📄 ztrsm_kernel_lt_1x2_penryn.s
字号:
subpd %xmm6, %xmm7 movddup -10 * SIZE(BB), %xmm2 movddup -9 * SIZE(BB), %xmm3 pshufd $0x4e, %xmm7, %xmm6 xorpd %xmm1, %xmm6 mulpd %xmm2, %xmm7 mulpd %xmm3, %xmm6 addpd %xmm6, %xmm7#endif#ifdef RT movddup -10 * SIZE(BB), %xmm2 movddup -9 * SIZE(BB), %xmm3 pshufd $0x4e, %xmm7, %xmm6 xorpd %xmm1, %xmm6 mulpd %xmm2, %xmm7 mulpd %xmm3, %xmm6 addpd %xmm6, %xmm7 movddup -12 * SIZE(BB), %xmm2 movddup -11 * SIZE(BB), %xmm3 movapd %xmm7, %xmm4 pshufd $0x4e, %xmm7, %xmm6 xorpd %xmm1, %xmm6 mulpd %xmm2, %xmm4 mulpd %xmm3, %xmm6 subpd %xmm4, %xmm5 subpd %xmm6, %xmm5 movddup -16 * SIZE(BB), %xmm2 movddup -15 * SIZE(BB), %xmm3 pshufd $0x4e, %xmm5, %xmm4 xorpd %xmm1, %xmm4 mulpd %xmm2, %xmm5 mulpd %xmm3, %xmm4 addpd %xmm4, %xmm5#endif#ifdef LN subl $2 * SIZE, CO1#endif movlpd %xmm5, 0 * SIZE(CO1) movhpd %xmm5, 1 * SIZE(CO1) movlpd %xmm7, 0 * SIZE(CO1, LDC) movhpd %xmm7, 1 * SIZE(CO1, LDC)#if defined(LN) || defined(LT) movapd %xmm5, -16 * SIZE(BB) movapd %xmm7, -14 * SIZE(BB)#else movapd %xmm5, -16 * SIZE(AA) movapd %xmm7, -14 * SIZE(AA)#endif#ifndef LN addl $2 * SIZE, CO1#endif#if defined(LT) || defined(RN) movl K, %eax subl KK, %eax sall $ZBASE_SHIFT, %eax addl %eax, 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 $ZBASE_SHIFT, %eax addl %eax, AORIG#endif decl %ebx # i -- jg .L10 ALIGN_4 .L99:#ifdef LN movl K, %eax sall $1 + ZBASE_SHIFT, %eax addl %eax, B#endif#if defined(LT) || defined(RN) movl BB, B#endif#ifdef RN addl $2, KK#endif#ifdef RT subl $2, KK#endif decl J # j -- jg .L01 ALIGN_4.L100: movl N, %eax testl $1, %eax jle .L999#if defined(LT) || defined(RN) movl A, AA#else movl A, %eax movl %eax, AORIG#endif#ifdef RT movl K, %eax sall $ZBASE_SHIFT, %eax subl %eax, B#endif#ifdef RT subl LDC, C#endif movl C, CO1#ifndef RT addl LDC, 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, %ebx ALIGN_4L110:#ifdef LN movl K, %eax sall $ZBASE_SHIFT, %eax subl %eax, AORIG#endif#if defined(LN) || defined(RT) movl KK, %eax movl AORIG, AA sall $ZBASE_SHIFT, %eax addl %eax, AA#endif movl B, BB#if defined(LN) || defined(RT) movl KK, %eax sall $ZBASE_SHIFT, %eax addl %eax, BB#endif movaps -16 * SIZE(AA), %xmm0 pxor %xmm2, %xmm2 movaps -16 * SIZE(BB), %xmm1 pxor %xmm3, %xmm3 pxor %xmm4, %xmm4#ifdef LN prefetcht0 -2 * SIZE(CO1)#else prefetcht0 1 * SIZE(CO1)#endif pxor %xmm5, %xmm5 pxor %xmm6, %xmm6 pxor %xmm7, %xmm7#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif sarl $3, %eax je L115 ALIGN_4L112: PREFETCH (PREFETCHSIZE + 0) * SIZE(AA) pshufd $0x4e, %xmm1, %xmm2 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm2 movaps -14 * SIZE(AA), %xmm0 ADD1 %xmm1, %xmm4 movaps -14 * SIZE(BB), %xmm1 ADD2 %xmm2, %xmm5 pshufd $0x4e, %xmm1, %xmm2 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm2 movaps -12 * SIZE(AA), %xmm0 ADD1 %xmm1, %xmm6 movaps -12 * SIZE(BB), %xmm1 ADD2 %xmm2, %xmm7 pshufd $0x4e, %xmm1, %xmm2 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm2 movaps -10 * SIZE(AA), %xmm0 ADD1 %xmm1, %xmm4 movaps -10 * SIZE(BB), %xmm1 ADD2 %xmm2, %xmm5 pshufd $0x4e, %xmm1, %xmm2 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm2 movaps -8 * SIZE(AA), %xmm0 ADD1 %xmm1, %xmm6 movaps -8 * SIZE(BB), %xmm1 ADD2 %xmm2, %xmm7 PREFETCH (PREFETCHSIZE + 8) * SIZE(AA) pshufd $0x4e, %xmm1, %xmm2 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm2 movaps -6 * SIZE(AA), %xmm0 ADD1 %xmm1, %xmm4 movaps -6 * SIZE(BB), %xmm1 ADD2 %xmm2, %xmm5 pshufd $0x4e, %xmm1, %xmm2 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm2 movaps -4 * SIZE(AA), %xmm0 ADD1 %xmm1, %xmm6 movaps -4 * SIZE(BB), %xmm1 ADD2 %xmm2, %xmm7 pshufd $0x4e, %xmm1, %xmm2 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm2 movaps -2 * SIZE(AA), %xmm0 ADD1 %xmm1, %xmm4 movaps -2 * SIZE(BB), %xmm1 ADD2 %xmm2, %xmm5 pshufd $0x4e, %xmm1, %xmm2 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm2 movaps 0 * SIZE(AA), %xmm0 ADD1 %xmm1, %xmm6 movaps 0 * SIZE(BB), %xmm1 ADD2 %xmm2, %xmm7 subl $-16 * SIZE, AA subl $-16 * SIZE, BB subl $1, %eax jne L112 ALIGN_4L115:#if defined(LT) || defined(RN) movl KK, %eax#else movl K, %eax subl KK, %eax#endif andl $7, %eax # if (k & 1) BRANCH je L118 ALIGN_4L116: pshufd $0x4e, %xmm1, %xmm2 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm2 movaps -14 * SIZE(AA), %xmm0 ADD1 %xmm1, %xmm4 movaps -14 * SIZE(BB), %xmm1 ADD2 %xmm2, %xmm5 addl $2 * SIZE, AA addl $2 * SIZE, BB decl %eax jg L116 ALIGN_4L118:#if defined(LN) || defined(RT) movl KK, %eax#ifdef LN subl $1, %eax#else subl $1, %eax#endif movl AORIG, AA sall $ZBASE_SHIFT, %eax leal (AA, %eax, 1), AA leal (B, %eax, 1), BB#endif addpd %xmm6, %xmm4 pcmpeqb %xmm1, %xmm1 addpd %xmm7, %xmm5 psllq $63, %xmm1#ifndef CONJ pshufd $0x40, %xmm1, %xmm0 shufps $0x04, %xmm1, %xmm1 pxor %xmm0, %xmm4#else#if defined(LN) || defined(LT) pshufd $0x40, %xmm1, %xmm0#else pshufd $0x04, %xmm1, %xmm0#endif shufps $0x40, %xmm1, %xmm1 pxor %xmm0, %xmm5#endif haddpd %xmm5, %xmm4#if defined(LN) || defined(LT) movapd -16 * SIZE(BB), %xmm5 subpd %xmm4, %xmm5#else movapd -16 * SIZE(AA), %xmm5 subpd %xmm4, %xmm5#endif#if defined(LN) || defined(LT) movddup -16 * SIZE(AA), %xmm2 movddup -15 * SIZE(AA), %xmm3 pshufd $0x4e, %xmm5, %xmm4 xorpd %xmm1, %xmm4 mulpd %xmm2, %xmm5 mulpd %xmm3, %xmm4 addpd %xmm4, %xmm5#endif#if defined(RN) || defined(RT) movddup -16 * SIZE(BB), %xmm2 movddup -15 * SIZE(BB), %xmm3 pshufd $0x4e, %xmm5, %xmm4 xorpd %xmm1, %xmm4 mulpd %xmm2, %xmm5 mulpd %xmm3, %xmm4 addpd %xmm4, %xmm5#endif#ifdef LN subl $2 * SIZE, CO1#endif movlpd %xmm5, 0 * SIZE(CO1) movhpd %xmm5, 1 * SIZE(CO1)#if defined(LN) || defined(LT) movapd %xmm5, -16 * SIZE(BB)#else movapd %xmm5, -16 * SIZE(AA)#endif#ifndef LN addl $2 * SIZE, CO1#endif#if defined(LT) || defined(RN) movl K, %eax subl KK, %eax sall $ZBASE_SHIFT, %eax addl %eax, AA addl %eax, BB#endif#ifdef LN subl $1, KK#endif#ifdef LT addl $1, KK#endif#ifdef RT movl K, %eax sall $ZBASE_SHIFT, %eax addl %eax, AORIG#endif decl %ebx # i -- jg L110#ifdef LN movl K, %eax sall $ZBASE_SHIFT, %eax addl %eax, B#endif#if defined(LT) || defined(RN) movl BB, 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 + -