atl_dmm6x1x72_sse2_k.c
来自「基于Blas CLapck的.用过的人知道是干啥的」· C语言 代码 · 共 1,063 行 · 第 1/2 页
C
1,063 行
#if KB > 30 movapd 240-120(pB0), rB0 movapd 240-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 240-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 240-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 240-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 240-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 240-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 32 movapd 256-120(pB0), rB0 movapd 256-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 256-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 256-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 256-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 256-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 256-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 34 movapd 272-120(pB0), rB0 movapd 272-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 272-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 272-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 272-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 272-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 272-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 36 movapd 288-120(pB0), rB0 movapd 288-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 288-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 288-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 288-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 288-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 288-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 38 movapd 304-120(pB0), rB0 movapd 304-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 304-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 304-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 304-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 304-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 304-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 40 movapd 320-120(pB0), rB0 movapd 320-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 320-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 320-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 320-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 320-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 320-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 42 movapd 336-120(pB0), rB0 movapd 336-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 336-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 336-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 336-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 336-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 336-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 44 movapd 352-120(pB0), rB0 movapd 352-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 352-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 352-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 352-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 352-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 352-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 46 movapd 368-120(pB0), rB0 movapd 368-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 368-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 368-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 368-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 368-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 368-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 48 movapd 384-120(pB0), rB0 movapd 384-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 384-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 384-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 384-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 384-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 384-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 50 movapd 400-120(pB0), rB0 movapd 400-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 400-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 400-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 400-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 400-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 400-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 52 movapd 416-120(pB0), rB0 movapd 416-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 416-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 416-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 416-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 416-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 416-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 54 movapd 432-120(pB0), rB0 movapd 432-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 432-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 432-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 432-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 432-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 432-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 56 movapd 448-120(pB0), rB0 movapd 448-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 448-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 448-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 448-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 448-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 448-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 58 movapd 464-120(pB0), rB0 movapd 464-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 464-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 464-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 464-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 464-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 464-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 60 movapd 480-120(pB0), rB0 movapd 480-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 480-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 480-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 480-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 480-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 480-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 62 movapd 496-120(pB0), rB0 movapd 496-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 496-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 496-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 496-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 496-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 496-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 64 movapd 512-120(pB0), rB0 movapd 512-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 512-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 512-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 512-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 512-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 512-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 66 movapd 528-120(pB0), rB0 movapd 528-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 528-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 528-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 528-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 528-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 528-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif pref2((pfA)) addl $52, pfA#if KB > 68 movapd 544-120(pB0), rB0 movapd 544-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 544-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 544-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 544-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 544-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 544-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 70 movapd 560-120(pB0), rB0 movapd 560-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 560-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 560-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 560-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 560-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 560-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif/* * 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, rB0 unpcklpd rC1, rC0 /* rC0 = c0a c1a */ unpcklpd rC5, rC4 /* rC4 = c4a c5a */ unpckhpd rC1, rA0 /* rA0 = c0b c1b */ unpckhpd rC5, rB0 /* rB0 = c4b c5b */ addpd rA0, rC0 /* rC0 = c0ab c1ab */ addpd rB0, rC4 /* rC4 = c4ab c5ab */ movapd rC2, rA0 unpcklpd rC3, rC2 /* rC2 = c2a c3a */ addl $NB6so, pA0 addl $NB6so, pA3 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 */ #ifdef DCPLX addl $96, pC #else addl $48, pC #endif/* * while (pA != stM); */ subb $6, stM jnz UMLOOP/* * pC += incCn; pA -= NBNB; pB += NB; */ addl COFF(%esp), pC addl $NBso, pB0#if MB == 0 subl COFF+8(%esp), pA0 lea NBso(pA0,ldab,2), pA3#else subl $MBKBso, pA0 subl $MBKBso, pA3#endif/* * while (pB != stN); */ sub $1, stN jnz UNLOOP/* * 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 $28, %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?