regarding the definition of a "gang" according to the user guide

[Morten Mikkelsen]

regarding the definition of a "gang" according to the user guide.

The user's guide states "The number of program instances in a gang is relatively small; in practice, it's no more than 2-4x the native SIMD width of the hardware it is executing on.".

However, I've tried reading the value of programCount and also tried inferring it's value in more complex scenarios. So far it's always coming back as exactly the native width of my hardware

which is 8 since I have avx. At no point so far has the gang size been 16 or 32 going by the "2-4x" statement in the user's guide.
I'm not sure if I'm missing something or if it only happens in scenarios I just haven't hit yet that it would go wider than the native simd width?

Thanks,

Morten.

[Dmitry Babokin]

If you run "ispc --help", there will be information about available targets:

    [--target=<t>] Select target ISA and width.

                        <t>={host, sse2-i32x4, sse2-i32x8, sse4-i32x4, sse4-i32x8, sse4-i16x8, sse4-i8x16, avx1-i32x4, avx1-i32x8, avx1-i32x16, avx1-i64x4, avx1.1-i32x8, avx1.1-i32x16, avx1.1-i64x4, avx2-i32x8, avx2-i32x16, avx2-i64x4, avx512knl-i32x16,

                        avx512skx-i32x16, generic-x1, generic-x4, generic-x8, generic-x16, generic-x32, generic-x64, *-generic-x16}

Note that there are targets that are "2-4x the native SIMD width", for example avx2-i32x16 (is 2x) or sse4-i8x16 (is 4x).

For 32 bit wide base type targets, i.e. (i32 in target notation), it doesn't make sense to go wider than 2x. And 2x is also profitable only in some scenarios.

Hope it helps.

Dmitry.

--
You received this message because you are subscribed to the Google Groups "Intel SPMD Program Compiler Developers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ispc-dev+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.