Hi Kim,

On Fri, 1 Mar 2013 11:16:22 -0500 Akshay Saraswat akshay.s@samsung.com wrote:

...

SHA-256 and SHA-1 accelerated using ACE hardware.

curious about the rationale: how much faster is this than software?

Time taken by s/w and h/w sha256 are as given below: Size: 0x1000 s/w - 11 ticks h/w - 10 ticks

Size: 0x2000 s/w - 11 ticks h/w - 10 ticks

Size: 0x4000 s/w - 13 ticks h/w - 10 ticks

Size: 0x8000 s/w - 16 ticks h/w - 10 ticks

Size: 0x800000 s/w - 1775 ticks h/w - 45 ticks

Size: 0x400000 s/w - 893 ticks h/w - 27 ticks

Size: 0x200000 s/w - 451 ticks h/w - 18 ticks

Size: 0x8000000 s/w - 28270 ticks h/w - 110 ticks

Size: 0x40000000 s/w - 225896 ticks h/w - 110 ticks

...

---

Changes since v2:

• Added falling back to software sha256 in case length exceeds buffer limit.
• Reduced one tab at lines 533, 559 and 571 in this patch.
• Removed space after a cast at line 506 in this patch.
• Removed blank line at line 561 in this patch.
• Removed space before semicolon at line 576 in this patch.

you should retain the list of changes since v1 here. Read:

http://www.denx.de/wiki/view/U-Boot/Patches#Sending_updated_patch_versions

...

+/* Hash status */ +#define ACE_HASH_BUFRDY_MASK (1 << 0) +#define ACE_HASH_BUFRDY_OFF (0 << 0) +#define ACE_HASH_BUFRDY_ON (1 << 0) +#define ACE_HASH_SEEDSETTING_MASK (1 << 1) +#define ACE_HASH_SEEDSETTING_OFF (0 << 1) +#define ACE_HASH_SEEDSETTING_ON (1 << 1) +#define ACE_HASH_PRNGBUSY_MASK (1 << 2) +#define ACE_HASH_PRNGBUSY_OFF (0 << 2) +#define ACE_HASH_PRNGBUSY_ON (1 << 2) +#define ACE_HASH_PARTIALDONE_MASK (1 << 4) +#define ACE_HASH_PARTIALDONE_OFF (0 << 4) +#define ACE_HASH_PARTIALDONE_ON (1 << 4)

based on the nomenclature of the above definitions, this hardware obviously has support for breaking up hash requests into multiple buffer submissions, so do that, and not this:

...

• } else if (buf_len > BUF_LIMIT) {

• /*
  

•  * ACE H/W cannot compute hash value for buffer > 8 MB.
  

•  * Falling back to software.
  

•  */
  

• if (hash_type == ACE_SHA_TYPE_SHA1)
  

• 	sha1_csum_wd(pbuf, buf_len, pout, CHUNKSZ_SHA1);
  

• else
  

• 	sha256_csum_wd(pbuf, buf_len, pout, CHUNKSZ_SHA256);
  

• return 0;
  

• }

Kim

Regards, Akshay Saraswat