XAxiDma_SimpleTransfer and Maximum Transfer Length

[gdg]

(Really) short version:

In my HLS design, I am trying a single AXI-DMA transfer of 126MB and the SDK function XAxiDma_SimpleTransfer fails with status 15 (max transfer size limit). How should I modify my design?

Detailed version:

In Vivado HLS, I am using an AXI DMA component to feed an HLS-home-made filter module which uses AXI stream interfaces. Here you can see the configuration of the AXI DMA and how the filter is connected to it.

I created a simple software test bench in the SDK to test the system in bare-metal:

// Pass the size of the image to the filter module.
XFilter_11x11_blur_axis_Set_height_V(&Filter_11x11_blur_axis, IMG_NROWS);
XFilter_11x11_blur_axis_Set_width_V(&Filter_11x11_blur_axis, IMG_NCOLS);
XFilter_11x11_blur_axis_Start(&Filter_11x11_blur_axis);

// Set DMA transfers (AXI DMA to/from the filter)
status = XAxiDma_SimpleTransfer(&axiDMA, (INTPTR) input_data, IMG_SIZE * sizeof(u32), XAXIDMA_DMA_TO_DEVICE);
// It fails with status 15. See the function implemention in the below snippet.
if (status != XST_SUCCESS)
{
	printf("ERROR: Sending data from DMA to IP core (status = %d)\n", status);
	return XST_FAILURE;
}

status = XAxiDma_SimpleTransfer(&axiDMA, (INTPTR) output_data, IMG_SIZE * sizeof(u32), XAXIDMA_DEVICE_TO_DMA);
if (status != XST_SUCCESS)
{
	printf("INFO: Receiving data from IP core to DMA (status = %d)\n", status);
	return XST_FAILURE;
}

// Wait for completion
while (XAxiDma_Busy(&axiDMA, XAXIDMA_DMA_TO_DEVICE)) { /* Wait */ }
while (XAxiDma_Busy(&axiDMA, XAXIDMA_DEVICE_TO_DMA)) { /* Wait */ }

while (!XFilter_11x11_blur_axis_IsDone(&Filter_11x11_blur_axis));
status = XFilter_11x11_blur_axis_Get_return(&Filter_11x11_blur_axis);

The test bench (i.e. the DMA transfer) works fine with medium-size images (HDPLUS 1600x900, FHD 1920x1080), but it fails at the first XAxiDma_SimpleTransfer (XAXIDMA_DMA_TO_DEVICE), when I try to transfer 4K and 8K UHD images (e.g., 7680x4320 ~ 126MB image). The return value is 15. In the xaxidma.c, I can see the function implementation:

u32 XAxiDma_SimpleTransfer(XAxiDma *InstancePtr, UINTPTR BuffAddr, u32 Length, int Direction)
{
  //...
  		if ((Length < 1) ||
			(Length > InstancePtr->TxBdRing.MaxTransferLen)) {
			return XST_INVALID_PARAM; // 15
		}
  //...
}

At this point, how can I transfer a big image through the AXI DMA? How can I increase the maximum transfer length? 

Thank you

[hbucher]

@gdg  It is fixed by the type of hardware configuration. Looking at xaxidma.c I see it initialized to either XAXIDMA_MAX_TRANSFER_LEN or XAXIDMA_MCHAN_MAX_TRANSFER_LEN 

if ((InstancePtr->RxNumChannels > 1) ||
		(InstancePtr->TxNumChannels > 1)) {
		MaxTransferLen =
			XAXIDMA_MCHAN_MAX_TRANSFER_LEN;
	}
	else {
		MaxTransferLen =
			XAXIDMA_MAX_TRANSFER_LEN;
	}

And they are defined in xaxidma.h as

#define XAXIDMA_MAX_TRANSFER_LEN	0x7FFFFF  /* Max length hw supports */
#define XAXIDMA_MCHAN_MAX_TRANSFER_LEN	0x00FFFF  /* Max length MCDMA
						     hw supports */

So it seams that if you have just one channel on both RX and TX then the max transfer size is 0x7F FFFF and 0xFFFF otherwise.

[gdg]

@hbucher,

so, given my system configuration the maximum size of an AXI transfer in bytes is 0x7FFFFF = 8388607 bytes ~ 8MB. That is exactly the size of the Full HD image (1920*1080 * 4bytes).

The question is still open. How can I increase the maximum size of the transfer or (automatically) implement multiple bursts through the AXI DMA?

[gdg]

I spotted a typo in a sentence:

In Vivado HLS, I am using an AXI DMA component to feed an HLS-home-made filter module which uses AXI stream interfaces.

The correct one is "In Vivado, I am using an AXI DMA".

[ I am scared of using the edit post, that behaves weird some times :-) ]

Please, note that the problem is still open.

[hbucher · Accepted solution]

@gdg

The answer is - you can't just increase the size because it is set in hardware.

You have to break it down in blocks by setting up a ring of descriptors - OR issue one simple transfer at a time. 

See the example here:

https://github.com/Xilinx/embeddedsw/blob/master/XilinxProcessorIPLib/drivers/axidma/examples/xaxidma_example_sg_poll.c

[gdg]

@hbucher,

do you have a reference or tutorial on the SDK software API of the AXI DMA with scatter gather?

The example that you pointed is also part of this example package

https://www.xilinx.com/support/answers/58080.html

but none of them provide details about the API.

For example, given the following function, I would like the explanations of the function calls. What is BD? What is a TX ring? ... etc.

static int TxSetup(XAxiDma * AxiDmaInstPtr)
{
	XAxiDma_BdRing *TxRingPtr;
	XAxiDma_Bd BdTemplate;
	int Delay = 0;
	int Coalesce = 1;
	int Status;
	u32 BdCount;

	TxRingPtr = XAxiDma_GetTxRing(&AxiDma);

	/* Disable all TX interrupts before TxBD space setup */

	XAxiDma_BdRingIntDisable(TxRingPtr, XAXIDMA_IRQ_ALL_MASK);

	/* Set TX delay and coalesce */
	XAxiDma_BdRingSetCoalesce(TxRingPtr, Coalesce, Delay);

	/* Setup TxBD space  */
	BdCount = XAxiDma_BdRingCntCalc(XAXIDMA_BD_MINIMUM_ALIGNMENT, TX_BD_SPACE_HIGH - TX_BD_SPACE_BASE + 1);

	Status = XAxiDma_BdRingCreate(TxRingPtr, TX_BD_SPACE_BASE, TX_BD_SPACE_BASE, XAXIDMA_BD_MINIMUM_ALIGNMENT, BdCount);
	if (Status != XST_SUCCESS) {
		xil_printf("ERROR: Failed create BD ring in txsetup\r\n");
		return XST_FAILURE;
	}

	/*
	 * We create an all-zero BD as the template.
	 */
	XAxiDma_BdClear(&BdTemplate);

	Status = XAxiDma_BdRingClone(TxRingPtr, &BdTemplate);
	if (Status != XST_SUCCESS) {
		xil_printf("ERROR: Failed bdring clone in txsetup %d\r\n", Status);
		return XST_FAILURE;
	}

	/* Start the TX channel */
	Status = XAxiDma_BdRingStart(TxRingPtr);
	if (Status != XST_SUCCESS) {
		xil_printf("ERROR: Failed start bdring txsetup %d\r\n", Status);
		return XST_FAILURE;
	}

	return XST_SUCCESS;
}

[hbucher]

@gdg

The best reference I could find is in the header files. They have a huge amount of documentation and usually is more than sufficient to understand. 

Look inside XAxiDma.h, XAxiDma_Bd.h and XAxiDma_BdRing.h

[gdg]

Ok, so I will go the hard way :-)

Thank you

[hbucher]

@gdg

This guy Mohammed Sadri actually has a very good set of tutorials on youtube

ZYNQ Training - Lesson 10 Part I - Using AXI DMA In Scatter-Gather Mode

https://www.youtube.com/watch?v=rDGOVszjeKs

This friend of mine, Leonardo, also has some good tutorials (his accent is just adorable)

VIVADO HLS Training AXI Stream interface #07

https://www.youtube.com/watch?v=3So1DPe2_4s

[gdg]

Those are good tutorial indeed. I followed many of their tutorials.

Thank you

[gilles007]

@gdg 

Hello I am trying to understand the basics of image processing withing the PL of FPGA.

Can you explain or enlighten me on how your IP-core looks like ?

Is your inputs and outputs to the IP-core with AXI4Streams ? And internally processed with hls::axiVIDEO2Mat(src, img) or a different way ?

Thank you

Gilles