Citation Context

...hich can be parallelized. To cheaply and conservatively ensure that parallelized code executes only on rectangular arrays of unaliased subarrays, our implementation adopts the dense flag technique of =-=[26]-=-. An extra one bit flag is added to the header of every multi-dimensional array. When the array is created using the multianewarray bytecode instruction, the flag is set to true. The array returned by...

Citation Context

...ions (if all dimensions have asterisks, the array is actually an ordinary, nondistributed, Fortran-like, multidimensional array—a valuable addition to Java in its own right, as many people have note=-=d [5, 6]).-=- The variables a, b, and c are all distributed array variables. The creation expressions on the right hand side of the initializes specify that the arrays here all have ranges x and y—they are all M...

Citation Context

...it would be straightforward to modify our code generator such that instead of using standard Java arrays, it generates the code in terms of a multidimensional array package, such as those proposed in =-=[25]-=-. Other techniques, such as recognizing vector operations and generating the appropriate method call, would require more significant modifications. 6. CONCLUSION We have demonstrated that it is feasib...

Citation Context

...ed arrays available in the Java language. An X10 multi-dimensional array is flattened and represented internally as a combination of an X10 array descriptor object and a single-dimensional Java array =-=[42]-=-. The array descriptor is part of a boxed array implementation which provides information on the distribution and shape of the array, as well as convenient support for collective operations. X10 also ...

Citation Context

...cluster the row objects together in memory [15]. However this accounts for only a minor part of the dierence. Other suggestion is to make a multidimensional Java array class avoiding array of arrays [=-=16-=-]. For sparse matrices we have illustrated the eect of having the possibility to manipulate only the rows of the structure without updating the rest of the structure. We still do not know if Java Spar...

Citation Context

...ons (if all dimensions have asterisks, the array is actually an ordinary, non-distributed, Fortran-like, multi-dimensional array—a valuable adjunct to Java in its own right, as many people have noted =-=[11,12]-=-). The constructors on the right-hand side of the initializers specify that the arrays here all have ranges x and y—they are all M by N arrays, block-distributed over p. We see that mapping of distrib...

Citation Context

... for integer or double. Thus, handling large data structures requires using primitive types. Allocation strategy of multi-dimensional arrays also has an important impact on memory footprint, because Java does not have support for true multidimensional arrays. Instead, it relies on arrays of arrays to simulate them. This leads to non rectangular arrays with variable shapes. Also, Java arrays are actually Objects even if they only contain primitive type data. It is very hard for a compiler to perform optimizations on such array. Although some solutions have been proposed to address these issues [15], none has made it in the official releases of Java. Table IV-C shows the memory usage of various JVMs when allocating a 2-dimensions array of 20M elements (double or byte). We also indicate the value measured on a C and a Fortran versions compiled with gcc. We have measured 3 different allocations strategies : [2][10M], [10M][2M] and [20M]. As expected, the lowest usage is obtained when allocating a single dimension array, as the memory can be allocated contiguously in memory. When allocating twodimensional arrays, we see that the Oracle JRockit JVM performs sometimes better than the C versio...