Batched ec additions msm #615
Conversation
…utilization logging with ifdef to allow some ability of benchmarking
Signed-off-by: Koren-Brand <koren@ingonyama.com>
Signed-off-by: Koren-Brand <koren@ingonyama.com>
| @@ -9,6 +9,8 @@ | |||
| #define MANAGER_SLEEP_USEC 10 | |||
| #define THREAD_SLEEP_USEC 1 | |||
|
|
|||
| // #define LOG_UTILIZATION | |||
There was a problem hiding this comment.
do you plan to keep it here?
| @@ -14,10 +14,16 @@ | |||
| #include "icicle/msm.h" | |||
| #include "tasks_manager.h" | |||
| #include "icicle/backend/msm_config.h" | |||
| #ifdef LOG_UTILIZATION | |||
There was a problem hiding this comment.
Do you want the code to have this flow for logging? it's kind of ugly in the code
There was a problem hiding this comment.
I can include timer either way if it's important
|
|
||
| using namespace icicle; | ||
| using namespace curve_config; | ||
|
|
||
| #define LOG_EC_BATCH_SIZE 2 |
There was a problem hiding this comment.
also the name is not that clear
There was a problem hiding this comment.
It was chosen as preparation for AVX but I can run a few checks to see what works better
| @@ -352,20 +421,23 @@ void Msm<A, P>::phase1_bucket_accumulator(const scalar_t* scalars, const A* base | |||
| // Handle required preprocess of base P | |||
| A base = | |||
| m_are_points_mont ? A::from_montgomery(bases[m_precompute_factor * i + j]) : bases[m_precompute_factor * i + j]; | |||
| // TODO move to preprocess before precompute to avoid repeating conversions | |||
There was a problem hiding this comment.
todo when? also for the one below
There was a problem hiding this comment.
Maybe not this week, I forgot to remove TODOs before the PR.
| @@ -6,7 +6,7 @@ | |||
| #include <cassert> | |||
| #include "timer.hpp" | |||
|
|
|||
| // #define DUMMY_TYPES | |||
| #define DUMMY_TYPES | |||
There was a problem hiding this comment.
didn't you want to remove this test? I don't know if you should but you wanted I think
icicle/tests/test_curve_api.cpp
Outdated
| const int N = 1 << logn; | ||
| const int precompute_factor = 2; | ||
| const int precompute_factor = 8; |
| (int)std::floor(std::log2((double)(nof_threads * TASKS_PER_THREAD - 1) / (double)(2 * m_num_bms))), 0)), | ||
| m_num_bm_segments(std::min((int)(1 << m_log_num_segments), (int)(m_num_bms * m_num_bkts))), | ||
| (int)std::floor( | ||
| std::log2((double)(nof_threads * TASKS_PER_THREAD * EC_BATCH_SIZE - 1) / (double)(2 * m_num_bms))), |
There was a problem hiding this comment.
do you really need the double casting?
There was a problem hiding this comment.
I can remove one of the (double)s if that is what you mean.
Signed-off-by: Koren-Brand <koren@ingonyama.com>
Describe the changes
Batched EC addition was added for MSM, improving performance. Also previous bug where c divides scalar bit width is fixed.
Linked Issues
Resolves #