atl_dmm6x1x72_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 3,260 行 · 第 1/5 页
C
3,260 行
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#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#if KB > 72 movapd 576-120(pB0), rB0 movapd 576-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 576-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 576-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 576-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 576-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 576-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 74 movapd 592-120(pB0), rB0 movapd 592-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 592-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 592-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 592-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 592-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 592-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 76 movapd 608-120(pB0), rB0 movapd 608-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 608-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 608-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 608-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 608-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 608-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif#if KB > 78 movapd 624-120(pB0), rB0 movapd 624-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 624-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 624-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 624-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 624-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 624-120(pA3,ldab,2), rB0 addpd rB0, rC5#endif/* * Get these bastard things summed up correctly */#ifdef ATL_SSE3 haddpd rC1, rC0 /* rC0 = c0ab c1ab */ addl $NB6so, pA0 haddpd rC3, rC2 /* rC2 = c2ab c3ab */ addl $NB6so, pA3 haddpd rC5, rC4 /* rC4 = c4ab c3ab */#else /* 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 */#endif/* * 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#endif/* * last iteration of UMLOOP unrolled for prefetch */#if MB == 0ULMLOOP:#endif#ifdef BETA0 xorpd rC0, rC0 xorpd rC1, rC1 xorpd rC2, rC2 xorpd rC3, rC3 xorpd rC4, rC4 xorpd rC5, rC5#else #ifdef DCPLX movsd (pC), rC0 movsd 16(pC), rC1 movsd 32(pC), rC2 movsd 48(pC), rC3 movsd 64(pC), rC4 movsd 80(pC), rC5 #else movsd (pC), rC0 movsd 8(pC), rC1 movsd 16(pC), rC2 movsd 24(pC), rC3 movsd 32(pC), rC4 movsd 40(pC), rC5 #endif #ifdef BETAX #ifndef ATL_GAS_x8664 movlpd (%esp), rA0 #endif mulsd rbeta, rC0 mulsd rbeta, rC1 mulsd rbeta, rC2 mulsd rbeta, rC3 mulsd rbeta, rC4 mulsd rbeta, rC5 #endif#endif movapd 0-120(pB0), rB0 movapd 0-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 0-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 0-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 0-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 0-120(pA3,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC4 mulpd 0-120(pA3,ldab,2), rB0 addpd rB0, rC5#if KB > 2 movapd 16-120(pB0), rB0 movapd 16-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 16-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 16-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 movapd 16-120(pA3), rA0 mulpd rB0, rA0 addpd rA0, rC3 movapd 16-120(pA3,ldab), rA0 mulpd rB0, rA0 prefB((pB0,ldab)) prefB(128(pB0,ldab)) addpd rA0, rC4 mulpd 16-120(pA3,ldab,2), rB0 addpd rB0, rC5#else prefB((pB0,ldab))#endif#if KB > 4 movapd 32-120(pB0), rB0 movapd 32-120(pA0), rA0 mulpd rB0, rA0 addpd rA0, rC0 movapd 32-120(pA0,ldab), rA0 mulpd rB0, rA0 addpd rA0, rC1 movapd 32-120(pA0,ldab,2), rA0 mulpd rB0, rA0 addpd rA0, rC2 m
⌨️ 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?