SDIO and SD fat access example?

[natblist]

Posted on February 09, 2012 at 01:06

Hi Chaps,

Going slowly mad trying to get Chan Fat working with the sdio libs on a stm32f4.I'm using the most recent (1.0.0) library for the F4, and the low level access demo within that library builds and (at least appears to ) work OK. I've spent many, many hours attempting to patch in 0.9 version of fatfs from chan (thks chan, you're a hero) - but I can't get it to work.If possible, I'd really really appreciate a demo project/source (or direction to - though I've scoured the web and can't find anything that works!)I'll post separately about the current problems I'm having with the my current build, but a working example would sort it.Many thanks,nat. #hse-sdio-stm32 #stm32-fat-chanfat-fatfs-sdio #sdcard-stm32f4-sdio-fatfs #sdcard-stm32f4-sdio-fatfs

[Tesla DeLorean]

Posted on March 27, 2013 at 16:46

Different 2GB MicroSD

50 KB block read/write, non contiguous regions, with SD_ReadMultiBlocks 51200 Bytes, 471825 Cycles 13.021777 MBps Read                                                             51200 Bytes, 31099889 Cycles 0.197557 MBps Write                                                             51200 Bytes, 476647 Cycles 12.890042 MBps Read                                                             51200 Bytes, 3929347 Cycles 1.563619 MBps Write                                                             51200 Bytes, 474559 Cycles 12.946757 MBps Read                                                             51200 Bytes, 3866183 Cycles 1.589164 MBps Write                                                             51200 Bytes, 474571 Cycles 12.946430 MBps Read                                                             51200 Bytes, 3845599 Cycles 1.597670 MBps Write                                                             51200 Bytes, 474535 Cycles                                                     12.947412 MBps Read                                                             51200 Bytes, 3876209 Cycles 1.585054 MBps Write                                                             Overwriting on the same spot gets folded by the card, claims 5.6-5.8 MBps   20 MB linear write in 4KB blocks with FATFS 20480000 Bytes, 1310214451 Cycles 1.875723 MBps Write (FatFs)                                                     20480000 Bytes, 1299831750 Cycles 1.890706 MBps Write (FatFs) 20480000 Bytes, 1279462454 Cycles 1.920807 MBps Write (FatFs)                                                     20480000 Bytes, 1300349142 Cycles 1.889954 MBps Write (FatFs) 500 KB linear write in 512 byte blocks with FATFS (Same card spectacular results) 512000 Bytes, 1386510924 Cycles 0.044313 MBps Write (FatFs)                                                     512000 Bytes, 1356673081 Cycles 0.045287 MBps Write (FatFs)                                                     512000 Bytes, 1344495713 Cycles 0.045697 MBps Write (FatFs)                                                     512000 Bytes, 1374083889 Cycles 0.044713 MBps Write (FatFs)                                                     512000 Bytes, 1346310799 Cycles 0.045636 MBps Write (FatFs) Cards vary wildly in their internal caching and buffering strategy. FATFS has sub cluster issues.

[Tesla DeLorean]

Posted on March 27, 2013 at 17:04

8GB SDXC (Class 10, Ultra 1)

50 KB block read/write, non contiguous regions, with SD_ReadMultiBlocks & WriteMultiBlocks 51200 Bytes, 562059 Cycles 10.931237 MBps Read 51200 Bytes, 752601 Cycles 8.163688 MBps Write 51200 Bytes, 444775 Cycles 13.813726 MBps Read 51200 Bytes, 701289 Cycles 8.761010 MBps Write 51200 Bytes, 444803 Cycles 13.812856 MBps Read 51200 Bytes, 702461 Cycles 8.746393 MBps Write 51200 Bytes, 445019 Cycles 13.806152 MBps Read 51200 Bytes, 704563 Cycles 8.720299 MBps Write 51200 Bytes, 445047 Cycles 13.805283 MBps Read 51200 Bytes, 3691509 Cycles 1.664360 MBps Write 20 MB linear write in 4KB blocks with FATFS (These seem a tad high, will need to do some more analysis) 20480000 Bytes, 65363911 Cycles 37.598729 MBps Write (FatFs) 20480000 Bytes, 58578512 Cycles 41.953951 MBps Write (FatFs) 20480000 Bytes, 78068204 Cycles 31.480166 MBps Write (FatFs) 20480000 Bytes, 61597732 Cycles 39.897573 MBps Write (FatFs) 20480000 Bytes, 74588402 Cycles 32.948822 MBps Write (FatFs) 500 KB linear write in 512 byte blocks with FATFS 512000 Bytes, 150580273 Cycles 0.408022 MBps Write (FatFs) 512000 Bytes, 153377418 Cycles 0.400580 MBps Write (FatFs) 512000 Bytes, 152458762 Cycles 0.402994 MBps Write (FatFs) 512000 Bytes, 153580688 Cycles 0.400050 MBps Write (FatFs) 512000 Bytes, 152551330 Cycles 0.402750 MBps Write (FatFs)

[gmgunderground]

Posted on March 27, 2013 at 21:41

Hi, can you post a link to your test project? Is very interesting transfer rate (MB == MegaByte ??).

I have made a test to write linear blocks in DMA without FileSystem and in 1s I can only write 160/180 KBytes. With a scope and one GPIO I have measured the time of SD_WaitWriteOperation

1.
GPIO_WriteBit(GPIOG, GPIO_Pin_1, Bit_SET);
2.
Status = SD_WaitWriteOperation(); 
// Check if the Transfer is finished
3.
GPIO_WriteBit(GPIOG, GPIO_Pin_1, Bit_RESET);

and I have find thath this is the bottle neck, 1.25ms, too much I think. I have verify the SDIO clock is at 24MHz, card is 4GB Class 10 Ultra (30MB/S in reading from manufacure info), is inpossible my evaluation board can't write more then 130K ! Some people on the forum say that they write at 8/9 MByte per second. Any idea?

[Tesla DeLorean]

Posted on March 27, 2013 at 22:42

After revisiting the 8GB SDHC/XC card and fixing some DMA Completion and Interrupt issues.

Raw Read/Write 51200 Bytes, 597976 Cycles 10.274660 MBps Read 51200 Bytes, 1605269 Cycles 3.827396 MBps Write 51200 Bytes, 566017 Cycles 10.854798 MBps Read 51200 Bytes, 1526177 Cycles 4.025745 MBps Write 51200 Bytes, 568901 Cycles 10.799770 MBps Read 51200 Bytes, 1528307 Cycles 4.020135 MBps Write 51200 Bytes, 566817 Cycles 10.839477 MBps Read 51200 Bytes, 1912465 Cycles 3.212608 MBps Write 51200 Bytes, 685941 Cycles 8.957039 MBps Read 51200 Bytes, 5843669 Cycles 1.051394 MBps Write 10MB at 4K w/FATFS CRC32 213CAAC7 Memory Image 10240000 Bytes, 412131978 Cycles 2.981569 MBps Write (FatFs) CRC32 213CAAC7 COUNTER.TXT 10240000 Bytes, 413281981 Cycles 2.973273 MBps Write (FatFs) CRC32 213CAAC7 COUNTER.TXT 500KB at 512 w/FATFS CRC32 B88E65CC Memory Image 512000 Bytes, 164962825 Cycles 0.372448 MBps Write (FatFs) CRC32 B88E65CC COUNTER.TXT 512000 Bytes, 165551795 Cycles 0.371123 MBps Write (FatFs) CRC32 B88E65CC COUNTER.TXT I added some code to actually validate the data written, so READ and WRITE paths are definitely working. I originally ported the STM3240G-EVAL code to the STM32F4-Discovery board. In terms of speed you might want to check the clocks, especially the SDIO clock, and the HSE which is 8 MHz on the Discovery vs 25 MHz on the EVAL. These benchmarks were done of an STM32F2 @ 120 MHz with a 48 MHz SDIO clock source, externally it's probably 24 MHz. The ''Cycles'' here are ticks of the 120 MHz clock. (1MBps represents 1,000,000 bytes in one second)

[gmgunderground]

Posted on March 27, 2013 at 22:50

Can you attach your code to the post ?

best regards

[Tesla DeLorean]

Posted on March 27, 2013 at 22:55

Can you attach your code to the post ?

Sorry, a bunch of proprietary stuff in there.

[gmgunderground]

Posted on March 27, 2013 at 23:07

I think if the clock on CLK line is 24MHz misured with the scope all is OK, the problem I think is the DMA performance, is right ?

[Tesla DeLorean]

Posted on March 28, 2013 at 01:06

I think DMA should haul data pretty effectively, you might want to review the speed of the AHB/APB buses over which it's transacting. On the SDIO controller side there is a pretty deep FIFO (128 bytes?)

The card initially comes up at 1-bit / 400 KHz, then should switch to 4-bit / 24 MHz. You could check the divisors there. Clocking the SDIO beyond 48 MHz is also possible, gets you some more bandwidth, but USB won't work. Wouldn't recommend, but overclocking does speed the cards up. Doing transactions of 512 will always kill you. The cards and the setup efficiencies work better with larger blocks, 4KB to 32KB being sweeter, but ideally at least a cluster, better still if it matches the block/boundaries of the underlying flash devices. The benchmark for your 512 byte test don't surprise me, try doing at least 4KB writes. Having a simple read cache, with write through, I think would make quite a difference, although I don't have time to invest in one. I have a better FAT implementation, which caches structures internally, but it's a bit of a hog for the STM32. The speed limitation for write tends to be one the cards internal controller dictates.

[gmgunderground]

Posted on March 28, 2013 at 19:21

After setting correctly the clock of APB2 and fixed some initialization issues, now I can write MicroSD at 7.2MB/s with FATFS!!! Incredibly fast! Next step is to try in bypass mode... someone has experience ?

[Bill Lewis]

Posted on March 30, 2013 at 21:09

In the FATFS-DBG example code, stm32f4_discovery.c

Function SD_LowLevel_DMA_TxConfig() This line:

 SDDMA_InitStructure.DMA_BufferSize = 0;

causes an assert to be thrown when calling:

DMA_Init(SD_SDIO_DMA_STREAM, &SDDMA_InitStructure);

-Bill