diff --git a/src/acc/acc_libsmm.h b/src/acc/acc_libsmm.h
index 06957d74074..94428d598e5 100644
--- a/src/acc/acc_libsmm.h
+++ b/src/acc/acc_libsmm.h
@@ -15,11 +15,11 @@
 #define DBCSR_TYPE_double dbcsr_type_real_8
 #define DBCSR_TYPE_float dbcsr_type_real_4
 
-#define LIBSMM_ACC_TRANSPOSE_ROUTINE_NAME_STRPTR ((const char**)&libsmm_acc_transpose_routine_name_ptr)
+#define LIBSMM_ACC_TRANSPOSE_ROUTINE_NAME_STRPTR ((const char**)((uintptr_t)&libsmm_acc_transpose_routine_name_str))
 #define LIBSMM_ACC_TRANSPOSE_ROUTINE_NAME_LENPTR (&libsmm_acc_transpose_routine_name_len)
 #define LIBSMM_ACC_TRANSPOSE_ROUTINE_NAME_STR (libsmm_acc_transpose_routine_name_str)
 
-#define LIBSMM_ACC_PROCESS_ROUTINE_NAME_STRPTR ((const char**)&libsmm_acc_process_routine_name_ptr)
+#define LIBSMM_ACC_PROCESS_ROUTINE_NAME_STRPTR ((const char**)((uintptr_t)&libsmm_acc_process_routine_name_str))
 #define LIBSMM_ACC_PROCESS_ROUTINE_NAME_LENPTR (&libsmm_acc_process_routine_name_len)
 #define LIBSMM_ACC_PROCESS_ROUTINE_NAME_STR (libsmm_acc_process_routine_name_str)
 
@@ -49,11 +49,9 @@ int libsmm_acc_process(const int* host_param_stack, const int* dev_param_stack,
 int c_calculate_norms(const double* mat, int nblks, const int* offsets, const int* nelems, float* norms, void* stream_ptr);
 
 static const char libsmm_acc_transpose_routine_name_str[] = "jit_kernel_transpose";
-static const char* const libsmm_acc_transpose_routine_name_ptr = libsmm_acc_transpose_routine_name_str;
 static const int libsmm_acc_transpose_routine_name_len = (int)sizeof(libsmm_acc_transpose_routine_name_str) - 1;
 
 static const char libsmm_acc_process_routine_name_str[] = "jit_kernel_multiply";
-static const char* const libsmm_acc_process_routine_name_ptr = libsmm_acc_process_routine_name_str;
 static const int libsmm_acc_process_routine_name_len = (int)sizeof(libsmm_acc_process_routine_name_str) - 1;
 
 #if defined(__cplusplus)
diff --git a/src/acc/cuda/Makefile b/src/acc/cuda/Makefile
index 2aedadeb979..6f6c66b2369 100644
--- a/src/acc/cuda/Makefile
+++ b/src/acc/cuda/Makefile
@@ -103,23 +103,15 @@ ifneq (,$(ELEM_TYPE))
   DFLAGS += -DELEM_TYPE=$(ELEM_TYPE)
 endif
 
-ifeq (1,$(INTEL))
-  CXX := icpc
-  CC := icc
-  AR := xiar
-else ifneq (0,$(INTEL))
-  CXX := icpx
-  CC := icx
-  AR := xiar
-else ifneq (0,$(GNU))
-  override CXX := g++
-  override CC := gcc
-  ifneq (Darwin,$(UNAME))
-    override AR := gcc-ar
+ifneq (0,$(INTEL))
+  ifneq (1,$(INTEL))
+    CXX := icpx
+    CC := icx
   else
-    override AR := ar
+    CXX := icpc
+    CC := icc
   endif
-  #override LD_LIBRARY_DIRS := $(NULL)
+  AR := $(if $(call which,xiar),xiar,ar)
 else
   CXX := g++
   CC := gcc
diff --git a/src/acc/opencl/Makefile b/src/acc/opencl/Makefile
index 0cbefd97d69..244a7b2692f 100644
--- a/src/acc/opencl/Makefile
+++ b/src/acc/opencl/Makefile
@@ -73,23 +73,15 @@ ifneq (,$(ELEM_TYPE))
   CFLAGS += -DELEM_TYPE=$(ELEM_TYPE)
 endif
 
-ifeq (1,$(INTEL))
-  CXX := icpc
-  CC := icc
-  AR := xiar
-else ifneq (0,$(INTEL))
-  CXX := icpx
-  CC := icx
-  AR := xiar
-else ifneq (0,$(GNU))
-  override CXX := g++
-  override CC := gcc
-  ifneq (Darwin,$(UNAME))
-    override AR := gcc-ar
+ifneq (0,$(INTEL))
+  ifneq (1,$(INTEL))
+    CXX := icpx
+    CC := icx
   else
-    override AR := ar
+    CXX := icpc
+    CC := icc
   endif
-  #override LD_LIBRARY_DIRS := $(NULL)
+  AR := $(if $(call which,xiar),xiar,ar)
 else
   CXX := g++
   CC := gcc
@@ -109,13 +101,14 @@ ifneq (0,$(DEV))
     CFLAGS += -D__DBCSR_ACC
     CFLAGS += -Wno-deprecated -Werror
     ifneq (2,$(DEV))
+      $(info DEBUG: $(CC) $(CXX))
       ifneq (,$(findstring clang,$(CC) $(CXX)))
         override CC := clang++ --analyze
       else
-        CC := $(CXX) -xc++
+        override CC := $(CXX) -xc++
       endif
     else
-      CC := $(CXX) -xc++
+      override CC := $(CXX) -xc++
     endif
     $(info CC: $(shell $(CC) --version | head -n1))
     OMP := 0
diff --git a/src/acc/opencl/acc_opencl.c b/src/acc/opencl/acc_opencl.c
index 57e1f050f0b..cab7e358cfb 100644
--- a/src/acc/opencl/acc_opencl.c
+++ b/src/acc/opencl/acc_opencl.c
@@ -380,8 +380,8 @@ int c_dbcsr_acc_init(void) {
     assert(EXIT_SUCCESS == result);
 #  endif
     if (~(1 + 2) & c_dbcsr_acc_opencl_config.wa) { /* environment is populated before touching the compute runtime */
-      static char* key_value[] = {"NEOReadDebugKeys=1", "ZE_FLAT_DEVICE_HIERARCHY=COMPOSITE", "EnableRecoverablePageFaults=0",
-        "DirectSubmissionOverrideBlitterSupport=0"};
+      static char a[] = "NEOReadDebugKeys=1", b[] = "ZE_FLAT_DEVICE_HIERARCHY=COMPOSITE", c[] = "EnableRecoverablePageFaults=0";
+      static char d[] = "DirectSubmissionOverrideBlitterSupport=0", *key_value[] = {a, b, c, d};
       if (NULL == env_neo) ACC_OPENCL_EXPECT(0 == LIBXSMM_PUTENV(key_value[0]));
       if ((4 & c_dbcsr_acc_opencl_config.wa) && NULL == getenv("ZE_FLAT_DEVICE_HIERARCHY")) {
         ACC_OPENCL_EXPECT(0 == LIBXSMM_PUTENV(key_value[1]));
diff --git a/src/acc/opencl/acc_opencl.h b/src/acc/opencl/acc_opencl.h
index fadcf20f976..ddd1834051f 100644
--- a/src/acc/opencl/acc_opencl.h
+++ b/src/acc/opencl/acc_opencl.h
@@ -124,7 +124,7 @@
 #  define ACC_OPENCL_ACTIVATE 0
 #endif
 /* Use DBCSR's profile for detailed timings */
-#if !defined(ACC_OPENCL_PROFILE) && 0
+#if !defined(ACC_OPENCL_PROFILE) && (defined(__OFFLOAD_PROFILING) || 0)
 #  define ACC_OPENCL_PROFILE
 #endif
 
@@ -362,7 +362,7 @@ typedef struct c_dbcsr_acc_opencl_config_t {
 extern c_dbcsr_acc_opencl_config_t c_dbcsr_acc_opencl_config;
 
 /** Determines host-pointer registration for modification. */
-c_dbcsr_acc_opencl_info_memptr_t* c_dbcsr_acc_opencl_info_hostptr(void* memory);
+c_dbcsr_acc_opencl_info_memptr_t* c_dbcsr_acc_opencl_info_hostptr(const void* memory);
 /** Determines device-pointer registration for modification (internal). */
 c_dbcsr_acc_opencl_info_memptr_t* c_dbcsr_acc_opencl_info_devptr_modify(
   ACC_OPENCL_LOCKTYPE* lock, void* memory, size_t elsize, const size_t* amount, size_t* offset);
diff --git a/src/acc/opencl/acc_opencl_mem.c b/src/acc/opencl/acc_opencl_mem.c
index 8a218634057..7d9bd86a4d0 100644
--- a/src/acc/opencl/acc_opencl_mem.c
+++ b/src/acc/opencl/acc_opencl_mem.c
@@ -60,7 +60,7 @@ void c_dbcsr_acc_opencl_pfree(ACC_OPENCL_LOCKTYPE* lock, const void* pointer, vo
 }
 
 
-c_dbcsr_acc_opencl_info_memptr_t* c_dbcsr_acc_opencl_info_hostptr(void* memory) {
+c_dbcsr_acc_opencl_info_memptr_t* c_dbcsr_acc_opencl_info_hostptr(const void* memory) {
   assert(NULL == memory || sizeof(c_dbcsr_acc_opencl_info_memptr_t) <= (uintptr_t)memory);
   return (NULL != memory ? (c_dbcsr_acc_opencl_info_memptr_t*)((uintptr_t)memory - sizeof(c_dbcsr_acc_opencl_info_memptr_t))
                          : (c_dbcsr_acc_opencl_info_memptr_t*)NULL);
@@ -291,6 +291,7 @@ int c_dbcsr_acc_opencl_memcpy_d2h(
   int result = EXIT_SUCCESS;
 #  if defined(ACC_OPENCL_MEM_DEVPTR)
   if (NULL != c_dbcsr_acc_opencl_config.device.clEnqueueMemcpyINTEL) {
+    assert(0 == c_dbcsr_acc_opencl_config.device.unified);
     result = c_dbcsr_acc_opencl_config.device.clEnqueueMemcpyINTEL(queue, finish, host_mem, dev_mem, nbytes, 0, NULL, NULL);
   }
   else
@@ -302,6 +303,7 @@ int c_dbcsr_acc_opencl_memcpy_d2h(
     int result_sync = EXIT_SUCCESS;
 #  if defined(ACC_OPENCL_MEM_DEVPTR)
     if (NULL != c_dbcsr_acc_opencl_config.device.clEnqueueMemcpyINTEL) {
+      assert(0 == c_dbcsr_acc_opencl_config.device.unified);
       result_sync = c_dbcsr_acc_opencl_config.device.clEnqueueMemcpyINTEL(queue, CL_TRUE, host_mem, dev_mem, nbytes, 0, NULL, NULL);
     }
     else
@@ -343,6 +345,7 @@ int c_dbcsr_acc_dev_mem_allocate(void** dev_mem, size_t nbytes) {
   assert(NULL != dev_mem && NULL != context);
 #  if defined(ACC_OPENCL_MEM_DEVPTR)
   if (NULL != c_dbcsr_acc_opencl_config.device.clDeviceMemAllocINTEL) {
+    assert(0 == c_dbcsr_acc_opencl_config.device.unified);
     *dev_mem = memptr = c_dbcsr_acc_opencl_config.device.clDeviceMemAllocINTEL(
       context, c_dbcsr_acc_opencl_config.device.id, NULL /*properties*/, nbytes, 0 /*alignment*/, &result);
     if (EXIT_SUCCESS != result) *dev_mem = NULL;
@@ -436,6 +439,7 @@ int c_dbcsr_acc_dev_mem_deallocate(void* dev_mem) {
 #  else
     assert(NULL != c_dbcsr_acc_opencl_config.device.context);
     if (NULL != c_dbcsr_acc_opencl_config.device.clMemFreeINTEL) {
+      assert(0 == c_dbcsr_acc_opencl_config.device.unified);
       result = c_dbcsr_acc_opencl_config.device.clMemFreeINTEL(c_dbcsr_acc_opencl_config.device.context, dev_mem);
     }
     else {
@@ -513,6 +517,7 @@ int c_dbcsr_acc_memcpy_h2d(const void* host_mem, void* dev_mem, size_t nbytes, v
     assert(NULL != str && NULL != str->queue);
 #  if defined(ACC_OPENCL_MEM_DEVPTR)
     if (NULL != c_dbcsr_acc_opencl_config.device.clEnqueueMemcpyINTEL) {
+      assert(0 == c_dbcsr_acc_opencl_config.device.unified);
       result = c_dbcsr_acc_opencl_config.device.clEnqueueMemcpyINTEL(str->queue, finish, dev_mem, host_mem, nbytes, 0, NULL, NULL);
     }
     else
@@ -592,6 +597,7 @@ int c_dbcsr_acc_memcpy_d2d(const void* devmem_src, void* devmem_dst, size_t nbyt
 #  if defined(ACC_OPENCL_MEM_DEVPTR)
     assert(NULL != c_dbcsr_acc_opencl_config.device.context);
     if (NULL != c_dbcsr_acc_opencl_config.device.clEnqueueMemcpyINTEL) {
+      assert(0 == c_dbcsr_acc_opencl_config.device.unified);
       result = c_dbcsr_acc_opencl_config.device.clEnqueueMemcpyINTEL(
         str->queue, CL_FALSE /*blocking*/, devmem_dst, devmem_src, nbytes, 0, NULL, &event);
     }
@@ -642,6 +648,7 @@ int c_dbcsr_acc_opencl_memset(void* dev_mem, int value, size_t offset, size_t nb
 #  if defined(ACC_OPENCL_MEM_DEVPTR)
     assert(NULL != c_dbcsr_acc_opencl_config.device.context);
     if (NULL != c_dbcsr_acc_opencl_config.device.clEnqueueMemFillINTEL) {
+      assert(0 == c_dbcsr_acc_opencl_config.device.unified);
       result = c_dbcsr_acc_opencl_config.device.clEnqueueMemFillINTEL(
         str->queue, (char*)dev_mem + offset, &value, size_of_value, nbytes, 0, NULL, &event);
     }