Searched refs:TRANSPOSE_TILE_DIM (Results 1 – 1 of 1) sorted by relevance
1098 const Ncv32u TRANSPOSE_TILE_DIM = 16; variable1122 __shared__ T tile[TRANSPOSE_TILE_DIM][TRANSPOSE_TILE_DIM+1]; in transpose()1139 Ncv32u xIndex = blockIdx_x * TRANSPOSE_TILE_DIM + threadIdx.x; in transpose()1140 Ncv32u yIndex = blockIdx_y * TRANSPOSE_TILE_DIM + threadIdx.y; in transpose()1145 for (Ncv32u i=0; i<TRANSPOSE_TILE_DIM; i+=TRANSPOSE_BLOCK_ROWS) in transpose()1156 xIndex = blockIdx_y * TRANSPOSE_TILE_DIM + threadIdx.x; in transpose()1157 yIndex = blockIdx_x * TRANSPOSE_TILE_DIM + threadIdx.y; in transpose()1162 for (Ncv32u i=0; i<TRANSPOSE_TILE_DIM; i+=TRANSPOSE_BLOCK_ROWS) in transpose()1185 dim3 grid((srcRoi.width + TRANSPOSE_TILE_DIM - 1) / TRANSPOSE_TILE_DIM, in transposeWrapperDevice()1186 (srcRoi.height + TRANSPOSE_TILE_DIM - 1) / TRANSPOSE_TILE_DIM); in transposeWrapperDevice()[all …]