use FNV-1a for kernel indexing instead of md5

[yan-ming]

make test

********************
Testing Time: 634.94s
********************
Failing Tests (6):
    CPPAMP :: Unit/HC/memcpy_symbol1.cpp
    CPPAMP :: Unit/HC/memcpy_symbol3.cpp
    CPPAMP :: Unit/HC/wg_size.cpp
    CPPAMP :: Unit/HSAIL/shfl_xor.cpp
    CPPAMP :: Unit/SharedLibrary/shared_library2.cpp
    CPPAMP :: Unit/SharedLibrary/shared_library3.cpp

  Expected Passes    : 663
  Expected Failures  : 25
  Unsupported Tests  : 10
  Unexpected Failures: 6

[whchung]

@yan-ming I was looking at failed cases in this pull request. And unfortunately it seems the following two tests are directly resulted from switching to FNV-1a and introducing 104-byte cutoff:

CPPAMP :: Unit/SharedLibrary/shared_library2.cpp
CPPAMP :: Unit/SharedLibrary/shared_library3.cpp

Those 2 tests have 2 nearly-identical kernels with a very small difference in the names. And that gives 104 heuristic fail. I've tested they would pass if enlarge the cutoff size to 256 bytes.

[yan-ming]

Hi @whchung , thanks for prompting this issue up. I agree there would be collisions when two nearly identical kernels are used at a time. Changing the cutoff size to 256 doesn't make them pass on my end. It looks like the proper size here is about 270, I think it might be safer if we have like 512 in this case?

[whchung]

@yan-ming you are right, I accidentally used MD5 in my previous experiment. 512 is a safer value. I've pushed a commit already.

[whchung]

@yan-ming further tests show we need to use 768 so both HSAIL backend and LC backend can work.

[yan-ming]

@whchung could you briefly explain how did you discover this value?

[whchung]

@yan-ming I made two builds of hcc, one with HSAIL backend (default one), and another with LC backend. Then run the 2 tests aforementioned. And gradually increase the cutoff value by 128 bytes.

This should be just a temporary thing. We'll push people working on LC backend and HSAIL finalizer to calculate hash and put into headers so we don't need to do them in HCC runtime anymore.

[yan-ming]

@whchung I see. I only use the LC backend for testing hcfft and the porting applications. Even 104 works for my cases.

I'm okay if we enlarge the cutoff size here, I agree with that the ultimate solution should be done in the finalizer.