PROCESSOR-SDK-J721E: TDA4X UDMA transfer errors on C66x

hi Vipul,

Vipul Kulkarni said:

1) Is there any limitation of size in UDMA transfer?

No, 512KB is no limitation in the size of the UDMA transfer. In fact, there are drivers which use UDMA to transfer MBs of data. So there should be something wrong in the way the UDMA is getting configured. Can you share the details of parameters? 

Vipul Kulkarni said:

2) Is there any available document, which gives the explanation about DMA transfer, size limitations etc.

Please refer to the TRM for this. 

Vipul Kulkarni said:

3) If there is a limitation of size then what you would suggest to do a memcpy of buffers with large size ( greater that size limitation )

Again, what is the size of the buffer that you want to do memory copy?

Regards,

Brijesh

Hello Birjesh,

Thanks for the reply.

Actually I want do a memcpy of 1280*944 YUYV format  image. For doing the same I am using ND UDMA copy. Following is the configuration for the UDMA. In the below configuration I am transferring 1280*2*512 ( just for checking the size limit )

static void memcpyC66(uint8_t *pOut, uint8_t *pIn, uint8_t *tmp_buf, int32_t width, int32_t height)
{
app_udma_copy_nd_prms_t prms_nd_in;
app_udma_copy_nd_prms_t prms_nd_out;
app_udma_ch_handle_t udmaChIn;
app_udma_ch_handle_t udmaChOut;

appUdmaCopyNDPrms_Init(&prms_nd_in);

appUdmaCopyNDPrms_Init(&prms_nd_out);

prms_nd_in.copy_mode = 2;
prms_nd_in.src_addr = appMemGetVirt2PhyBufPtr((uint64_t) pIn, APP_MEM_HEAP_DDR);
prms_nd_in.dest_addr = appMemGetVirt2PhyBufPtr((uint64_t) tmp_buf, APP_MEM_HEAP_DDR);
prms_nd_in.icnt0 = width;
prms_nd_in.icnt1 = height;
prms_nd_in.icnt2 = 32;
prms_nd_in.icnt3 = 1;
prms_nd_in.dim1 = width*2;
prms_nd_in.dim2 = (height * width*2);
prms_nd_in.dim3 = (height * width*2 * 32);

prms_nd_in.dicnt0 = prms_nd_in.icnt0;
prms_nd_in.dicnt1 = prms_nd_in.icnt1;
prms_nd_in.dicnt2 = 32; /* Ping-pong */
prms_nd_in.dicnt3 = 1;
prms_nd_in.ddim1 = width*2;
prms_nd_in.ddim2 = (height * width*2);
prms_nd_in.ddim3 = 0;

prms_nd_out.copy_mode = 2;
prms_nd_out.src_addr = appMemGetVirt2PhyBufPtr((uint64_t) tmp_buf, APP_MEM_HEAP_DDR);
prms_nd_out.dest_addr = appMemGetVirt2PhyBufPtr((uint64_t) pOut, APP_MEM_HEAP_DDR);
prms_nd_out.icnt0 = width;
prms_nd_out.icnt1 = height;
prms_nd_out.icnt2 = 32; /* Ping-pong */
prms_nd_out.icnt3 = 1;
prms_nd_out.dim1 = width*2;
prms_nd_out.dim2 = (height * width*2);
prms_nd_out.dim3 = 0;

prms_nd_out.dicnt0 = prms_nd_out.icnt0;
prms_nd_out.dicnt1 = prms_nd_out.icnt1;
prms_nd_out.dicnt2 = 32;
prms_nd_out.dicnt3 = 1;
prms_nd_out.ddim1 = width*2;
prms_nd_out.ddim2 = (height * width*2);
prms_nd_out.ddim3 = (height * width*2 * 32);

udmaChIn = appUdmaCopyNDGetHandle(0);
udmaChOut = appUdmaCopyNDGetHandle(1);

appUdmaCopyNDInit(udmaChIn, &prms_nd_in);
appUdmaCopyNDInit(udmaChOut, &prms_nd_out);

int i;
for(i=0; i<32; i++)
{
appUdmaCopyNDTrigger(udmaChIn);
appUdmaCopyNDWait(udmaChIn);
appUdmaCopyNDTrigger(udmaChOut);
appUdmaCopyNDWait(udmaChOut);
}
appUdmaCopyNDDeinit(udmaChIn);
appUdmaCopyNDDeinit(udmaChOut);
}

As per this configuration, I am currently doing memcpy of 1280 * 512 * 2 ( just for checking usage ). in each transfer I am executing 1280*2*16 and doing for 32 iterations. So I have selected icnt2 as 32. Please take a look at this and let me know if anything is not correct.

Thanks,

Vipul Kulkarni.

Hi Vipul,

But why do you want to copy small size 32 times? Why can't you copy entire 1280*944*2 bytes in one shot using 2D copy function? 

Regards,

Brijesh

Hi Birjesh,

Actually, first I tried normal 2d copy of 1280*944*2 but it was giving memory fault core dumped error and it was dumping only 260 kb buffer. After that I tried Nd copy.

Can you tell what might be the reason of the memory fault with 2d copy.

Thanks,

Vipul Kulkarni

Hi Birjesh,

This is the configuration for 2D transfer:

appUdmaCopy2DPrms_Init(&prms_2d);
prms_2d.width = width; //1280
prms_2d.height = height; //944
prms_2d.dest_pitch = width*2; //1280*2
prms_2d.src_pitch = width*2; //1280*2
prms_2d.dest_addr = appMemGetVirt2PhyBufPtr((uint64_t) pOut, APP_MEM_HEAP_DDR);
prms_2d.src_addr = appMemGetVirt2PhyBufPtr((uint64_t) pIn, APP_MEM_HEAP_DDR);
appUdmaCopy2D(NULL, &prms_2d, 1);

With this config. when I was doing transfer for the images which has size less than 260 kb then error was not there but above that some meory fault error was there.

Can you please help regarding the issue.

Thanks

Hi Vipul,

I see you are setting pitch as width x 2 whereas width is same as width. When you are doing DMA transfer width should be in terms of bytes. 

I am not exactly if this is the reason for the memory fault, but its definitely not correct. 

Are your buffer address, pIn and pOut correct and well separated? can you check them?

Regards,

Brijesh 

Hi Birjesh,

Yes Birjesh. I have printed the addresses for input and output, addresses are in the range of ddr shared memory.

about pitch, yes you are right but for smaller resolution image is getting dumped correctly. I tried for 1280*2*64 for this resolution image is getting dumped correctly and no memory fault error also.

Thanks,

Vipul

Hi Birjesh, 

Can you please tell if any existing example which does 2d or ND dma tranafer from ddr to ddr.

Thank You.

Vipul

Hi Vipul,

Other than UDMA Test example application, i dont see these APIs being used anywhere in the SDK. Can you please refer to this example API appUdmaTest2DCopy in file ti-processor-sdk-rtos-j721e-evm-08_06_00_12\vision_apps\utils\udma\src\app_udma_test.c ?

Regards,

Brijesh

Hi Birjesh,

Yes, I refered the same example, but with that only some memory fault error is there.

Thanks,

Vipul

On which core are you seeing this error? Are the buffers getting allocated properly? Are you using the same API to test this feature? 

Regards,

Brijesh

Hi Birjesh,

Currently I am trying this on C66x core and bufferes also getting allocated properly. I have printed the address for the input and output ptr and the addresses are as below :

For input  =  0xAE000000 which is nothing but the starting address of ddr shared memory for C66x

For output = 0xAE140000 which is after 1280*512*2 size on c66x.

And the api which i used is appUdma2dCopy().

Thanks,

Vipul.

Hi Birjesh,

Actually By mistakenly I have marked it as resolved

but when I checked memory map, the starting address of OpenVX memory is  0XAC040000 and length is of 31.62 MB (0x01FA0000) and starting address of DDR memory is from AE000000

So according to this it should not collide with each other.

Thanks,

Vipul.

Hi Vipul,

It should be allocated from the local heap of C6x, which starts from 0xDC000000 for C6x_1. Looking at the code, it should have been allocated from this section, when you use appMemAlloc with APP_UDMA_HEAP_ID Heap Id, can you please check and confirm if you are using this? 

Regards,

Brijesh

Vipul Kulkarni said:

but when I checked memory map, the starting address of OpenVX memory is  0XAC040000 and length is of 31.62 MB (0x01FA0000) and starting address of DDR memory is from AE000000

ok, which SDK release are you using? Have rearranged memory map from the SDK release?

Hi Birjesh,

Thanks for quick reply.

Actually I have printed the address after calling the function appMemGetVirt2PhyBufPtr(pIn, APP_MEM_HEAP_DDR). but the address is same as previous i.e 0xAE000000. I think it should start from 0xDC000000 as you mentioned, because we are giving APP_MEM_HEAP_DDR.

Thanks,

Vipul.

Hi Birjesh,

SDK version is 7.1.we did some memory map changes previously.

Thanks,

Vipul

oh, that's very old release. is it possible to try it out on latest release? 

Even on SDK7.1, can you try using local heap on C6x, like below? 

Regards,

Brijesh

Hi Birjesh,

Yes we will try to allocate in the local heap once. Will let you know.

Thanks,

Vipul

Thanks, i will move this ticket to waiting state.

Hi Birjesh,

As you mentioned we have tried by allocating the memories in local heap region , but still the issue is same. we have printed the addresses and that also looks correct.

Is it because of any other reason this issues comes?

Thanks,

Vipul.

Hi Vipul,

Do you have JTAG + CCS connection on this board? Then, can you connect to C6x core and see where exactly it crashes? Also, can you dump TR memory and share it?

Regards,

Brijesh