Hadar/b const mult #723
Hadar/b const mult #723
Conversation
| @@ -25,6 +25,11 @@ namespace bls12_377 { | |||
| static constexpr point_field_t weierstrass_b = {0x00000001, 0x00000000, 0x00000000, 0x00000000, | |||
| 0x00000000, 0x00000000, 0x00000000, 0x00000000, | |||
| 0x00000000, 0x00000000, 0x00000000, 0x00000000}; | |||
| static constexpr point_field_t weierstrass_3b = {0x00000003, 0x00000000, 0x00000000, 0x00000000, | |||
There was a problem hiding this comment.
do we really need that extra definition?
using constexpr in the code does not yield the same performance?
There was a problem hiding this comment.
it probably will, i can try
| if constexpr (Gen::is_b_u32) { | ||
| r = mul_unsigned<Field{Gen::weierstrass_3b}.limbs_storage.limbs[0], Field>(xs); | ||
| if constexpr (Gen::is_b_neg) | ||
| return neg(r); |
There was a problem hiding this comment.
I think it will be more efficient to write here another multiplication by neg(b) which can be calculated in constexpr
There was a problem hiding this comment.
I don't understand.. xs is known only at runtime, this cannot be calculated in constexpr
There was a problem hiding this comment.
according to your code, the neg(r) is calculated in runtime
you can write instead of "return neg(r)"
return mul_unsigned<neg(Field{Gen::weierstrass_3b}.limbs_storage.limbs[0], Field)>(xs);
that way the neg operation will be calculated at compile time
There was a problem hiding this comment.
that will ruin the whole point of the optimization... 3b is stored in absolute value because it's small that way
| if constexpr (Gen::is_b_u32_g2_re) { | ||
| r = FF::template mul_unsigned<FF{Gen::weierstrass_3b_g2_re}.limbs_storage.limbs[0], FF>(xs); | ||
| if constexpr (Gen::is_b_neg_g2_re) | ||
| return FF::neg(r); |
There was a problem hiding this comment.
I think it will be more efficient to write here another multiplication by neg(b) which can be calculated in constexpr
|
. |
This PR optimizes the multiplication by the constant 3b in the EC addition.
It seems to significantly improve the CPU MSM performance and the CUDA MSM performance a tiny bit.