I am trying to compile ccminer to work with a GPU that has compute capability 3.5.
When I compile with the default nvcc_ARCH (compute capability 5.0 + 5.2), the code compiles but when I try to benchmark
./ccminer --benchmark I get the following error:
Cuda error in func 'quark_blake512_cpu_setBlock_80' at line 815 : invalid device symbol.
For what I have been able to figure out, this error is related to compute capability errors, and that is why I tried changing the setting in the makefile.
I have changed Makefile.am accordingly so that nvcc_ARCH matches only the 3.5 capability (only 3.5), and I get this error:
ptxas /tmp/tmpxft_000015a2_00000000-6_cuda_x11_echo.ptx, line 10165; error : Instruction 'lop3' requires .target sm_50 or higher
I have also tried maintaining the original nvcc_ARCH (5.0 and 5.2) and appending the 3.5 line (+), and this is the error I get:
ptxas /tmp/tmpxft_00006277_00000000-6_cuda_x11_echo.compute_35.ptx, line 9182; error : Instruction 'lop3' requires .target sm_50 or higher
With all 4 lines of nvcc_ARCH uncommented, this is the error I get when compiling:
Algo256/cuda_bmw256.cu(337): error: identifier "__ldg" is undefined
What am I missing here? Am I trying to do this the wrong way? Is this problem not related with ccminer and I am looking at the wrong place?