[Libreoffice-commits] core.git: Branch 'libreoffice-5-1-5' - include/opencl opencl/source sc/source

[Laszlo Nemeth]

 include/opencl/openclwrapper.hxx    |    2 -
 opencl/source/openclwrapper.cxx     |   37 ++++++++++++++++++++++++++++++++++--
 sc/source/core/data/formulacell.cxx |    5 +---
 3 files changed, 38 insertions(+), 6 deletions(-)

New commits:
commit cb309928872735c577d2b59fe9dd8690424410b5
Author: Laszlo Nemeth <nemeth at numbertext.org>
Date:   Mon Jul 25 16:28:39 2016 +0300

    Need to try to avoid TDR also with NVIDIA cards on Windows 7 or earlier
    
    (TDR is Timeout detection and recovery, was introduced in Vista.)
    
    Change-Id: If88f8e9e2aff2a5ffd633607ee6aebb5614c5caf
    Reviewed-on: https://gerrit.libreoffice.org/27523
    Reviewed-by: Caolán McNamara <caolanm at redhat.com>
    Reviewed-by: Tor Lillqvist <tml at collabora.com>
    Reviewed-by: Jan Holesovsky <kendy at collabora.com>
    Tested-by: Jan Holesovsky <kendy at collabora.com>

diff --git a/include/opencl/openclwrapper.hxx b/include/opencl/openclwrapper.hxx
index 0583400..ac86bea 100644
--- a/include/opencl/openclwrapper.hxx
+++ b/include/opencl/openclwrapper.hxx
@@ -49,7 +49,7 @@ struct OPENCL_DLLPUBLIC GPUEnv
     int mnCmdQueuePos;
     bool mnKhrFp64Flag;
     bool mnAmdFp64Flag;
-    cl_uint mnPreferredVectorWidthFloat;
+    bool mbNeedsTDRAvoidance;
 
     static bool isOpenCLEnabled();
 };
diff --git a/opencl/source/openclwrapper.cxx b/opencl/source/openclwrapper.cxx
index 2551b05..de1ed26 100644
--- a/opencl/source/openclwrapper.cxx
+++ b/opencl/source/openclwrapper.cxx
@@ -42,6 +42,10 @@
 #define OPENCL_DLL_NAME "libOpenCL.so"
 #endif
 
+#ifdef _WIN32_WINNT_WINBLUE
+#include <VersionHelpers.h>
+#endif
+
 #define DEVICE_NAME_LENGTH 1024
 #define DRIVER_VERSION_LENGTH 1024
 #define PLATFORM_VERSION_LENGTH 1024
@@ -456,6 +460,8 @@ void checkDeviceForDoubleSupport(cl_device_id deviceId, bool& bKhrFp64, bool& bA
 bool initOpenCLRunEnv( GPUEnv *gpuInfo )
 {
     OpenCLZone zone;
+    cl_uint nPreferredVectorWidthFloat;
+    char pName[64];
 
     bool bKhrFp64 = false;
     bool bAmdFp64 = false;
@@ -465,11 +471,38 @@ bool initOpenCLRunEnv( GPUEnv *gpuInfo )
     gpuInfo->mnKhrFp64Flag = bKhrFp64;
     gpuInfo->mnAmdFp64Flag = bAmdFp64;
 
-    gpuInfo->mnPreferredVectorWidthFloat = 0;
+    gpuInfo->mbNeedsTDRAvoidance = false;
 
     clGetDeviceInfo(gpuInfo->mpDevID, CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT, sizeof(cl_uint),
-                    &gpuInfo->mnPreferredVectorWidthFloat, nullptr);
+                    &nPreferredVectorWidthFloat, nullptr);
+    clGetPlatformInfo(gpuInfo->mpPlatformID, CL_PLATFORM_NAME, 64,
+             pName, nullptr);
+
+    bool bIsNotWinOrIsWin8OrGreater = true;
+
+// the Win32 SDK 8.1 deprecates GetVersionEx()
+#ifdef _WIN32_WINNT_WINBLUE
+    bIsNotWinOrIsWin8OrGreater = IsWindows8OrGreater();
+#elif defined (_WIN32)
+    OSVERSIONINFO aVersionInfo;
+    memset( &aVersionInfo, 0, sizeof(aVersionInfo) );
+    aVersionInfo.dwOSVersionInfoSize = sizeof( aVersionInfo );
+    if (GetVersionEx( &aVersionInfo ))
+    {
+        // Windows 7 or lower?
+        if (aVersionInfo.dwMajorVersion < 6 ||
+           (aVersionInfo.dwMajorVersion == 6 && aVersionInfo.dwMinorVersion < 2))
+            bIsNotWinOrIsWin8OrGreater = false;
+    }
+#endif
 
+    // Heuristic: Certain old low-end OpenCL implementations don't
+    // work for us with too large group lengths. Looking at the preferred
+    // float vector width seems to be a way to detect these devices, except
+    // the non-working NVIDIA cards on Windows older than version 8.
+    gpuInfo->mbNeedsTDRAvoidance = ( nPreferredVectorWidthFloat == 4 ) ||
+        ( !bIsNotWinOrIsWin8OrGreater &&
+          OUString::createFromAscii(pName).indexOf("NVIDIA") > -1 );
     return false;
 }
 
diff --git a/sc/source/core/data/formulacell.cxx b/sc/source/core/data/formulacell.cxx
index daf8f65..147c097 100644
--- a/sc/source/core/data/formulacell.cxx
+++ b/sc/source/core/data/formulacell.cxx
@@ -4039,9 +4039,8 @@ bool ScFormulaCell::InterpretFormulaGroup()
 #ifdef WNT
     // Heuristic: Certain old low-end OpenCL implementations don't
     // work for us with too large group lengths. 1000 was determined
-    // empirically to be a good compromise. Looking at the preferred
-    // float vector width seems to be a way to detect these devices.
-    if (opencl::gpuEnv.mnPreferredVectorWidthFloat == 4)
+    // empirically to be a good compromise.
+    if (opencl::gpuEnv.mbNeedsTDRAvoidance)
         nMaxGroupLength = 1000;
 #endif