📄 trsm_kernel_lt_4x4_sse3.s
字号:
movapd 2 * SIZE(AO), %xmm8 addpd %xmm9, %xmm3 movddup 0 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm4 movddup 1 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm5 movddup 2 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm6 movddup 3 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 movapd 4 * SIZE(AO), %xmm8 addpd %xmm9, %xmm7 movddup 4 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm0 movddup 5 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm1 movddup 6 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm2 movddup 7 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 movapd 6 * SIZE(AO), %xmm8 addpd %xmm9, %xmm3 movddup 4 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm4 movddup 5 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm5 movddup 6 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm6 movddup 7 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 movapd 32 * SIZE(AO), %xmm8 addpd %xmm9, %xmm7 movddup 32 * SIZE(BO), %xmm9 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm0 movddup 9 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm1 movddup 10 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm2 movddup 11 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 movapd 10 * SIZE(AO), %xmm10 addpd %xmm11, %xmm3 movddup 8 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm4 movddup 9 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm5 movddup 10 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm6 movddup 11 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 movapd 12 * SIZE(AO), %xmm10 addpd %xmm11, %xmm7 movddup 12 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm0 movddup 13 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm1 movddup 14 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm2 movddup 15 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 movapd 14 * SIZE(AO), %xmm10 addpd %xmm11, %xmm3 movddup 12 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm4 movddup 13 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm5 movddup 14 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm6 movddup 15 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 movapd 40 * SIZE(AO), %xmm10 addpd %xmm11, %xmm7 movddup 40 * SIZE(BO), %xmm11 mulpd %xmm12, %xmm13 PREFETCH (PREFETCHSIZE + 16) * SIZE(AO) addpd %xmm13, %xmm0 movddup 17 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm1 movddup 18 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm2 movddup 19 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 movapd 18 * SIZE(AO), %xmm12 addpd %xmm13, %xmm3 movddup 16 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm4 movddup 17 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm5 movddup 18 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm6 movddup 19 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 movapd 20 * SIZE(AO), %xmm12 addpd %xmm13, %xmm7 movddup 20 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm0 movddup 21 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm1 movddup 22 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm2 movddup 23 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 movapd 22 * SIZE(AO), %xmm12 addpd %xmm13, %xmm3 movddup 20 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm4 movddup 21 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm5 movddup 22 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 addpd %xmm13, %xmm6 movddup 23 * SIZE(BO), %xmm13 mulpd %xmm12, %xmm13 movapd 48 * SIZE(AO), %xmm12 addpd %xmm13, %xmm7 movddup 48 * SIZE(BO), %xmm13 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm0 movddup 25 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm1 movddup 26 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm2 movddup 27 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 movapd 26 * SIZE(AO), %xmm14 addpd %xmm15, %xmm3 movddup 24 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm4 movddup 25 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm5 movddup 26 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm6 movddup 27 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 movapd 28 * SIZE(AO), %xmm14 addpd %xmm15, %xmm7 movddup 28 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm0 movddup 29 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm1 movddup 30 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm2 movddup 31 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 movapd 30 * SIZE(AO), %xmm14 addpd %xmm15, %xmm3 movddup 28 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm4 movddup 29 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm5 movddup 30 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 addpd %xmm15, %xmm6 movddup 31 * SIZE(BO), %xmm15 mulpd %xmm14, %xmm15 movapd 56 * SIZE(AO), %xmm14 addpd %xmm15, %xmm7 movddup 56 * SIZE(BO), %xmm15 addq $32 * SIZE, BO addq $32 * SIZE, AO decq %rax BRANCH jne .L12#endif ALIGN_4.L15:#if defined(LT) || defined(RN) movq KK, %rax#else movq K, %rax subq KK, %rax#endif andq $7, %rax # if (k & 1) BRANCH je .L19 ALIGN_4.L16: mulpd %xmm8, %xmm9 movapd 2 * SIZE(AO), %xmm10 addpd %xmm9, %xmm0 movddup 1 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 movddup 0 * SIZE(BO), %xmm11 addpd %xmm9, %xmm1 movddup 2 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 addpd %xmm9, %xmm2 movddup 3 * SIZE(BO), %xmm9 mulpd %xmm8, %xmm9 movapd 4 * SIZE(AO), %xmm8 addpd %xmm9, %xmm3 movddup 4 * SIZE(BO), %xmm9 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm4 movddup 1 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm5 movddup 2 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm6 movddup 3 * SIZE(BO), %xmm11 mulpd %xmm10, %xmm11 addpd %xmm11, %xmm7 addq $4 * SIZE, AO # aoffset += 4 addq $4 * SIZE, BO # boffset1 += 8 decq %rax jg .L16 ALIGN_4.L19:#if defined(LN) || defined(RT) movq KK, %rax subq $4, %rax leaq (, %rax, SIZE), %rax movq AORIG, AO leaq (AO, %rax, 4), AO leaq (B, %rax, 4), BO#endif#if defined(LN) || defined(LT) movapd %xmm0, %xmm8 unpcklpd %xmm1, %xmm0 unpckhpd %xmm1, %xmm8 movapd %xmm2, %xmm10 unpcklpd %xmm3, %xmm2 unpckhpd %xmm3, %xmm10 movapd %xmm4, %xmm12 unpcklpd %xmm5, %xmm4 unpckhpd %xmm5, %xmm12 movapd %xmm6, %xmm14 unpcklpd %xmm7, %xmm6 unpckhpd %xmm7, %xmm14 movapd 0 * SIZE(BO), %xmm1 movapd 2 * SIZE(BO), %xmm3 movapd 4 * SIZE(BO), %xmm5 movapd 6 * SIZE(BO), %xmm7 movapd 8 * SIZE(BO), %xmm9 movapd 10 * SIZE(BO), %xmm11 movapd 12 * SIZE(BO), %xmm13 movapd 14 * SIZE(BO), %xmm15 subpd %xmm0, %xmm1 subpd %xmm2, %xmm3 subpd %xmm8, %xmm5 subpd %xmm10, %xmm7 subpd %xmm4, %xmm9 subpd %xmm6, %xmm11 subpd %xmm12, %xmm13 subpd %xmm14, %xmm15#else movapd 0 * SIZE(AO), %xmm8 movapd 2 * SIZE(AO), %xmm9 movapd 4 * SIZE(AO), %xmm10 movapd 6 * SIZE(AO), %xmm11 movapd 8 * SIZE(AO), %xmm12 movapd 10 * SIZE(AO), %xmm13 movapd 12 * SIZE(AO), %xmm14 movapd 14 * SIZE(AO), %xmm15 subpd %xmm0, %xmm8 subpd %xmm4, %xmm9 subpd %xmm1, %xmm10 subpd %xmm5, %xmm11 subpd %xmm2, %xmm12 subpd %xmm6, %xmm13 subpd %xmm3, %xmm14 subpd %xmm7, %xmm15#endif#ifdef LN movddup 15 * SIZE(AO), %xmm0 mulpd %xmm0, %xmm13 mulpd %xmm0, %xmm15 movddup 14 * SIZE(AO), %xmm2 mulpd %xmm13, %xmm2 subpd %xmm2, %xmm9 movddup 14 * SIZE(AO), %xmm2 mulpd %xmm15, %xmm2 subpd %xmm2, %xmm11 movddup 13 * SIZE(AO), %xmm4 mulpd %xmm13, %xmm4 subpd %xmm4, %xmm5 movddup 13 * SIZE(AO), %xmm4 mulpd %xmm15, %xmm4 subpd %xmm4, %xmm7 movddup 12 * SIZE(AO), %xmm6 mulpd %xmm13, %xmm6 subpd %xmm6, %xmm1 movddup 12 * SIZE(AO), %xmm6 mulpd %xmm15, %xmm6 subpd %xmm6, %xmm3 movddup 10 * SIZE(AO), %xmm0 mulpd %xmm0, %xmm9 mulpd %xmm0, %xmm11 movddup 9 * SIZE(AO), %xmm2 mulpd %xmm9, %xmm2 subpd %xmm2, %xmm5 movddup 9 * SIZE(AO), %xmm2 mulpd %xmm11, %xmm2 subpd %xmm2, %xmm7 movddup 8 * SIZE(AO), %xmm4 mulpd %xmm9, %xmm4 subpd %xmm4, %xmm1 movddup 8 * SIZE(AO), %xmm4 mulpd %xmm11, %xmm4 subpd %xmm4, %xmm3 movddup 5 * SIZE(AO), %xmm0 mulpd %xmm0, %xmm5 mulpd %xmm0, %xmm7 movddup 4 * SIZE(AO), %xmm2 mulpd %xmm5, %xmm2 subpd %xmm2, %xmm1 movddup 4 * SIZE(AO), %xmm2 mulpd %xmm7, %xmm2 subpd %xmm2, %xmm3 movddup 0 * SIZE(AO), %xmm0 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm3#endif#ifdef LT movddup 0 * SIZE(AO), %xmm0 mulpd %xmm0, %xmm1 mulpd %xmm0, %xmm3 movddup 1 * SIZE(AO), %xmm2 mulpd %xmm1, %xmm2 subpd %xmm2, %xmm5 movddup 1 * SIZE(AO), %xmm2 mulpd %xmm3, %xmm2 subpd %xmm2, %xmm7 movddup 2 * SIZE(AO), %xmm4 mulpd %xmm1, %xmm4 subpd %xmm4, %xmm9 movddup 2 * SIZE(AO), %xmm4 mulpd %xmm3, %xmm4 subpd %xmm4, %xmm11 movddup 3 * SIZE(AO), %xmm6 mulpd %xmm1, %xmm6 subpd %xmm6, %xmm13 movddup 3 * SIZE(AO), %xmm6 mulpd %xmm3, %xmm6 subpd %xmm6, %xmm15 movddup 5 * SIZE(AO), %xmm0 mulpd %xmm0, %xmm5 mulpd %xmm0, %xmm7 movddup 6 * SIZE(AO), %xmm2 mulpd %xmm5, %xmm2 subpd %xmm2, %xmm9 movddup 6 * SIZE(AO), %xmm2 mulpd %xmm7, %xmm2 subpd %xmm2, %xmm11 movddup 7 * SIZE(AO), %xmm4 mulpd %xmm5, %xmm4 subpd %xmm4, %xmm13 movddup 7 * SIZE(AO), %xmm4 mulpd %xmm7, %xmm4 subpd %xmm4, %xmm15 movddup 10 * SIZE(AO), %xmm0 mulpd %xmm0, %xmm9 mulpd %xmm0, %xmm11 movddup 11 * SIZE(AO), %xmm2 mulpd %xmm9, %xmm2 subpd %xmm2, %xmm13 movddup 11 * SIZE(AO), %xmm2 mulpd %xmm11, %xmm2 subpd %xmm2, %xmm15 movddup 15 * SIZE(AO), %xmm0 mulpd %xmm0, %xmm13 mulpd %xmm0, %xmm15#endif#ifdef RN movddup 0 * SIZE(BO), %xmm0 mulpd %xmm0, %xmm8 mulpd %xmm0, %xmm9 movddup 1 * SIZE(BO), %xmm1 mulpd %xmm8, %xmm1 subpd %xmm1, %xmm10 movddup 1 * SIZE(BO), %xmm1 mulpd %xmm9, %xmm1 subpd %xmm1, %xmm11 movddup 2 * SIZE(BO), %xmm2 mulpd %xmm8, %xmm2 subpd %xmm2, %xmm12 movddup 2 * SIZE(BO), %xmm2 mulpd %xmm9, %xmm2 subpd %xmm2, %xmm13 movddup 3 * SIZE(BO), %xmm3 mulpd %xmm8, %xmm3 subpd %xmm3, %xmm14 movddup 3 * SIZE(BO), %xmm3 mulpd %xmm9, %xmm3 subpd %xmm3, %xmm15 movddup 5 * SIZE(BO), %xmm0 mulpd %xmm0, %xmm10 mulpd %xmm0, %xmm11 movddup 6 * SIZE(BO), %xmm1 mulpd %xmm10, %xmm1 subpd %xmm1, %xmm12 movddup 6 * SIZE(BO), %xmm1 mulpd %xmm11, %xmm1 subpd %xmm1, %xmm13 movddup 7 * SIZE(BO), %xmm2 mulpd %xmm10, %xmm2 subpd %xmm2, %xmm14 movddup 7 * SIZE(BO), %xmm2 mulpd %xmm11, %xmm2 subpd %xmm2, %xmm15 movddup 10 * SIZE(BO), %xmm0 mulpd %xmm0, %xmm12 mulpd %xmm0, %xmm13 movddup 11 * SIZE(BO), %xmm1 mulpd %xmm12, %xmm1 subpd %xmm1, %xmm14 movddup 11 * SIZE(BO), %xmm1 mulpd %xmm13, %xmm1 subpd %xmm1, %xmm15 movddup 15 * SIZE(BO), %xmm0 mulpd %xmm0, %xmm14 mulpd %xmm0, %xmm15#endif#ifdef RT movddup 15 * SIZE(BO), %xmm0 mulpd %xmm0, %xmm14 mulpd %xmm0, %xmm15 movddup 14 * SIZE(BO), %xmm1 mulpd %xmm14, %xmm1 subpd %xmm1, %xmm12 movddup 14 * SIZE(BO), %xmm1 mulpd %xmm15, %xmm1 subpd %xmm1, %xmm13 movddup 13 * SIZE(BO), %xmm2 mulpd %xmm14, %xmm2 subpd %xmm2, %xmm10 movddup 13 * SIZE(BO), %xmm2 mulpd %xmm15, %xmm2 subpd %xmm2, %xmm11 movddup 12 * SIZE(BO), %xmm3 mulpd %xmm14, %xmm3 subpd %xmm3, %xmm8 movddup 12 * SIZE(BO), %xmm3 mulpd %xmm15, %xmm3 subpd %xmm3, %xmm9 movddup 10 * SIZE(BO), %xmm0 mulpd %xmm0, %xmm12 mulpd %xmm0, %xmm13 movddup 9 * SIZE(BO), %xmm1 mulpd %xmm12, %xmm1 subpd %xmm1, %xmm10 movddup 9 * SIZE(BO), %xmm1 mulpd %xmm13, %xmm1 subpd %xmm1, %xmm11 movddup 8 * SIZE(BO), %xmm2 mulpd %xmm12, %xmm2 subpd %xmm2, %xmm8 movddup 8 * SIZE(BO), %xmm2 mulpd %xmm13, %xmm2 subpd %xmm2, %xmm9 movddup 5 * SIZE(BO), %xmm0 mulpd %xmm0, %xmm10 mulpd %xmm0, %xmm11 movddup 4 * SIZE(BO), %xmm1 mulpd %xmm10, %xmm1 subpd %xmm1, %xmm8 movddup 4 * SIZE(BO), %xmm1 mulpd %xmm11, %xmm1 subpd %xmm1, %xmm9 movddup 0 * SIZE(BO), %xmm0 mulpd %xmm0, %xmm8 mulpd %xmm0, %xmm9#endif#ifdef LN subq $4 * SIZE, CO1 subq $4 * SIZE, CO2#endif#if defined(LN) || defined(LT) movsd %xmm1, 0 * SIZE(CO1) movsd %xmm5, 1 * SIZE(CO1) movsd %xmm9, 2 * SIZE(CO1) movsd %xmm13, 3 * SIZE(CO1) movhpd %xmm1, 0 * SIZE(CO2) movhpd %xmm5, 1 * SIZE(CO2) movhpd %xmm9, 2 * SIZE(CO2) movhpd %xmm13, 3 * SIZE(CO2) movsd %xmm3, 0 * SIZE(CO1, LDC, 2) movsd %xmm7, 1 * SIZE(CO1, LDC, 2) movsd %xmm11, 2 * SIZE(CO1, LDC, 2) movsd %xmm15, 3 * SIZE(CO1, LDC, 2) movhpd %xmm3, 0 * SIZE(CO2, LDC, 2) movhpd %xmm7, 1 * SIZE(CO2, LDC, 2) movhpd %xmm11, 2 * SIZE(CO2, LDC, 2) movhpd %xmm15, 3 * SIZE(CO2, LDC, 2)#else movsd %xmm8, 0 * SIZE(CO1) movhpd %xmm8, 1 * SIZE(CO1) movsd %xmm9, 2 * SIZE(CO1) movhpd %xmm9, 3 * SIZE(CO1) movsd %xmm10, 0 * SIZE(CO2) movhpd %xmm10, 1 * SIZE(CO2) movsd %xmm11, 2 * SIZE(CO2) movhpd %xmm11, 3 * SIZE(CO2) movsd %xmm12, 0 * SIZE(CO1, LDC, 2) movhpd %xmm12, 1 * SIZE(CO1, LDC, 2) movsd %xmm13, 2 * SIZE(CO1, LDC, 2) movhpd %xmm13, 3 * SIZE(CO1, LDC, 2) movsd %xmm14, 0 * SIZE(CO2, LDC, 2) movhpd %xmm14, 1 * SIZE(CO2, LDC, 2) movsd %xmm15, 2 * SIZE(CO2, LDC, 2) movhpd %xmm15, 3 * SIZE(CO2, LDC, 2)#endif#if defined(LN) || defined(LT) movapd %xmm1, 0 * SIZE(BO) movapd %xmm3, 2 * SIZE(BO) movapd %xmm5, 4 * SIZE(BO) movapd %xmm7, 6 * SIZE(BO) movapd %xmm9, 8 * SIZE(BO) movapd %xmm11, 10 * SIZE(BO) movapd %xmm13, 12 * SIZE(BO) movapd %xmm15, 14 * SIZE(BO)#else movapd %xmm8, 0 * SIZE(AO) movapd %xmm9, 2 * SIZE(AO) movapd %xmm10, 4 * SIZE(AO) movapd %xmm11, 6 * SIZE(AO) movapd %xmm12, 8 * SIZE(AO) movapd %xmm13, 10 * SIZE(AO) movapd %xmm14, 12 * SIZE(AO) movapd %xmm15, 14 * 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
⌨️ 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 + -