phoenix_transpose_impl.h

Go to the documentation of this file.

/***************************************
        Auteur : Pierre Aubert
        Mail : pierre.aubert@lapp.in2p3.fr
        Licence : CeCILL-C
****************************************/
 
#ifndef __PHOENIX_TRANSPOSE_IMPL_H__
#define __PHOENIX_TRANSPOSE_IMPL_H__
 
#include <algorithm>
#include "phoenix_transpose.h"


template<typename T, typename U>
void phoenix_transpose_copy(T* __restrict__ ptabOutput, const U * __restrict__ ptabInput,
                        size_t nbRowOut, size_t nbColOut, size_t outputPadding, bool isTransposed)
{
        T* tabOutput = phoenix_assume_aligned(ptabOutput);
        const U* tabInput = phoenix_assume_aligned(ptabInput);
        size_t outputRowSize(nbColOut + outputPadding);
        if(!isTransposed){
                for(size_t i(0lu); i < nbRowOut; ++i){
                        for(size_t j(0lu); j < nbColOut; ++j){
                                tabOutput[i*outputRowSize + j] = tabInput[i*nbColOut + j];
                        }
                }
        }else{          //The bad slices rejection from real data have to be done here !!!!!
                for(size_t j(0lu); j < nbColOut; ++j){
                        for(size_t i(0lu); i < nbRowOut; ++i){
                                tabOutput[i*outputRowSize + j] = tabInput[i + j*nbRowOut];
                        }
                }
        }
}


template<typename T, typename U>
void phoenix_transpose_block(T* __restrict__ ptabOutput, const U * __restrict__ ptabInput,
                        size_t nbRowOut, size_t nbColOut, size_t nbBlockRowOut, size_t nbBlockColOut, size_t outputPadding)
{
        T* tabOutput = phoenix_assume_aligned(ptabOutput);
        const U* tabInput = phoenix_assume_aligned(ptabInput);
//      T* tabOutput = ptabOutput;
//      const U* tabInput = ptabInput;
//      std::cerr << "phoenix_transpose_block : nbRowOut = " << nbRowOut << ", nbColOut = " << nbColOut << ", nbBlockRowOut = " << nbBlockRowOut << ", nbBlockColOut = " << nbBlockColOut << std::endl;
        size_t outputRowSize(nbColOut + outputPadding);
        for(size_t i(0lu); i < nbColOut; i += nbBlockColOut){
                for(size_t j(0lu); j < nbRowOut; j += nbBlockRowOut){
                        // transpose the block beginning at [i,j]
                        for(size_t k(i); k < std::min(i + nbBlockColOut, nbColOut); ++k){
                                for(size_t l(j); l < std::min(j + nbBlockRowOut, nbRowOut); ++l){
//                                      std::cerr << "phoenix_transpose_block : (i = "<<i<<", j = "<<j<<") tabOutput["<<k<<" + "<<l<<"*"<<outputRowSize<<"] = tabInput["<<l<<" + "<<k<<"*"<<nbRowOut<<"]" << std::endl;
                                        tabOutput[l*outputRowSize + k] = tabInput[k*nbRowOut + l];
                                }
                        }
                }
        }
}


template<typename T, typename U>
void phoenix_transpose_block_copy(T* __restrict__ ptabOutput, const U * __restrict__ ptabInput,
                        size_t nbRowOut, size_t nbColOut, size_t nbBlockRowOut, size_t nbBlockColOut, size_t outputPadding, bool isTransposed)
{
        if(!isTransposed){
                T* tabOutput = phoenix_assume_aligned(ptabOutput);
                const U* tabInput = phoenix_assume_aligned(ptabInput);
                size_t outputRowSize(nbColOut + outputPadding);
                for(size_t i(0lu); i < nbColOut; i += nbBlockColOut){
                        for(size_t j(0lu); j < nbRowOut; j += nbBlockRowOut){
                                // transpose the block beginning at [i,j]
                                for(size_t k(i); k < std::min(i + nbBlockColOut, nbColOut); ++k){
                                        for(size_t l(j); l < std::min(j + nbBlockRowOut, nbRowOut); ++l){
                                                tabOutput[l*outputRowSize + k] = tabInput[l*nbRowOut + k];
                                        }
                                }
                        }
                }
        }else{          //The bad slices rejection from real data have to be done here !!!!!
                phoenix_transpose_block(ptabOutput, ptabInput,
                        nbRowOut, nbColOut, nbBlockRowOut, nbBlockColOut, outputPadding);
        }
}
 
 
#endif