seq/seqhip/sellhip.hip.cxx

80   PetscInt  i, row, slice_id, row_in_slice;
86     row_in_slice = row % sliceheight;
88     for (i = sliidx[slice_id] + row_in_slice; i < sliidx[slice_id + 1]; i += sliceheight) sum += aval[i] * x[acolidx[i]];
95   PetscInt  i, row, slice_id, row_in_slice;
101     row_in_slice = row % sliceheight;
103     for (i = sliidx[slice_id] + row_in_slice; i < sliidx[slice_id + 1]; i += sliceheight) sum += aval[i] * x[acolidx[i]];
334   PetscInt             i, row, slice_id, row_in_slice;
339     row_in_slice = row % sliceheight;
342     for (i = sliidx[slice_id] + row_in_slice + sliceheight * threadIdx.y; i < sliidx[slice_id + 1]; i += sliceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]];
362   PetscInt             i, row, slice_id, row_in_slice;
367     row_in_slice = row % sliceheight;
370     for (i = sliidx[slice_id] + row_in_slice + sliceheight * threadIdx.y; i < sliidx[slice_id + 1]; i += sliceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]];
388   PetscInt             i, row, slice_id, row_in_slice;
393     row_in_slice = row % sliceheight;
396     for (i = sliidx[slice_id] + row_in_slice + sliceheight * threadIdx.y; i < sliidx[slice_id + 1]; i += sliceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]];
412   PetscInt             i, row, slice_id, row_in_slice;
417     row_in_slice = row % sliceheight;
420     for (i = sliidx[slice_id] + row_in_slice + sliceheight * threadIdx.y; i < sliidx[slice_id + 1]; i += sliceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]];
434   PetscInt             i, row, slice_id, row_in_slice;
439     row_in_slice = row % sliceheight;
442     for (i = sliidx[slice_id] + row_in_slice + sliceheight * threadIdx.y; i < sliidx[slice_id + 1]; i += sliceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]];
454   PetscInt             i, row, slice_id, row_in_slice;
459     row_in_slice = row % sliceheight;
462     for (i = sliidx[slice_id] + row_in_slice + sliceheight * threadIdx.y; i < sliidx[slice_id + 1]; i += sliceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]];
482   PetscInt             i, row, slice_id, row_in_slice;
487     row_in_slice = row % sliceheight;
490     for (i = sliidx[slice_id] + row_in_slice + sliceheight * threadIdx.y; i < sliidx[slice_id + 1]; i += sliceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]];
508   PetscInt             i, row, slice_id, row_in_slice;
513     row_in_slice = row % sliceheight;
516     for (i = sliidx[slice_id] + row_in_slice + sliceheight * threadIdx.y; i < sliidx[slice_id + 1]; i += sliceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]];
532   PetscInt             i, row, slice_id, row_in_slice;
537     row_in_slice = row % sliceheight;
540     for (i = sliidx[slice_id] + row_in_slice + sliceheight * threadIdx.y; i < sliidx[slice_id + 1]; i += sliceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]];
554   PetscInt             i, row, slice_id, row_in_slice;
559     row_in_slice = row % sliceheight;
562     for (i = sliidx[slice_id] + row_in_slice + sliceheight * threadIdx.y; i < sliidx[slice_id + 1]; i += sliceheight * blockDim.y) shared[threadIdx.y * blockDim.x + threadIdx.x] += aval[i] * x[acolidx[i]];