atl_dmm6x1x60_sse2.c
来自「基于Blas CLapck的.用过的人知道是干啥的」· C语言 代码 · 共 1,489 行 · 第 1/3 页
C
1,489 行
movapd 128(pA), rA0 mulpd 128(pB), rA0 addpd rA0, rC0 movapd 128+NBso(pA), rA0 mulpd 128(pB), rA0 addpd rA0, rC1 movapd 128+NB2so(pA), rA0 mulpd 128(pB), rA0 addpd rA0, rC2 movapd 128+NB3so(pA), rA0 mulpd 128(pB), rA0 addpd rA0, rC3 movapd 128+NB4so(pA), rA0 mulpd 128(pB), rA0 addpd rA0, rC4 movapd 128+NB5so(pA), rA0 mulpd 128(pB), rA0 addpd rA0, rC5 movapd 144(pA), rA0 mulpd 144(pB), rA0 addpd rA0, rC0 movapd 144+NBso(pA), rA0 mulpd 144(pB), rA0 addpd rA0, rC1 movapd 144+NB2so(pA), rA0 mulpd 144(pB), rA0 addpd rA0, rC2 movapd 144+NB3so(pA), rA0 mulpd 144(pB), rA0 addpd rA0, rC3 movapd 144+NB4so(pA), rA0 mulpd 144(pB), rA0 addpd rA0, rC4 movapd 144+NB5so(pA), rA0 mulpd 144(pB), rA0 addpd rA0, rC5 movapd 160(pA), rA0 mulpd 160(pB), rA0 addpd rA0, rC0 movapd 160+NBso(pA), rA0 mulpd 160(pB), rA0 addpd rA0, rC1 movapd 160+NB2so(pA), rA0 mulpd 160(pB), rA0 addpd rA0, rC2 movapd 160+NB3so(pA), rA0 mulpd 160(pB), rA0 addpd rA0, rC3 movapd 160+NB4so(pA), rA0 mulpd 160(pB), rA0 addpd rA0, rC4 movapd 160+NB5so(pA), rA0 mulpd 160(pB), rA0 addpd rA0, rC5 movapd 176(pA), rA0 mulpd 176(pB), rA0 addpd rA0, rC0 movapd 176+NBso(pA), rA0 mulpd 176(pB), rA0 addpd rA0, rC1 movapd 176+NB2so(pA), rA0 mulpd 176(pB), rA0 addpd rA0, rC2 movapd 176+NB3so(pA), rA0 mulpd 176(pB), rA0 addpd rA0, rC3 movapd 176+NB4so(pA), rA0 mulpd 176(pB), rA0 addpd rA0, rC4 movapd 176+NB5so(pA), rA0 mulpd 176(pB), rA0 addpd rA0, rC5 movapd 192(pA), rA0 mulpd 192(pB), rA0 addpd rA0, rC0 movapd 192+NBso(pA), rA0 mulpd 192(pB), rA0 addpd rA0, rC1 movapd 192+NB2so(pA), rA0 mulpd 192(pB), rA0 addpd rA0, rC2 movapd 192+NB3so(pA), rA0 mulpd 192(pB), rA0 addpd rA0, rC3 movapd 192+NB4so(pA), rA0 mulpd 192(pB), rA0 addpd rA0, rC4 movapd 192+NB5so(pA), rA0 mulpd 192(pB), rA0 addpd rA0, rC5 movapd 208(pA), rA0 mulpd 208(pB), rA0 addpd rA0, rC0 movapd 208+NBso(pA), rA0 mulpd 208(pB), rA0 addpd rA0, rC1 movapd 208+NB2so(pA), rA0 mulpd 208(pB), rA0 addpd rA0, rC2 movapd 208+NB3so(pA), rA0 mulpd 208(pB), rA0 addpd rA0, rC3 movapd 208+NB4so(pA), rA0 mulpd 208(pB), rA0 addpd rA0, rC4 movapd 208+NB5so(pA), rA0 mulpd 208(pB), rA0 addpd rA0, rC5 movapd 224(pA), rA0 mulpd 224(pB), rA0 addpd rA0, rC0 movapd 224+NBso(pA), rA0 mulpd 224(pB), rA0 addpd rA0, rC1 movapd 224+NB2so(pA), rA0 mulpd 224(pB), rA0 addpd rA0, rC2 movapd 224+NB3so(pA), rA0 mulpd 224(pB), rA0 addpd rA0, rC3 movapd 224+NB4so(pA), rA0 mulpd 224(pB), rA0 addpd rA0, rC4 movapd 224+NB5so(pA), rA0 mulpd 224(pB), rA0 addpd rA0, rC5 movapd 240(pA), rA0 mulpd 240(pB), rA0 addpd rA0, rC0 movapd 240+NBso(pA), rA0 mulpd 240(pB), rA0 addpd rA0, rC1 movapd 240+NB2so(pA), rA0 mulpd 240(pB), rA0 addpd rA0, rC2 movapd 240+NB3so(pA), rA0 mulpd 240(pB), rA0 addpd rA0, rC3 movapd 240+NB4so(pA), rA0 mulpd 240(pB), rA0 addpd rA0, rC4 movapd 240+NB5so(pA), rA0 mulpd 240(pB), rA0 addpd rA0, rC5 movapd 256(pA), rA0 mulpd 256(pB), rA0 addpd rA0, rC0 movapd 256+NBso(pA), rA0 mulpd 256(pB), rA0 addpd rA0, rC1 movapd 256+NB2so(pA), rA0 mulpd 256(pB), rA0 addpd rA0, rC2 movapd 256+NB3so(pA), rA0 mulpd 256(pB), rA0 addpd rA0, rC3 movapd 256+NB4so(pA), rA0 mulpd 256(pB), rA0 addpd rA0, rC4 movapd 256+NB5so(pA), rA0 mulpd 256(pB), rA0 addpd rA0, rC5 movapd 272(pA), rA0 mulpd 272(pB), rA0 addpd rA0, rC0 movapd 272+NBso(pA), rA0 mulpd 272(pB), rA0 addpd rA0, rC1 movapd 272+NB2so(pA), rA0 mulpd 272(pB), rA0 addpd rA0, rC2 movapd 272+NB3so(pA), rA0 mulpd 272(pB), rA0 addpd rA0, rC3 movapd 272+NB4so(pA), rA0 mulpd 272(pB), rA0 addpd rA0, rC4 movapd 272+NB5so(pA), rA0 mulpd 272(pB), rA0 addpd rA0, rC5 movapd 288(pA), rA0 mulpd 288(pB), rA0 addpd rA0, rC0 movapd 288+NBso(pA), rA0 mulpd 288(pB), rA0 addpd rA0, rC1 movapd 288+NB2so(pA), rA0 mulpd 288(pB), rA0 addpd rA0, rC2 movapd 288+NB3so(pA), rA0 mulpd 288(pB), rA0 addpd rA0, rC3 movapd 288+NB4so(pA), rA0 mulpd 288(pB), rA0 addpd rA0, rC4 movapd 288+NB5so(pA), rA0 mulpd 288(pB), rA0 addpd rA0, rC5 movapd 304(pA), rA0 mulpd 304(pB), rA0 addpd rA0, rC0 movapd 304+NBso(pA), rA0 mulpd 304(pB), rA0 addpd rA0, rC1 movapd 304+NB2so(pA), rA0 mulpd 304(pB), rA0 addpd rA0, rC2 movapd 304+NB3so(pA), rA0 mulpd 304(pB), rA0 addpd rA0, rC3 movapd 304+NB4so(pA), rA0 mulpd 304(pB), rA0 addpd rA0, rC4 movapd 304+NB5so(pA), rA0 mulpd 304(pB), rA0 addpd rA0, rC5 movapd 320(pA), rA0 mulpd 320(pB), rA0 addpd rA0, rC0 movapd 320+NBso(pA), rA0 mulpd 320(pB), rA0 addpd rA0, rC1 movapd 320+NB2so(pA), rA0 mulpd 320(pB), rA0 addpd rA0, rC2 movapd 320+NB3so(pA), rA0 mulpd 320(pB), rA0 addpd rA0, rC3 movapd 320+NB4so(pA), rA0 mulpd 320(pB), rA0 addpd rA0, rC4 movapd 320+NB5so(pA), rA0 mulpd 320(pB), rA0 addpd rA0, rC5 movapd 336(pA), rA0 mulpd 336(pB), rA0 addpd rA0, rC0 movapd 336+NBso(pA), rA0 mulpd 336(pB), rA0 addpd rA0, rC1 movapd 336+NB2so(pA), rA0 mulpd 336(pB), rA0 addpd rA0, rC2 movapd 336+NB3so(pA), rA0 mulpd 336(pB), rA0 addpd rA0, rC3 movapd 336+NB4so(pA), rA0 mulpd 336(pB), rA0 addpd rA0, rC4 movapd 336+NB5so(pA), rA0 mulpd 336(pB), rA0 addpd rA0, rC5 movapd 352(pA), rA0 mulpd 352(pB), rA0 addpd rA0, rC0 movapd 352+NBso(pA), rA0 mulpd 352(pB), rA0 addpd rA0, rC1 movapd 352+NB2so(pA), rA0 mulpd 352(pB), rA0 addpd rA0, rC2 movapd 352+NB3so(pA), rA0 mulpd 352(pB), rA0 addpd rA0, rC3 movapd 352+NB4so(pA), rA0 mulpd 352(pB), rA0 addpd rA0, rC4 movapd 352+NB5so(pA), rA0 mulpd 352(pB), rA0 addpd rA0, rC5 movapd 368(pA), rA0 mulpd 368(pB), rA0 addpd rA0, rC0 movapd 368+NBso(pA), rA0 mulpd 368(pB), rA0 addpd rA0, rC1 movapd 368+NB2so(pA), rA0 mulpd 368(pB), rA0 addpd rA0, rC2 movapd 368+NB3so(pA), rA0 mulpd 368(pB), rA0 addpd rA0, rC3 movapd 368+NB4so(pA), rA0 mulpd 368(pB), rA0 addpd rA0, rC4 movapd 368+NB5so(pA), rA0 mulpd 368(pB), rA0 addpd rA0, rC5 movapd 384(pA), rA0 mulpd 384(pB), rA0 addpd rA0, rC0 movapd 384+NBso(pA), rA0 mulpd 384(pB), rA0 addpd rA0, rC1 movapd 384+NB2so(pA), rA0 mulpd 384(pB), rA0 addpd rA0, rC2 movapd 384+NB3so(pA), rA0 mulpd 384(pB), rA0 addpd rA0, rC3 movapd 384+NB4so(pA), rA0 mulpd 384(pB), rA0 addpd rA0, rC4 movapd 384+NB5so(pA), rA0 mulpd 384(pB), rA0 addpd rA0, rC5 movapd 400(pA), rA0 mulpd 400(pB), rA0 addpd rA0, rC0 movapd 400+NBso(pA), rA0 mulpd 400(pB), rA0 addpd rA0, rC1 movapd 400+NB2so(pA), rA0 mulpd 400(pB), rA0 addpd rA0, rC2 movapd 400+NB3so(pA), rA0 mulpd 400(pB), rA0 addpd rA0, rC3 movapd 400+NB4so(pA), rA0 mulpd 400(pB), rA0 addpd rA0, rC4 movapd 400+NB5so(pA), rA0 mulpd 400(pB), rA0 addpd rA0, rC5 prefB(256+NBso(pB)) prefB(320+NBso(pB)) movapd 416(pA), rA0 mulpd 416(pB), rA0 addpd rA0, rC0 movapd 416+NBso(pA), rA0 mulpd 416(pB), rA0 addpd rA0, rC1 movapd 416+NB2so(pA), rA0 mulpd 416(pB), rA0 addpd rA0, rC2 movapd 416+NB3so(pA), rA0 mulpd 416(pB), rA0 addpd rA0, rC3 movapd 416+NB4so(pA), rA0 mulpd 416(pB), rA0 addpd rA0, rC4 movapd 416+NB5so(pA), rA0 mulpd 416(pB), rA0 addpd rA0, rC5 movapd 432(pA), rA0 mulpd 432(pB), rA0 addpd rA0, rC0 movapd 432+NBso(pA), rA0 mulpd 432(pB), rA0 addpd rA0, rC1 movapd 432+NB2so(pA), rA0 mulpd 432(pB), rA0 addpd rA0, rC2 movapd 432+NB3so(pA), rA0 mulpd 432(pB), rA0 addpd rA0, rC3 movapd 432+NB4so(pA), rA0 mulpd 432(pB), rA0 addpd rA0, rC4 movapd 432+NB5so(pA), rA0 mulpd 432(pB), rA0 addpd rA0, rC5 movapd 448(pA), rA0 mulpd 448(pB), rA0 addpd rA0, rC0 movapd 448+NBso(pA), rA0 mulpd 448(pB), rA0 addpd rA0, rC1 movapd 448+NB2so(pA), rA0 mulpd 448(pB), rA0 addpd rA0, rC2 movapd 448+NB3so(pA), rA0 mulpd 448(pB), rA0 addpd rA0, rC3 movapd 448+NB4so(pA), rA0 mulpd 448(pB), rA0 addpd rA0, rC4 movapd 448+NB5so(pA), rA0 mulpd 448(pB), rA0 addpd rA0, rC5 movapd 464(pA), rA0 mulpd 464(pB), rA0 addpd rA0, rC0 movapd 464+NBso(pA), rA0 mulpd 464(pB), rA0 addpd rA0, rC1 movapd 464+NB2so(pA), rA0 mulpd 464(pB), rA0 addpd rA0, rC2 movapd 464+NB3so(pA), rA0 mulpd 464(pB), rA0 addpd rA0, rC3 prefC((pC,incCn)) movapd 464+NB4so(pA), rA0 mulpd 464(pB), rA0 addpd rA0, rC4 movapd 464+NB5so(pA), rA0 mulpd 464(pB), rA0 addpd rA0, rC5/* * Get these bastard things summed up correctly */ /* rC0 = c0a c0b */ /* rC1 = c1a c1b */ /* rC2 = c2a c2b */ /* rC3 = c3a c3b *//* */ /* rC4 = c4a c4b */ /* rC5 = c5a c5b */ movapd rC0, rA0 movapd rC4, rA1 unpcklpd rC1, rC0 /* rC0 = c0a c1a */ unpcklpd rC5, rC4 /* rC4 = c4a c5a */ unpckhpd rC1, rA0 /* rA0 = c0b c1b */ unpckhpd rC5, rA1 /* rA1 = c4b c5b */ addpd rA0, rC0 /* rC0 = c0ab c1ab */ addpd rA1, rC4 /* rC4 = c4ab c5ab */ movapd rC2, rA0 unpcklpd rC3, rC2 /* rC2 = c2a c3a */ prefB(384+NBso(pB)) prefB(448+NBso(pB)) unpckhpd rC3, rA0 /* rA0 = c2b c3b */ addpd rA0, rC2 /* rC2 = c2ab c3ab *//* * Write results back to C */ #ifdef DCPLX movlpd rC0, (pC) movhpd rC0, 16(pC) movlpd rC2, 32(pC) movhpd rC2, 48(pC) movlpd rC4, 64(pC) movhpd rC4, 80(pC) #else movupd rC0, (pC) movupd rC2, 16(pC) movupd rC4, 32(pC) #endif/* * pC += 6; pA += 6*NB *//* addl $NB6so, pA *//* * pC += incCn; pA -= NBNB; pB += NB; */ addl incCn, pC subl $NBNBso-NB6so, pA addl $NBso, pB/* * while (pB != stN); */ cmp pB, stN jne NLOOP/* * Restore callee-saved iregs */ #ifdef BETAX movl 32(%esp), %ebp movl 28(%esp), %ebx movl 24(%esp), %esi movl 20(%esp), %edi movl 16(%esp), %esp #else movl 12(%esp), %ebp movl 8(%esp), %ebx movl 4(%esp), %esi movl (%esp), %edi addl $16, %esp #endif ret
⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?