📄 gemm_kernel_8x4_penryn.s
字号:
movddup ALPHA, %xmm3 pshufd $0x39, %xmm9, %xmm2 pshufd $0x4e, %xmm10, %xmm0 pshufd $0x93, %xmm11, %xmm11 movaps %xmm8, %xmm10 unpcklps %xmm0, %xmm8 unpckhps %xmm0, %xmm10 movaps %xmm2, %xmm1 unpcklps %xmm11, %xmm2 unpckhps %xmm11, %xmm1 movaps %xmm8, %xmm9 unpcklps %xmm2, %xmm8 unpckhps %xmm2, %xmm9 movaps %xmm10, %xmm11 unpcklps %xmm1, %xmm10 unpckhps %xmm1, %xmm11 pshufd $0x93, %xmm9, %xmm9 pshufd $0x4e, %xmm10, %xmm10 pshufd $0x39, %xmm11, %xmm11 mulps %xmm3, %xmm8 mulps %xmm3, %xmm9 mulps %xmm3, %xmm10 mulps %xmm3, %xmm11#ifndef TRMMKERNEL movsd 0 * SIZE(CO1), %xmm0 movhps 2 * SIZE(CO1), %xmm0 movsd 0 * SIZE(CO2), %xmm2 movhps 2 * SIZE(CO2), %xmm2 movsd 0 * SIZE(CO1, LDC, 2), %xmm4 movhps 2 * SIZE(CO1, LDC, 2), %xmm4 movsd 0 * SIZE(CO2, LDC, 2), %xmm6 movhps 2 * SIZE(CO2, LDC, 2), %xmm6 addps %xmm0, %xmm8 addps %xmm2, %xmm9 addps %xmm4, %xmm10 addps %xmm6, %xmm11#endif movsd %xmm8, 0 * SIZE(CO1) movhps %xmm8, 2 * SIZE(CO1) movsd %xmm9, 0 * SIZE(CO2) movhps %xmm9, 2 * SIZE(CO2) movsd %xmm10, 0 * SIZE(CO1, LDC, 2) movhps %xmm10, 2 * SIZE(CO1, LDC, 2) movsd %xmm11, 0 * SIZE(CO2, LDC, 2) movhps %xmm11, 2 * SIZE(CO2, LDC, 2)#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, 4), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $4, KK#endif addq $4 * SIZE, CO1 addq $4 * SIZE, CO2 ALIGN_4 .L30: testq $2, M jle,pt .L40#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq B, BO#else movq B, BO movq KK, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 4), BO#endif movaps -32 * SIZE(AO), %xmm0 pxor %xmm3, %xmm3 pxor %xmm4, %xmm4 movaps -32 * SIZE(BO), %xmm2 pxor %xmm5, %xmm5 pxor %xmm6, %xmm6 movaps %xmm4, %xmm8 movaps %xmm4, %xmm9 movaps %xmm4, %xmm10 movaps %xmm4, %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 $4, %rax#endif movq %rax, KKK#endif sarq $2, %rax jle,pn .L35 ALIGN_3.L32: PREFETCH (PREFETCHSIZE + 0) * SIZE(AO) pshufd $0x44, %xmm0, %xmm1 addps %xmm3, %xmm8 pshufd $0x50, %xmm2, %xmm3 mulps %xmm1, %xmm3 addps %xmm4, %xmm9 pshufd $0xfa, %xmm2, %xmm4 movaps -28 * SIZE(BO), %xmm2 mulps %xmm1, %xmm4 pshufd $0xee, %xmm0, %xmm1 movaps -28 * SIZE(AO), %xmm0 addps %xmm3, %xmm10 pshufd $0x50, %xmm2, %xmm3 mulps %xmm1, %xmm3 addps %xmm4, %xmm11 pshufd $0xfa, %xmm2, %xmm4 movaps -24 * SIZE(BO), %xmm2 mulps %xmm1, %xmm4 pshufd $0x44, %xmm0, %xmm1 addps %xmm3, %xmm8 pshufd $0x50, %xmm2, %xmm3 mulps %xmm1, %xmm3 addps %xmm4, %xmm9 pshufd $0xfa, %xmm2, %xmm4 movaps -20 * SIZE(BO), %xmm2 mulps %xmm1, %xmm4 pshufd $0xee, %xmm0, %xmm1 movaps -24 * SIZE(AO), %xmm0 addps %xmm3, %xmm10 pshufd $0x50, %xmm2, %xmm3 mulps %xmm1, %xmm3 addps %xmm4, %xmm11 pshufd $0xfa, %xmm2, %xmm4 movaps -16 * SIZE(BO), %xmm2 mulps %xmm1, %xmm4 subq $-8 * SIZE, AO subq $-16 * SIZE, BO subq $1, %rax jg,pt .L32 ALIGN_3.L35:#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif andq $3, %rax je,pt .L38 ALIGN_3.L36: pshufd $0x44, %xmm0, %xmm1 movsd -30 * SIZE(AO), %xmm0 addps %xmm3, %xmm8 pshufd $0x50, %xmm2, %xmm3 mulps %xmm1, %xmm3 addps %xmm4, %xmm9 pshufd $0xfa, %xmm2, %xmm4 movaps -28 * SIZE(BO), %xmm2 mulps %xmm1, %xmm4 addq $2 * SIZE, AO addq $4 * SIZE, BO subq $1, %rax jg,pt .L36 ALIGN_3.L38: movddup ALPHA, %xmm2 addps %xmm10, %xmm8 addps %xmm11, %xmm9 addps %xmm3, %xmm8 addps %xmm4, %xmm9 mulps %xmm2, %xmm8 mulps %xmm2, %xmm9#ifndef TRMMKERNEL movsd 0 * SIZE(CO1), %xmm0 movhps 0 * SIZE(CO2), %xmm0 movsd 0 * SIZE(CO1, LDC, 2), %xmm1 movhps 0 * SIZE(CO2, LDC, 2), %xmm1 addps %xmm0, %xmm8 addps %xmm1, %xmm9#endif movsd %xmm8, 0 * SIZE(CO1) movhps %xmm8, 0 * SIZE(CO2) movsd %xmm9, 0 * SIZE(CO1, LDC, 2) movhps %xmm9, 0 * SIZE(CO2, LDC, 2)#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, 4), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $2, KK#endif addq $2 * SIZE, CO1 addq $2 * SIZE, CO2 ALIGN_4 .L40: testq $1, M jle,pt .L49#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq B, BO#else movq B, BO movq KK, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 1), AO leaq (BO, %rax, 4), BO#endif movsd -32 * SIZE(AO), %xmm0 pxor %xmm8, %xmm8 movaps -32 * SIZE(BO), %xmm2 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 $4, %rax#endif movq %rax, KKK#endif sarq $2, %rax jle,pn .L45 ALIGN_3.L42: PREFETCH (PREFETCHSIZE + 0) * SIZE(AO) pshufd $0x00, %xmm0, %xmm1 movss -31 * SIZE(AO), %xmm0 mulps %xmm1, %xmm2 addps %xmm2, %xmm8 movaps -28 * SIZE(BO), %xmm2 pshufd $0x00, %xmm0, %xmm1 movss -30 * SIZE(AO), %xmm0 mulps %xmm1, %xmm2 addps %xmm2, %xmm9 movaps -24 * SIZE(BO), %xmm2 pshufd $0x00, %xmm0, %xmm1 movss -29 * SIZE(AO), %xmm0 mulps %xmm1, %xmm2 addps %xmm2, %xmm8 movaps -20 * SIZE(BO), %xmm2 pshufd $0x00, %xmm0, %xmm1 movss -28 * SIZE(AO), %xmm0 mulps %xmm1, %xmm2 addps %xmm2, %xmm9 movaps -16 * SIZE(BO), %xmm2 subq $ -4 * SIZE, AO subq $-16 * SIZE, BO subq $1, %rax jg,pt .L42 ALIGN_3.L45:#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif andq $3, %rax je,pt .L48 ALIGN_3.L46: pshufd $0x00, %xmm0, %xmm1 movss -31 * SIZE(AO), %xmm0 mulps %xmm1, %xmm2 addps %xmm2, %xmm8 movaps -28 * SIZE(BO), %xmm2 addq $1 * SIZE, AO addq $4 * SIZE, BO subq $1, %rax jg,pt .L46 ALIGN_3.L48: movddup ALPHA, %xmm2 addps %xmm9, %xmm8 mulps %xmm2, %xmm8 pshufd $0xff, %xmm8, %xmm11 pshufd $0xaa, %xmm8, %xmm10 pshufd $0x55, %xmm8, %xmm9 pshufd $0x00, %xmm8, %xmm8#ifndef TRMMKERNEL addss 0 * SIZE(CO1), %xmm8 addss 0 * SIZE(CO2), %xmm9 addss 0 * SIZE(CO1, LDC, 2), %xmm10 addss 0 * SIZE(CO2, LDC, 2), %xmm11#endif movss %xmm8, 0 * SIZE(CO1) movss %xmm9, 0 * SIZE(CO2) movss %xmm10, 0 * SIZE(CO1, LDC, 2) movss %xmm11, 0 * SIZE(CO2, LDC, 2)#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, 4), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $1, KK#endif ALIGN_4.L49:#if defined(TRMMKERNEL) && !defined(LEFT) addq $4, KK#endif movq BO, B leaq (C, LDC, 4), C subq $1, J jg,pt .L10 ALIGN_4.L50: testq $2, N jle .L90#if defined(TRMMKERNEL) && defined(LEFT) movq OFFSET, %rax movq %rax, KK#endif movq C, CO1 leaq (C, LDC, 1), CO2 movq A, AO movq K, %rax salq $BASE_SHIFT + 1, %rax leaq (B, %rax), BB movq M, I sarq $3, I jle,pn .L60 ALIGN_4.L51:#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq B, BO#else movq B, BO movq KK, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 8), AO leaq (BO, %rax, 2), BO#endif prefetcht2 -32 * SIZE(BB) subq $-8 * SIZE, BB movaps -32 * SIZE(AO), %xmm0 pxor %xmm3, %xmm3 movaps -28 * SIZE(AO), %xmm1 pxor %xmm4, %xmm4 movaps -32 * SIZE(BO), %xmm2 pxor %xmm5, %xmm5 pxor %xmm6, %xmm6 prefetcht0 7 * SIZE(CO1) movaps %xmm4, %xmm8 movaps %xmm4, %xmm9 prefetcht0 7 * SIZE(CO2) movaps %xmm4, %xmm10 movaps %xmm4, %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 $8, %rax#else addq $2, %rax#endif movq %rax, KKK#endif sarq $2, %rax jle,pn .L55 ALIGN_3.L52: PREFETCH (PREFETCHSIZE + 0) * SIZE(AO) addps %xmm3, %xmm8 pshufd $0x00, %xmm2, %xmm3 mulps %xmm0, %xmm3 addps %xmm4, %xmm9 pshufd $0x55, %xmm2, %xmm4 mulps %xmm0, %xmm4 movaps -24 * SIZE(AO), %xmm0 addps %xmm5, %xmm10 pshufd $0x00, %xmm2, %xmm5 mulps %xmm1, %xmm5 addps %xmm6, %xmm11 pshufd $0x55, %xmm2, %xmm6 mulps %xmm1, %xmm6 movaps -20 * SIZE(AO), %xmm1 addps %xmm3, %xmm8 pshufd $0xaa, %xmm2, %xmm3 mulps %xmm0, %xmm3 addps %xmm4, %xmm9 pshufd $0xff, %xmm2, %xmm4 mulps %xmm0, %xmm4 movaps -16 * SIZE(AO), %xmm0 PREFETCH (PREFETCHSIZE + 16) * SIZE(AO) addps %xmm5, %xmm10 pshufd $0xaa, %xmm2, %xmm5 mulps %xmm1, %xmm5 addps %xmm6, %xmm11 pshufd $0xff, %xmm2, %xmm6 movaps -28 * SIZE(BO), %xmm2 mulps %xmm1, %xmm6 movaps -12 * SIZE(AO), %xmm1 addps %xmm3, %xmm8 pshufd $0x00, %xmm2, %xmm3 mulps %xmm0, %xmm3 addps %xmm4, %xmm9 pshufd $0x55, %xmm2, %xmm4 mulps %xmm0, %xmm4 movaps -8 * SIZE(AO), %xmm0 addps %xmm5, %xmm10 pshufd $0x00, %xmm2, %xmm5 mulps %xmm1, %xmm5 addps %xmm6, %xmm11 pshufd $0x55, %xmm2, %xmm6 mulps %xmm1, %xmm6 movaps -4 * SIZE(AO), %xmm1 addps %xmm3, %xmm8 pshufd $0xaa, %xmm2, %xmm3 mulps %xmm0, %xmm3 addps %xmm4, %xmm9 pshufd $0xff, %xmm2, %xmm4 mulps %xmm0, %xmm4 movaps 0 * SIZE(AO), %xmm0 addps %xmm5, %xmm10 pshufd $0xaa, %xmm2, %xmm5 mulps %xmm1, %xmm5 addps %xmm6, %xmm11 pshufd $0xff, %xmm2, %xmm6 movaps -24 * SIZE(BO), %xmm2 mulps %xmm1, %xmm6 movaps 4 * SIZE(AO), %xmm1 subq $-32 * SIZE, AO subq $ -8 * SIZE, BO subq $1, %rax jg,pt .L52 ALIGN_3.L55:#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif andq $3, %rax # if (k & 1) je,pt .L58 ALIGN_3.L56: addps %xmm3, %xmm8 pshufd $0x00, %xmm2, %xmm3 mulps %xmm0, %xmm3 addps %xmm4, %xmm9 pshufd $0x55, %xmm2, %xmm4 mulps %xmm0, %xmm4 movaps -24 * SIZE(AO), %xmm0 addps %xmm5, %xmm10 pshufd $0x00, %xmm2, %xmm5 mulps %xmm1, %xmm5 addps %xmm6, %xmm11 pshufd $0x55, %xmm2, %xmm6 movsd -30 * SIZE(BO), %xmm2 mulps %xmm1, %xmm6 movaps -20 * SIZE(AO), %xmm1 addq $8 * SIZE, AO addq $2 * SIZE, BO subq $1, %rax jg,pt .L56 ALIGN_3.L58: movddup ALPHA, %xmm7 addps %xmm3, %xmm8 addps %xmm4, %xmm9 addps %xmm5, %xmm10 addps %xmm6, %xmm11 mulps %xmm7, %xmm8 mulps %xmm7, %xmm9 mulps %xmm7, %xmm10 mulps %xmm7, %xmm11#ifndef TRMMKERNEL movsd 0 * SIZE(CO1), %xmm0 movhps 2 * SIZE(CO1), %xmm0 movsd 4 * SIZE(CO1), %xmm1 movhps 6 * SIZE(CO1), %xmm1 movsd 0 * SIZE(CO2), %xmm2 movhps 2 * SIZE(CO2), %xmm2 movsd 4 * SIZE(CO2), %xmm3 movhps 6 * SIZE(CO2), %xmm3 addps %xmm0, %xmm8 addps %xmm1, %xmm10 addps %xmm2, %xmm9 addps %xmm3, %xmm11#endif movsd %xmm8, 0 * SIZE(CO1) movhps %xmm8, 2 * SIZE(CO1) movsd %xmm10, 4 * SIZE(CO1) movhps %xmm10, 6 * SIZE(CO1) movsd %xmm9, 0 * SIZE(CO2) movhps %xmm9, 2 * SIZE(CO2) movsd %xmm11, 4 * SIZE(CO2) movhps %xmm11, 6 * 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, 8), AO leaq (BO, %rax, 2), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $8, KK#endif addq $8 * SIZE, CO1 addq $8 * SIZE, CO2 decq I jg,pt .L51 ALIGN_4 .L60: testq $4, M jle,pt .L70#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq B, BO#else movq B, BO movq KK, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 4), AO leaq (BO, %rax, 2), BO#endif movaps -32 * SIZE(AO), %xmm0 pxor %xmm3, %xmm3 pxor %xmm4, %xmm4 movaps -32 * SIZE(BO), %xmm2 pxor %xmm8, %xmm8 pxor %xmm9, %xmm9 pxor %xmm10, %xmm10 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 $4, %rax#else addq $2, %rax#endif movq %rax, KKK#endif sarq $2, %rax jle,pn .L65 ALIGN_3.L62: PREFETCH (PREFETCHSIZE + 0) * SIZE(AO) addps %xmm3, %xmm8 pshufd $0x00, %xmm2, %xmm3 mulps %xmm0, %xmm3 addps %xmm4, %xmm9 pshufd $0x55, %xmm2, %xmm4 mulps %xmm0, %xmm4 movaps -28 * SIZE(AO), %xmm0 addps %xmm3, %xmm10 pshufd $0xaa, %xmm2, %xmm3 mulps %xmm0, %xmm3 addps %xmm4, %xmm11 pshufd $0xff, %xmm2, %xmm4 movaps -28 * SIZE(BO), %xmm2 mulps %xmm0, %xmm4 movaps -24 * SIZE(AO), %xmm0 addps %xmm3, %xmm8 pshufd $0x00, %xmm2, %xmm3 mulps %xmm0, %xmm3 addps %xmm4, %xmm9 pshufd $0x55, %xmm2, %xmm4 mulps %xmm0, %xmm4 movaps -20 * SIZE(AO), %xmm0 addps %xmm3, %xmm10 pshufd $0xaa, %xmm2, %xmm3 mulps %xmm0, %xmm3 addps %xmm4, %xmm11 pshufd $0xff, %xmm2, %xmm4 movaps -24 * SIZE(BO), %xmm2 mulps %xmm0, %xmm4 movaps -16 * SIZE(AO), %xmm0 subq $-16 * SIZE, AO subq $ -8 * SIZE, BO subq $1, %rax jg,pt .L62 ALIGN_3.L65:#ifndef TRMMKERNEL movq K, %rax#else movq KKK, %rax#endif andq $3, %rax je,pt .L68 ALIGN_3.L66: addps %xmm3, %xmm8 pshufd $0x00, %xmm2, %xmm3 mulps %xmm0, %xmm3 addps %xmm4, %xmm9 pshufd $0x55, %xmm2, %xmm4 movsd -30 * SIZE(BO), %xmm2 mulps %xmm0, %xmm4 movaps -28 * SIZE(AO), %xmm0 addq $4 * SIZE, AO addq $2 * SIZE, BO subq $1, %rax jg,pt .L66 ALIGN_3.L68: movddup ALPHA, %xmm7 addps %xmm10, %xmm8 addps %xmm11, %xmm9 addps %xmm3, %xmm8 addps %xmm4, %xmm9 mulps %xmm7, %xmm8 mulps %xmm7, %xmm9#ifndef TRMMKERNEL movsd 0 * SIZE(CO1), %xmm0 movhps 2 * SIZE(CO1), %xmm0 movsd 0 * SIZE(CO2), %xmm2 movhps 2 * SIZE(CO2), %xmm2 addps %xmm0, %xmm8 addps %xmm2, %xmm9#endif movsd %xmm8, 0 * SIZE(CO1) movhps %xmm8, 2 * SIZE(CO1) movsd %xmm9, 0 * SIZE(CO2) movhps %xmm9, 2 * 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, 4), AO leaq (BO, %rax, 2), BO#endif#if defined(TRMMKERNEL) && defined(LEFT) addq $4, KK#endif addq $4 * SIZE, CO1 addq $4 * SIZE, CO2 ALIGN_4 .L70: testq $2, M jle,pt .L80#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movq B, BO#else movq B, BO movq KK, %rax leaq (, %rax, SIZE), %rax leaq (AO, %rax, 2), AO leaq (BO, %rax, 2), BO#endif movaps -32 * SIZE(AO), %xmm0 pxor %xmm3, %xmm3 movaps -32 * SIZE(BO), %xmm2 pxor %xmm8, %xmm8
⌨️ 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 + -