diff options
| author | tradke <tradke@10716dce-81a3-4424-a2c8-48026a0d3035> | 2000-11-23 09:28:15 +0000 |
|---|---|---|
| committer | tradke <tradke@10716dce-81a3-4424-a2c8-48026a0d3035> | 2000-11-23 09:28:15 +0000 |
| commit | e05d659b4d6dbbab4a201e37d5cb634c1322e5e7 (patch) | |
| tree | 252db94e48109148a7f86567de29084bbb4752d6 | |
| parent | 6e18d9894709ffedb3f15e5054678a273ca65f92 (diff) | |
Optimized hyperslab extraction for downsampled cases.
git-svn-id: http://svn.cactuscode.org/arrangements/CactusPUGH/PUGHSlab/trunk@38 10716dce-81a3-4424-a2c8-48026a0d3035
| -rw-r--r-- | src/Hyperslab.c | 46 |
1 files changed, 36 insertions, 10 deletions
diff --git a/src/Hyperslab.c b/src/Hyperslab.c index b9fd31e..dba8d8a 100644 --- a/src/Hyperslab.c +++ b/src/Hyperslab.c @@ -38,7 +38,7 @@ #define PICKUP_HYPERSLAB_DATA(cctk_type, vdim, vdata, hdata, point, \ startpoint, endpoint, downsample,points_per_dim)\ { \ - int i, dim, vindex; \ + int i, dim, vindex, dim0_elements; \ cctk_type *typed_vdata = (cctk_type *) vdata; \ cctk_type *typed_hdata = (cctk_type *) hdata; \ \ @@ -46,12 +46,14 @@ /* set point to local startpoint */ \ memcpy (point, startpoint, vdim * sizeof (point[0])); \ \ + dim0_elements = endpoint[0] - startpoint[0]; \ + \ /* do the nested loops starting with the innermost */ \ - dim = 0; \ + dim = 1; \ while (1) \ { \ /* check for end of current loop */ \ - if (point[dim] >= endpoint[dim]) \ + if (vdim > 1 && point[dim] >= endpoint[dim]) \ { \ /* increment outermost loopers */ \ for (dim++; dim < vdim; dim++) \ @@ -70,23 +72,47 @@ } \ \ /* reset innermost loopers */ \ - for (dim--; dim >= 0; dim--) \ + for (dim--; dim > 0; dim--) \ { \ point[dim] = startpoint[dim]; \ } \ - dim = 0; \ + dim = 1; \ } \ \ - /* copy the data point */ \ + /* get the linear offset */ \ vindex = point[0]; \ for (i = 1; i < vdim; i++) \ { \ vindex += point[i] * points_per_dim[i]; \ } \ - *typed_hdata++ = typed_vdata[vindex]; \ \ - /* increment current looper */ \ - point[dim] += downsample[dim]; \ + /* copy the data in lowest dimension */ \ + /* if possible (no downsampling was requested) use plain memcpy() \ + otherwise copy element-wise in a loop */ \ + if (downsample[0] == 1) \ + { \ + memcpy (typed_hdata, typed_vdata, \ + dim0_elements * sizeof (*typed_hdata)); \ + typed_hdata += dim0_elements; \ + } \ + else \ + { \ + for (i = 0; i < dim0_elements; i += downsample[0]) \ + { \ + *typed_hdata++ = typed_vdata[vindex + i]; \ + } \ + } \ + \ + if (vdim > 1) \ + { \ + /* increment current looper */ \ + point[dim] += downsample[dim]; \ + } \ + else \ + { \ + /* exit loop if hyperslab dim is only 1D */ \ + break; \ + } \ \ } /* end of nested loops over all dimensions */ \ } @@ -451,7 +477,7 @@ int Hyperslab_GetLocalHyperslab (cGH *GH, int vindex, int vtimelvl, for (vdim = 1; vdim < vinfo.dim; vdim++) { points_per_dim[vdim] = points_per_dim[vdim-1] * - GA->extras->lnsize[vdim-1]; + GA->extras->lnsize[vdim-1]; } /* now get the hyperslab data points using that wonderful macro... */ |