BigInt Multiplication

BigInt Multiplication provides an example for multiplying two 64-bit arrays. Performs a total of leftLength x rightLength multiplication operations. It does not require the result array to be reset while the calculation is performed and sequential access is provided while reaching the result. Performing multiplication sequentially may be more amenable to parallel computation in the future. Especially if features like AVX, 64 bit multiplication and writing the result as low and high bits and addition with carry are included. Of course, loading the registers (ymm, zmm) from memory in reverse order can bring an extra performance increase.

Decimal digits count of 64-bit arrays

When performing mathematical operations, we may need operations with higher precision than numbers such as double or float. For example, if we want to find the roots of a quartic function as a double, the input of the function must be greater than the double precision.

Length	Byte Count	Decimal Count
8	64	155
16	128	309
32	256	617
64	512	1234
128	1024	2467
256	2048	4933
512	4096	9865
1024	8192	19719
2048	16384	39447

Calculation Process

For example, if leftLength is 8 and rightLength is 5, the order of operations is listed in the table below. Result Index is the index where the multiplication result will be written. Left Index, Right Index values are multiplied and each product is added and written to Result Index. Rows can be divided into 3 regions. Start [0-4], middle [5-7], finish [8, 11]. Thus, the code is run without making too many checks in loops.

Result Index	Left Index, Right Index
0	0,0
1	0,1 1,0
2	0,2 1,1 2,0
3	0,3 1,2 2,1 3,0
4	0,4 1,3 2,2 3,1 4,0
5	X 1,4 2,3 3,2 4,1 5,0
6	X X 2,4 3,3 4,2 5,1 6,0
7	X X X 3,4 4,3 5,2 6,1 7,0
8	X X X X 4,4 5,3 6,2 7,1 X
9	X X X X X 5,4 6,3 7,2 X X
10	X X X X X X 6,4 7,3 X X X
11	X X X X X X X 7,4 X X X X

Performance Results

In the performance table below, two random 64-bit numbers of length leftLength and rightLength were multiplied. Each line was run 300 times and the minimum ElapsedTicks values were written with StopWatch(.NET). ExpectedTick(max) belongs to BigInteger(.NET).

leftLength	rightLength	expectedTick(max)	actualTick	Percent
16	16	9	6	66%
16	32	18	12	66%
16	64	36	21	58%
16	128	72	40	55%
16	256	141	78	55%
16	512	278	144	51%
16	1024	561	265	47%
16	2048	1128	528	46%
32	16	18	12	66%
32	32	30	23	76%
32	64	58	45	77%
32	128	116	82	70%
32	256	233	170	72%
32	512	467	334	71%
32	1024	931	640	68%
32	2048	1871	1323	70%
64	16	48	23	47%
64	32	85	46	54%
64	64	144	85	59%
64	128	289	183	63%
64	256	330	356	107%
64	512	654	696	106%
64	1024	1308	1390	106%
64	2048	2612	2709	103%
128	16	61	37	60%
128	32	100	84	84%
128	64	162	176	108%
128	128	253	355	140%
128	256	503	723	143%
128	512	1009	1405	139%
128	1024	1996	2834	141%
128	2048	4039	5794	143%
256	16	123	72	58%
256	32	200	169	84%
256	64	325	351	108%
256	128	504	707	140%
256	256	774	1467	189%
256	512	1543	2963	192%
256	1024	3046	5860	192%
256	2048	5796	11690	201%
512	16	243	143	58%
512	32	400	317	79%
512	64	650	686	105%
512	128	1010	1451	143%
512	256	1545	2954	191%
512	512	2341	5972	255%
512	1024	4377	12077	275%
512	2048	8760	23290	265%
1024	16	488	266	54%
1024	32	803	671	83%
1024	64	1278	1398	109%
1024	128	2024	2849	140%
1024	256	3083	5869	190%
1024	512	4467	11852	265%
1024	1024	6603	23285	352%
1024	2048	12974	47030	362%
2048	16	930	493	53%
2048	32	1602	1293	80%
2048	64	2607	2777	106%
2048	128	4032	5703	141%
2048	256	5770	11730	203%
2048	512	8731	22610	258%
2048	1024	13186	46620	353%
2048	2048	20239	94075	464%

Conclusion

This method can provide better performance than BigInteger for values smaller than 1234 digits. Perhaps using this method, two 64-bit numbers can be multiplied simultaneously with AVX-256 and all performance lines can be achieved.