lkml.org 
[lkml]   [2020]   [Sep]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 5/5] perf: arm_spe: Decode SVE events
On Mon, Sep 28, 2020 at 03:47:56PM +0100, Dave Martin wrote:
> On Mon, Sep 28, 2020 at 02:59:34PM +0100, André Przywara wrote:
> > On 28/09/2020 14:21, Dave Martin wrote:
> >
> > Hi Dave,
> >
> > > On Tue, Sep 22, 2020 at 11:12:25AM +0100, Andre Przywara wrote:
> > >> The Scalable Vector Extension (SVE) is an ARMv8 architecture extension
> > >> that introduces very long vector operations (up to 2048 bits).
> > >
> > > (8192, in fact, though don't expect to see that on real hardware any
> > > time soon... qemu and the Arm fast model can do it, though.)
> > >
> > >> The SPE profiling feature can tag SVE instructions with additional
> > >> properties like predication or the effective vector length.
> > >>
> > >> Decode the new operation type bits in the SPE decoder to allow the perf
> > >> tool to correctly report about SVE instructions.
> > >
> > >
> > > I don't know anything about SPE, so just commenting on a few minor
> > > things that catch my eye here.
> >
> > Many thanks for taking a look!
> > Please note that I actually missed a prior submission by Wei, so the
> > code changes here will end up in:
> > https://lore.kernel.org/patchwork/patch/1288413/
> >
> > But your two points below magically apply to his patch as well, so....
> >
> > >
> > >> Signed-off-by: Andre Przywara <andre.przywara@arm.com>
> > >> ---
> > >> .../arm-spe-decoder/arm-spe-pkt-decoder.c | 48 ++++++++++++++++++-
> > >> 1 file changed, 47 insertions(+), 1 deletion(-)
> > >>
> > >> diff --git a/tools/perf/util/arm-spe-decoder/arm-spe-pkt-decoder.c b/tools/perf/util/arm-spe-decoder/arm-spe-pkt-decoder.c
> > >> index a033f34846a6..f0c369259554 100644
> > >> --- a/tools/perf/util/arm-spe-decoder/arm-spe-pkt-decoder.c
> > >> +++ b/tools/perf/util/arm-spe-decoder/arm-spe-pkt-decoder.c
> > >> @@ -372,8 +372,35 @@ int arm_spe_pkt_desc(const struct arm_spe_pkt *packet, char *buf,
> > >> }
> > >> case ARM_SPE_OP_TYPE:
> > >> switch (idx) {
> > >> - case 0: return snprintf(buf, buf_len, "%s", payload & 0x1 ?
> > >> + case 0: {
> > >> + size_t blen = buf_len;
> > >> +
> > >> + if ((payload & 0x89) == 0x08) {
> > >> + ret = snprintf(buf, buf_len, "SVE");
> > >> + buf += ret;
> > >> + blen -= ret;
> > >
> > > (Nit: can ret be < 0 ? I've never been 100% clear on this myself for
> > > the s*printf() family -- if this assumption is widespread in perf tool
> > > a lready that I guess just go with the flow.)
> >
> > Yeah, some parts of the code in here check for -1, actually, but doing
> > this on every call to snprintf would push this current code over the
> > edge - and I cowardly avoided a refactoring ;-)
> >
> > Please note that his is perf userland, and also we are printing constant
> > strings here.
> > Although admittedly this starts to sounds like an excuse now ...
> >
> > > I wonder if this snprintf+increment+decrement sequence could be wrapped
> > > up as a helper, rather than having to be repeated all over the place.
> >
> > Yes, I was hoping nobody would notice ;-)
>
> It's probably not worth losing sleep over.
>
> snprintf(3) says, under NOTES:
>
> Until glibc 2.0.6, they would return -1 when the output was
> truncated.
>
> which is probably ancient enough history that we don't care. C11 does
> say that a negative return value can happen "if an encoding error
> occurred". _Probably_ not a problem if perf tool never calls
> setlocale(), but ...

I have one patch which tried to fix the snprintf+increment sequence
[1], to be honest, the change seems urgly for me. I agree it's better
to use a helper to wrap up.

[1] https://lore.kernel.org/patchwork/patch/1288410/

> > >> + if (payload & 0x2)
> > >> + ret = snprintf(buf, buf_len, " FP");
> > >> + else
> > >> + ret = snprintf(buf, buf_len, " INT");
> > >> + buf += ret;
> > >> + blen -= ret;
> > >> + if (payload & 0x4) {
> > >> + ret = snprintf(buf, buf_len, " PRED");
> > >> + buf += ret;
> > >> + blen -= ret;
> > >> + }
> > >> + /* Bits [7..4] encode the vector length */
> > >> + ret = snprintf(buf, buf_len, " EVLEN%d",
> > >> + 32 << ((payload >> 4) & 0x7));
> > >
> > > Isn't this just extracting 3 bits (0x7)?
> >
> > Ah, right, the comment is wrong. It's actually bits [6:4].
> >
> > > And what unit are we aiming
> > > for here: is it the number of bytes per vector, or something else? I'm
> > > confused by the fact that this will go up in steps of 32, which doesn't
> > > seem to match up to the architecure.
> >
> > So this is how SPE encodes the effective vector length in its payload:
> > the format is described in section "D10.2.7 Operation Type packet" in a
> > (recent) ARMv8 ARM. I put the above statement in a C file and ran all
> > input values through it, it produced the exact *bit* length values as in
> > the spec.
> >
> > Is there any particular pattern you are concerned about?
> > I admit this is somewhat hackish, I can do an extra function to put some
> > comments in there.
>
> Mostly I'm curious because the encoding doesn't match the SVE
> architecture: SVE requires 4 bits to specify the vector length, not 3.
> This might have been a deliberate limitation in the SPE spec., but it
> raises questions about what should happen when 3 bits is not enough.
>
> For SVE, valid vector lengths are 16 bytes * n
> or equivalently 128 bits * n), where 1 <= n <= 16.
>
> The code here though cannot print EVLEN16 or EVLEN48 etc. This might
> not be a bug, but I'd like to understand where it comes from...

In the SPE's spec, the defined values for EVL are:

0b'000 -> EVLEN: 32 bits.
0b'001 -> EVLEN: 64 bits.
0b'010 -> EVLEN: 128 bits.
0b'011 -> EVLEN: 256 bits.
0b'100 -> EVLEN: 512 bits.
0b'101 -> EVLEN: 1024 bits.
0b'110 -> EVLEN: 2048 bits.

Note that 0b'111 is reserved. In theory, I think SPE Operation packet
can support up to 4196 bits (32 << 7) when the EVL field is 0b'111; but
it's impossible to express vector length for 8192 bits as you mentioned.

Thanks,
Leo

> > > I notice that bit 7 has to be zero to get into this if() though.
> > >
> > >> + buf += ret;
> > >> + blen -= ret;
> > >> + return buf_len - blen;
> > >> + }
> > >> +
> > >> + return snprintf(buf, buf_len, "%s", payload & 0x1 ?
> > >> "COND-SELECT" : "INSN-OTHER");
> > >> + }
> > >> case 1: {
> > >> size_t blen = buf_len;
> > >>
> > >> @@ -403,6 +430,25 @@ int arm_spe_pkt_desc(const struct arm_spe_pkt *packet, char *buf,
> > >> ret = snprintf(buf, buf_len, " NV-SYSREG");
> > >> buf += ret;
> > >> blen -= ret;
> > >> + } else if ((payload & 0x0a) == 0x08) {
> > >> + ret = snprintf(buf, buf_len, " SVE");
> > >> + buf += ret;
> > >> + blen -= ret;
> > >> + if (payload & 0x4) {
> > >> + ret = snprintf(buf, buf_len, " PRED");
> > >> + buf += ret;
> > >> + blen -= ret;
> > >> + }
> > >> + if (payload & 0x80) {
> > >> + ret = snprintf(buf, buf_len, " SG");
> > >> + buf += ret;
> > >> + blen -= ret;
> > >> + }
> > >> + /* Bits [7..4] encode the vector length */
> > >> + ret = snprintf(buf, buf_len, " EVLEN%d",
> > >> + 32 << ((payload >> 4) & 0x7));
> > >
> > > Same comment as above. Maybe have a common helper for decoding the
> > > vector length bits so it can be fixed in a single place?
> >
> > Yup. Although I wonder if this is the smallest of the problems with this
> > function going forward.
> >
> > Cheers,
> > Andre
>
> Fair enough.
>
> Cheers
> ---Dave

\
 
 \ /
  Last update: 2020-09-29 04:19    [W:0.061 / U:0.468 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site