estebanpw/cudagecko

A cuda parallel implementation of the GECKO algorithm for exhaustive ungapped sequence comparisons

cudagecko's Issues

in original gecko/unlimited the second hit is badly filtered

see FilterHits.c:

lastPosition is not updated on the first written hit (default one)
and still used in the second hit comparison

consider a floating point implementation

Floating point operations have higher throughput in CUDA. Although for the current hashing scheme (which requires values of up to 2^64, exactly 64 bits) a floating point encoding of 64 bits would not suffice (since numbers besides the range 2^53 can not represent exact integers). This would in turn cause matching errors at the large keys.

Therefore, it would be interesting to study a workaround or to reduce the k-size (which increases sensitivity) at the cost of more processing, although maybe compensated by the performance boost of floating point operations.

Someone to check on other GPUs

It would be interesting if someone could provide data on runtimes using other GPUs.
The only requirement would be that the PCIexpress is version 3 at least, otherwise it will be difficult to compare.

fine-grained aligner

add a fine-grained aligner mode that uses reduction to check the exact endings of the high-scoring alignments

add support mode for non-pooled pinned memory

The idea behind pinned memory is speeding up data transfers between host and device. Combined with pooled memory (only one malloc) we can save lots of system calls.

-> However for 4 GB of GRAM the host requires ~5 GB of pinned memory, which can be expensive in some systems.

TODO: implement a malloc/free on demand that can enabled via parameters