diff options
| author | David Phillips <david@sighup.nz> | 2018-04-24 16:57:48 +1200 |
|---|---|---|
| committer | David Phillips <david@sighup.nz> | 2018-04-24 17:11:30 +1200 |
| commit | 0755368f42c8f233c67528e398352a9de2bb33bf (patch) | |
| tree | a7773108812640dafaa20ae7ad8ea5e084d9356c /trampoline.c | |
| parent | 2fec05364409258a28a77ede1552f1cdc3eabd24 (diff) | |
| download | sand-leek-0755368f42c8f233c67528e398352a9de2bb33bf.tar.xz | |
Improve host-side checks, fix MSB bug, misc tidying
Host-side check now uses from-scratch libssl SHA to confirm GPU work.
Bug when most significant bit of the exponent is set turns out to
be because of libssl's storage of such exponents requiring 5 bytes.
This cannot be dynamically adjusted without adjusting the pre-calculated
value which is hardcoded into the CL kernel for performance reasons
(INITIAL_DATA_LEN)
Diffstat (limited to 'trampoline.c')
| -rw-r--r-- | trampoline.c | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/trampoline.c b/trampoline.c index 24573f1..8dd9040 100644 --- a/trampoline.c +++ b/trampoline.c @@ -395,7 +395,7 @@ int tramp_run_kernel() cl_event event; cl_int ret = 0; size_t workgroup_sizes[2]; - workgroup_sizes[0] = 65536; + workgroup_sizes[0] = 32768; workgroup_sizes[1] = 1; ret = clEnqueueNDRangeKernel(command_queue, kernel, 2, NULL, workgroup_sizes, NULL, 0, NULL, &event); |