📄 gemm_kernel_4x4_barcelona.s
字号:
#else addq $2, %rax#endif movq %rax, KKK#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: movddup ALPHA, %xmm7#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %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 #ifndef TRMMKERNEL movupd (CO1), %xmm0 movupd (CO2), %xmm2#endif mulpd %xmm7, %xmm8 mulpd %xmm7, %xmm9#ifndef TRMMKERNEL addpd %xmm0, %xmm8 addpd %xmm2, %xmm9#endif movlpd %xmm8, (CO1) movhpd %xmm8, 1 * SIZE(CO1) movlpd %xmm9, (CO2) movhpd %xmm9, 1 * SIZE(CO2)#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq K, %rax subq KKK, %rax leaq (,%rax, SIZE), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 2), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $2, KK#endif addq $2 * SIZE, CO1 # coffset += 4 addq $2 * SIZE, CO2 # coffset += 4 ALIGN_4 .L70: testq $1, M je .L79 ALIGN_4.L71:#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq B, BO#else movq KK, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 1), AO leaq (B, %rax, 2), 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#ifndef TRMMKERNEL movq K, %rax#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) movq K, %rax subq KK, %rax movq %rax, KKK #else movq KK, %rax#ifdef LEFT addq $1, %rax#else addq $2, %rax#endif movq %rax, KKK#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: movddup ALPHA, %xmm7#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %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#ifndef TRMMKERNEL movlpd (CO1), %xmm0 movhpd (CO2), %xmm0#endif mulpd %xmm7, %xmm8#ifndef TRMMKERNEL addpd %xmm0, %xmm8#endif movlpd %xmm8, (CO1) movhpd %xmm8, (CO2)#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq K, %rax subq KKK, %rax leaq (,%rax, SIZE), %rax leaq (AO, %rax, 1), AO leaq (BO, %rax, 2), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $1, KK#endif ALIGN_4 .L79:#if defined(TRMMKERNEL) && !defined(LEFT) addq $2, KK#endif movq BO, B leaq (C, LDC, 2), C ALIGN_4.L80: testq $1, N je .L999 ALIGN_4.L81:#if defined(TRMMKERNEL) && defined(LEFT) movq OFFSET, %rax movq %rax, KK#endif movq C, CO1 # coffset1 = c movq A, AO # aoffset = a movq M, I sarq $2, I # i = (m >> 2) jle .L100 ALIGN_4.L91:#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq B, BO#else movq KK, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 4), AO leaq (B, %rax, 1), BO#endif movapd -8 * SIZE(AO), %xmm2 pxor %xmm8, %xmm8 movapd -16 * SIZE(AO), %xmm0 pxor %xmm9, %xmm9 movddup -16 * SIZE(BO), %xmm1 pxor %xmm12, %xmm12 movddup -14 * SIZE(BO), %xmm3 pxor %xmm13, %xmm13 movddup -15 * SIZE(BO), %xmm5 prefetchw 3 * SIZE(CO1)#ifndef TRMMKERNEL movq K, %rax#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) movq K, %rax subq KK, %rax movq %rax, KKK #else movq KK, %rax#ifdef LEFT addq $4, %rax#else addq $1, %rax#endif movq %rax, KKK#endif andq $-4, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 4), AO leaq (BO, %rax, 1), BO negq %rax je,pn .L96 ALIGN_4.L92: mulpd %xmm1, %xmm0 mulpd -14 * SIZE(AO, %rax, 4), %xmm1 addpd %xmm0, %xmm8 movapd -12 * SIZE(AO, %rax, 4), %xmm0 addpd %xmm1, %xmm12 movddup -12 * SIZE(BO, %rax, 1), %xmm1 mulpd %xmm5, %xmm0 mulpd -10 * SIZE(AO, %rax, 4), %xmm5 addpd %xmm0, %xmm9 movapd (AO, %rax, 4), %xmm0 addpd %xmm5, %xmm13 movddup -13 * SIZE(BO, %rax, 1), %xmm5 mulpd %xmm3, %xmm2 mulpd -6 * SIZE(AO, %rax, 4), %xmm3 addpd %xmm2, %xmm8 movapd -4 * SIZE(AO, %rax, 4), %xmm2 addpd %xmm3, %xmm12 movddup -10 * SIZE(BO, %rax, 1), %xmm3 mulpd %xmm5, %xmm2 mulpd -2 * SIZE(AO, %rax, 4), %xmm5 addpd %xmm2, %xmm9 movapd 8 * SIZE(AO, %rax, 4), %xmm2 addpd %xmm5, %xmm13 movddup -11 * SIZE(BO, %rax, 1), %xmm5 addq $4 * SIZE, %rax jl,pt .L92 ALIGN_4.L96: movddup ALPHA, %xmm7#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif andq $3, %rax # if (k & 1) je .L99 leaq (, %rax, SIZE), %rax leaq (AO, %rax, 4), AO leaq (BO, %rax, 1), BO negq %rax ALIGN_4.L97: mulpd %xmm1, %xmm0 mulpd -14 * SIZE(AO, %rax, 4), %xmm1 addpd %xmm0, %xmm8 movapd -12 * SIZE(AO, %rax, 4), %xmm0 addpd %xmm1, %xmm12 movddup -15 * SIZE(BO, %rax, 1), %xmm1 addq $SIZE, %rax jl .L97 ALIGN_4.L99: addpd %xmm9, %xmm8 addpd %xmm13, %xmm12#ifndef TRMMKERNEL movupd (CO1), %xmm0 movupd 2 * SIZE(CO1), %xmm1#endif mulpd %xmm7, %xmm8 mulpd %xmm7, %xmm12#ifndef TRMMKERNEL addpd %xmm0, %xmm8 addpd %xmm1, %xmm12#endif movlpd %xmm8, (CO1) movhpd %xmm8, 1 * SIZE(CO1) movlpd %xmm12, 2 * SIZE(CO1) movhpd %xmm12, 3 * SIZE(CO1)#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq K, %rax subq KKK, %rax leaq (,%rax, SIZE), %rax leaq (AO, %rax, 4), AO leaq (BO, %rax, 1), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $4, KK#endif addq $4 * SIZE, CO1 # coffset += 4 decq I # i -- jg .L91 ALIGN_4 .L100: testq $2, M je .L110 ALIGN_4.L101:#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq B, BO#else movq KK, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 2), AO leaq (B, %rax, 1), BO#endif movddup -16 * SIZE(BO), %xmm0 pxor %xmm8, %xmm8 movddup -15 * SIZE(BO), %xmm1 pxor %xmm9, %xmm9 movddup -14 * SIZE(BO), %xmm2 pxor %xmm10, %xmm10 movddup -13 * SIZE(BO), %xmm3 pxor %xmm11, %xmm11#ifndef TRMMKERNEL movq K, %rax#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) movq K, %rax subq KK, %rax movq %rax, KKK #else movq KK, %rax#ifdef LEFT addq $2, %rax#else addq $1, %rax#endif movq %rax, KKK#endif andq $-4, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 1), BO negq %rax je,pn .L106 ALIGN_4.L102: mulpd -16 * SIZE(AO, %rax, 2), %xmm0 addpd %xmm0, %xmm8 movddup -12 * SIZE(BO, %rax, 1), %xmm0 mulpd -14 * SIZE(AO, %rax, 2), %xmm1 addpd %xmm1, %xmm9 movddup -11 * SIZE(BO, %rax, 1), %xmm1 mulpd -12 * SIZE(AO, %rax, 2), %xmm2 addpd %xmm2, %xmm10 movddup -10 * SIZE(BO, %rax, 1), %xmm2 mulpd -10 * SIZE(AO, %rax, 2), %xmm3 addpd %xmm3, %xmm11 movddup -9 * SIZE(BO, %rax, 1), %xmm3 addq $4 * SIZE, %rax jl,pt .L102 ALIGN_4.L106: movddup ALPHA, %xmm7#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif andq $3, %rax # if (k & 1) je .L109 leaq (, %rax, SIZE), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 1), BO negq %rax ALIGN_4.L107: movddup -16 * SIZE(BO, %rax, 1), %xmm0 mulpd -16 * SIZE(AO, %rax, 2), %xmm0 addpd %xmm0, %xmm8 addq $SIZE, %rax jl .L107 ALIGN_4.L109: addpd %xmm9, %xmm8 addpd %xmm11, %xmm10 addpd %xmm10, %xmm8#ifndef TRMMKERNEL movupd (CO1), %xmm0#endif mulpd %xmm7, %xmm8#ifndef TRMMKERNEL addpd %xmm0, %xmm8#endif movlpd %xmm8, (CO1) movhpd %xmm8, 1 * SIZE(CO1)#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq K, %rax subq KKK, %rax leaq (,%rax, SIZE), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 1), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $2, KK#endif addq $2 * SIZE, CO1 # coffset += 4 ALIGN_4 .L110: testq $1, M je .L999 ALIGN_4.L111:#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq B, BO#else movq KK, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 1), AO leaq (B, %rax, 1), BO#endif movapd -16 * SIZE(AO), %xmm0 pxor %xmm8, %xmm8 movapd -14 * SIZE(AO), %xmm1 pxor %xmm9, %xmm9#ifndef TRMMKERNEL movq K, %rax#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) movq K, %rax subq KK, %rax movq %rax, KKK #else movq KK, %rax#ifdef LEFT addq $1, %rax#else addq $1, %rax#endif movq %rax, KKK#endif andq $-4, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 1), AO leaq (BO, %rax, 1), BO negq %rax je,pn .L116 ALIGN_4.L112: mulpd -16 * SIZE(BO, %rax, 1), %xmm0 addpd %xmm0, %xmm8 movapd -12 * SIZE(AO, %rax, 1), %xmm0 mulpd -14 * SIZE(BO, %rax, 1), %xmm1 addpd %xmm1, %xmm9 movapd -10 * SIZE(AO, %rax, 1), %xmm1 addq $4 * SIZE, %rax jl,pt .L112 ALIGN_4.L116: movddup ALPHA, %xmm7#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif andq $3, %rax # if (k & 1) je .L118 leaq (, %rax, SIZE), %rax leaq (AO, %rax, 1), AO leaq (BO, %rax, 1), BO negq %rax ALIGN_4.L117: mulsd -16 * SIZE(BO, %rax, 1), %xmm0 addsd %xmm0, %xmm8 movsd -15 * SIZE(AO, %rax, 1), %xmm0 addq $SIZE, %rax jl .L117 ALIGN_4.L118: addpd %xmm9, %xmm8 haddpd %xmm8, %xmm8#ifndef TRMMKERNEL movsd (CO1), %xmm0#endif mulsd %xmm7, %xmm8#ifndef TRMMKERNEL addsd %xmm0, %xmm8#endif movlpd %xmm8, (CO1) ALIGN_4 .L999: movq (%rsp), %rbx movq 8(%rsp), %rbp movq 16(%rsp), %r12 movq 24(%rsp), %r13 movq 32(%rsp), %r14 movq 40(%rsp), %r15#ifdef WINDOWS_ABI movq 48(%rsp), %rdi movq 56(%rsp), %rsi movups 64(%rsp), %xmm6 movups 80(%rsp), %xmm7 movups 96(%rsp), %xmm8 movups 112(%rsp), %xmm9 movups 128(%rsp), %xmm10 movups 144(%rsp), %xmm11 movups 160(%rsp), %xmm12 movups 176(%rsp), %xmm13 movups 192(%rsp), %xmm14 movups 208(%rsp), %xmm15#endif addq $STACKSIZE, %rsp 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 + -