add info

3 files changed, 12 insertions, 79 deletions

diff --git a/backend/src/driver/cl_gen_driver.h b/backend/src/driver/cl_gen_driver.h
index 5e475b17..82c344e3 100644
--- a/backend/src/driver/cl_gen_driver.h
+++ b/backend/src/driver/cl_gen_driver.h
@@ -57,8 +57,7 @@ cl_int GenSupportImageFmt(const cl_device_id device, cl_mem_object_type image_ty
 cl_int GenCreateBuffer(cl_mem mem, const cl_device_id device);
 cl_int GenReleaseMem(cl_mem mem, const cl_device_id device);
 cl_int GenEnqueueMapBuffer(cl_command_queue queue, cl_mem mem, void** ret_addr, cl_bool block,
-                           cl_map_flags flags, size_t offset, size_t size, cl_uint num_events,
-                           const cl_event *event_list, cl_event event_ret);
+                           cl_map_flags flags, size_t offset, size_t size, cl_command_queue_work_item item);
 cl_int GenEnqueueUnmapMem(cl_command_queue queue, cl_mem mem, void *mapped_ptr,
                           cl_uint num_events, const cl_event *event_list, cl_event event_ret);
 cl_int GenEnqueueNDRangeKernel(cl_command_queue queue, cl_kernel kernel, const uint32_t work_dim,
diff --git a/backend/src/driver/cl_gen_driver.hpp b/backend/src/driver/cl_gen_driver.hpp
index 6f880174..39607c16 100644
--- a/backend/src/driver/cl_gen_driver.hpp
+++ b/backend/src/driver/cl_gen_driver.hpp
@@ -254,6 +254,7 @@ struct GenGPUCommandQueue {
 struct GenGPUMem {
   drm_intel_bo *bo;
   GenGPUContext* gpuCtx;
+  bool bindUserPtr;
   void* alignedHostPtr;
   void* mappedAddr;
   bool writeMap;
@@ -262,8 +263,8 @@ struct GenGPUMem {
   bool mappedGtt;
   volatile int mapRef;
   pthread_mutex_t mutex;
-  GenGPUMem(GenGPUContext* ctx) : bo(NULL), gpuCtx(ctx), alignedHostPtr(NULL), mappedAddr(NULL),
-    writeMap(false), realSize(0), tiling(GPGPU_NO_TILE), mappedGtt(false), mapRef(0) {
+  GenGPUMem(GenGPUContext* ctx) : bo(NULL), gpuCtx(ctx), bindUserPtr(false), alignedHostPtr(NULL),
+    mappedAddr(NULL), writeMap(false), realSize(0), tiling(GPGPU_NO_TILE), mappedGtt(false), mapRef(0) {
     pthread_mutex_init(&this->mutex, NULL);
   }
   ~GenGPUMem(void);
diff --git a/backend/src/driver/cl_gen_mem.cpp b/backend/src/driver/cl_gen_mem.cpp
index 49975000..f69a9f5d 100644
--- a/backend/src/driver/cl_gen_mem.cpp
+++ b/backend/src/driver/cl_gen_mem.cpp
@@ -442,13 +442,17 @@ bool GenGPUWorkItemMapBuf::submit(void)
 
 extern "C"
 cl_int GenEnqueueMapBuffer(cl_command_queue queue, cl_mem mem, void** ret_addr, cl_bool block,
-                           cl_map_flags flags, size_t offset, size_t size, cl_uint num_events,
-                           const cl_event *event_list, cl_event event_ret)
+                           cl_map_flags flags, size_t offset, size_t size, cl_command_queue_work_item item)
 {
+  assert(item);
+
   if (mem->type != CL_MEM_OBJECT_BUFFER) {
     return CL_INVALID_VALUE;
   }
+
   GenGPUMem* genMem = reinterpret_cast<GenGPUMem*>(getGenMemPrivate(mem, queue->device));
+  GenGPUCommandQueue* gpuQueue = (GenGPUCommandQueue*)getGenCommandQueuePrivate(queue);
+  GBE_ASSERT(gpuQueue);
   void* retAddr = NULL;
 
   if (genMem == NULL)
@@ -458,90 +462,19 @@ cl_int GenEnqueueMapBuffer(cl_command_queue queue, cl_mem mem, void** ret_addr,
     return CL_MEM_OBJECT_ALLOCATION_FAILURE;
   }
 
-  GenGPUCommandQueue* gpuQueue = (GenGPUCommandQueue*)getGenCommandQueuePrivate(queue);
-  GBE_ASSERT(gpuQueue);
-
-  if (block) {
-    /* According to spec, when in block mode, we need to ensure all the
-       commands in queue are flushed. */
-    gpuQueue->waitForFlush();
-
-    if (event_list) { // Need to wait for events.
-      if (gpuQueue->waitForEvents(event_list, num_events) == false) {
-        /* canceled or some errors. */
-        return CL_MAP_FAILURE;
-      }
-    }
-
+  if (block || item->depend_events == NULL) { // We do not need to do it async, just map it.
     retAddr = genDoMapBuffer(genMem, mem, flags, offset, size);
     if (retAddr == NULL) {
-      if (event_ret) {
-        event_ret->set_status(event_ret, -1); // Set error for that event.
-      }
-      return CL_MAP_FAILURE;
-    }
-
-    if (ret_addr)
-      *ret_addr = retAddr;
-
-    if (event_ret) {
-      event_ret->set_status(event_ret, CL_COMPLETE);
-    }
-    return CL_SUCCESS;
-  } else if (event_list == NULL) {
-    /* We do not have any events to wait, map it in sync mode. */
-    retAddr = genDoMapBuffer(genMem, mem, flags, offset, size);
-    if (retAddr == NULL) {
-      if (event_ret) {
-        event_ret->set_status(event_ret, -1); // Set error for that event.
-      }
-      return CL_MAP_FAILURE;
-    }
-
-    if (ret_addr)
-      *ret_addr = retAddr;
-
-    if (event_ret) {
-      event_ret->set_status(event_ret, CL_COMPLETE);
-    }
-    return CL_SUCCESS;
-  } else {
-    GBE_ASSERT(num_events > 0);
-    /* A awkward case, we have events to wait, so we can not do the map operation sync.
-       But the spec require us to return the mapped address immediately. We can just
-       allocate a host mem and return that address. After really mapped, we then
-       copy back the data. */
-    void* addr = NULL;
-    if (!mem->flags & CL_MEM_USE_HOST_PTR) {
-      addr = GBE_MALLOC(size);
-      if (addr == NULL)
-        return CL_OUT_OF_HOST_MEMORY;
-    }
-
-    GenGPUWorkItemMapBuf* mapItem = GBE_NEW(GenGPUWorkItemMapBuf, mem, genMem, addr, flags,
-                                            offset, size, event_ret, event_list, num_events);
-    if (mapItem == NULL) {
-      if (addr)
-        GBE_FREE(addr);
-      return CL_OUT_OF_HOST_MEMORY;
-    }
-
-    /* Final,enqueue it in the queue worker thread. */
-    if (gpuQueue->enqueueWorkItem(mapItem) == false) {
-      GBE_DELETE(mapItem);
-      if (addr)
-        GBE_FREE(addr);
       return CL_MAP_FAILURE;
     }
 
     if (ret_addr)
       *ret_addr = retAddr;
 
+    item->status = CL_COMPLETE;
     return CL_SUCCESS;
   }
 
-  // Never come to here.
-  GBE_ASSERT(0);
   return CL_MAP_FAILURE;
 }