[REVIEW] Remove unnecessary norm_add1 array from tSNE

[zbjornson]

Reduces memory usage by sizeof(float)*nRows, reduces global loads/stores and fixes an FP error.

@drobison00 I think we can actually fold this get_norm kernel into attractive_kernel_bh and get rid of the norms array also. AFAICT the denominator just works out to

      Y1[i] * Y1[i] + Y2[i] * Y2[i] + 1.0f   // norm_add1[i]
    + Y1[j] * Y1[j] + Y2[j] * Y2[j]   // norm[j]
    - 2.0f * (Y1[i] * Y1[j] + Y2[i] * Y2[j])   // the other terms

Then it'll be easier to deal with FP error in a single place because we can reorder intermediates, etc.

Did that in a separate branch for consideration: 4fa9e89

It's these norm values that get to be huge, and I'm seeing lots of cases where norm[i] + 1 != norm_add1[i] because of insufficient precision.

[GPUtester]

Can one of the admins verify this patch?

[JohnZed]

Add to whitelist