📄 izamax_hummer.s
字号:
/*********************************************************************//* *//* Optimized BLAS libraries *//* By Kazushige Goto <kgoto@tacc.utexas.edu> *//* *//* Copyright (c) The University of Texas, 2005. All rights reserved. *//* UNIVERSITY EXPRESSLY DISCLAIMS ANY AND ALL WARRANTIES CONCERNING *//* THIS SOFTWARE AND DOCUMENTATION, INCLUDING ANY WARRANTIES OF *//* MERCHANTABILITY, FITNESS FOR ANY PARTICULAR PURPOSE, *//* NON-INFRINGEMENT AND WARRANTIES OF PERFORMANCE, AND ANY WARRANTY *//* THAT MIGHT OTHERWISE ARISE FROM COURSE OF DEALING OR USAGE OF *//* TRADE. NO WARRANTY IS EITHER EXPRESS OR IMPLIED WITH RESPECT TO *//* THE USE OF THE SOFTWARE OR DOCUMENTATION. *//* Under no circumstances shall University be liable for incidental, *//* special, indirect, direct or consequential damages or loss of *//* profits, interruption of business, or related expenses which may *//* arise from use of Software or Documentation, including but not *//* limited to those resulting from defects in Software and/or *//* Documentation, or loss or inaccuracy of data of any kind. *//*********************************************************************/#define ASSEMBLER#include "common.h" #define N r3#define X r4#define INCX r5 #define INCX2 r6#define X2 r7#define XX r8#define RET r9#define NN r10#define C1 f1#define C2 f0#define C3 f2#define C4 f3#define A1 f4#define A2 f5#define A3 f6#define A4 f7#define A5 f8#define A6 f9#define A7 f10#define A8 f11#define F1 f12#define F2 f13#define F3 f14#define F4 f15#define T1 f16#define T2 f17#define T3 f18#define T4 f19#define B1 f20#define B2 f21#define B3 f22#define B4 f23#define B5 f24#define B6 f25#define B7 f26#define B8 f27 PROLOGUE PROFCODE li r10, -16 stfpdux f14, SP, r10 stfpdux f15, SP, r10 stfpdux f16, SP, r10 stfpdux f17, SP, r10 stfpdux f18, SP, r10 stfpdux f19, SP, r10 stfpdux f20, SP, r10 stfpdux f21, SP, r10 stfpdux f22, SP, r10 stfpdux f23, SP, r10 stfpdux f24, SP, r10 stfpdux f25, SP, r10 stfpdux f26, SP, r10 stfpdux f27, SP, r10#ifdef F_INTERFACE LDINT N, 0(N) LDINT INCX, 0(INCX)#endif slwi INCX, INCX, BASE_SHIFT add INCX2, INCX, INCX li RET, 0 cmpwi cr0, N, 0 ble LL(999) cmpwi cr0, INCX, 0 mr NN, N ble LL(999) mr XX, X LFD A1, 0 * SIZE(X) LFD A2, 1 * SIZE(X) add X, X, INCX2 li RET, 1 fabs A1, A1 fabs A2, A2 subi INCX2, INCX2, SIZE addi N, N, -1 cmpwi cr0, N, 0 fadd C1, A1, A2 ble LL(999) fsmfp C1, C1 li INCX, SIZE fpmr C2, C1 sub X, X, INCX2 fpmr C3, C1 srawi. r0, N, 3 fpmr C4, C1 mtspr CTR, r0 beq- LL(105) LFDUX A1, X, INCX2 LFDUX A2, X, INCX LFDUX A3, X, INCX2 LFDUX A4, X, INCX LFSDUX A1, X, INCX2 LFSDUX A2, X, INCX LFSDUX A3, X, INCX2 LFSDUX A4, X, INCX LFDUX A5, X, INCX2 LFDUX A6, X, INCX LFDUX A7, X, INCX2 LFDUX A8, X, INCX LFSDUX A5, X, INCX2 LFSDUX A6, X, INCX LFSDUX A7, X, INCX2 LFSDUX A8, X, INCX bdz LL(103) .align 4LL(102): fpabs B1, A1 LFDUX A1, X, INCX2 fpabs B2, A2 LFDUX A2, X, INCX fpabs B3, A3 LFDUX A3, X, INCX2 fpabs B4, A4 LFDUX A4, X, INCX fpabs B5, A5 LFSDUX A1, X, INCX2 fpabs B6, A6 LFSDUX A2, X, INCX fpabs B7, A7 LFSDUX A3, X, INCX2 fpabs B8, A8 LFSDUX A4, X, INCX fpadd T1, B1, B2 LFDUX A5, X, INCX2 fpadd T2, B3, B4 LFDUX A6, X, INCX fpadd T3, B5, B6 LFDUX A7, X, INCX2 fpadd T4, B7, B8 LFDUX A8, X, INCX fpsub F1, C1, T1 LFSDUX A5, X, INCX2 fpsub F2, C2, T2 LFSDUX A6, X, INCX fpsub F3, C3, T3 LFSDUX A7, X, INCX2 fpsub F4, C4, T4 LFSDUX A8, X, INCX fpsel C1, F1, C1, T1 fpsel C2, F2, C2, T2 fpsel C3, F3, C3, T3 fpsel C4, F4, C4, T4 bdnz LL(102) .align 4LL(103): fpabs B1, A1 fpabs B2, A2 fpabs B3, A3 fpabs B4, A4 fpabs B5, A5 fpabs B6, A6 fpabs B7, A7 fpabs B8, A8 fpadd T1, B1, B2 fpadd T2, B3, B4 fpadd T3, B5, B6 fpadd T4, B7, B8 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 .align 4LL(105): andi. r0, N, 7 beq LL(120) andi. r0, N, 4 beq LL(106) LFDUX A1, X, INCX2 LFDUX A2, X, INCX LFDUX A3, X, INCX2 LFDUX A4, X, INCX LFSDUX A1, X, INCX2 LFSDUX A2, X, INCX LFSDUX A3, X, INCX2 LFSDUX A4, X, INCX fpabs A1, A1 fpabs A2, A2 fpabs A3, A3 fpabs A4, A4 fpadd A1, A1, A2 fpadd A3, A3, A4 fpsub F1, C1, A1 fpsub F2, C2, A3 fpsel C1, F1, C1, A1 fpsel C2, F2, C2, A3 .align 4LL(106): andi. r0, N, 2 beq LL(107) LFDUX A1, X, INCX2 LFDUX A2, X, INCX LFSDUX A1, X, INCX2 LFSDUX A2, X, INCX fpabs A1, A1 fpabs A2, A2 fpadd A1, A1, A2 fpsub F1, C1, A1 fpsel C1, F1, C1, A1 .align 4LL(107): andi. r0, N, 1 beq LL(120) LFDUX A1, X, INCX2 LFDUX A2, X, INCX fabs A1, A1 fabs A2, A2 fadd A1, A1, A2 fsub F1, C1, A1 fsel C1, F1, C1, A1 .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 li RET, 0 fsub F1, C1, C2 fsel C1, F1, C1, C2 fsmfp C1, C1 sub XX, XX, INCX2 srawi. r0, NN, 3 mtspr CTR, r0 beq- LL(125) LFDUX A1, XX, INCX2 LFDUX A2, XX, INCX LFDUX A3, XX, INCX2 LFDUX A4, XX, INCX LFSDUX A1, XX, INCX2 LFSDUX A2, XX, INCX LFSDUX A3, XX, INCX2 LFSDUX A4, XX, INCX LFDUX A5, XX, INCX2 LFDUX A6, XX, INCX LFDUX A7, XX, INCX2 LFDUX A8, XX, INCX LFSDUX A5, XX, INCX2 LFSDUX A6, XX, INCX LFSDUX A7, XX, INCX2 LFSDUX A8, XX, INCX fpabs T1, A1 fpabs T2, A2 fpabs T3, A3 fpabs T4, A4 fpadd B1, T1, T2 fpadd B2, T3, T4 bdz LL(123) .align 4LL(122): LFDUX A1, XX, INCX2 fpabs T1, A5 addi RET, RET, 1 fcmpu cr0, C1, B1 LFDUX A2, XX, INCX beq cr0, LL(999) LFDUX A3, XX, INCX2 fpabs T2, A6 addi RET, RET, 1 fcmpu cr0, C1, B2 LFDUX A4, XX, INCX beq cr0, LL(999) LFSDUX A1, XX, INCX2 fpabs T3, A7 addi RET, RET, 1 fscmp cr0, C1, B1 LFSDUX A2, XX, INCX beq cr0, LL(999) LFSDUX A3, XX, INCX2 fpabs T4, A8 addi RET, RET, 1 fscmp cr0, C1, B2 LFSDUX A4, XX, INCX beq cr0, LL(999) fpadd B3, T1, T2 fpadd B4, T3, T4 LFDUX A5, XX, INCX2 fpabs T1, A1 addi RET, RET, 1 fcmpu cr0, C1, B3 LFDUX A6, XX, INCX beq cr0, LL(999) LFDUX A7, XX, INCX2 fpabs T2, A2 addi RET, RET, 1 fcmpu cr0, C1, B4 LFDUX A8, XX, INCX beq cr0, LL(999) LFSDUX A5, XX, INCX2 fpabs T3, A3 addi RET, RET, 1 fscmp cr0, C1, B3 LFSDUX A6, XX, INCX beq cr0, LL(999) LFSDUX A7, XX, INCX2 fpabs T4, A4 addi RET, RET, 1 fscmp cr0, C1, B4 LFSDUX A8, XX, INCX beq cr0, LL(999) fpadd B1, T1, T2 fpadd B2, T3, T4 bdnz LL(122) .align 4LL(123): fpabs T1, A5 addi RET, RET, 1 fcmpu cr0, C1, B1 beq cr0, LL(999) fpabs T2, A6 addi RET, RET, 1 fcmpu cr0, C1, B2 beq cr0, LL(999) fpabs T3, A7 addi RET, RET, 1 fscmp cr0, C1, B1 beq cr0, LL(999) fpabs T4, A8 addi RET, RET, 1 fscmp cr0, C1, B2 beq cr0, LL(999) fpadd B3, T1, T2 fpadd B4, T3, T4 addi RET, RET, 1 fcmpu cr0, C1, B3 beq cr0, LL(999) addi RET, RET, 1 fcmpu cr0, C1, B4 beq cr0, LL(999) addi RET, RET, 1 fscmp cr0, C1, B3 beq cr0, LL(999) addi RET, RET, 1 fscmp cr0, C1, B4 beq cr0, LL(999) .align 4LL(125): andi. r0, NN, 4 beq LL(126) LFDUX A1, XX, INCX2 LFDUX A2, XX, INCX LFDUX A3, XX, INCX2 LFDUX A4, XX, INCX LFSDUX A1, XX, INCX2 LFSDUX A2, XX, INCX LFSDUX A3, XX, INCX2 LFSDUX A4, XX, INCX fpabs A1, A1 fpabs A2, A2 fpabs A3, A3 fpabs A4, A4 fpadd A1, A1, A2 fpadd A3, A3, A4 addi RET, RET, 1 fcmpu cr0, C1, A1 beq cr0, LL(999) addi RET, RET, 1 fcmpu cr0, C1, A3 beq cr0, LL(999) addi RET, RET, 1 fscmp cr0, C1, A1 beq cr0, LL(999) addi RET, RET, 1 fscmp cr0, C1, A3 beq cr0, LL(999) .align 4LL(126): andi. r0, NN, 2 beq LL(127) LFDUX A1, XX, INCX2 LFDUX A2, XX, INCX LFDUX A3, XX, INCX2 LFDUX A4, XX, INCX fabs A1, A1 fabs A2, A2 fabs A3, A3 fabs A4, A4 fadd A1, A1, A2 fadd A3, A3, A4 addi RET, RET, 1 fcmpu cr0, C1, A1 beq cr0, LL(999) addi RET, RET, 1 fcmpu cr0, C1, A3 beq cr0, LL(999) .align 4LL(127): addi RET, RET, 1 .align 4LL(999): li r10, 16 addi SP, SP, -16 mr r3, RET 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 + -