📄 ztrsm_kernel_lt_2x2_sse2.s
字号:
pshufd $0x4e, %xmm1, %xmm0 pshufd $0x4e, %xmm3, %xmm2 xorpd %xmm15, %xmm0 xorpd %xmm15, %xmm2 mulpd %xmm12, %xmm1 mulpd %xmm13, %xmm0 mulpd %xmm12, %xmm3 mulpd %xmm13, %xmm2 addpd %xmm0, %xmm1 addpd %xmm2, %xmm3#endif#ifdef LT movlpd 0 * SIZE(AO), %xmm8 movhpd 0 * SIZE(AO), %xmm8 movlpd 1 * SIZE(AO), %xmm9 movhpd 1 * SIZE(AO), %xmm9 movlpd 2 * SIZE(AO), %xmm10 movhpd 2 * SIZE(AO), %xmm10 movlpd 3 * SIZE(AO), %xmm11 movhpd 3 * SIZE(AO), %xmm11 movlpd 6 * SIZE(AO), %xmm12 movhpd 6 * SIZE(AO), %xmm12 movlpd 7 * SIZE(AO), %xmm13 movhpd 7 * SIZE(AO), %xmm13 pshufd $0x4e, %xmm1, %xmm0 pshufd $0x4e, %xmm3, %xmm2 xorpd %xmm15, %xmm0 xorpd %xmm15, %xmm2 mulpd %xmm8, %xmm1 mulpd %xmm9, %xmm0 mulpd %xmm8, %xmm3 mulpd %xmm9, %xmm2 addpd %xmm0, %xmm1 addpd %xmm2, %xmm3 movapd %xmm1, %xmm0 movapd %xmm3, %xmm2 pshufd $0x4e, %xmm1, %xmm4 pshufd $0x4e, %xmm3, %xmm6 xorpd %xmm15, %xmm4 xorpd %xmm15, %xmm6 mulpd %xmm10, %xmm0 mulpd %xmm10, %xmm2 mulpd %xmm11, %xmm4 mulpd %xmm11, %xmm6 subpd %xmm0, %xmm5 subpd %xmm2, %xmm7 subpd %xmm4, %xmm5 subpd %xmm6, %xmm7 pshufd $0x4e, %xmm5, %xmm4 pshufd $0x4e, %xmm7, %xmm6 xorpd %xmm15, %xmm4 xorpd %xmm15, %xmm6 mulpd %xmm12, %xmm5 mulpd %xmm13, %xmm4 mulpd %xmm12, %xmm7 mulpd %xmm13, %xmm6 addpd %xmm4, %xmm5 addpd %xmm6, %xmm7#endif#ifdef RN movlpd 0 * SIZE(B), %xmm8 movhpd 0 * SIZE(B), %xmm8 movlpd 1 * SIZE(B), %xmm9 movhpd 1 * SIZE(B), %xmm9 movlpd 2 * SIZE(B), %xmm10 movhpd 2 * SIZE(B), %xmm10 movlpd 3 * SIZE(B), %xmm11 movhpd 3 * SIZE(B), %xmm11 movlpd 6 * SIZE(B), %xmm12 movhpd 6 * SIZE(B), %xmm12 movlpd 7 * SIZE(B), %xmm13 movhpd 7 * SIZE(B), %xmm13 pshufd $0x4e, %xmm1, %xmm0 pshufd $0x4e, %xmm5, %xmm4 xorpd %xmm15, %xmm0 xorpd %xmm15, %xmm4 mulpd %xmm8, %xmm1 mulpd %xmm9, %xmm0 mulpd %xmm8, %xmm5 mulpd %xmm9, %xmm4 addpd %xmm0, %xmm1 addpd %xmm4, %xmm5 movapd %xmm1, %xmm0 movapd %xmm5, %xmm2 pshufd $0x4e, %xmm1, %xmm4 pshufd $0x4e, %xmm5, %xmm6 xorpd %xmm15, %xmm4 xorpd %xmm15, %xmm6 mulpd %xmm10, %xmm0 mulpd %xmm10, %xmm2 mulpd %xmm11, %xmm4 mulpd %xmm11, %xmm6 subpd %xmm0, %xmm3 subpd %xmm2, %xmm7 subpd %xmm4, %xmm3 subpd %xmm6, %xmm7 pshufd $0x4e, %xmm3, %xmm2 pshufd $0x4e, %xmm7, %xmm6 xorpd %xmm15, %xmm2 xorpd %xmm15, %xmm6 mulpd %xmm12, %xmm3 mulpd %xmm13, %xmm2 mulpd %xmm12, %xmm7 mulpd %xmm13, %xmm6 addpd %xmm2, %xmm3 addpd %xmm6, %xmm7#endif#ifdef RT movlpd 6 * SIZE(B), %xmm8 movhpd 6 * SIZE(B), %xmm8 movlpd 7 * SIZE(B), %xmm9 movhpd 7 * SIZE(B), %xmm9 movlpd 4 * SIZE(B), %xmm10 movhpd 4 * SIZE(B), %xmm10 movlpd 5 * SIZE(B), %xmm11 movhpd 5 * SIZE(B), %xmm11 movlpd 0 * SIZE(B), %xmm12 movhpd 0 * SIZE(B), %xmm12 movlpd 1 * SIZE(B), %xmm13 movhpd 1 * SIZE(B), %xmm13 pshufd $0x4e, %xmm3, %xmm2 pshufd $0x4e, %xmm7, %xmm6 xorpd %xmm15, %xmm2 xorpd %xmm15, %xmm6 mulpd %xmm8, %xmm3 mulpd %xmm9, %xmm2 mulpd %xmm8, %xmm7 mulpd %xmm9, %xmm6 addpd %xmm2, %xmm3 addpd %xmm6, %xmm7 movapd %xmm3, %xmm0 movapd %xmm7, %xmm2 pshufd $0x4e, %xmm3, %xmm4 pshufd $0x4e, %xmm7, %xmm6 xorpd %xmm15, %xmm4 xorpd %xmm15, %xmm6 mulpd %xmm10, %xmm0 mulpd %xmm10, %xmm2 mulpd %xmm11, %xmm4 mulpd %xmm11, %xmm6 subpd %xmm0, %xmm1 subpd %xmm2, %xmm5 subpd %xmm4, %xmm1 subpd %xmm6, %xmm5 pshufd $0x4e, %xmm1, %xmm0 pshufd $0x4e, %xmm5, %xmm4 xorpd %xmm15, %xmm0 xorpd %xmm15, %xmm4 mulpd %xmm12, %xmm1 mulpd %xmm13, %xmm0 mulpd %xmm12, %xmm5 mulpd %xmm13, %xmm4 addpd %xmm0, %xmm1 addpd %xmm4, %xmm5#endif#ifdef LN subq $4 * SIZE, CO1 subq $4 * SIZE, CO2#endif movsd %xmm1, 0 * SIZE(CO1) movhpd %xmm1, 1 * SIZE(CO1) movsd %xmm5, 2 * SIZE(CO1) movhpd %xmm5, 3 * SIZE(CO1) movsd %xmm3, 0 * SIZE(CO2) movhpd %xmm3, 1 * SIZE(CO2) movsd %xmm7, 2 * SIZE(CO2) movhpd %xmm7, 3 * SIZE(CO2)#if defined(LN) || defined(LT) movapd %xmm1, 0 * SIZE(B) movapd %xmm3, 2 * SIZE(B) movapd %xmm5, 4 * SIZE(B) movapd %xmm7, 6 * SIZE(B) movlpd %xmm1, 0 * SIZE(BO) movlpd %xmm1, 1 * SIZE(BO) movhpd %xmm1, 2 * SIZE(BO) movhpd %xmm1, 3 * SIZE(BO) movlpd %xmm3, 4 * SIZE(BO) movlpd %xmm3, 5 * SIZE(BO) movhpd %xmm3, 6 * SIZE(BO) movhpd %xmm3, 7 * SIZE(BO) movlpd %xmm5, 8 * SIZE(BO) movlpd %xmm5, 9 * SIZE(BO) movhpd %xmm5, 10 * SIZE(BO) movhpd %xmm5, 11 * SIZE(BO) movlpd %xmm7, 12 * SIZE(BO) movlpd %xmm7, 13 * SIZE(BO) movhpd %xmm7, 14 * SIZE(BO) movhpd %xmm7, 15 * SIZE(BO)#else movapd %xmm1, 0 * SIZE(AO) movapd %xmm5, 2 * SIZE(AO) movapd %xmm3, 4 * SIZE(AO) movapd %xmm7, 6 * SIZE(AO)#endif#ifndef LN addq $4 * SIZE, CO1 addq $4 * SIZE, CO2#endif#if defined(LT) || defined(RN) movq K, %rax subq KK, %rax salq $ZBASE_SHIFT, %rax leaq (AO, %rax, 2), AO#ifdef LT addq $8 * SIZE, B#endif#endif#ifdef LN subq $2, KK movq BORIG, B#endif#ifdef LT addq $2, KK#endif#ifdef RT movq K, %rax movq BORIG, B salq $1 + ZBASE_SHIFT, %rax addq %rax, AORIG#endif decq I # i -- jg .L10 ALIGN_4 .L30: testq $1, M jle .L99#ifdef LN movq K, %rax salq $0 + ZBASE_SHIFT, %rax subq %rax, AORIG#endif#if defined(LN) || defined(RT) movq KK, %rax movq AORIG, AO salq $ZBASE_SHIFT, %rax addq %rax, AO#endif leaq BUFFER, BO#if defined(LN) || defined(RT) movq KK, %rax salq $1 + ZBASE_SHIFT, %rax leaq (BO, %rax, 2), BO#endif pxor %xmm0, %xmm0 pxor %xmm1, %xmm1 pxor %xmm2, %xmm2 pxor %xmm3, %xmm3#if defined(LT) || defined(RN) movq KK, %rax#else movq K, %rax subq KK, %rax#endif sarq $2, %rax je .L42.L41: movapd 0 * SIZE(AO), %xmm8 movapd 0 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm0 movapd 2 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm1 movapd 4 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm2 movapd 6 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm3 movapd 2 * SIZE(AO), %xmm8 movapd 8 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm0 movapd 10 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm1 movapd 12 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm2 movapd 14 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm3 movapd 4 * SIZE(AO), %xmm8 movapd 16 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm0 movapd 18 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm1 movapd 20 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm2 movapd 22 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm3 movapd 6 * SIZE(AO), %xmm8 movapd 24 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm0 movapd 26 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm1 movapd 28 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm2 movapd 30 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm3 addq $ 8 * SIZE, AO addq $32 * SIZE, BO decq %rax jne .L41.L42:#if defined(LT) || defined(RN) movq KK, %rax#else movq K, %rax subq KK, %rax#endif movapd POSINV, %xmm15 andq $3, %rax # if (k & 1) BRANCH jle .L44.L43: movapd 0 * SIZE(AO), %xmm8 movapd 0 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm0 movapd 2 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm1 movapd 4 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm2 movapd 6 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm3 addq $2 * SIZE, AO # aoffset += 4 addq $8 * SIZE, BO # boffset1 += 8 decq %rax jg .L43 ALIGN_4.L44:#if defined(LN) || defined(RT) movq KK, %rax#ifdef LN subq $1, %rax#else subq $2, %rax#endif movq AORIG, AO movq BORIG, B leaq BUFFER, BO salq $ZBASE_SHIFT, %rax leaq (AO, %rax, 1), AO leaq (B, %rax, 2), B leaq (BO, %rax, 4), BO#endif SHUFPD_1 %xmm1, %xmm1 SHUFPD_1 %xmm3, %xmm3#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \ defined(NR) || defined(NC) || defined(TR) || defined(TC) xorpd %xmm15, %xmm1 xorpd %xmm15, %xmm3#else xorpd %xmm15, %xmm0 xorpd %xmm15, %xmm2#endif#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \ defined(RR) || defined(RC) || defined(CR) || defined(CC) subpd %xmm1, %xmm0 subpd %xmm3, %xmm2#else addpd %xmm1, %xmm0 addpd %xmm3, %xmm2#endif#if defined(LN) || defined(LT) movapd 0 * SIZE(B), %xmm1 movapd 2 * SIZE(B), %xmm3 subpd %xmm0, %xmm1 subpd %xmm2, %xmm3#else movapd 0 * SIZE(AO), %xmm1 movapd 2 * SIZE(AO), %xmm3 subpd %xmm0, %xmm1 subpd %xmm2, %xmm3#endif#ifndef CONJ SHUFPD_1 %xmm15, %xmm15#endif#if defined(LN) || defined(LT) movlpd 0 * SIZE(AO), %xmm8 movhpd 0 * SIZE(AO), %xmm8 movlpd 1 * SIZE(AO), %xmm9 movhpd 1 * SIZE(AO), %xmm9 pshufd $0x4e, %xmm1, %xmm0 pshufd $0x4e, %xmm3, %xmm2 xorpd %xmm15, %xmm0 xorpd %xmm15, %xmm2 mulpd %xmm8, %xmm1 mulpd %xmm9, %xmm0 mulpd %xmm8, %xmm3 mulpd %xmm9, %xmm2 addpd %xmm0, %xmm1 addpd %xmm2, %xmm3#endif#ifdef RN movlpd 0 * SIZE(B), %xmm8 movhpd 0 * SIZE(B), %xmm8 movlpd 1 * SIZE(B), %xmm9 movhpd 1 * SIZE(B), %xmm9 movlpd 2 * SIZE(B), %xmm10 movhpd 2 * SIZE(B), %xmm10 movlpd 3 * SIZE(B), %xmm11 movhpd 3 * SIZE(B), %xmm11 movlpd 6 * SIZE(B), %xmm12 movhpd 6 * SIZE(B), %xmm12 movlpd 7 * SIZE(B), %xmm13 movhpd 7 * SIZE(B), %xmm13 pshufd $0x4e, %xmm1, %xmm0 xorpd %xmm15, %xmm0 mulpd %xmm8, %xmm1 mulpd %xmm9, %xmm0 addpd %xmm0, %xmm1 movapd %xmm1, %xmm0 pshufd $0x4e, %xmm1, %xmm4 xorpd %xmm15, %xmm4 mulpd %xmm10, %xmm0 mulpd %xmm11, %xmm4 subpd %xmm0, %xmm3 subpd %xmm4, %xmm3 pshufd $0x4e, %xmm3, %xmm2 xorpd %xmm15, %xmm2 mulpd %xmm12, %xmm3 mulpd %xmm13, %xmm2 addpd %xmm2, %xmm3#endif#ifdef RT movlpd 6 * SIZE(B), %xmm8 movhpd 6 * SIZE(B), %xmm8 movlpd 7 * SIZE(B), %xmm9 movhpd 7 * SIZE(B), %xmm9 movlpd 4 * SIZE(B), %xmm10 movhpd 4 * SIZE(B), %xmm10 movlpd 5 * SIZE(B), %xmm11 movhpd 5 * SIZE(B), %xmm11 movlpd 0 * SIZE(B), %xmm12 movhpd 0 * SIZE(B), %xmm12 movlpd 1 * SIZE(B), %xmm13 movhpd 1 * SIZE(B), %xmm13 pshufd $0x4e, %xmm3, %xmm2 xorpd %xmm15, %xmm2 mulpd %xmm8, %xmm3 mulpd %xmm9, %xmm2 addpd %xmm2, %xmm3 movapd %xmm3, %xmm0 pshufd $0x4e, %xmm3, %xmm4 xorpd %xmm15, %xmm4 mulpd %xmm10, %xmm0 mulpd %xmm11, %xmm4 subpd %xmm0, %xmm1 subpd %xmm4, %xmm1 pshufd $0x4e, %xmm1, %xmm0 xorpd %xmm15, %xmm0 mulpd %xmm12, %xmm1 mulpd %xmm13, %xmm0 addpd %xmm0, %xmm1#endif#ifdef LN subq $2 * SIZE, CO1 subq $2 * SIZE, CO2#endif movsd %xmm1, 0 * SIZE(CO1) movhpd %xmm1, 1 * SIZE(CO1) movsd %xmm3, 0 * SIZE(CO2) movhpd %xmm3, 1 * SIZE(CO2)#if defined(LN) || defined(LT) movapd %xmm1, 0 * SIZE(B) movapd %xmm3, 2 * SIZE(B) movlpd %xmm1, 0 * SIZE(BO) movlpd %xmm1, 1 * SIZE(BO) movhpd %xmm1, 2 * SIZE(BO) movhpd %xmm1, 3 * SIZE(BO) movlpd %xmm3, 4 * SIZE(BO) movlpd %xmm3, 5 * SIZE(BO) movhpd %xmm3, 6 * SIZE(BO) movhpd %xmm3, 7 * SIZE(BO)#else movapd %xmm1, 0 * SIZE(AO) movapd %xmm3, 2 * SIZE(AO)#endif#ifndef LN addq $2 * SIZE, CO1 addq $2 * SIZE, CO2#endif#if defined(LT) || defined(RN) movq K, %rax subq KK, %rax salq $ZBASE_SHIFT, %rax leaq (AO, %rax, 1), AO#ifdef LT addq $4 * SIZE, B#endif#endif#ifdef LN subq $1, KK movq BORIG, B#endif#ifdef LT addq $1, KK#endif#ifdef RT movq K, %rax movq BORIG, B salq $0 + ZBASE_SHIFT, %rax addq %rax, AORIG#endif ALIGN_4 .L99:#ifdef LN leaq (, K, SIZE), %rax leaq (B, %rax, 4), B#endif#if defined(LT) || defined(RN) movq K, %rax subq KK, %rax leaq (,%rax, SIZE), %rax leaq (B, %rax, 2 * COMPSIZE), B#endif#ifdef RN addq $2, KK#endif#ifdef RT subq $2, KK#endif decq J # j -- jg .L01.L100: testq $1, N jle .L999.L101:#ifdef LN movq OFFSET, %rax addq M, %rax movq %rax, KK#endif /* Copying to Sub Buffer */ leaq BUFFER, BO#ifdef RT movq K, %rax salq $0 + ZBASE_SHIFT, %rax subq %rax, B#endif#if defined(LN) || defined(RT) movq KK, %rax movq B, BORIG salq $ZBASE_SHIFT, %rax leaq (B, %rax, 1), B leaq (BO, %rax, 2), BO#endif #if defined(LT) movq OFFSET, %rax movq %rax, KK#endif#if defined(LT) || defined(RN) movq KK, %rax#else movq K, %rax subq KK, %rax#endif sarq $2, %rax jle .L103 ALIGN_4 .L102: movlpd 0 * SIZE(B), %xmm0 movlpd 1 * SIZE(B), %xmm1 movlpd 2 * SIZE(B), %xmm2 movlpd 3 * SIZE(B), %xmm3 movlpd 4 * SIZE(B), %xmm4 movlpd 5 * SIZE(B), %xmm5 movlpd 6 * SIZE(B), %xmm6 movlpd 7 * SIZE(B), %xmm7 movlpd %xmm0, 0 * SIZE(BO) movlpd %xmm0, 1 * SIZE(BO) movlpd %xmm1, 2 * SIZE(BO) movlpd %xmm1, 3 * SIZE(BO) movlpd %xmm2, 4 * SIZE(BO) movlpd %xmm2, 5 * SIZE(BO) movlpd %xmm3, 6 * SIZE(BO) movlpd %xmm3, 7 * SIZE(BO) movlpd %xmm4, 8 * SIZE(BO) movlpd %xmm4, 9 * SIZE(BO) movlpd %xmm5, 10 * SIZE(BO) movlpd %xmm5, 11 * SIZE(BO) movlpd %xmm6, 12 * SIZE(BO) movlpd %xmm6, 13 * SIZE(BO) movlpd %xmm7, 14 * SIZE(BO) movlpd %xmm7, 15 * SIZE(BO) subq $-16 * SIZE, BO
⌨️ 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 + -