From 5edf682c81cd58d50bb9fd071a18a634612c8bc0 Mon Sep 17 00:00:00 2001
From: Pavel <68122101+red-prig@users.noreply.github.com>
Date: Tue, 6 Dec 2022 21:12:24 +0300
Subject: [PATCH] VOP1?13 and etc

---
 spirv/emit_exp.pas     | 16 +++++++++++++---
 spirv/emit_op.pas      | 10 ++++++++++
 spirv/emit_post_op.pas | 37 +++++++++++++++++++++++++++++++++++++
 spirv/emit_vop1.pas    | 18 ++++++++++++++++++
 spirv/emit_vop2.pas    | 18 ++++++++++++++++++
 spirv/emit_vop3.pas    | 40 +++++++++++++++++++++++++++-------------
 spirv/srType.pas       | 20 ++++++++++++++++++++
 spirv/srTypes.pas      |  1 +
 8 files changed, 144 insertions(+), 16 deletions(-)

diff --git a/spirv/emit_exp.pas b/spirv/emit_exp.pas
index b8194374..3fa116d3 100644
--- a/spirv/emit_exp.pas
+++ b/spirv/emit_exp.pas
@@ -38,15 +38,20 @@ Var
  rtype:TsrDataType;
  f,i,p:DWORD;
 
+ push_count:DWORD;
 begin
  //if (VM<>0) and (EXEC<>0) = set pixel else (if DONE=1) discard pixel /(PS only)
 
+ push_count:=0;
+
  pOpBlock:=nil;
  if (FSPI.EXP.VM<>0) and (FSPI.EXP.DONE<>0) then
  begin
   pOpBlock:=AllocBlockOp;
   pOpBlock^.SetInfo(btOther,Cursor.Adr,Cursor.Adr);
+
   PushBlockOp(line,pOpBlock,nil);
+  Inc(push_count);
 
   exc:=MakeRead(get_exec0,dtBool);
   node:=AddSpirvOp(OpMakeExp);
@@ -57,16 +62,21 @@ begin
  if (TpsslExportType(FSPI.EXP.TGT)=etNull) //only set kill mask
    or (FSPI.EXP.EN=0) then //nop
  begin
-  if (pOpBlock<>nil) then //is pushed
+
+  While (push_count<>0) do
   begin
    Main^.PopBlock;
+   Dec(push_count);
   end;
+
   Exit;
  end;
 
  pOpBlock:=AllocBlockOp; //down
  pOpBlock^.SetInfo(btOther,Cursor.Adr,Cursor.Adr);
+
  PushBlockOp(line,pOpBlock,nil);
+ Inc(push_count);
 
  //output
 
@@ -172,10 +182,10 @@ begin
   OpStore(line,dout,dst);
  end;
 
- if (pOpBlock<>nil) then //is pushed
+ While (push_count<>0) do
  begin
   Main^.PopBlock;
-  Main^.PopBlock;
+  Dec(push_count);
  end;
 end;
 
diff --git a/spirv/emit_op.pas b/spirv/emit_op.pas
index 0f4f95cd..24b5671c 100644
--- a/spirv/emit_op.pas
+++ b/spirv/emit_op.pas
@@ -141,6 +141,8 @@ type
   function  OpAndTo(src0,src1:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
   function  OpAndTo(src0:PsrRegNode;src1:QWORD;ppLine:PPspirvOp=nil):PsrRegNode;
   //
+  function  OpBitCountTo(src:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
+  //
   function  OpImageSampleImplicitLod(pLine:PspirvOp;img:PsrNode;dst,coord:PsrRegNode):PSpirvOp;
   function  OpImageSampleExplicitLod(pLine:PspirvOp;img:PsrNode;dst,coord:PsrRegNode):PSpirvOp;
   function  OpImageSampleDrefImplicitLod(pLine:PspirvOp;img:PsrNode;dst,coord,pcf:PsrRegNode):PSpirvOp;
@@ -1259,6 +1261,14 @@ end;
 
 //
 
+function TEmitOp.OpBitCountTo(src:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
+begin
+ Result:=NewReg(src^.dtype);
+ _set_line(ppLine,_Op1(_get_line(ppLine),Op.OpBitCount,Result,src));
+end;
+
+//
+
 function TEmitOp.OpImageSampleImplicitLod(pLine:PspirvOp;img:PsrNode;dst,coord:PsrRegNode):PSpirvOp;
 Var
  node:PSpirvOp;
diff --git a/spirv/emit_post_op.pas b/spirv/emit_post_op.pas
index 7f77f09d..d18ef686 100644
--- a/spirv/emit_post_op.pas
+++ b/spirv/emit_post_op.pas
@@ -49,6 +49,8 @@ type
   function  OnNot1(node:PSpirvOp):Integer;
   function  OnBranchConditional1(node:PSpirvOp):Integer;
   //
+  function  OpBitCount1(node:PSpirvOp):Integer;
+  //
   function  OnSelect1(node:PSpirvOp):Integer;
   //
   procedure MakeVecConst(rtype:TsrDataType;dst:PsrRegNode;src:PPsrRegNode);
@@ -103,6 +105,8 @@ begin
 
   Op.OpBranchConditional:Result:=OnBranchConditional1(node);
 
+  Op.OpBitCount         :Result:=OpBitCount1(node);
+
   else;
  end;
 end;
@@ -779,6 +783,39 @@ begin
  Inc(Result);
 end;
 
+function TEmitPostOp.OpBitCount1(node:PSpirvOp):Integer;
+var
+ dst,src:PsrRegNode;
+ data:QWORD;
+
+ procedure _SetConst(dtype:TsrDataType;value:QWORD);
+ begin
+  dst^.pWriter:=ConstList.Fetch(dtype,value);
+  node^.mark_not_used;
+  node^.pDst:=nil;
+  Inc(Result);
+ end;
+
+begin
+ Result:=0;
+ dst:=node^.pDst^.AsType(ntReg);
+ src:=RegDown(node^.ParamNode(0)^.AsReg);
+
+ if (dst=nil) or (src=nil) then Exit;
+
+ if src^.is_const then
+ begin
+  //need a const calc
+  data:=src^.AsConst^.GetData;
+  data:=PopCnt(data); //BitCount
+
+  _SetConst(dst^.dtype,data);
+  Exit;
+ end;
+
+end;
+
+
 function try_get_comp_bridge(var src:PsrRegNode):Integer; forward;
 
 function TEmitPostOp.OpConvert1(node:PSpirvOp):Integer;
diff --git a/spirv/emit_vop1.pas b/spirv/emit_vop1.pas
index 0b730475..163c89d1 100644
--- a/spirv/emit_vop1.pas
+++ b/spirv/emit_vop1.pas
@@ -21,6 +21,7 @@ type
   procedure emit_V_CVT_F16_F32;
   procedure emit_V_CVT_F32_F16;
   procedure emit_V_CVT_OFF_F32_I4;
+  procedure emit_V_CVT_FLR_I32_F32;
   procedure emit_V_CVT_F32_UBYTE0;
   procedure emit_V_EXT_F32(OpId:DWORD);
   procedure emit_V_RSQ_CLAMP_F32;
@@ -109,6 +110,21 @@ begin
  Op2(Op.OpFDiv,dtFloat32,dst,src,num_16);
 end;
 
+procedure TEmit_VOP1.emit_V_CVT_FLR_I32_F32; //ConvertFloatToSignedInt(floor(vsrc.f))
+Var
+ dst:PsrRegSlot;
+ src:PsrRegNode;
+ flr:PsrRegNode;
+begin
+ dst:=get_vdst8(FSPI.VOP1.VDST);
+ src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtFloat32);
+
+ flr:=NewReg(dtFloat32);
+ _OpGlsl1(line,GlslOp.Floor,flr,src);
+
+ Op1(Op.OpConvertFToS,dtInt32,dst,flr);
+end;
+
 procedure TEmit_VOP1.emit_V_CVT_F32_UBYTE0;
 Var
  dst:PsrRegSlot;
@@ -220,6 +236,8 @@ begin
 
   V_CVT_OFF_F32_I4: emit_V_CVT_OFF_F32_I4;
 
+  V_CVT_FLR_I32_F32: emit_V_CVT_FLR_I32_F32;
+
   V_CVT_F32_UBYTE0: emit_V_CVT_F32_UBYTE0;
 
   V_FRACT_F32: emit_V_EXT_F32(GlslOp.Fract);
diff --git a/spirv/emit_vop2.pas b/spirv/emit_vop2.pas
index 00a0c0a5..493ada79 100644
--- a/spirv/emit_vop2.pas
+++ b/spirv/emit_vop2.pas
@@ -35,6 +35,7 @@ type
   procedure emit_V_MAC_LEGACY_F32;
   procedure emit_V_MADAK_F32;
   procedure emit_V_MADMK_F32;
+  procedure emit_V_BCNT_U32_B32;
   procedure emit_V_MMX(OpId:DWORD;rtype:TsrDataType);
  end;
 
@@ -374,6 +375,21 @@ begin
  OpFmaF32(dst,src[0],src[1],src[2]);
 end;
 
+procedure TEmit_VOP2.emit_V_BCNT_U32_B32; //vdst = bit_count(vsrc0) + vsrc1.u
+Var
+ dst:PsrRegSlot;
+ src:array[0..1] of PsrRegNode;
+begin
+ dst:=get_vdst8(FSPI.VOP2.VDST);
+
+ src[0]:=fetch_ssrc9(FSPI.VOP2.SRC0 ,dtUint32);
+ src[1]:=fetch_vsrc8(FSPI.VOP2.VSRC1,dtUint32);
+
+ src[0]:=OpBitCountTo(src[0]);
+
+ Op2(Op.OpIAdd,dtUint32,dst,src[0],src[1]);
+end;
+
 procedure TEmit_VOP2.emit_V_MMX(OpId:DWORD;rtype:TsrDataType);
 Var
  dst:PsrRegSlot;
@@ -427,6 +443,8 @@ begin
   V_MADAK_F32: emit_V_MADAK_F32;
   V_MADMK_F32: emit_V_MADMK_F32;
 
+  V_BCNT_U32_B32: emit_V_BCNT_U32_B32;
+
   V_MIN_LEGACY_F32: emit_V_MMX(GlslOp.NMin,dtFloat32);
   V_MAX_LEGACY_F32: emit_V_MMX(GlslOp.NMax,dtFloat32);
 
diff --git a/spirv/emit_vop3.pas b/spirv/emit_vop3.pas
index db491aad..9a54b31b 100644
--- a/spirv/emit_vop3.pas
+++ b/spirv/emit_vop3.pas
@@ -35,10 +35,10 @@ type
   procedure emit_V_SUBREV_F32;
   procedure emit_V_CVT_PKRTZ_F16_F32;
   procedure emit_V_MMX_F32(OpId:DWORD);
-  procedure emit_V_MUL_LO_I32;
+  procedure emit_V_MUL_LO(rtype:TsrDataType);
   procedure emit_V_MUL_I32_I24;
   procedure emit_V_MUL_U32_U24;
-  procedure emit_V_MUL_HI_U32;
+  procedure emit_V_MUL_HI(rtype:TsrDataType);
   procedure emit_V_MAC_F32;
 
   procedure emit_V_BFE_U32;
@@ -303,7 +303,7 @@ begin
  emit_dst_clamp_f(dst);
 end;
 
-procedure TEmit_VOP3.emit_V_MUL_LO_I32;
+procedure TEmit_VOP3.emit_V_MUL_LO(rtype:TsrDataType);
 Var
  dst:PsrRegSlot;
  src:array[0..1] of PsrRegNode;
@@ -315,8 +315,8 @@ begin
  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,dtInt32);
+ src[0]:=fetch_ssrc9(FSPI.VOP3a.SRC0,rtype);
+ src[1]:=fetch_ssrc9(FSPI.VOP3a.SRC1,rtype);
 
  OpIMul(dst,src[0],src[1]);
 end;
@@ -375,11 +375,12 @@ begin
  OpIMul(dst,src[0],src[1]);
 end;
 
-procedure TEmit_VOP3.emit_V_MUL_HI_U32;
+procedure TEmit_VOP3.emit_V_MUL_HI(rtype:TsrDataType);
 Var
  dst:PsrRegSlot;
  src:array[0..1] of PsrRegNode;
  tmp_r,dst_r:PsrRegNode;
+ tst:TsrDataType;
 begin
  dst:=get_vdst8(FSPI.VOP3a.VDST);
 
@@ -388,13 +389,23 @@ begin
  Assert(FSPI.VOP3a.CLAMP=0,'FSPI.VOP3a.CLAMP');
  Assert(FSPI.VOP3a.NEG  =0,'FSPI.VOP3a.NEG');
 
- src[0]:=fetch_ssrc9(FSPI.VOP3a.SRC0,dtUInt32);
- src[1]:=fetch_ssrc9(FSPI.VOP3a.SRC1,dtUInt32);
+ src[0]:=fetch_ssrc9(FSPI.VOP3a.SRC0,rtype);
+ src[1]:=fetch_ssrc9(FSPI.VOP3a.SRC1,rtype);
 
- tmp_r:=NewReg(dtStruct2u);
- _Op2(line,Op.OpUMulExtended,tmp_r,src[0],src[1]);
+ tst:=rtype.AsStruct2;
+ Assert(tst<>dtUnknow);
 
- dst_r:=dst^.New(line,dtUInt32);
+ tmp_r:=NewReg(tst);
+
+ if (rtype.Sign=0) then
+ begin
+  _Op2(line,Op.OpUMulExtended,tmp_r,src[0],src[1]);
+ end else
+ begin
+  _Op2(line,Op.OpSMulExtended,tmp_r,src[0],src[1]);
+ end;
+
+ dst_r:=dst^.New(line,rtype);
 
  OpExtract(line,dst_r,tmp_r,1);
 end;
@@ -1025,8 +1036,11 @@ begin
 
   //VOP3 only
 
-  V_MUL_LO_I32: emit_V_MUL_LO_I32;
-  V_MUL_HI_U32: emit_V_MUL_HI_U32;
+  V_MUL_LO_U32: emit_V_MUL_LO(dtUint32);
+  V_MUL_HI_U32: emit_V_MUL_HI(dtUint32);
+
+  V_MUL_LO_I32: emit_V_MUL_LO(dtInt32);
+  V_MUL_HI_I32: emit_V_MUL_HI(dtInt32);
 
   V_BFE_U32: emit_V_BFE_U32;
   V_BFI_B32: emit_V_BFI_B32;
diff --git a/spirv/srType.pas b/spirv/srType.pas
index fcb80576..e7161af7 100644
--- a/spirv/srType.pas
+++ b/spirv/srType.pas
@@ -36,6 +36,7 @@ type
   dtVec3b,
   dtVec4b,
 
+  dtStruct2i,
   dtStruct2u,
 
   dtVec2u8,
@@ -92,6 +93,7 @@ type
   function BitSize:Byte;
   function High:QWORD;
   function AsVector(_count:Byte):TsrDataType;
+  function AsStruct2:TsrDataType;
  end;
 
  Pvec2f=^Tvec2f;
@@ -163,6 +165,7 @@ begin
   dtVec3b      :Result:='bvec3';
   dtVec4b      :Result:='bvec4';
 
+  dtStruct2i   :Result:='rec2i';
   dtStruct2u   :Result:='rec2u';
 
   dtVec2u8     :Result:='u8vec2';
@@ -233,6 +236,7 @@ begin
   dtInt64,
   dtUint64,
 
+  dtStruct2i,
   dtStruct2u,
 
   dtVec2u8,
@@ -305,6 +309,7 @@ begin
   dtVec3b,
   dtVec4b,
 
+  dtStruct2i,
   dtStruct2u,
 
   dtVec2u8,
@@ -369,6 +374,7 @@ begin
   dtVec3u,
   dtVec4u:Result:=dtUint32;
 
+  dtStruct2i,
   dtVec2i,
   dtVec3i,
   dtVec4i:Result:=dtInt32;
@@ -396,6 +402,7 @@ begin
   dtVec2i,
   dtVec2h,
   dtVec2f,
+  dtStruct2i,
   dtStruct2u:Result:=2;
 
   dtVec3b,
@@ -425,6 +432,8 @@ begin
   dtInt32,
   dtInt64,
 
+  dtStruct2i,
+
   dtHalf16,
   dtFloat32,
   dtFloat64,
@@ -469,6 +478,7 @@ begin
   dtVec4i16,
   dtVec2f,
   dtVec4h,
+  dtStruct2i,
   dtStruct2u:Result:=64;
 
   dtVec3u,
@@ -559,6 +569,16 @@ begin
  end;
 end;
 
+function TsrDataTypeHelper.AsStruct2:TsrDataType;
+begin
+ Result:=dtUnknow;
+ Case Self of
+  dtInt32 :Result:=dtStruct2i;
+  dtUint32:Result:=dtStruct2u;
+  else;
+ end;
+end;
+
 function CompareType(rtype1,rtype2:TsrDataType):Boolean;
 begin
  Case rtype1 of
diff --git a/spirv/srTypes.pas b/spirv/srTypes.pas
index 0ff40819..b086fe45 100644
--- a/spirv/srTypes.pas
+++ b/spirv/srTypes.pas
@@ -568,6 +568,7 @@ begin
 
   //
 
+  dtStruct2i,
   dtStruct2u:
     begin
      Result:=_FetchStruct2(dtype);