PoissonSolver3DCylindricalGPU adalah pustaka yang dikembangkan untuk menyelesaikan persamaan Poisson 3 dimensi dalam sistem koordinat silinder dengan akselerator GPU
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 

2242 lines
81 KiB

/**************************************************************************
* Copyright(c) 2018, *
* Kelompok penelitian komputasi berkinerja tinggi *
* Pusat Penelitian Informatika *
* Lembaga Ilmu Pengetahuan Indonesia *
* All rights reserved. *
* *
* Contributors are mentioned in the code where appropriate. *
* *
* Permission to use, copy, modify and distribute this software and its *
* documentation strictly for non-commercial purposes is hereby granted *
* without fee, provided that the above copyright notice appears in all *
* copies and that both the copyright notice and this permission notice *
* appear in the supporting documentation. The authors make no claims *
* about the suitability of this software for any purpose. It is *
* provided "as is" without express or implied warranty. *
**************************************************************************/
#include "PoissonSolver3DGPU.h"
#include <cuda.h>
#include <math.h>
// GPU constant variables
__device__ __constant__ int d_coef_StartPos;
__device__ __constant__ int d_grid_StartPos;
__device__ __constant__ float d_h2;
__device__ __constant__ float d_ih2;
__device__ __constant__ float d_tempRatioZ;
/* GPU kernels start */
/// Relaksasi menggunakan penyelesaian iteratif Red-Black Gauss-Seidel (bagian Red)
///
/// \param VPotential float* Array potensial
/// \param RhoChargeDensity float* Array rapat arus
/// \param RRow int Jumlah baris di arah sumbu \f$ r \f$
/// \param ZColumn int Jumlah kolom di arah sumbu \f$ z \f$
/// \param PhiSlice int Jumlah irisan di arah sumbu \f$ \phi \f$
/// \param coef1 float* Array untuk koefisien \f$ V_{x+1,y,z} \f$
/// \param coef2 float* Array untuk koefisien \f$ V_{x-1,y,z} \f$
/// \param coef3 float* Array untuk koefisien \f$ z \f$
/// \param coef4 float* Array untuk koefisien \f$ f(r,\phi,z) \f$
__global__ void relaxationGaussSeidelRed
(
float *VPotential,
float *RhoChargeDensity,
const int RRow,
const int ZColumn,
const int PhiSlice,
float *coef1,
float *coef2,
float *coef3,
float *coef4
)
{
int index_x, index_y, index, index_left, index_right, index_up, index_down, index_front, index_back, index_coef;
index_x = blockIdx.x * blockDim.x + threadIdx.x;
index_y = blockIdx.y * blockDim.y + threadIdx.y;
index = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + index_x;
index_left = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + (index_x - 1);
index_right = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + (index_x + 1);
index_up = d_grid_StartPos + blockIdx.z * RRow * ZColumn + (index_y - 1) * ZColumn + index_x;
index_down = d_grid_StartPos + blockIdx.z * RRow * ZColumn + (index_y + 1) * ZColumn + index_x;
index_front = d_grid_StartPos + ((blockIdx.z - 1 + PhiSlice) % PhiSlice) * RRow * ZColumn + index_y * ZColumn + index_x;
index_back = d_grid_StartPos + ((blockIdx.z + 1) % PhiSlice) * RRow * ZColumn + index_y * ZColumn + index_x;
index_coef = d_coef_StartPos + index_y;
if (index_x != 0 && index_x < (ZColumn - 1) && index_y != 0 && index_y < (RRow - 1))
{
//calculate red
if ((blockIdx.z % 2 == 0 && (index_x + index_y) % 2 == 0) || (blockIdx.z % 2 != 0 && (index_x + index_y) % 2 != 0))
{
VPotential[index] = (coef2[index_coef] * VPotential[index_up] +
coef1[index_coef] * VPotential[index_down] +
d_tempRatioZ * (VPotential[index_left] + VPotential[index_right]) +
coef3[index_coef] * (VPotential[index_front] + VPotential[index_back]) +
d_h2 * RhoChargeDensity[index]) * coef4[index_coef];
}
}
}
/// Relaksasi menggunakan penyelesaian iteratif Red-Black Gauss-Seidel (bagian Black)
///
/// \param VPotential float* Array potensial
/// \param RhoChargeDensity float* Array rapat arus
/// \param RRow int Jumlah baris di arah sumbu \f$ r \f$
/// \param ZColumn int Jumlah kolom di arah sumbu \f$ z \f$
/// \param PhiSlice int Jumlah irisan di arah sumbu \f$ \phi \f$
/// \param coef1 float* Array untuk koefisien \f$ V_{x+1,y,z} \f$
/// \param coef2 float* Array untuk koefisien \f$ V_{x-1,y,z} \f$
/// \param coef3 float* Array untuk koefisien \f$ z \f$
/// \param coef4 float* Array untuk koefisien \f$ f(r,\phi,z) \f$
__global__ void relaxationGaussSeidelBlack
(
float *VPotential,
float *RhoChargeDensity,
const int RRow,
const int ZColumn,
const int PhiSlice,
float *coef1,
float *coef2,
float *coef3,
float *coef4
)
{
int index_x, index_y, index, index_left, index_right, index_up, index_down, index_front, index_back, index_coef;
index_x = blockIdx.x * blockDim.x + threadIdx.x;
index_y = blockIdx.y * blockDim.y + threadIdx.y;
index = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + index_x;
index_left = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + (index_x - 1);
index_right = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + (index_x + 1);
index_up = d_grid_StartPos + blockIdx.z * RRow * ZColumn + (index_y - 1) * ZColumn + index_x;
index_down = d_grid_StartPos + blockIdx.z * RRow * ZColumn + (index_y + 1) * ZColumn + index_x;
index_front = d_grid_StartPos + ((blockIdx.z - 1 + PhiSlice) % PhiSlice) * RRow * ZColumn + index_y * ZColumn + index_x;
index_back = d_grid_StartPos + ((blockIdx.z + 1) % PhiSlice) * RRow * ZColumn + index_y * ZColumn + index_x;
index_coef = d_coef_StartPos + index_y;
if (index_x != 0 && index_x < (ZColumn - 1) && index_y != 0 && index_y < (RRow - 1))
{
//calculate black
if ((blockIdx.z % 2 == 0 && (index_x + index_y) % 2 != 0) || (blockIdx.z % 2 != 0 && (index_x + index_y) % 2 == 0))
{
VPotential[index] = (coef2[index_coef] * VPotential[index_up] +
coef1[index_coef] * VPotential[index_down] +
d_tempRatioZ * (VPotential[index_left] + VPotential[index_right]) +
coef3[index_coef] * (VPotential[index_front] + VPotential[index_back]) +
d_h2 * RhoChargeDensity[index]) * coef4[index_coef];
}
}
}
/// Menghitung residu dari hasil proses relaksasi
///
/// Rumus:
///
///
/// \param VPotential float* Array potensial
/// \param RhoChargeDensity float* Array rapat arus
/// \param DeltaResidue float* Array residu
/// \param RRow int Jumlah baris di arah sumbu \f$ r \f$
/// \param ZColumn int Jumlah kolom di arah sumbu \f$ z \f$
/// \param PhiSlice int Jumlah irisan di arah sumbu \f$ \phi \f$
/// \param coef1 float* Array untuk koefisien \f$ V_{x+1,y,z} \f$
/// \param coef2 float* Array untuk koefisien \f$ V_{x-1,y,z} \f$
/// \param coef3 float* Array untuk koefisien \f$ z \f$
/// \param icoef4 float* Array untuk koefisien invers dari \f$ f(r,\phi,z) \f$
__global__ void residueCalculation
(
float *VPotential,
float *RhoChargeDensity,
float *DeltaResidue,
const int RRow,
const int ZColumn,
const int PhiSlice,
float *coef1,
float *coef2,
float *coef3,
float *icoef4
)
{
int index_x, index_y, index, index_left, index_right, index_up, index_down, index_front, index_back, index_coef;
index_x = blockIdx.x * blockDim.x + threadIdx.x;
index_y = blockIdx.y * blockDim.y + threadIdx.y;
index = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + index_x;
index_left = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + (index_x - 1);
index_right = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + (index_x + 1);
index_up = d_grid_StartPos + blockIdx.z * RRow * ZColumn + (index_y - 1) * ZColumn + index_x;
index_down = d_grid_StartPos + blockIdx.z * RRow * ZColumn + (index_y + 1) * ZColumn + index_x;
index_front = d_grid_StartPos + ((blockIdx.z - 1 + PhiSlice) % PhiSlice) * RRow * ZColumn + index_y * ZColumn + index_x;
index_back = d_grid_StartPos + ((blockIdx.z + 1) % PhiSlice) * RRow * ZColumn + index_y * ZColumn + index_x;
index_coef = d_coef_StartPos + index_y;
if (index_x != 0 && index_x < (ZColumn - 1) && index_y != 0 && index_y < (RRow - 1))
{
DeltaResidue[index] = d_ih2 * (coef2[index_coef] * VPotential[index_up] +
coef1[index_coef] * VPotential[index_down] +
d_tempRatioZ * (VPotential[index_left] + VPotential[index_right]) +
coef3[index_coef] * (VPotential[index_front] + VPotential[index_back]) -
icoef4[index_coef] * VPotential[index]) + RhoChargeDensity[index];
}
}
/// Restriksi dari finer grid ke coarser grid dengan operator Half Weighting
///
/// \f$ I_h^{2h} = \frac{1}{8} \begin{bmatrix}[ccc] 0 & 1 & 0 \\ 1 & 4 & 1\\ 0 & 1 & 0 \end{bmatrix}
///
/// \param RhoChargeDensity float* Array rapat arus
/// \param DeltaResidue float* Array residu hasil relaksasi
/// \param RRow const int Jumlah baris di arah sumbu \f$ r \f$
/// \param ZColumn const int Jumlah kolom di arah sumbu \f$ z \f$
/// \param PhiSlice const int Jumlah irisan di arah sumbu \f$ \phi \f$
__global__ void restriction2DHalf
(
float *RhoChargeDensity,
float *DeltaResidue,
const int RRow,
const int ZColumn,
const int PhiSlice
)
{
int index_x, index_y, index;
int finer_RRow, finer_ZColumn, finer_grid_StartPos;
int finer_index_x, finer_index_y, finer_index, finer_index_left, finer_index_right, finer_index_up, finer_index_down;
index_x = blockIdx.x * blockDim.x + threadIdx.x;
index_y = blockIdx.y * blockDim.y + threadIdx.y;
index = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + index_x;
finer_RRow = 2 * RRow - 1;
finer_ZColumn = 2 * ZColumn - 1;
finer_grid_StartPos = d_grid_StartPos - (finer_RRow * finer_ZColumn * PhiSlice);
finer_index_x = index_x * 2;
finer_index_y = index_y * 2;
finer_index = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + finer_index_y * finer_ZColumn + finer_index_x;
finer_index_left = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + finer_index_y * finer_ZColumn + (finer_index_x - 1);
finer_index_right = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + finer_index_y * finer_ZColumn + (finer_index_x + 1);
finer_index_up = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + (finer_index_y - 1) * finer_ZColumn + finer_index_x;
finer_index_down = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + (finer_index_y + 1) * finer_ZColumn + finer_index_x;
if (index_x != 0 && index_x < (ZColumn - 1) && index_y != 0 && index_y < (RRow - 1))
{
RhoChargeDensity[index] = 0.5 * DeltaResidue[finer_index] +
0.125 * (DeltaResidue[finer_index_left] + DeltaResidue[finer_index_right] + DeltaResidue[finer_index_up] + DeltaResidue[finer_index_down]);
}
}
/// Restriksi dari finer grid ke coarser grid dengan operator Full Weighting
///
/// \f$ I_h^{2h} = \frac{1}{16} \begin{bmatrix}[ccc] 1 & 2 & 1 \\ 2 & 4 & 2\\ 1 & 2 & 1 \end{bmatrix}
///
/// \param RhoChargeDensity float* Array rapat arus
/// \param DeltaResidue float* Array residu hasil relaksasi
/// \param RRow const int Jumlah baris di arah sumbu \f$ r \f$
/// \param ZColumn const int Jumlah kolom di arah sumbu \f$ z \f$
/// \param PhiSlice const int Jumlah irisan di arah sumbu \f$ \phi \f$
__global__ void restriction2DFull
(
float *RhoChargeDensity,
float *DeltaResidue,
const int RRow,
const int ZColumn,
const int PhiSlice
)
{
int index_x, index_y, index;
int finer_RRow, finer_ZColumn, finer_grid_StartPos;
int finer_index_x, finer_index_y, finer_index, finer_index_left, finer_index_right, finer_index_up, finer_index_down;
int finer_index_up_left, finer_index_up_right, finer_index_down_left, finer_index_down_right;
index_x = blockIdx.x * blockDim.x + threadIdx.x;
index_y = blockIdx.y * blockDim.y + threadIdx.y;
index = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + index_x;
finer_RRow = 2 * RRow - 1;
finer_ZColumn = 2 * ZColumn - 1;
finer_grid_StartPos = d_grid_StartPos - (finer_RRow * finer_ZColumn * PhiSlice);
finer_index_x = index_x * 2;
finer_index_y = index_y * 2;
finer_index = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + finer_index_y * finer_ZColumn + finer_index_x;
finer_index_left = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + finer_index_y * finer_ZColumn + (finer_index_x - 1);
finer_index_right = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + finer_index_y * finer_ZColumn + (finer_index_x + 1);
finer_index_up = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + (finer_index_y - 1) * finer_ZColumn + finer_index_x;
finer_index_down = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + (finer_index_y + 1) * finer_ZColumn + finer_index_x;
finer_index_up_left = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + (finer_index_y - 1) * finer_ZColumn + (finer_index_x - 1);
finer_index_up_right = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + (finer_index_y - 1) * finer_ZColumn + (finer_index_x + 1);
finer_index_down_left = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + (finer_index_y + 1) * finer_ZColumn + (finer_index_x - 1);
finer_index_down_right = finer_grid_StartPos + blockIdx.z * finer_RRow * finer_ZColumn + (finer_index_y + 1) * finer_ZColumn + (finer_index_x + 1);
if (index_x != 0 && index_x < (ZColumn - 1) && index_y != 0 && index_y < (RRow - 1))
{
RhoChargeDensity[index] = 0.25 * DeltaResidue[finer_index] +
0.125 * (DeltaResidue[finer_index_left] + DeltaResidue[finer_index_right] + DeltaResidue[finer_index_up] + DeltaResidue[finer_index_down]) +
0.0625 * (DeltaResidue[finer_index_up_left] + DeltaResidue[finer_index_up_right] + DeltaResidue[finer_index_down_left] + DeltaResidue[finer_index_down_right]);
} else {
RhoChargeDensity[index] = DeltaResidue[finer_index];
}
}
__global__ void zeroingVPotential
(
float *VPotential,
const int RRow,
const int ZColumn,
const int PhiSlice
)
{
int index_x, index_y, index;
index_x = blockIdx.x * blockDim.x + threadIdx.x;
index_y = blockIdx.y * blockDim.y + threadIdx.y;
index = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + index_x;
if (index_x != 0 && index_x < (ZColumn - 1) && index_y != 0 && index_y < (RRow - 1))
{
// zeroing V
VPotential[index] = 0;
}
if (index_x == ZColumn - 2) {
index_x++;
index = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + index_x;
VPotential[index] = 0;
}
}
__global__ void zeroingBoundaryTopBottom
(
float *VPotential,
int RRow,
int ZColumn,
int PhiSlice
)
{
int index_x, index_top, index_bottom;
index_x = blockIdx.x * blockDim.x + threadIdx.x;
index_top = d_grid_StartPos + blockIdx.z * RRow * ZColumn + 0 * ZColumn + index_x;
index_bottom = d_grid_StartPos + blockIdx.z * RRow * ZColumn + (ZColumn - 1) * ZColumn + index_x;
if (index_x < RRow)
{
VPotential[index_top] = 0.0;
VPotential[index_bottom] = 0.0;
}
}
__global__ void zeroingBoundaryLeftRight
(
float *VPotential,
int RRow,
int ZColumn,
int PhiSlice
)
{
int index_y, index_left, index_right;
index_y = blockIdx.x * blockDim.x + threadIdx.x;
index_left = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + 0;
index_right = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + (RRow - 1);
if (index_y < ZColumn)
{
VPotential[index_left] = 0.0;
VPotential[index_right] = 0.0;
}
}
__global__ void prolongation2DHalf
(
float *VPotential,
const int RRow,
const int ZColumn,
const int PhiSlice
)
{
int index_x, index_y, index;
int coarser_RRow = (RRow >> 1) + 1;
int coarser_ZColumn = (ZColumn >> 1) + 1;
int coarser_grid_StartPos = d_grid_StartPos + RRow * ZColumn * PhiSlice;
int coarser_index_self;
int coarser_index_up, coarser_index_down, coarser_index_left, coarser_index_right;
int coarser_index_up_left, coarser_index_up_right, coarser_index_down_left, coarser_index_down_right;
index_x = blockIdx.x * blockDim.x + threadIdx.x;
index_y = blockIdx.y * blockDim.y + threadIdx.y;
index = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + index_x;
if (index_x != 0 && index_x < (ZColumn - 1) && index_y != 0 && index_y < (RRow - 1))
{
// x odd, y odd
if ((index_x % 2 != 0) && (index_y % 2 != 0))
{
coarser_index_up_left = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2);
coarser_index_up_right = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2 + 1);
coarser_index_down_left = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2 + 1) * coarser_ZColumn + (index_x / 2);
coarser_index_down_right = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2 + 1) * coarser_ZColumn + (index_x / 2 + 1);
VPotential[index] += 0.25 * (VPotential[coarser_index_up_left] + VPotential[coarser_index_up_right] + VPotential[coarser_index_down_left] + VPotential[coarser_index_down_right]);
}
// x even, y odd
else if ((index_x % 2 == 0) && (index_y % 2 != 0))
{
coarser_index_up = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2);
coarser_index_down = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2 + 1) * coarser_ZColumn + (index_x / 2);
VPotential[index] += 0.5 * (VPotential[coarser_index_up] + VPotential[coarser_index_down]);
}
// x odd, y even
else if ((index_x % 2 != 0) && (index_y % 2 == 0))
{
coarser_index_left = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2);
coarser_index_right = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2 + 1);
VPotential[index] += 0.5 * (VPotential[coarser_index_left] + VPotential[coarser_index_right]);
}
// x even, y even
else
{
coarser_index_self = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2);
VPotential[index] += VPotential[coarser_index_self];
}
}
}
__global__ void prolongation2DHalfNoAdd
(
float *VPotential,
const int RRow,
const int ZColumn,
const int PhiSlice
)
{
int index_x, index_y, index;
int coarser_RRow = (RRow >> 1) + 1;
int coarser_ZColumn = (ZColumn >> 1) + 1;
int coarser_grid_StartPos = d_grid_StartPos + RRow * ZColumn * PhiSlice;
int coarser_index_self;
int coarser_index_up, coarser_index_down, coarser_index_left, coarser_index_right;
int coarser_index_up_left, coarser_index_up_right, coarser_index_down_left, coarser_index_down_right;
index_x = blockIdx.x * blockDim.x + threadIdx.x;
index_y = blockIdx.y * blockDim.y + threadIdx.y;
index = d_grid_StartPos + blockIdx.z * RRow * ZColumn + index_y * ZColumn + index_x;
if (index_x != 0 && index_x < (ZColumn - 1) && index_y != 0 && index_y < (RRow - 1))
{
// x odd, y odd
if ((index_x % 2 != 0) && (index_y % 2 != 0))
{
coarser_index_up_left = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2);
coarser_index_up_right = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2 + 1);
coarser_index_down_left = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2 + 1) * coarser_ZColumn + (index_x / 2);
coarser_index_down_right = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2 + 1) * coarser_ZColumn + (index_x / 2 + 1);
VPotential[index] = 0.25 * (VPotential[coarser_index_up_left] + VPotential[coarser_index_up_right] + VPotential[coarser_index_down_left] + VPotential[coarser_index_down_right]);
}
// x even, y odd
else if ((index_x % 2 == 0) && (index_y % 2 != 0))
{
coarser_index_up = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2);
coarser_index_down = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2 + 1) * coarser_ZColumn + (index_x / 2);
VPotential[index] = 0.5 * (VPotential[coarser_index_up] + VPotential[coarser_index_down]);
}
// x odd, y even
else if ((index_x % 2 != 0) && (index_y % 2 == 0))
{
coarser_index_left = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2);
coarser_index_right = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2 + 1);
VPotential[index] = 0.5 * (VPotential[coarser_index_left] + VPotential[coarser_index_right]);
}
// x even, y even
else
{
coarser_index_self = coarser_grid_StartPos + blockIdx.z * coarser_RRow * coarser_ZColumn + (index_y / 2) * coarser_ZColumn + (index_x / 2);
VPotential[index] = VPotential[coarser_index_self];
}
}
}
__global__ void errorCalculation
(
float *VPotentialPrev,
float *VPotential,
float *EpsilonError,
const int RRow,
const int ZColumn,
const int PhiSlice
)
{
int index_x, index_y, index;
float error;
float sum_error;
index_x = blockIdx.x * blockDim.x + threadIdx.x;
index_y = blockIdx.y * blockDim.y + threadIdx.y;
index = blockIdx.z * RRow * ZColumn + index_y * ZColumn + index_x;
if (index_x != 0 && index_x < (ZColumn - 1) && index_y != 0 && index_y < (RRow - 1))
{
error = VPotential[index] - VPotentialPrev[index];
sum_error = error * error;
__syncthreads();
atomicAdd( EpsilonError, sum_error );
}
}
/* GPU kernels end */
/* Error related functions start */
float GetErrorNorm2
(
float * VPotential,
float * VPotentialPrev,
const int rows,
const int cols,
float weight
)
{
float error = 0.0;
float sum_error = 0.0;
for (int i=0;i<rows;i++)
for (int j=0;j <cols;j++)
{
error = (VPotential[i * cols + j] - VPotentialPrev[i * cols + j]) / weight;
sum_error += (error * error);
}
return sum_error / (rows * cols);
}
float GetAbsMax
(
float *VPotentialExact,
int size
)
{
float mymax = 0.0;
for (int i=0;i< size;i++)
if (abs(VPotentialExact[i]) > mymax) mymax = abs(VPotentialExact[i]);
return mymax;
}
/* Error related functions end */
/* Restrict Boundary for FCycle start -- just CPU enough */
void Restrict_Boundary
(
float *VPotential,
const int RRow,
const int ZColumn,
const int PhiSlice,
const int Offset
)
{
int i,ii,j,jj;
int finer_RRow = 2 * RRow - 1;
int finer_ZColumn = 2 * ZColumn - 1;
int finer_Offset = Offset - (finer_RRow * finer_ZColumn * PhiSlice);
int sliceStart;
int finer_SliceStart;
//printf("(%d,%d,%d) -> (%d,%d,%d)\n",RRow,ZColumn,Offset,finer_RRow,finer_ZColumn,finer_Offset);
// do for each slice
for ( int m = 0;m < PhiSlice;m++)
{
sliceStart = m * (RRow * ZColumn);
finer_SliceStart = m * (finer_RRow * finer_ZColumn);
// copy boundary
for ( j = 0, jj =0; j < ZColumn ; j++,jj+=2)
{
VPotential[Offset + sliceStart + (0 * ZColumn) + j] =
VPotential[finer_Offset + finer_SliceStart + (0 * finer_ZColumn) + jj];
VPotential[Offset + sliceStart + ((RRow - 1) * ZColumn) + j] =
VPotential[finer_Offset + finer_SliceStart + ((finer_RRow -1) * finer_ZColumn) + jj];
}
for ( i = 0, ii =0; i < RRow ; i++,ii+=2) {
VPotential[Offset + sliceStart + (i * ZColumn)] =
VPotential[finer_Offset + finer_SliceStart + (ii * finer_ZColumn)];
VPotential[Offset + sliceStart + (i * ZColumn) + (ZColumn - 1)] =
VPotential[finer_Offset + finer_SliceStart + (ii * finer_ZColumn) + (finer_ZColumn - 1)];
}
}
/**
// top left (0,0)
// boundary in top and down
for ( j = 1, jj =2; j < ZColumn-1 ; j++,jj+=2)
{
VPotential[Offset + sliceStart + (0 * ZColumn) + j] =
0.5 * VPotential[finer_Offset + finer_SliceStart + (0 * finer_ZColumn) + jj] +
0.25 * VPotential[finer_Offset + finer_SliceStart + (0 * finer_ZColumn) + jj - 1] +
0.25 * VPotential[finer_Offset + finer_SliceStart + (0 * finer_ZColumn) + jj + 1];
VPotential[Offset + sliceStart + ((RRow - 1) * ZColumn) + j] =
0.5 * VPotential[finer_Offset + finer_SliceStart + ((finer_RRow -1) * finer_ZColumn) + jj] +
0.25 * VPotential[finer_Offset + finer_SliceStart + ((finer_RRow -1) * finer_ZColumn) + jj - 1] +
0.25 * VPotential[finer_Offset + finer_SliceStart + ((finer_RRow -1) * finer_ZColumn) + jj + 1];
}
// boundary in left and right
for ( i = 1, ii =2; i < RRow - 1 ; i++,ii+=2) {
VPotential[Offset + sliceStart + (i * ZColumn)] =
0.5 * VPotential[finer_Offset + finer_SliceStart + (ii * finer_ZColumn)] +
0.25 * VPotential[finer_Offset + finer_SliceStart + ((ii-1) * finer_ZColumn)] +
0.25 * VPotential[finer_Offset + finer_SliceStart + ((ii + 1) * finer_ZColumn)];
VPotential[Offset + sliceStart + (i * ZColumn) + (ZColumn - 1)] =
0.5 * VPotential[finer_Offset + finer_SliceStart + (ii * finer_ZColumn) + jj + (finer_ZColumn - 1)] +
0.25 * VPotential[finer_Offset + finer_SliceStart + ((ii -1) * finer_ZColumn) + (finer_ZColumn - 1)] +
0.25 * VPotential[finer_Offset + finer_SliceStart + ((ii +1) * finer_ZColumn) + (finer_ZColumn - 1)];
}
// top left (0,0)
VPotential[Offset + sliceStart + (0 * ZColumn) + 0] =
0.5 * VPotential[finer_Offset + finer_SliceStart] +
0.25 * VPotential[finer_Offset + finer_SliceStart + (0 * finer_ZColumn) + 1] +
0.25 * VPotential[finer_Offset + finer_SliceStart + (1 * finer_ZColumn)];
// top right
VPotential[Offset + sliceStart + (0 * ZColumn) + (ZColumn - 1) ] =
0.5 * VPotential[finer_Offset + finer_SliceStart + (0 * finer_ZColumn) + (finer_ZColumn -1) ] +
0.25 * VPotential[finer_Offset + finer_SliceStart + (0 * finer_ZColumn) + (finer_ZColumn - 2)] +
0.25 * VPotential[finer_Offset + finer_SliceStart + (1 * finer_ZColumn) + (finer_ZColumn - 1)];
// bottom left
VPotential[Offset + sliceStart + ((RRow - 1) * ZColumn) + 0] =
0.5 * VPotential[finer_Offset + finer_SliceStart + ((finer_RRow - 1) * finer_ZColumn) + 0] +
0.25 * VPotential[finer_Offset + finer_SliceStart + ((finer_RRow - 1) * finer_ZColumn) + 1] +
0.25 * VPotential[finer_Offset + finer_SliceStart + ((finer_RRow - 2) * finer_ZColumn) + 0];
// bottom right
VPotential[Offset + sliceStart + ((RRow - 1) * ZColumn) + (ZColumn - 1)] =
0.5 * VPotential[finer_Offset + finer_SliceStart + ((finer_RRow - 1) * finer_ZColumn) + (finer_ZColumn - 1)] +
0.25 * VPotential[finer_Offset + finer_SliceStart + ((finer_RRow - 1) * finer_ZColumn) + (finer_ZColumn - 2)] +
0.25 * VPotential[finer_Offset + finer_SliceStart + ((finer_RRow - 2) * finer_ZColumn) + (finer_ZColumn - 1)];
}
**/
}
/* Restrict Boundary for FCycle end */
/** Print matrix **/
void PrintMatrix
(
float *Mat,
const int Row,
const int Column
)
{
printf("Matrix (%d,%d)\n",Row,Column);
for (int i=0;i<Row;i++)
{
for (int j=0;j<Column;j++)
{
printf("%11.4g ",Mat[i*Column + j]);
}
printf("\n");
}
}
/* Cycle functions start */
void VCycleSemiCoarseningGPU
(
float *d_VPotential,
float *d_RhoChargeDensity,
float *d_DeltaResidue,
float *d_coef1,
float *d_coef2,
float *d_coef3,
float *d_coef4,
float *d_icoef4,
float gridSizeR,
float ratioZ,
float ratioPhi,
int RRow,
int ZColumn,
int PhiSlice,
int gridFrom,
int gridTo,
int nPre,
int nPost
)
{
int grid_RRow;
int grid_ZColumn;
int grid_PhiSlice = PhiSlice;
int grid_StartPos;
int coef_StartPos;
int iOne, jOne;
float h, h2, ih2;
float tempRatioZ;
float tempRatioPhi;
float radius;
// V-Cycle => Finest Grid
iOne = 1 << (gridFrom - 1);
jOne = 1 << (gridFrom - 1);
//grid_RRow = ((RRow - 1) / iOne) + 1;
//grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
// change accordingly to gridFrom
grid_StartPos = 0;
coef_StartPos = 0;
for (int step = 1; step < gridFrom; step++)
{
grid_RRow = ((RRow - 1) / (1 << (step - 1))) + 1;
grid_ZColumn = ((ZColumn - 1) / (1 << (step - 1))) + 1;
grid_StartPos += grid_RRow * grid_ZColumn * grid_PhiSlice;
coef_StartPos += grid_RRow;
}
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
// pre-compute constant memory
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// red-black gauss seidel relaxation (nPre times)
for (int i = 0; i < nPre; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
}
// residue calculation
residueCalculation<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_icoef4 );
// V-Cycle => from finer to coarsest grid
for (int step = gridFrom + 1; step <= gridTo; step++)
{
iOne = 1 << (step - 1);
jOne = 1 << (step - 1);
grid_StartPos += grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos += grid_RRow;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
// pre-compute constant memory
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// restriction
restriction2DFull<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice );
// zeroing V
zeroingVPotential<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
// zeroing boundaries
dim3 grid_BlockPerGridTopBottom((grid_RRow < 16) ? 1 : ((grid_RRow / 16) + 1), 1, PhiSlice);
dim3 grid_BlockPerGridLeftRight((grid_ZColumn < 16) ? 1 : ((grid_ZColumn / 16) + 1), 1, PhiSlice);
dim3 grid_ThreadPerBlockBoundary(16);
zeroingBoundaryTopBottom<<< grid_BlockPerGridTopBottom, grid_ThreadPerBlockBoundary >>>( d_VPotential, grid_RRow, grid_ZColumn, PhiSlice );
zeroingBoundaryLeftRight<<< grid_BlockPerGridLeftRight, grid_ThreadPerBlockBoundary >>>( d_VPotential, grid_RRow, grid_ZColumn, PhiSlice );
// red-black gauss seidel relaxation (nPre times)
for (int i = 0; i < nPre; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
}
// residue calculation
if (step < gridTo)
{
residueCalculation<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_icoef4 );
}
}
// V-Cycle => from coarser to finer grid
for (int step = (gridTo - 1); step >= gridFrom; step--)
{
iOne = iOne / 2;
jOne = jOne / 2;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
grid_StartPos -= grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos -= grid_RRow;
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// prolongation
prolongation2DHalf<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
// red-black gauss seidel relaxation (nPost times)
for (int i = 0; i < nPost; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
}
}
}
/* Cycle functions end */
/*extern function */
extern "C" void PoissonMultigrid3DSemiCoarseningGPUError
(
float *VPotential,
float *RhoChargeDensity,
const int RRow,
const int ZColumn,
const int PhiSlice,
const int Symmetry,
float *fparam,
int *iparam,
bool isExactPresent,
float *errorConv,
float *errorExact,
float *VPotentialExact //allocation in the client
)
{
// variables for CPU memory
float *temp_VPotential;
float *VPotentialPrev;
float *EpsilonError;
// variables for GPU memory
float *d_VPotential;
float *d_RhoChargeDensity;
float *d_DeltaResidue;
float *d_VPotentialPrev;
float *d_EpsilonError;
float *d_coef1;
float *d_coef2;
float *d_coef3;
float *d_coef4;
float *d_icoef4;
// variables for coefficent calculations
float *coef1;
float *coef2;
float *coef3;
float *coef4;
float *icoef4;
float tempRatioZ;
float tempRatioPhi;
float radius;
int gridFrom;
int gridTo;
int loops;
// variables passed from ALIROOT
float gridSizeR = fparam[0];
float gridSizePhi = fparam[1];
float gridSizeZ = fparam[2];
float ratioPhi = fparam[3];
float ratioZ = fparam[4];
float convErr = fparam[5];
float IFCRadius = fparam[6];
int nPre = iparam[0];
int nPost = iparam[1];
int maxLoop = iparam[2];
int nCycle = iparam[3];
// variables for calculating GPU memory allocation
int grid_RRow;
int grid_ZColumn;
int grid_PhiSlice = PhiSlice;
int grid_Size = 0;
int grid_StartPos;
int coef_Size = 0;
int coef_StartPos;
int iOne, jOne;
float h, h2, ih2;
// variables for calculating multigrid maximum depth
int depth_RRow = 0;
int depth_ZColumn = 0;
int temp_RRow = RRow;
int temp_ZColumn = ZColumn;
// calculate depth for multigrid
while (temp_RRow >>= 1) depth_RRow++;
while (temp_ZColumn >>= 1) depth_ZColumn++;
loops = (depth_RRow > depth_ZColumn) ? depth_ZColumn : depth_RRow;
loops = (loops > maxLoop) ? maxLoop : loops;
gridFrom = 1;
gridTo = loops;
// calculate GPU memory allocation for multigrid
for (int step = gridFrom; step <= gridTo; step++)
{
grid_RRow = ((RRow - 1) / (1 << (step - 1))) + 1;
grid_ZColumn = ((ZColumn - 1) / (1 << (step - 1))) + 1;
grid_Size += grid_RRow * grid_ZColumn * grid_PhiSlice;
coef_Size += grid_RRow;
}
// allocate memory for temporary output
temp_VPotential = (float *) malloc(grid_Size * sizeof(float));
VPotentialPrev = (float *) malloc(RRow * ZColumn * PhiSlice * sizeof(float));
EpsilonError = (float *) malloc(1 * sizeof(float));
// allocate memory for relaxation coefficient
coef1 = (float *) malloc(coef_Size * sizeof(float));
coef2 = (float *) malloc(coef_Size * sizeof(float));
coef3 = (float *) malloc(coef_Size * sizeof(float));
coef4 = (float *) malloc(coef_Size * sizeof(float));
icoef4 = (float *) malloc(coef_Size * sizeof(float));
// pre-compute relaxation coefficient
coef_StartPos = 0;
iOne = 1 << (gridFrom - 1);
jOne = 1 << (gridFrom - 1);
for (int step = gridFrom; step <= gridTo; step++)
{
grid_RRow = ((RRow - 1) / iOne) + 1;
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
for (int i = 1; i < grid_RRow - 1; i++)
{
radius = IFCRadius + i * h;
coef1[coef_StartPos + i] = 1.0 + h / (2 * radius);
coef2[coef_StartPos + i] = 1.0 - h / (2 * radius);
coef3[coef_StartPos + i] = tempRatioPhi / (radius * radius);
coef4[coef_StartPos + i] = 0.5 / (1.0 + tempRatioZ + coef3[coef_StartPos + i]);
icoef4[coef_StartPos + i] = 1.0 / coef4[coef_StartPos + i];
}
coef_StartPos += grid_RRow;
iOne = 2 * iOne;
jOne = 2 * jOne;
}
// device memory allocation
cudaMalloc( &d_VPotential, grid_Size * sizeof(float) );
cudaMalloc( &d_VPotentialPrev, RRow * ZColumn * PhiSlice * sizeof(float) );
cudaMalloc( &d_EpsilonError, 1 * sizeof(float) );
cudaMalloc( &d_DeltaResidue, grid_Size * sizeof(float) );
cudaMalloc( &d_RhoChargeDensity, grid_Size * sizeof(float) );
cudaMalloc( &d_coef1, coef_Size * sizeof(float) );
cudaMalloc( &d_coef2, coef_Size * sizeof(float) );
cudaMalloc( &d_coef3, coef_Size * sizeof(float) );
cudaMalloc( &d_coef4, coef_Size * sizeof(float) );
cudaMalloc( &d_icoef4, coef_Size * sizeof(float) );
// set memory to zero
cudaMemset( d_VPotential, 0, grid_Size * sizeof(float) );
cudaMemset( d_DeltaResidue, 0, grid_Size * sizeof(float) );
cudaMemset( d_RhoChargeDensity, 0, grid_Size * sizeof(float) );
cudaMemset( d_VPotentialPrev, 0, RRow * ZColumn * PhiSlice * sizeof(float) );
cudaMemset( d_EpsilonError, 0, 1 * sizeof(float) );
// copy data from host to device
cudaMemcpy( d_VPotential, VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyHostToDevice ); //check
cudaMemcpy( d_RhoChargeDensity, RhoChargeDensity, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyHostToDevice ); //check
cudaMemcpy( d_coef1, coef1, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_coef2, coef2, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_coef3, coef3, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_coef4, coef4, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_icoef4, icoef4, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_VPotentialPrev, VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyHostToDevice );
// max exact
// float maxAbsExact = GetAbsMax(VPotentialExact, RRow * PhiSlice * ZColumn);
float maxAbsExact = 1.0;
if (isExactPresent == true)
maxAbsExact = GetAbsMax(VPotentialExact, RRow * PhiSlice * ZColumn);
dim3 error_BlockPerGrid((RRow < 16) ? 1 : (RRow / 16), (ZColumn < 16) ? 1 : (ZColumn / 16), PhiSlice);
dim3 error_ThreadPerBlock(16, 16);
for (int cycle = 0; cycle < nCycle; cycle++)
{
cudaMemcpy( temp_VPotential, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
if (isExactPresent == true) errorExact[cycle] = GetErrorNorm2(temp_VPotential, VPotentialExact, RRow * PhiSlice,ZColumn, maxAbsExact);
VCycleSemiCoarseningGPU(d_VPotential, d_RhoChargeDensity, d_DeltaResidue, d_coef1, d_coef2, d_coef3, d_coef4, d_icoef4, gridSizeR, ratioZ, ratioPhi, RRow, ZColumn, PhiSlice, gridFrom, gridTo, nPre, nPost);
errorCalculation<<< error_BlockPerGrid, error_ThreadPerBlock >>> ( d_VPotentialPrev, d_VPotential, d_EpsilonError, RRow, ZColumn, PhiSlice);
cudaMemcpy( EpsilonError, d_EpsilonError, 1 * sizeof(float), cudaMemcpyDeviceToHost );
errorConv[cycle] = *EpsilonError / (RRow * ZColumn * PhiSlice);
if (errorConv[cycle] < convErr)
{
//errorConv
nCycle = cycle;
break;
}
cudaMemcpy( d_VPotentialPrev, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToDevice );
cudaMemset( d_EpsilonError, 0, 1 * sizeof(float) );
}
iparam[3] = nCycle;
// for (int cycle = 0; cycle < nCycle; cycle++)
// {
// cudaMemcpy( temp_VPotentialPrev, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
// VCycleSemiCoarseningGPU(d_VPotential, d_RhoChargeDensity, d_DeltaResidue, d_coef1, d_coef2, d_coef3, d_coef4, d_icoef4, gridSizeR, ratioZ, ratioPhi, RRow, ZColumn, PhiSlice, gridFrom, gridTo, nPre, nPost);
// cudaMemcpy( temp_VPotential, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
// errorConv[cycle] = GetErrorNorm2(temp_VPotential, temp_VPotentialPrev, RRow * PhiSlice, ZColumn, 1.0);
// //errorExact[cycle] = GetErrorNorm2(temp_VPotential, VPotentialExact, RRow * PhiSlice,ZColumn, 1.0);
// }
// copy result from device to host
cudaMemcpy( temp_VPotential, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
memcpy(VPotential, temp_VPotential, RRow * ZColumn * PhiSlice * sizeof(float));
// free device memory
cudaFree( d_VPotential );
cudaFree( d_DeltaResidue );
cudaFree( d_RhoChargeDensity );
cudaFree( d_VPotentialPrev );
cudaFree( d_EpsilonError );
cudaFree( d_coef1 );
cudaFree( d_coef2 );
cudaFree( d_coef3 );
cudaFree( d_coef4 );
cudaFree( d_icoef4 );
// free host memory
free( coef1 );
free( coef2 );
free( coef3 );
free( coef4 );
free( icoef4 );
free( temp_VPotential );
free( VPotentialPrev );
}
extern "C" void PoissonMultigrid3DSemiCoarseningGPUErrorWCycle
(
float *VPotential,
float *RhoChargeDensity,
const int RRow,
const int ZColumn,
const int PhiSlice,
const int Symmetry,
float *fparam,
int *iparam,
float *errorConv,
float *errorExact,
float *VPotentialExact //allocation in the client
)
{
// variables for CPU memory
float *temp_VPotential;
float *VPotentialPrev;
float *EpsilonError;
// variables for GPU memory
float *d_VPotential;
float *d_RhoChargeDensity;
float *d_DeltaResidue;
float *d_coef1;
float *d_coef2;
float *d_coef3;
float *d_coef4;
float *d_icoef4;
float *d_VPotentialPrev;
float *d_EpsilonError;
// variables for coefficent calculations
float *coef1;
float *coef2;
float *coef3;
float *coef4;
float *icoef4;
float tempRatioZ;
float tempRatioPhi;
float radius;
int gridFrom;
int gridTo;
int loops;
// variables passed from ALIROOT
float gridSizeR = fparam[0];
//float gridSizePhi = fparam[1];
//float gridSizeZ = fparam[2];
float ratioPhi = fparam[3];
float ratioZ = fparam[4];
float convErr = fparam[5];
float IFCRadius = fparam[6];
int nPre = iparam[0];
int nPost = iparam[1];
int maxLoop = iparam[2];
int nCycle = iparam[3];
// variables for calculating GPU memory allocation
int grid_RRow;
int grid_ZColumn;
int grid_PhiSlice = PhiSlice;
int grid_Size = 0;
int grid_StartPos;
int coef_Size = 0;
int coef_StartPos;
int iOne, jOne;
float h, h2, ih2;
// variables for calculating multigrid maximum depth
int depth_RRow = 0;
int depth_ZColumn = 0;
int temp_RRow = RRow;
int temp_ZColumn = ZColumn;
// calculate depth for multigrid
while (temp_RRow >>= 1) depth_RRow++;
while (temp_ZColumn >>= 1) depth_ZColumn++;
loops = (depth_RRow > depth_ZColumn) ? depth_ZColumn : depth_RRow;
loops = (loops > maxLoop) ? maxLoop : loops;
gridFrom = 1;
gridTo = loops;
// calculate GPU memory allocation for multigrid
for (int step = gridFrom; step <= gridTo; step++)
{
grid_RRow = ((RRow - 1) / (1 << (step - 1))) + 1;
grid_ZColumn = ((ZColumn - 1) / (1 << (step - 1))) + 1;
grid_Size += grid_RRow * grid_ZColumn * grid_PhiSlice;
coef_Size += grid_RRow;
}
// allocate memory for temporary output
temp_VPotential = (float *) malloc(grid_Size * sizeof(float));
VPotentialPrev = (float *) malloc(RRow * ZColumn * PhiSlice * sizeof(float));
EpsilonError = (float *) malloc(1 * sizeof(float));
// allocate memory for relaxation coefficient
coef1 = (float *) malloc(coef_Size * sizeof(float));
coef2 = (float *) malloc(coef_Size * sizeof(float));
coef3 = (float *) malloc(coef_Size * sizeof(float));
coef4 = (float *) malloc(coef_Size * sizeof(float));
icoef4 = (float *) malloc(coef_Size * sizeof(float));
// pre-compute relaxation coefficient
coef_StartPos = 0;
iOne = 1 << (gridFrom - 1);
jOne = 1 << (gridFrom - 1);
for (int step = gridFrom; step <= gridTo; step++)
{
grid_RRow = ((RRow - 1) / iOne) + 1;
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
for (int i = 1; i < grid_RRow - 1; i++)
{
radius = IFCRadius + i * h;
coef1[coef_StartPos + i] = 1.0 + h / (2 * radius);
coef2[coef_StartPos + i] = 1.0 - h / (2 * radius);
coef3[coef_StartPos + i] = tempRatioPhi / (radius * radius);
coef4[coef_StartPos + i] = 0.5 / (1.0 + tempRatioZ + coef3[coef_StartPos + i]);
icoef4[coef_StartPos + i] = 1.0 / coef4[coef_StartPos + i];
}
coef_StartPos += grid_RRow;
iOne = 2 * iOne;
jOne = 2 * jOne;
}
// device memory allocation
cudaMalloc( &d_VPotential, grid_Size * sizeof(float) );
cudaMalloc( &d_DeltaResidue, grid_Size * sizeof(float) );
cudaMalloc( &d_VPotentialPrev, RRow * ZColumn * PhiSlice * sizeof(float) );
cudaMalloc( &d_EpsilonError, 1 * sizeof(float) );
cudaMalloc( &d_RhoChargeDensity, grid_Size * sizeof(float) );
cudaMalloc( &d_coef1, coef_Size * sizeof(float) );
cudaMalloc( &d_coef2, coef_Size * sizeof(float) );
cudaMalloc( &d_coef3, coef_Size * sizeof(float) );
cudaMalloc( &d_coef4, coef_Size * sizeof(float) );
cudaMalloc( &d_icoef4, coef_Size * sizeof(float) );
// set memory to zero
cudaMemset( d_VPotential, 0, grid_Size * sizeof(float) );
cudaMemset( d_DeltaResidue, 0, grid_Size * sizeof(float) );
cudaMemset( d_RhoChargeDensity, 0, grid_Size * sizeof(float) );
cudaMemset( d_VPotentialPrev, 0, RRow * ZColumn * PhiSlice * sizeof(float) );
cudaMemset( d_EpsilonError, 0, 1 * sizeof(float) );
// copy data from host to device
cudaMemcpy( d_VPotential, VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyHostToDevice ); //check
cudaMemcpy( d_RhoChargeDensity, RhoChargeDensity, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyHostToDevice ); //check
cudaMemcpy( d_coef1, coef1, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_coef2, coef2, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_coef3, coef3, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_coef4, coef4, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_icoef4, icoef4, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_VPotentialPrev, VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyHostToDevice );
// max exact float maxAbsExact = GetAbsMax(VPotentialExact,RRow * PhiSlice * ZColumn);
float maxAbsExact = GetAbsMax(VPotentialExact, RRow * PhiSlice * ZColumn);
dim3 error_BlockPerGrid((RRow < 16) ? 1 : (RRow / 16), (ZColumn < 16) ? 1 : (ZColumn / 16), PhiSlice);
dim3 error_ThreadPerBlock(16, 16);
for (int cycle = 0; cycle < nCycle; cycle++)
{
/*V-Cycle starts*/
// error conv
// cudaMemcpy( temp_VPotentialPrev, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
cudaMemcpy( temp_VPotential, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
errorExact[cycle] = GetErrorNorm2(temp_VPotential,VPotentialExact,RRow * PhiSlice,ZColumn,maxAbsExact);
// V-Cycle => Finest Grid
iOne = 1 << (gridFrom - 1);
jOne = 1 << (gridFrom - 1);
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
grid_StartPos = 0;
coef_StartPos = 0;
// pre-compute constant memory
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// red-black gauss seidel relaxation (nPre times)
for (int i = 0; i < nPre; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
//cudaDeviceSynchronize();
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
//cudaDeviceSynchronize();
}
// residue calculation
residueCalculation<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_icoef4 );
//cudaDeviceSynchronize();
// V-Cycle => from finer to coarsest grid
for (int step = gridFrom + 1; step <= gridTo; step++)
{
iOne = 1 << (step - 1);
jOne = 1 << (step - 1);
grid_StartPos += grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos += grid_RRow;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
// pre-compute constant memory
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// restriction
restriction2DFull<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice );
//cudaDeviceSynchronize();
// zeroing V
zeroingVPotential<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
//cudaDeviceSynchronize();
// red-black gauss seidel relaxation (nPre times)
for (int i = 0; i < nPre; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
//cudaDeviceSynchronize();
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
//cudaDeviceSynchronize();
}
// residue calculation
if (step < gridTo)
{
residueCalculation<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_icoef4 );
//cudaDeviceSynchronize();
}
}
/////////// innner w cycle
/// up one down one
// up one
{
int step = (gridTo - 1);
iOne = iOne / 2;
jOne = jOne / 2;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
grid_StartPos -= grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos -= grid_RRow;
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// prolongation
prolongation2DHalf<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
// cudaDeviceSynchronize();
// red-black gauss seidel relaxation (nPost times)
for (int i = 0; i < nPost; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
// cudaDeviceSynchronize();
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
// cudaDeviceSynchronize();
}
}
// down one
{
residueCalculation<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_icoef4 );
iOne = iOne * 2;
jOne = jOne * 2;
grid_StartPos += grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos += grid_RRow;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
// pre-compute constant memory
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// restriction
restriction2DFull<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice );
//cudaDeviceSynchronize();
// zeroing V
zeroingVPotential<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
//cudaDeviceSynchronize();
// red-black gauss seidel relaxation (nPre times)
for (int i = 0; i < nPre; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
//cudaDeviceSynchronize();
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
//cudaDeviceSynchronize();
}
}
/// end up one down on
/// up two down two
// up two from gridTo - 1, to gridTo -3
for (int step = (gridTo - 1); step >= gridTo - 3; step--)
{
iOne = iOne / 2;
jOne = jOne / 2;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
grid_StartPos -= grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos -= grid_RRow;
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// prolongation
prolongation2DHalf<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
// cudaDeviceSynchronize();
// red-black gauss seidel relaxation (nPost times)
for (int i = 0; i < nPost; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
// cudaDeviceSynchronize();
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
// cudaDeviceSynchronize();
}
}
// down to from gridTo - 1, to gridTo -3
for (int step = gridTo - 3; step <= gridTo - 1; step++)
{
iOne = iOne * 2;
jOne = jOne * 2;
grid_StartPos += grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos += grid_RRow;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
// pre-compute constant memory
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// restriction
restriction2DFull<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice );
//cudaDeviceSynchronize();
// zeroing V
zeroingVPotential<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
//cudaDeviceSynchronize();
// red-black gauss seidel relaxation (nPre times)
for (int i = 0; i < nPre; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
//cudaDeviceSynchronize();
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
//cudaDeviceSynchronize();
}
// residue calculation
if (step < gridTo)
{
residueCalculation<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_icoef4 );
//cudaDeviceSynchronize();
}
}
/// up one down one
{
int step = (gridTo - 1);
iOne = iOne / 2;
jOne = jOne / 2;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
grid_StartPos -= grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos -= grid_RRow;
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// prolongation
prolongation2DHalf<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
// cudaDeviceSynchronize();
// red-black gauss seidel relaxation (nPost times)
for (int i = 0; i < nPost; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
// cudaDeviceSynchronize();
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
// cudaDeviceSynchronize();
}
}
// down one
{
residueCalculation<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_icoef4 );
iOne = iOne * 2;
jOne = jOne * 2;
grid_StartPos += grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos += grid_RRow;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
// pre-compute constant memory
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// restriction
restriction2DFull<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_RhoChargeDensity, d_DeltaResidue, grid_RRow, grid_ZColumn, grid_PhiSlice );
//cudaDeviceSynchronize();
// zeroing V
zeroingVPotential<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
//cudaDeviceSynchronize();
// red-black gauss seidel relaxation (nPre times)
for (int i = 0; i < nPre; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
//cudaDeviceSynchronize();
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
//cudaDeviceSynchronize();
}
}
/// end up one down one
/////////// end inner w cyle
// V-Cycle => from coarser to finer grid
for (int step = (gridTo - 1); step >= gridFrom; step--)
{
iOne = iOne / 2;
jOne = jOne / 2;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
grid_StartPos -= grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos -= grid_RRow;
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// prolongation
prolongation2DHalf<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
// cudaDeviceSynchronize();
// red-black gauss seidel relaxation (nPost times)
for (int i = 0; i < nPost; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
// cudaDeviceSynchronize();
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
// cudaDeviceSynchronize();
}
}
/*V-Cycle ends*/
errorCalculation<<< error_BlockPerGrid, error_ThreadPerBlock >>> ( d_VPotentialPrev, d_VPotential, d_EpsilonError, RRow, ZColumn, PhiSlice);
cudaMemcpy( EpsilonError, d_EpsilonError, 1 * sizeof(float), cudaMemcpyDeviceToHost );
errorConv[cycle] = *EpsilonError / (RRow * ZColumn * PhiSlice);
if (errorConv[cycle] < convErr)
{
//errorConv
nCycle = cycle;
iparam[3] = nCycle;
break;
}
cudaMemcpy( d_VPotentialPrev, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToDevice );
cudaMemset( d_EpsilonError, 0, 1 * sizeof(float) );
}
cudaDeviceSynchronize();
// copy result from device to host
cudaMemcpy( temp_VPotential, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
memcpy(VPotential, temp_VPotential, RRow * ZColumn * PhiSlice * sizeof(float));
// free device memory
cudaFree( d_VPotential );
cudaFree( d_VPotentialPrev );
cudaFree( d_EpsilonError );
cudaFree( d_DeltaResidue );
cudaFree( d_RhoChargeDensity );
cudaFree( d_coef1 );
cudaFree( d_coef2 );
cudaFree( d_coef3 );
cudaFree( d_coef4 );
cudaFree( d_icoef4 );
// free host memory
free( coef1 );
free( coef2 );
free( coef3 );
free( coef4 );
free( icoef4 );
free( temp_VPotential );
//free( temp_VPotentialPrev );
}
/*extern function */
extern "C" void PoissonMultigrid3DSemiCoarseningGPUErrorFCycle
(
float *VPotential,
float *RhoChargeDensity,
const int RRow,
const int ZColumn,
const int PhiSlice,
const int Symmetry,
float *fparam,
int *iparam,
bool isExactPresent,
float *errorConv,
float *errorExact,
float *VPotentialExact //allocation in the client
)
{
// variables for CPU memory
float *temp_VPotential;
float *VPotentialPrev;
float *EpsilonError;
// variables for GPU memory
float *d_VPotential;
float *d_RhoChargeDensity;
float *d_DeltaResidue;
float *d_coef1;
float *d_coef2;
float *d_coef3;
float *d_coef4;
float *d_icoef4;
float *d_VPotentialPrev;
float *d_EpsilonError;
// variables for coefficent calculations
float *coef1;
float *coef2;
float *coef3;
float *coef4;
float *icoef4;
float tempRatioZ;
float tempRatioPhi;
float radius;
int gridFrom;
int gridTo;
int loops;
// variables passed from ALIROOT
float gridSizeR = fparam[0];
//float gridSizePhi = fparam[1];
//float gridSizeZ = fparam[2];
float ratioPhi = fparam[3];
float ratioZ = fparam[4];
float convErr = fparam[5];
float IFCRadius = fparam[6];
int nPre = iparam[0];
int nPost = iparam[1];
int maxLoop = iparam[2];
int nCycle = iparam[3];
// variables for calculating GPU memory allocation
int grid_RRow;
int grid_ZColumn;
int grid_PhiSlice = PhiSlice;
int grid_Size = 0;
int grid_StartPos;
int coef_Size = 0;
int coef_StartPos;
int iOne, jOne;
float h, h2, ih2;
// variables for calculating multigrid maximum depth
int depth_RRow = 0;
int depth_ZColumn = 0;
int temp_RRow = RRow;
int temp_ZColumn = ZColumn;
// calculate depth for multigrid
while (temp_RRow >>= 1) depth_RRow++;
while (temp_ZColumn >>= 1) depth_ZColumn++;
loops = (depth_RRow > depth_ZColumn) ? depth_ZColumn : depth_RRow;
loops = (loops > maxLoop) ? maxLoop : loops;
gridFrom = 1;
gridTo = loops;
// calculate GPU memory allocation for multigrid
for (int step = gridFrom; step <= gridTo; step++)
{
grid_RRow = ((RRow - 1) / (1 << (step - 1))) + 1;
grid_ZColumn = ((ZColumn - 1) / (1 << (step - 1))) + 1;
grid_Size += grid_RRow * grid_ZColumn * grid_PhiSlice;
coef_Size += grid_RRow;
}
// allocate memory for temporary output
temp_VPotential = (float *) malloc(grid_Size * sizeof(float));
VPotentialPrev = (float *) malloc(grid_Size * sizeof(float));
EpsilonError = (float *) malloc(1 * sizeof(float));
for (int i=0;i<grid_Size;i++) temp_VPotential[i] = 0.0;
// allocate memory for relaxation coefficient
coef1 = (float *) malloc(coef_Size * sizeof(float));
coef2 = (float *) malloc(coef_Size * sizeof(float));
coef3 = (float *) malloc(coef_Size * sizeof(float));
coef4 = (float *) malloc(coef_Size * sizeof(float));
icoef4 = (float *) malloc(coef_Size * sizeof(float));
// pre-compute relaxation coefficient
// restrict boundary
coef_StartPos = 0;
grid_StartPos = 0;
iOne = 1 << (gridFrom - 1);
jOne = 1 << (gridFrom - 1);
for (int step = gridFrom; step <= gridTo; step++)
{
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / iOne) + 1;
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
for (int i = 1; i < grid_RRow - 1; i++)
{
radius = IFCRadius + i * h;
coef1[coef_StartPos + i] = 1.0 + h / (2 * radius);
coef2[coef_StartPos + i] = 1.0 - h / (2 * radius);
coef3[coef_StartPos + i] = tempRatioPhi / (radius * radius);
coef4[coef_StartPos + i] = 0.5 / (1.0 + tempRatioZ + coef3[coef_StartPos + i]);
icoef4[coef_StartPos + i] = 1.0 / coef4[coef_StartPos + i];
}
// call restrict boundary
if (step == gridFrom) {
// Copy original VPotential to tempPotential
memcpy(temp_VPotential, VPotential, RRow * ZColumn * PhiSlice * sizeof(float));
}
// else
//{
// Restrict_Boundary(temp_VPotential, grid_RRow, grid_ZColumn, PhiSlice, grid_StartPos);
//}
coef_StartPos += grid_RRow;
grid_StartPos += grid_RRow * grid_ZColumn * PhiSlice;
iOne = 2 * iOne;
jOne = 2 * jOne;
}
// device memory allocation
cudaMalloc( &d_VPotential, grid_Size * sizeof(float) );
cudaMalloc( &d_DeltaResidue, grid_Size * sizeof(float) );
cudaMalloc( &d_RhoChargeDensity, grid_Size * sizeof(float) );
cudaMalloc( &d_coef1, coef_Size * sizeof(float) );
cudaMalloc( &d_coef2, coef_Size * sizeof(float) );
cudaMalloc( &d_coef3, coef_Size * sizeof(float) );
cudaMalloc( &d_coef4, coef_Size * sizeof(float) );
cudaMalloc( &d_icoef4, coef_Size * sizeof(float) );
cudaMalloc( &d_VPotentialPrev, grid_Size * sizeof(float) );
cudaMalloc( &d_EpsilonError, 1 * sizeof(float) );
// set memory to zero
cudaMemset( d_VPotential, 0, grid_Size * sizeof(float) );
cudaMemset( d_DeltaResidue, 0, grid_Size * sizeof(float) );
cudaMemset( d_RhoChargeDensity, 0, grid_Size * sizeof(float) );
cudaMemset( d_VPotentialPrev, 0, grid_Size * sizeof(float) );
cudaMemset( d_EpsilonError, 0, 1 * sizeof(float) );
// set memory to zero
cudaMemset( d_VPotential, 0, grid_Size * sizeof(float) );
cudaMemset( d_DeltaResidue, 0, grid_Size * sizeof(float) );
cudaMemset( d_RhoChargeDensity, 0, grid_Size * sizeof(float) );
// copy data from host to devicei
// case of FCycle you need to copy all boundary for all
cudaMemcpy( d_VPotential, temp_VPotential, grid_Size * sizeof(float), cudaMemcpyHostToDevice ); //check
// cudaMemcpy( d_VPotential, VPotential, grid_Size * isizeof(float), cudaMemcpyHostToDevice ); //check
cudaMemcpy( d_RhoChargeDensity, RhoChargeDensity, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyHostToDevice ); //check
// cudaMemcpy( d_RhoChargeDensity, temp_VPotentialPrev, grid_Size * sizeof(float), cudaMemcpyHostToDevice ); //check
cudaMemcpy( d_coef1, coef1, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_coef2, coef2, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_coef3, coef3, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_coef4, coef4, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
cudaMemcpy( d_icoef4, icoef4, coef_Size * sizeof(float), cudaMemcpyHostToDevice );
// cudaMemcpy( d_VPotentialPrev, temp_VPotential, grid_Size * sizeof(float), cudaMemcpyHostToDevice );
// cudaMemcpy( d_VPotentialPrev, VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyHostToDevice );
// max exact
float maxAbsExact = 1.0;
if (isExactPresent == true)
maxAbsExact = GetAbsMax(VPotentialExact, RRow * PhiSlice * ZColumn);
// init iOne,grid_RRow, grid_ZColumn, grid_StartPos, coef_StartPos
iOne = 1 << (gridFrom - 1);
jOne = 1 << (gridFrom - 1);
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
grid_StartPos = 0;
coef_StartPos = 0;
//// Restrict Boundary and Rho
for (int step = gridFrom + 1; step <= gridTo; step++)
{
iOne = 1 << (step - 1);
jOne = 1 << (step - 1);
grid_StartPos += grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos += grid_RRow;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
// pre-compute constant memory
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// restriction
restriction2DFull<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_RhoChargeDensity, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice );
// restrict boundary (already done in cpu)
/// cudaMemcpy( temp_VPotential, d_RhoChargeDensity + grid_StartPos , grid_RRow * grid_ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
// PrintMatrix(temp_VPotential,grid_RRow * PhiSlice,grid_ZColumn);
restriction2DFull<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
}
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
// relax on the coarsest
// red-black gauss seidel relaxation (nPre times)
// printf("rho\n");
// cudaMemcpy( temp_VPotential, d_RhoChargeDensity + grid_StartPos , grid_RRow * grid_ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
// PrintMatrix(temp_VPotential,grid_RRow,grid_ZColumn);
// printf("v\n");
// cudaMemcpy( temp_VPotential, d_VPotential + grid_StartPos , grid_RRow * grid_ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
// PrintMatrix(temp_VPotential,grid_RRow,grid_ZColumn);
for (int i = 0; i < nPre; i++)
{
relaxationGaussSeidelRed<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
cudaDeviceSynchronize();
relaxationGaussSeidelBlack<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, d_RhoChargeDensity, grid_RRow, grid_ZColumn, grid_PhiSlice, d_coef1, d_coef2, d_coef3, d_coef4 );
cudaDeviceSynchronize();
}
// printf("v after relax\n");
// cudaMemcpy( temp_VPotential, d_VPotential + grid_StartPos , grid_RRow * grid_ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
// PrintMatrix(temp_VPotential,grid_RRow,grid_ZColumn);
// V-Cycle => from coarser to finer grid
for (int step = gridTo -1 ; step >= gridFrom; step--)
{
iOne = iOne / 2;
jOne = jOne / 2;
grid_RRow = ((RRow - 1) / iOne) + 1;
grid_ZColumn = ((ZColumn - 1) / jOne) + 1;
grid_StartPos -= grid_RRow * grid_ZColumn * PhiSlice;
coef_StartPos -= grid_RRow;
h = gridSizeR * iOne;
h2 = h * h;
ih2 = 1.0 / h2;
tempRatioZ = ratioZ * iOne * iOne / (jOne * jOne);
tempRatioPhi = ratioPhi * iOne * iOne;
// copy constant to device memory
cudaMemcpyToSymbol( d_grid_StartPos, &grid_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_coef_StartPos, &coef_StartPos, 1 * sizeof(int), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_h2, &h2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_ih2, &ih2, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
cudaMemcpyToSymbol( d_tempRatioZ, &tempRatioZ, 1 * sizeof(float), 0, cudaMemcpyHostToDevice );
// set kernel grid size and block size
dim3 grid_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 grid_ThreadPerBlock(16, 16);
prolongation2DHalfNoAdd<<< grid_BlockPerGrid, grid_ThreadPerBlock >>>( d_VPotential, grid_RRow, grid_ZColumn, grid_PhiSlice );
// just
// max exact
cudaMemcpy( d_VPotentialPrev + grid_StartPos, d_VPotential + grid_StartPos, grid_RRow * grid_ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToDevice );
float maxAbsExact = 1.0;
if (isExactPresent == true)
maxAbsExact = GetAbsMax(VPotentialExact, RRow * PhiSlice * ZColumn);
dim3 error_BlockPerGrid((grid_RRow < 16) ? 1 : (grid_RRow / 16), (grid_ZColumn < 16) ? 1 : (grid_ZColumn / 16), PhiSlice);
dim3 error_ThreadPerBlock(16, 16);
for (int cycle = 0; cycle < nCycle; cycle++)
{
if (step == gridFrom) {
cudaMemcpy( temp_VPotential, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
if (isExactPresent == true )errorExact[cycle] = GetErrorNorm2(temp_VPotential, VPotentialExact, RRow * PhiSlice,ZColumn, maxAbsExact);
}
//cudaDeviceSynchronize();
VCycleSemiCoarseningGPU(d_VPotential, d_RhoChargeDensity, d_DeltaResidue, d_coef1, d_coef2, d_coef3, d_coef4, d_icoef4, gridSizeR, ratioZ, ratioPhi, RRow, ZColumn, PhiSlice, step, gridTo, nPre, nPost);
//if (step == gridFrom) {
//cudaMemcpy( temp_VPotential, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
//errorConv[cycle] = GetErrorNorm2(temp_VPotential, VPotentialPrev, RRow * PhiSlice,ZColumn, 1.0);
errorCalculation<<< error_BlockPerGrid, error_ThreadPerBlock >>> ( d_VPotentialPrev + grid_StartPos, d_VPotential + grid_StartPos, d_EpsilonError, grid_RRow, grid_ZColumn, PhiSlice);
cudaMemcpy( EpsilonError, d_EpsilonError, 1 * sizeof(float), cudaMemcpyDeviceToHost );
errorConv[cycle] = *EpsilonError / (grid_RRow * grid_ZColumn * PhiSlice);
if (errorConv[cycle]< convErr)
{
nCycle = cycle;
break;
}
cudaMemcpy( d_VPotentialPrev + grid_StartPos, d_VPotential + grid_StartPos, grid_RRow * grid_ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToDevice );
cudaMemset( d_EpsilonError, 0, 1 * sizeof(float) );
}
}
iparam[3] = nCycle;
// copy result from device to host
cudaMemcpy( temp_VPotential, d_VPotential, RRow * ZColumn * PhiSlice * sizeof(float), cudaMemcpyDeviceToHost );
memcpy(VPotential, temp_VPotential, RRow * ZColumn * PhiSlice * sizeof(float));
// free device memory
cudaFree( d_VPotential );
cudaFree( d_DeltaResidue );
cudaFree( d_RhoChargeDensity );
cudaFree( d_coef1 );
cudaFree( d_coef2 );
cudaFree( d_coef3 );
cudaFree( d_coef4 );
cudaFree( d_icoef4 );
// free host memory
free( coef1 );
free( coef2 );
free( coef3 );
free( coef4 );
free( icoef4 );
free( temp_VPotential );
free( VPotentialPrev );
}