cpuminer, branch cell-be-optimizations Cell BE optimized miner for litecoin http://yyz.us/bitcoin/ https://git.amat.us/cpuminer/atom/?h=cell-be-optimizations 2013-03-28T03:50:04Z Force work_restart to be 128-byte aligned. 2013-03-28T03:50:04Z David Barksdale amatus.amongus@gmail.com 2013-03-28T03:50:04Z urn:sha1:cc439de3a7a7199d8aa5bb7e37c38e07a743f4de This fixes a bug on my system where mfc_get was failing on work_restart_pointer because it was not 128-byte aligned. Initial partial conversion to SPU assembly 2012-02-20T22:03:44Z Siarhei Siamashka siarhei.siamashka@gmail.com 2012-02-20T22:03:44Z urn:sha1:3c51c7107674e3fb62734d6fe80405db5484ee5e Only the first loop is converted. The performance is increased up to ~5.97 khash/sec per SPE core. TODO: try to update it to use better data layout and more unrolling from shakti's variant. CURLOPT_NOSIGNAL fix (borrowed from pooler's code) 2012-02-20T21:56:51Z Siarhei Siamashka siarhei.siamashka@gmail.com 2012-02-20T21:56:51Z urn:sha1:a62cd79b85c69aea5b980f91d2c563b082b06af0 Prevents the miner from eventually abnormally terminating due to curl misbehaviour. Fix strict aliasing problems 2012-01-01T02:04:54Z Siarhei Siamashka siarhei.siamashka@gmail.com 2012-01-01T02:04:00Z urn:sha1:b690870238b0b468355ec0d5b83c5b43b00f8aa6 Currently the cast between uint32_t and uint64_t pointers breaks strict aliasing rules and needs -fno-strict-aliasing option as a workaround, otherwise the code gets miscompiled. But -fno-strict-aliasing can seriously inhibit optimization possibilities. For example, performance of 1 thread on Cell PPU (using Altivec instructions): CFLAGS="-O3 -mcpu=cell -fno-strict-aliasing" - 1.79 khash/sec CFLAGS="-O3 -mcpu=cell -fstrict-aliasing" - 2.60 khash/sec Use the SHA256 code tweaked by pooler 2012-01-01T01:27:09Z Siarhei Siamashka siarhei.siamashka@gmail.com 2011-12-31T23:47:25Z urn:sha1:6d1e5bf3c2300623153479fd43c08a5874dacbbc Because the data endiannes has changed (native instead of little endian) and also SHA256 functions arguments are now different, this required lots of changes all over the place. Improves Altivec performance on Cell PPU from ~3.4 khash/s to ~3.6 khash/s (two threads). Seems to have no effect on SPU performance though. Set hashes_done variable to 0 in order not to confuse valgrind 2011-12-31T17:53:05Z Siarhei Siamashka siarhei.siamashka@gmail.com 2011-12-31T17:53:05Z urn:sha1:4280aca375f60981d70826091f1ece82e7cf56c5 This variable is set by SPU code, but valgrind can't see it and complains. Added Cell/BE optimizations 2011-12-29T04:47:40Z Siarhei Siamashka siarhei.siamashka@gmail.com 2011-12-27T01:51:24Z urn:sha1:ea4520474ba5a966260fc5652c5248e6eab50d58 Linux on PS3 gets a huge boost in litecoin mining performance. Cell/BE support should be detected and enabled automatically by autotools. The miner threads are first allocated to the available SPU cores (typically 6). The remaining threads are allocated on PPU. There will be 8 threads total on PS3: 6 SPU threads and 2 PPU threads. Each SPU core provides ~5.4 khash/s if compiled with spu-elf-gcc 4.6 The performance may vary for different gcc versions, older ones are typically slower. Use gcc vector extensions for SIMD scrypt key derivation function 2011-12-29T04:46:17Z Siarhei Siamashka siarhei.siamashka@gmail.com 2011-12-28T23:38:05Z urn:sha1:7a9481d266b58bd02a7d8033f359ca344cfd65d6 The code can be compiled for different architectures from the same source starting with gcc 4.7. But SSE2/Altivec/SPU targets have compatibility wrappers, which also allow the use of older versions of gcc. Two hashes are processed at the same time, so twice bigger scratch buffer is needed (~256K vs. ~128K). Speedup on Cell PPU (32-bit), single thread, 3.2GHz: ~0.58 khash/s -> ~1.79 khash/sec SHA256 code moved to a separate header file 2011-12-29T04:29:30Z Siarhei Siamashka siarhei.siamashka@gmail.com 2011-12-28T21:51:06Z urn:sha1:a0139edc7e2c3aa8252405ceff8a23a065965716 Fix big endian bugs 2011-12-22T19:35:19Z Siarhei Siamashka siarhei.siamashka@gmail.com 2011-12-22T19:31:21Z urn:sha1:285cc8d4d023ea858116b9e985667b44ef7e69cc Now it seems to work correctly and provide performance ~0.58 khash/s per thread on Cell PPU.