Lines Matching refs:dst1
120 // for size >= 16), and multiply-accumulate into dst1 and dst5 (or
121 // dst1-dst2 and dst5-dst6 for size >= 8 and dst1-dst4 and dst5-dst8
123 .macro extmlal dst1, dst2, dst3, dst4, dst5, dst6, dst7, dst8, src1, src2, src3, src4, src5, src6, offset, size
126 smlal \dst1\().4s, v20.4h, v0.h[\offset]
386 // Evaluate the filter twice in parallel, from the inputs src1-src9 into dst1-dst2
387 // (src1-src8 into dst1, src2-src9 into dst2).
388 .macro convolve4 dst1, dst2, src1, src2, src3, src4, src5, src6, src7, src8, src9, tmp1, tmp2
389 smull \dst1\().4s, \src1\().4h, v0.h[0]
393 smlal \dst1\().4s, \src3\().4h, v0.h[2]
397 smlal \dst1\().4s, \src5\().4h, v0.h[4]
401 smlal \dst1\().4s, \src7\().4h, v0.h[6]
405 add \dst1\().4s, \dst1\().4s, \tmp1\().4s
409 // Evaluate the filter twice in parallel, from the inputs src1-src9 into dst1-dst4
410 // (src1-src8 into dst1-dst2, src2-src9 into dst3-dst4).
411 .macro convolve8 dst1, dst2, dst3, dst4, src1, src2, src3, src4, src5, src6, src7, src8, src9
412 smull \dst1\().4s, \src1\().4h, v0.h[0]
416 smlal \dst1\().4s, \src2\().4h, v0.h[1]
420 smlal \dst1\().4s, \src3\().4h, v0.h[2]
424 smlal \dst1\().4s, \src4\().4h, v0.h[3]
428 smlal \dst1\().4s, \src5\().4h, v0.h[4]
432 smlal \dst1\().4s, \src6\().4h, v0.h[5]
436 smlal \dst1\().4s, \src7\().4h, v0.h[6]
440 smlal \dst1\().4s, \src8\().4h, v0.h[7]