📄 trsm_kernel_rt_4x4_barcelona.s
字号:
mulpd -14 * SIZE(AO, %rax, 4), %xmm1 addpd %xmm0, %xmm8 movapd -12 * SIZE(AO, %rax, 4), %xmm0 addpd %xmm1, %xmm12 movddup -14 * SIZE(BO, %rax, 2), %xmm1 mulpd %xmm5, %xmm2 mulpd -14 * SIZE(AO, %rax, 4), %xmm5 addpd %xmm2, %xmm9 addpd %xmm5, %xmm13 movddup -13 * SIZE(BO, %rax, 2), %xmm5 movapd %xmm0, %xmm2 addq $SIZE, %rax jl .L57 ALIGN_4.L59:#if defined(LN) || defined(RT) movq KK, %rax#ifdef LN subq $4, %rax#else subq $2, %rax#endif leaq (, %rax, SIZE), %rax movq AORIG, AO leaq (AO, %rax, 4), AO leaq (B, %rax, 2), BO#endif#if defined(LN) || defined(LT) movapd %xmm8, %xmm0 unpcklpd %xmm9, %xmm8 unpckhpd %xmm9, %xmm0 movapd %xmm12, %xmm4 unpcklpd %xmm13, %xmm12 unpckhpd %xmm13, %xmm4 movapd -16 * SIZE(BO), %xmm9 movapd -14 * SIZE(BO), %xmm13 movapd -12 * SIZE(BO), %xmm1 movapd -10 * SIZE(BO), %xmm5 subpd %xmm8, %xmm9 subpd %xmm0, %xmm13 subpd %xmm12, %xmm1 subpd %xmm4, %xmm5#else movapd -16 * SIZE(AO), %xmm0 movapd -14 * SIZE(AO), %xmm1 movapd -12 * SIZE(AO), %xmm2 movapd -10 * SIZE(AO), %xmm3 subpd %xmm8, %xmm0 subpd %xmm12, %xmm1 subpd %xmm9, %xmm2 subpd %xmm13, %xmm3#endif#ifdef LN movddup -1 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm5 movddup -2 * SIZE(AO), %xmm10 mulpd %xmm5, %xmm10 subpd %xmm10, %xmm1 movddup -3 * SIZE(AO), %xmm12 mulpd %xmm5, %xmm12 subpd %xmm12, %xmm13 movddup -4 * SIZE(AO), %xmm14 mulpd %xmm5, %xmm14 subpd %xmm14, %xmm9 movddup -6 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm1 movddup -7 * SIZE(AO), %xmm10 mulpd %xmm1, %xmm10 subpd %xmm10, %xmm13 movddup -8 * SIZE(AO), %xmm12 mulpd %xmm1, %xmm12 subpd %xmm12, %xmm9 movddup -11 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm13 movddup -12 * SIZE(AO), %xmm10 mulpd %xmm13, %xmm10 subpd %xmm10, %xmm9 movddup -16 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm9#endif#ifdef LT movddup -16 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm9 movddup -15 * SIZE(AO), %xmm10 mulpd %xmm9, %xmm10 subpd %xmm10, %xmm13 movddup -14 * SIZE(AO), %xmm12 mulpd %xmm9, %xmm12 subpd %xmm12, %xmm1 movddup -13 * SIZE(AO), %xmm14 mulpd %xmm9, %xmm14 subpd %xmm14, %xmm5 movddup -11 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm13 movddup -10 * SIZE(AO), %xmm10 mulpd %xmm13, %xmm10 subpd %xmm10, %xmm1 movddup -9 * SIZE(AO), %xmm12 mulpd %xmm13, %xmm12 subpd %xmm12, %xmm5 movddup -6 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm1 movddup -5 * SIZE(AO), %xmm10 mulpd %xmm1, %xmm10 subpd %xmm10, %xmm5 movddup -1 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm5#endif#ifdef RN movddup -16 * SIZE(BO), %xmm8 mulpd %xmm8, %xmm0 mulpd %xmm8, %xmm1 movddup -15 * SIZE(BO), %xmm9 mulpd %xmm0, %xmm9 subpd %xmm9, %xmm2 movddup -15 * SIZE(BO), %xmm9 mulpd %xmm1, %xmm9 subpd %xmm9, %xmm3 movddup -13 * SIZE(BO), %xmm8 mulpd %xmm8, %xmm2 mulpd %xmm8, %xmm3#endif#ifdef RT movddup -13 * SIZE(BO), %xmm8 mulpd %xmm8, %xmm2 mulpd %xmm8, %xmm3 movddup -14 * SIZE(BO), %xmm9 mulpd %xmm2, %xmm9 subpd %xmm9, %xmm0 movddup -14 * SIZE(BO), %xmm9 mulpd %xmm3, %xmm9 subpd %xmm9, %xmm1 movddup -16 * SIZE(BO), %xmm8 mulpd %xmm8, %xmm0 mulpd %xmm8, %xmm1#endif#ifdef LN subq $4 * SIZE, CO1 subq $4 * SIZE, CO2#endif#if defined(LN) || defined(LT) movlpd %xmm9, 0 * SIZE(CO1) movlpd %xmm13, 1 * SIZE(CO1) movlpd %xmm1, 2 * SIZE(CO1) movlpd %xmm5, 3 * SIZE(CO1) movhpd %xmm9, 0 * SIZE(CO2) movhpd %xmm13, 1 * SIZE(CO2) movhpd %xmm1, 2 * SIZE(CO2) movhpd %xmm5, 3 * SIZE(CO2)#else movlpd %xmm0, 0 * SIZE(CO1) movhpd %xmm0, 1 * SIZE(CO1) movlpd %xmm1, 2 * SIZE(CO1) movhpd %xmm1, 3 * SIZE(CO1) movlpd %xmm2, 0 * SIZE(CO2) movhpd %xmm2, 1 * SIZE(CO2) movlpd %xmm3, 2 * SIZE(CO2) movhpd %xmm3, 3 * SIZE(CO2)#endif#if defined(LN) || defined(LT) movaps %xmm9, -16 * SIZE(BO) movaps %xmm13,-14 * SIZE(BO) movaps %xmm1, -12 * SIZE(BO) movaps %xmm5, -10 * SIZE(BO)#else movaps %xmm0, -16 * SIZE(AO) movaps %xmm1, -14 * SIZE(AO) movaps %xmm2, -12 * SIZE(AO) movaps %xmm3, -10 * 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 leaq (,%rax, SIZE), %rax leaq (AO, %rax, 4), AO leaq (BO, %rax, 2), BO#endif#ifdef LN subq $4, KK#endif#ifdef LT addq $4, KK#endif#ifdef RT movq K, %rax salq $2 + BASE_SHIFT, %rax addq %rax, AORIG#endif decq I # i -- jg .L51 ALIGN_4 .L60: testq $2, M je .L70#ifdef LN movq K, %rax salq $1 + BASE_SHIFT, %rax subq %rax, AORIG#endif#if defined(LN) || defined(RT) movq KK, %rax movq AORIG, AO leaq (, %rax, SIZE), %rax leaq (AO, %rax, 2), AO#endif movq B, BO#if defined(LN) || defined(RT) movq KK, %rax leaq (, %rax, SIZE), %rax leaq (BO, %rax, 2), BO#endif movapd -16 * SIZE(AO), %xmm0 pxor %xmm8, %xmm8 movapd -12 * SIZE(AO), %xmm2 pxor %xmm9, %xmm9 movddup -16 * SIZE(BO), %xmm1 pxor %xmm10, %xmm10 movddup -15 * SIZE(BO), %xmm3 pxor %xmm11, %xmm11#if defined(LT) || defined(RN) movq KK, %rax#else movq K, %rax subq KK, %rax#endif andq $-4, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 2), BO negq %rax je,pn .L66 ALIGN_4.L62: mulpd %xmm0, %xmm1 addpd %xmm1, %xmm8 movddup -14 * SIZE(BO, %rax, 2), %xmm1 mulpd %xmm0, %xmm3 movapd -14 * SIZE(AO, %rax, 2), %xmm0 addpd %xmm3, %xmm9 movddup -13 * SIZE(BO, %rax, 2), %xmm3 mulpd %xmm0, %xmm1 addpd %xmm1, %xmm10 movddup -12 * SIZE(BO, %rax, 2), %xmm1 mulpd %xmm0, %xmm3 movapd -8 * SIZE(AO, %rax, 2), %xmm0 addpd %xmm3, %xmm11 movddup -11 * SIZE(BO, %rax, 2), %xmm3 mulpd %xmm2, %xmm1 addpd %xmm1, %xmm8 movddup -10 * SIZE(BO, %rax, 2), %xmm1 mulpd %xmm2, %xmm3 movapd -10 * SIZE(AO, %rax, 2), %xmm2 addpd %xmm3, %xmm9 movddup -9 * SIZE(BO, %rax, 2), %xmm3 mulpd %xmm2, %xmm1 addpd %xmm1, %xmm10 movddup -8 * SIZE(BO, %rax, 2), %xmm1 mulpd %xmm2, %xmm3 movapd -4 * SIZE(AO, %rax, 2), %xmm2 addpd %xmm3, %xmm11 movddup -7 * SIZE(BO, %rax, 2), %xmm3 addq $4 * SIZE, %rax jl,pt .L62 ALIGN_4.L66:#if defined(LT) || defined(RN) movq KK, %rax#else movq K, %rax subq KK, %rax#endif andq $3, %rax # if (k & 1) je .L69 leaq (, %rax, SIZE), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 2), BO negq %rax ALIGN_4.L67: mulpd %xmm0, %xmm1 addpd %xmm1, %xmm8 movddup -14 * SIZE(BO, %rax, 2), %xmm1 mulpd %xmm0, %xmm3 movapd -14 * SIZE(AO, %rax, 2), %xmm0 addpd %xmm3, %xmm9 movddup -13 * SIZE(BO, %rax, 2), %xmm3 addq $SIZE, %rax jl .L67 ALIGN_4.L69: addpd %xmm10, %xmm8 addpd %xmm11, %xmm9#if defined(LN) || defined(RT) movq KK, %rax#ifdef LN subq $2, %rax#else subq $2, %rax#endif leaq (, %rax, SIZE), %rax movq AORIG, AO leaq (AO, %rax, 2), AO leaq (B, %rax, 2), BO#endif#if defined(LN) || defined(LT) movapd %xmm8, %xmm0 unpcklpd %xmm9, %xmm8 unpckhpd %xmm9, %xmm0 movapd -16 * SIZE(BO), %xmm9 movapd -14 * SIZE(BO), %xmm13 subpd %xmm8, %xmm9 subpd %xmm0, %xmm13#else movapd -16 * SIZE(AO), %xmm0 movapd -14 * SIZE(AO), %xmm2 subpd %xmm8, %xmm0 subpd %xmm9, %xmm2#endif#ifdef LN movddup -13 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm13 movddup -14 * SIZE(AO), %xmm10 mulpd %xmm13, %xmm10 subpd %xmm10, %xmm9 movddup -16 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm9#endif#ifdef LT movddup -16 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm9 movddup -15 * SIZE(AO), %xmm10 mulpd %xmm9, %xmm10 subpd %xmm10, %xmm13 movddup -13 * SIZE(AO), %xmm8 mulpd %xmm8, %xmm13#endif#ifdef RN movddup -16 * SIZE(BO), %xmm8 mulpd %xmm8, %xmm0 movddup -15 * SIZE(BO), %xmm9 mulpd %xmm0, %xmm9 subpd %xmm9, %xmm2 movddup -13 * SIZE(BO), %xmm8 mulpd %xmm8, %xmm2#endif#ifdef RT movddup -13 * SIZE(BO), %xmm8 mulpd %xmm8, %xmm2 movddup -14 * SIZE(BO), %xmm9 mulpd %xmm2, %xmm9 subpd %xmm9, %xmm0 movddup -16 * SIZE(BO), %xmm8 mulpd %xmm8, %xmm0#endif#ifdef LN subq $2 * SIZE, CO1 subq $2 * SIZE, CO2#endif#if defined(LN) || defined(LT) movlpd %xmm9, 0 * SIZE(CO1) movlpd %xmm13, 1 * SIZE(CO1) movhpd %xmm9, 0 * SIZE(CO2) movhpd %xmm13, 1 * SIZE(CO2)#else movlpd %xmm0, 0 * SIZE(CO1) movhpd %xmm0, 1 * SIZE(CO1) movlpd %xmm2, 0 * SIZE(CO2) movhpd %xmm2, 1 * SIZE(CO2)#endif#if defined(LN) || defined(LT) movaps %xmm9, -16 * SIZE(BO) movaps %xmm13, -14 * SIZE(BO)#else movaps %xmm0, -16 * SIZE(AO) movaps %xmm2, -14 * 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 leaq (,%rax, SIZE), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 2), BO#endif#ifdef LN subq $2, KK#endif#ifdef LT addq $2, KK#endif#ifdef RT movq K, %rax salq $1 + BASE_SHIFT, %rax addq %rax, AORIG#endif ALIGN_4 .L70: testq $1, M je .L79 ALIGN_4.L71:#ifdef LN movq K, %rax salq $0 + BASE_SHIFT, %rax subq %rax, AORIG#endif#if defined(LN) || defined(RT) movq KK, %rax movq AORIG, AO leaq (, %rax, SIZE), %rax leaq (AO, %rax, 1), AO#endif movq B, BO#if defined(LN) || defined(RT) movq KK, %rax salq $1 + BASE_SHIFT, %rax leaq (BO, %rax, 1), BO#endif movddup -16 * SIZE(AO), %xmm0 pxor %xmm8, %xmm8 movddup -15 * SIZE(AO), %xmm1 pxor %xmm9, %xmm9 movddup -14 * SIZE(AO), %xmm2 pxor %xmm10, %xmm10 movddup -13 * SIZE(AO), %xmm3 pxor %xmm11, %xmm11#if defined(LT) || defined(RN) movq KK, %rax#else movq K, %rax subq KK, %rax#endif andq $-4, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 1), AO leaq (BO, %rax, 2), BO negq %rax je,pn .L76 ALIGN_4.L72: mulpd -16 * SIZE(BO, %rax, 2), %xmm0 addpd %xmm0, %xmm8 movddup -12 * SIZE(AO, %rax, 1), %xmm0 mulpd -14 * SIZE(BO, %rax, 2), %xmm1 addpd %xmm1, %xmm9 movddup -11 * SIZE(AO, %rax, 1), %xmm1 mulpd -12 * SIZE(BO, %rax, 2), %xmm2 addpd %xmm2, %xmm10 movddup -10 * SIZE(AO, %rax, 1), %xmm2 mulpd -10 * SIZE(BO, %rax, 2), %xmm3 addpd %xmm3, %xmm11 movddup -9 * SIZE(AO, %rax, 1), %xmm3 addq $4 * SIZE, %rax jl,pt .L72 ALIGN_4.L76:#if defined(LT) || defined(RN) movq KK, %rax#else movq K, %rax subq KK, %rax#endif andq $3, %rax # if (k & 1) je .L78 leaq (, %rax, SIZE), %rax leaq (AO, %rax, 1), AO leaq (BO, %rax, 2), BO negq %rax ALIGN_4.L77: mulpd -16 * SIZE(BO, %rax, 2), %xmm0 addpd %xmm0, %xmm8 movddup -15 * SIZE(AO, %rax, 1), %xmm0 addq $SIZE, %rax jl .L77 ALIGN_4.L78: addpd %xmm9, %xmm8 addpd %xmm11, %xmm10 addpd %xmm10, %xmm8#if defined(LN) || defined(RT) movq KK, %rax#ifdef LN subq $1, %rax#else subq $2, %rax#endif leaq (, %rax, SIZE), %rax movq AORIG, AO leaq (AO, %rax, 1), AO leaq (B, %rax, 2), BO#endif#if defined(LN) || defined(LT) movapd -16 * SIZE(BO), %xmm2#else movapd -16 * SIZE(AO), %xmm2#endif subpd %xmm8, %xmm2#if defined(LN) || defined(LT) movddup -16 * SIZE(AO), %xmm0 mulpd %xmm0, %xmm2#endif#ifdef RN movapd %xmm2, %xmm0 unpckhpd %xmm0, %xmm0 mulsd -16 * SIZE(BO), %xmm2 movsd -15 * SIZE(BO), %xmm4 mulsd %xmm2, %xmm4 subsd %xmm4, %xmm0 mulsd -13 * SIZE(BO), %xmm0 unpcklpd %xmm0, %xmm2#endif#ifdef RT movapd %xmm2, %xmm0 unpckhpd %xmm0, %xmm0 mulsd -13 * SIZE(BO), %xmm0 movlpd -14 * SIZE(BO), %xmm4 mulsd %xmm0, %xmm4 subsd %xmm4, %xmm2 mulsd -16 * SIZE(BO), %xmm2 unpcklpd %xmm0, %xmm2#endif#ifdef LN subq $1 * SIZE, CO1 subq $1 * SIZE, CO2#endif movlpd %xmm2, 0 * SIZE(CO1) movhpd %xmm2, 0 * SIZE(CO2)#if defined(LN) || defined(LT) movaps %xmm2, -16 * SIZE(BO)#else movaps %xmm2, -16 * SIZE(AO)#endif#ifndef LN addq $1 * SIZE, CO1 addq $1 * SIZE, CO2#endif#if defined(LT) || defined(RN) movq K, %rax subq KK, %rax leaq (,%rax, SIZE), %rax leaq (AO, %rax, 1), AO leaq (BO, %rax, 2), BO#endif#ifdef LN subq $1, KK#endif
⌨️ 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 + -