[Libreoffice-commits] core.git: include/opencl opencl/source

[Kohei Yoshida]

 include/opencl/openclwrapper.hxx |   10 +++++
 opencl/source/openclwrapper.cxx  |   66 ++++++++++++++++++++++++++++++---------
 2 files changed, 60 insertions(+), 16 deletions(-)

New commits:
commit e12fcd050a6647232df3003d0a5cca01e564ac22
Author: Kohei Yoshida <kohei.yoshida at collabora.com>
Date:   Wed Jan 7 16:24:06 2015 -0500

    Set up foundation for supporting multiple OpenCL command queues per device.
    
    Change-Id: Ia63c8bd6552cdbc2b5eabadaa044b9f1eca5f664

diff --git a/include/opencl/openclwrapper.hxx b/include/opencl/openclwrapper.hxx
index 412c00b..d869819 100644
--- a/include/opencl/openclwrapper.hxx
+++ b/include/opencl/openclwrapper.hxx
@@ -33,6 +33,7 @@ if( status != CL_SUCCESS )    \
 }
 
 #define MAX_CLFILE_NUM 50
+#define OPENCL_CMDQUEUE_SIZE 1 // number of command queues per OpenCL device.
 
 #include <cstdio>
 
@@ -53,9 +54,10 @@ struct GPUEnv
     cl_context mpContext;
     cl_device_id *mpArryDevsID;
     cl_device_id mpDevID;
-    cl_command_queue mpCmdQueue;
+    cl_command_queue mpCmdQueue[OPENCL_CMDQUEUE_SIZE];
     cl_program mpArryPrograms[MAX_CLFILE_NUM]; //one program object maps one kernel source file
     int mnIsUserCreated; // 1: created , 0:no create and needed to create by opencl wrapper
+    int mnCmdQueuePos;
     bool mnKhrFp64Flag;
     bool mnAmdFp64Flag;
 };
@@ -79,6 +81,12 @@ OPENCL_DLLPUBLIC bool switchOpenCLDevice(const OUString* pDeviceId, bool bAutoSe
 
 OPENCL_DLLPUBLIC void getOpenCLDeviceInfo(size_t& rDeviceId, size_t& rPlatformId);
 
+/**
+ * Set the current command queue position in case of multiple command queues
+ * for a given device.
+ */
+OPENCL_DLLPUBLIC void setOpenCLCmdQueuePosition( int nPos );
+
 }
 
 #endif
diff --git a/opencl/source/openclwrapper.cxx b/opencl/source/openclwrapper.cxx
index 09e7bc0..613e503 100644
--- a/opencl/source/openclwrapper.cxx
+++ b/opencl/source/openclwrapper.cxx
@@ -89,8 +89,10 @@ OString maCacheFolder = getCacheFolder();
 void setKernelEnv( KernelEnv *envInfo )
 {
     envInfo->mpkContext = gpuEnv.mpContext;
-    envInfo->mpkCmdQueue = gpuEnv.mpCmdQueue;
     envInfo->mpkProgram = gpuEnv.mpArryPrograms[0];
+
+    assert(gpuEnv.mnCmdQueuePos < OPENCL_CMDQUEUE_SIZE);
+    envInfo->mpkCmdQueue = gpuEnv.mpCmdQueue[gpuEnv.mnCmdQueuePos];
 }
 
 namespace {
@@ -259,7 +261,7 @@ struct OpenCLEnv
     cl_platform_id mpOclPlatformID;
     cl_context mpOclContext;
     cl_device_id mpOclDevsID;
-    cl_command_queue mpOclCmdQueue;
+    cl_command_queue mpOclCmdQueue[OPENCL_CMDQUEUE_SIZE];
 };
 
 bool initOpenCLAttr( OpenCLEnv * env )
@@ -270,10 +272,14 @@ bool initOpenCLAttr( OpenCLEnv * env )
     gpuEnv.mpContext = env->mpOclContext;
     gpuEnv.mpPlatformID = env->mpOclPlatformID;
     gpuEnv.mpDevID = env->mpOclDevsID;
-    gpuEnv.mpCmdQueue = env->mpOclCmdQueue;
 
     gpuEnv.mnIsUserCreated = 1;
 
+    for (int i = 0; i < OPENCL_CMDQUEUE_SIZE; ++i)
+        gpuEnv.mpCmdQueue[i] = env->mpOclCmdQueue[i];
+
+    gpuEnv.mnCmdQueuePos = 0; // default to 0.
+
     return false;
 }
 
@@ -284,11 +290,16 @@ void releaseOpenCLEnv( GPUEnv *gpuInfo )
         return;
     }
 
-    if ( gpuEnv.mpCmdQueue )
+    for (int i = 0; i < OPENCL_CMDQUEUE_SIZE; ++i)
     {
-        clReleaseCommandQueue( gpuEnv.mpCmdQueue );
-        gpuEnv.mpCmdQueue = NULL;
+        if (gpuEnv.mpCmdQueue[i])
+        {
+            clReleaseCommandQueue(gpuEnv.mpCmdQueue[i]);
+            gpuEnv.mpCmdQueue[i] = NULL;
+        }
     }
+    gpuEnv.mnCmdQueuePos = 0;
+
     if ( gpuEnv.mpContext )
     {
         clReleaseContext( gpuEnv.mpContext );
@@ -761,25 +772,41 @@ bool switchOpenCLDevice(const OUString* pDevice, bool bAutoSelect, bool bForceEv
         return false;
     }
 
-    cl_command_queue command_queue = clCreateCommandQueue(
+    cl_command_queue command_queue[OPENCL_CMDQUEUE_SIZE];
+    for (int i = 0; i < OPENCL_CMDQUEUE_SIZE; ++i)
+    {
+        command_queue[i] = clCreateCommandQueue(
             context, pDeviceId, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &nState);
 
-    if(command_queue == NULL || nState != CL_SUCCESS)
-    {
-        if(command_queue != NULL)
-            clReleaseCommandQueue(command_queue);
+        if (command_queue[i] == NULL || nState != CL_SUCCESS)
+        {
+            // Release all command queues created so far.
+            for (int j = 0; j <= i; ++j)
+            {
+                if (command_queue[j])
+                {
+                    clReleaseCommandQueue(command_queue[j]);
+                    command_queue[j] = NULL;
+                }
+            }
 
-        clReleaseContext(context);
-        SAL_WARN("opencl", "failed to set/switch opencl device");
-        return false;
+            clReleaseContext(context);
+            SAL_WARN("opencl", "failed to set/switch opencl device");
+            return false;
+        }
     }
 
+    setOpenCLCmdQueuePosition(0); // Call this just to avoid the method being deleted from unused function deleter.
+
     releaseOpenCLEnv(&gpuEnv);
     OpenCLEnv env;
     env.mpOclPlatformID = platformId;
     env.mpOclContext = context;
     env.mpOclDevsID = pDeviceId;
-    env.mpOclCmdQueue = command_queue;
+
+    for (int i = 0; i < OPENCL_CMDQUEUE_SIZE; ++i)
+        env.mpOclCmdQueue[i] = command_queue[i];
+
     initOpenCLAttr(&env);
 
     // why do we need this at all?
@@ -804,6 +831,15 @@ void getOpenCLDeviceInfo(size_t& rDeviceId, size_t& rPlatformId)
     findDeviceInfoFromDeviceId(id, rDeviceId, rPlatformId);
 }
 
+void setOpenCLCmdQueuePosition( int nPos )
+{
+    if (nPos < 0 || nPos >= OPENCL_CMDQUEUE_SIZE)
+        // Out of range. Ignore this.
+        return;
+
+    gpuEnv.mnCmdQueuePos = nPos;
+}
+
 }
 
 /* vim:set shiftwidth=4 softtabstop=4 expandtab: */