📄 trsm_kernel_rt_2x2.s
字号:
prefetchw 2 * SIZE(%edi, %ebp, 1)#elif defined(HAVE_SSE) prefetchnta 2 * SIZE(%edi) prefetchnta 2 * SIZE(%edi, %ebp, 1)#endif#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif sarl $2, %eax je .L16 ALIGN_4.MainLoop:#if defined(HAVE_3DNOW) prefetch (PREFETCH_OFFSET) * SIZE(BB) nop#elif defined(HAVE_SSE) prefetchnta (PREFETCH_OFFSET) * SIZE(BB)#if (L2_SIZE == 524288) prefetcht0 (PREFETCH_OFFSET) * SIZE(AA)#endif#endif fmul %st, %st(1) FMUL 1 * SIZE(BB) fxch %st(1) faddp %st, %st(4) FLD 0 * SIZE(BB) fxch %st(1) faddp %st, %st(5) FLD 1 * SIZE(AA) fmul %st, %st(1) FMUL 1 * SIZE(BB) fxch %st(1) faddp %st, %st(6) FLD 2 * SIZE(BB) fxch %st(1) faddp %st, %st(7) FLD 2 * SIZE(AA) fmul %st, %st(1) FMUL 3 * SIZE(BB) fxch %st(1) faddp %st, %st(4) FLD 2 * SIZE(BB) fxch %st(1) faddp %st, %st(5) FLD 3 * SIZE(AA) fmul %st, %st(1) FMUL 3 * SIZE(BB) fxch %st(1) faddp %st, %st(6) FLD 8 * SIZE(BB) fxch %st(1) faddp %st, %st(7) FLD 8 * SIZE(AA) fxch %st(2)#if !defined(HAVE_3DNOW) && defined(HAVE_SSE) && defined(DOUBLE) prefetchnta (PREFETCH_OFFSET + 4) * SIZE(BB)#if (L2_SIZE == 524288) prefetcht0 (PREFETCH_OFFSET + 4) * SIZE(AA)#endif#endif fmul %st, %st(3) FMUL 5 * SIZE(BB) fxch %st(3) faddp %st, %st(4) FLD 4 * SIZE(BB) fxch %st(3) faddp %st, %st(5) FLD 5 * SIZE(AA) fmul %st, %st(3) FMUL 5 * SIZE(BB) fxch %st(3) faddp %st, %st(6) FLD 6 * SIZE(BB) fxch %st(3) faddp %st, %st(7) FLD 6 * SIZE(AA) fmul %st, %st(3) FMUL 7 * SIZE(BB) fxch %st(3) faddp %st, %st(4) FLD 6 * SIZE(BB) fxch %st(3) faddp %st, %st(5) FLD 7 * SIZE(AA) fmul %st, %st(3) FMUL 7 * SIZE(BB) fxch %st(3) faddp %st, %st(6) FLD 12 * SIZE(BB) fxch %st(3) faddp %st, %st(7) FLD 12 * SIZE(AA) fxch %st(2) subl $-8 * SIZE, BB subl $-8 * SIZE, AA decl %eax # l -- jne .MainLoop ALIGN_4.L16:#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif and $3, %eax je .L21 ALIGN_4.SubLoop: fmul %st, %st(1) FMUL 1 * SIZE(BB) fxch %st(1) faddp %st, %st(4) FLD 0 * SIZE(BB) fxch %st(1) faddp %st, %st(5) FLD 1 * SIZE(AA) fmul %st, %st(1) FMUL 1 * SIZE(BB) fxch %st(1) faddp %st, %st(6) FLD 2 * SIZE(BB) fxch %st(1) faddp %st, %st(7) FLD 2 * SIZE(AA) addl $2 * SIZE,BB addl $2 * SIZE,AA decl %eax jne .SubLoop ALIGN_4.L21: ffreep %st(0) ffreep %st(0) ffreep %st(0) ffreep %st(0)#if defined(LN) || defined(RT) movl KK, %eax#ifdef LN subl $2, %eax#else subl $2, %eax#endif leal (, %eax, SIZE), %eax movl AORIG, AA leal (AA, %eax, 2), 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) FLD 2 * SIZE(BB) fsubp %st, %st(3) FLD 3 * SIZE(BB) fsubp %st, %st(4)#else FLD 0 * SIZE(AA) fsubp %st, %st(1) FLD 1 * SIZE(AA) fsubp %st, %st(3) FLD 2 * SIZE(AA) fsubp %st, %st(2) FLD 3 * SIZE(AA) fsubp %st, %st(4)#endif#ifdef LN FLD 3 * SIZE(AA) fmul %st, %st(3) fmulp %st, %st(4) FLD 2 * SIZE(AA) fmul %st(3), %st FLD 2 * SIZE(AA) fmul %st(5), %st fsubrp %st, %st(3) fsubrp %st, %st(1) FLD 0 * SIZE(AA) fmul %st, %st(1) fmulp %st, %st(2)#endif#ifdef LT FLD 0 * SIZE(AA) fmul %st, %st(1) fmulp %st, %st(2) FLD 1 * SIZE(AA) fmul %st(1), %st FLD 1 * SIZE(AA) fmul %st(3), %st fsubrp %st, %st(5) fsubrp %st, %st(3) FLD 3 * SIZE(AA) fmul %st, %st(3) fmulp %st, %st(4)#endif#ifdef RN FLD 0 * SIZE(BB) fmul %st, %st(1) fmulp %st, %st(3) FLD 1 * SIZE(BB) fmul %st(1), %st FLD 1 * SIZE(BB) fmul %st(4), %st fsubrp %st, %st(5) fsubrp %st, %st(2) FLD 3 * SIZE(BB) fmul %st, %st(2) fmulp %st, %st(4)#endif#ifdef RT FLD 3 * SIZE(BB) fmul %st, %st(2) fmulp %st, %st(4) FLD 2 * SIZE(BB) fmul %st(2), %st FLD 2 * SIZE(BB) fmul %st(5), %st fsubrp %st, %st(4) fsubrp %st, %st(1) FLD 0 * SIZE(BB) fmul %st, %st(1) fmulp %st, %st(3)#endif#ifdef LN subl $2 * SIZE, %edi#endif#if defined(LN) || defined(LT) FSTU 0 * SIZE(BB) fxch %st(1) FSTU 1 * SIZE(BB) fxch %st(2) FSTU 2 * SIZE(BB) fxch %st(3) FSTU 3 * SIZE(BB) FST 1 * SIZE(%edi,%ebp) FST 0 * SIZE(%edi) FST 0 * SIZE(%edi,%ebp) FST 1 * SIZE(%edi)#else FSTU 0 * SIZE(AA) fxch %st(2) FSTU 1 * SIZE(AA) fxch %st(1) FSTU 2 * SIZE(AA) fxch %st(3) FSTU 3 * SIZE(AA) FST 1 * SIZE(%edi,%ebp) FST 1 * SIZE(%edi) FST 0 * SIZE(%edi) FST 0 * SIZE(%edi,%ebp)#endif#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, 2), 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 .MainHead ALIGN_4.L12: movl M, %eax # m # MEMORY andl $1, %eax je .L27#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, 2), BB#else movl %ebx, BB#endif fldz fldz FLD 0 * SIZE(AA) # temp1 = *(aoffset + 0)#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif sarl $1,%eax # k >> 1 # MEMORY je .L54 ALIGN_4.L55: 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) 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.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 + -