Re: [Freedreno] [PATCH 2/3] drm/msm/dpu: Set DATABUS_WIDEN on command mode encoders
From: Abhinav Kumar
Date: Thu Jun 22 2023 - 19:37:28 EST
On 6/22/2023 4:14 PM, Dmitry Baryshkov wrote:
On 23/06/2023 01:37, Abhinav Kumar wrote:
On 6/21/2023 4:46 PM, Dmitry Baryshkov wrote:
On 22/06/2023 02:01, Abhinav Kumar wrote:
On 6/21/2023 9:36 AM, Dmitry Baryshkov wrote:
On 21/06/2023 18:17, Marijn Suijten wrote:
On 2023-06-20 14:38:34, Jessica Zhang wrote:
<snip>
+ if (phys_enc->hw_intf->ops.enable_widebus)
+ phys_enc->hw_intf->ops.enable_widebus(phys_enc->hw_intf);
No. Please provide a single function which takes necessary
configuration, including compression and wide_bus_enable.
There are two ways to look at this. Your point is coming from the
perspective that its programming the same register but just a
different
bit. But that will also make it a bit confusing.
My point is to have a high-level function that configures the
INTF for
the CMD mode. This way it can take a structure with necessary
configuration bits.
Hi Dmitry,
After discussing this approach with Abhinav, we still have a few
questions about it:
Currently, only 3 of the 32 bits for INTF_CONFIG2 are being used
(the
rest are reserved with no plans of being programmed in the
future). Does
this still justify the use of a struct to pass in the necessary
configuration?
No. The point Dmitry is making is **not** about this concidentally
using the same register, but about adding a common codepath to enable
compression on this hw_intf (regardless of the registers it needs to
touch).
Actually to setup INTF for CMD stream (which is equal to setting up
compression at this point).
Yes it should be setup intf for cmd and not enable compression.
Widebus and compression are different features and we should be able
to control them independently.
We just enable them together for DSI. So a separation is necessary.
But I am still not totally convinced we even need to go down the
path for having an op called setup_intf_cmd() which takes in a
struct like
struct dpu_cmd_intf_cfg {
bool data_compress;
bool widebus_en;
};
As we have agreed that we will not touch the video mode timing
engine path, it leaves us with only two bits.
And like I said, its not that these two bits always go together. We
want to be able to control them independently which means that its
not necessary both bits program the same register one by one. We
might just end up programming one of them if we just use widebus.
Thats why I am still leaning on keeping this approach.
I do not like the idea of having small functions being called between
modules. So, yes there will a config of two booleans, but it is
preferable (and more scalable) compared to separate callbacks.
I definitely agree with the scalable part and I even cross checked
that the number of usable bitfields of this register is going up from
one chipset to the other although once again that depends on whether
we will use those features.
For that reason I am not opposed to the struct idea.
But there is also another pattern i am seeing which worries me. Usable
bitfields sometimes even reduce. For those cases, if we go with a
pre-defined struct it ends up with redundant members as those
bitfields go away.
With the function op based approach, we just call the op if the
feature bit / core revision.
So I wanted to check once more about the fact that we should consider
not just expansion but also reduction.
As we have to support all generations, there is no actual reduction.
Newer SoCs do not have particular feature/bit, but older ones do. By
having multiple optional ops we just move this knowledge from
ops->complex_callback() to _setup_block_ops(). But more importantly the
caller gets more complicated. Instead of just calling ops->set_me_up(),
it has to check all the optional callbacks and call the one by one.
Alright, I am thinking that perhaps because this register is kind of
unique that its actually controlling a specific setting in the datapath,
downstream also has separate ops for this.
But thats fine, we can go ahead with the struct based approach.
Not to mention that it allows us to program required registers
directly (by setting values) rather than using RMW cycles and thus
depending on the value being previously programmed to these registers.
This will not change. We will still have to use RMW cycles to preserve
the reset values of some of the fields as those are the right values
for the registers and shouldnt be touched.
I'd like to point to the dpu_hw_intf_setup_timing_engine(), a close
rival callback, setting up the INTF for video mode. It does not do RMW
cycles, it just writes all the registers.
Yes as it ends up ORing pretty much all the usable bits. So it works fine.
In the worst case, there will be a single RMW instead of having multiple
of them.
Similar to how dpu_hw_intf_setup_timing_engine() programs the
hw_intf - including widebus! - for video-mode.
Or even more generically, have a struct similar to intf_timing_params
that says how the intf needs to be configured - without the caller
knowing about INTF_CONFIG2.
struct dpu_hw_intf_cfg is a very good example of how we can use a
single
struct and a single callback to configure multiple registers at once
based on some input parameters.
In addition, it seems that video mode does all its INTF_CONFIG2
configuration separately in dpu_hw_intf_setup_timing_engine(). If we
have a generic set_intf_config2() op, it might be good to have it as
part of a larger cleanup where we have both video and command
mode use
the generic op. What are your thoughts on this?
Not in that way, but if there is a generic enable_compression() or
configure_compression() callback (or even more generic, similar to
setup_intf_cfg in dpu_hw_ctl) that would work for both video-mode and
command-mode, maybe that is beneficial.
I'd rather not do this. Let's just 'setup timing enging' vs 'setup
CMD'. For example, it might also include setting up other INTF
parameters for CMD mode (if anything is required later on).
Agreed on setup CMD but I dont know whether we need a setup CMD at all.
Seems like an overkill.
- Marijn