Le samedi 05 mars 2005 à 21:13 +0100, Christoph Ludwig a écrit :
Can you point me to any data about the speedup of bigint operations due to ET? How complex need the expressions to be and what is the bitlength threshold so that programs profit from ET? (I am aware that ET can have a huge effect on the speed of matrix operations, but there you typically need to allocate / copy / deallocate much more data if you have temporary objects.)
Does anyone know what speedup one can expect if the move-semantic proposal is applied to bigint types? (Say, for simple expressions like x = y + z and x = y * z.)
I can't tell you what it would be for this particular implementation, but I can give you some results I got when using GMP some times ago. I hope everybody will agree that GMP is a high quality implementation of big integers, and hence the results are relevant. The tests were comparing the cost of a superfluous temporary when using basic arithmetic operations on 256 bit wide integers. So it applies to your question, since we can suppose that a correct ET implementation would not use a temporary for a single operation. When using a temporary, a multiplication was executed around two times slower, and a addition was executed around six times slower. Since GMP is heavily optimized, the cost of a temporary (allocating memory, copying data, freeing memory) for each operation is no more negligible. Quite the contrary in fact, since it accounts for more than 80% of the cost of an addition. And the cost of a temporary becomes even worse, when the additional memory requirements kick you out of L1 cache. Especially since a big expression with temporaries can quickly require twice the memory needed for the same expression without temporaries. Best regards, Guillaume The body of the test for "c = a * b;" was mpz_t d; mpz_init(d); mpz_mul(d, a, b); mpz_set(c, d); mpz_clear(d); when a temporary d was used to store the result of the multiplication. Without temporary, it was simply mpz_mul(c, a, b);