From 5364cf1093c53f04f22f650bedbca1a5b27fe423 Mon Sep 17 00:00:00 2001
From: Junyan He <junyan.he@intel.com>
Date: Tue, 12 Apr 2016 15:37:50 +0800
Subject: add finish

---
 backend/src/driver/cl_gen_command_queue.cpp | 79 +++++++++++++++++++++++++++--
 backend/src/driver/cl_gen_driver.c          |  2 +
 backend/src/driver/cl_gen_driver.h          |  2 +
 backend/src/driver/cl_gen_driver.hpp        |  2 +
 4 files changed, 82 insertions(+), 3 deletions(-)

(limited to 'backend/src/driver')

diff --git a/backend/src/driver/cl_gen_command_queue.cpp b/backend/src/driver/cl_gen_command_queue.cpp
index 5ee7730e..1616487e 100644
--- a/backend/src/driver/cl_gen_command_queue.cpp
+++ b/backend/src/driver/cl_gen_command_queue.cpp
@@ -178,15 +178,63 @@ void GenGPUCommandQueue::waitOnEvent(GenGPUWorkItem* item)
   }
 
   while (this->quit == false) {
-    if ((item && item->state == CL_COMPLETE) ||
-        (!item && (this->workItems.empty() || this->inExec == true))) {
+    if (item && item->state != CL_COMPLETE) {
       pthread_cond_wait(&this->cond, &this->mutex);
+    } else {
+      break;
     }
   }
 
   pthread_mutex_unlock(&this->mutex);
 }
 
+void GenGPUCommandQueue::waitForFlush(void)
+{
+  pthread_mutex_lock(&this->mutex);
+  if (this->quit) { // already destroy the queue?
+    pthread_mutex_unlock(&this->mutex);
+    return;
+  }
+
+  bool needWait = false;
+  while (this->quit == false) {
+    needWait = false;
+    for (list<GenGPUWorkItem*>::iterator it = this->workItems.begin();
+         it != this->workItems.end(); it++) {
+      if ((*it)->state > CL_SUBMITTED) {
+        needWait = true;
+        break;
+      }
+    }
+
+    if (needWait) {
+      pthread_cond_wait(&this->cond, &this->mutex);
+    } else {
+      break;
+    }
+  }
+
+  pthread_mutex_unlock(&this->mutex);
+}
+
+void GenGPUCommandQueue::waitForFinish(void)
+{
+  pthread_mutex_lock(&this->mutex);
+  if (this->quit) { // already destroy the queue?
+    pthread_mutex_unlock(&this->mutex);
+    return;
+  }
+
+  while (this->quit == false) {
+    if (this->workItems.empty() && this->inExec == false)
+      break;
+
+    pthread_cond_wait(&this->cond, &this->mutex);
+  }
+
+  pthread_mutex_unlock(&this->mutex);
+}
+
 void GenGPUCommandQueue::userEventChange(void)
 {
   pthread_mutex_lock(&this->mutex);
@@ -263,10 +311,35 @@ cl_int GenReleaseCommandQueue(cl_command_queue queue)
 {
   GenGPUCommandQueue* gpuQueue = (GenGPUCommandQueue*)getGenCommandQueuePrivate(queue);
   if (gpuQueue == NULL) {
-    return CL_INVALID_VALUE;
+    return CL_INVALID_COMMAND_QUEUE;
   }
 
   GBE_DELETE(gpuQueue);
   setGenCommandQueuePrivate(queue, NULL);
   return CL_SUCCESS;
 }
+
+extern "C"
+cl_int GenFlushCommandQueue(cl_command_queue queue)
+{
+  GenGPUCommandQueue* gpuQueue = (GenGPUCommandQueue*)getGenCommandQueuePrivate(queue);
+  if (gpuQueue == NULL) {
+    return CL_INVALID_COMMAND_QUEUE;
+  }
+
+  gpuQueue->waitForFlush();
+  return CL_SUCCESS;
+}
+
+extern "C"
+cl_int GenFinishCommandQueue(cl_command_queue queue)
+{
+  GenGPUCommandQueue* gpuQueue = (GenGPUCommandQueue*)getGenCommandQueuePrivate(queue);
+  if (gpuQueue == NULL) {
+    return CL_INVALID_COMMAND_QUEUE;
+  }
+
+  gpuQueue->waitForFinish();
+  return CL_SUCCESS;
+}
+
diff --git a/backend/src/driver/cl_gen_driver.c b/backend/src/driver/cl_gen_driver.c
index 00cb8070..c8f962f7 100644
--- a/backend/src/driver/cl_gen_driver.c
+++ b/backend/src/driver/cl_gen_driver.c
@@ -25,6 +25,8 @@ _cl_driver clgenDriver = {
   .release_context = GenReleaseContext,
   .create_command_queue = GenCreateCommandQueue,
   .release_command_queue = GenReleaseCommandQueue,
+  .flush_command_queue = GenFlushCommandQueue,
+  .finish_command_queue = GenFinishCommandQueue,
   .build_program = GenBuildProgram,
   .get_program_kernel_names = GenGetProgramKernelNames,
   .release_program = GenReleaseProgram,
diff --git a/backend/src/driver/cl_gen_driver.h b/backend/src/driver/cl_gen_driver.h
index 7d66103b..ba572a42 100644
--- a/backend/src/driver/cl_gen_driver.h
+++ b/backend/src/driver/cl_gen_driver.h
@@ -34,6 +34,8 @@ cl_int GenCreateContext(cl_context context, const cl_device_id device);
 cl_int GenReleaseContext(cl_context context, const cl_device_id device);
 cl_int GenCreateCommandQueue(cl_command_queue queue);
 cl_int GenReleaseCommandQueue(cl_command_queue queue);
+cl_int GenFlushCommandQueue(cl_command_queue queue);
+cl_int GenFinishCommandQueue(cl_command_queue queue);
 cl_int GenBuildProgram(cl_program program, const cl_device_id device);
 cl_int GenGetProgramKernelNames(cl_program program, const cl_device_id device, char *names,
                                 cl_uint name_sz, cl_uint* ret_sz, cl_uint* ker_num);
diff --git a/backend/src/driver/cl_gen_driver.hpp b/backend/src/driver/cl_gen_driver.hpp
index a671f38b..ff58c7d3 100644
--- a/backend/src/driver/cl_gen_driver.hpp
+++ b/backend/src/driver/cl_gen_driver.hpp
@@ -256,6 +256,8 @@ struct GenGPUCommandQueue {
   bool enqueueWorkItem(GenGPUWorkItem* item);
   void userEventChange(void);
   void waitOnEvent(GenGPUWorkItem* item);
+  void waitForFlush(void);
+  void waitForFinish(void);
   GenGPUCommandQueue(dri_bufmgr *bufmgr, drm_intel_context *ctx);
   ~GenGPUCommandQueue(void);
 };
-- 
cgit v1.2.3