📄 gemm_kernel_8x4_barcelona.s
字号:
prefetchw (WPREFETCHSIZE + 0) * SIZE(BO) pshufd $0x00, %xmm3, %xmm0 pshufd $0x55, %xmm3, %xmm1 pshufd $0xaa, %xmm3, %xmm2 pshufd $0xff, %xmm3, %xmm3 prefetchw (WPREFETCHSIZE + 16) * SIZE(BO) pshufd $0x00, %xmm7, %xmm4 pshufd $0x55, %xmm7, %xmm5 pshufd $0xaa, %xmm7, %xmm6 pshufd $0xff, %xmm7, %xmm7 movaps %xmm0, 0 * SIZE(BO) movaps %xmm1, 4 * SIZE(BO) movaps %xmm2, 8 * SIZE(BO) movaps %xmm3, 12 * SIZE(BO) movaps %xmm4, 16 * SIZE(BO) movaps %xmm5, 20 * SIZE(BO) movaps %xmm6, 24 * SIZE(BO) movaps %xmm7, 28 * SIZE(BO) addq $ 8 * SIZE, B addq $32 * SIZE, BO decq %rax jne .L102 ALIGN_4.L103: movq K, %rax andq $7, %rax BRANCH jle .L110 ALIGN_4.L104: movss 0 * SIZE(B), %xmm3 pshufd $0x00, %xmm3, %xmm0 movaps %xmm0, 0 * SIZE(BO) addq $ 1 * SIZE, B addq $ 4 * SIZE, BO decq %rax jne .L104 ALIGN_4 .L110: movq C, CO1 # coffset1 = c movq A, AO # aoffset = a movq M, I sarq $3, I # i = (m >> 3) jle .L120 ALIGN_4.L111:#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) leaq BUFFER, BO#else leaq BUFFER, BO movq KK, %rax leaq (, %rax, 8), %rax leaq (AO, %rax, 4), AO leaq (BO, %rax, 2), BO#endif movaps -32 * SIZE(AO), %xmm8 movaps -16 * SIZE(AO), %xmm10 movaps 0 * SIZE(AO), %xmm12 movaps 16 * SIZE(AO), %xmm14 movaps 0 * SIZE(BO), %xmm9 movaps 16 * SIZE(BO), %xmm11 movaps 32 * SIZE(BO), %xmm13 movaps 48 * SIZE(BO), %xmm15 pxor %xmm0, %xmm0 pxor %xmm1, %xmm1 prefetchw 4 * SIZE(CO1) pxor %xmm4, %xmm4 pxor %xmm5, %xmm5#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 $8, %rax#else addq $1, %rax#endif movq %rax, KKK#endif sarq $3, %rax je .L115 ALIGN_4.L112: mulps %xmm9, %xmm8#if defined(OPTERON) || defined(BARCELONA) PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)#endif mulps -28 * SIZE(AO), %xmm9 addps %xmm8, %xmm0 movaps -24 * SIZE(AO), %xmm8 addps %xmm9, %xmm4 movaps 4 * SIZE(BO), %xmm9 mulps %xmm9, %xmm8 mulps -20 * SIZE(AO), %xmm9 addps %xmm8, %xmm0 movaps 32 * SIZE(AO), %xmm8 addps %xmm9, %xmm4 movaps 8 * SIZE(BO), %xmm9#if defined(OPTERON) || defined(BARCELONA) PREFETCH (PREFETCHSIZE + 16) * SIZE(AO)#endif mulps %xmm9, %xmm10 mulps -12 * SIZE(AO), %xmm9 addps %xmm10, %xmm0 movaps -8 * SIZE(AO), %xmm10 addps %xmm9, %xmm4 movaps 12 * SIZE(BO), %xmm9 mulps %xmm9, %xmm10 mulps -4 * SIZE(AO), %xmm9 addps %xmm10, %xmm0 movaps 48 * SIZE(AO), %xmm10 addps %xmm9, %xmm4 movaps 32 * SIZE(BO), %xmm9#if defined(OPTERON) || defined(BARCELONA) PREFETCH (PREFETCHSIZE + 32) * SIZE(AO)#endif mulps %xmm11, %xmm12 mulps 4 * SIZE(AO), %xmm11 addps %xmm12, %xmm0 movaps 8 * SIZE(AO), %xmm12 addps %xmm11, %xmm4 movaps 20 * SIZE(BO), %xmm11 mulps %xmm11, %xmm12 mulps 12 * SIZE(AO), %xmm11 addps %xmm12, %xmm0 movaps 64 * SIZE(AO), %xmm12 addps %xmm11, %xmm4 movaps 24 * SIZE(BO), %xmm11#if defined(OPTERON) || defined(BARCELONA) PREFETCH (PREFETCHSIZE + 48) * SIZE(AO)#endif mulps %xmm11, %xmm14 mulps 20 * SIZE(AO), %xmm11 addps %xmm14, %xmm0 movaps 24 * SIZE(AO), %xmm14 addps %xmm11, %xmm4 movaps 28 * SIZE(BO), %xmm11 mulps %xmm11, %xmm14 mulps 28 * SIZE(AO), %xmm11 addps %xmm14, %xmm0 movaps 80 * SIZE(AO), %xmm14 addps %xmm11, %xmm4 movaps 48 * SIZE(BO), %xmm11 addq $64 * SIZE, AO addq $32 * SIZE, BO decq %rax jne .L112 ALIGN_4.L115:#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif movaps ALPHA, %xmm15 andq $7, %rax # if (k & 1) BRANCH je .L118 ALIGN_4.L116: mulps %xmm9, %xmm8 mulps -28 * SIZE(AO), %xmm9 addps %xmm8, %xmm0 movaps -24 * SIZE(AO), %xmm8 addps %xmm9, %xmm4 movaps 4 * SIZE(BO), %xmm9 addq $8 * SIZE, AO # aoffset += 4 addq $4 * SIZE, BO # boffset1 += 8 decq %rax jg .L116 ALIGN_4.L118:#ifndef TRMMKERNEL movlps 0 * SIZE(CO1), %xmm8 movhps 2 * SIZE(CO1), %xmm8 movlps 4 * SIZE(CO1), %xmm9 movhps 6 * SIZE(CO1), %xmm9#endif mulps %xmm15, %xmm0 mulps %xmm15, %xmm4#ifndef TRMMKERNEL addps %xmm8, %xmm0 addps %xmm9, %xmm4#endif movlps %xmm0, 0 * SIZE(CO1) movhps %xmm0, 2 * SIZE(CO1) movlps %xmm4, 4 * SIZE(CO1) movhps %xmm4, 6 * SIZE(CO1)#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq K, %rax subq KKK, %rax leaq (,%rax, 8), %rax leaq (AO, %rax, 4), AO leaq (BO, %rax, 2), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $8, KK#endif addq $8 * SIZE, CO1 # coffset += 4 decq I # i -- jg .L111 ALIGN_4 .L120: testq $4, M je .L130#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) leaq BUFFER, BO#else leaq BUFFER, BO movq KK, %rax leaq (, %rax, 8), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 2), BO#endif movaps -32 * SIZE(AO), %xmm8 movaps -16 * SIZE(AO), %xmm10 movaps 0 * SIZE(BO), %xmm9 movaps 16 * SIZE(BO), %xmm11 pxor %xmm0, %xmm0 pxor %xmm1, %xmm1 pxor %xmm2, %xmm2 pxor %xmm3, %xmm3#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 sarq $3, %rax je .L125 ALIGN_4.L122: mulps %xmm8, %xmm9#if defined(OPTERON) || defined(BARCELONA) PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)#endif movaps -28 * SIZE(AO), %xmm8 mulps 4 * SIZE(BO), %xmm8 addps %xmm9, %xmm0 movaps 32 * SIZE(BO), %xmm9 addps %xmm8, %xmm1 movaps -24 * SIZE(AO), %xmm8 mulps 8 * SIZE(BO), %xmm8 addps %xmm8, %xmm2 movaps -20 * SIZE(AO), %xmm8 mulps 12 * SIZE(BO), %xmm8 addps %xmm8, %xmm3 movaps 0 * SIZE(AO), %xmm8#if defined(OPTERON) || defined(BARCELONA) PREFETCH (PREFETCHSIZE + 16) * SIZE(AO)#endif mulps %xmm10, %xmm11 movaps -12 * SIZE(AO), %xmm10 mulps 20 * SIZE(BO), %xmm10 addps %xmm11, %xmm0 movaps 48 * SIZE(BO), %xmm11 addps %xmm10, %xmm1 movaps -8 * SIZE(AO), %xmm10 mulps 24 * SIZE(BO), %xmm10 addps %xmm10, %xmm2 movaps -4 * SIZE(AO), %xmm10 mulps 28 * SIZE(BO), %xmm10 addps %xmm10, %xmm3 movaps 16 * SIZE(AO), %xmm10 addq $32 * SIZE, AO addq $32 * SIZE, BO decq %rax jne .L122 ALIGN_4.L125:#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif movaps ALPHA, %xmm15 andq $7, %rax # if (k & 1) BRANCH je .L128 ALIGN_4.L126: mulps %xmm8, %xmm9 movaps -28 * SIZE(AO), %xmm8 addps %xmm9, %xmm0 movaps 4 * SIZE(BO), %xmm9 addq $4 * SIZE, AO # aoffset += 4 addq $4 * SIZE, BO # boffset1 += 8 decq %rax jg .L126 ALIGN_4.L128:#ifndef TRMMKERNEL movlps 0 * SIZE(CO1), %xmm8 movhps 2 * SIZE(CO1), %xmm8#endif addps %xmm1, %xmm0 addps %xmm3, %xmm2 addps %xmm2, %xmm0 mulps %xmm15, %xmm0#ifndef TRMMKERNEL addps %xmm8, %xmm0#endif movlps %xmm0, 0 * SIZE(CO1) movhps %xmm0, 2 * SIZE(CO1)#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq K, %rax subq KKK, %rax leaq (,%rax, 8), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 2), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $4, KK#endif addq $4 * SIZE, CO1 # coffset += 4 ALIGN_4 .L130: testq $2, M je .L140#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) leaq BUFFER, BO#else leaq BUFFER, BO movq KK, %rax leaq (, %rax, 8), %rax leaq (AO, %rax, 1), AO leaq (BO, %rax, 2), BO#endif movaps -32 * SIZE(AO), %xmm8 movaps -24 * SIZE(AO), %xmm10 movaps 0 * SIZE(BO), %xmm9 movaps 16 * SIZE(BO), %xmm11 pxor %xmm0, %xmm0 pxor %xmm1, %xmm1 pxor %xmm2, %xmm2 pxor %xmm3, %xmm3#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 sarq $3, %rax je .L135 ALIGN_4.L132: mulps %xmm8, %xmm9#if defined(OPTERON) || defined(BARCELONA) PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)#endif movlps -30 * SIZE(AO), %xmm8 addps %xmm9, %xmm0 movlps 4 * SIZE(BO), %xmm9 mulps %xmm8, %xmm9 movlps -28 * SIZE(AO), %xmm8 addps %xmm9, %xmm1 movlps 8 * SIZE(BO), %xmm9 mulps %xmm8, %xmm9 movlps -26 * SIZE(AO), %xmm8 addps %xmm9, %xmm0 movlps 12 * SIZE(BO), %xmm9 mulps %xmm8, %xmm9 movlps -16 * SIZE(AO), %xmm8 addps %xmm9, %xmm1 movlps 32 * SIZE(BO), %xmm9 mulps %xmm10, %xmm11 movlps -22 * SIZE(AO), %xmm10 addps %xmm11, %xmm0 movlps 20 * SIZE(BO), %xmm11 mulps %xmm10, %xmm11 movlps -20 * SIZE(AO), %xmm10 addps %xmm11, %xmm1 movlps 24 * SIZE(BO), %xmm11 mulps %xmm10, %xmm11 movlps -18 * SIZE(AO), %xmm10 addps %xmm11, %xmm0 movlps 28 * SIZE(BO), %xmm11 mulps %xmm10, %xmm11 movlps -8 * SIZE(AO), %xmm10 addps %xmm11, %xmm1 movlps 48 * SIZE(BO), %xmm11 addq $16 * SIZE, AO addq $32 * SIZE, BO decq %rax jne .L132 ALIGN_4.L135:#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif movaps ALPHA, %xmm15 andq $7, %rax # if (k & 1) BRANCH je .L138 ALIGN_4.L136: mulps %xmm8, %xmm9 movlps -30 * SIZE(AO), %xmm8 addps %xmm9, %xmm0 movlps 4 * SIZE(BO), %xmm9 addq $2 * SIZE, AO # aoffset += 4 addq $4 * SIZE, BO # boffset1 += 8 decq %rax jg .L136 ALIGN_4.L138: addps %xmm1, %xmm0 mulps %xmm15, %xmm0#ifndef TRMMKERNEL movlps 0 * SIZE(CO1), %xmm8 addps %xmm8, %xmm0#endif movlps %xmm0, 0 * SIZE(CO1)#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq K, %rax subq KKK, %rax leaq (,%rax, 8), %rax leaq (AO, %rax, 1), AO leaq (BO, %rax, 2), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $2, KK#endif addq $2 * SIZE, CO1 # coffset += 4 ALIGN_4 .L140: testq $1, M je .L999#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) leaq BUFFER, BO#else leaq BUFFER, BO movq KK, %rax leaq (, %rax, 4), %rax leaq (AO, %rax, 1), AO leaq (BO, %rax, 4), BO#endif movss -32 * SIZE(AO), %xmm8 movss -28 * SIZE(AO), %xmm10 movss 0 * SIZE(BO), %xmm9 movss 16 * SIZE(BO), %xmm11 pxor %xmm0, %xmm0 pxor %xmm1, %xmm1 pxor %xmm2, %xmm2 pxor %xmm3, %xmm3#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 sarq $3, %rax je .L145 ALIGN_4.L142: mulss %xmm8, %xmm9#if defined(OPTERON) || defined(BARCELONA) PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)#endif movss -31 * SIZE(AO), %xmm8 mulss 4 * SIZE(BO), %xmm8 addss %xmm9, %xmm0 movss 32 * SIZE(BO), %xmm9 addss %xmm8, %xmm1 movss -30 * SIZE(AO), %xmm8 mulss 8 * SIZE(BO), %xmm8 addss %xmm8, %xmm2 movss -29 * SIZE(AO), %xmm8 mulss 12 * SIZE(BO), %xmm8 addss %xmm8, %xmm3 movss -24 * SIZE(AO), %xmm8 mulss %xmm10, %xmm11 movss -27 * SIZE(AO), %xmm10 mulss 20 * SIZE(BO), %xmm10 addss %xmm11, %xmm0 movss 48 * SIZE(BO), %xmm11 addss %xmm10, %xmm1 movss -26 * SIZE(AO), %xmm10 mulss 24 * SIZE(BO), %xmm10 addss %xmm10, %xmm2 movss -25 * SIZE(AO), %xmm10 mulss 28 * SIZE(BO), %xmm10 addss %xmm10, %xmm3 movss -20 * SIZE(AO), %xmm10 addq $ 8 * SIZE, AO addq $32 * SIZE, BO decq %rax jne .L142 ALIGN_4.L145:#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif movss ALPHA
⌨️ 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 + -