IVF performance vs. CVF

[dlemmon1]

I am now getting acquainted with porting issues (pains), but at this point I'd like to look at the benefits. In one project I experienced nearly 2x performance increase with IVF vs. CVF (Xeon 3GHz). Has anyone done any benchmarks of IVF versus CVF?

[garyscott]

http://www.polyhedron.co.uk/compare/win32/f90bench_p4.html

[garyscott]

I wonder what some of the differences are that contribute to Intel's consistent performance advantage (CVF being a very strong contender of course). Does argument passing convention (CDECL vs STDCALL) impact it, is it mostly just attention to detail? Did the GEM intermediate form (assuming that was partially for convenience to allow targeting multiple processors) hinder producing the most efficient x86 code in any way? Is there insider information involved, or just plain familiarity and knowledge of the processor?

[Steven_L_Intel1]

It was really just attention to detail. Lots of it. CVF's code generator, good as it was, never got a lot of effort put into generating the best possible code for X86.Keep in mind that CVF's code generation and optimization has been essentially unchanged for almost three years. At Compaq, we never had the resources to devote to large-scale performance analysis and code tuning for X86 as Intel devotes to the subject. I'm not aware of any "inside knowledge" being applied (nor can I imagine what that might be.) What Intel does bring to the table is a thorough understanding of its processors' strengths based on lots of analysis.

Some of the benchmarks benefit from the use of SSE2 vectorization, which CVF doesn't do. The calling mechanism is just noise.

There will be some applications where CVF still does better, but those should be few, and become fewer over time. A lot of the "GEM" optimizations that were in the Compaq compilers have been added to the Intel compilers now, and you're seeing the results.