Changes between Version 27 and Version 28 of MatrixMultiply
- Timestamp:
- Jun 23, 2010 5:21:17 PM (14 years ago)
Legend:
- Unmodified
- Added
- Removed
- Modified
-
MatrixMultiply
v27 v28 2 2 We have implemented single/double precision matrix multiply program for RV770/Cypress. In our implementation, we use two input streams for computing C=AB. One is transposed input matrix A (i.e. column major) and other is input matrix B in normal format (i.e. row major). Output matrix C is also row major. We adopted 8x8 block for single precision and 4x4 for double precision. Here is benchmark result for each case. Note only kernel execution time is measured. 3 3 4 Add double-double precision performance. We used 2x2 block in this case. On Cypress architecture GPU, we take advantage of FMA_64 instruction.4 Add double-double (DD) precision performance. We used 2x2 block in this case. On Cypress architecture GPU, we take advantage of FMA_64 instruction. For MAD peak in DD, we assume one DD operation takes 20 DP operations(ops) without FMA and 15 ops with FMA. Precicely, DD add takes ~ 20 ops and DD mul wihtou FMA takes ~ 20. DD mul with FMA takes ~ 8 ops. 5 5 6 6 == Peformance Summary == 7 || board || Pmax || Nmax || prec || reg. usage || MAD peak || 8 || HD4850 || 736 || 3328 || SP || 25 || 1040 || 9 || HD5870 || 2140 || 7424 || SP || 25 || 2720 || 10 || HD4850 || 177 || 1408 || DP || 19 || 208 || 11 || HD5870 || 475 || 2048 || DP || 19 || 544 || 12 || HD4850 || 7.5 || 768 || DDP || 21 || || 13 || HD5870 || 20 || 1024 || DDP || 21 || || 14 || HD5870 FMA || 31 || 1024 || DDP || 18 || || 7 || board || Pmax || Nmax || prec || reg. usage || MAD peak || note || 8 || HD4850 || 736 || 3328 || SP || 25 || 1040 |||| 9 || HD5870 || 2140 || 7424 || SP || 25 || 2720 |||| 10 || HD4850 || 177 || 1408 || DP || 19 || 208 |||| 11 || HD5870 || 475 || 2048 || DP || 19 || 544 |||| 12 || HD4850 || 7.5 || 768 || DDP || 21 || ~10.4 |||| 13 || HD5870 || 20 || 1024 || DDP || 21 || ~27.2 |||| 14 || HD5870 || 31 || 1024 || DDP || 18 || ~36.2 || FMA || 15 16 15 17 16 18 Pmax & MAD in GFLOPS