📄 gemv_n.s
字号:
LDF [A2 + 1 * SIZE], a6 LDF [A3 + 1 * SIZE], a7 LDF [A4 + 1 * SIZE], a8 LDF [Y1 + 1 * SIZE], y2 add A1, 2 * SIZE, A1 add A2, 2 * SIZE, A2 add A3, 2 * SIZE, A3 add A4, 2 * SIZE, A4 FMUL a1, x1, t1 FMUL a2, x2, t2 FMUL a3, x3, t3 FMUL a4, x4, t4 FADD y1, t1, y1 FMUL a5, x1, t1 FADD y1, t2, y1 FMUL a6, x2, t2 FADD y1, t3, y1 FMUL a7, x3, t3 FADD y1, t4, y1 FMUL a8, x4, t4 FADD y2, t1, y2 FADD y2, t2, y2 FADD y2, t3, y2 FADD y2, t4, y2 STF y1, [Y1 + 0 * SIZE] STF y2, [Y1 + 1 * SIZE] add Y1, 2 * SIZE, Y1.LL18: andcc M, 1, I ble,pn %icc, .LL19 nop LDF [A1 + 0 * SIZE], a1 LDF [A2 + 0 * SIZE], a2 LDF [A3 + 0 * SIZE], a3 LDF [A4 + 0 * SIZE], a4 LDF [Y1 + 0 * SIZE], y1 FMUL a1, x1, t1 FMUL a2, x2, t2 FMUL a3, x3, t3 FMUL a4, x4, t4 FADD y1, t1, y1 FADD y1, t2, y1 FADD y1, t3, y1 FADD y1, t4, y1 STF y1, [Y1].LL19: deccc J bg %icc, .LL11 nop.LL20: andcc N, 2, J ble,pn %icc, .LL30 nop.LL21: mov YY, Y1 mov A, A1 add A, LDA, A2 add A2, LDA, A LDF STACK_ALPHA, ALPHA LDF [X], x1 add X, INCX, X LDF [X], x2 add X, INCX, X FMUL ALPHA, x1, x1 FMUL ALPHA, x2, x2 sra M, 3, I cmp I, 0 ble,pn %icc, .LL26 nop LDF [Y1 + 0 * SIZE], y1 LDF [Y1 + 1 * SIZE], y2 LDF [Y1 + 2 * SIZE], y3 LDF [Y1 + 3 * SIZE], y4 LDF [Y1 + 4 * SIZE], y5 LDF [Y1 + 5 * SIZE], y6 LDF [Y1 + 6 * SIZE], y7 LDF [Y1 + 7 * SIZE], y8 LDF [A1 + 0 * SIZE], a1 LDF [A1 + 1 * SIZE], a2 LDF [A1 + 2 * SIZE], a3 LDF [A1 + 3 * SIZE], a4 LDF [A1 + 4 * SIZE], a5 LDF [A1 + 5 * SIZE], a6 LDF [A1 + 6 * SIZE], a7 LDF [A1 + 7 * SIZE], a8 LDF [A2 + 0 * SIZE], a9 LDF [A2 + 1 * SIZE], a10 LDF [A2 + 2 * SIZE], a11 LDF [A2 + 3 * SIZE], a12 LDF [A2 + 4 * SIZE], a13 LDF [A2 + 5 * SIZE], a14 LDF [A2 + 6 * SIZE], a15 LDF [A2 + 7 * SIZE], a16 FMUL a1, x1, t1 deccc I LDF [A1 + 8 * SIZE], a1 FMUL a2, x1, t2 LDF [A1 + 9 * SIZE], a2 FMUL a3, x1, t3 LDF [A1 + 10 * SIZE], a3 FMUL a4, x1, t4 ble,pn %icc, .LL23 LDF [A1 + 11 * SIZE], a4.LL22: FADD y1, t1, y1 prefetch [A1 + PREFETCHSIZE * SIZE], 1 FMUL a5, x1, t1 LDF [A1 + 12 * SIZE], a5 FADD y2, t2, y2 FMUL a6, x1, t2 LDF [A1 + 13 * SIZE], a6 FADD y3, t3, y3 FMUL a7, x1, t3 LDF [A1 + 14 * SIZE], a7 FADD y4, t4, y4 FMUL a8, x1, t4 LDF [A1 + 15 * SIZE], a8 FADD y5, t1, y5 FMUL a9, x2, t1 LDF [A2 + 8 * SIZE], a9 FADD y6, t2, y6 FMUL a10, x2, t2 LDF [A2 + 9 * SIZE], a10 FADD y7, t3, y7 FMUL a11, x2, t3 LDF [A2 + 10 * SIZE], a11 FADD y8, t4, y8 FMUL a12, x2, t4 LDF [A2 + 11 * SIZE], a12 FADD y1, t1, y1 prefetch [A2 + PREFETCHSIZE * SIZE], 1 FMUL a13, x2, t1 LDF [A2 + 12 * SIZE], a13 FADD y2, t2, y2 FMUL a14, x2, t2 LDF [A2 + 13 * SIZE], a14 FADD y3, t3, y3 FMUL a15, x2, t3 LDF [A2 + 14 * SIZE], a15 FADD y4, t4, y4 FMUL a16, x2, t4 LDF [A2 + 15 * SIZE], a16 FADD y5, t1, y5 FMUL a1, x1, t1 LDF [A1 + 16 * SIZE], a1 FADD y6, t2, y6 FMUL a2, x1, t2 LDF [A1 + 17 * SIZE], a2 FADD y7, t3, y7 FMUL a3, x1, t3 LDF [A1 + 18 * SIZE], a3 FADD y8, t4, y8 FMUL a4, x1, t4 LDF [A1 + 19 * SIZE], a4 STF y1, [Y1 + 0 * SIZE] STF y2, [Y1 + 1 * SIZE] STF y3, [Y1 + 2 * SIZE] STF y4, [Y1 + 3 * SIZE] STF y5, [Y1 + 4 * SIZE] STF y6, [Y1 + 5 * SIZE] STF y7, [Y1 + 6 * SIZE] STF y8, [Y1 + 7 * SIZE] LDF [Y1 + 8 * SIZE], y1 add A1, 8 * SIZE, A1 LDF [Y1 + 9 * SIZE], y2 add A2, 8 * SIZE, A2 LDF [Y1 + 10 * SIZE], y3 deccc I LDF [Y1 + 11 * SIZE], y4 LDF [Y1 + 12 * SIZE], y5 LDF [Y1 + 13 * SIZE], y6 LDF [Y1 + 14 * SIZE], y7 LDF [Y1 + 15 * SIZE], y8 bg,pn %icc, .LL22 add Y1, 8 * SIZE, Y1.LL23: FADD y1, t1, y1 FMUL a5, x1, t1 FADD y2, t2, y2 FMUL a6, x1, t2 FADD y3, t3, y3 FMUL a7, x1, t3 FADD y4, t4, y4 FMUL a8, x1, t4 FADD y5, t1, y5 FMUL a9, x2, t1 FADD y6, t2, y6 FMUL a10, x2, t2 FADD y7, t3, y7 FMUL a11, x2, t3 FADD y8, t4, y8 FMUL a12, x2, t4 FADD y1, t1, y1 FMUL a13, x2, t1 FADD y2, t2, y2 FMUL a14, x2, t2 FADD y3, t3, y3 FMUL a15, x2, t3 FADD y4, t4, y4 FMUL a16, x2, t4 STF y1, [Y1 + 0 * SIZE] FADD y5, t1, y5 STF y2, [Y1 + 1 * SIZE] FADD y6, t2, y6 STF y3, [Y1 + 2 * SIZE] FADD y7, t3, y7 STF y4, [Y1 + 3 * SIZE] FADD y8, t4, y8 STF y5, [Y1 + 4 * SIZE] add A1, 8 * SIZE, A1 STF y6, [Y1 + 5 * SIZE] add A2, 8 * SIZE, A2 STF y7, [Y1 + 6 * SIZE] nop STF y8, [Y1 + 7 * SIZE] add Y1, 8 * SIZE, Y1.LL26: andcc M, 4, I ble,pn %icc, .LL27 nop LDF [A1 + 0 * SIZE], a1 LDF [A1 + 1 * SIZE], a2 LDF [A1 + 2 * SIZE], a3 LDF [A1 + 3 * SIZE], a4 LDF [A2 + 0 * SIZE], a5 LDF [A2 + 1 * SIZE], a6 LDF [A2 + 2 * SIZE], a7 LDF [A2 + 3 * SIZE], a8 LDF [Y1 + 0 * SIZE], y1 add A1, 4 * SIZE, A1 LDF [Y1 + 1 * SIZE], y2 add A2, 4 * SIZE, A2 LDF [Y1 + 2 * SIZE], y3 LDF [Y1 + 3 * SIZE], y4 FMUL a1, x1, t1 FMUL a2, x1, t2 FMUL a3, x1, t3 FMUL a4, x1, t4 FADD y1, t1, y1 FMUL a5, x2, t1 FADD y2, t2, y2 FMUL a6, x2, t2 FADD y3, t3, y3 FMUL a7, x2, t3 FADD y4, t4, y4 FMUL a8, x2, t4 FADD y1, t1, y1 FADD y2, t2, y2 FADD y3, t3, y3 FADD y4, t4, y4 STF y1, [Y1 + 0 * SIZE] STF y2, [Y1 + 1 * SIZE] STF y3, [Y1 + 2 * SIZE] STF y4, [Y1 + 3 * SIZE] add Y1, 4 * SIZE, Y1.LL27: andcc M, 2, I ble,pn %icc, .LL28 nop LDF [A1 + 0 * SIZE], a1 LDF [A2 + 0 * SIZE], a2 LDF [Y1 + 0 * SIZE], y1 LDF [A1 + 1 * SIZE], a5 LDF [A2 + 1 * SIZE], a6 add A1, 2 * SIZE, A1 LDF [Y1 + 1 * SIZE], y2 add A2, 2 * SIZE, A2 FMUL a1, x1, t1 FMUL a2, x2, t2 FADD y1, t1, y1 FMUL a5, x1, t1 FADD y1, t2, y1 FMUL a6, x2, t2 FADD y2, t1, y2 FADD y2, t2, y2 STF y1, [Y1 + 0 * SIZE] STF y2, [Y1 + 1 * SIZE] add Y1, 2 * SIZE, Y1.LL28: andcc M, 1, I ble,pn %icc, .LL30 nop LDF [A1 + 0 * SIZE], a1 LDF [A2 + 0 * SIZE], a2 LDF [Y1 + 0 * SIZE], y1 FMUL a1, x1, t1 FMUL a2, x2, t2 FADD y1, t1, y1 FADD y1, t2, y1 STF y1, [Y1].LL30: andcc N, 1, J ble,pn %icc, .LL990 nop.LL31: mov YY, Y1 mov A, A1 LDF STACK_ALPHA, ALPHA LDF [X], x1 add X, INCX, X FMUL ALPHA, x1, x1 sra M, 3, I cmp I, 0 ble,pn %icc, .LL36 nop LDF [Y1 + 0 * SIZE], y1 LDF [Y1 + 1 * SIZE], y2 LDF [Y1 + 2 * SIZE], y3 LDF [Y1 + 3 * SIZE], y4 LDF [Y1 + 4 * SIZE], y5 LDF [Y1 + 5 * SIZE], y6 LDF [Y1 + 6 * SIZE], y7 LDF [Y1 + 7 * SIZE], y8 LDF [A1 + 0 * SIZE], a1 LDF [A1 + 1 * SIZE], a2 LDF [A1 + 2 * SIZE], a3 LDF [A1 + 3 * SIZE], a4 LDF [A1 + 4 * SIZE], a5 LDF [A1 + 5 * SIZE], a6 LDF [A1 + 6 * SIZE], a7 LDF [A1 + 7 * SIZE], a8 FMUL a1, x1, t1 deccc I LDF [A1 + 8 * SIZE], a1 FMUL a2, x1, t2 LDF [A1 + 9 * SIZE], a2 FMUL a3, x1, t3 LDF [A1 + 10 * SIZE], a3 FMUL a4, x1, t4 ble,pn %icc, .LL33 LDF [A1 + 11 * SIZE], a4.LL32: FADD y1, t1, y1 prefetch [A1 + PREFETCHSIZE * SIZE], 1 FMUL a5, x1, t1 LDF [A1 + 12 * SIZE], a5 FADD y2, t2, y2 FMUL a6, x1, t2 LDF [A1 + 13 * SIZE], a6 FADD y3, t3, y3 FMUL a7, x1, t3 LDF [A1 + 14 * SIZE], a7 FADD y4, t4, y4 FMUL a8, x1, t4 LDF [A1 + 15 * SIZE], a8 FADD y5, t1, y5 FMUL a1, x1, t1 LDF [A1 + 16 * SIZE], a1 FADD y6, t2, y6 FMUL a2, x1, t2 LDF [A1 + 17 * SIZE], a2 FADD y7, t3, y7 FMUL a3, x1, t3 LDF [A1 + 18 * SIZE], a3 FADD y8, t4, y8 FMUL a4, x1, t4 LDF [A1 + 19 * SIZE], a4 STF y1, [Y1 + 0 * SIZE] STF y2, [Y1 + 1 * SIZE] STF y3, [Y1 + 2 * SIZE] STF y4, [Y1 + 3 * SIZE] STF y5, [Y1 + 4 * SIZE] STF y6, [Y1 + 5 * SIZE] STF y7, [Y1 + 6 * SIZE] STF y8, [Y1 + 7 * SIZE] LDF [Y1 + 8 * SIZE], y1 LDF [Y1 + 9 * SIZE], y2 LDF [Y1 + 10 * SIZE], y3 LDF [Y1 + 11 * SIZE], y4 LDF [Y1 + 12 * SIZE], y5 deccc I LDF [Y1 + 13 * SIZE], y6 add A1, 8 * SIZE, A1 LDF [Y1 + 14 * SIZE], y7 add Y1, 8 * SIZE, Y1 bg,pn %icc, .LL32 LDF [Y1 + 7 * SIZE], y8.LL33: FADD y1, t1, y1 FMUL a5, x1, t1 FADD y2, t2, y2 FMUL a6, x1, t2 FADD y3, t3, y3 FMUL a7, x1, t3 FADD y4, t4, y4 FMUL a8, x1, t4 STF y1, [Y1 + 0 * SIZE] FADD y5, t1, y5 STF y2, [Y1 + 1 * SIZE] FADD y6, t2, y6 STF y3, [Y1 + 2 * SIZE] FADD y7, t3, y7 STF y4, [Y1 + 3 * SIZE] FADD y8, t4, y8 STF y5, [Y1 + 4 * SIZE] STF y6, [Y1 + 5 * SIZE] STF y7, [Y1 + 6 * SIZE] add A1, 8 * SIZE, A1 STF y8, [Y1 + 7 * SIZE] add Y1, 8 * SIZE, Y1.LL36: andcc M, 4, I ble,pn %icc, .LL37 nop LDF [A1 + 0 * SIZE], a1 LDF [A1 + 1 * SIZE], a2 LDF [A1 + 2 * SIZE], a3 LDF [A1 + 3 * SIZE], a4 LDF [Y1 + 0 * SIZE], y1 add A1, 4 * SIZE, A1 LDF [Y1 + 1 * SIZE], y2 LDF [Y1 + 2 * SIZE], y3 LDF [Y1 + 3 * SIZE], y4 FMUL a1, x1, t1 FMUL a2, x1, t2 FMUL a3, x1, t3 FMUL a4, x1, t4 FADD y1, t1, y1 FADD y2, t2, y2 FADD y3, t3, y3 FADD y4, t4, y4 STF y1, [Y1 + 0 * SIZE] STF y2, [Y1 + 1 * SIZE] STF y3, [Y1 + 2 * SIZE] STF y4, [Y1 + 3 * SIZE] add Y1, 4 * SIZE, Y1.LL37: andcc M, 2, I ble,pn %icc, .LL38 nop LDF [A1 + 0 * SIZE], a1 LDF [Y1 + 0 * SIZE], y1 LDF [A1 + 1 * SIZE], a5 LDF [Y1 + 1 * SIZE], y2 add A1, 2 * SIZE, A1 FMUL a1, x1, t1 FADD y1, t1, y1 FMUL a5, x1, t1 FADD y2, t1, y2 STF y1, [Y1 + 0 * SIZE] STF y2, [Y1 + 1 * SIZE] add Y1, 2 * SIZE, Y1.LL38: andcc M, 1, I ble,pn %icc, .LL990 nop LDF [A1 + 0 * SIZE], a1 LDF [Y1 + 0 * SIZE], y1 FMUL a1, x1, t1 FADD y1, t1, y1 STF y1, [Y1].LL990: cmp INCY, SIZE be %icc, .LL999 mov Y, Y1 sra M, 3, I cmp I, 0 ble,pn %icc, .LL995 nop.LL991: LDF [BUFFER + 0 * SIZE], a1 LDF [Y], y1 add Y, INCY, Y LDF [BUFFER + 1 * SIZE], a2 LDF [Y], y2 add Y, INCY, Y LDF [BUFFER + 2 * SIZE], a3 LDF [Y], y3 add Y, INCY, Y LDF [BUFFER + 3 * SIZE], a4 LDF [Y], y4 add Y, INCY, Y LDF [BUFFER + 4 * SIZE], a5 FADD y1, a1, y1 LDF [Y], y5 add Y, INCY, Y LDF [BUFFER + 5 * SIZE], a6 FADD y2, a2, y2 LDF [Y], y6 add Y, INCY, Y LDF [BUFFER + 6 * SIZE], a7 FADD y3, a3, y3 LDF [Y], y7 add Y, INCY, Y LDF [BUFFER + 7 * SIZE], a8 FADD y4, a4, y4 LDF [Y], y8 add Y, INCY, Y STF y1, [Y1] FADD y5, a5, y5 add Y1, INCY, Y1 STF y2, [Y1] FADD y6, a6, y6 add Y1, INCY, Y1 STF y3, [Y1] FADD y7, a7, y7 add Y1, INCY, Y1 STF y4, [Y1] FADD y8, a8, y8 add Y1, INCY, Y1 STF y5, [Y1] add Y1, INCY, Y1 STF y6, [Y1] add Y1, INCY, Y1 STF y7, [Y1] add Y1, INCY, Y1 STF y8, [Y1] add Y1, INCY, Y1 deccc I bg,pn %icc, .LL991 add BUFFER, 8 * SIZE, BUFFER.LL995: andcc M, 7, I ble,pn %icc, .LL999 nop andcc M, 4, I ble,pn %icc, .LL996 nop LDF [BUFFER + 0 * SIZE], a1 LDF [BUFFER + 1 * SIZE], a2 LDF [BUFFER + 2 * SIZE], a3 LDF [BUFFER + 3 * SIZE], a4 add BUFFER, 4 * SIZE, BUFFER LDF [Y], y1 add Y, INCY, Y LDF [Y], y2 add Y, INCY, Y LDF [Y], y3 add Y, INCY, Y LDF [Y], y4 add Y, INCY, Y FADD y1, a1, y1 FADD y2, a2, y2 FADD y3, a3, y3 FADD y4, a4, y4 STF y1, [Y1] add Y1, INCY, Y1 STF y2, [Y1] add Y1, INCY, Y1 STF y3, [Y1] add Y1, INCY, Y1 STF y4, [Y1] add Y1, INCY, Y1 .LL996: andcc M, 2, I ble,pn %icc, .LL997 nop LDF [BUFFER + 0 * SIZE], a1 LDF [BUFFER + 1 * SIZE], a2 add BUFFER, 2 * SIZE, BUFFER LDF [Y], y1 add Y, INCY, Y LDF [Y], y2 add Y, INCY, Y FADD y1, a1, y1 FADD y2, a2, y2 STF y1, [Y1] add Y1, INCY, Y1 STF y2, [Y1] add Y1, INCY, Y1 .LL997: andcc M, 1, I ble,pn %icc, .LL999 nop LDF [BUFFER + 0 * SIZE], a1 LDF [Y], y1 FADD y1, a1, y1 STF y1, [Y1].LL999: return %i7 + 8 clr %o0 EPILOGUE
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -