Tuned bandwidth numbers for some 5-d operators looks wrong

[maddyscientist]

Flops look correct, but the bandwidth numbers for the 5-d local kernels look way off. A quick look at the code suggests that these aren't special cased, and the default DslashCuda::bytes() is being used, but I'll need to check for more than 30 seconds to work this out exactly.

Tuned block=(128,3,1), shared=24577 giving 71.00 Gflop/s, 1116.09 GB/s for N4quda16MDWFDslashPCCudaI7double2S1_EE with type=single-GPU,reconstruct=18,Dslash4pre
Tuned block=(32,4,1), shared=0 giving 170.86 Gflop/s, 347.94 GB/s for N4quda16MDWFDslashPCCudaI7double2S1_EE with type=single-GPU,reconstruct=18,Dslash4
Tuned block=(32,4,1), shared=16385 giving 155.89 Gflop/s, 181.87 GB/s for N4quda16MDWFDslashPCCudaI7double2S1_EE with type=single-GPU,reconstruct=18,Dslash5inv
Tuned block=(32,2,1), shared=24577 giving 64.99 Gflop/s, 895.91 GB/s for N4quda16MDWFDslashPCCudaI7double2S1_EE with type=single-GPU,reconstruct=18,Xpay,Dslash5
Executing 10 kernel loops...

[mathiaswagner]

What volume was this? Did you check limiting cases?

[mathiaswagner]

Closed with #317.