diff options
Diffstat (limited to 'nerv/lib/matrix/cukernel.cu')
| -rw-r--r-- | nerv/lib/matrix/cukernel.cu | 36 |
1 files changed, 36 insertions, 0 deletions
diff --git a/nerv/lib/matrix/cukernel.cu b/nerv/lib/matrix/cukernel.cu index a19030a..210e6bf 100644 --- a/nerv/lib/matrix/cukernel.cu +++ b/nerv/lib/matrix/cukernel.cu @@ -1,5 +1,41 @@ #define NERV_GENERIC_CUKERNEL +#include "cumatrix.h" + +#ifdef __NERV_FUTURE_CUDA_7 +__device__ double atomicAdd_nvidia(double* address, double val) { + /* nvidia provided this implementation + atmoicAdd is not included in CUDA for double */ + unsigned long long int* address_as_ull = + (unsigned long long int*)address; + unsigned long long int old = *address_as_ull, assumed; + do { + assumed = old; + old = atomicCAS(address_as_ull, assumed, + __double_as_longlong(val + + __longlong_as_double(assumed))); + } while (assumed != old); + return __longlong_as_double(old); +} + +__device__ float atomicAdd_nvidia(float* address, float val) { + /* nvidia provided this implementation + I tried the included atomocAdd, but the select_liner layer result seems + unreproduceable, but sadly, even if I used this implementation, the + select_linear layer result is still unreproduceable */ + int* address_as_ull = (int*)address; + int old = *address_as_ull, assumed; + do { + assumed = old; + old = atomicCAS(address_as_ull, assumed, + __float_as_int(val + + __int_as_float(assumed))); + } while (assumed != old); + return __int_as_float(old); +} +#endif + + #define cudak_(NAME) cudak_float_ ## NAME #define MATRIX_USE_FLOAT #include "generic/elem_type.h" |