AMD "Bdver3" Steamroller GCC Testing With Kaveri
Phoronix: AMD "Bdver3" Steamroller GCC Testing With Kaveri
Following the GCC 4.9 Kaveri benchmarks and Clang 3.4 benchmarks with Kaveri, here's some other benchmarks from GCC looking at the impact of "-march=bdver3" and other CPU tuning flags from the AMD A10-7850K APU on Ubuntu Linux...
Those are nice improvements to the bdver versions. Bdver3 looks slightly slower than bdver2 but within the margin of error.
Can you do the same on clang? Their x86 target code looks very centered around Intel architectures, so it would be interesting if that leaves any impact or turns out to be irrelevant.
Edit: Btw, from reading sources. As far as I can tell bdver1 and bdver2 shares the same instruction timing, but bdver3 has its own scheduling (at least in gcc 4.9), but shares almost the same ISA extensions as bdver2. In clang the AMD isa extensions doesn't seem to matter, but there are different instruction timings.
Last edited by carewolf; 02-03-2014 at 06:38 PM.
It's going to take more than compiler optimizations to polish that turd.