| Project/Area Number |
25330144
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Multi-year Fund |
|---|
| Section | 一般 |
|---|
| Research Field |
High performance computing
|
|---|
| Research Institution | Kogakuin University |
|---|
Principal Investigator |
Tanaka Teruo 工学院大学, 情報学部(情報工学部), 教授 (90622837)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
長谷川 秀彦 筑波大学, 図書館情報メディア系, 教授 (20164824)
|
|---|
| Co-Investigator(Renkei-kenkyūsha) |
FUJII Akihiro 工学院大学, 情報学部(情報工学部), 准教授 (10383986)
|
|---|
| Project Period (FY) |
2013-04-01 – 2017-03-31
|
| Project Status |
Completed (Fiscal Year 2016)
|
| Budget Amount *help |
¥4,290,000 (Direct Cost: ¥3,300,000、Indirect Cost: ¥990,000)
Fiscal Year 2016: ¥650,000 (Direct Cost: ¥500,000、Indirect Cost: ¥150,000)
Fiscal Year 2015: ¥1,040,000 (Direct Cost: ¥800,000、Indirect Cost: ¥240,000)
Fiscal Year 2014: ¥1,040,000 (Direct Cost: ¥800,000、Indirect Cost: ¥240,000)
Fiscal Year 2013: ¥1,560,000 (Direct Cost: ¥1,200,000、Indirect Cost: ¥360,000)
|
|---|
| Keywords | 倍々精度計算 / 任意多倍長計算 / AVX / GMP / 通信量削減 / MPI / 自動チューニング / 性能パラメタ / 多倍長演算 / 倍々精度演算 / 集団通信 / ４倍精度演算 / 反復計算ライブラリ / 疎行列計算 / マルチコア / ベクトル演算 |
|---|
| Outline of Final Research Achievements |
1. We developed DD-AVX, a library of Double-Double (DD) precision matrix and/or vector operations accelerated by AVX and AVX2 SIMD instructions. We also developed Xev-GMP, a directive-based automatic code generation for a C code with multiple-precision floating-point-operation data from a C code with double precision data. The GMP code uses the GNU Multiple Precision Arithmetic Library. Xev-GMP can also support MPI libraries.
2. We proposed and evaluated CBCGR method which reduces MPI collective communication for the Chebyshev base conjugate gradient method in Massively parallel processing. We also devised a calculation pattern to remove the overlap in the Matrix Power Kernel method.
3. We implemented the simultaneous estimation method of two performance parameters on automatic tuning base pp-OpenAT. The enhanced version of pp-OpenAT is made public. Furthermore, we suggested a repetitive linear search method for the practical use of performance parameters more than three.
|
