Cycles: set num_sort_partition_elements to 8192 for oneAPI
The default value of 65536 wasn't optimal on Intel GPUs, switching to 8192 gives a 0 to 15% performance improvement depending on the scenes.
This commit is contained in:
parent
f9a0e825c6
commit
4d4f8bbfe4
|
@ -52,6 +52,11 @@ int OneapiDeviceQueue::num_concurrent_busy_states(const size_t /*state_size*/) c
|
|||
return 4 * max(8 * max_num_threads, 65536);
|
||||
}
|
||||
|
||||
int OneapiDeviceQueue::num_sort_partition_elements() const
|
||||
{
|
||||
return 8192;
|
||||
}
|
||||
|
||||
void OneapiDeviceQueue::init_execution()
|
||||
{
|
||||
oneapi_device_->load_texture_info();
|
||||
|
|
|
@ -28,6 +28,8 @@ class OneapiDeviceQueue : public DeviceQueue {
|
|||
|
||||
virtual int num_concurrent_busy_states(const size_t state_size) const override;
|
||||
|
||||
virtual int num_sort_partition_elements() const override;
|
||||
|
||||
virtual void init_execution() override;
|
||||
|
||||
virtual bool enqueue(DeviceKernel kernel,
|
||||
|
|
Loading…
Reference in New Issue