MIMG?96 and other

spirv/emit_mimg.pas

@@ -42,7 +42,8 @@ type
   procedure emit_image_sample_gather(Tgrp:PsrNode;info:PsrImageInfo);
   procedure emit_image_load(Tgrp:PsrNode;info:PsrImageInfo);
   procedure emit_image_store(Tgrp:PsrNode;info:PsrImageInfo);
-  procedure emit_get_resinfo(Tgrp:PsrNode;info:PsrImageInfo);
+  procedure emit_image_get_resinfo(Tgrp:PsrNode;info:PsrImageInfo);
+  procedure emit_image_get_lod(Tgrp:PsrNode;info:PsrImageInfo);
  end;
 
 implementation
@@ -1006,7 +1007,7 @@ begin
 
 end;
 
-procedure TEmit_MIMG.emit_get_resinfo(Tgrp:PsrNode;info:PsrImageInfo);
+procedure TEmit_MIMG.emit_image_get_resinfo(Tgrp:PsrNode;info:PsrImageInfo);
 var
  offset:DWORD;
  dst,lod:PsrRegNode;
@@ -1037,6 +1038,51 @@ begin
  AddCapability(Capability.ImageQuery);
 end;
 
+procedure TEmit_MIMG.emit_image_get_lod(Tgrp:PsrNode;info:PsrImageInfo);
+var
+ src:array[0..3] of PsrRegSlot;
+
+ pLayout:PsrDataLayout;
+ Sgrp:PsrNode;
+
+ dst,cmb:PsrRegNode;
+
+ param:TImgSampleParam;
+
+ mask:Byte;
+begin
+ if not get_srsrc(FSPI.MIMG.SSAMP,4,@src) then Assert(false);
+
+ pLayout:=GroupingSharp(src,rtSSharp4);
+ Sgrp:=FetchSampler(pLayout);
+
+ cmb:=OpSampledImage(line,Tgrp,Sgrp,info^.dtype,info^.tinfo);
+
+ Assert(FSPI.MIMG.DMASK=1,'FSPI.MIMG.DMASK<>1');
+
+ //gather
+ param:=Default(TImgSampleParam);
+ param.coord:=GatherCoord_f(param.roffset,info);
+ //gather
+
+ dst:=NewReg(param.coord^.dtype);
+
+ _Op2(line,Op.OpImageQueryLod,dst,PsrRegNode(cmb),param.coord);
+
+ case param.coord^.dtype.Count of
+  1:mask:=1;
+  2:mask:=3;
+  3:mask:=7;
+  4:mask:=15;
+  else
+    mask:=0;
+ end;
+
+ DistribDmask(mask,dst,info);
+
+ AddCapability(Capability.ImageQuery);
+end;
+
 procedure TEmit_MIMG.emit_MIMG;
 var
  src:array[0..7] of PsrRegSlot;
@@ -1111,7 +1157,15 @@ begin
      info.tinfo.Sampled:=1;
      Tgrp:=FetchImage(pLayout,info.dtype,info.tinfo);
 
-     emit_GET_RESINFO(Tgrp,@info);
+     emit_image_get_resinfo(Tgrp,@info);
+    end;
+
+  IMAGE_GET_LOD:
+    begin
+     info.tinfo.Sampled:=1;
+     Tgrp:=FetchImage(pLayout,info.dtype,info.tinfo);
+
+     emit_image_get_lod(Tgrp,@info);
     end;
 
   else

spirv/emit_op.pas

@@ -74,7 +74,7 @@ type
   function  OpAbsDiff(pLine:PspirvOp;dst,src0,src1:PsrRegNode):PSpirvOp;
   procedure OpWQM32(dst:PsrRegSlot;src:PsrRegNode);
   //
-  procedure OpBFEU32(dst:PsrRegSlot;base,src0,src1:PsrRegNode);
+  procedure OpBFE_32(dst:PsrRegSlot;base,src0,src1:PsrRegNode);
   procedure OpBFIB32(dst:PsrRegSlot;bitmsk,src0,src1:PsrRegNode);
   //
   function  OpBFITo(src0,src1,src2,src3:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
@@ -638,9 +638,9 @@ begin
  Op1(srOpUtils.OpWQM32,dtUnknow,dst,src);
 end;
 
-procedure TEmitOp.OpBFEU32(dst:PsrRegSlot;base,src0,src1:PsrRegNode);
+procedure TEmitOp.OpBFE_32(dst:PsrRegSlot;base,src0,src1:PsrRegNode);
 begin
- Op3(srOpUtils.OpBFEU32,dtUint32,dst,base,src0,src1);
+ Op3(srOpUtils.OpBFE_32,base^.dtype,dst,base,src0,src1);
 end;
 
 procedure TEmitOp.OpBFIB32(dst:PsrRegSlot;bitmsk,src0,src1:PsrRegNode);

spirv/emit_post.pas

@@ -63,6 +63,7 @@ type
  end;
 
 function  EnumLineRegs(cb:TRegsCb;pLine:PSpirvOp):Integer;
+function  EnumFirstReg(cb:TRegsCb;pLine:PSpirvOp):Integer;
 function  EnumBlockOpForward(cb:TPostCb;pBlock:PsrOpBlock):Integer;
 function  EnumBlockOpBackward(cb:TPostCb;pBlock:PsrOpBlock):Integer;
 
@@ -101,6 +102,25 @@ begin
  end;
 end;
 
+function EnumFirstReg(cb:TRegsCb;pLine:PSpirvOp):Integer;
+var
+ node:POpParamNode;
+ pReg:PsrRegNode;
+begin
+ Result:=0;
+ if (cb=nil) or (pLine=nil) then Exit;
+ node:=pLine^.ParamFirst;
+ if (node<>nil) then
+ begin
+  if node^.Value^.IsType(ntReg) then
+  begin
+   pReg:=node^.AsReg;
+   Result:=Result+cb(pLine,pReg);
+   node^.Value:=pReg;
+  end;
+ end;
+end;
+
 function EnumBlockOpForward(cb:TPostCb;pBlock:PsrOpBlock):Integer;
 var
  node,prev:PSpirvOp;
@@ -429,7 +449,8 @@ begin
  if not node^.pDst^.IsType(ntReg) then Exit; //is reg
 
  Case node^.OpId of
-  Op.OpBitFieldUExtract  ,
+  Op.OpBitFieldSExtract  ,
+  Op.OpBitFieldUExtract  :Result:=EnumFirstReg(@RegSTStrict,node);
   Op.OpSelect            :Result:=EnumLineRegs(@RegSTStrict,node);
   Op.OpIAddCarry         ,
   Op.OpISubBorrow        ,

spirv/emit_post_op.pas

@@ -38,7 +38,7 @@ type
   function  OnWQM32__1(node:PSpirvOp):Integer;
   function  OnPackOfs1(node:PSpirvOp):Integer;
   function  _Fetch_PackAnc(node:PsrRegNode;index,count:Byte):PsrRegNode;
-  function  OnBFEU32_1(node:PSpirvOp):Integer;
+  function  OnBFE_32_1(node:PSpirvOp):Integer;
   function  OnBFIB32_1(node:PSpirvOp):Integer;
   function  OnMakeCub1(node:PSpirvOp):Integer;
   //
@@ -92,7 +92,7 @@ begin
   srOpUtils.OpAbsDiff   :Result:=OnAbsDiff1(node);
   srOpUtils.OpWQM32     :Result:=OnWQM32__1(node);
   srOpUtils.OpPackOfs   :Result:=OnPackOfs1(node);
-  srOpUtils.OpBFEU32    :Result:=OnBFEU32_1(node);
+  srOpUtils.OpBFE_32    :Result:=OnBFE_32_1(node);
   srOpUtils.OpBFIB32    :Result:=OnBFIB32_1(node);
   srOpUtils.OpMakeCub   :Result:=OnMakeCub1(node);
 
@@ -1889,7 +1889,7 @@ begin
 
 end;
 
-function TEmitPostOp.OnBFEU32_1(node:PSpirvOp):Integer;
+function TEmitPostOp.OnBFE_32_1(node:PSpirvOp):Integer;
 var
  dst:PsrRegNode;
  rBase,rIndex,rCount:PsrRegNode;
@@ -1897,6 +1897,7 @@ var
  num_31:PsrRegNode;
  data:array[0..1] of QWORD;
  index,count:Byte;
+ dtype:TsrDataType;
 begin
  Result:=0;
  dst:=node^.pDst^.AsType(ntReg);
@@ -1908,6 +1909,8 @@ begin
 
  if (rBase=nil) or (rIndex=nil) or (rCount=nil) then Exit;
 
+ dtype:=rBase^.dtype;
+
  //else
  node^.mark_not_used;
  node^.pDst:=nil;
@@ -1942,7 +1945,7 @@ begin
     num_31:=NewReg_q(dtUInt32,data[1],@Node);
     //
     rsl:=OpAndTo(rsl,num_31,@node);
-    rsl^.PrepType(ord(dtUInt32));
+    rsl^.PrepType(ord(dtype));
    end;
 
    dst^.pWriter:=rsl;
@@ -2002,7 +2005,13 @@ begin
   rCount^.PrepType(ord(dtUInt32));
  end;
 
- _Op3(node,Op.OpBitFieldUExtract,dst,rBase,rIndex,rCount);
+ case dtype of
+  dtUint32:_Op3(node,Op.OpBitFieldUExtract,dst,rBase,rIndex,rCount);
+  dtInt32 :_Op3(node,Op.OpBitFieldSExtract,dst,rBase,rIndex,rCount);
+  else
+   Assert(False);
+ end;
+
 end;
 
 function TEmitPostOp.OnBFIB32_1(node:PSpirvOp):Integer;

spirv/emit_vop1.pas

@@ -24,6 +24,9 @@ type
   procedure emit_V_CVT_FLR_I32_F32;
   procedure emit_V_CVT_RPI_I32_F32;
   procedure emit_V_CVT_F32_UBYTE0;
+  procedure emit_V_CVT_F32_UBYTE1;
+  procedure emit_V_CVT_F32_UBYTE2;
+  procedure emit_V_CVT_F32_UBYTE3;
   procedure emit_V_EXT_F32(OpId:DWORD);
   procedure emit_V_RSQ_CLAMP_F32;
   procedure emit_V_SIN_COS(OpId:DWORD);
@@ -152,6 +155,51 @@ begin
  Op1(Op.OpConvertUToF,dtFloat32,dst,src);
 end;
 
+procedure TEmit_VOP1.emit_V_CVT_F32_UBYTE1;
+Var
+ dst:PsrRegSlot;
+ src:PsrRegNode;
+begin
+ dst:=get_vdst8(FSPI.VOP1.VDST);
+ src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtUInt32);
+
+ src:=OpShrTo(src,8);
+ src:=OpAndTo(src,$FF);
+ src^.PrepType(ord(dtUInt32));
+
+ Op1(Op.OpConvertUToF,dtFloat32,dst,src);
+end;
+
+procedure TEmit_VOP1.emit_V_CVT_F32_UBYTE2;
+Var
+ dst:PsrRegSlot;
+ src:PsrRegNode;
+begin
+ dst:=get_vdst8(FSPI.VOP1.VDST);
+ src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtUInt32);
+
+ src:=OpShrTo(src,16);
+ src:=OpAndTo(src,$FF);
+ src^.PrepType(ord(dtUInt32));
+
+ Op1(Op.OpConvertUToF,dtFloat32,dst,src);
+end;
+
+procedure TEmit_VOP1.emit_V_CVT_F32_UBYTE3;
+Var
+ dst:PsrRegSlot;
+ src:PsrRegNode;
+begin
+ dst:=get_vdst8(FSPI.VOP1.VDST);
+ src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtUInt32);
+
+ src:=OpShrTo(src,24);
+ src:=OpAndTo(src,$FF);
+ src^.PrepType(ord(dtUInt32));
+
+ Op1(Op.OpConvertUToF,dtFloat32,dst,src);
+end;
+
 procedure TEmit_VOP1.emit_V_EXT_F32(OpId:DWORD);
 Var
  dst:PsrRegSlot;
@@ -254,7 +302,9 @@ begin
   V_CVT_RPI_I32_F32: emit_V_CVT_RPI_I32_F32;
 
   V_CVT_F32_UBYTE0: emit_V_CVT_F32_UBYTE0;
-
+  V_CVT_F32_UBYTE1: emit_V_CVT_F32_UBYTE1;
+  V_CVT_F32_UBYTE2: emit_V_CVT_F32_UBYTE2;
+  V_CVT_F32_UBYTE3: emit_V_CVT_F32_UBYTE3;
 
   V_FRACT_F32: emit_V_EXT_F32(GlslOp.Fract);
   V_TRUNC_F32: emit_V_EXT_F32(GlslOp.Trunc);

spirv/emit_vop3.pas

@@ -47,6 +47,7 @@ type
   procedure emit_V_MBCNT_HI_U32_B32;
 
   procedure emit_V_BFE_U32;
+  procedure emit_V_BFE_I32;
   procedure emit_V_BFI_B32;
   procedure emit_V_MAD_F32;
   procedure emit_V_MAD_LEGACY_F32;
@@ -555,7 +556,26 @@ begin
  src[1]:=fetch_ssrc9(FSPI.VOP3a.SRC1,dtUint32);
  src[2]:=fetch_ssrc9(FSPI.VOP3a.SRC2,dtUint32);
 
- OpBFEU32(dst,src[0],src[1],src[2]);
+ OpBFE_32(dst,src[0],src[1],src[2]);
+end;
+
+procedure TEmit_VOP3.emit_V_BFE_I32;
+Var
+ dst:PsrRegSlot;
+ src:array[0..2] of PsrRegNode;
+begin
+ dst:=get_vdst8(FSPI.VOP3a.VDST);
+
+ Assert(FSPI.VOP3a.OMOD =0,'FSPI.VOP3a.OMOD');
+ Assert(FSPI.VOP3a.ABS  =0,'FSPI.VOP3a.ABS');
+ Assert(FSPI.VOP3a.CLAMP=0,'FSPI.VOP3a.CLAMP');
+ Assert(FSPI.VOP3a.NEG  =0,'FSPI.VOP3a.NEG');
+
+ src[0]:=fetch_ssrc9(FSPI.VOP3a.SRC0,dtInt32);
+ src[1]:=fetch_ssrc9(FSPI.VOP3a.SRC1,dtUint32);
+ src[2]:=fetch_ssrc9(FSPI.VOP3a.SRC2,dtUint32);
+
+ OpBFE_32(dst,src[0],src[1],src[2]);
 end;
 
 procedure TEmit_VOP3.emit_V_BFI_B32;
@@ -1235,7 +1255,9 @@ begin
   V_MUL_HI_I32: emit_V_MUL_HI(dtInt32);
 
   V_BFE_U32: emit_V_BFE_U32;
+  V_BFE_I32: emit_V_BFE_I32;
   V_BFI_B32: emit_V_BFI_B32;
+
   V_MAD_F32: emit_V_MAD_F32;
   V_MAD_LEGACY_F32: emit_V_MAD_LEGACY_F32;
 

spirv/srOpUtils.pas

@@ -19,7 +19,7 @@ const
  OpAbsDiff=DWORD(-3);
  OpWQM32  =DWORD(-4);
 
- OpBFEU32 =DWORD(-5);
+ OpBFE_32 =DWORD(-5);
  OpBFIB32 =DWORD(-6);
 
  OpPackAnc=DWORD(-7);