[Gretl-devel] arma estimation performance on Atom with 64-bit OpenBLAS

I'm abstracting here from information provided by Oleh. He's estimating an ARMAX model (ARMA with an exogenous regressor) using current gretl snapshots for Windows, both 64-bit and 32-bit. The relevant gretl commands are: open phillips_aus.gdt # from the POE4 data collection arma 1 1 ; inf 0 diff(u) Trimming down Oleh's information gives these three cases: 1. host: DESKTOP-DE5ESQO, gretl in 64-bit mode, running Windows 10, using OpenBLAS (64-bit) with Atom blascore Function evaluations: 132 Evaluations of gradient: 22 Log-likelihood -62.75480 2. host: DESKTOP-DE5ESQO, gretl in 32-bit mode, running Windows 10, using OpenBLAS (32-bit) with Atom blascore Function evaluations: 52 Evaluations of gradient: 17 Log-likelihood -62.56261 3. host: HER-PC, gretl in 64-bit mode, running Windows 8 using OpenBLAS (64-bit) with Nehalem blascore Function evaluations: 52 Evaluations of gradient: 17 Log-likelihood -62.56261 The important contrast is 32-bit gretl vs 64-bit gretl on Oleh's DESKTOP-DE5ESQO, with an Intel "Atom"- or "Bay Trail"-type processor (specifically Pentium N3540). As you can see, the 64-bit version is performing badly: it takes a lot more iterations to achieve a worse log-likelihood on the ARMAX problem. Oleh's case 3 (Nehalem processor, 64-bit) is basically the same as all cases I've seen on machines I'm running: it seems "right". And case 2 is practically indistinguishable from case 3. So what's wrong with 64-bit gretl on Atom, running Windows? I'm assuming (perhaps wrongly) that this is an OpenBLAS issue, so I'll try pursuing that further. But if anyone else is able to test on Atom in 64-bit mode, or to contribute relevant knowledge, please do! Allin