1 files changed, 11 insertions, 7 deletions

diff --git a/backend/src/driver/cl_gen_mem.cpp b/backend/src/driver/cl_gen_mem.cpp
index 3d8cd219..ea263662 100644
--- a/backend/src/driver/cl_gen_mem.cpp
+++ b/backend/src/driver/cl_gen_mem.cpp
@@ -387,7 +387,7 @@ cl_int GenReleaseMem(cl_mem mem, const cl_device_id device)
   return CL_SUCCESS;
 }
 
-static void* genDoMapBuffer(GenGPUMem* genMem, cl_mem mem, cl_map_flags flags, size_t offset, size_t size)
+static void* genDoMapBufferSync(GenGPUMem* genMem, cl_mem mem, cl_map_flags flags, size_t offset, size_t size)
 {
   cl_mem_buffer buffer = cl_mem_to_buffer(mem);
 
@@ -401,16 +401,20 @@ static void* genDoMapBuffer(GenGPUMem* genMem, cl_mem mem, cl_map_flags flags, s
     return NULL;
   }
 
-  if (mem->flags & CL_MEM_USE_HOST_PTR) {
-    retAddr = (char*)mem->host_ptr + buffer->sub_offset + offset;
+  if (mem->flags & (CL_MEM_USE_HOST_PTR|CL_MEM_ALLOC_HOST_PTR)) {
+    if (genMem->bindedHostPtr) {
+      GBE_ASSERT(genMem->bindedHostPtr == mem->host_ptr - buffer->sub_offset);
+    }
+    retAddr = (char*)mem->host_ptr + offset;
   } else {
     retAddr = (char*)mappedAddr + buffer->sub_offset + offset;
   }
 
   /* We might need to call back the data in GPU side to host. */
-  if ((mem->flags & CL_MEM_USE_HOST_PTR) && genMem->alignedHostPtr == NULL) {
+  if ((mem->flags & (CL_MEM_USE_HOST_PTR|CL_MEM_ALLOC_HOST_PTR))
+      && genMem->bindedHostPtr == NULL) {
     char* ptr = (char*)mappedAddr + buffer->sub_offset + offset;
-    memcpy((char*)mem->host_ptr + buffer->sub_offset + offset, ptr, size);
+    memcpy(retAddr, ptr, size);
   }
 
   return (void*)retAddr;
@@ -442,7 +446,7 @@ bool GenGPUWorkItemMapBuf::complete(void)
 
 bool GenGPUWorkItemMapBuf::submit(void)
 {
-  void* retAddr = genDoMapBuffer(genMem, mem, flags, offset, size);
+  void* retAddr = genDoMapBufferSync(genMem, mem, flags, offset, size);
   if (retAddr == NULL) {
     return false;
   }
@@ -477,7 +481,7 @@ cl_int GenEnqueueMapBuffer(cl_command_queue queue, cl_mem mem, void** ret_addr,
   }
 
   if (block || item->depend_events == NULL) { // We do not need to do it async, just map it.
-    retAddr = genDoMapBuffer(genMem, mem, flags, offset, size);
+    retAddr = genDoMapBufferSync(genMem, mem, flags, offset, size);
     if (retAddr == NULL) {
       return CL_MAP_FAILURE;
     }