📄 qtrsm_kernel_ln_2x2.s
字号:
fmul %st, %st(3) fmulp %st, %st(4)#endif#ifdef RN FLD -8 * SIZE(BO) fmul %st, %st(1) fmulp %st, %st(3) FLD -7 * SIZE(BO) fmul %st(1), %st FLD -7 * SIZE(BO) fmul %st(4), %st fsubrp %st, %st(5) fsubrp %st, %st(2) FLD -5 * SIZE(BO) fmul %st, %st(2) fmulp %st, %st(4)#endif#ifdef RT FLD -5 * SIZE(BO) fmul %st, %st(2) fmulp %st, %st(4) FLD -6 * SIZE(BO) fmul %st(2), %st FLD -6 * SIZE(BO) fmul %st(5), %st fsubrp %st, %st(4) fsubrp %st, %st(1) FLD -8 * SIZE(BO) fmul %st, %st(1) fmulp %st, %st(3)#endif#ifdef LN subl $2 * SIZE, CO#endif#if defined(LN) || defined(LT) fld %st FST -8 * SIZE(BO) fxch %st(1) fld %st FST -7 * SIZE(BO) fxch %st(2) fld %st FST -6 * SIZE(BO) fxch %st(3) fld %st FST -5 * SIZE(BO) FST 1 * SIZE(CO, LDC) FST 0 * SIZE(CO) FST 0 * SIZE(CO, LDC) FST 1 * SIZE(CO)#else fld %st FST -8 * SIZE(AO) fxch %st(2) fld %st FST -7 * SIZE(AO) fxch %st(1) fld %st FST -6 * SIZE(AO) fxch %st(3) fld %st FST -5 * SIZE(AO) FST 1 * SIZE(CO, LDC) FST 1 * SIZE(CO) FST 0 * SIZE(CO) FST 0 * SIZE(CO, LDC)#endif#ifndef LN addl $2 * SIZE, CO#endif#if defined(LT) || defined(RN) movl K, %eax subl KK, %eax sall $BASE_SHIFT, %eax leal (AO, %eax, 2), AO leal (BO, %eax, 2), BO#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 I jne .L11 ALIGN_4.L29:#ifdef LN movl K, %eax sall $BASE_SHIFT, %eax leal (B, %eax, 2), B#endif#if defined(LT) || defined(RN) movl BO, B#endif#ifdef RN addl $2, KK#endif#ifdef RT subl $2, KK#endif decl J jne .L01 ALIGN_4.L30: movl N, %eax testl $1, %eax je .L999#if defined(LT) || defined(RN) movl A, AO#else movl A, %eax movl %eax, AORIG#endif#ifdef RT movl K, %eax sall $0 + BASE_SHIFT, %eax subl %eax, B#endif#ifdef RT subl LDC, C#endif movl C, CO#ifndef RT addl %eax, 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, %eax andl $1, %eax je .L40 ALIGN_4.L41:#ifdef LN movl K, %eax sall $0 + BASE_SHIFT, %eax subl %eax, AORIG#endif#if defined(LN) || defined(RT) movl KK, %eax sall $BASE_SHIFT, %eax movl AORIG, AO leal (AO, %eax, 1), AO leal (B, %eax, 1), BO#else movl B, BO#endif fldz#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif sarl $2, %eax je .L45 ALIGN_4.L42: PREFETCH (PREFETCHSIZE + 0) * SIZE(AO) FLD -8 * SIZE(AO) FLD -8 * SIZE(BO) fmulp %st, %st(1) faddp %st, %st(1) FLD -7 * SIZE(AO) FLD -7 * SIZE(BO) fmulp %st, %st(1) faddp %st, %st(1) FLD -6 * SIZE(AO) FLD -6 * SIZE(BO) fmulp %st, %st(1) faddp %st, %st(1) FLD -5 * SIZE(AO) FLD -5 * SIZE(BO) fmulp %st, %st(1) faddp %st, %st(1) addl $4 * SIZE,AO addl $4 * SIZE,BO decl %eax jne .L42 ALIGN_4.L45:#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif and $3, %eax je .L48 ALIGN_4.L46: FLD -8 * SIZE(AO) FLD -8 * SIZE(BO) fmulp %st, %st(1) faddp %st, %st(1) addl $1 * SIZE,AO addl $1 * SIZE,BO decl %eax jne .L46 ALIGN_4.L48:#if defined(LN) || defined(RT) movl KK, %eax#ifdef LN subl $1, %eax#else subl $1, %eax#endif sall $BASE_SHIFT, %eax movl AORIG, AO leal (AO, %eax, 1), AO leal (B, %eax, 1), BO#endif#if defined(LN) || defined(LT) FLD -8 * SIZE(BO) fsubp %st, %st(1)#else FLD -8 * SIZE(AO) fsubp %st, %st(1)#endif#ifdef LN FLD -8 * SIZE(AO) fmulp %st, %st(1)#endif#ifdef LT FLD -8 * SIZE(AO) fmulp %st, %st(1)#endif#ifdef RN FLD -8 * SIZE(BO) fmulp %st, %st(1)#endif#ifdef RT FLD -8 * SIZE(BO) fmulp %st, %st(1)#endif#ifdef LN subl $1 * SIZE, CO#endif#if defined(LN) || defined(LT) fld %st FST -8 * SIZE(BO)#else fld %st FST -8 * SIZE(AO)#endif FST 0 * SIZE(CO)#ifndef LN addl $1 * SIZE, CO#endif#if defined(LT) || defined(RN) movl K, %eax subl KK, %eax sall $BASE_SHIFT, %eax leal (AO, %eax, 1), AO leal (BO, %eax, 1), BO#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.L40: movl M, I sarl $1, I je .L49 ALIGN_4.L31:#ifdef LN movl K, %eax sall $1 + BASE_SHIFT, %eax subl %eax, AORIG#endif#if defined(LN) || defined(RT) movl KK, %eax sall $BASE_SHIFT, %eax movl AORIG, AO leal (AO, %eax, 2), AO leal (B, %eax, 1), BO#else movl B, BO#endif fldz fldz#if defined(HAVE_3DNOW) prefetchw 2 * SIZE(CO)#elif defined(HAVE_SSE) prefetchnta 2 * SIZE(CO)#endif#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif sarl $2, %eax je .L35 ALIGN_4.L32: PREFETCH (PREFETCHSIZE + 0) * SIZE(AO) FLD -8 * SIZE(BO) FLD -8 * SIZE(AO) fmul %st(1), %st faddp %st, %st(2) FLD -7 * SIZE(AO) fmulp %st, %st(1) faddp %st, %st(2) FLD -7 * SIZE(BO) FLD -6 * SIZE(AO) fmul %st(1), %st faddp %st, %st(2) FLD -5 * SIZE(AO) fmulp %st, %st(1) faddp %st, %st(2) FLD -6 * SIZE(BO) FLD -4 * SIZE(AO) fmul %st(1), %st faddp %st, %st(2) FLD -3 * SIZE(AO) fmulp %st, %st(1) faddp %st, %st(2) FLD -5 * SIZE(BO) FLD -2 * SIZE(AO) fmul %st(1), %st faddp %st, %st(2) FLD -1 * SIZE(AO) fmulp %st, %st(1) faddp %st, %st(2) addl $8 * SIZE,AO addl $4 * SIZE,BO decl %eax jne .L32 ALIGN_4.L35:#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif and $3, %eax je .L38 ALIGN_4.L36: FLD -8 * SIZE(BO) FLD -8 * SIZE(AO) fmul %st(1), %st faddp %st, %st(2) FLD -7 * SIZE(AO) fmulp %st, %st(1) faddp %st, %st(2) addl $2 * SIZE,AO addl $1 * SIZE,BO decl %eax jne .L36 ALIGN_4.L38:#if defined(LN) || defined(RT) movl KK, %eax#ifdef LN subl $2, %eax#else subl $1, %eax#endif sall $BASE_SHIFT, %eax movl AORIG, AO leal (AO, %eax, 2), AO leal (B, %eax, 1), BO#endif#if defined(LN) || defined(LT) FLD -8 * SIZE(BO) fsubp %st, %st(1) FLD -7 * SIZE(BO) fsubp %st, %st(2)#else FLD -8 * SIZE(AO) fsubp %st, %st(1) FLD -7 * SIZE(AO) fsubp %st, %st(3)#endif#ifdef LN FLD -5 * SIZE(AO) fmulp %st, %st(2) FLD -6 * SIZE(AO) fmul %st(2), %st fsubrp %st, %st(1) FLD -8 * SIZE(AO) fmulp %st, %st(1)#endif#ifdef LT FLD -8 * SIZE(AO) fmulp %st, %st(1) FLD -7 * SIZE(AO) fmul %st(1), %st fsubrp %st, %st(2) FLD -5 * SIZE(AO) fmulp %st, %st(2)#endif#ifdef RN FLD -8 * SIZE(BO) fmul %st, %st(1) fmulp %st, %st(2)#endif#ifdef RT FLD -8 * SIZE(BO) fmul %st, %st(1) fmulp %st, %st(2)#endif#ifdef LN subl $2 * SIZE, CO#endif#if defined(LN) || defined(LT) fld %st FST -8 * SIZE(BO) fxch %st(1) fld %st FST -7 * SIZE(BO)#else fld %st FST -8 * SIZE(AO) fxch %st(1) fld %st FST -7 * SIZE(AO)#endif FST 1 * SIZE(CO) FST 0 * SIZE(CO)#ifndef LN addl $2 * SIZE, CO#endif#if defined(LT) || defined(RN) movl K, %eax subl KK, %eax sall $BASE_SHIFT, %eax leal (AO, %eax, 2), AO leal (BO, %eax, 1), BO#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 I jne .L31 ALIGN_4.L49:#ifdef LN movl K, %eax sall $BASE_SHIFT, %eax leal (B, %eax, 1), B#endif#if defined(LT) || defined(RN) movl BO, B#endif#ifdef RN addl $1, KK#endif#ifdef RT subl $1, KK#endif ALIGN_4.L999: 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 + -