[Bf-blender-cvs] [b3adce77667] blender-v2.79a-release: Cycles: Workaround for performance loss with the CUDA 9.0 SDK.

[Stefan Werner]

Commit: b3adce77667498c20cd44e7cf6d76f8424a6dd89
Author: Stefan Werner
Date:   Tue Nov 21 10:29:11 2017 +0100
Branches: blender-v2.79a-release
https://developer.blender.org/rBb3adce77667498c20cd44e7cf6d76f8424a6dd89

Cycles: Workaround for performance loss with the CUDA 9.0 SDK.

CUDA 9.0.176 apparently caused some slow down on high-end Pascal cards that can be mitigated by increasing the number of registers. See https://developer.blender.org/F1142667 for a detailed comparison.

===================================================================

M	intern/cycles/kernel/kernels/cuda/kernel_config.h

===================================================================

diff --git a/intern/cycles/kernel/kernels/cuda/kernel_config.h b/intern/cycles/kernel/kernels/cuda/kernel_config.h
index 9fa39dc9ebb..efea522d803 100644
--- a/intern/cycles/kernel/kernels/cuda/kernel_config.h
+++ b/intern/cycles/kernel/kernels/cuda/kernel_config.h
@@ -76,6 +76,11 @@
 #  define CUDA_KERNEL_MAX_REGISTERS 48
 #  define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 63
 
+/* CUDA 9.0 seems to cause slowdowns on high-end Pascal cards unless we increase the number of registers */
+#  if __CUDACC_VER_MAJOR__ == 9 && __CUDA_ARCH__ >= 600
+#    #define CUDA_KERNEL_MAX_REGISTERS 64
+#  endif
+
 /* unknown architecture */
 #else
 #  error "Unknown or unsupported CUDA architecture, can't determine launch bounds"