diff options
Diffstat (limited to 'target/hexagon/imported/mmvec/ext.idef')
| -rw-r--r-- | target/hexagon/imported/mmvec/ext.idef | 321 |
1 files changed, 320 insertions, 1 deletions
diff --git a/target/hexagon/imported/mmvec/ext.idef b/target/hexagon/imported/mmvec/ext.idef index 8ca5a606e1..ead32c243b 100644 --- a/target/hexagon/imported/mmvec/ext.idef +++ b/target/hexagon/imported/mmvec/ext.idef @@ -1,5 +1,5 @@ /* - * Copyright(c) 2019-2021 Qualcomm Innovation Center, Inc. All Rights Reserved. + * Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved. * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -62,6 +62,9 @@ EXTINSN(V6_##TAG, SYNTAX, ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VS), \ DESCR, DO_FOR_EACH_CODE(WIDTH, CODE)) +#define ITERATOR_INSN_SHIFT3_SLOT(WIDTH,TAG,SYNTAX,DESCR,CODE) \ +EXTINSN(V6_##TAG, SYNTAX, ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VS,A_CVI_VS_3SRC,A_NOTE_SHIFT_RESOURCE,A_NOTE_NOVP,A_NOTE_VA_UNARY), \ +DESCR, DO_FOR_EACH_CODE(WIDTH, CODE)) #define ITERATOR_INSN_SHIFT_SLOT_VV_LATE(WIDTH,TAG,SYNTAX,DESCR,CODE) \ EXTINSN(V6_##TAG, SYNTAX, ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VS), \ @@ -116,6 +119,10 @@ ITERATOR_INSN_MPY_SLOT_LATE(WIDTH,TAG, SYNTAX2,DESCR,CODE) EXTINSN(V6_##TAG, SYNTAX, ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VX_DV), \ DESCR, DO_FOR_EACH_CODE(WIDTH, CODE)) +#define ITERATOR_INSN_MPY_SLOT_DOUBLE_VEC_VX_FWD(WIDTH,TAG,SYNTAX,DESCR,CODE) \ +EXTINSN(V6_##TAG, SYNTAX, ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VX_DV), \ +DESCR, DO_FOR_EACH_CODE(WIDTH, CODE)) + #define ITERATOR_INSN2_MPY_SLOT_DOUBLE_VEC(WIDTH,TAG,SYNTAX,SYNTAX2,DESCR,CODE) \ ITERATOR_INSN_MPY_SLOT_DOUBLE_VEC(WIDTH,TAG,SYNTAX2,DESCR,CODE) @@ -976,6 +983,22 @@ NARROWING_SHIFT(16,vasrhubrndsat,fSETBYTE,ub,h,:rnd:sat,fVSATUB,fVROUND,0x7) NARROWING_SHIFT(16,vasrhbsat,fSETBYTE,b,h,:sat,fVSATB,fVNOROUND,0x7) NARROWING_SHIFT(16,vasrhbrndsat,fSETBYTE,b,h,:rnd:sat,fVSATB,fVROUND,0x7) +#define NARROWING_VECTOR_SHIFT(ITERSIZE,TAG,DSTM,DSTTYPE,SRCTYPE,SRCTYPE2,SYNOPTS,SATFUNC,RNDFUNC,SHAMTMASK) \ +ITERATOR_INSN_SHIFT3_SLOT(ITERSIZE,TAG, \ +"Vd32." #DSTTYPE "=vasr(Vuu32." #SRCTYPE ",Vv32." #SRCTYPE2 ")" #SYNOPTS, \ +"Vector shift by vector right and shuffle", \ + fHIDE(int )shamt = VvV.SRCTYPE2[2*i+0] & SHAMTMASK; \ + DSTM(0,VdV.SRCTYPE[i],SATFUNC(RNDFUNC(VuuV.v[0].SRCTYPE[i],shamt) >> shamt)); \ + shamt = VvV.SRCTYPE2[2*i+1] & SHAMTMASK; \ + DSTM(1,VdV.SRCTYPE[i],SATFUNC(RNDFUNC(VuuV.v[1].SRCTYPE[i],shamt) >> shamt))) + +/* WORD TO HALF*/ +NARROWING_VECTOR_SHIFT(32,vasrvwuhsat,fSETHALF,uh,w,uh,:sat,fVSATUH,fVNOROUND,0xF) +NARROWING_VECTOR_SHIFT(32,vasrvwuhrndsat,fSETHALF,uh,w,uh,:rnd:sat,fVSATUH,fVROUND,0xF) +/* HALF TO BYTE*/ +NARROWING_VECTOR_SHIFT(16,vasrvuhubsat,fSETBYTE,ub,uh,ub,:sat,fVSATUB,fVNOROUND,0x7) +NARROWING_VECTOR_SHIFT(16,vasrvuhubrndsat,fSETBYTE,ub,uh,ub,:rnd:sat,fVSATUB,fVROUND,0x7) + NARROWING_SHIFT_NOV1(16,vasruhubsat,fSETBYTE,ub,uh,:sat,fVSATUB,fVNOROUND,0x7) NARROWING_SHIFT_NOV1(16,vasruhubrndsat,fSETBYTE,ub,uh,:rnd:sat,fVSATUB,fVROUND,0x7) @@ -1360,6 +1383,9 @@ ITERATOR_INSN2_MPY_SLOT_DOUBLE_VEC(16,vmpyhvsrs,"Vd32=vmpyh(Vu32,Vv32):<<1:rnd:s +ITERATOR_INSN_MPY_SLOT(16,vmpyuhvs, "Vd32.uh=vmpy(Vu32.uh,Vv32.uh):>>16", +"Vector by Vector Unsigned Halfword Multiply with 16 bit rightshift", + VdV.uh[i] = fGETUHALF(1,fMPY16UU(VuV.uh[i],VvV.uh[i]))) ITERATOR_INSN2_MPY_SLOT_DOUBLE_VEC(32,vmpyhus, "Vdd32=vmpyhus(Vu32,Vv32)","Vdd32.w=vmpy(Vu32.h,Vv32.uh)", @@ -2038,6 +2064,24 @@ ITERATOR_INSN_ANY_SLOT_DOUBLE_VEC(8,vcombine,"Vdd32=vcombine(Vu32,Vv32)", /////////////////////////////////////////////////////////////////////////// +EXTINSN(V6_vcombine_tmp, "Vdd32.tmp=vcombine(Vu32,Vv32)", ATTRIBS(A_EXTENSION,A_CVI,A_CVI_REMAP,A_CVI_TMP,A_NO_INTRINSIC), +"Vector assign tmp, Any two to Vector Pair ", +{ + fHIDE(int i;) + fVFOREACH(8, i) { + VddV.v[0].ub[i] = VvV.ub[i]; + VddV.v[1].ub[i] = VuV.ub[i]; + } +}) + +EXTINSN(V6_vassign_tmp, "Vd32.tmp=Vu32", ATTRIBS(A_EXTENSION,A_CVI,A_CVI_REMAP,A_CVI_TMP,A_NO_INTRINSIC), +"Vector assign tmp, Any two to Vector Pair ", +{ + fHIDE(int i;) + fVFOREACH(32, i) { + VdV.w[i]=VuV.w[i]; + } +}) /********************************************************* * GENERAL PERMUTE NETWORKS @@ -2507,6 +2551,281 @@ EXTINSN(V6_vscattermhw , "vscatter(Rt32,Mu2,Vvv32.w).h=Vw32", ATTRIBS(A_EXTENSIO }) +ITERATOR_INSN_MPY_SLOT_DOUBLE_VEC_VX_FWD(32, v6mpyvubs10_vxx, "Vxx32.w+=v6mpy(Vuu32.ub,Vvv32.b,#u2):v", "", + fHIDE(size2s_t c00;) + fGET10BIT(c00, VvvV.v[0].uw[i], 0) + fHIDE(size2s_t c01;) + fGET10BIT(c01, VvvV.v[0].uw[i], 1) + fHIDE(size2s_t c02;) + fGET10BIT(c02, VvvV.v[0].uw[i], 2) + + fHIDE(size2s_t c10;) + fGET10BIT(c10, VvvV.v[1].uw[i], 0) + fHIDE(size2s_t c11;) + fGET10BIT(c11, VvvV.v[1].uw[i], 1) + fHIDE(size2s_t c12;) + fGET10BIT(c12, VvvV.v[1].uw[i], 2) + + if (uiV == 0) { + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c10); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c11); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c12); + + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c00); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c02); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c10); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c11); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c12); + + } else if (uiV == 1) { + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c00); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c01); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c02); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c10); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c11); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c12); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c00); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c02); + + } else if (uiV == 2) { + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c10); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c12); + + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c00); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c01); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c02); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c10); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c11); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c12); + + } else if (uiV == 3) { + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c00); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c01); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c02); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c10); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c12); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c00); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c01); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c02); + } +) +ITERATOR_INSN_MPY_SLOT_DOUBLE_VEC_VX_FWD(32, v6mpyhubs10_vxx, "Vxx32.w+=v6mpy(Vuu32.ub,Vvv32.b,#u2):h", "", + fHIDE(size2s_t c00;) + fGET10BIT(c00, VvvV.v[0].uw[i], 0) + fHIDE(size2s_t c01;) + fGET10BIT(c01, VvvV.v[0].uw[i], 1) + fHIDE(size2s_t c02;) + fGET10BIT(c02, VvvV.v[0].uw[i], 2) + fHIDE(size2s_t c10;) + fGET10BIT(c10, VvvV.v[1].uw[i], 0) + fHIDE(size2s_t c11;) + fGET10BIT(c11, VvvV.v[1].uw[i], 1) + fHIDE(size2s_t c12;) + fGET10BIT(c12, VvvV.v[1].uw[i], 2) + + if (uiV == 0) { + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c10); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c11); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c12); + + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c00); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c02); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c10); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c11); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c12); + + } else if (uiV == 1) { + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c00); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c01); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c02); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c10); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c11); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c12); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c00); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c02); + + } else if (uiV == 2) { + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c10); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c12); + + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c00); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c01); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c02); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c10); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c11); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c12); + + } else if (uiV == 3) { + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c00); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c01); + VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c02); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c10); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c12); + + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c00); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c01); + VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c02); + } +) + + +ITERATOR_INSN_MPY_SLOT_DOUBLE_VEC(32, v6mpyvubs10, "Vdd32.w=v6mpy(Vuu32.ub,Vvv32.b,#u2):v", "", + fHIDE(short c00;) + fGET10BIT(c00, VvvV.v[0].uw[i], 0) + fHIDE(short c01;) + fGET10BIT(c01, VvvV.v[0].uw[i], 1) + fHIDE(short c02;) + fGET10BIT(c02, VvvV.v[0].uw[i], 2) + fHIDE(short c10;) + fGET10BIT(c10, VvvV.v[1].uw[i], 0) + fHIDE(short c11;) + fGET10BIT(c11, VvvV.v[1].uw[i], 1) + fHIDE(short c12;) + fGET10BIT(c12, VvvV.v[1].uw[i], 2) + + + + if (uiV == 0) { + VddV.v[1].w[i] = fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c10); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c11); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c12); + + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c00); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c02); + + VddV.v[0].w[i] = fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c10); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c11); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c12); + + } else if (uiV == 1) { + VddV.v[1].w[i] = fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c00); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c01); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c02); + + VddV.v[0].w[i] = fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c10); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c11); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c12); + + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c00); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c02); + + } else if (uiV == 2) { + VddV.v[1].w[i] = fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c10); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c12); + + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c00); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c01); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c02); + + VddV.v[0].w[i] = fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c10); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c11); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c12); + + } else if (uiV == 3) { + VddV.v[1].w[i] = fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c00); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c01); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c02); + + VddV.v[0].w[i] = fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c10); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c12); + + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c00); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c01); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c02); + } +) + +ITERATOR_INSN_MPY_SLOT_DOUBLE_VEC(32, v6mpyhubs10, "Vdd32.w=v6mpy(Vuu32.ub,Vvv32.b,#u2):h", "", + fHIDE(short c00;) + fGET10BIT(c00, VvvV.v[0].uw[i], 0) + fHIDE(short c01;) + fGET10BIT(c01, VvvV.v[0].uw[i], 1) + fHIDE(short c02;) + fGET10BIT(c02, VvvV.v[0].uw[i], 2) + fHIDE(short c10;) + fGET10BIT(c10, VvvV.v[1].uw[i], 0) + fHIDE(short c11;) + fGET10BIT(c11, VvvV.v[1].uw[i], 1) + fHIDE(short c12;) + fGET10BIT(c12, VvvV.v[1].uw[i], 2) + + if (uiV == 0) { + VddV.v[1].w[i] = fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c10); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c11); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c12); + + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c00); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c02); + + VddV.v[0].w[i] = fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c10); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c11); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c12); + + } else if (uiV == 1) { + VddV.v[1].w[i] = fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c00); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c01); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c02); + + VddV.v[0].w[i] = fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c10); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c11); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c12); + + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c00); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c02); + + } else if (uiV == 2) { + VddV.v[1].w[i] = fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c10); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c12); + + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c00); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c01); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c02); + + VddV.v[0].w[i] = fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c10); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c11); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c12); + + } else if (uiV == 3) { + VddV.v[1].w[i] = fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c00); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c01); + VddV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c02); + + VddV.v[0].w[i] = fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c10); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c12); + + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c00); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c01); + VddV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c02); + } +) + EXTINSN(V6_vscattermhwq, "if (Qs4) vscatter(Rt32,Mu2,Vvv32.w).h=Vw32", ATTRIBS(A_EXTENSION,A_CVI,A_CVI_SCATTER,A_CVI_VA_DV,A_CVI_VM,A_MEMLIKE), "Scatter halfwords conditional", { |