use fftw for inplace transposition

src/RMHD_converter.cpp

@@ -122,13 +122,12 @@ int RMHD_converter::convert(
 
     fftwf_clip_zero_padding(this->f4r, this->r3);
 
-    // new array where mixed components will be placed
-    float *rtmp = fftwf_alloc_real( 2*this->f3r->local_size);
-
     // mix components
-    this->f3r->interleave(this->r3, rtmp, 2);
+    this->f3r->interleave(this->r3, 2);
 
-    this->s->shuffle(rtmp, this->r3, ofile);
+    // new array where mixed components will be placed
+    float *rtmp = fftwf_alloc_real( 2*this->f3r->local_size);
+    this->s->shuffle(this->r3, rtmp, ofile);
 
     fftwf_free(rtmp);
     return EXIT_SUCCESS;

src/field_descriptor.cpp

@@ -239,13 +239,23 @@ int field_descriptor::transpose(
 
 int field_descriptor::interleave(
         float *input,
-        float *output,
         int dim)
 {
-    // TODO: implement inplace interleaver
-    for (int k = 0; k < this->local_size; k++)
-        for (int j = 0; j < dim; j++)
-            output[k*dim + j] = input[j*this->local_size + k];
+    fftwf_iodim howmany_dims[2];
+    howmany_dims[0].n  = dim;
+    howmany_dims[0].is = this->local_size;
+    howmany_dims[0].os = 1;
+    howmany_dims[1].n  = this->local_size;
+    howmany_dims[1].is = 1;
+    howmany_dims[1].os = dim;
+    const int howmany_rank = sizeof(howmany_dims)/sizeof(howmany_dims[0]);
+
+    fftwf_plan tmp = fftwf_plan_guru_r2r(
+            /*rank*/0, /*dims*/NULL,
+            howmany_rank, howmany_dims,
+            input, input, /*kind*/NULL, FFTW_ESTIMATE);
+    fftwf_execute(tmp);
+    fftwf_destroy_plan(tmp);
     return EXIT_SUCCESS;
 }
 

src/field_descriptor.hpp

@@ -57,7 +57,6 @@ class field_descriptor
 
         int interleave(
                 float *input,
-                float *output,
                 int dim);
         int interleave(
                 fftw_complex *input,