdot, err, handle, n, x_, incx, y_, incy, res_);
+            CUBLAS_ERROR_FUNC_SYNC(cublasSdot, err, handle, n, x_, incx, y_, incy, res_);
             // Higher level BLAS functions expect CUBLAS_POINTER_MODE_HOST
             // to be set, therfore we need to reset this to the default value
             // in order to avoid CUDA_ERROR_ILLEGAL_ADRESS errors
@@ -418,7 +418,7 @@ inline void rotmg(const char *func_name, Func func, sycl::queue &queue, sycl::bu
             auto y1_ = sc.get_mem<cuDataType *>(y1_acc);
             auto param_ = sc.get_mem<cuDataType *>(param_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, d1_, d2_, x1_, y1_, param_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, d1_, d2_, x1_, y1_, param_);
             // Higher level BLAS functions expect CUBLAS_POINTER_MODE_HOST
             // to be set, therfore we need to reset this to the default value
             // in order to avoid CUDA_ERROR_ILLEGAL_ADRESS errors
@@ -466,7 +466,7 @@ inline void iamax(const char *func_name, Func func, sycl::queue &queue, int64_t
             cublasStatus_t err;
             // For negative incx, iamax returns 0. This behaviour is similar to that of
             // reference netlib BLAS.
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, int_res_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, int_res_);
             // Higher level BLAS functions expect CUBLAS_POINTER_MODE_HOST
             // to be set, therfore we need to reset this to the default value
             // in order to avoid CUDA_ERROR_ILLEGAL_ADRESS errors
@@ -503,7 +503,7 @@ inline void swap(const char *func_name, Func func, sycl::queue &queue, int64_t n
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, y_, incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, y_, incy);
         });
     });
 }
@@ -549,7 +549,7 @@ inline void iamin(const char *func_name, Func func, sycl::queue &queue, int64_t
             cublasStatus_t err;
             // For negative incx, iamin returns 0. This behaviour is similar to that of
             // implemented as a reference IAMIN.
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, int_res_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, int_res_);
             // Higher level BLAS functions expect CUBLAS_POINTER_MODE_HOST
             // to be set, therfore we need to reset this to the default value
             // in order to avoid CUDA_ERROR_ILLEGAL_ADRESS errors
@@ -593,7 +593,7 @@ inline void nrm2(const char *func_name, Func func, sycl::queue &queue, int64_t n
             auto res_ = sc.get_mem<cuDataType2 *>(res_acc);
             cublasStatus_t err;
             // NRM2 does not support negative index
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, std::abs(incx), res_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, std::abs(incx), res_);
             // Higher level BLAS functions expect CUBLAS_POINTER_MODE_HOST
             // to be set, therfore we need to reset this to the default value
             // in order to avoid CUDA_ERROR_ILLEGAL_ADRESS errors
@@ -635,7 +635,7 @@ inline sycl::event asum(const char *func_name, Func func, sycl::queue &queue, in
             auto res_ = reinterpret_cast<cuDataType2 *>(result);
             cublasStatus_t err;
             // ASUM does not support negative index
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, std::abs(incx), res_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, std::abs(incx), res_);
         });
     });
     return done;
@@ -668,8 +668,8 @@ inline sycl::event scal(const char *func_name, Func func, sycl::queue &queue, in
             auto x_ = reinterpret_cast<cuDataType2 *>(x);
             cublasStatus_t err;
             // SCAL does not support negative incx
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, (cuDataType1 *)&a, x_,
-                                std::abs(incx));
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, (cuDataType1 *)&a, x_,
+                                     std::abs(incx));
         });
     });
     return done;
@@ -704,8 +704,8 @@ inline sycl::event axpy(const char *func_name, Func func, sycl::queue &queue, in
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, (cuDataType *)&alpha, x_, incx, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy);
         });
     });
     return done;
@@ -764,7 +764,7 @@ inline sycl::event rotg(const char *func_name, Func func, sycl::queue &queue, T1
             auto c_ = reinterpret_cast<cuDataType2 *>(c);
             auto s_ = reinterpret_cast<cuDataType1 *>(s);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, a_, b_, c_, s_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, a_, b_, c_, s_);
         });
     });
     return done;
@@ -799,7 +799,7 @@ inline sycl::event rotm(const char *func_name, Func func, sycl::queue &queue, in
             auto y_ = reinterpret_cast<cuDataType *>(y);
             auto param_ = reinterpret_cast<cuDataType *>(param);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, y_, incy, param_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, y_, incy, param_);
         });
     });
     return done;
@@ -832,7 +832,7 @@ inline sycl::event copy(const char *func_name, Func func, sycl::queue &queue, in
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, y_, incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, y_, incy);
         });
     });
     return done;
@@ -867,7 +867,7 @@ inline sycl::event dot(const char *func_name, Func func, sycl::queue &queue, int
             auto y_ = reinterpret_cast<const cuDataType *>(y);
             auto res_ = reinterpret_cast<cuDataType *>(result);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, y_, incy, res_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, y_, incy, res_);
         });
     });
     return done;
@@ -906,8 +906,8 @@ inline sycl::event rot(const char *func_name, Func func, sycl::queue &queue, int
             auto x_ = reinterpret_cast<cuDataType1 *>(x);
             auto y_ = reinterpret_cast<cuDataType1 *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, y_, incy,
-                                (cuDataType2 *)&c, (cuDataType3 *)&s);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, y_, incy,
+                                     (cuDataType2 *)&c, (cuDataType3 *)&s);
         });
     });
     return done;
@@ -943,7 +943,7 @@ sycl::event sdsdot(sycl::queue &queue, int64_t n, float sb, const float *x, int6
             auto y_ = reinterpret_cast<const float *>(y);
             auto res_ = reinterpret_cast<float *>(result);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC(cublasSdot, err, handle, n, x_, incx, y_, incy, res_);
+            CUBLAS_ERROR_FUNC_SYNC(cublasSdot, err, handle, n, x_, incx, y_, incy, res_);
         });
     });
     done.wait();
@@ -973,7 +973,7 @@ inline sycl::event rotmg(const char *func_name, Func func, sycl::queue &queue, T
             auto y1_ = reinterpret_cast<const cuDataType *>(&y1);
             auto param_ = reinterpret_cast<cuDataType *>(param);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, d1_, d2_, x1_, y1_, param_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, d1_, d2_, x1_, y1_, param_);
         });
     });
     return done;
@@ -1014,7 +1014,7 @@ inline sycl::event iamax(const char *func_name, Func func, sycl::queue &queue, i
             cublasStatus_t err;
             // For negative incx, iamax returns 0. This behaviour is similar to that of
             // reference iamax.
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, int_res_p_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, int_res_p_);
         });
     });
     done.wait();
@@ -1049,7 +1049,7 @@ inline sycl::event swap(const char *func_name, Func func, sycl::queue &queue, in
             auto x_ = reinterpret_cast<cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, y_, incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, y_, incy);
         });
     });
     return done;
@@ -1092,7 +1092,7 @@ inline sycl::event iamin(const char *func_name, Func func, sycl::queue &queue, i
             cublasStatus_t err;
             // For negative incx, iamin returns 0. This behaviour is similar to that of
             // implemented iamin.
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, incx, int_res_p_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, incx, int_res_p_);
         });
     });
     done.wait();
@@ -1130,7 +1130,7 @@ inline sycl::event nrm2(const char *func_name, Func func, sycl::queue &queue, in
             auto res_ = reinterpret_cast<cuDataType2 *>(result);
             cublasStatus_t err;
             // NRM2 does not support negative index
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, n, x_, std::abs(incx), res_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, n, x_, std::abs(incx), res_);
         });
     });
     return done;
diff --git a/src/blas/backends/cublas/cublas_level2.cpp b/src/blas/backends/cublas/cublas_level2.cpp
index 1a3ef88bf..8f711243b 100644
--- a/src/blas/backends/cublas/cublas_level2.cpp
+++ b/src/blas/backends/cublas/cublas_level2.cpp
@@ -46,9 +46,9 @@ inline void gemv(const char *func_name, Func func, sycl::queue &queue, transpose
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_operation(trans), m, n,
-                                (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_operation(trans), m,
+                                     n, (cuDataType *)&alpha, a_, lda, x_, incx,
+                                     (cuDataType *)&beta, y_, incy);
         });
     });
 }
@@ -83,9 +83,9 @@ inline void gbmv(const char *func_name, Func func, sycl::queue &queue, transpose
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_operation(trans), m, n, kl,
-                                ku, (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta,
-                                y_, incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_operation(trans), m,
+                                     n, kl, ku, (cuDataType *)&alpha, a_, lda, x_, incx,
+                                     (cuDataType *)&beta, y_, incy);
         });
     });
 }
@@ -120,8 +120,8 @@ inline void ger(const char *func_name, Func func, sycl::queue &queue, int64_t m,
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, m, n, (cuDataType *)&alpha, x_, incx,
-                                y_, incy, a_, lda);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, m, n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy, a_, lda);
         });
     });
 }
@@ -157,9 +157,9 @@ inline void hbmv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                k, (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, k, (cuDataType *)&alpha,
+                                     a_, lda, x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
 }
@@ -192,9 +192,9 @@ inline void hemv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, a_,
+                                     lda, x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
 }
@@ -227,8 +227,9 @@ inline void her(const char *func_name, Func func, sycl::queue &queue, uplo upper
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuScalarType *)&alpha, x_, incx, a_, lda);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuScalarType *)&alpha,
+                                     x_, incx, a_, lda);
         });
     });
 }
@@ -261,8 +262,9 @@ inline void her2(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, y_, incy, a_, lda);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy, a_, lda);
         });
     });
 }
@@ -296,8 +298,9 @@ inline void hpmv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, a_, x_, incx, (cuDataType *)&beta, y_, incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, a_,
+                                     x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
 }
@@ -330,8 +333,9 @@ inline void hpr(const char *func_name, Func func, sycl::queue &queue, uplo upper
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuScalarType *)&alpha, x_, incx, a_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuScalarType *)&alpha,
+                                     x_, incx, a_);
         });
     });
 }
@@ -363,8 +367,9 @@ inline void hpr2(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, y_, incy, a_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy, a_);
         });
     });
 }
@@ -397,9 +402,9 @@ inline void sbmv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                k, (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, k, (cuDataType *)&alpha,
+                                     a_, lda, x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
 }
@@ -433,9 +438,9 @@ inline void symv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, a_,
+                                     lda, x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
 }
@@ -466,8 +471,9 @@ inline void syr(const char *func_name, Func func, sycl::queue &queue, uplo upper
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, a_, lda);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, a_, lda);
         });
     });
 }
@@ -501,8 +507,9 @@ inline void syr2(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, y_, incy, a_, lda);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy, a_, lda);
         });
     });
 }
@@ -539,8 +546,9 @@ inline void spmv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, a_, x_, incx, (cuDataType *)&beta, y_, incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, a_,
+                                     x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
 }
@@ -571,8 +579,9 @@ inline void spr(const char *func_name, Func func, sycl::queue &queue, uplo upper
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, a_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, a_);
         });
     });
 }
@@ -604,8 +613,9 @@ inline void spr2(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             auto y_ = sc.get_mem<cuDataType *>(y_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, y_, incy, a_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy, a_);
         });
     });
 }
@@ -636,9 +646,9 @@ inline void tbmv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, k,
-                                a_, lda, x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, k, a_, lda, x_, incx);
         });
     });
 }
@@ -672,9 +682,9 @@ inline void tbsv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, k,
-                                a_, lda, x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, k, a_, lda, x_, incx);
         });
     });
 }
@@ -708,9 +718,9 @@ inline void tpmv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, a_,
-                                x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, a_, x_, incx);
         });
     });
 }
@@ -743,9 +753,9 @@ inline void tpsv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, a_,
-                                x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, a_, x_, incx);
         });
     });
 }
@@ -778,9 +788,9 @@ inline void trmv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, a_,
-                                lda, x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, a_, lda, x_, incx);
         });
     });
 }
@@ -813,9 +823,9 @@ inline void trsv(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto x_ = sc.get_mem<cuDataType *>(x_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, a_,
-                                lda, x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, a_, lda, x_, incx);
         });
     });
 }
@@ -854,9 +864,9 @@ inline sycl::event gemv(const char *func_name, Func func, sycl::queue &queue, tr
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_operation(trans), m, n,
-                                (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_operation(trans), m,
+                                     n, (cuDataType *)&alpha, a_, lda, x_, incx,
+                                     (cuDataType *)&beta, y_, incy);
         });
     });
     return done;
@@ -894,9 +904,9 @@ inline sycl::event gbmv(const char *func_name, Func func, sycl::queue &queue, tr
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_operation(trans), m, n, kl,
-                                ku, (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta,
-                                y_, incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_operation(trans), m,
+                                     n, kl, ku, (cuDataType *)&alpha, a_, lda, x_, incx,
+                                     (cuDataType *)&beta, y_, incy);
         });
     });
     return done;
@@ -934,8 +944,8 @@ inline sycl::event ger(const char *func_name, Func func, sycl::queue &queue, int
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<const cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, m, n, (cuDataType *)&alpha, x_, incx,
-                                y_, incy, a_, lda);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, m, n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy, a_, lda);
         });
     });
     return done;
@@ -975,9 +985,9 @@ inline sycl::event hbmv(const char *func_name, Func func, sycl::queue &queue, up
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                k, (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, k, (cuDataType *)&alpha,
+                                     a_, lda, x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
     return done;
@@ -1012,9 +1022,9 @@ inline sycl::event hemv(const char *func_name, Func func, sycl::queue &queue, up
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, a_,
+                                     lda, x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
     return done;
@@ -1050,8 +1060,9 @@ inline sycl::event her(const char *func_name, Func func, sycl::queue &queue, upl
             auto a_ = reinterpret_cast<cuDataType *>(a);
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuScalarType *)&alpha, x_, incx, a_, lda);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuScalarType *)&alpha,
+                                     x_, incx, a_, lda);
         });
     });
     return done;
@@ -1087,8 +1098,9 @@ inline sycl::event her2(const char *func_name, Func func, sycl::queue &queue, up
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<const cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, y_, incy, a_, lda);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy, a_, lda);
         });
     });
     return done;
@@ -1124,8 +1136,9 @@ inline sycl::event hpmv(const char *func_name, Func func, sycl::queue &queue, up
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, a_, x_, incx, (cuDataType *)&beta, y_, incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, a_,
+                                     x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
     return done;
@@ -1161,8 +1174,9 @@ inline sycl::event hpr(const char *func_name, Func func, sycl::queue &queue, upl
             auto a_ = reinterpret_cast<cuDataType *>(a);
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuScalarType *)&alpha, x_, incx, a_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuScalarType *)&alpha,
+                                     x_, incx, a_);
         });
     });
     return done;
@@ -1198,8 +1212,9 @@ inline sycl::event hpr2(const char *func_name, Func func, sycl::queue &queue, up
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<const cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, y_, incy, a_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy, a_);
         });
     });
     return done;
@@ -1236,9 +1251,9 @@ inline sycl::event sbmv(const char *func_name, Func func, sycl::queue &queue, up
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                k, (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, k, (cuDataType *)&alpha,
+                                     a_, lda, x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
     return done;
@@ -1274,9 +1289,9 @@ inline sycl::event symv(const char *func_name, Func func, sycl::queue &queue, up
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, a_, lda, x_, incx, (cuDataType *)&beta, y_,
-                                incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, a_,
+                                     lda, x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
     return done;
@@ -1311,8 +1326,9 @@ inline sycl::event syr(const char *func_name, Func func, sycl::queue &queue, upl
             auto a_ = reinterpret_cast<cuDataType *>(a);
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, a_, lda);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, a_, lda);
         });
     });
     return done;
@@ -1350,8 +1366,9 @@ inline sycl::event syr2(const char *func_name, Func func, sycl::queue &queue, up
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<const cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, y_, incy, a_, lda);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy, a_, lda);
         });
     });
     return done;
@@ -1390,8 +1407,9 @@ inline sycl::event spmv(const char *func_name, Func func, sycl::queue &queue, up
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, a_, x_, incx, (cuDataType *)&beta, y_, incy);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, a_,
+                                     x_, incx, (cuDataType *)&beta, y_, incy);
         });
     });
     return done;
@@ -1426,8 +1444,9 @@ inline sycl::event spr(const char *func_name, Func func, sycl::queue &queue, upl
             auto a_ = reinterpret_cast<cuDataType *>(a);
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, a_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, a_);
         });
     });
     return done;
@@ -1462,8 +1481,9 @@ inline sycl::event spr2(const char *func_name, Func func, sycl::queue &queue, up
             auto x_ = reinterpret_cast<const cuDataType *>(x);
             auto y_ = reinterpret_cast<const cuDataType *>(y);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower), n,
-                                (cuDataType *)&alpha, x_, incx, y_, incy, a_);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), n, (cuDataType *)&alpha, x_,
+                                     incx, y_, incy, a_);
         });
     });
     return done;
@@ -1499,9 +1519,9 @@ inline sycl::event tbmv(const char *func_name, Func func, sycl::queue &queue, up
             auto a_ = reinterpret_cast<const cuDataType *>(a);
             auto x_ = reinterpret_cast<cuDataType *>(x);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, k,
-                                a_, lda, x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, k, a_, lda, x_, incx);
         });
     });
     return done;
@@ -1539,9 +1559,9 @@ inline sycl::event tbsv(const char *func_name, Func func, sycl::queue &queue, up
             auto a_ = reinterpret_cast<const cuDataType *>(a);
             auto x_ = reinterpret_cast<cuDataType *>(x);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, k,
-                                a_, lda, x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, k, a_, lda, x_, incx);
         });
     });
     return done;
@@ -1578,9 +1598,9 @@ inline sycl::event tpmv(const char *func_name, Func func, sycl::queue &queue, up
             auto a_ = reinterpret_cast<const cuDataType *>(a);
             auto x_ = reinterpret_cast<cuDataType *>(x);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, a_,
-                                x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, a_, x_, incx);
         });
     });
     return done;
@@ -1617,9 +1637,9 @@ inline sycl::event tpsv(const char *func_name, Func func, sycl::queue &queue, up
             auto a_ = reinterpret_cast<const cuDataType *>(a);
             auto x_ = reinterpret_cast<cuDataType *>(x);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, a_,
-                                x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, a_, x_, incx);
         });
     });
     return done;
@@ -1656,9 +1676,9 @@ inline sycl::event trmv(const char *func_name, Func func, sycl::queue &queue, up
             auto a_ = reinterpret_cast<const cuDataType *>(a);
             auto x_ = reinterpret_cast<cuDataType *>(x);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, a_,
-                                lda, x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, a_, lda, x_, incx);
         });
     });
     return done;
@@ -1695,9 +1715,9 @@ inline sycl::event trsv(const char *func_name, Func func, sycl::queue &queue, up
             auto a_ = reinterpret_cast<const cuDataType *>(a);
             auto x_ = reinterpret_cast<cuDataType *>(x);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), get_cublas_diag_type(unit_diag), n, a_,
-                                lda, x_, incx);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), n, a_, lda, x_, incx);
         });
     });
     return done;
diff --git a/src/blas/backends/cublas/cublas_level3.cpp b/src/blas/backends/cublas/cublas_level3.cpp
index 07e48333c..5ea4e2152 100644
--- a/src/blas/backends/cublas/cublas_level3.cpp
+++ b/src/blas/backends/cublas/cublas_level3.cpp
@@ -47,9 +47,9 @@ inline void gemm(const char *func_name, Func func, sycl::queue &queue, transpose
             auto b_ = sc.get_mem<cuDataType *>(b_acc);
             auto c_ = sc.get_mem<cuDataType *>(c_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_operation(transa),
-                                get_cublas_operation(transb), m, n, k, (cuDataType *)&alpha, a_,
-                                lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_operation(transa),
+                                     get_cublas_operation(transb), m, n, k, (cuDataType *)&alpha,
+                                     a_, lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
         });
     });
 }
@@ -94,10 +94,10 @@ inline void gemm_ex(DATATYPE_A DT_A, DATATYPE_B DT_B, DATATYPE_C DT_C, sycl::que
             auto b_ = sc.get_mem<cuDataType_B *>(b_acc);
             auto c_ = sc.get_mem<cuDataType_C *>(c_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC(cublasGemmEx, err, handle, get_cublas_operation(transa),
-                              get_cublas_operation(transb), m, n, k, (cuDataType_C *)&alpha, a_,
-                              DT_A, lda, b_, DT_B, ldb, (cuDataType_C *)&beta, c_, DT_C, ldc, DT_C,
-                              CUBLAS_GEMM_DEFAULT);
+            CUBLAS_ERROR_FUNC_SYNC(cublasGemmEx, err, handle, get_cublas_operation(transa),
+                                   get_cublas_operation(transb), m, n, k, (cuDataType_C *)&alpha,
+                                   a_, DT_A, lda, b_, DT_B, ldb, (cuDataType_C *)&beta, c_, DT_C,
+                                   ldc, DT_C, CUBLAS_GEMM_DEFAULT);
         });
     });
 }
@@ -139,9 +139,9 @@ inline void symm(const char *func_name, Func func, sycl::queue &queue, side left
             auto b_ = sc.get_mem<cuDataType *>(b_acc);
             auto c_ = sc.get_mem<cuDataType *>(c_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_side_mode(left_right),
-                                get_cublas_fill_mode(upper_lower), m, n, (cuDataType *)&alpha, a_,
-                                lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_side_mode(left_right),
+                                     get_cublas_fill_mode(upper_lower), m, n, (cuDataType *)&alpha,
+                                     a_, lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
         });
     });
 }
@@ -178,9 +178,9 @@ inline void hemm(const char *func_name, Func func, sycl::queue &queue, side left
             auto b_ = sc.get_mem<cuDataType *>(b_acc);
             auto c_ = sc.get_mem<cuDataType *>(c_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_side_mode(left_right),
-                                get_cublas_fill_mode(upper_lower), m, n, (cuDataType *)&alpha, a_,
-                                lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_side_mode(left_right),
+                                     get_cublas_fill_mode(upper_lower), m, n, (cuDataType *)&alpha,
+                                     a_, lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
         });
     });
 }
@@ -211,9 +211,10 @@ inline void syrk(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto c_ = sc.get_mem<cuDataType *>(c_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), n, k, (cuDataType *)&alpha, a_, lda,
-                                (cuDataType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     n, k, (cuDataType *)&alpha, a_, lda, (cuDataType *)&beta, c_,
+                                     ldc);
         });
     });
 }
@@ -249,9 +250,10 @@ inline void herk(const char *func_name, Func func, sycl::queue &queue, uplo uppe
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto c_ = sc.get_mem<cuDataType *>(c_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), n, k, (cuScalarType *)&alpha, a_, lda,
-                                (cuScalarType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     n, k, (cuScalarType *)&alpha, a_, lda, (cuScalarType *)&beta,
+                                     c_, ldc);
         });
     });
 }
@@ -286,9 +288,10 @@ inline void syr2k(const char *func_name, Func func, sycl::queue &queue, uplo upp
             auto b_ = sc.get_mem<cuDataType *>(b_acc);
             auto c_ = sc.get_mem<cuDataType *>(c_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), n, k, (cuDataType *)&alpha, a_, lda,
-                                b_, ldb, (cuDataType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     n, k, (cuDataType *)&alpha, a_, lda, b_, ldb,
+                                     (cuDataType *)&beta, c_, ldc);
         });
     });
 }
@@ -325,9 +328,10 @@ inline void her2k(const char *func_name, Func func, sycl::queue &queue, uplo upp
             auto b_ = sc.get_mem<cuDataType *>(b_acc);
             auto c_ = sc.get_mem<cuDataType *>(c_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), n, k, (cuDataType *)&alpha, a_, lda,
-                                b_, ldb, (cuScalarType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     n, k, (cuDataType *)&alpha, a_, lda, b_, ldb,
+                                     (cuScalarType *)&beta, c_, ldc);
         });
     });
 }
@@ -364,10 +368,10 @@ inline void trmm(const char *func_name, Func func, sycl::queue &queue, side left
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto b_ = sc.get_mem<cuDataType *>(b_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_side_mode(left_right),
-                                get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
-                                get_cublas_diag_type(unit_diag), m, n, (cuDataType *)&alpha, a_,
-                                lda, b_, ldb, b_, ldb);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_side_mode(left_right),
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), m, n, (cuDataType *)&alpha,
+                                     a_, lda, b_, ldb, b_, ldb);
         });
     });
 }
@@ -400,10 +404,10 @@ inline void trsm(const char *func_name, Func func, sycl::queue &queue, side left
             auto a_ = sc.get_mem<cuDataType *>(a_acc);
             auto b_ = sc.get_mem<cuDataType *>(b_acc);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_side_mode(left_right),
-                                get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
-                                get_cublas_diag_type(unit_diag), m, n, (cuDataType *)&alpha, a_,
-                                lda, b_, ldb);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_side_mode(left_right),
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), m, n, (cuDataType *)&alpha,
+                                     a_, lda, b_, ldb);
         });
     });
 }
@@ -442,9 +446,9 @@ inline sycl::event gemm(const char *func_name, Func func, sycl::queue &queue, tr
             auto b_ = reinterpret_cast<const cuDataType *>(b);
             auto c_ = reinterpret_cast<cuDataType *>(c);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_operation(transa),
-                                get_cublas_operation(transb), m, n, k, (cuDataType *)&alpha, a_,
-                                lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_operation(transa),
+                                     get_cublas_operation(transb), m, n, k, (cuDataType *)&alpha,
+                                     a_, lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
         });
     });
     return done;
@@ -488,10 +492,10 @@ inline sycl::event gemm_ex_usm(DATATYPE_A DT_A, DATATYPE_B DT_B, DATATYPE_C DT_C
             auto b_ = reinterpret_cast<const cuDataType_B *>(b);
             auto c_ = reinterpret_cast<cuDataType_C *>(c);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC(cublasGemmEx, err, handle, get_cublas_operation(transa),
-                              get_cublas_operation(transb), m, n, k, (cuDataType_C *)&alpha, a_,
-                              DT_A, lda, b_, DT_B, ldb, (cuDataType_C *)&beta, c_, DT_C, ldc, DT_C,
-                              CUBLAS_GEMM_DEFAULT);
+            CUBLAS_ERROR_FUNC_SYNC(cublasGemmEx, err, handle, get_cublas_operation(transa),
+                                   get_cublas_operation(transb), m, n, k, (cuDataType_C *)&alpha,
+                                   a_, DT_A, lda, b_, DT_B, ldb, (cuDataType_C *)&beta, c_, DT_C,
+                                   ldc, DT_C, CUBLAS_GEMM_DEFAULT);
         });
     });
     return done;
@@ -537,9 +541,9 @@ inline sycl::event symm(const char *func_name, Func func, sycl::queue &queue, si
             auto b_ = reinterpret_cast<const cuDataType *>(b);
             auto c_ = reinterpret_cast<cuDataType *>(c);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_side_mode(left_right),
-                                get_cublas_fill_mode(upper_lower), m, n, (cuDataType *)&alpha, a_,
-                                lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_side_mode(left_right),
+                                     get_cublas_fill_mode(upper_lower), m, n, (cuDataType *)&alpha,
+                                     a_, lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
         });
     });
     return done;
@@ -579,9 +583,9 @@ inline sycl::event hemm(const char *func_name, Func func, sycl::queue &queue, si
             auto b_ = reinterpret_cast<const cuDataType *>(b);
             auto c_ = reinterpret_cast<cuDataType *>(c);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_side_mode(left_right),
-                                get_cublas_fill_mode(upper_lower), m, n, (cuDataType *)&alpha, a_,
-                                lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_side_mode(left_right),
+                                     get_cublas_fill_mode(upper_lower), m, n, (cuDataType *)&alpha,
+                                     a_, lda, b_, ldb, (cuDataType *)&beta, c_, ldc);
         });
     });
     return done;
@@ -616,9 +620,10 @@ inline sycl::event syrk(const char *func_name, Func func, sycl::queue &queue, up
             auto a_ = reinterpret_cast<const cuDataType *>(a);
             auto c_ = reinterpret_cast<cuDataType *>(c);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), n, k, (cuDataType *)&alpha, a_, lda,
-                                (cuDataType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     n, k, (cuDataType *)&alpha, a_, lda, (cuDataType *)&beta, c_,
+                                     ldc);
         });
     });
     return done;
@@ -657,9 +662,10 @@ inline sycl::event herk(const char *func_name, Func func, sycl::queue &queue, up
             auto a_ = reinterpret_cast<const cuDataType *>(a);
             auto c_ = reinterpret_cast<cuDataType *>(c);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), n, k, (cuScalarType *)&alpha, a_, lda,
-                                (cuScalarType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     n, k, (cuScalarType *)&alpha, a_, lda, (cuScalarType *)&beta,
+                                     c_, ldc);
         });
     });
     return done;
@@ -697,9 +703,10 @@ inline sycl::event syr2k(const char *func_name, Func func, sycl::queue &queue, u
             auto b_ = reinterpret_cast<const cuDataType *>(b);
             auto c_ = reinterpret_cast<cuDataType *>(c);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), n, k, (cuDataType *)&alpha, a_, lda,
-                                b_, ldb, (cuDataType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     n, k, (cuDataType *)&alpha, a_, lda, b_, ldb,
+                                     (cuDataType *)&beta, c_, ldc);
         });
     });
     return done;
@@ -740,9 +747,10 @@ inline sycl::event her2k(const char *func_name, Func func, sycl::queue &queue, u
             auto b_ = reinterpret_cast<const cuDataType *>(b);
             auto c_ = reinterpret_cast<cuDataType *>(c);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_fill_mode(upper_lower),
-                                get_cublas_operation(trans), n, k, (cuDataType *)&alpha, a_, lda,
-                                b_, ldb, (cuScalarType *)&beta, c_, ldc);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle,
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     n, k, (cuDataType *)&alpha, a_, lda, b_, ldb,
+                                     (cuScalarType *)&beta, c_, ldc);
         });
     });
     return done;
@@ -783,10 +791,10 @@ inline sycl::event trmm(const char *func_name, Func func, sycl::queue &queue, si
             auto a_ = reinterpret_cast<const cuDataType *>(a);
             auto b_ = reinterpret_cast<cuDataType *>(b);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_side_mode(left_right),
-                                get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
-                                get_cublas_diag_type(unit_diag), m, n, (cuDataType *)&alpha, a_,
-                                lda, b_, ldb, b_, ldb);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_side_mode(left_right),
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), m, n, (cuDataType *)&alpha,
+                                     a_, lda, b_, ldb, b_, ldb);
         });
     });
     return done;
@@ -823,10 +831,10 @@ inline sycl::event trsm(const char *func_name, Func func, sycl::queue &queue, si
             auto a_ = reinterpret_cast<const cuDataType *>(a);
             auto b_ = reinterpret_cast<cuDataType *>(b);
             cublasStatus_t err;
-            CUBLAS_ERROR_FUNC_T(func_name, func, err, handle, get_cublas_side_mode(left_right),
-                                get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
-                                get_cublas_diag_type(unit_diag), m, n, (cuDataType *)&alpha, a_,
-                                lda, b_, ldb);
+            CUBLAS_ERROR_FUNC_T_SYNC(func_name, func, err, handle, get_cublas_side_mode(left_right),
+                                     get_cublas_fill_mode(upper_lower), get_cublas_operation(trans),
+                                     get_cublas_diag_type(unit_diag), m, n, (cuDataType *)&alpha,
+                                     a_, lda, b_, ldb);
         });
     });
     return done;