smyalygames/FiniteVolumeGPU
1
0
Fork 0
mirror of https://github.com/smyalygames/FiniteVolumeGPU.git synced 2025-05-18 14:34:13 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity

Refactoring

Browse Source
This commit is contained in:
André R. Brodtkorb 2018-11-01 21:34:53 +01:00
parent d9eb72d78c
commit 0671bd747a
6 changed files with 178 additions and 212 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

55
GPUSimulators/cuda/SWE2D_FORCE.cu
View File

@ -28,27 +28,18 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
*/ */
__device__ __device__
void computeFluxF(float Q[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2], void computeFluxF(float Q[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
float F[3][BLOCK_HEIGHT+1][BLOCK_WIDTH+1], float F[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
const float g_, const float dx_, const float dt_) { const float g_, const float dx_, const float dt_) {
//Index of thread within block
const int tx = threadIdx.x;
const int ty = threadIdx.y;
//Compute fluxes along the x axis //Compute fluxes along the x axis
{ for (int j=threadIdx.y; j<BLOCK_HEIGHT+2; j+=BLOCK_HEIGHT) {
int j=ty; for (int i=threadIdx.x; i<BLOCK_WIDTH+1; i+=BLOCK_WIDTH) {
const int l = j + 1; //Skip ghost cells
for (int i=tx; i<BLOCK_WIDTH+1; i+=BLOCK_WIDTH) {
const int k = i;
// Q at interface from the right and left // Q at interface from the right and left
const float3 Qp = make_float3(Q[0][l][k+1], const float3 Qp = make_float3(Q[0][j][i+1],
Q[1][l][k+1], Q[1][j][i+1],
Q[2][l][k+1]); Q[2][j][i+1]);
const float3 Qm = make_float3(Q[0][l][k], const float3 Qm = make_float3(Q[0][j][i],
Q[1][l][k], Q[1][j][i],
Q[2][l][k]); Q[2][j][i]);
// Computed flux // Computed flux
const float3 flux = FORCE_1D_flux(Qm, Qp, g_, dx_, dt_); const float3 flux = FORCE_1D_flux(Qm, Qp, g_, dx_, dt_);
@ -65,27 +56,19 @@ void computeFluxF(float Q[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
*/ */
__device__ __device__
void computeFluxG(float Q[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2], void computeFluxG(float Q[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
float G[3][BLOCK_HEIGHT+1][BLOCK_WIDTH+1], float G[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
const float g_, const float dy_, const float dt_) { const float g_, const float dy_, const float dt_) {
//Index of thread within block
const int tx = threadIdx.x;
const int ty = threadIdx.y;
//Compute fluxes along the y axis //Compute fluxes along the y axis
for (int j=ty; j<BLOCK_HEIGHT+1; j+=BLOCK_HEIGHT) { for (int j=threadIdx.y; j<BLOCK_HEIGHT+1; j+=BLOCK_HEIGHT) {
const int l = j; for (int i=threadIdx.x; i<BLOCK_WIDTH+2; i+=BLOCK_WIDTH) {
{
int i=tx;
const int k = i + 1; //Skip ghost cells
// Q at interface from the right and left // Q at interface from the right and left
// Note that we swap hu and hv // Note that we swap hu and hv
const float3 Qp = make_float3(Q[0][l+1][k], const float3 Qp = make_float3(Q[0][j+1][i],
Q[2][l+1][k], Q[2][j+1][i],
Q[1][l+1][k]); Q[1][j+1][i]);
const float3 Qm = make_float3(Q[0][l][k], const float3 Qm = make_float3(Q[0][j][i],
Q[2][l][k], Q[2][j][i],
Q[1][l][k]); Q[1][j][i]);
// Computed flux // Computed flux
// Note that we swap back // Note that we swap back
@ -120,7 +103,7 @@ __global__ void FORCEKernel(
const unsigned int vars = 3; const unsigned int vars = 3;
__shared__ float Q[3][h+2][w+2]; __shared__ float Q[3][h+2][w+2];
__shared__ float F[3][h+1][w+1]; __shared__ float F[3][h+2][w+2];
//Read into shared memory //Read into shared memory
readBlock<w, h, gc>( h0_ptr_, h0_pitch_, Q[0], nx_+2, ny_+2); readBlock<w, h, gc>( h0_ptr_, h0_pitch_, Q[0], nx_+2, ny_+2);

60
GPUSimulators/cuda/SWE2D_HLL.cu
View File

@ -31,24 +31,20 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
*/ */
__device__ __device__
void computeFluxF(float Q[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2], void computeFluxF(float Q[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
float F[3][BLOCK_HEIGHT+1][BLOCK_WIDTH+1], float F[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
const float g_) { const float g_) {
//Index of thread within block for (int j=threadIdx.y; j<BLOCK_HEIGHT+2; j+=BLOCK_HEIGHT) {
const int tx = threadIdx.x; for (int i=threadIdx.x; i<BLOCK_WIDTH+1; i+=BLOCK_WIDTH) {
const int ty = threadIdx.y; // Q at interface from the right and left
const float3 Q_r = make_float3(Q[0][j][i+1],
{ Q[1][j][i+1],
const int j=ty; Q[2][j][i+1]);
const int l = j + 1; //Skip ghost cells const float3 Q_l = make_float3(Q[0][j][i],
for (int i=tx; i<BLOCK_WIDTH+1; i+=BLOCK_WIDTH) { Q[1][j][i],
const int k = i; Q[2][j][i]);
const float3 Q_l = make_float3(Q[0][l][k ], Q[1][l][k ], Q[2][l][k ]); // Computed flux
const float3 Q_r = make_float3(Q[0][l][k+1], Q[1][l][k+1], Q[2][l][k+1]);
const float3 flux = HLL_flux(Q_l, Q_r, g_); const float3 flux = HLL_flux(Q_l, Q_r, g_);
//Write to shared memory
F[0][j][i] = flux.x; F[0][j][i] = flux.x;
F[1][j][i] = flux.y; F[1][j][i] = flux.y;
F[2][j][i] = flux.z; F[2][j][i] = flux.z;
@ -65,27 +61,23 @@ void computeFluxF(float Q[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
*/ */
__device__ __device__
void computeFluxG(float Q[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2], void computeFluxG(float Q[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
float G[3][BLOCK_HEIGHT+1][BLOCK_WIDTH+1], float G[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
const float g_) { const float g_) {
//Index of thread within block //Compute fluxes along the y axis
const int tx = threadIdx.x; for (int j=threadIdx.y; j<BLOCK_HEIGHT+1; j+=BLOCK_HEIGHT) {
const int ty = threadIdx.y; for (int i=threadIdx.x; i<BLOCK_WIDTH+2; i+=BLOCK_WIDTH) {
// Q at interface from the right and left
for (int j=ty; j<BLOCK_HEIGHT+1; j+=BLOCK_HEIGHT) { // Note that we swap hu and hv
const int l = j; const float3 Q_r = make_float3(Q[0][j+1][i],
{ Q[2][j+1][i],
const int i=tx; Q[1][j+1][i]);
const int k = i + 1; //Skip ghost cells const float3 Q_l = make_float3(Q[0][j][i],
Q[2][j][i],
//NOte that hu and hv are swapped ("transposing" the domain)! Q[1][j][i]);
const float3 Q_l = make_float3(Q[0][l ][k], Q[2][l ][k], Q[1][l ][k]);
const float3 Q_r = make_float3(Q[0][l+1][k], Q[2][l+1][k], Q[1][l+1][k]);
// Computed flux // Computed flux
const float3 flux = HLL_flux(Q_l, Q_r, g_);
//Write to shared memory
//Note that we here swap hu and hv back to the original //Note that we here swap hu and hv back to the original
const float3 flux = HLL_flux(Q_l, Q_r, g_);
G[0][j][i] = flux.x; G[0][j][i] = flux.x;
G[1][j][i] = flux.z; G[1][j][i] = flux.z;
G[2][j][i] = flux.y; G[2][j][i] = flux.y;
@ -128,7 +120,7 @@ __global__ void HLLKernel(
//Shared memory variables //Shared memory variables
__shared__ float Q[3][h+2][w+2]; __shared__ float Q[3][h+2][w+2];
__shared__ float F[3][h+1][w+1]; __shared__ float F[3][h+2][w+2];
//Read into shared memory //Read into shared memory
readBlock<w, h, gc>( h0_ptr_, h0_pitch_, Q[0], nx_+2, ny_+2); readBlock<w, h, gc>( h0_ptr_, h0_pitch_, Q[0], nx_+2, ny_+2);

82
GPUSimulators/cuda/SWE2D_HLL2.cu
View File

@ -33,33 +33,26 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
*/ */
__device__ __device__
void computeFluxF(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4], void computeFluxF(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float Qx[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2], float Qx[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float F[3][BLOCK_HEIGHT+1][BLOCK_WIDTH+1], float F[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
const float g_, const float dx_, const float dt_) { const float g_, const float dx_, const float dt_) {
//Index of thread within block for (int j=threadIdx.y; j<BLOCK_HEIGHT+4; j+=BLOCK_HEIGHT) {
const int tx = threadIdx.x; for (int i=threadIdx.x+1; i<BLOCK_WIDTH+2; i+=BLOCK_WIDTH) {
const int ty = threadIdx.y;
{
const int j=ty;
const int l = j + 2; //Skip ghost cells
for (int i=tx; i<BLOCK_WIDTH+1; i+=BLOCK_WIDTH) {
const int k = i + 1;
// Reconstruct point values of Q at the left and right hand side // Reconstruct point values of Q at the left and right hand side
// of the cell for both the left (i) and right (i+1) cell // of the cell for both the left (i) and right (i+1) cell
const float3 Q_rl = make_float3(Q[0][l][k+1] - 0.5f*Qx[0][j][i+1], const float3 Q_rl = make_float3(Q[0][j][i+1] - 0.5f*Qx[0][j][i+1],
Q[1][l][k+1] - 0.5f*Qx[1][j][i+1], Q[1][j][i+1] - 0.5f*Qx[1][j][i+1],
Q[2][l][k+1] - 0.5f*Qx[2][j][i+1]); Q[2][j][i+1] - 0.5f*Qx[2][j][i+1]);
const float3 Q_rr = make_float3(Q[0][l][k+1] + 0.5f*Qx[0][j][i+1], const float3 Q_rr = make_float3(Q[0][j][i+1] + 0.5f*Qx[0][j][i+1],
Q[1][l][k+1] + 0.5f*Qx[1][j][i+1], Q[1][j][i+1] + 0.5f*Qx[1][j][i+1],
Q[2][l][k+1] + 0.5f*Qx[2][j][i+1]); Q[2][j][i+1] + 0.5f*Qx[2][j][i+1]);
const float3 Q_ll = make_float3(Q[0][l][k] - 0.5f*Qx[0][j][i], const float3 Q_ll = make_float3(Q[0][j][i] - 0.5f*Qx[0][j][i],
Q[1][l][k] - 0.5f*Qx[1][j][i], Q[1][j][i] - 0.5f*Qx[1][j][i],
Q[2][l][k] - 0.5f*Qx[2][j][i]); Q[2][j][i] - 0.5f*Qx[2][j][i]);
const float3 Q_lr = make_float3(Q[0][l][k] + 0.5f*Qx[0][j][i], const float3 Q_lr = make_float3(Q[0][j][i] + 0.5f*Qx[0][j][i],
Q[1][l][k] + 0.5f*Qx[1][j][i], Q[1][j][i] + 0.5f*Qx[1][j][i],
Q[2][l][k] + 0.5f*Qx[2][j][i]); Q[2][j][i] + 0.5f*Qx[2][j][i]);
//Evolve half a timestep (predictor step) //Evolve half a timestep (predictor step)
const float3 Q_r_bar = Q_rl + dt_/(2.0f*dx_) * (F_func(Q_rl, g_) - F_func(Q_rr, g_)); const float3 Q_r_bar = Q_rl + dt_/(2.0f*dx_) * (F_func(Q_rl, g_) - F_func(Q_rr, g_));
@ -85,34 +78,27 @@ void computeFluxF(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
*/ */
__device__ __device__
void computeFluxG(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4], void computeFluxG(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float Qy[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2], float Qy[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float G[3][BLOCK_HEIGHT+1][BLOCK_WIDTH+1], float G[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
const float g_, const float dy_, const float dt_) { const float g_, const float dy_, const float dt_) {
//Index of thread within block for (int j=threadIdx.y+1; j<BLOCK_HEIGHT+2; j+=BLOCK_HEIGHT) {
const int tx = threadIdx.x; for (int i=threadIdx.x; i<BLOCK_WIDTH+4; i+=BLOCK_WIDTH) {
const int ty = threadIdx.y;
for (int j=ty; j<BLOCK_HEIGHT+1; j+=BLOCK_HEIGHT) {
const int l = j + 1;
{
int i=tx;
const int k = i + 2; //Skip ghost cells
// Reconstruct point values of Q at the left and right hand side // Reconstruct point values of Q at the left and right hand side
// of the cell for both the left (i) and right (i+1) cell // of the cell for both the left (i) and right (i+1) cell
//NOte that hu and hv are swapped ("transposing" the domain)! //NOte that hu and hv are swapped ("transposing" the domain)!
const float3 Q_rl = make_float3(Q[0][l+1][k] - 0.5f*Qy[0][j+1][i], const float3 Q_rl = make_float3(Q[0][j+1][i] - 0.5f*Qy[0][j+1][i],
Q[2][l+1][k] - 0.5f*Qy[2][j+1][i], Q[2][j+1][i] - 0.5f*Qy[2][j+1][i],
Q[1][l+1][k] - 0.5f*Qy[1][j+1][i]); Q[1][j+1][i] - 0.5f*Qy[1][j+1][i]);
const float3 Q_rr = make_float3(Q[0][l+1][k] + 0.5f*Qy[0][j+1][i], const float3 Q_rr = make_float3(Q[0][j+1][i] + 0.5f*Qy[0][j+1][i],
Q[2][l+1][k] + 0.5f*Qy[2][j+1][i], Q[2][j+1][i] + 0.5f*Qy[2][j+1][i],
Q[1][l+1][k] + 0.5f*Qy[1][j+1][i]); Q[1][j+1][i] + 0.5f*Qy[1][j+1][i]);
const float3 Q_ll = make_float3(Q[0][l][k] - 0.5f*Qy[0][j][i], const float3 Q_ll = make_float3(Q[0][j][i] - 0.5f*Qy[0][j][i],
Q[2][l][k] - 0.5f*Qy[2][j][i], Q[2][j][i] - 0.5f*Qy[2][j][i],
Q[1][l][k] - 0.5f*Qy[1][j][i]); Q[1][j][i] - 0.5f*Qy[1][j][i]);
const float3 Q_lr = make_float3(Q[0][l][k] + 0.5f*Qy[0][j][i], const float3 Q_lr = make_float3(Q[0][j][i] + 0.5f*Qy[0][j][i],
Q[2][l][k] + 0.5f*Qy[2][j][i], Q[2][j][i] + 0.5f*Qy[2][j][i],
Q[1][l][k] + 0.5f*Qy[1][j][i]); Q[1][j][i] + 0.5f*Qy[1][j][i]);
//Evolve half a timestep (predictor step) //Evolve half a timestep (predictor step)
const float3 Q_r_bar = Q_rl + dt_/(2.0f*dy_) * (F_func(Q_rl, g_) - F_func(Q_rr, g_)); const float3 Q_r_bar = Q_rl + dt_/(2.0f*dy_) * (F_func(Q_rl, g_) - F_func(Q_rr, g_));
@ -163,8 +149,8 @@ __global__ void HLL2Kernel(
//Shared memory variables //Shared memory variables
__shared__ float Q[3][h+4][w+4]; __shared__ float Q[3][h+4][w+4];
__shared__ float Qx[3][h+2][w+2]; __shared__ float Qx[3][h+4][w+4];
__shared__ float F[3][h+1][w+1]; __shared__ float F[3][h+4][w+4];
//Read into shared memory //Read into shared memory
readBlock<w, h, gc>( h0_ptr_, h0_pitch_, Q[0], nx_+2, ny_+2); readBlock<w, h, gc>( h0_ptr_, h0_pitch_, Q[0], nx_+2, ny_+2);

37
GPUSimulators/cuda/SWE2D_KP07.cu
View File

@ -96,6 +96,39 @@ void computeFluxG(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
__device__ void minmodSlopeX(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float Qx[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
const float theta_) {
//Reconstruct slopes along x axis
for (int p=0; p<3; ++p) {
{
const int j = threadIdx.y+2;
for (int i=threadIdx.x+1; i<BLOCK_WIDTH+3; i+=BLOCK_WIDTH) {
Qx[p][j-2][i-1] = minmodSlope(Q[p][j][i-1], Q[p][j][i], Q[p][j][i+1], theta_);
}
}
}
}
/**
* Reconstructs a minmod slope for a whole block along the ordinate
*/
__device__ void minmodSlopeY(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float Qy[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2],
const float theta_) {
//Reconstruct slopes along y axis
for (int p=0; p<3; ++p) {
const int i = threadIdx.x + 2;
for (int j=threadIdx.y+1; j<BLOCK_HEIGHT+3; j+=BLOCK_HEIGHT) {
{
Qy[p][j-1][i-2] = minmodSlope(Q[p][j-1][i], Q[p][j][i], Q[p][j+1][i], theta_);
}
}
}
}
/** /**
* This unsplit kernel computes the 2D numerical scheme with a TVD RK2 time integration scheme * This unsplit kernel computes the 2D numerical scheme with a TVD RK2 time integration scheme
*/ */
@ -159,8 +192,8 @@ __global__ void KP07Kernel(
//Reconstruct slopes along x and axis //Reconstruct slopes along x and axis
minmodSlopeX<w, h, gc, vars>(Q, Qx, theta_); minmodSlopeX(Q, Qx, theta_);
minmodSlopeY<w, h, gc, vars>(Q, Qy, theta_); minmodSlopeY(Q, Qy, theta_);
__syncthreads(); __syncthreads();

110
GPUSimulators/cuda/SWE2D_KP07_dimsplit.cu
View File

@ -31,77 +31,65 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
__device__ __device__
void computeFluxF(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4], void computeFluxF(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float Qx[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2], float Qx[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float F[3][BLOCK_HEIGHT+1][BLOCK_WIDTH+1], float F[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
const float g_, const float dx_, const float dt_) { const float g_, const float dx_, const float dt_) {
//Index of thread within block for (int j=threadIdx.y; j<BLOCK_HEIGHT+4; j+=BLOCK_HEIGHT) {
const int tx = threadIdx.x; for (int i=threadIdx.x+1; i<BLOCK_WIDTH+2; i+=BLOCK_WIDTH) {
const int ty = threadIdx.y; // Reconstruct point values of Q at the left and right hand side
// of the cell for both the left (i) and right (i+1) cell
int j=ty; const float3 Q_rl = make_float3(Q[0][j][i+1] - 0.5f*Qx[0][j][i+1],
const int l = j + 2; //Skip ghost cells Q[1][j][i+1] - 0.5f*Qx[1][j][i+1],
for (int i=tx; i<BLOCK_WIDTH+1; i+=BLOCK_WIDTH) { Q[2][j][i+1] - 0.5f*Qx[2][j][i+1]);
const int k = i + 1; const float3 Q_rr = make_float3(Q[0][j][i+1] + 0.5f*Qx[0][j][i+1],
// Reconstruct point values of Q at the left and right hand side Q[1][j][i+1] + 0.5f*Qx[1][j][i+1],
// of the cell for both the left (i) and right (i+1) cell Q[2][j][i+1] + 0.5f*Qx[2][j][i+1]);
const float3 Q_rl = make_float3(Q[0][l][k+1] - 0.5f*Qx[0][j][i+1],
Q[1][l][k+1] - 0.5f*Qx[1][j][i+1],
Q[2][l][k+1] - 0.5f*Qx[2][j][i+1]);
const float3 Q_rr = make_float3(Q[0][l][k+1] + 0.5f*Qx[0][j][i+1],
Q[1][l][k+1] + 0.5f*Qx[1][j][i+1],
Q[2][l][k+1] + 0.5f*Qx[2][j][i+1]);
const float3 Q_ll = make_float3(Q[0][l][k] - 0.5f*Qx[0][j][i],
Q[1][l][k] - 0.5f*Qx[1][j][i],
Q[2][l][k] - 0.5f*Qx[2][j][i]);
const float3 Q_lr = make_float3(Q[0][l][k] + 0.5f*Qx[0][j][i],
Q[1][l][k] + 0.5f*Qx[1][j][i],
Q[2][l][k] + 0.5f*Qx[2][j][i]);
//Evolve half a timestep (predictor step)
const float3 Q_r_bar = Q_rl + dt_/(2.0f*dx_) * (F_func(Q_rl, g_) - F_func(Q_rr, g_));
const float3 Q_l_bar = Q_lr + dt_/(2.0f*dx_) * (F_func(Q_ll, g_) - F_func(Q_lr, g_));
// Compute flux based on prediction const float3 Q_ll = make_float3(Q[0][j][i] - 0.5f*Qx[0][j][i],
const float3 flux = CentralUpwindFlux(Q_l_bar, Q_r_bar, g_); Q[1][j][i] - 0.5f*Qx[1][j][i],
Q[2][j][i] - 0.5f*Qx[2][j][i]);
//Write to shared memory const float3 Q_lr = make_float3(Q[0][j][i] + 0.5f*Qx[0][j][i],
F[0][j][i] = flux.x; Q[1][j][i] + 0.5f*Qx[1][j][i],
F[1][j][i] = flux.y; Q[2][j][i] + 0.5f*Qx[2][j][i]);
F[2][j][i] = flux.z;
//Evolve half a timestep (predictor step)
const float3 Q_r_bar = Q_rl + dt_/(2.0f*dx_) * (F_func(Q_rl, g_) - F_func(Q_rr, g_));
const float3 Q_l_bar = Q_lr + dt_/(2.0f*dx_) * (F_func(Q_ll, g_) - F_func(Q_lr, g_));
// Compute flux based on prediction
const float3 flux = CentralUpwindFlux(Q_l_bar, Q_r_bar, g_);
//Write to shared memory
F[0][j][i] = flux.x;
F[1][j][i] = flux.y;
F[2][j][i] = flux.z;
}
} }
} }
__device__ __device__
void computeFluxG(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4], void computeFluxG(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float Qy[3][BLOCK_HEIGHT+2][BLOCK_WIDTH+2], float Qy[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float G[3][BLOCK_HEIGHT+1][BLOCK_WIDTH+1], float G[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
const float g_, const float dy_, const float dt_) { const float g_, const float dy_, const float dt_) {
//Index of thread within block for (int j=threadIdx.y+1; j<BLOCK_HEIGHT+2; j+=BLOCK_HEIGHT) {
const int tx = threadIdx.x; for (int i=threadIdx.x; i<BLOCK_WIDTH+4; i+=BLOCK_WIDTH) {
const int ty = threadIdx.y;
for (int j=ty; j<BLOCK_HEIGHT+1; j+=BLOCK_HEIGHT) {
const int l = j + 1;
{
int i=tx;
const int k = i + 2; //Skip ghost cells
// Reconstruct point values of Q at the left and right hand side // Reconstruct point values of Q at the left and right hand side
// of the cell for both the left (i) and right (i+1) cell // of the cell for both the left (i) and right (i+1) cell
//NOte that hu and hv are swapped ("transposing" the domain)! //NOte that hu and hv are swapped ("transposing" the domain)!
const float3 Q_rl = make_float3(Q[0][l+1][k] - 0.5f*Qy[0][j+1][i], const float3 Q_rl = make_float3(Q[0][j+1][i] - 0.5f*Qy[0][j+1][i],
Q[2][l+1][k] - 0.5f*Qy[2][j+1][i], Q[2][j+1][i] - 0.5f*Qy[2][j+1][i],
Q[1][l+1][k] - 0.5f*Qy[1][j+1][i]); Q[1][j+1][i] - 0.5f*Qy[1][j+1][i]);
const float3 Q_rr = make_float3(Q[0][l+1][k] + 0.5f*Qy[0][j+1][i], const float3 Q_rr = make_float3(Q[0][j+1][i] + 0.5f*Qy[0][j+1][i],
Q[2][l+1][k] + 0.5f*Qy[2][j+1][i], Q[2][j+1][i] + 0.5f*Qy[2][j+1][i],
Q[1][l+1][k] + 0.5f*Qy[1][j+1][i]); Q[1][j+1][i] + 0.5f*Qy[1][j+1][i]);
const float3 Q_ll = make_float3(Q[0][l][k] - 0.5f*Qy[0][j][i], const float3 Q_ll = make_float3(Q[0][j][i] - 0.5f*Qy[0][j][i],
Q[2][l][k] - 0.5f*Qy[2][j][i], Q[2][j][i] - 0.5f*Qy[2][j][i],
Q[1][l][k] - 0.5f*Qy[1][j][i]); Q[1][j][i] - 0.5f*Qy[1][j][i]);
const float3 Q_lr = make_float3(Q[0][l][k] + 0.5f*Qy[0][j][i], const float3 Q_lr = make_float3(Q[0][j][i] + 0.5f*Qy[0][j][i],
Q[2][l][k] + 0.5f*Qy[2][j][i], Q[2][j][i] + 0.5f*Qy[2][j][i],
Q[1][l][k] + 0.5f*Qy[1][j][i]); Q[1][j][i] + 0.5f*Qy[1][j][i]);
//Evolve half a timestep (predictor step) //Evolve half a timestep (predictor step)
const float3 Q_r_bar = Q_rl + dt_/(2.0f*dy_) * (F_func(Q_rl, g_) - F_func(Q_rr, g_)); const float3 Q_r_bar = Q_rl + dt_/(2.0f*dy_) * (F_func(Q_rl, g_) - F_func(Q_rr, g_));
@ -153,8 +141,8 @@ __global__ void KP07DimsplitKernel(
//Shared memory variables //Shared memory variables
__shared__ float Q[3][h+4][w+4]; __shared__ float Q[3][h+4][w+4];
__shared__ float Qx[3][h+2][w+2]; __shared__ float Qx[3][h+4][w+4];
__shared__ float F[3][h+1][w+1]; __shared__ float F[3][h+4][w+4];

46
GPUSimulators/cuda/SWE2D_WAF.cu
View File

@ -34,23 +34,15 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
*/ */
__device__ __device__
void computeFluxF(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4], void computeFluxF(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float F[3][BLOCK_HEIGHT+1][BLOCK_WIDTH+1], float F[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
const float g_, const float dx_, const float dt_) { const float g_, const float dx_, const float dt_) {
//Index of thread within block for (int j=threadIdx.y; j<BLOCK_HEIGHT+4; j+=BLOCK_HEIGHT) {
const int tx = threadIdx.x; for (int i=threadIdx.x+1; i<BLOCK_WIDTH+2; i+=BLOCK_WIDTH) {
const int ty = threadIdx.y;
{
int j=ty;
const int l = j + 2; //Skip ghost cells
for (int i=tx; i<BLOCK_WIDTH+1; i+=BLOCK_WIDTH) {
const int k = i + 1;
// Q at interface from the right and left // Q at interface from the right and left
const float3 Ql2 = make_float3(Q[0][l][k-1], Q[1][l][k-1], Q[2][l][k-1]); const float3 Ql2 = make_float3(Q[0][j][i-1], Q[1][j][i-1], Q[2][j][i-1]);
const float3 Ql1 = make_float3(Q[0][l][k ], Q[1][l][k ], Q[2][l][k ]); const float3 Ql1 = make_float3(Q[0][j][i ], Q[1][j][i ], Q[2][j][i ]);
const float3 Qr1 = make_float3(Q[0][l][k+1], Q[1][l][k+1], Q[2][l][k+1]); const float3 Qr1 = make_float3(Q[0][j][i+1], Q[1][j][i+1], Q[2][j][i+1]);
const float3 Qr2 = make_float3(Q[0][l][k+2], Q[1][l][k+2], Q[2][l][k+2]); const float3 Qr2 = make_float3(Q[0][j][i+2], Q[1][j][i+2], Q[2][j][i+2]);
// Computed flux // Computed flux
const float3 flux = WAF_1D_flux(Ql2, Ql1, Qr1, Qr2, g_, dx_, dt_); const float3 flux = WAF_1D_flux(Ql2, Ql1, Qr1, Qr2, g_, dx_, dt_);
@ -73,24 +65,16 @@ void computeFluxF(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
*/ */
__device__ __device__
void computeFluxG(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4], void computeFluxG(float Q[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
float G[3][BLOCK_HEIGHT+1][BLOCK_WIDTH+1], float G[3][BLOCK_HEIGHT+4][BLOCK_WIDTH+4],
const float g_, const float dy_, const float dt_) { const float g_, const float dy_, const float dt_) {
//Index of thread within block for (int j=threadIdx.y+1; j<BLOCK_HEIGHT+2; j+=BLOCK_HEIGHT) {
const int tx = threadIdx.x; for (int i=threadIdx.x; i<BLOCK_WIDTH+4; i+=BLOCK_WIDTH) {
const int ty = threadIdx.y;
//Compute fluxes along the y axis
for (int j=ty; j<BLOCK_HEIGHT+1; j+=BLOCK_HEIGHT) {
const int l = j + 1;
{
int i=tx;
const int k = i + 2; //Skip ghost cells
// Q at interface from the right and left // Q at interface from the right and left
// Note that we swap hu and hv // Note that we swap hu and hv
const float3 Ql2 = make_float3(Q[0][l-1][k], Q[2][l-1][k], Q[1][l-1][k]); const float3 Ql2 = make_float3(Q[0][j-1][i], Q[2][j-1][i], Q[1][j-1][i]);
const float3 Ql1 = make_float3(Q[0][l ][k], Q[2][l ][k], Q[1][l ][k]); const float3 Ql1 = make_float3(Q[0][j ][i], Q[2][j ][i], Q[1][j ][i]);
const float3 Qr1 = make_float3(Q[0][l+1][k], Q[2][l+1][k], Q[1][l+1][k]); const float3 Qr1 = make_float3(Q[0][j+1][i], Q[2][j+1][i], Q[1][j+1][i]);
const float3 Qr2 = make_float3(Q[0][l+2][k], Q[2][l+2][k], Q[1][l+2][k]); const float3 Qr2 = make_float3(Q[0][j+2][i], Q[2][j+2][i], Q[1][j+2][i]);
// Computed flux // Computed flux
// Note that we swap back // Note that we swap back
@ -138,7 +122,7 @@ __global__ void WAFKernel(
//Shared memory variables //Shared memory variables
__shared__ float Q[3][h+4][w+4]; __shared__ float Q[3][h+4][w+4];
__shared__ float F[3][h+1][w+1]; __shared__ float F[3][h+4][w+4];