imx28 ABPX ABPH Peripherals, Speed vs. imx6

[wolfgang_gaerbe]

Dear All,

Using imx28 bare.metal, we found out that the typical access penalties for certain peripherals are higher than expected:

GPIO:

(this values are somewhat expected)

ABPX-BUS: write: 2-3 X Cycles -> 85-125ns

                  read:  4-6 X Cycles ->170-250ns

This would equal 50-100 waitstates if clocked with 454MHz.

DEBUG UART:

(this was some surprise)

ABPX-BUS: read, write: 6 X Cycles -> 250ns

Using 32 Byte tx - fifo - this means that filling up the fifo may take around 8us, if you do that in a while loop checking for fifo full - 16us.

Bascially - using the Debug Uart is like having a 4MHz peripheral with corresponding waitstates.

This would equal 114 waitstates if clocked with 454MHz.

SPI / SSP:

(some lecture on H Bus)

ABPH-BUS: read, write: 8H Cycles -> 50ns

Because of multiplexed data/address on H bus + some clocks to overcome the different time domains - the actual access time

is around 50ns - which would correspond to access a 24MHz peripheral.

This would equal 23 waitstates if clocked with 454MHz.

This is the reason why using DMA for X and H bus is highly recommended - but not available for debug uart.

Another option is to separate consecutive writes - but thats no option for reads.

Has somebody experience with imx6 peripherals, busses ?

Is the performance similar as described here ?

rgds.

[wolfgang_gaerbe]

AHB / PTP ACCESS

Because of access problems to ENET 1588PTP HW - we assume that there is extra delay needed - if you address AHB blocks running at lower speed than expected.

Reading ATIME register (with ptp clock =40MHz)- we need to wait 72ns after issued capture command from control register.

This would mean that there is no synchronization /halt / auto generated waitstate if you address or readout  ahb i/o which runs slower than "expected".

Even write -> reread written register -> read dependent register doesn´t work.

So a verified written register is no proof that a precondition is given.

rgds.

[RaymondChoi]

Hi Wolfgang,

I am sorry that we do not have such data for MX6.

Best regards

Raymond