Split formula group for OpenCL up into smaller bits when necessary

Will make it less demanding on low-end hardware, where the device
driver is unresponsive for too long when an OpenCL kernel handling lots
of data is executing. This makes Windows restart the driver which is
problematic.

I tried several approaches of splitting, both at higher levels in sc
and at the lowest level just before creating and executing the OpenCL
kernel(s). This seems to be the most minimal and local approach. Doing
it at the lower level would have required too much poking into our
obscure OpenCL code, like passing an offset parameter to every kernel.

Use a simple heuristic to find out whether to split. On the
problematic low-end devices, CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT is
4, while for more performant devices it is 1 or 8.

1 files changed, 1 insertions, 0 deletions

diff --git a/include/opencl/openclwrapper.hxx b/include/opencl/openclwrapper.hxx
index 04fe1e3725e7..fe677729460b 100644
--- a/include/opencl/openclwrapper.hxx
+++ b/include/opencl/openclwrapper.hxx
@@ -51,6 +51,7 @@ struct GPUEnv
     int mnCmdQueuePos;
     bool mnKhrFp64Flag;
     bool mnAmdFp64Flag;
+    cl_uint mnPreferredVectorWidthFloat;
 };
 
 extern OPENCL_DLLPUBLIC GPUEnv gpuEnv;