mirror of
https://github.com/GOSTSec/ccminer
synced 2025-01-08 22:07:56 +00:00
b9da6c67f5
the main improvement is to reduce asm calls to read global mem
but, a few more regs are used (68 mini vs 64 on SM 5.2)
so reduce the forced launch bounds to allow 80 or 128 regs per thread
Note: cuda 6.5 seems not able to store with v4.u32... (7.5 is fine)
st.global.v4.u32 [%rd2], {%r3783, %r3824, %r3823, %r3822};
st.global.v2.u32 [%rd2+16], {%r3821, %r3820};
st.global.u32 [%rd2+24], %r3819;
st.global.u32 [%rd2+28], %r3818;
st.global.u32 [%rd2+44], %r3814;
st.global.u32 [%rd2+40], %r3815;
...
todo, check alexis variant.. but wanted to keep this code before in git...
|
||
|---|---|---|
| .. | ||
| cuda_bmw512_sm3.cuh | ||
| cuda_bmw512.cu | ||
| cuda_jh512.cu | ||
| cuda_quark_blake512_sp.cuh | ||
| cuda_quark_blake512.cu | ||
| cuda_quark_compactionTest.cu | ||
| cuda_quark_groestl512_sm2.cuh | ||
| cuda_quark_groestl512.cu | ||
| cuda_quark_keccak512.cu | ||
| cuda_quark.h | ||
| cuda_skein512.cu | ||
| groestl_functions_quad.h | ||
| groestl_simple.cuh | ||
| groestl_transf_quad.h | ||
| nist5.cu | ||
| quarkcoin.cu | ||