中田真秀(なかたまほ)のブログ

研究について、日常について、その他。

memory alignment in MetaTrader4 : 4bytes alignment for "double"

2014-03-20 11:23:12 | 日記
I noticed that LAPACK interface for MetaTrader4 doesn't work with OpenBLAS.

The reason is that OpenBLAS implicitly assumes double type data is aligned by 8bytes = 64bit.
However MT4 aligns as 4bytes! So if SSE2 (or something like that) is used, it gives access violation.

There is no reason to align double array in 4bytes, so MT4 sucks, but adding dummy array looks fine as
a workaround.





Gaussian on GPU by NVIDIA @ GTC

2014-03-11 12:17:51 | 日記

S4613 - Enabling Gaussian 09 on GPGPUs

Roberto Gomperts ( Principal Engineer, NVIDIA )

In 2011 Gaussian, Inc., NVIDIA Corp. and PGI started a long-term project to enable all the performance critical paths of Gaussian on GPGPUs. While the ultimate goal is to show significant performance improvement by using accelerators in conjunction with CPUs, the initial efforts are directed towards creating an infrastructure that will leverage the current CPU code base and at the same time minimize the additional maintenance effort associated with running on GPUs. Here we present the current status of this work for Direct Hartree-Fock and triples-correction calculations as applied in for example Coupled Cluster calculations that uses mostly the directives based OpenACC framework.

Session Level: Intermediate
Session Type: Talk
Tags: Quantum Chemistry; Programming Languages & Compilers; Supercomputing


だそう。direct Hartree-Fock, CCSD(T)の計算をOpenACCディレクティブベースでやる?

クロスアビリティもGaussian on GPUをやっており、もっとcomprehensiveだと思うがプログラムは非常に複雑っぽい

Singular Spectrum Analysis and Forex on MetaTrader4 (Build 610)

2014-03-11 11:12:23 | 日記
Hi MetaTrader4 Freaks,
Here I have coded Singular Spectrum ANalysis sample code on MetaTrader4

Singular Spectrum Analysis (SSA and aka Catapillar method) is a powerful method to extract trends and smoothing for time series, without knowing specific structure. This algorithm embeds time series to trajectory matrix followed by singular value decomposition. Then choose some eigentriples (usually some largest ones), finally, reconstruct to (new) time series.

This implementation is naive one, so that others can modify for their purposes.

TODO
* O(n^2) implemetation of SSA.
etc. etc. etc.