diff options
| author | Brecht Van Lommel <brechtvanlommel@gmail.com> | 2017-10-08 20:08:44 +0300 |
|---|---|---|
| committer | Brecht Van Lommel <brechtvanlommel@gmail.com> | 2017-10-08 22:12:16 +0300 |
| commit | e360d003ea45ee233c6f10c03ff57c956929b383 (patch) | |
| tree | 177745a20e1fc69cfee4765cc29ac055bed383d2 /intern/cycles/device/device_cuda.cpp | |
| parent | 5aa08eb3cc7c9c5a6282d938fd2ffa6266a40ff7 (diff) | |
Cycles: schedule more work for non-display and compute preemption CUDA cards.
This change affects CUDA GPUs not connected to a display or connected to a
display but supporting compute preemption so that the display does not
freeze. I couldn't find an official list, but compute preemption seems to be
only supported with GTX 1070+ and Linux (not GTX 1060- or Windows).
This helps improve small tile rendering performance further if there are
sufficient samples x number of pixels in a single tile to keep the GPU busy.
Diffstat (limited to 'intern/cycles/device/device_cuda.cpp')
| -rw-r--r-- | intern/cycles/device/device_cuda.cpp | 21 |
1 files changed, 16 insertions, 5 deletions
diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp index 56a56c5217c..216c85f24e7 100644 --- a/intern/cycles/device/device_cuda.cpp +++ b/intern/cycles/device/device_cuda.cpp @@ -1313,9 +1313,14 @@ public: CUdeviceptr d_work_tiles = cuda_device_ptr(work_tiles.device_pointer); /* Prepare work size. More step samples render faster, but for now we - * remain conservative to avoid driver timeouts. */ + * remain conservative for GPUs connected to a display to avoid driver + * timeouts and display freezing. */ int min_blocks, num_threads_per_block; cuda_assert(cuOccupancyMaxPotentialBlockSize(&min_blocks, &num_threads_per_block, cuPathTrace, NULL, 0, 0)); + if(!info.display_device) { + min_blocks *= 8; + } + uint step_samples = divide_up(min_blocks * num_threads_per_block, wtile->w * wtile->h);; /* Render all samples. */ @@ -2109,7 +2114,6 @@ void device_cuda_info(vector<DeviceInfo>& devices) for(int num = 0; num < count; num++) { char name[256]; - int attr; if(cuDeviceGetName(name, 256, num) != CUDA_SUCCESS) continue; @@ -2141,14 +2145,21 @@ void device_cuda_info(vector<DeviceInfo>& devices) (unsigned int)pci_location[1], (unsigned int)pci_location[2]); - /* if device has a kernel timeout, assume it is used for display */ - if(cuDeviceGetAttribute(&attr, CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT, num) == CUDA_SUCCESS && attr == 1) { + /* If device has a kernel timeout and no compute preemption, we assume + * it is connected to a display and will freeze the display while doing + * computations. */ + int timeout_attr = 0, preempt_attr = 0; + cuDeviceGetAttribute(&timeout_attr, CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT, num); + cuDeviceGetAttribute(&preempt_attr, CU_DEVICE_ATTRIBUTE_COMPUTE_PREEMPTION_SUPPORTED, num); + + if(timeout_attr && !preempt_attr) { info.description += " (Display)"; info.display_device = true; display_devices.push_back(info); } - else + else { devices.push_back(info); + } } if(!display_devices.empty()) |