[PATCH] clk: sunxi-ng: sun50i: h6: Modify GPU clock configuration to support DFS
Clément Péron
peron.clem at gmail.com
Tue Jun 28 05:58:59 PDT 2022
Hi Roman, Jernej,
On Sat, 25 Jun 2022 at 16:02, Roman Stratiienko <r.stratiienko at gmail.com> wrote:
>
> PS:
>
> For better DFS resolution P or GPU_CLK divider can be preprogrammed
> and GPU_PLL can be marked with fixed out divider. It's safe. Not safe
> to touch these regs during runtime.
> But do we really need that resolution?
>
> сб, 25 июн. 2022 г. в 16:27, Roman Stratiienko <r.stratiienko at gmail.com>:
> >
> > Hi,
> >
> > DVFS was tested as DVFS using devfreq driver, not the script.
> >
> > The following OPP table was used:
> > https://github.com/clementperon/linux/commit/add3ef683238095d2721de03601d5b01f2d9ce22
I now remember when I tried to enable GPU devfreq on H6 and noticed
instability, I made a recap of my searches:
https://lore.kernel.org/lkml/CAJiuCce58Gaxf_Qg2cnMwvOgUqYU__eKb3MDX1Fe_+47htg2bA@mail.gmail.com/
And found that Megous gave a possible explanation on linux-sunxi IRC:
20:12 <megi> looks like gpu pll on H6 is NKMP clock, and those are
implemented in such a way in mainline that they are prone to
overshooting the frequency during output divider reduction
20:13 <megi> so disabling P divider may help
20:13 <megi> or fixing the dividers
20:14 <megi> and just allowing N to change
20:22 <megi> hmm, I haven't looked at this for quite some time, but H6
BSP way of setting PLL factors actually makes the most sense out of
everything I've seen/tested so far
20:23 <megi> it waits for lock not after setting NK factors, but after
reducing the M factor (pre-divider)
20:24 <megi> I might as well re-run my CPU PLL tester with this
algorithm, to see if it fixes the lockups
20:26 <megi> it makes sense to wait for PLL to stabilize "after"
changing all the factors that actually affect the VCO, and not just
some of them
20:27 <megi> warpme_: ^
20:28 <megi> it may be the same thing that plagues the CPU PLL rate
changes at runtime
Regards,
Clement
> >
> > As is already mentioned in the commit message, P causes the issues as well.
> >
> > Regards,
> > Roman
> >
> > сб, 25 июн. 2022 г. в 13:43, Jernej Škrabec <jernej.skrabec at gmail.com>:
> >
> > >
> > > Hi Roman,
> > >
> > > Dne petek, 24. junij 2022 ob 18:52:11 CEST je Roman Stratiienko napisal(a):
> > > > Using simple bash script it was discovered that not all CCU registers
> > > > can be safely used for DFS, e.g.:
> > > >
> > > > while true
> > > > do
> > > > devmem 0x3001030 4 0xb0003e02
> > > > devmem 0x3001030 4 0xb0001e02
> > > > done
> > > >
> > > > Script above changes the GPU_PLL multiplier register value. While the
> > > > script is running, the user should interact with the user interface.
> > > >
> > > > Using this method the following results were obtained:
> > > > | Register | Name | Bits | Values | Result |
> > > > | -- | -- | -- | -- | -- |
> > > > | 0x3001030 | GPU_PLL.MULT | 15..8 | 20-62 | OK |
> > > > | 0x3001030 | GPU_PLL.INDIV | 1 | 0-1 | OK |
> > > > | 0x3001030 | GPU_PLL.OUTDIV | 0 | 0-1 | FAIL |
> > > > | 0x3001670 | GPU_CLK.DIV | 3..0 | ANY | FAIL |
> > > >
> > > > Once bits that caused system failure disabled (kept default 0),
> > > > it was discovered that GPU_CLK.MUX was used during DFS for some
> > > > reason and was causing the failure too.
> > > >
> > > > After disabling GPU_PLL.OUTDIV the system started to fail during
> > > > booting for some reason until the maximum frequency of GPU_PLL
> > > > clock was limited to 756MHz.
> > > >
> > > > After all the changes made DVFS started to work seamlessly.
> > >
> > > I appreciate testing effort, but I don't think userspace approach is good way
> > > for testing DVFS. I see 2 issues:
> > > - As name already suggest, voltage also plays crucial role for stability. You
> > > didn't say on which board you tested this, but I assume it has PMIC. Did you
> > > make sure GPU voltage regulator is always at 1.04 V, which is needed for 756
> > > MHz?
> > > - Kernel clock driver always goes through proper procedure for clock rate
> > > change, which involves several steps. Bypassing them might also cause some
> > > stability problems.
> > >
> > > I agree that GPU PLL should be limited to 756 MHz max. This seems to be
> > > maximum operating point specified at vendor DT. But I managed to extract some
> > > more information from vendor GPU driver. More specifically, from this snippet,
> > > located in modules/gpu/mali-midgard/kernel_mode/driver/drivers/gpu/arm/
> > > midgard/platform/sunxi/mali_kbase_config_sunxi.c:
> > >
> > > pll_freq = target->freq;
> > > while (pll_freq < 288000000)
> > > pll_freq *= 2;
> > >
> > > err = clk_set_rate(sunxi_mali->gpu_pll_clk, pll_freq);
> > > <...>
> > > err = clk_set_rate(kbdev->clock, target->freq);
> > > <...>
> > >
> > > Apparently, minimum stable PLL frequency is 288 MHz (this should be added) and
> > > divider in peripheral clock can really be used, although preferably not.
> > > Vendor GPU operating points specify only 2 lower than 288 MHz points - at 264
> > > MHz and 216 MHz. I'm fully aware that they may not be really stable and given
> > > that these two and next two all share minimum voltage of 810 mV, power and
> > > thermal savings are probably not that great, so we can skip them and pin
> > > peripheral divider to 1, as you already did.
> > >
> > > Another discrepancy I see is that vendor DT has two operating points, at 336
> > > MHz and 384 MHz, which also use factor P (also known as d2 in vendor clock
> > > source). This can be again an oversight or alternatively, it can be that P
> > > factor can actually be used, but just with lower frequencies.
> > >
> > > Can you please make another test with GPU operating points specified in DT and
> > > check if it works with P factor left in?
> > >
> > > For reference, vendor DT has following operating points (kHz, uV):
> > > 756000 1040000
> > > 624000 950000
> > > 576000 930000
> > > 540000 910000
> > > 504000 890000
> > > 456000 870000
> > > 432000 860000
> > > 420000 850000
> > > 408000 840000
> > > 384000 830000
> > > 360000 820000
> > > 336000 810000
> > > 312000 810000
> > > 264000 810000
> > > 216000 810000
> > >
> > > Best regards,
> > > Jernej
> > >
> > > >
> > > > Signed-off-by: Roman Stratiienko <r.stratiienko at gmail.com>
> > > > ---
> > > > drivers/clk/sunxi-ng/ccu-sun50i-h6.c | 12 +++++-------
> > > > 1 file changed, 5 insertions(+), 7 deletions(-)
> > > >
> > > > diff --git a/drivers/clk/sunxi-ng/ccu-sun50i-h6.c
> > > > b/drivers/clk/sunxi-ng/ccu-sun50i-h6.c index 2ddf0a0da526f..d941238cd178a
> > > > 100644
> > > > --- a/drivers/clk/sunxi-ng/ccu-sun50i-h6.c
> > > > +++ b/drivers/clk/sunxi-ng/ccu-sun50i-h6.c
> > > > @@ -95,13 +95,14 @@ static struct ccu_nkmp pll_periph1_clk = {
> > > > },
> > > > };
> > > >
> > > > +/* For GPU PLL, using an output divider for DFS causes system to fail */
> > > > #define SUN50I_H6_PLL_GPU_REG 0x030
> > > > static struct ccu_nkmp pll_gpu_clk = {
> > > > .enable = BIT(31),
> > > > .lock = BIT(28),
> > > > .n = _SUNXI_CCU_MULT_MIN(8, 8, 12),
> > > > .m = _SUNXI_CCU_DIV(1, 1), /* input divider */
> > > > - .p = _SUNXI_CCU_DIV(0, 1), /* output divider
> > > */
> > > > + .max_rate = 756000000UL,
> > > > .common = {
> > > > .reg = 0x030,
> > > > .hw.init = CLK_HW_INIT("pll-gpu", "osc24M",
> > > > @@ -294,12 +295,9 @@ static SUNXI_CCU_M_WITH_MUX_GATE(deinterlace_clk,
> > > > "deinterlace", static SUNXI_CCU_GATE(bus_deinterlace_clk,
> > > > "bus-deinterlace", "psi-ahb1-ahb2", 0x62c, BIT(0), 0);
> > > >
> > > > -static const char * const gpu_parents[] = { "pll-gpu" };
> > > > -static SUNXI_CCU_M_WITH_MUX_GATE(gpu_clk, "gpu", gpu_parents, 0x670,
> > > > - 0, 3, /* M */
> > > > - 24, 1, /* mux */
> > > > - BIT(31), /* gate */
> > > > - CLK_SET_RATE_PARENT);
> > > > +/* GPU_CLK divider kept disabled to avoid interferences with DFS */
> > > > +static SUNXI_CCU_GATE(gpu_clk, "gpu", "pll-gpu", 0x670,
> > > > + BIT(31), CLK_SET_RATE_PARENT);
> > > >
> > > > static SUNXI_CCU_GATE(bus_gpu_clk, "bus-gpu", "psi-ahb1-ahb2",
> > > > 0x67c, BIT(0), 0);
> > >
> > >
> > >
> > >
More information about the linux-arm-kernel
mailing list