diff options
| author | Thomas Dinges <blender@dingto.org> | 2013-12-24 23:20:35 +0400 |
|---|---|---|
| committer | Thomas Dinges <blender@dingto.org> | 2013-12-24 23:20:57 +0400 |
| commit | ad0a3de3ce7f84486047b11343405072f0972107 (patch) | |
| tree | ae4ac8d0d79d9ddc0507a2f539e68bd4ca6501e8 | |
| parent | ec1f90864fc452ad5217b13f66179e2d167a9b82 (diff) | |
Cycles / OpenCL: Let the OpenCL runtime determine its optimal work-group size automatically, by passing a NULL pointer here.
This is recommended in the Intel OpenCL optimization docs (http://software.intel.com/en-us/vcsource/samples/optimizing-opencl) and I can confirm a small performance increase here (1-2% on nVidia OpenCL, up to 8% on Intel OpenCL).
| -rw-r--r-- | intern/cycles/device/device_opencl.cpp | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/intern/cycles/device/device_opencl.cpp b/intern/cycles/device/device_opencl.cpp index 3a88b68b292..9117b70d749 100644 --- a/intern/cycles/device/device_opencl.cpp +++ b/intern/cycles/device/device_opencl.cpp @@ -931,7 +931,7 @@ public: size_t global_size[2] = {global_size_round_up(local_size[0], w), global_size_round_up(local_size[1], h)}; /* run kernel */ - ciErr = clEnqueueNDRangeKernel(cqCommandQueue, kernel, 2, NULL, global_size, local_size, 0, NULL, NULL); + ciErr = clEnqueueNDRangeKernel(cqCommandQueue, kernel, 2, NULL, global_size, NULL, 0, NULL, NULL); opencl_assert(ciErr); opencl_assert(clFlush(cqCommandQueue)); } |