On 14:56-20240202, Krzysztof Kozlowski wrote: > On 02/02/2024 13:52, Nishanth Menon wrote: > > On 11:47-20240202, Krzysztof Kozlowski wrote: > >> On 01/02/2024 19:42, Brandon Brnich wrote: > >>> Wave521c has capability to use SRAM carveout to store reference data with > >>> purpose of reducing memory bandwidth. To properly use this pool, the driver > >>> expects to have an sram and sram-size node. Without sram-size node, driver > >>> will default value to zero, making sram node irrelevant. > >> > >> I am sorry, but what driver expects should not be rationale for new > >> property. This justification suggests clearly it is not a property for DT. > >> > > > > Yup, the argumentation in the commit message is from the wrong > > perspective. bindings are OS agnostic hardware description, and what > > driver does with the description is driver's problem. > > > > I will at least paraphrase my understanding: > > In this case, however, the hardware block will limp along with > > the usage of DDR (as is the current description), due to the > > latencies involved for DDR accesses. However, the hardware block > > has capability to use a substantially lower latency SRAM to provide > > proper performance and hence for example, deal with higher resolution > > data streams. This SRAM is instantiated at SoC level rather than > > embedded within the hardware block itself. > > That sounds like OS policy. Why would different boards with the same > component have this set differently? Based on amount of available > memory? This, I believe, is runtime configuration because it might > depend on user-space you run. Based on purpose (e.g. optimize for > decoding or general usage)? Again, run-time because same hardware board > can be used for different purposes. > Why is this OS policy? It is a hardware capability. Traditionally many similar hardware blocks would have allocated local SRAM for worst case inside the hardware block itself and don't need to use DDR in the first place. However, for this hardware block, it has capability to use some part of one of the many SRAM blocks in the SoC, not be shared for some part of the system - so from a hardware description perspective, we will need to call that out as to which SRAM is available for the hardware block. Why would different boards need this differently? simply because different cameras have different resolution and framerates - and you dont want to pay the worst case sram penalty for all product configuration. Further, Linux is not the only thing that runs on these SoCs.. these are mixed systems with autonomous operations of uC cores who may or maynot (typically not) even need to communicate with MPU to state which part of resource they are hogging (hence the board level definition). -- Regards, Nishanth Menon Key (0xDDB5849D1736249D) / Fingerprint: F8A2 8693 54EB 8232 17A3 1A34 DDB5 849D 1736 249D
