I'm trying to implement a rather complex (at least for me!) system on a Cortex M4 mcu: LPC4370. This one has HighSpeed ADC (up to 80Msps), DMA and DSP (Single Instruction Multiple Data) instructions. Here's what I want to do:

  • Let the ADC sample in a continuous manner (at least at 10 Msps)
  • Move the data to SRAM
  • Process them in real-time w/ cortex M4 DSP (pulse shaping filtering)

MCU clock is 204 MHz and for now let's assume ADC fs is not a design spec, but ideally I would like it to be as high as it can. So I need the code to be as fast as I can. Here's the MCU memory set up.

enter image description here And here the AHB MULTILAYER MATRIX

enter image description here As of now I am considering more the general firmware architecture, and specifically memory management. Some considerations:

  • I don't want core M4 and DMA to fight for memory: I need the DMA to be able to write data while the M4 is doing the processing
  • most of the code and all acquired data should be in SRAM for faster execution
  • Instruction fetch should not interfere with data storing (DMA) and processing (M4)

In the LPC4370 user guide (chap.2):

To optimize the CPU performance, the ARM Cortex-M4 has three buses for Instruction (code) (I) access, Data (D) access, and System (S) access. The I- and D-bus access memory space is located below 0x2000 0000, the S-bus accesses the memory space staring from 0x2000 0000. When instructions and data are kept in separate memories, then code and data accesses can be done in parallel in one cycle. When code and data are kept in the same memory, then instructions that load or store data may take two cycles.

My idea at the moment is to hold the sampled data in two different buffers placed in two different memory areas (like LocRam128 and LocRam72) and "ping pong" DMA and M4 between these two areas. Only problem is that these are the two areas used for instruction (I). So I guest code should be placed here also, and this is not good to me. I wonder how could I use RAMAHB32 effectively since it is only connected to M4 system bus (S) and not to data (D) nor instructions (I).

Any hints?

  • 3
    \$\begingroup\$ To be honest, this is going to be a tough question to answer. I've written a "High speed ADC" routine for an M4. But it tops out at about 1800kHZ total sample throughput. I was doing a ton of math on the samples which really limited the throughput (not to mention other housekeeping). Honestly, you need to start this with a real sample rate requirement. You may find you are CPU bound rather than memory contention bound. \$\endgroup\$
    – pgvoorhees
    Commented Sep 30, 2019 at 15:24
  • \$\begingroup\$ Hi @pgvoorhees thanks for the reply. I got what you mean, and that's why I don't want to set the sample frequency strictly. Do you know if there's a way to use the system (S) bus? \$\endgroup\$
    – a_bet
    Commented Sep 30, 2019 at 15:27
  • \$\begingroup\$ Let me reiterate: this question is impossible to answer without knowing what you are going to do with the samples. If you are going to do nothing with them, than 80Mhz is a very achievable goal: Simply allocate a buffer in memory and continuously overwrite it using DMA and never let the processor intervene. If you want to do something other than that, you must specify what it is because the highest possible sample rate and strategy to achieve that sample rate changes with the end need. \$\endgroup\$
    – pgvoorhees
    Commented Sep 30, 2019 at 18:13
  • \$\begingroup\$ Hi @pgvoorhees I added some info (I cannot describe the project in details): I need at least 10 Mhz and I want to do pulse shaping on my data. I'm already trying to write my processing function in the best way possible using SIMD instrinsics. \$\endgroup\$
    – a_bet
    Commented Oct 1, 2019 at 15:48
  • 1
    \$\begingroup\$ I think you are missing the point that @pgvoorhees is trying to make. You are asking about subtle details of memory management and DMA for a very specific processor, but it appears that the processor in question is not suitable for other reasons. So why waste time thinking about this processor? \$\endgroup\$ Commented Oct 1, 2019 at 15:54

1 Answer 1


Ok, since you are unable to share more details. I'll give you some general points:

  1. the Scatter-gather functionality in the DMA module is going to save your bacon; Take time to understand how it works and how to use it.
  2. If you are worried about memory accesses, just go ahead and place your ping and pong buffers in different memories. Scatter-gather will help facilitate this.
  3. After the above, don't worry about bus contention until you get there. Realistically: if, while using separate memories, bus contention is STILL your bottle neck, then you've got the wrong chip. Pure and simple.
  4. Invest in a Segger j-trace and implement streaming trace in your debug board. This will help you when you need to troubleshoot timing issues. Yes it's expensive.
  5. Take time to experiment with your processing loop, size your ping and pong buffers based on processing loop time. You may also have to get creative with doing partial workloads to meet your deadlines.
  6. I needed to rewrite some of the CMSIS DSP functions to be faster.
  7. Don't be afraid to dig in to the CMSIS libraries, they are very readable AND they provide a good example of SIMD processing.
  8. when I was benchmarking my code, i found that leaving my firmware .data section in flash didn't give me a super-significant performance hit over ram-located .data. That surprised me.
  9. Use fixed-point data everywhere, convert to float only at the end and only if necessary.

Hope this helps.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.