smyalygames/FiniteVolumeGPU
1
0
Fork 0
mirror of https://github.com/smyalygames/FiniteVolumeGPU.git synced 2026-01-14 15:48:43 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

Fixed order again

Browse Source
This commit is contained in:
André R. Brodtkorb
2018-11-15 16:47:13 +01:00
parent dcb849b705
commit 7592ad5b9f
22 changed files with 758 additions and 619 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
GPUSimulators/cuda/SWE2D_LxF.cu
View File

@@ -118,16 +118,18 @@ void LxFKernel(
const unsigned int w = BLOCK_WIDTH;
const unsigned int h = BLOCK_HEIGHT;
const unsigned int gc = 1;
const unsigned int gc_x = 1;
const unsigned int gc_y = 1;
const unsigned int vars = 3;
__shared__ float Q[3][h+2][w+2];
__shared__ float F[3][h ][w+1];
__shared__ float G[3][h+1][w ];
__shared__ float Q[vars][h+2][w+2];
__shared__ float F[vars][h ][w+1];
__shared__ float G[vars][h+1][w ];
//Read from global memory
readBlock<w, h, gc, 1, 1>( h0_ptr_, h0_pitch_, Q[0], nx_, ny_, boundary_conditions_);
readBlock<w, h, gc, 1, -1>(hu0_ptr_, hu0_pitch_, Q[1], nx_, ny_, boundary_conditions_);
readBlock<w, h, gc, -1, 1>(hv0_ptr_, hv0_pitch_, Q[2], nx_, ny_, boundary_conditions_);
readBlock<w, h, gc_x, gc_y, 1, 1>( h0_ptr_, h0_pitch_, Q[0], nx_, ny_, boundary_conditions_);
readBlock<w, h, gc_x, gc_y, -1, 1>(hu0_ptr_, hu0_pitch_, Q[1], nx_, ny_, boundary_conditions_);
readBlock<w, h, gc_x, gc_y, 1, -1>(hv0_ptr_, hv0_pitch_, Q[2], nx_, ny_, boundary_conditions_);
//Compute fluxes along the x and y axis
computeFluxF<w, h>(Q, F, g_, dx_, dt_);
@@ -149,9 +151,9 @@ void LxFKernel(
__syncthreads();
//Write to main memory
writeBlock<w, h, gc>( h1_ptr_, h1_pitch_, Q[0], nx_, ny_, 0, 1);
writeBlock<w, h, gc>(hu1_ptr_, hu1_pitch_, Q[1], nx_, ny_, 0, 1);
writeBlock<w, h, gc>(hv1_ptr_, hv1_pitch_, Q[2], nx_, ny_, 0, 1);
writeBlock<w, h, gc_x, gc_y>( h1_ptr_, h1_pitch_, Q[0], nx_, ny_, 0, 1);
writeBlock<w, h, gc_x, gc_y>(hu1_ptr_, hu1_pitch_, Q[1], nx_, ny_, 0, 1);
writeBlock<w, h, gc_x, gc_y>(hv1_ptr_, hv1_pitch_, Q[2], nx_, ny_, 0, 1);
}
} // extern "C"