Hi Nicolas,
On 9/28/23 01:50, Chautru, Nicolas wrote:
Hi Maxime, Hemant,
I wanted initially to keep it fairly open hence a hash table for the windows
profiles, but it is also possible to expose something more descriptive, that
would work as well actually.
Ie.
+ /** FFT windowing width for 2048 FFT. */
+ uint16_t fft_window_width[RTE_BBDEV_MAX_FFT_WIN];
The provides the width of each windows shape which is enough to distinguish
major variants and to estimate noise factor.
That sounds much better IMHO.
Regarding the array and values sizes:
1. Should it only covers 2048 points FFT? I see some references about
4096 FFT for 5G and satellites communications
2. Is uint16_t enough for all the usecases?
Since this array is quite big, could it be exposed to the application
via dedicated APIs instead of a field? An API to query the length of the
array so that the application can allocate required meory, and an API to
copy the data in the allocated mem?
Maybe overkill, but I feel different FFT size could be supported in the
future, so that would be both future proof and more memory efficient for
apps that don't need this.
Let me know of opinion.
Thanks for suggesting this,
Maxime
Thanks
Nic
-----Original Message-----
From: Maxime Coquelin <maxime.coque...@redhat.com>
Sent: Tuesday, September 26, 2023 3:00 AM
To: Chautru, Nicolas <nicolas.chau...@intel.com>;
hemant.agra...@nxp.com; dev@dpdk.org
Cc: david.march...@redhat.com; Vargas, Hernan
<hernan.var...@intel.com>
Subject: Re: [PATCH v1 1/7] bbdev: add FFT version member in driver info
On 9/22/23 18:41, Chautru, Nicolas wrote:
Hi Maxime,
-----Original Message-----
From: Maxime Coquelin <maxime.coque...@redhat.com>
Sent: Friday, September 22, 2023 1:15 AM
To: Chautru, Nicolas <nicolas.chau...@intel.com>;
hemant.agra...@nxp.com; dev@dpdk.org
Cc: david.march...@redhat.com; Vargas, Hernan
<hernan.var...@intel.com>
Subject: Re: [PATCH v1 1/7] bbdev: add FFT version member in driver
info
Hi Nicolas,
On 9/19/23 22:51, Chautru, Nicolas wrote:
Hi Maxime,
This is neither part of 3GPP per se, nor specific to VRB device. Let
me provide
more context.
The SRS processing chain
(https://doc.dpdk.org/guides/prog_guide/bbdev.html#bbdev-fft-operatio
n) includes a pointwise multiplication by time window.
The generic API include some control of these windowing function but
still
the actual shape need to be programmed onto any device (ie.
rectangular, taped, sinc, different width or offset, any abritraty
shape defined as an array of scalars). These degrees of liberties
cannot be exposed through a generic API (information is multi-kB, ie
the data itself) and can be user specific (external to the HW IP itself or
outside of Intel control).
Thanks for the explanations. I also did my homework as my FFT
knowledge was buried quite deep in my memory. :)
So this is a vendor-specific way to express generic paramaters.
Unsure this is that vendor specific. At least the interface allows to know a
hash of the table being loaded (which is just pointwise data really, non-
proprietary format). I did not state the content is a simple md5sum of the bin
file being loaded from linux.
Ok, I think it would be better to provide an API to get the table directly, and
have the format being described in the documentation.
With that, we can also provide the hash as you'd like, but the method to
calculate the hash should also be provided. Or the application can perform
the hash itself if it needs it.
The fact that it is several KB is not an issue, as this information would only
be
queried once at init time if really needed.
An non-DPDK alternative could be to pass such information to the pod via the
device plugin (as a mounted file for instance, or variable).
Regarding VRB device, is this table per device or per VF?
Could it be configured by the application directly, or has it to be
done through the PF?
This is configured for the device at platform level, ie. through operator.
Common to all application/devices. This captures the windows shape
assumptions.
Thanks for the information!
As an illustration for VRB device pf_bb_config provides to user an
option to
include such windowing data as an input ("FFT LUT bin file"), but
more generally at platform level for any bb device this big Look-Up
Table or big array can be configured on the host during platform
initialization for a given deployment or vendor.
What is required here is for the user application to have knowledge
of what
version of such array is being used on the given platform, as this
information would be relevant to processing done outside of bbdev
(notably for noise estimate). Through that mechanism, the user can
now map through that API which possible file was being used, and act
accordingly.
The content itself is not specified, for VRB we just use the md5sum
of that
binary file (which is just a big array of int16 for point wise
multiplication) so that this can be used to share knowledge between
initialized platform configuration and at run-time user application
assumption.
It is also important to under that the user/vendor may use any array
or
shape (based on their algorithm) regardless of Intel or IP, and still
be able to share information mapping between what is configured on
the platform (multiple versions possible) and what the application
enumerates.
I can add more details in the documentation indeed but above should
arguably make sense. The name FFT_version naming may be quite vague,
this is more related to the FFT pointwise windowing array variant
assumed on the platform. I did not want to impose for it to be an
md5sum necessarily, hence the vagueness, as it could be any hash
shared between the device programming and the user application
related to the semi-static FFT processing programming.
Let me know if unclear or if any other thought,
I think this is clear now to me.
In my opinion, this is not good to have this part of the BBDEV API,
as every vendor will have their own way to represent this.
Other alternative is to have a vendor specific API. This is far from
ideal and should be avoided as much as possible, but in this case the
application has to know anyways which device it is driving. It would
be at least clear the field has to be interpreted in a vendor-specific way.
@Hemant, I would be interested in your opinion. (I don't know if NXP
has or plans to have FFT accelerator IP)
Yes looking forward to it.
Thanks,
Maxime
Regards,
Maxime
Thanks
Nic
-----Original Message-----
From: Maxime Coquelin <maxime.coque...@redhat.com>
Sent: Tuesday, September 19, 2023 2:56 AM
To: Chautru, Nicolas <nicolas.chau...@intel.com>; dev@dpdk.org
Cc: hemant.agra...@nxp.com; david.march...@redhat.com; Vargas,
Hernan
<hernan.var...@intel.com>
Subject: Re: [PATCH v1 1/7] bbdev: add FFT version member in driver
info
On 9/19/23 03:21, Nicolas Chautru wrote:
This can be used to distinguish different version of the flexible
pointwise windowing applied to the FFT and expose this to the
application.
Does this version relates to a standard, or is this specific to the
implementation of your VRB devices?
Signed-off-by: Nicolas Chautru <nicolas.chau...@intel.com>
---
lib/bbdev/rte_bbdev.h | 2 ++
1 file changed, 2 insertions(+)
diff --git a/lib/bbdev/rte_bbdev.h b/lib/bbdev/rte_bbdev.h index
a5bcc09f10..d6e54ee9a4 100644
--- a/lib/bbdev/rte_bbdev.h
+++ b/lib/bbdev/rte_bbdev.h
@@ -349,6 +349,8 @@ struct rte_bbdev_driver_info {
const struct rte_bbdev_op_cap *capabilities;
/** Device cpu_flag requirements */
const enum rte_cpu_flag_t *cpu_flag_reqs;
+ /** Versioning number for the FFT operation type. */
+ uint16_t fft_version;
};
/** Macro used at end of bbdev PMD list */
