< prof >

	fft in x-direction are reduced by half

Author: Qianruipku

source/module_pw/fft.cpp

@@ -36,7 +36,7 @@ FFT::~FFT()
 #endif
 }
 
-void FFT:: initfft(int nx_in, int bigny_in, int nz_in, int ns_in, int nplane_in, int nproc_in, bool gamma_only_in, bool mpifft_in)
+void FFT:: initfft(int nx_in, int bigny_in, int nz_in, int lix_in, int rix_in, int ns_in, int nplane_in, int nproc_in, bool gamma_only_in, bool mpifft_in)
 {
 	this->gamma_only = gamma_only_in;
 	this->nx = nx_in;
@@ -45,6 +45,8 @@ void FFT:: initfft(int nx_in, int bigny_in, int nz_in, int ns_in, int nplane_in,
 	else	this->ny = this->bigny;
 	this->nz = nz_in;
 	this->ns = ns_in;
+	this->lix = lix_in;
+	this->rix = rix_in;
 	this->nplane = nplane_in;
 	this->nproc = nproc_in;
 	this->mpifft = mpifft_in;
@@ -330,7 +332,11 @@ void FFT::fftxyfor(std::complex<double>* & in, std::complex<double>* & out)
 {
 	int npy = this->nplane * this-> ny;
 	fftw_execute_dft( this->planxfor, (fftw_complex *)in, (fftw_complex *)out);
-	for (int i=0; i<this->nx;++i)
+	for (int i=0; i<=this->lix;++i)
+	{
+		fftw_execute_dft( this->planyfor, (fftw_complex*)&in[i*npy], (fftw_complex*)&out[i*npy] );
+	}
+	for (int i=this->rix; i<this->nx;++i)
 	{
 		fftw_execute_dft( this->planyfor, (fftw_complex*)&in[i*npy], (fftw_complex*)&out[i*npy] );
 	}
@@ -341,7 +347,11 @@ void FFT::fftxybac(std::complex<double>* & in, std::complex<double>* & out)
 {
 	int npy = this->nplane * this-> ny;
 
-	for (int i=0; i<this->nx;++i)
+	for (int i=0; i<=this->lix;++i)
+	{
+		fftw_execute_dft( this->planybac, (fftw_complex*)&in[i*npy], (fftw_complex*)&out[i*npy] );
+	}
+	for (int i=this->rix; i<this->nx;++i)
 	{
 		fftw_execute_dft( this->planybac, (fftw_complex*)&in[i*npy], (fftw_complex*)&out[i*npy] );
 	}

source/module_pw/fft.h

@@ -27,7 +27,8 @@ class FFT
 
 	FFT();
 	~FFT();
-	void initfft(int nx_in, int bigny_in, int nz_in, int ns_in, int nplane_in, int nproc_in, bool gamma_only_in, bool mpifft_in = false);
+	void initfft(int nx_in, int bigny_in, int nz_in, int lix_in, int rix_in, int ns_in, int nplane_in, 
+				 int nproc_in, bool gamma_only_in, bool mpifft_in = false);
 	void setupFFT();
 	void cleanFFT();
 
@@ -55,6 +56,7 @@ class FFT
 	int nxy;
 	int bigny;
 	int bignxy;
+	int lix,rix;// lix: the left edge of the pw ball; rix: the right edge of the pw ball
 	int ns; //number of sticks
 	int nplane; //number of x-y planes
 	int maxgrids; // max between nz * ns and bignxy * nplane

source/module_pw/pw_basis.cpp

@@ -54,7 +54,7 @@ void PW_Basis::setuptransform()
     this->distribute_r();
     this->distribute_g();
     this->getstartgr();
-    this->ft.initfft(this->nx,this->bigny,this->nz,this->nst,this->nplane,this->poolnproc,this->gamma_only);
+    this->ft.initfft(this->nx,this->bigny,this->nz,this->lix,this->rix,this->nst,this->nplane,this->poolnproc,this->gamma_only);
     this->ft.setupFFT();
 }
 

source/module_pw/pw_basis.h

@@ -184,7 +184,7 @@ class PW_Basis
 	// FFT dimensions for wave functions.
 	int nx, ny, nz, nxyz, nxy;
     int bigny, bignxyz, bignxy; // Gamma_only: ny = int(bigny/2)-1 , others: ny = bigny
-    int liy,riy;// liy: the left edge of the pw ball; riy: the right edge of the pw ball
+    int lix,rix;// lix: the left edge of the pw ball; rix: the right edge of the pw ball
     int maxgrids; // max between nz * ns and bignxy * nplane
     FFT ft;
 

source/module_pw/pw_distributeg.cpp

@@ -54,6 +54,7 @@ void PW_Basis::count_pw_st(
         iy_start = 0;
         iy_end = this->ny - 1;
     }
+    this->lix = this->rix = 0;
     for (int ix = -ibox[0]; ix <= ibox[0]; ++ix)
     {
         for (int iy = iy_start; iy <= iy_end; ++iy)
@@ -81,6 +82,8 @@ void PW_Basis::count_pw_st(
                     if (length == 0) st_bottom2D[index] = iz; // length == 0 means this point is the bottom of stick (x, y).
                     ++tot_npw;
                     ++length;
+                    if(ix < this->rix) this->rix = ix;
+                    if(ix > this->lix) this->lix = ix;
                 }
             }
             if (length > 0)
@@ -90,6 +93,8 @@ void PW_Basis::count_pw_st(
             }
         }
     }
+    if(rix <= 0) rix += this->nx;
+    std::cout<<"lix "<<lix<<" ; rix "<<rix<<std::endl;
     return;
 }
 

source/module_pw/pw_distributeg_method1.cpp

@@ -156,6 +156,8 @@ void PW_Basis::distribution_method1()
 #ifdef __MPI
     MPI_Bcast(&tot_npw, 1, MPI_INT, 0, POOL_WORLD);
     MPI_Bcast(&this->nstot, 1, MPI_INT, 0, POOL_WORLD);
+    MPI_Bcast(&lix, 1, MPI_INT, 0, POOL_WORLD);
+    MPI_Bcast(&rix, 1, MPI_INT, 0, POOL_WORLD);
     if (this->poolrank != 0)
     {
         st_bottom2D = new int[this->nxy];                      // minimum z of stick.
@@ -224,7 +226,6 @@ void PW_Basis::collect_st(
         iy_start = 0;
         iy_end = this->ny - 1;
     }
-    this->liy = this->riy = 0;
     for (int ix = -ibox[0]; ix <= ibox[0]; ++ix)
     {
         for (int iy = iy_start; iy <= iy_end; ++iy)
@@ -259,17 +260,13 @@ void PW_Basis::collect_st(
                     temp_st_i[is] = x;
                     temp_st_j[is] = y;
                     temp_st_length[is] = static_cast<double>(st_length2D[index]);
-                    if(iy < this->riy) this->riy = iy;
-                    if(iy > this->liy) this->liy = iy;
                     ++is;
                     std::cout << "is   " << is << '\n'; 
                 }   
             }
         }
     }
     assert(is == this->nstot);
-    if(riy <= 0) riy += this->ny;
-    std::cout<<"liy "<<liy<<" ; riy "<<riy<<std::endl;
     std::cout<<"collect sticks done\n";
 
     // As we will distribute the longest sticks preferentially in Step(3), we rearrange st_* in the order of length decreasing.

source/module_pw/pw_distributeg_method2.cpp

@@ -105,6 +105,8 @@ void PW_Basis::distribution_method2()
 #ifdef __MPI
     MPI_Bcast(&tot_npw, 1, MPI_INT, 0, POOL_WORLD);
     MPI_Bcast(&this->nstot, 1, MPI_INT, 0, POOL_WORLD);
+    MPI_Bcast(&lix, 1, MPI_INT, 0, POOL_WORLD);
+    MPI_Bcast(&rix, 1, MPI_INT, 0, POOL_WORLD);
     MPI_Bcast(this->nst_per, this->poolnproc, MPI_INT, 0 , POOL_WORLD);
     if (this->poolrank != 0)
     {