74 #define SWAP_ROWS_F16(A,COL,i,j)                  \  argument
87        tmpb=vldrhq_z_f16(&data[j*_numCols + _w],p0);\
90        vstrhq_p(&data[j*_numCols + _w], tmpa, p0);  \
115 #define MAC_ROW_F16(COL,A,i,v,B,j)                   \  argument
128        tmpb = vldrhq_z_f16(&dataB[j*_numCols + _w],p0);\
136 #define MAS_ROW_F16(COL,A,i,v,B,j)                   \  argument
150        tmpb = vldrhq_z_f16(&dataB[j*_numCols + _w],p0);\
161 #define SWAP_ROWS_F16(A,COL,i,j)       \
170   dataJ += j*_numCols + (COL);          \
197 #define MAC_ROW_F16(COL,A,i,v,B,j)                \
206   dataB += j*_numCols + (COL);                     \
214 #define MAS_ROW_F16(COL,A,i,v,B,j)                \
223   dataB += j*_numCols + (COL);                     \
244 #define SWAP_ROWS_F32(A,COL,i,j)                  \  argument
257        tmpb=vldrwq_z_f32(&data[j*_numCols + _w],p0);\
260        vstrwq_p(&data[j*_numCols + _w], tmpa, p0);  \
266 #define MAC_ROW_F32(COL,A,i,v,B,j)                   \  argument
279        tmpb = vldrwq_z_f32(&dataB[j*_numCols + _w],p0);\
287 #define MAS_ROW_F32(COL,A,i,v,B,j)                   \  argument
301        tmpb = vldrwq_z_f32(&dataB[j*_numCols + _w],p0);\
330 #define SWAP_ROWS_F32(A,COL,i,j)       \
339   dataJ += j*_numCols + (COL);          \
351 #define MAC_ROW_F32(COL,A,i,v,B,j)     \
363   dataB += j*_numCols + (COL);          \
385 #define MAS_ROW_F32(COL,A,i,v,B,j)     \
397   dataB += j*_numCols + (COL);          \
451 #define SWAP_ROWS_F32(A,COL,i,j)       \
461   dataJ += j*_numCols + (COL);          \
488 #define MAC_ROW_F32(COL,A,i,v,B,j)     \
497   dataB = dataB + j*_numCols + (COL);   \
505 #define MAS_ROW_F32(COL,A,i,v,B,j)     \
514   dataB = dataB + j*_numCols + (COL);   \
533 #define SWAP_COLS_F32(A,COL,i,j)               \  argument
542      data[_w*_numCols + i] = data[_w*_numCols + j];\
543      data[_w*_numCols + j] = tmp;                \
550 #define SWAP_ROWS_F64(A,COL,i,j)       \  argument
559   dataJ += j*_numCols + (COL);          \
570 #define SWAP_COLS_F64(A,COL,i,j)               \  argument
579      data[_w*_numCols + i] = data[_w*_numCols + j];\
580      data[_w*_numCols + j] = tmp;                \
602 #define MAC_ROW_F64(COL,A,i,v,B,j)      \  argument
611   dataB += j*_numCols + (COL);           \
619 #define MAS_ROW_F64(COL,A,i,v,B,j)      \  argument
628   dataB += j*_numCols + (COL);           \