📄 znrm2_hummer.s
字号:
lfpdux f18, SP, r10 lfpdux f17, SP, r10 lfpdux f16, SP, r10 lfpdux f15, SP, r10 lfpdux f14, SP, r10 fmul C1, ALPHA, C1 addi SP, SP, 16 blr#endif .align 4LL(99): li r10, 16 lfpdux f29, SP, r10 lfpdux f28, SP, r10 lfpdux f27, SP, r10 lfpdux f26, SP, r10 lfpdux f25, SP, r10 lfpdux f24, SP, r10 lfpdux f23, SP, r10 lfpdux f22, SP, r10 lfpdux f21, SP, r10 lfpdux f20, SP, r10 lfpdux f19, SP, r10 lfpdux f18, SP, r10 lfpdux f17, SP, r10 lfpdux f16, SP, r10 lfpdux f15, SP, r10 lfpdux f14, SP, r10 addi SP, SP, 16 blr .align 4LL(100): sub X, X, INCX2 addi X2, X, SIZE srawi. r0, N, 3 mtspr CTR, r0 beq- LL(105) LFDUX A1, X, INCX2 LFDUX A2, X2, INCX2 LFDUX A3, X, INCX2 LFDUX A4, X2, INCX2 LFSDUX A1, X, INCX2 LFSDUX A2, X2, INCX2 LFSDUX A3, X, INCX2 LFSDUX A4, X2, INCX2 LFDUX A5, X, INCX2 LFDUX A6, X2, INCX2 LFDUX A7, X, INCX2 LFDUX A8, X2, INCX2 LFSDUX A5, X, INCX2 fpabs T1, A1 LFSDUX A6, X2, INCX2 fpabs T2, A2 LFSDUX A7, X, INCX2 fpabs T3, A3 LFSDUX A8, X2, INCX2 fpabs T4, A4 bdz LL(103) .align 4LL(102): fpsub F1, C1, T1 LFDUX A1, X, INCX2 fpsub F2, C2, T2 LFDUX A2, X2, INCX2 fpsub F3, C3, T3 LFDUX A3, X, INCX2 fpsub F4, C4, T4 LFDUX A4, X2, INCX2 fpabs T5, A5 LFSDUX A1, X, INCX2 fpabs T6, A6 LFSDUX A2, X2, INCX2 fpabs T7, A7 LFSDUX A3, X, INCX2 fpabs T8, A8 LFSDUX A4, X2, INCX2 fpsel C1, F1, C1, T1 LFDUX A5, X, INCX2 fpsel C2, F2, C2, T2 LFDUX A6, X2, INCX2 fpsel C3, F3, C3, T3 LFDUX A7, X, INCX2 fpsel C4, F4, C4, T4 LFDUX A8, X2, INCX2 fpsub F5, C1, T5 LFSDUX A5, X, INCX2 fpsub F6, C2, T6 LFSDUX A6, X2, INCX2 fpsub F7, C3, T7 LFSDUX A7, X, INCX2 fpsub F8, C4, T8 LFSDUX A8, X2, INCX2 fpabs T1, A1 fpabs T2, A2 fpabs T3, A3 fpabs T4, A4 fpsel C1, F5, C1, T5 fpsel C2, F6, C2, T6 fpsel C3, F7, C3, T7 fpsel C4, F8, C4, T8 bdnz LL(102) .align 4LL(103): fpabs T5, A5 fpabs T6, A6 fpabs T7, A7 fpabs T8, A8 fpsub F1, C1, T1 fpsub F2, C2, T2 fpsub F3, C3, T3 fpsub F4, C4, T4 fpsel C1, F1, C1, T1 fpsel C2, F2, C2, T2 fpsel C3, F3, C3, T3 fpsel C4, F4, C4, T4 fpsub F5, C1, T5 fpsub F6, C2, T6 fpsub F7, C3, T7 fpsub F8, C4, T8 fpsel C1, F5, C1, T5 fpsel C2, F6, C2, T6 fpsel C3, F7, C3, T7 fpsel C4, F8, C4, T8 .align 4LL(105): andi. r0, N, 7 beq LL(120) andi. r0, N, 4 beq LL(106) LFDUX A1, X, INCX2 LFDUX A2, X2, INCX2 LFDUX A3, X, INCX2 LFDUX A4, X2, INCX2 LFSDUX A1, X, INCX2 LFSDUX A2, X2, INCX2 LFSDUX A3, X, INCX2 LFSDUX A4, X2, INCX2 fpabs A1, A1 fpabs A2, A2 fpabs A3, A3 fpabs A4, A4 fpsub F1, C1, A1 fpsub F2, C2, A2 fpsub F3, C3, A3 fpsub F4, C4, A4 fpsel C1, F1, C1, A1 fpsel C2, F2, C2, A2 fpsel C3, F3, C3, A3 fpsel C4, F4, C4, A4 .align 4LL(106): andi. r0, N, 2 beq LL(107) LFDUX A1, X, INCX2 LFDUX A2, X2, INCX2 LFDUX A3, X, INCX2 LFDUX A4, X2, INCX2 fabs A1, A1 fabs A2, A2 fabs A3, A3 fabs A4, A4 fsub F1, C1, A1 fsub F2, C2, A2 fsub F3, C3, A3 fsub F4, C4, A4 fsel C1, F1, C1, A1 fsel C2, F2, C2, A2 fsel C3, F3, C3, A3 fsel C4, F4, C4, A4 .align 4LL(107): andi. r0, N, 1 beq LL(120) LFDUX A1, X, INCX2 LFDUX A2, X2, INCX2 fabs A1, A1 fabs A2, A2 fsub F1, C1, A1 fsub F2, C2, A2 fsel C1, F1, C1, A1 fsel C2, F2, C2, A2 .align 4LL(120): fpsub F1, C1, C2 fpsub F2, C3, C4 fpsel C1, F1, C1, C2 fpsel C3, F2, C3, C4 fpsub F1, C1, C3 fpsel C1, F1, C1, C3 fsmtp C2, C1 fsub F1, C1, C2 fsel ALPHA, F1, C1, C2 li r10, 0 lfs ALPHA_R, 8(SP) # load 1.0 fdiv ALPHA_R, ALPHA_R, ALPHA lfpsx C1, SP, r10 # Zero clear fpmr C2, C1 fpmr C3, C1 fpmr C4, C1 fsmfp ALPHA_R, ALPHA_R sub XX, XX, INCX2 addi X2, XX, SIZE srawi. r0, N, 3 mtspr CTR, r0 beq- LL(125) LFDUX A1, XX, INCX2 LFDUX A2, X2, INCX2 LFDUX A3, XX, INCX2 LFDUX A4, X2, INCX2 LFSDUX A1, XX, INCX2 LFSDUX A2, X2, INCX2 LFSDUX A3, XX, INCX2 LFSDUX A4, X2, INCX2 LFDUX A5, XX, INCX2 LFDUX A6, X2, INCX2 LFDUX A7, XX, INCX2 LFDUX A8, X2, INCX2 LFSDUX A5, XX, INCX2 fpmul T1, ALPHA_R, A1 LFSDUX A6, X2, INCX2 fpmul T2, ALPHA_R, A2 LFSDUX A7, XX, INCX2 fpmul T3, ALPHA_R, A3 LFSDUX A8, X2, INCX2 fpmul T4, ALPHA_R, A4 bdz LL(123) .align 4LL(122): fpmadd C1, T1, T1, C1 LFDUX A1, XX, INCX2 fpmul T1, ALPHA_R, A5 LFDUX A2, X2, INCX2 fpmadd C2, T2, T2, C2 LFDUX A3, XX, INCX2 fpmul T2, ALPHA_R, A6 LFDUX A4, X2, INCX2 fpmadd C3, T3, T3, C3 LFSDUX A1, XX, INCX2 fpmul T3, ALPHA_R, A7 LFSDUX A2, X2, INCX2 fpmadd C4, T4, T4, C4 LFSDUX A3, XX, INCX2 fpmul T4, ALPHA_R, A8 LFSDUX A4, X2, INCX2 fpmadd C1, T1, T1, C1 LFDUX A5, XX, INCX2 fpmul T1, ALPHA_R, A1 LFDUX A6, X2, INCX2 fpmadd C2, T2, T2, C2 LFDUX A7, XX, INCX2 fpmul T2, ALPHA_R, A2 LFDUX A8, X2, INCX2 fpmadd C3, T3, T3, C3 LFSDUX A5, XX, INCX2 fpmul T3, ALPHA_R, A3 LFSDUX A6, X2, INCX2 fpmadd C4, T4, T4, C4 LFSDUX A7, XX, INCX2 fpmul T4, ALPHA_R, A4 LFSDUX A8, X2, INCX2 bdnz LL(122) .align 4LL(123): fpmadd C1, T1, T1, C1 fpmul T1, ALPHA_R, A5 fpmadd C2, T2, T2, C2 fpmul T2, ALPHA_R, A6 fpmadd C3, T3, T3, C3 fpmul T3, ALPHA_R, A7 fpmadd C4, T4, T4, C4 fpmul T4, ALPHA_R, A8 fpmadd C1, T1, T1, C1 fpmadd C2, T2, T2, C2 fpmadd C3, T3, T3, C3 fpmadd C4, T4, T4, C4 .align 4LL(125): andi. r0, N, 7 beq LL(998) andi. r0, N, 4 beq LL(126) LFDUX A1, XX, INCX2 LFDUX A2, X2, INCX2 LFDUX A3, XX, INCX2 LFDUX A4, X2, INCX2 LFSDUX A1, XX, INCX2 LFSDUX A2, X2, INCX2 LFSDUX A3, XX, INCX2 LFSDUX A4, X2, INCX2 fpmul A1, ALPHA_R, A1 fpmul A2, ALPHA_R, A2 fpmul A3, ALPHA_R, A3 fpmul A4, ALPHA_R, A4 fpmadd C1, A1, A1, C1 fpmadd C2, A2, A2, C2 fpmadd C3, A3, A3, C3 fpmadd C4, A4, A4, C4 .align 4LL(126): andi. r0, N, 2 beq LL(127) LFDUX A1, XX, INCX2 LFDUX A2, X2, INCX2 LFDUX A3, XX, INCX2 LFDUX A4, X2, INCX2 fmul A1, ALPHA_R, A1 fmul A2, ALPHA_R, A2 fmul A3, ALPHA_R, A3 fmul A4, ALPHA_R, A4 fmadd C1, A1, A1, C1 fmadd C2, A2, A2, C2 fmadd C3, A3, A3, C3 fmadd C4, A4, A4, C4 .align 4LL(127): andi. r0, N, 1 beq LL(998) LFDUX A1, XX, INCX2 LFDUX A2, X2, INCX2 fmul A1, ALPHA_R, A1 fmul A2, ALPHA_R, A2 fmadd C1, A1, A1, C1 fmadd C2, A2, A2, C2 .align 4LL(998): fpadd C1, C1, C2 lis r3, 0x3f00 fpadd C3, C3, C4 lis r4, 0x4040 stw r3, 4(SP) stw r4, 8(SP) fpadd C1, C1, C3 lfs f10, 0(SP) fsmtp C2, C1 lfs f11, 4(SP) fadd C1, C2, C1 lfs f12, 8(SP) fcmpu cr0, f10, C1 beq cr0, LL(99)#ifndef HUMMER_EMULATOR frsqrte f9, C1 li r10, 16 fmul f2, f1, f9 lfpdux f29, SP, r10 fmul f3, f9, f11 lfpdux f28, SP, r10 fnmsub f7, f2, f9, f12 lfpdux f27, SP, r10 fmul f9, f3, f7 lfpdux f26, SP, r10 fadd f13, f11, f11 lfpdux f25, SP, r10 fmul f12, f1, f9 lfpdux f24, SP, r10 fmul f11, f12, f11 lfpdux f23, SP, r10 lfpdux f22, SP, r10 fnmsub f1, f12, f9, f13 lfpdux f21, SP, r10 lfpdux f20, SP, r10 lfpdux f19, SP, r10 lfpdux f18, SP, r10 fmadd f1, f11, f1, f12 lfpdux f17, SP, r10 lfpdux f16, SP, r10 lfpdux f15, SP, r10 lfpdux f14, SP, r10 addi SP, SP, 16 fmul C1, ALPHA, C1 blr#else fsqrt C1, C1 li r10, 16 lfpdux f29, SP, r10 lfpdux f28, SP, r10 lfpdux f27, SP, r10 lfpdux f26, SP, r10 lfpdux f25, SP, r10 lfpdux f24, SP, r10 lfpdux f23, SP, r10 lfpdux f22, SP, r10 lfpdux f21, SP, r10 lfpdux f20, SP, r10 lfpdux f19, SP, r10 lfpdux f18, SP, r10 lfpdux f17, SP, r10 lfpdux f16, SP, r10 lfpdux f15, SP, r10 lfpdux f14, SP, r10 fmul C1, ALPHA, C1 addi SP, SP, 16 blr#endif .align 4LL(999): li r10, 16 lfpdux f29, SP, r10 lfpdux f28, SP, r10 lfpdux f27, SP, r10 lfpdux f26, SP, r10 lfpdux f25, SP, r10 lfpdux f24, SP, r10 lfpdux f23, SP, r10 lfpdux f22, SP, r10 lfpdux f21, SP, r10 lfpdux f20, SP, r10 lfpdux f19, SP, r10 lfpdux f18, SP, r10 lfpdux f17, SP, r10 lfpdux f16, SP, r10 lfpdux f15, SP, r10 lfpdux f14, SP, r10 addi SP, SP, 16 blr 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 + -