Lines Matching refs:aval
78 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmult_seqsell_basic_kernel() argument
88 …ice_id] + row_in_slice; i < sliidx[slice_id + 1]; i += sliceheight) sum += aval[i] * x[acolidx[i]]; in matmult_seqsell_basic_kernel()
93 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmultadd_seqsell_basic_kernel() argument
103 …ice_id] + row_in_slice; i < sliidx[slice_id + 1]; i += sliceheight) sum += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_basic_kernel()
112 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmult_seqsell_tiled_kernel9() argument
124 …IZE * threadIdx.y; i < sliidx[slice_id + 1]; i += WARP_SIZE * BLOCKY) t += aval[i] * x[acolidx[i]]; in matmult_seqsell_tiled_kernel9()
141 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmultadd_seqsell_tiled_kernel9() argument
153 …IZE * threadIdx.y; i < sliidx[slice_id + 1]; i += WARP_SIZE * BLOCKY) t += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_tiled_kernel9()
189 …onst PetscInt *chunk_slice_map, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmult_seqsell_tiled_kernel8() argument
213 if (row < nrows && gid < totalentries) t = aval[gid] * x[acolidx[gid]]; in matmult_seqsell_tiled_kernel8()
220 if (row < nrows) t += aval[gid] * x[acolidx[gid]]; in matmult_seqsell_tiled_kernel8()
244 …onst PetscInt *chunk_slice_map, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmultadd_seqsell_tiled_kernel8() argument
268 if (row < nrows && gid < totalentries) t = aval[gid] * x[acolidx[gid]]; in matmultadd_seqsell_tiled_kernel8()
275 if (row < nrows) t += aval[gid] * x[acolidx[gid]]; in matmultadd_seqsell_tiled_kernel8()
298 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmult_seqsell_tiled_kernel7() argument
305 …for (i = sliidx[slice_id] + threadIdx.x; i < sliidx[slice_id + 1]; i += WARP_SIZE) t += aval[i] * … in matmult_seqsell_tiled_kernel7()
313 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmultadd_seqsell_tiled_kernel7() argument
320 …for (i = sliidx[slice_id] + threadIdx.x; i < sliidx[slice_id + 1]; i += WARP_SIZE) t += aval[i] * … in matmultadd_seqsell_tiled_kernel7()
331 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in PETSC_PRAGMA_DIAGNOSTIC_IGNORED_END()
342 …liceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]]; in PETSC_PRAGMA_DIAGNOSTIC_IGNORED_END()
359 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmult_seqsell_tiled_kernel5() argument
370 …liceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]]; in matmult_seqsell_tiled_kernel5()
385 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmult_seqsell_tiled_kernel4() argument
396 …liceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]]; in matmult_seqsell_tiled_kernel4()
409 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmult_seqsell_tiled_kernel3() argument
420 …liceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]]; in matmult_seqsell_tiled_kernel3()
431 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmult_seqsell_tiled_kernel2() argument
442 …liceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]]; in matmult_seqsell_tiled_kernel2()
451 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmultadd_seqsell_tiled_kernel6() argument
462 …liceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_tiled_kernel6()
479 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmultadd_seqsell_tiled_kernel5() argument
490 …liceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_tiled_kernel5()
505 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmultadd_seqsell_tiled_kernel4() argument
516 …liceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_tiled_kernel4()
529 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmultadd_seqsell_tiled_kernel3() argument
540 …liceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_tiled_kernel3()
551 …nt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar *aval, const PetscInt *sl… in matmultadd_seqsell_tiled_kernel2() argument
562 …liceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_tiled_kernel2()
578 MatScalar *aval; in MatMult_SeqSELLHIP() local
593 aval = hipstruct->val; in MatMult_SeqSELLHIP()
606 …tiled_kernel9<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
608 …tiled_kernel9<4><<<nblocks, dim3(WARP_SIZE, 4)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
610 …tiled_kernel9<8><<<nblocks, dim3(WARP_SIZE, 8)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
612 …led_kernel9<16><<<nblocks, dim3(WARP_SIZE, 16)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
614 …tiled_kernel9<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
620 …ll_tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
622 …ll_tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 4)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
624 …ll_tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 8)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
626 …l_tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 16)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
629 …ll_tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
635 …matmult_seqsell_tiled_kernel6<<<nblocks, block32>>>(nrows, sliceheight, acolidx, aval, sliidx, x, … in MatMult_SeqSELLHIP()
639 …matmult_seqsell_tiled_kernel5<<<nblocks, block16>>>(nrows, sliceheight, acolidx, aval, sliidx, x, … in MatMult_SeqSELLHIP()
643 …matmult_seqsell_tiled_kernel4<<<nblocks, block8>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y… in MatMult_SeqSELLHIP()
647 …matmult_seqsell_tiled_kernel3<<<nblocks, block4>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y… in MatMult_SeqSELLHIP()
651 …matmult_seqsell_tiled_kernel2<<<nblocks, block2>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y… in MatMult_SeqSELLHIP()
655 …matmult_seqsell_basic_kernel<<<nblocks, blocksize>>>(nrows, sliceheight, acolidx, aval, sliidx, x,… in MatMult_SeqSELLHIP()
668 …, 2)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
670 …, 4)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
672 …, 8)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
674 … 16)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
676 …, 2)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
683 …ll_tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
686 …tiled_kernel9<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
690 …tiled_kernel9<8><<<nblocks, dim3(WARP_SIZE, 8)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
693 …led_kernel9<16><<<nblocks, dim3(WARP_SIZE, 16)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y); in MatMult_SeqSELLHIP()
715 MatScalar *aval = hipstruct->val; in MatMultAdd_SeqSELLHIP() local
741 …ed_kernel9<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
743 …ed_kernel9<4><<<nblocks, dim3(WARP_SIZE, 4)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
745 …ed_kernel9<8><<<nblocks, dim3(WARP_SIZE, 8)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
747 …_kernel9<16><<<nblocks, dim3(WARP_SIZE, 16)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
749 …ed_kernel9<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
760 …)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
762 …)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
764 …)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
766 …)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
768 …)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
774 …tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
776 …tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 4)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
778 …tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 8)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
780 …iled_kernel7<<<nblocks, dim3(WARP_SIZE, 16)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
783 …tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
789 …matmultadd_seqsell_tiled_kernel6<<<nblocks, block32>>>(nrows, sliceheight, acolidx, aval, sliidx, … in MatMultAdd_SeqSELLHIP()
793 …matmultadd_seqsell_tiled_kernel5<<<nblocks, block16>>>(nrows, sliceheight, acolidx, aval, sliidx, … in MatMultAdd_SeqSELLHIP()
797 …matmultadd_seqsell_tiled_kernel4<<<nblocks, block8>>>(nrows, sliceheight, acolidx, aval, sliidx, x… in MatMultAdd_SeqSELLHIP()
801 …matmultadd_seqsell_tiled_kernel3<<<nblocks, block4>>>(nrows, sliceheight, acolidx, aval, sliidx, x… in MatMultAdd_SeqSELLHIP()
805 …matmultadd_seqsell_tiled_kernel2<<<nblocks, block2>>>(nrows, sliceheight, acolidx, aval, sliidx, x… in MatMultAdd_SeqSELLHIP()
809 …matmultadd_seqsell_basic_kernel<<<nblocks, blocksize>>>(nrows, sliceheight, acolidx, aval, sliidx,… in MatMultAdd_SeqSELLHIP()
822 …)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
824 …)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
826 …)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
828 …)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
830 …)>>>(nrows, sliceheight, chunksperblock, nchunks, chunk_slice_map, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
837 …tiled_kernel7<<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
840 …ed_kernel9<2><<<nblocks, dim3(WARP_SIZE, 2)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
844 …ed_kernel9<8><<<nblocks, dim3(WARP_SIZE, 8)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()
847 …_kernel9<16><<<nblocks, dim3(WARP_SIZE, 16)>>>(nrows, sliceheight, acolidx, aval, sliidx, x, y, z); in MatMultAdd_SeqSELLHIP()