amd3dx.h

这是一组DCT和iDCT的代码

#define _k3d_name_xlat_m6 _mm6
#define _k3d_name_xlat_m7 _mm7
#define _k3d_name_xlat_M0 _mm0
#define _k3d_name_xlat_M1 _mm1
#define _k3d_name_xlat_M2 _mm2
#define _k3d_name_xlat_M3 _mm3
#define _k3d_name_xlat_M4 _mm4
#define _k3d_name_xlat_M5 _mm5
#define _k3d_name_xlat_M6 _mm6
#define _k3d_name_xlat_M7 _mm7
#define _k3d_name_xlat_mm0 _mm0
#define _k3d_name_xlat_mm1 _mm1
#define _k3d_name_xlat_mm2 _mm2
#define _k3d_name_xlat_mm3 _mm3
#define _k3d_name_xlat_mm4 _mm4
#define _k3d_name_xlat_mm5 _mm5
#define _k3d_name_xlat_mm6 _mm6
#define _k3d_name_xlat_mm7 _mm7
#define _k3d_name_xlat_MM0 _mm0
#define _k3d_name_xlat_MM1 _mm1
#define _k3d_name_xlat_MM2 _mm2
#define _k3d_name_xlat_MM3 _mm3
#define _k3d_name_xlat_MM4 _mm4
#define _k3d_name_xlat_MM5 _mm5
#define _k3d_name_xlat_MM6 _mm6
#define _k3d_name_xlat_MM7 _mm7
#define _k3d_name_xlat_eax _eax
#define _k3d_name_xlat_ebx _ebx
#define _k3d_name_xlat_ecx _ecx
#define _k3d_name_xlat_edx _edx
#define _k3d_name_xlat_esi _esi
#define _k3d_name_xlat_edi _edi
#define _k3d_name_xlat_ebp _ebp
#define _k3d_name_xlat_EAX _eax
#define _k3d_name_xlat_EBX _ebx
#define _k3d_name_xlat_ECX _ecx
#define _k3d_name_xlat_EDX _edx
#define _k3d_name_xlat_ESI _esi
#define _k3d_name_xlat_EDI _edi
#define _k3d_name_xlat_EBP _ebp
#define _k3d_name_xlat__eax _eax
#define _k3d_name_xlat__ebx _ebx
#define _k3d_name_xlat__ecx _ecx
#define _k3d_name_xlat__edx _edx
#define _k3d_name_xlat__esi _esi
#define _k3d_name_xlat__edi _edi
#define _k3d_name_xlat__ebp _ebp
#define _k3d_name_xlat__EAX _eax
#define _k3d_name_xlat__EBX _ebx
#define _k3d_name_xlat__ECX _ecx
#define _k3d_name_xlat__EDX _edx
#define _k3d_name_xlat__ESI _esi
#define _k3d_name_xlat__EDI _edi
#define _k3d_name_xlat__EBP _ebp

#define _k3d_xglue3(a,b,c) a##b##c
#define _k3d_glue3(a,b,c) _k3d_xglue3(a,b,c)
#define _k3d_MODRM(dst, src) _k3d_glue3(_k3d,_k3d_name_xlat_##dst,_k3d_name_xlat_##src)
#define _k3d_MODRMbo(dst, src) _k3d_glue3(_k3dbo,_k3d_name_xlat_##dst,_k3d_name_xlat_##src)

/* Operand defines for prefetch and prefetchw */

#define _k3d_pref_eax 0x00
#define _k3d_pref_ecx 0x01
#define _k3d_pref_edx 0x02
#define _k3d_pref_ebx 0x03
#define _k3d_pref_esi 0x06
#define _k3d_pref_edi 0x07
#define _k3d_pref_EAX 0x00
#define _k3d_pref_ECX 0x01
#define _k3d_pref_EDX 0x02
#define _k3d_pref_EBX 0x03
#define _k3d_pref_ESI 0x06
#define _k3d_pref_EDI 0x07
#define _k3d_prefw_eax 0x08
#define _k3d_prefw_ecx 0x09
#define _k3d_prefw_edx 0x0A
#define _k3d_prefw_ebx 0x0B
#define _k3d_prefw_esi 0x0E
#define _k3d_prefw_edi 0x0F
#define _k3d_prefw_EAX 0x08
#define _k3d_prefw_ECX 0x09
#define _k3d_prefw_EDX 0x0A
#define _k3d_prefw_EBX 0x0B
#define _k3d_prefw_ESI 0x0E
#define _k3d_prefw_EDI 0x0F

/* Defines for 3DNow! instructions */
#define PF2ID(dst, src)         db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0x1d
#define PFACC(dst, src)         db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xae
#define PFADD(dst, src)         db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0x9e
#define PFCMPEQ(dst, src)       db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xb0
#define PFCMPGE(dst, src)       db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0x90
#define PFCMPGT(dst, src)       db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xa0
#define PFMAX(dst, src)         db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xa4
#define PFMIN(dst, src)         db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0x94
#define PFMUL(dst, src)         db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xb4
#define PFRCP(dst, src)         db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0x96
#define PFRCPIT1(dst, src)      db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xa6
#define PFRCPIT2(dst, src)      db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xb6
#define PFRSQRT(dst, src)       db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0x97
#define PFRSQIT1(dst, src)      db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xa7
#define PFSUB(dst, src)         db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0x9a
#define PFSUBR(dst, src)        db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xaa
#define PI2FD(dst, src)         db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0x0d
#define FEMMS                   db 0x0f, 0x0e
#define PAVGUSB(dst, src)       db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xbf
#define PMULHRW(dst, src)       db 0x0f, 0x0f, _k3d_MODRM(dst, src), 0xb7
#define PREFETCH(src)           db 0x0f, 0x0d, _k3d_pref_##src
#define PREFETCHW(src)          db 0x0f, 0x0d, _k3d_prefw_##src

/* Defines for new, K7 opcodes */
#define PFNACC(dst,src)         db 0x0f, 0x0f, _k3d_MODRM(dst,src), 0x8a
#define PFPNACC(dst,src)        db 0x0f, 0x0f, _k3d_MODRM(dst,src), 0x8e
#define PSWAPD(dst,src)         db 0x0f, 0x0f, _k3d_MODRM(dst,src), 0xbb
#define PMINUB(dst,src)         db 0x0f, 0xda, _k3d_MODRM(dst,src)
#define PMAXUB(dst,src)         db 0x0f, 0xde, _k3d_MODRM(dst,src)
#define PMINSW(dst,src)         db 0x0f, 0xea, _k3d_MODRM(dst,src)
#define PMAXSW(dst,src)         db 0x0f, 0xee, _k3d_MODRM(dst,src)
#define PMULHUW(dst,src)        db 0x0f, 0xe4, _k3d_MODRM(dst,src)
#define PAVGB(dst,src)          db 0x0f, 0xe0, _k3d_MODRM(dst,src)
#define PAVGW(dst,src)          db 0x0f, 0xe3, _k3d_MODRM(dst,src)
#define PSADBW(dst,src)         db 0x0f, 0xf6, _k3d_MODRM(dst,src)
#define PMOVMSKB(dst,src)       db 0x0f, 0xd7, _k3d_MODRM(dst,src)
#define PMASKMOVQ(dst,src)      db 0x0f, 0xf7, _k3d_MODRM(dst,src)
#define MASKMOVQ(dst,src)       db 0x0f, 0xf7, _k3d_MODRM(dst,src)
#define PINSRW(dst,src,msk)     db 0x0f, 0xc4, _k3d_MODRM(dst,src), msk
#define PEXTRW(dst,src,msk)     db 0x0f, 0xc5, _k3d_MODRM(dst,src), msk
#define PSHUFW(dst,src,msk)     db 0x0f, 0x70, _k3d_MODRM(dst,src), msk
#define MOVNTQ(dst,src)         db 0x0f, 0xe7, _k3d_MODRM(src,dst)
#define SFENCE                  db 0x0f, 0xae, 0xf8

/* Memory/offset versions of the opcodes */
#define PF2IDM(dst,src,off)     db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0x1d
#define PFACCM(dst,src,off)     db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xae
#define PFADDM(dst,src,off)     db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0x9e
#define PFCMPEQM(dst,src,off)   db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xb0
#define PFCMPGEM(dst,src,off)   db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0x90
#define PFCMPGTM(dst,src,off)   db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xa0
#define PFMAXM(dst,src,off)     db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xa4
#define PFMINM(dst,src,off)     db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0x94
#define PFMULM(dst,src,off)     db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xb4
#define PFRCPM(dst,src,off)     db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0x96
#define PFRCPIT1M(dst,src,off)  db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xa6
#define PFRCPIT2M(dst,src,off)  db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xb6
#define PFRSQRTM(dst,src,off)   db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0x97
#define PFRSQIT1M(dst,src,off)  db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xa7
#define PFSUBM(dst,src,off)     db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0x9a
#define PFSUBRM(dst,src,off)    db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xaa
#define PI2FDM(dst,src,off)     db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0x0d
#define PAVGUSBM(dst,src,off)   db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xbf
#define PMULHRWM(dst,src,off)   db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xb7


/* Memory/offset versions of the new, K7 opcodes */
#define PFNACCM(dst,src,off)        db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0x8a
#define PFPNACCM(dst,src,off)       db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0x8e
#define PSWAPDM(dst,src,off)        db 0x0f, 0x0f, _k3d_MODRMbo(dst,src), off, 0xbb
#define PMINUBM(dst,src,off)        db 0x0f, 0xda, _k3d_MODRMbo(dst,src), off
#define PMAXUBM(dst,src,off)        db 0x0f, 0xde, _k3d_MODRMbo(dst,src), off
#define PMINSWM(dst,src,off)        db 0x0f, 0xea, _k3d_MODRMbo(dst,src), off
#define PMAXSWM(dst,src,off)        db 0x0f, 0xee, _k3d_MODRMbo(dst,src), off
#define PMULHUWM(dst,src,off)       db 0x0f, 0xe4, _k3d_MODRMbo(dst,src), off
#define PAVGBM(dst,src,off)         db 0x0f, 0xe0, _k3d_MODRMbo(dst,src), off
#define PAVGWM(dst,src,off)         db 0x0f, 0xe3, _k3d_MODRMbo(dst,src), off
#define PSADBWM(dst,src,off)        db 0x0f, 0xf6, _k3d_MODRMbo(dst,src), off
#define PMOVMSKBM(dst,src,off)      db 0x0f, 0xd7, _k3d_MODRMbo(dst,src), off
#define PMASKMOVQM(dst,src,off)     db 0x0f, 0xf7, _k3d_MODRMbo(dst,src), off
#define MASKMOVQM(dst,src,off)      db 0x0f, 0xf7, _k3d_MODRMbo(dst,src), off
#define MOVNTQM(dst,src,off)        db 0x0f, 0xe7, _k3d_MODRMbo(src,dst), off
#define PINSRWM(dst,src,off,msk)    db 0x0f, 0xc4, _k3d_MODRMbo(dst,src), off, msk
#define PSHUFWM(dst,src,off,msk)    db 0x0f, 0x70, _k3d_MODRMbo(dst,src), off, msk


/* Defines for 3DNow! instructions for use in pragmas */
#define p_pf2id(dst,src)        0x0f 0x0f _k3d_MODRM(dst,src) 0x1d
#define p_pfacc(dst,src)        0x0f 0x0f _k3d_MODRM(dst,src) 0xae
#define p_pfadd(dst,src)        0x0f 0x0f _k3d_MODRM(dst,src) 0x9e
#define p_pfcmpeq(dst,src)      0x0f 0x0f _k3d_MODRM(dst,src) 0xb0
#define p_pfcmpge(dst,src)      0x0f 0x0f _k3d_MODRM(dst,src) 0x90
#define p_pfcmpgt(dst,src)      0x0f 0x0f _k3d_MODRM(dst,src) 0xa0
#define p_pfmax(dst,src)        0x0f 0x0f _k3d_MODRM(dst,src) 0xa4
#define p_pfmin(dst,src)        0x0f 0x0f _k3d_MODRM(dst,src) 0x94
#define p_pfmul(dst,src)        0x0f 0x0f _k3d_MODRM(dst,src) 0xb4
#define p_pfrcp(dst,src)        0x0f 0x0f _k3d_MODRM(dst,src) 0x96
#define p_pfrcpit1(dst,src)     0x0f 0x0f _k3d_MODRM(dst,src) 0xa6
#define p_pfrcpit2(dst,src)     0x0f 0x0f _k3d_MODRM(dst,src) 0xb6
#define p_pfrsqrt(dst,src)      0x0f 0x0f _k3d_MODRM(dst,src) 0x97
#define p_pfrsqit1(dst,src)     0x0f 0x0f _k3d_MODRM(dst,src) 0xa7
#define p_pfsub(dst,src)        0x0f 0x0f _k3d_MODRM(dst,src) 0x9a
#define p_pfsubr(dst,src)       0x0f 0x0f _k3d_MODRM(dst,src) 0xaa
#define p_pi2fd(dst,src)        0x0f 0x0f _k3d_MODRM(dst,src) 0x0d
#define p_femms                 0x0f 0x0e
#define p_pavgusb(dst,src)      0x0f 0x0f _k3d_MODRM(dst,src) 0xbf
#define p_pmulhrw(dst,src)      0x0f 0x0f _k3d_MODRM(dst,src) 0xb7
#define p_prefetch(src)         0x0f 0x0d _k3d_pref_##src
#define p_prefetchw(src)        0x0f 0x0d _k3d_prefw_##src
#define P_PFNACC(dst,src)       0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x8a
#define P_PFPNACC(dst,src)      0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x8e
#define P_PSWAPD(dst,src)       0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xbb
#define P_PMINUB(dst,src)       0x0f 0xda _k3d_MODRMbo(dst,src) off
#define P_PMAXUB(dst,src)       0x0f 0xde _k3d_MODRMbo(dst,src) off
#define P_PMINSW(dst,src)       0x0f 0xea _k3d_MODRMbo(dst,src) off
#define P_PMAXSW(dst,src)       0x0f 0xee _k3d_MODRMbo(dst,src) off
#define P_PMULHUW(dst,src)      0x0f 0xe4 _k3d_MODRMbo(dst,src) off
#define P_PAVGB(dst,src)        0x0f 0xe0 _k3d_MODRMbo(dst,src) off
#define P_PAVGW(dst,src)        0x0f 0xe3 _k3d_MODRMbo(dst,src) off
#define P_PSADBW(dst,src)       0x0f 0xf6 _k3d_MODRMbo(dst,src) off
#define P_PMOVMSKB(dst,src)     0x0f 0xd7 _k3d_MODRMbo(dst,src) off
#define P_PMASKMOVQ(dst,src)    0x0f 0xf7 _k3d_MODRMbo(dst,src) off
#define P_MASKMOVQ(dst,src)     0x0f 0xf7 _k3d_MODRMbo(dst,src) off
#define P_PINSRW(dst,src,msk)   0x0f 0xc4 _k3d_MODRMbo(dst,src) off msk
#define P_PEXTRW(dst,src,msk)   0x0f 0xc5 _k3d_MODRMbo(dst,src) off msk
#define P_PSHUFW(dst,src,msk)   0x0f 0x70 _k3d_MODRMbo(dst,src) off msk
#define P_MOVNTQ(dst,src)       0x0f 0xe7 _k3d_MODRMbo(src,dst) off

#define P_PF2IDM(dst,src,off)    0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x1d
#define P_PFACCM(dst,src,off)    0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xae
#define P_PFADDM(dst,src,off)    0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x9e
#define P_PFCMPEQM(dst,src,off)  0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xb0
#define P_PFCMPGEM(dst,src,off)  0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x90
#define P_PFCMPGTM(dst,src,off)  0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xa0
#define P_PFMAXM(dst,src,off)    0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xa4
#define P_PFMINM(dst,src,off)    0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x94
#define P_PFMULM(dst,src,off)    0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xb4
#define P_PFRCPM(dst,src,off)    0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x96
#define P_PFRCPIT1M(dst,src,off) 0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xa6
#define P_PFRCPIT2M(dst,src,off) 0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xb6
#define P_PFRSQRTM(dst,src,off)  0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x97
#define P_PFRSQIT1M(dst,src,off) 0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xa7
#define P_PFSUBM(dst,src,off)    0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x9a
#define P_PFSUBRM(dst,src,off)   0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xaa
#define P_PI2FDM(dst,src,off)    0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x0d
#define P_PAVGUSBM(dst,src,off)  0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xbf
#define P_PMULHRWM(dst,src,off)  0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xb7
#define P_PFNACCM(dst,src,off)   0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x8a
#define P_PFPNACCM(dst,src,off)  0x0f 0x0f _k3d_MODRMbo(dst,src) off 0x8e
#define P_PSWAPDM(dst,src,off)   0x0f 0x0f _k3d_MODRMbo(dst,src) off 0xbb
#define P_PMINUBM(dst,src,off)   0x0f 0xda _k3d_MODRMbo(dst,src) off
#define P_PMAXUBM(dst,src,off)   0x0f 0xde _k3d_MODRMbo(dst,src) off
#define P_PMINSWM(dst,src,off)   0x0f 0xea _k3d_MODRMbo(dst,src) off
#define P_PMAXSWM(dst,src,off)   0x0f 0xee _k3d_MODRMbo(dst,src) off
#define P_PMULHUWM(dst,src,off)  0x0f 0xe4 _k3d_MODRMbo(dst,src) off
#define P_PAVGBM(dst,src,off)    0x0f 0xe0 _k3d_MODRMbo(dst,src) off
#define P_PAVGWM(dst,src,off)    0x0f 0xe3 _k3d_MODRMbo(dst,src) off
#define P_PSADBWM(dst,src,off)   0x0f 0xf6 _k3d_MODRMbo(dst,src) off
#define P_PMOVMSKBM(dst,src,off) 0x0f 0xd7 _k3d_MODRMbo(dst,src) off
#define P_MOVNTQM(dst,src,off)   0x0f 0xe7 _k3d_MODRMbo(src,dst) off
#define P_PMASKMOVQM(dst,src,off)   0x0f 0xf7 _k3d_MODRMbo(dst,src) off
#define P_MASKMOVQM(dst,src,off)   0x0f 0xf7 _k3d_MODRMbo(dst,src) off
#define P_PINSRWM(dst,src,off,msk)  0x0f 0xc4 _k3d_MODRMbo(dst,src) off msk
#define P_PSHUFWM(dst,src,off,msk)  0x0f 0x70 _k3d_MODRMbo(dst,src) off msk


#define P_PF2ID(dst,src)            p_pf2id(dst,src)
#define P_PFACC(dst,src)            p_pfacc(dst,src)
#define P_PFADD(dst,src)            p_pfadd(dst,src)
#define P_PFCMPEQ(dst,src)          p_pfcmpeq(dst,src)
#define P_PFCMPGE(dst,src)          p_pfcmpge(dst,src)
#define P_PFCMPGT(dst,src)          p_pfcmpgt(dst,src)
#define P_PFMAX(dst,src)            p_pfmax(dst,src)
#define P_PFMIN(dst,src)            p_pfmin(dst,src)
#define P_PFMUL(dst,src)            p_pfmul(dst,src)
#define P_PFRCP(dst,src)            p_pfrcp(dst,src)
#define P_PFRCPIT1(dst,src)         p_pfrcpit1(dst,src)
#define P_PFRCPIT2(dst,src)         p_pfrcpit2(dst,src)
#define P_PFRSQRT(dst,src)          p_pfrsqrt(dst,src)
#define P_PFRSQIT1(dst,src)         p_pfrsqit1(dst,src)
#define P_PFSUB(dst,src)            p_pfsub(dst,src)
#define P_PFSUBR(dst,src)           p_pfsubr(dst,src)
#define P_PI2FD(dst,src)            p_pi2fd(dst,src)
#define P_FEMMS                     p_femms
#define P_PAVGUSB(dst,src)          p_pavgusb(dst,src)
#define P_PMULHRW(dst,src)          p_pmulhrw(dst,src)
#define P_PREFETCH(src)             p_prefetch(src)
#define P_PREFETCHW(src)            p_prefetchw(src)
#define p_pf2idm(dst,src,off)       P_PF2IDM(dst,src,off)
#define p_pfaccm(dst,src,off)       P_PFACCM(dst,src,off)
#define p_pfaddm(dst,src,off)       P_PFADDM(dst,src,off)
#define p_pfcmpeqm(dst,src,off)     P_PFCMPEQM(dst,src,off)
#define p_pfcmpgem(dst,src,off)     P_PFCMPGEM(dst,src,off)
#define p_pfcmpgtm(dst,src,off)     P_PFCMPGTM(dst,src,off)
#define p_pfmaxm(dst,src,off)       P_PFMAXM(dst,src,off)
#define p_pfminm(dst,src,off)       P_PFMINM(dst,src,off)
#define p_pfmulm(dst,src,off)       P_PFMULM(dst,src,off)
#define p_pfrcpm(dst,src,off)       P_PFRCPM(dst,src,off)
#define p_pfrcpit1m(dst,src,off)    P_PFRCPIT1M(dst,src,off)
#define p_pfrcpit2m(dst,src,off)    P_PFRCPIT2M(dst,src,off)
#define p_pfrsqrtm(dst,src,off)     P_PFRSQRTM(dst,src,off)
#define p_pfrsqit1m(dst,src,off)    P_PFRSQIT1M(dst,src,off)
#define p_pfsubm(dst,src,off)       P_PFSUBM(dst,src,off)
#define p_pfsubrm(dst,src,off)      P_PFSUBRM(dst,src,off)
#define p_pi2fdm(dst,src,off)       P_PI2FDM(dst,src,off)
#define p_pavgusbm(dst,src,off)     P_PAVGUSBM(dst,src,off)
#define p_pmulhrwm(dst,src,off)     P_PMULHRWM(dst,src,off)

#define p_pfnacc(dst,src)           P_PFNACC(dst,src)           
#define p_pfpnacc(dst,src)          P_PFPNACC(dst,src)          
#define p_pswapd(dst,src)           P_PSWAPD(dst,src)           
#define p_pminub(dst,src)           P_PMINUB(dst,src)           
#define p_pmaxub(dst,src)           P_PMAXUB(dst,src)           
#define p_pminsw(dst,src)           P_PMINSW(dst,src)           
#define p_pmaxsw(dst,src)           P_PMAXSW(dst,src)           
#define p_pmulhuw(dst,src)          P_PMULHUW(dst,src)          
#define p_pavgb(dst,src)            P_PAVGB(dst,src)            
#define p_avgw(dst,src)             P_PAVGW(dst,src)            
#define p_psadbw(dst,src)           P_PSADBW(dst,src)           
#define p_pmovmskb(dst,src)         P_PMOVMSKB(dst,src)         
#define p_pmaskmovq(dst,src)        P_PMASKMOVQ(dst,src)        
#define p_maskmovq(dst,src)			P_MASKMOVQ(dst,src)        
#define p_pinsrw(dst,src,msk)       P_PINSRW(dst,src,msk)