ecmult_multi: reduce strauss memory usage by 30% #1761

pull jonasnick wants to merge 1 commits into bitcoin-core:master from jonasnick:strauss-mem changing 1 files +19 −4
  1. jonasnick commented at 2:28 pm on October 17, 2025: contributor
    This is a draft because I’m not sure about the cleanest way to implement it.
  2. ecmult_multi: reduce strauss memory usage by 30% 26166c4f5f
  3. real-or-random added the label performance on Oct 27, 2025
  4. real-or-random commented at 8:02 am on October 27, 2025: contributor

    This is a draft because I’m not sure about the cleanest way to implement it.

    The current approach looks clean. What other approaches do you have in mind?

  5. real-or-random added the label tweak/refactor on Oct 27, 2025
  6. in src/ecmult_impl.h:239 in 26166c4f5f 
    236+}
237+
238 struct secp256k1_strauss_point_state {
239-    int wnaf_na_1[129];
240-    int wnaf_na_lam[129];
241+    int8_t wnaf_na_1[129];
    real-or-random commented at 8:10 am on October 27, 2025: 
    0    int_least8_t wnaf_na_1[129];

    This is technically better to retain support for platforms that don’t offer int8_t. I don’t think there are any relevant platforms of that kind, but why not… 🤷

    edit: Now that I think about it again, a platform that does not offer int8_t will most likely one for which CHAR_BITS != 8. And then it’s neither relevant nor it’s likely to offer int32_t or int64_t which we need anyway. So I guess there’s no difference between int8_t and int_least8_t.

    hebasto commented at 4:58 pm on October 27, 2025:

    This is technically better to retain support for platforms that don’t offer int8_t.

    int8_t is already used in src/assumptions.h.

  7. hebasto commented at 4:47 pm on October 27, 2025: member
    Concept ACK.
  8. hebasto commented at 6:00 pm on October 27, 2025: member

    On my x86_64 system, this PR reduces the memory allocated on the scratch space from 2224 bytes to 1452 bytes per point.

    Please ping me once it’s undrafted.

  9. jonasnick commented at 6:52 pm on October 27, 2025: contributor

    The current approach requires a temporary array int wnaf_tmp[256]; to provide to secp256k1_ecmult_wnaf which looks unclean. The alternatives are

    1. copy almost all of secp256k1_ecmult_wnaf into secp256k1_ecmult_wnaf_small, or
    2. remove secp256k1_ecmult_wnaf_small and write an secp256k1_ecmult macro.

    Both options seem to be worse.

  10. hebasto commented at 10:00 pm on October 27, 2025: member

    The current approach requires a temporary array int wnaf_tmp[256]; to provide to secp256k1_ecmult_wnaf which looks unclean. The alternatives are

    1. copy almost all of secp256k1_ecmult_wnaf into secp256k1_ecmult_wnaf_small, or

    2. remove secp256k1_ecmult_wnaf_small and write an secp256k1_ecmult macro.

    Both options seem to be worse.

    I might suggest a third option: https://github.com/hebasto/secp256k1/commit/5c0d6eeeb2ec9dc8de92f59cf7647a30e6826dcb.

  11. in src/ecmult_impl.h:228 in 26166c4f5f 
    219@@ -220,9 +220,24 @@ static int secp256k1_ecmult_wnaf(int *wnaf, int len, const secp256k1_scalar *a,
220     return last_set_bit + 1;
221 }
222 
223+/* Same as secp256k1_ecmult_wnaf, but stores to int8_t array. Requires w <= 8. */
224+static int secp256k1_ecmult_wnaf_small(int8_t *wnaf, int len, const secp256k1_scalar *a, int w) {
225+    int wnaf_tmp[256];
226+    int ret, i;
227+
228+    VERIFY_CHECK(2 <= w && w <= 8);
    real-or-random commented at 9:37 am on November 7, 2025: 
    0    VERIFY_CHECK(2 <= w && w <= 7);
    jonasnick commented at 2:56 pm on November 12, 2025:

    I don’t see why w = 8 wouldn’t work. The documentation of wnaf says

    0 *  - each wnaf[i] is either 0, or an odd integer between -(1<<(w-1) - 1) and (1<<(w-1) - 1)

    So for w = 8, wnaf[i] is in [-127, 127] which fits in an int8_t.

    real-or-random commented at 8:18 pm on November 12, 2025:

    Sorry, yes, you’re right. I was getting confused. secp256k1_ecmult_wnaf itself needs w <= 31 (and not 32), if only because it performs a carry << w shift (for int carry) which is certainly UB if int is 32 bits. (In fact, if carry == 1, then even 1 << 31 is UB. This is another edge case that we should fix! Let me add this to the other issue.)

    But since your function only copies the results, everything is fine.

    jonasnick commented at 8:22 pm on November 12, 2025:

    In fact, if carry == 1, then even 1 « 31 is UB. This is another edge case that we should fix! Let me add this to the other issue

    Oh, great catch!

  12. real-or-random commented at 9:56 am on November 7, 2025: contributor

    I think the current approach in the PR is good. It may not be elegant to have a tmp array, but it’s simple and correct. We’d need to benchmark if the tmp array makes a difference in the end. But I think this PR needs a benchmark in general to make sure that using int8_t does not increase the running time (much).

    If we want to avoid it, here’s yet another variant: https://github.com/real-or-random/secp256k1/commit/f83731bce8880787351533c41d14ac334b38680c It uses a macro to define different variants of secp256k1_ecmult_wnaf parametrized in the output type. The macro is not elegant either, but this variant is better for type safety than just turning secp256k1_ecmult_wnaf into a macro.

    In fact, the current secp256k1_ecmult_wnaf needs the unstated and unchecked assumption that int has at least 32 value bits when it VERIFY_CHECKs that w <= 31. In practice, we call it only with WINDOW_A == 5 and WINDOW_G == ECMULT_WINDOW_SIZE where the latter is configurable in the range 2..24.

    A consequence of this “bug” is that the code fails on a 16-bit platform if you set ECMULT_WINDOW_SIZE > 16. I don’t think we need to support this, but code without unchecked assumptions is bad. So I suggest that we rewrite the function to use int32_t instead of int even if we don’t use my macro approach. Alternatively, we could add the assumption that INT_MAX >= INT32_MAX but this forbids 16-bit platforms, and the code seems to work on them in principle; see #792 (comment).

  13. real-or-random commented at 10:17 am on November 7, 2025: contributor

    I might suggest a third option: hebasto@5c0d6ee.

    Sorry, I forgot to comment on that option. That’s also clean, but it introduces a lot of code complexity.

    The way I see it:

    1. If the current approach is fine performance-wise, let’s take it.
    2. If not, use either @hebasto’s approach or mine. If you ask me, I prefer mine slightly because it’s less code and more “direct” even though it uses a macro.
    3. If none of this is satisfactory, we can still duplicate the code.
  14. jonasnick commented at 7:26 pm on November 12, 2025: contributor
    Thanks for demonstrating creative alternative solutions :) If it weren’t for the layers of indirection, I’d consider @hebasto’s approach to be the most elegant. The PR’s current approach is just so much simpler. And I ran benchmarks with bench_ecmult, which showed at most a 0.1us slowdown (for some number of points) on my Intel i7 machine.
  15. jonasnick marked this as ready for review on Nov 12, 2025
  16. real-or-random approved
  17. real-or-random commented at 8:23 pm on November 12, 2025: contributor
    utACK 26166c4f5fd5688c31e488449ee2325eb8f2fb36
  18. siv2r commented at 2:47 pm on November 17, 2025: contributor

    tACK 26166c4

    Points Master Avg (µs) PR Avg (µs) Improvement
    8,191 5.22 5.10 2.3%
    16,383 6.13 5.44 12.6%
    20,479 6.88 5.78 19.0%
    32,767 8.26 7.50 10.0% 
      0Using strauss_wnaf:
  1Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
  2
  3ecmult_gen                    ,     5.30      ,     5.48      ,     5.71   
  4ecmult_const                  ,    12.1       ,    12.3       ,    12.4    
  5ecmult_const_xonly            ,    13.6       ,    13.8       ,    13.9    
  6ecmult_1p                     ,     9.82      ,    10.0       ,    10.1    
  7ecmult_0p_g                   ,     6.82      ,     6.92      ,     7.00   
  8ecmult_1p_g                   ,     5.82      ,     5.89      ,     5.95   
  9ecmult_multi_0p_g             ,     6.84      ,     6.93      ,     6.99   
 10ecmult_multi_1p_g             ,     5.79      ,     5.90      ,     6.00   
 11ecmult_multi_2p_g             ,     5.50      ,     5.57      ,     5.65   
 12ecmult_multi_3p_g             ,     5.39      ,     5.45      ,     5.58   
 13ecmult_multi_4p_g             ,     5.21      ,     5.30      ,     5.37   
 14ecmult_multi_5p_g             ,     5.14      ,     5.25      ,     5.30   
 15ecmult_multi_6p_g             ,     5.10      ,     5.20      ,     5.27   
 16ecmult_multi_7p_g             ,     5.07      ,     5.15      ,     5.20   
 17ecmult_multi_8p_g             ,     5.03      ,     5.12      ,     5.17   
 18ecmult_multi_9p_g             ,     4.92      ,     5.09      ,     5.16   
 19ecmult_multi_10p_g            ,     5.01      ,     5.09      ,     5.14   
 20ecmult_multi_11p_g            ,     5.00      ,     5.08      ,     5.13   
 21ecmult_multi_12p_g            ,     4.98      ,     5.06      ,     5.12   
 22ecmult_multi_13p_g            ,     4.97      ,     5.05      ,     5.10   
 23ecmult_multi_14p_g            ,     4.93      ,     5.03      ,     5.13   
 24ecmult_multi_15p_g            ,     4.94      ,     5.03      ,     5.08   
 25ecmult_multi_17p_g            ,     4.94      ,     5.00      ,     5.06   
 26ecmult_multi_19p_g            ,     4.92      ,     4.99      ,     5.07   
 27ecmult_multi_21p_g            ,     4.89      ,     4.96      ,     5.05   
 28ecmult_multi_23p_g            ,     4.90      ,     4.98      ,     5.03   
 29ecmult_multi_25p_g            ,     4.88      ,     4.96      ,     5.03   
 30ecmult_multi_27p_g            ,     4.85      ,     4.95      ,     5.00   
 31ecmult_multi_29p_g            ,     4.90      ,     4.97      ,     5.07   
 32ecmult_multi_31p_g            ,     4.76      ,     4.94      ,     5.04   
 33ecmult_multi_35p_g            ,     4.84      ,     4.94      ,     5.09   
 34ecmult_multi_39p_g            ,     4.85      ,     4.92      ,     4.99   
 35ecmult_multi_43p_g            ,     4.79      ,     5.07      ,     6.24   
 36ecmult_multi_47p_g            ,     4.71      ,     4.88      ,     4.98   
 37ecmult_multi_51p_g            ,     4.68      ,     4.87      ,     5.00   
 38ecmult_multi_55p_g            ,     4.81      ,     4.90      ,     4.97   
 39ecmult_multi_59p_g            ,     4.80      ,     4.93      ,     5.00   
 40ecmult_multi_63p_g            ,     4.81      ,     4.92      ,     4.98   
 41ecmult_multi_71p_g            ,     4.80      ,     4.91      ,     4.99   
 42ecmult_multi_79p_g            ,     4.84      ,     4.93      ,     5.02   
 43ecmult_multi_87p_g            ,     4.76      ,     4.93      ,     5.02   
 44ecmult_multi_95p_g            ,     4.86      ,     4.96      ,     5.04   
 45ecmult_multi_103p_g           ,     4.88      ,     4.94      ,     5.02   
 46ecmult_multi_111p_g           ,     4.90      ,     4.96      ,     5.02   
 47ecmult_multi_119p_g           ,     4.88      ,     4.96      ,     5.01   
 48ecmult_multi_127p_g           ,     4.90      ,     4.98      ,     5.02   
 49ecmult_multi_143p_g           ,     4.94      ,     5.00      ,     5.05   
 50ecmult_multi_159p_g           ,     4.88      ,     4.99      ,     5.05   
 51ecmult_multi_175p_g           ,     4.88      ,     4.98      ,     5.05   
 52ecmult_multi_191p_g           ,     4.92      ,     5.01      ,     5.06   
 53ecmult_multi_207p_g           ,     4.94      ,     5.03      ,     5.09   
 54ecmult_multi_223p_g           ,     4.92      ,     4.99      ,     5.08   
 55ecmult_multi_239p_g           ,     4.98      ,     5.04      ,     5.10   
 56ecmult_multi_255p_g           ,     4.84      ,     5.01      ,     5.10   
 57ecmult_multi_287p_g           ,     4.98      ,     5.05      ,     5.11   
 58ecmult_multi_319p_g           ,     4.96      ,     5.06      ,     5.14   
 59ecmult_multi_351p_g           ,     5.00      ,     5.06      ,     5.15   
 60ecmult_multi_383p_g           ,     4.99      ,     5.06      ,     5.13   
 61ecmult_multi_415p_g           ,     4.95      ,     5.06      ,     5.14   
 62ecmult_multi_447p_g           ,     4.96      ,     5.07      ,     5.16   
 63ecmult_multi_479p_g           ,     5.01      ,     5.09      ,     5.15   
 64ecmult_multi_511p_g           ,     5.00      ,     5.07      ,     5.13   
 65ecmult_multi_575p_g           ,     4.94      ,     5.05      ,     5.13   
 66ecmult_multi_639p_g           ,     4.99      ,     5.07      ,     5.13   
 67ecmult_multi_703p_g           ,     4.99      ,     5.07      ,     5.16   
 68ecmult_multi_767p_g           ,     4.96      ,     5.07      ,     5.15   
 69ecmult_multi_831p_g           ,     5.00      ,     5.07      ,     5.14   
 70ecmult_multi_895p_g           ,     5.00      ,     5.08      ,     5.15   
 71ecmult_multi_959p_g           ,     4.98      ,     5.08      ,     5.14   
 72ecmult_multi_1023p_g          ,     5.00      ,     5.09      ,     5.16   
 73ecmult_multi_1151p_g          ,     5.01      ,     5.09      ,     5.17   
 74ecmult_multi_1279p_g          ,     5.04      ,     5.10      ,     5.16   
 75ecmult_multi_1407p_g          ,     4.92      ,     5.09      ,     5.15   
 76ecmult_multi_1535p_g          ,     5.00      ,     5.11      ,     5.16   
 77ecmult_multi_1663p_g          ,     5.03      ,     5.11      ,     5.19   
 78ecmult_multi_1791p_g          ,     5.04      ,     5.13      ,     5.18   
 79ecmult_multi_1919p_g          ,     5.02      ,     5.11      ,     5.20   
 80ecmult_multi_2047p_g          ,     4.99      ,     5.10      ,     5.18   
 81ecmult_multi_2303p_g          ,     4.97      ,     5.11      ,     5.20   
 82ecmult_multi_2559p_g          ,     5.02      ,     5.11      ,     5.19   
 83ecmult_multi_2815p_g          ,     5.04      ,     5.12      ,     5.18   
 84ecmult_multi_3071p_g          ,     5.01      ,     5.12      ,     5.24   
 85ecmult_multi_3327p_g          ,     5.05      ,     5.12      ,     5.16   
 86ecmult_multi_3583p_g          ,     5.03      ,     5.10      ,     5.18   
 87ecmult_multi_3839p_g          ,     5.07      ,     5.13      ,     5.23   
 88ecmult_multi_4095p_g          ,     5.02      ,     5.09      ,     5.19   
 89ecmult_multi_4607p_g          ,     5.04      ,     5.10      ,     5.15   
 90ecmult_multi_5119p_g          ,     5.01      ,     5.10      ,     5.18   
 91ecmult_multi_5631p_g          ,     5.02      ,     5.15      ,     5.29   
 92ecmult_multi_6143p_g          ,     5.06      ,     5.13      ,     5.19   
 93ecmult_multi_6655p_g          ,     5.06      ,     5.16      ,     5.22   
 94ecmult_multi_7167p_g          ,     5.10      ,     5.21      ,     5.40   
 95ecmult_multi_7679p_g          ,     5.15      ,     5.19      ,     5.25   
 96ecmult_multi_8191p_g          ,     5.16      ,     5.22      ,     5.29   
 97ecmult_multi_9215p_g          ,     5.30      ,     5.34      ,     5.41   
 98ecmult_multi_10239p_g         ,     5.34      ,     5.44      ,     5.54   
 99ecmult_multi_11263p_g         ,     5.46      ,     5.55      ,     5.66   
100ecmult_multi_12287p_g         ,     5.45      ,     5.64      ,     6.08   
101ecmult_multi_13311p_g         ,     5.66      ,     5.81      ,     6.44   
102ecmult_multi_14335p_g         ,     5.77      ,     5.90      ,     6.32   
103ecmult_multi_15359p_g         ,     5.89      ,     5.95      ,     6.08   
104ecmult_multi_16383p_g         ,     6.06      ,     6.13      ,     6.25   
105ecmult_multi_18431p_g         ,     6.44      ,     6.52      ,     6.82   
106ecmult_multi_20479p_g         ,     6.82      ,     6.88      ,     7.07   
107ecmult_multi_22527p_g         ,     7.09      ,     7.13      ,     7.18   
108ecmult_multi_24575p_g         ,     7.48      ,     7.57      ,     7.70   
109ecmult_multi_26623p_g         ,     7.74      ,     7.82      ,     8.02   
110ecmult_multi_28671p_g         ,     7.90      ,     7.95      ,     8.09   
111ecmult_multi_30719p_g         ,     8.05      ,     8.12      ,     8.27   
112ecmult_multi_32767p_g         ,     8.21      ,     8.26      ,     8.41   

      0Using strauss_wnaf:
  1Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)    
  2
  3ecmult_gen                    ,     5.24      ,     5.45      ,     5.82   
  4ecmult_const                  ,    12.0       ,    12.1       ,    12.3    
  5ecmult_const_xonly            ,    13.5       ,    13.7       ,    13.9    
  6ecmult_1p                     ,     9.81      ,     9.98      ,    10.1    
  7ecmult_0p_g                   ,     6.71      ,     6.84      ,     6.98   
  8ecmult_1p_g                   ,     5.80      ,     5.86      ,     5.92   
  9ecmult_multi_0p_g             ,     6.81      ,     6.87      ,     6.92   
 10ecmult_multi_1p_g             ,     5.80      ,     5.88      ,     5.97   
 11ecmult_multi_2p_g             ,     5.39      ,     5.53      ,     5.59   
 12ecmult_multi_3p_g             ,     5.30      ,     5.39      ,     5.44   
 13ecmult_multi_4p_g             ,     5.21      ,     5.29      ,     5.37   
 14ecmult_multi_5p_g             ,     5.13      ,     5.23      ,     5.28   
 15ecmult_multi_6p_g             ,     5.07      ,     5.17      ,     5.23   
 16ecmult_multi_7p_g             ,     5.07      ,     5.15      ,     5.20   
 17ecmult_multi_8p_g             ,     5.04      ,     5.13      ,     5.19   
 18ecmult_multi_9p_g             ,     4.99      ,     5.11      ,     5.18   
 19ecmult_multi_10p_g            ,     4.99      ,     5.07      ,     5.12   
 20ecmult_multi_11p_g            ,     4.97      ,     5.06      ,     5.18   
 21ecmult_multi_12p_g            ,     4.94      ,     5.05      ,     5.10   
 22ecmult_multi_13p_g            ,     4.96      ,     5.05      ,     5.13   
 23ecmult_multi_14p_g            ,     4.98      ,     5.03      ,     5.11   
 24ecmult_multi_15p_g            ,     4.94      ,     5.02      ,     5.10   
 25ecmult_multi_17p_g            ,     4.91      ,     5.01      ,     5.08   
 26ecmult_multi_19p_g            ,     4.89      ,     5.00      ,     5.07   
 27ecmult_multi_21p_g            ,     4.90      ,     4.98      ,     5.04   
 28ecmult_multi_23p_g            ,     4.89      ,     4.97      ,     5.05   
 29ecmult_multi_25p_g            ,     4.84      ,     4.96      ,     5.06   
 30ecmult_multi_27p_g            ,     4.87      ,     4.94      ,     5.03   
 31ecmult_multi_29p_g            ,     4.83      ,     4.92      ,     4.98   
 32ecmult_multi_31p_g            ,     4.82      ,     4.92      ,     4.99   
 33ecmult_multi_35p_g            ,     4.82      ,     4.93      ,     5.01   
 34ecmult_multi_39p_g            ,     4.86      ,     4.92      ,     4.97   
 35ecmult_multi_43p_g            ,     4.82      ,     4.91      ,     4.99   
 36ecmult_multi_47p_g            ,     4.81      ,     4.91      ,     4.96   
 37ecmult_multi_51p_g            ,     4.84      ,     4.91      ,     4.99   
 38ecmult_multi_55p_g            ,     4.79      ,     4.90      ,     4.97   
 39ecmult_multi_59p_g            ,     4.85      ,     4.91      ,     4.96   
 40ecmult_multi_63p_g            ,     4.81      ,     4.89      ,     4.96   
 41ecmult_multi_71p_g            ,     4.79      ,     4.88      ,     4.97   
 42ecmult_multi_79p_g            ,     4.79      ,     4.90      ,     4.97   
 43ecmult_multi_87p_g            ,     4.81      ,     4.90      ,     4.97   
 44ecmult_multi_95p_g            ,     4.82      ,     4.90      ,     4.99   
 45ecmult_multi_103p_g           ,     4.86      ,     4.92      ,     4.99   
 46ecmult_multi_111p_g           ,     4.82      ,     4.92      ,     4.99   
 47ecmult_multi_119p_g           ,     4.84      ,     4.94      ,     5.00   
 48ecmult_multi_127p_g           ,     4.81      ,     4.92      ,     5.01   
 49ecmult_multi_143p_g           ,     4.86      ,     4.95      ,     5.01   
 50ecmult_multi_159p_g           ,     4.92      ,     4.97      ,     5.06   
 51ecmult_multi_175p_g           ,     4.84      ,     4.95      ,     5.01   
 52ecmult_multi_191p_g           ,     4.88      ,     4.96      ,     5.04   
 53ecmult_multi_207p_g           ,     4.91      ,     4.99      ,     5.04   
 54ecmult_multi_223p_g           ,     4.93      ,     4.99      ,     5.05   
 55ecmult_multi_239p_g           ,     4.88      ,     4.96      ,     5.02   
 56ecmult_multi_255p_g           ,     4.88      ,     4.97      ,     5.03   
 57ecmult_multi_287p_g           ,     4.91      ,     4.97      ,     5.04   
 58ecmult_multi_319p_g           ,     4.92      ,     4.97      ,     5.04   
 59ecmult_multi_351p_g           ,     4.92      ,     4.98      ,     5.04   
 60ecmult_multi_383p_g           ,     4.88      ,     4.98      ,     5.04   
 61ecmult_multi_415p_g           ,     4.93      ,     4.99      ,     5.07   
 62ecmult_multi_447p_g           ,     4.81      ,     4.98      ,     5.05   
 63ecmult_multi_479p_g           ,     4.90      ,     4.98      ,     5.05   
 64ecmult_multi_511p_g           ,     4.92      ,     5.02      ,     5.07   
 65ecmult_multi_575p_g           ,     4.95      ,     5.01      ,     5.06   
 66ecmult_multi_639p_g           ,     4.94      ,     5.02      ,     5.08   
 67ecmult_multi_703p_g           ,     4.96      ,     5.04      ,     5.10   
 68ecmult_multi_767p_g           ,     4.92      ,     5.03      ,     5.09   
 69ecmult_multi_831p_g           ,     4.93      ,     5.06      ,     5.11   
 70ecmult_multi_895p_g           ,     4.90      ,     5.02      ,     5.11   
 71ecmult_multi_959p_g           ,     4.94      ,     5.02      ,     5.11   
 72ecmult_multi_1023p_g          ,     4.95      ,     5.04      ,     5.12   
 73ecmult_multi_1151p_g          ,     4.96      ,     5.05      ,     5.11   
 74ecmult_multi_1279p_g          ,     4.87      ,     5.06      ,     5.33   
 75ecmult_multi_1407p_g          ,     4.93      ,     5.04      ,     5.10   
 76ecmult_multi_1535p_g          ,     4.99      ,     5.04      ,     5.11   
 77ecmult_multi_1663p_g          ,     4.98      ,     5.04      ,     5.14   
 78ecmult_multi_1791p_g          ,     4.99      ,     5.05      ,     5.11   
 79ecmult_multi_1919p_g          ,     4.95      ,     5.05      ,     5.09   
 80ecmult_multi_2047p_g          ,     5.01      ,     5.03      ,     5.13   
 81ecmult_multi_2303p_g          ,     4.97      ,     5.05      ,     5.10   
 82ecmult_multi_2559p_g          ,     4.92      ,     5.04      ,     5.11   
 83ecmult_multi_2815p_g          ,     5.00      ,     5.05      ,     5.11   
 84ecmult_multi_3071p_g          ,     4.86      ,     5.00      ,     5.08   
 85ecmult_multi_3327p_g          ,     4.89      ,     5.02      ,     5.12   
 86ecmult_multi_3583p_g          ,     5.00      ,     5.06      ,     5.12   
 87ecmult_multi_3839p_g          ,     4.96      ,     5.04      ,     5.11   
 88ecmult_multi_4095p_g          ,     4.97      ,     5.04      ,     5.11   
 89ecmult_multi_4607p_g          ,     4.99      ,     5.07      ,     5.19   
 90ecmult_multi_5119p_g          ,     4.91      ,     5.06      ,     5.13   
 91ecmult_multi_5631p_g          ,     4.98      ,     5.07      ,     5.15   
 92ecmult_multi_6143p_g          ,     4.99      ,     5.07      ,     5.13   
 93ecmult_multi_6655p_g          ,     5.01      ,     5.12      ,     5.34   
 94ecmult_multi_7167p_g          ,     5.00      ,     5.06      ,     5.14   
 95ecmult_multi_7679p_g          ,     5.03      ,     5.07      ,     5.13   
 96ecmult_multi_8191p_g          ,     5.00      ,     5.10      ,     5.30   
 97ecmult_multi_9215p_g          ,     5.05      ,     5.09      ,     5.12   
 98ecmult_multi_10239p_g         ,     5.08      ,     5.19      ,     5.38   
 99ecmult_multi_11263p_g         ,     5.06      ,     5.15      ,     5.21   
100ecmult_multi_12287p_g         ,     5.08      ,     5.16      ,     5.22   
101ecmult_multi_13311p_g         ,     5.15      ,     5.21      ,     5.28   
102ecmult_multi_14335p_g         ,     5.21      ,     5.30      ,     5.61   
103ecmult_multi_15359p_g         ,     5.22      ,     5.30      ,     5.36   
104ecmult_multi_16383p_g         ,     5.33      ,     5.44      ,     5.83   
105ecmult_multi_18431p_g         ,     5.48      ,     5.51      ,     5.59   
106ecmult_multi_20479p_g         ,     5.65      ,     5.78      ,     6.26   
107ecmult_multi_22527p_g         ,     5.87      ,     5.92      ,     5.96   
108ecmult_multi_24575p_g         ,     6.14      ,     6.24      ,     6.64   
109ecmult_multi_26623p_g         ,     6.43      ,     6.52      ,     6.79   
110ecmult_multi_28671p_g         ,     6.66      ,     6.74      ,     6.86   
111ecmult_multi_30719p_g         ,     6.92      ,     6.98      ,     7.22   
112ecmult_multi_32767p_g         ,     7.14      ,     7.50      ,     8.38
  19. hebasto approved
  20. hebasto commented at 8:23 pm on November 17, 2025: member

    ACK 26166c4f5fd5688c31e488449ee2325eb8f2fb36, I have reviewed the code and it looks OK.

    On my machine (AMD Ryzen AI 7 350), my benchmarking results for this PR differ from others:

    • a tiny (<0.5%) but consistent slowdown for a small number of points
    • for a large number of points, the speedup is less than 2%
  21. real-or-random approved
  22. real-or-random commented at 8:35 am on November 18, 2025: contributor
    ACK 26166c4f5fd5688c31e488449ee2325eb8f2fb36 benchmarks show no significant difference (only tried low point counts)
  23. real-or-random merged this on Nov 18, 2025
  24. real-or-random closed this on Nov 18, 2025


jonasnick real-or-random hebasto siv2r

Labels
performance tweak/refactor

github-metadata-mirror

This is a metadata mirror of the GitHub repository bitcoin-core/secp256k1. This site is not affiliated with GitHub. Content is generated from a GitHub metadata backup.
generated: 2025-11-21 04:15 UTC
This site is hosted by @0xB10C
More mirrored repositories can be found on mirror.b10c.me