On Thu, 12 Aug 1999, Jason Thorpe wrote:

> Note that the 164SX has a wimpy memory bus speed compared to the 164LX
> and the original PC164 (which had the best of all of them, actually).  This
> might contribute.

Please, you are a bit off scale here -- my pciide driver reads about 4MB/s
and a simple memcpy() can copy 108MB/s in my SX. (This actually sucks,
when I tried memcpy() in linux on the same machine, it copied over
200MB/s).

> If you get better transfer rates on the *same machine* under a different
> OS (as was implied earlier in the thread, I think), then there may be
> a problem with how we initialize the Cypress controller... but that

Yes, this is what I was saying, I had 7MB/s using my Linux driver, and I
have 3.5MB/s with the NetBSD driver, same machine, same cypress interface.

> would be the only issue; the vast majority of the PCI IDE code is
> shared between all controller types.

The original question was why my newly written Promise driver produces
good performance on i386 but not alpha, and a general alpha problem was
suspected as the Cypress driver also has bad performance. But since the
CMD646 is reported to give expected performance, this seems to be sort of
ruled out.

I guess I'll just have to keep on digging, unless someone else has some
idea what might be wrong.

>         -- Jason R. Thorpe <thorpej@nas.nasa.gov>

/Andreas