Skip to content

Commit

Permalink
GridEncoding: additional safe-guard against half-precision for arch 61
Browse files Browse the repository at this point in the history
  • Loading branch information
Tom94 committed Feb 22, 2022
1 parent 953a8d0 commit ef26ff4
Showing 1 changed file with 1 addition and 1 deletion.
2 changes: 1 addition & 1 deletion include/tiny-cuda-nn/encodings/grid.h
Original file line number Diff line number Diff line change
Expand Up @@ -524,7 +524,7 @@ class GridEncoding : public Encoding<T> {
template <typename T, uint32_t N_POS_DIMS=3, uint32_t N_FEATURES_PER_LEVEL=2>
class GridEncodingTemplated : public GridEncoding<T> {
public:
#if TCNN_MIN_GPU_ARCH >= 60
#if TCNN_MIN_GPU_ARCH >= 62 || TCNN_MIN_GPU_ARCH == 60
// The GPUs that we tested this on do not have an efficient 1D fp16
// atomicAdd feature. Thus, we accumulate gradients at fp32 if we're
// forced to use 1D atomicAdds. As soon as 2D or higher is possible,
Expand Down

0 comments on commit ef26ff4

Please sign in to comment.