📄 trsm_kernel_lt_2x2.s
字号:
faddp %st, %st(2) FLD 1 * SIZE(AA) # temp1 = *(aoffset + 0) FLD 2 * SIZE(BB) # temp2 = *(boffset + 0) rep fmul %st(1) faddp %st, %st(2) FMUL 3 * SIZE(BB) # temp2 = *(boffset + 0) faddp %st, %st(2) FLD 2 * SIZE(AA) # temp1 = *(aoffset + 0) addl $2 * SIZE, AA addl $4 * SIZE, BB decl %eax jne .L55 ALIGN_4.L54:#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif andl $1,%eax # k & 1 je .L33 ALIGN_4 FLD 0 * SIZE(BB) # temp2 = *(boffset + 0) rep fmul %st(1) faddp %st, %st(2) FMUL 1 * SIZE(BB) # temp2 = *(boffset + 0) faddp %st, %st(2) FLD 1 * SIZE(AA) # temp1 = *(aoffset + 0) addl $1 * SIZE, AA addl $2 * SIZE, BB ALIGN_4.L33: ffreep %st(0)#if defined(LN) || defined(RT) movl KK, %eax#ifdef LN subl $1, %eax#else subl $2, %eax#endif leal (, %eax, SIZE), %eax movl AORIG, AA leal (AA, %eax, 1), AA leal (%ebx, %eax, 2), BB#endif#if defined(LN) || defined(LT) FLD 0 * SIZE(BB) fsubp %st, %st(1) FLD 1 * SIZE(BB) fsubp %st, %st(2)#else FLD 0 * SIZE(AA) fsubp %st, %st(1) FLD 1 * SIZE(AA) fsubp %st, %st(2)#endif#if defined(LN) || defined(LT) FLD 0 * SIZE(AA) fmul %st, %st(1) fmulp %st, %st(2)#endif#ifdef RN FLD 0 * SIZE(BB) fmulp %st, %st(1) FLD 1 * SIZE(BB) fmul %st(1), %st fsubrp %st, %st(2) FLD 3 * SIZE(BB) fmulp %st, %st(2)#endif#ifdef RT FLD 3 * SIZE(BB) fmulp %st, %st(2) FLD 2 * SIZE(BB) fmul %st(2), %st fsubrp %st, %st(1) FLD 0 * SIZE(BB) fmulp %st, %st(1)#endif#ifdef LN subl $1 * SIZE, %edi#endif#if defined(LN) || defined(LT) FSTU 0 * SIZE(BB) fxch %st(1) FSTU 1 * SIZE(BB)#else FSTU 0 * SIZE(AA) fxch %st(1) FSTU 1 * SIZE(AA)#endif FST 0 * SIZE(%edi,%ebp) FST 0 * SIZE(%edi)#ifndef LN addl $1 * SIZE, %edi#endif#if defined(LT) || defined(RN) movl K, %eax subl KK, %eax leal (,%eax, SIZE), %eax leal (AA, %eax, 1), AA leal (BB, %eax, 2), BB#endif#ifdef LN subl $1, KK#endif#ifdef LT addl $1, KK#endif#ifdef RT movl K, %eax sall $0 + BASE_SHIFT, %eax addl %eax, AORIG#endif ALIGN_4.L27:#ifdef LN movl K, %eax leal ( , %eax, SIZE), %eax leal (%ebx, %eax, 2), %ebx#endif#if defined(LT) || defined(RN) movl BB, %ebx#endif#ifdef RN addl $2, KK#endif#ifdef RT subl $2, KK#endif decl J # j-- # MEMORY jne .L34 ALIGN_4.L8: movl N, %eax # n # MEMORY andl $1, %eax je .End #if defined(LT) || defined(RN) movl A, AA#else movl A, %eax movl %eax, AORIG#endif#ifdef RT movl K, %eax sall $0 + BASE_SHIFT, %eax subl %eax, %ebx#endif#ifdef RT subl %ebp, C#endif movl C, %edi # c # MEMORY#ifndef RT addl %ebp, C#endif#ifdef LN movl OFFSET, %eax addl M, %eax movl %eax, KK#endif #ifdef LT movl OFFSET, %eax movl %eax, KK#endif movl M, %esi # m # MEMORY sarl $1, %esi # m >> 1 je .L36 ALIGN_4 .L46:#ifdef LN movl K, %eax sall $1 + BASE_SHIFT, %eax subl %eax, AORIG#endif#if defined(LN) || defined(RT) movl KK, %eax leal (, %eax, SIZE), %eax movl AORIG, AA leal (AA, %eax, 2), AA leal (%ebx, %eax, 1), BB#else movl %ebx, BB#endif fldz fldz FLD 0 * SIZE(BB) # temp1 = *(boffset + 0)#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif sarl $1, %eax je .L56 ALIGN_4.L57: FLD 0 * SIZE(AA) # temp2 = *(aoffset + 0) fmul %st(1) faddp %st, %st(2) FMUL 1 * SIZE(AA) # temp2 = *(aoffset + 0) faddp %st, %st(2) FLD 1 * SIZE(BB) # temp1 = *(boffset + 0) FLD 2 * SIZE(AA) # temp2 = *(aoffset + 0) fmul %st(1) faddp %st, %st(2) FMUL 3 * SIZE(AA) # temp2 = *(aoffset + 0) faddp %st, %st(2) FLD 2 * SIZE(BB) # temp1 = *(boffset + 0) addl $4 * SIZE,AA addl $2 * SIZE,BB dec %eax jne .L57 ALIGN_4.L56:#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif andl $1, %eax je .L45 ALIGN_4 FLD 0 * SIZE(AA) # temp2 = *(aoffset + 0) fmul %st(1) faddp %st, %st(2) FMUL 1 * SIZE(AA) # temp2 = *(aoffset + 0) faddp %st, %st(2) FLD 3 * SIZE(BB) # temp1 = *(boffset + 0) addl $2 * SIZE,AA addl $1 * SIZE,BB ALIGN_4.L45: ffreep %st(0)#if defined(LN) || defined(RT) movl KK, %eax#ifdef LN subl $2, %eax#else subl $1, %eax#endif leal (, %eax, SIZE), %eax movl AORIG, AA leal (AA, %eax, 2), AA leal (%ebx, %eax, 1), BB#endif#if defined(LN) || defined(LT) FLD 0 * SIZE(BB) fsubp %st, %st(1) FLD 1 * SIZE(BB) fsubp %st, %st(2)#else FLD 0 * SIZE(AA) fsubp %st, %st(1) FLD 1 * SIZE(AA) fsubp %st, %st(2)#endif#ifdef LN FLD 3 * SIZE(AA) fmulp %st, %st(2) FLD 2 * SIZE(AA) fmul %st(2), %st fsubrp %st, %st(1) FLD 0 * SIZE(AA) fmulp %st, %st(1)#endif#ifdef LT FLD 0 * SIZE(AA) fmulp %st, %st(1) FLD 1 * SIZE(AA) fmul %st(1), %st fsubrp %st, %st(2) FLD 3 * SIZE(AA) fmulp %st, %st(2)#endif#ifdef RN FLD 0 * SIZE(BB) fmul %st, %st(1) fmulp %st, %st(2)#endif#ifdef RT FLD 0 * SIZE(BB) fmul %st, %st(1) fmulp %st, %st(2)#endif#ifdef LN subl $2 * SIZE, %edi#endif#if defined(LN) || defined(LT) FSTU 0 * SIZE(BB) fxch %st(1) FSTU 1 * SIZE(BB)#else FSTU 0 * SIZE(AA) fxch %st(1) FSTU 1 * SIZE(AA)#endif FST 1 * SIZE(%edi) FST 0 * SIZE(%edi)#ifndef LN addl $2 * SIZE, %edi#endif#if defined(LT) || defined(RN) movl K, %eax subl KK, %eax leal (,%eax, SIZE), %eax leal (AA, %eax, 2), AA leal (BB, %eax, 1), BB#endif#ifdef LN subl $2, KK#endif#ifdef LT addl $2, KK#endif#ifdef RT movl K, %eax sall $1 + BASE_SHIFT, %eax addl %eax, AORIG#endif decl %esi # i -- jne .L46 ALIGN_4.L36: movl M, %eax # m # MEMORY andl $1, %eax # m & 1 je .L99#ifdef LN movl K, %eax sall $0 + BASE_SHIFT, %eax subl %eax, AORIG#endif#if defined(LN) || defined(RT) movl KK, %eax leal (, %eax, SIZE), %eax movl AORIG, AA leal (AA, %eax, 1), AA leal (%ebx, %eax, 1), BB#else movl %ebx, BB#endif fldz#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif test %eax, %eax jle .L52 ALIGN_3.L51: FLD (AA) FMUL (BB) addl $1 * SIZE,AA addl $1 * SIZE,BB faddp %st,%st(1) decl %eax jne .L51 ALIGN_4.L52:#if defined(LN) || defined(RT) movl KK, %eax#ifdef LN subl $1, %eax#else subl $1, %eax#endif leal (, %eax, SIZE), %eax movl AORIG, AA leal (AA, %eax, 1), AA leal (%ebx, %eax, 1), BB#endif#if defined(LN) || defined(LT) FLD 0 * SIZE(BB) fsubp %st, %st(1)#else FLD 0 * SIZE(AA) fsubp %st, %st(1)#endif#if defined(LN) || defined(LT) FMUL 0 * SIZE(AA)#else FMUL 0 * SIZE(BB)#endif#ifdef LN subl $1 * SIZE, %edi#endif#if defined(LN) || defined(LT) FSTU 0 * SIZE(BB)#else FSTU 0 * SIZE(AA)#endif FST 0 * SIZE(%edi)#ifndef LN addl $1 * SIZE, %edi#endif#if defined(LT) || defined(RN) movl K, %eax subl KK, %eax leal (,%eax, SIZE), %eax leal (AA, %eax, 1), AA leal (BB, %eax, 1), BB#endif#ifdef LN subl $1, KK#endif#ifdef LT addl $1, KK#endif#ifdef RT movl K, %eax sall $0 + BASE_SHIFT, %eax addl %eax, AORIG#endif ALIGN_4.L99:#ifdef LN movl K, %eax leal (%ebx, %eax, SIZE), %ebx#endif#if defined(LT) || defined(RN) movl BB, %ebx#endif#ifdef RN addl $1, KK#endif#ifdef RT subl $1, KK#endif ALIGN_4.End: popl %ebx popl %esi popl %edi popl %ebp addl $ARGS, %esp ret 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 + -