Searched refs:chunksperblock (Results 1 – 2 of 2) sorted by relevance
| /petsc/src/mat/impls/sell/seq/seqcuda/ |
| H A D | sellcuda.cu | 17 PetscInt chunksperblock; member
185 …seqsell_tiled_kernel8(PetscInt nrows, PetscInt sliceheight, PetscInt chunksperblock, PetscInt tota… in matmult_seqsell_tiled_kernel8() argument
196 for (int iter = 0; iter < chunksperblock; iter++) { in matmult_seqsell_tiled_kernel8()
197 cid = blockIdx.x * chunksperblock + iter; /* chunk id */ in matmult_seqsell_tiled_kernel8()
217 …if (iter == chunksperblock - 1 || (cid + 2) * BLOCKY * 32 > sliidx[start_slice + 1]) { /* last ite… in matmult_seqsell_tiled_kernel8()
240 …seqsell_tiled_kernel8(PetscInt nrows, PetscInt sliceheight, PetscInt chunksperblock, PetscInt tota… in matmultadd_seqsell_tiled_kernel8() argument
251 for (int iter = 0; iter < chunksperblock; iter++) { in matmultadd_seqsell_tiled_kernel8()
252 cid = blockIdx.x * chunksperblock + iter; /* chunk id */ in matmultadd_seqsell_tiled_kernel8()
272 …if (iter == chunksperblock - 1 || (cid + 2) * BLOCKY * 32 > sliidx[start_slice + 1]) { /* last ite… in matmultadd_seqsell_tiled_kernel8()
579 PetscInt chunksperblock, nchunks, *chunk_slice_map; in MatMult_SeqSELLCUDA() local
[all …]
|
| /petsc/src/mat/impls/sell/seq/seqhip/ |
| H A D | sellhip.hip.cxx | 20 PetscInt chunksperblock; member
189 …seqsell_tiled_kernel8(PetscInt nrows, PetscInt sliceheight, PetscInt chunksperblock, PetscInt tota… in matmult_seqsell_tiled_kernel8() argument
200 for (int iter = 0; iter < chunksperblock; iter++) { in matmult_seqsell_tiled_kernel8()
201 cid = blockIdx.x * chunksperblock + iter; /* chunk id */ in matmult_seqsell_tiled_kernel8()
221 …if (iter == chunksperblock - 1 || (cid + 2) * BLOCKY * WARP_SIZE > sliidx[start_slice + 1]) { /* l… in matmult_seqsell_tiled_kernel8()
244 …seqsell_tiled_kernel8(PetscInt nrows, PetscInt sliceheight, PetscInt chunksperblock, PetscInt tota… in matmultadd_seqsell_tiled_kernel8() argument
255 for (int iter = 0; iter < chunksperblock; iter++) { in matmultadd_seqsell_tiled_kernel8()
256 cid = blockIdx.x * chunksperblock + iter; /* chunk id */ in matmultadd_seqsell_tiled_kernel8()
276 …if (iter == chunksperblock - 1 || (cid + 2) * BLOCKY * WARP_SIZE > sliidx[start_slice + 1]) { /* l… in matmultadd_seqsell_tiled_kernel8()
584 PetscInt chunksperblock, nchunks, *chunk_slice_map; in MatMult_SeqSELLHIP() local
[all …]
|