Quote from: drazgor on June 07, 2013, 09:23:36 AM

Don't have that card so I can't test if it helps, but you could provide a patch to make it even easier to implement.

I'd be glad to, however it's not worth implementing in it's current form; I removed some of the checks that might need to be there for other card series!

Code snippet of the diff:

Code:

--- phatk121016.cl 2013-06-07 09:38:40.000000000 +0100
+++ phatk121016-modified.cl 2013-06-07 10:41:05.000000000 +0100
@@ -57,27 +57,12 @@
// SHA-256 Ch function, but provides it in exactly one instruction. If
// detected, use it for Ch. Otherwise, construct Ch out of simpler logical
// primitives.
-
- #ifdef BFI_INT
- // Well, slight problem... It turns out BFI_INT isn't actually exposed to
- // OpenCL (or CAL IL for that matter) in any way. However, there is
- // a similar instruction, BYTE_ALIGN_INT, which is exposed to OpenCL via
- // amd_bytealign, takes the same inputs, and provides the same output.
- // We can use that as a placeholder for BFI_INT and have the application
- // patch it after compilation.

- // This is the BFI_INT function
- #define Ch(x, y, z) amd_bytealign(x,y,z)
- // Ma can also be implemented in terms of BFI_INT...
- #define Ma(z, x, y) amd_bytealign(z^x,y,x)
- #else // BFI_INT
- // Later SDKs optimise this to BFI INT without patching and GCN
- // actually fails if manually patched with BFI_INT
-
- #define Ch(x, y, z) bitselect((u)z, (u)y, (u)x)
+ //We have an SDK which automatically optimizes to BFI INT, so lets do this
+ #define Ch(x, y, z) bitselect(z, y, x)
#define Ma(x, y, z) bitselect((u)x, (u)y, (u)z ^ (u)x)
#define rotr(x, y) amd_bitalign((u)x, (u)x, (u)y)
- #endif
+
#else // BITALIGN
#define Ch(x, y, z) (z ^ (x & (y ^ z)))
#define Ma(x, y, z) ((x & z) | (y & (x | z)))

...and a pastebin for those that prefer it:

http://pastebin.com/tZ7DeV2c