AM2432: am2432 multiplication in PRU

Rasty Slutsker

Expert 1890 points

Part Number: AM2432
Other Parts Discussed in Thread: AM2612

Tool/software:

I need some advice with PRU coding.

I see differences in code generation for pru in am243x comparing to am572x

It looks code generated for am572x vs am243x is bit different.

Can someone shed some light concerning integer multiplication (both 16 and 32 bits) mechanisms and efficiency on am243x?

Thanks

Rasty

5 months ago

0 PratheeshGangadhar 5 months ago

TI__Guru 50291 points

I suggest using multiply accumulate widget for this. Hopefully this sample helps: open-pru/academy/mac/mac_multiply at main · TexasInstruments/open-pru

0 Rasty Slutsker 5 months ago in reply to PratheeshGangadhar

Expert 1890 points

Thank you very much for fast reply.

1. I do not see am243x in the list of examples.

2. Is there some "theory of operation"? What happens behind "XIN"?

3. Is there some compatibility list: am243x vs am572x vs am26xx ? can I assume that the same functionality available in all?

Thanks

Rasty

0 PratheeshGangadhar 5 months ago in reply to Rasty Slutsker

TI__Guru 50291 points

For theory refer to TRM : AM64x/AM243x Technical Reference Manual (Rev. H)

The PRU_ICSSG broadside interface enables PRU cores on the AM243x to communicate with the integrated Multiplier with Accumulation (MPY/MAC) accelerator. This interface uses the XFR instruction to directly access a dedicated set of PRU registers (R25-R29) that control the MAC, provide operands, and retrieve results, allowing for operations to complete in a single clock cycle.

PRU MAC and Broadside Interface Operation

Each of the six PRU cores within a PRU_ICSSG (PRU0/PRU1, RTU_PRU0/RTU_PRU1, TX_PRU0/TX_PRU1) has a dedicated unsigned MPY/MAC unit. The MAC is directly connected to the PRU's internal registers R25 through R29. The broadside interface and XFR instructions are used to control the MAC's mode and to import the multiplication results back into the PRU.

The key features of the PRU MPY/MAC are:

Configurable for "Multiply Only" and "Multiply and Accumulate" functionality.
One clock cycle per operation.
Uses the broadside interface and XFR instructions (XIN, XOUT) for importing results and initiating the accumulate function.
Allows firmware to seed the accumulator.

Register Mapping for MAC Operations

The MAC's functionality is directly mapped to a specific set of PRU registers, which are accessed via the broadside interface.

Register	Function
R25	Configures the MAC for either Multiply Only or Multiply and Accumulate mode.
R28, R29	Provide the 32-bit operands for the multiplication.
R26, R27	Receive the 64-bit result, which includes a carry flag.

On the AM243x, each of the two PRU_ICSSG subsystems includes these MPY/MAC accelerators for its PRU, RTU_PRU, and TX_PRU cores.

; Step 1: Enable multiply only MAC_MODE
LDI R25, MULTIPLY_ONLY ; Set R25[0] = 1 for MAC mode
XOUT DEVICE_ID, &R25, 1 ; Enable MAC mode

; Multiply the operands
MOV R28, R10 ; Load a into first operand register
MOV R29, R11 ; Load b into second operand register
NOP ; It takes 1 PRU clock to multiply a * b

; Read final result
XIN DEVICE_ID, &R25, 1 ; Read MAC status
XIN DEVICE_ID, &R26, 8 ; Read result into R27: R26

0 PratheeshGangadhar 5 months ago in reply to PratheeshGangadhar

TI__Guru 50291 points

Rasty Slutsker said:
Is there some compatibility list: am243x vs am572x vs am26xx ? can I assume that the same functionality available in all?

Yes, it is available in all

0 Rasty Slutsker 5 months ago in reply to PratheeshGangadhar

Expert 1890 points

OK. Final question. Why code generated by different version of compiler looks differently? newer version introduces _mpyi wrapper call, why older injects XIN instruction directly? Is it possible to control this behavior? I'm asking because we would like to wite code in C, rather than assembly.

Thanks

Rasty

+1 Nick Saulnier 5 months ago in reply to Rasty Slutsker

TI__Guru** 108080 points

Hello Rasty,

Features per PRU subsystem

For a full list of hardware features per PRU subsystem, please refer to app note "PRU Subsystem Features Comparison"
https://www.ti.com/lit/sprac90

Programming the MAC from C

We also have example code for doing this on both AM57x and AM243x/AM64x (these parts are basically identical, AM243x is just missing the A53 cluster). Please refer to the AM64x example in the older PRU Software Support Package (PSSP):
https://git.ti.com/cgit/pru-software-support-package/pru-software-support-package/tree/examples/am64x/PRU_MAC_Multiply_Accum

The open-pru repo is still very early in development. It will take us some time to port the existing assembly example to AM243x & AM64x, and port the PSSP C code example to the open-pru repo.

Later this month we will be publishing the first version of the AM243x PRU Academy (you will see it as a new module in the AM243x academy: https://dev.ti.com/tirex/explore/node?node=A__AEIJm0rwIeU.2P1OBWwlaA__AM24X-ACADEMY__ZPSnq-h__LATEST ). Here's the Academy's MAC page:

https://e2e.ti.com/cfs-file/__key/communityserver-discussions-components-files/908/Multiplier-with-Accumulation-_2800_MPY-or-MAC_2900_-_1420_-AM64x-Academy.mhtml

Any additional thoughts on programming in C?

First off, please note that we have full support for loading C firmware from the Linux remoteproc driver (e.g., if you were using AM64x). However, there can be some issues with loading C code with the MCU+ PRUICSS driver. Fixing that behavior is still on the TODO list. If you run into problems, please reach out to us.

FYI: PRU Academy notifications

If you want to get notified when we push the PRU academy (and future updates to the PRU academy), please reach out to me in a direct message.

Regards,

Nick

0 Rasty Slutsker 5 months ago in reply to Nick Saulnier

Expert 1890 points

Thanks you very much for clarifications.

0 Rasty Slutsker 5 months ago in reply to Nick Saulnier

Expert 1890 points

Is it possible to get examples of inline assembly?

0 Nick Saulnier 5 months ago in reply to Rasty Slutsker

TI__Guru** 108080 points

I will add inline assembly examples over the next week or so to OpenPRU & the PRU Academy.

If you need to see them this week, you can refer to the AM64x code in the existing PRU Getting Started Labs in the PSSP:
https://software-dl.ti.com/processor-sdk-linux/esd/AM64X/11_01_05_03/exports/docs/common/PRU-ICSS/PRU-Getting-Started-Labs_Lab2.html
https://git.ti.com/cgit/pru-software-support-package/pru-software-support-package/tree/labs/Getting_Started_Labs

Regards,

Nick

0 Daniel Watson 5 months ago in reply to Nick Saulnier

Prodigy 185 points

Hi Nick!
I am having problems loading a PRU firmware written in C on the AM2612 using the PRUICSS driver, maybe you could provide me with an workaround?

LP-AM261: Problems when loading C application on the PRU programmatically (stuck in __TI_decompress_lzss) - Arm-based microcontrollers forum - Arm-based microcontrollers - TI E2E support forums

Best regards
Daniel

0 Rasty Slutsker 5 months ago in reply to Daniel Watson

Expert 1890 points

Hi Daniel

I have opposite problem, cannot load via JTAG AM2432: cannot load program into PRU

Thanks

Rasty

0 Nick Saulnier 5 months ago in reply to Rasty Slutsker

TI__Guru** 108080 points

Hello Rasty,

Following up on a couple of comments from offline:

1) inconsistent implementation of MAC operations

Have you checked different optimization levels to see if you get different outputs?

It has been a few years since I messed around with the PSSP MAC example, so I cannot remember what the assembly looked like. It looks like I used optimization level 2 (-O2) in the makefile: https://git.ti.com/cgit/pru-software-support-package/pru-software-support-package/tree/examples/am64x/PRU_MAC_Multiply_Accum/Makefile

but you can change that optimization from none to 4
C code optimization: -O (off, 0, 1, 2, 3, 4)

2) Suggested assembly implementation of MAC operations

I can either point you to the existing resources in several different places, or we can circle back after I publish the PRU academy in a couple weeks and I can point to an inline assembly example that you could combine with the MAC assembly example listed above. Let me know which you prefer.

Regards,

Nick

0 Rasty Slutsker 5 months ago in reply to Nick Saulnier

Expert 1890 points

1) please you look into the first message of this thread. I did this reseach - code generation does not not depend on compiler flags.

2) I can wait a couple of weeks.

0 Nick Saulnier 5 months ago in reply to Rasty Slutsker

TI__Guru** 108080 points

Hello Rasty,

Sounds good. Please set a calendar reminder to ping me in early October.

Regards,

Nick