Cycles: Make kernel compilable for 3.7 compute capability
authorSergey Sharybin <sergey.vfx@gmail.com>
Thu, 28 Jan 2016 10:56:09 +0000 (11:56 +0100)
committerSergey Sharybin <sergey.vfx@gmail.com>
Thu, 28 Jan 2016 10:56:09 +0000 (11:56 +0100)
It is used by GK210 GPUs which could be found in, i.e. Tesla K80.

intern/cycles/kernel/kernels/cuda/kernel.cu

index c8940b981bbf58d0cdf60031daf1198369aff6ce..29067a666c45b218307e4ee85bb7ac0a509bc181 100644 (file)
 #define CUDA_KERNEL_MAX_REGISTERS 63
 #define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 63
 
+/* 3.7 */
+#elif __CUDA_ARCH__ == 370
+#define CUDA_MULTIPRESSOR_MAX_REGISTERS 65536
+#define CUDA_MULTIPROCESSOR_MAX_BLOCKS 16
+#define CUDA_BLOCK_MAX_THREADS 1024
+#define CUDA_THREAD_MAX_REGISTERS 255
+
+/* tunable parameters */
+#define CUDA_THREADS_BLOCK_WIDTH 16
+#define CUDA_KERNEL_MAX_REGISTERS 63
+#define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 63
+
 /* 5.0, 5.2 and 5.3 */
 #elif __CUDA_ARCH__ == 500 || __CUDA_ARCH__ == 520 || __CUDA_ARCH__ == 530
 #define CUDA_MULTIPRESSOR_MAX_REGISTERS 65536