Lines Matching refs:chunksperblock

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
664chunksperblock = hipstruct->chunksperblock ? hipstruct->chunksperblock : 1 + (hipstruct->totalent… in MatMult_SeqSELLHIP()
665 nblocks = 1 + (nchunks - 1) / chunksperblock; in MatMult_SeqSELLHIP()
668 …iled_kernel8<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMult_SeqSELLHIP()
670 …iled_kernel8<4><<<nblocks, dim3(WARP_SIZE, 4)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMult_SeqSELLHIP()
672 …iled_kernel8<8><<<nblocks, dim3(WARP_SIZE, 8)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMult_SeqSELLHIP()
674 …ed_kernel8<16><<<nblocks, dim3(WARP_SIZE, 16)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMult_SeqSELLHIP()
676 …iled_kernel8<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMult_SeqSELLHIP()
720 PetscInt chunksperblock, nchunks, *chunk_slice_map; in MatMultAdd_SeqSELLHIP() local
756chunksperblock = hipstruct->chunksperblock ? hipstruct->chunksperblock : 1 + (hipstruct->totalent… in MatMultAdd_SeqSELLHIP()
757 nblocks = 1 + (nchunks - 1) / chunksperblock; in MatMultAdd_SeqSELLHIP()
760 …iled_kernel8<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMultAdd_SeqSELLHIP()
762 …iled_kernel8<4><<<nblocks, dim3(WARP_SIZE, 4)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMultAdd_SeqSELLHIP()
764 …iled_kernel8<8><<<nblocks, dim3(WARP_SIZE, 8)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMultAdd_SeqSELLHIP()
766 …ed_kernel8<16><<<nblocks, dim3(WARP_SIZE, 16)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMultAdd_SeqSELLHIP()
768 …iled_kernel8<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMultAdd_SeqSELLHIP()
818chunksperblock = hipstruct->chunksperblock ? hipstruct->chunksperblock : 1 + (hipstruct->totalent… in MatMultAdd_SeqSELLHIP()
819 nblocks = 1 + (nchunks - 1) / chunksperblock; in MatMultAdd_SeqSELLHIP()
822 …iled_kernel8<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMultAdd_SeqSELLHIP()
824 …iled_kernel8<4><<<nblocks, dim3(WARP_SIZE, 4)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMultAdd_SeqSELLHIP()
826 …iled_kernel8<8><<<nblocks, dim3(WARP_SIZE, 8)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMultAdd_SeqSELLHIP()
828 …ed_kernel8<16><<<nblocks, dim3(WARP_SIZE, 16)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMultAdd_SeqSELLHIP()
830 …iled_kernel8<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, chunksperblock, nchunks, chun… in MatMultAdd_SeqSELLHIP()
883 …cOptionsGetInt(NULL, NULL, "-mat_sell_spmv_hip_chunksperblock", &hipstruct->chunksperblock, &flg)); in MatSetFromOptions_SeqSELLHIP()