sandey/blender
1
0
Fork 0
forked from bartvdbraak/blender
Code Pull Requests Activity

Cycles: Improved thread order for better CUDA performance.

Browse Source 
This patch puts threads that render the same pixel closer together,
as opposed to threads that render the same sample. Thus threads
within a warp are more coherent in memory access and control flow,
leading to performance improvements.

Example benchmarks on a Quadro RTX4000 (WDDM) on Windows 10:
Koro:                 4:23 ->  3:46
BMW:                  1:18 ->  1:25
Barbershop Interior: 17:52 -> 14:55
Classroom:            4:37 ->  3:45

Performance differences on OpenCL/AMD were hit and miss, some scenes
became faster, others lost significantly. Therefore, this is kept as
CUDA only change for now.
This commit is contained in:
Stefan Werner 2019-03-14 11:45:58 +01:00
parent 4887baf7d6
commit 47da8dcbca
1 changed files with 6 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
intern/cycles/kernel/kernel_work_stealing.h
View File

@ -66,9 +66,15 @@ ccl_device_inline void get_work_pixel(ccl_global const WorkTile *tile,
ccl_private uint *y,
ccl_private uint *sample)
{
#ifdef __KERNEL_CUDA__
/* Keeping threads for the same pixel together improves performance on CUDA. */
uint sample_offset = global_work_index % tile->num_samples;
uint pixel_offset = global_work_index / tile->num_samples;
#else /* __KERNEL_CUDA__ */
uint tile_pixels = tile->w * tile->h;
uint sample_offset = global_work_index / tile_pixels;
uint pixel_offset = global_work_index - sample_offset * tile_pixels;
#endif /* __KERNEL_CUDA__ */
uint y_offset = pixel_offset / tile->w;
uint x_offset = pixel_offset - y_offset * tile->w;