Update shader recompiler

spirv/SprvEmit.pas

@@ -200,7 +200,8 @@ uses
  emit_EXP,
  emit_VINTRP,
  emit_SMRD,
- emit_MIMG;
+ emit_MIMG,
+ emit_DS;
 
 function TSprvEmit.Alloc(Size:ptruint):Pointer;
 begin
@@ -1721,7 +1722,7 @@ begin
             else
                      TEmit_VOP3(Self)._emit_VOP3a;
            end;
-  W_DS    :_emit_DS;
+  W_DS    :TEmit_DS(Self)._emit_DS;
   W_MUBUF :TEmit_MUBUF(Self)._emit_MUBUF;
   W_MTBUF :TEmit_MTBUF(Self)._emit_MTBUF;
   W_EXP   :TEmit_EXP(Self)._emit_EXP;

spirv/emit_ds.pas

@@ -0,0 +1,64 @@
+unit emit_DS;
+
+{$mode objfpc}{$H+}
+
+interface
+
+uses
+  sysutils,
+  ps4_pssl,
+  srTypes,
+  srReg,
+  spirv,
+  SprvEmit,
+  emit_op;
+
+type
+ TEmit_DS=object(TEmitOp)
+  procedure _emit_DS;
+  procedure _emit_DS_SWIZZLE_B32;
+ end;
+
+implementation
+
+procedure TEmit_DS._emit_DS_SWIZZLE_B32; //TODO
+Var
+ dst:PsrRegSlot;
+begin
+ dst:=FRegsStory.get_vdst8(FSPI.DS.VDST);
+ SetConst(dst,dtFloat32,0);
+end;
+
+procedure TEmit_DS._emit_DS;
+begin
+
+ Case FSPI.DS.OP of
+
+  DS_NOP:;
+
+  //DS_SWIZZLE_B32 v4 v3 v0 v0 OFFSET:0x80AA GDS:0
+
+  DS_SWIZZLE_B32: //TODO
+   begin;
+     _emit_DS_SWIZZLE_B32;
+   end;
+
+  else
+   Assert(false,'DS?');
+ end;
+
+end;
+
+{
+OFFSET0:Byte;
+OFFSET1:Byte;
+GDS:bit1;
+ADDR:Byte;  (vbindex)
+DATA0:Byte; (vsrc0)
+DATA1:Byte; (vsrc1)
+VDST:Byte;
+}
+
+end.
+
+

spirv/emit_op.pas

@@ -66,6 +66,7 @@ type
    procedure emit_OpFract(dst:PsrRegSlot;src:PsrRegNode);
    procedure emit_OpSqrt(dst:PsrRegSlot;src:PsrRegNode);
    procedure emit_OpInverseSqrt(dst:PsrRegSlot;src:PsrRegNode);
+   procedure emit_OpLog2(dst:PsrRegSlot;src:PsrRegNode);
    procedure emit_OpExp2(dst:PsrRegSlot;src:PsrRegNode);
    procedure emit_OpSin(dst:PsrRegSlot;src:PsrRegNode);
    procedure emit_OpFMin(dst:PsrRegSlot;src0,src1:PsrRegNode);
@@ -584,6 +585,11 @@ begin
  emit_OpExt1(GlslOp.InverseSqrt,dtFloat32,dst,src);
 end;
 
+procedure TEmitOp.emit_OpLog2(dst:PsrRegSlot;src:PsrRegNode);
+begin
+ emit_OpExt1(GlslOp.Log2,dtFloat32,dst,src);
+end;
+
 procedure TEmitOp.emit_OpExp2(dst:PsrRegSlot;src:PsrRegNode);
 begin
  emit_OpExt1(GlslOp.Exp2,dtFloat32,dst,src);

spirv/emit_sopp.pas

@@ -25,7 +25,9 @@ type
   procedure _emit_S_BRANCH;
   function  IsBegLoop(Adr:TSrcAdr):Boolean;
   function  IsEndLoop(Adr:TSrcAdr):Boolean;
+  function  IsUnknow(Adr:TSrcAdr):Boolean;
   procedure emit_cond_block(pSlot:PsrRegSlot;n:Boolean;adr:TSrcAdr);
+  procedure emit_block_unknow(adr:TSrcAdr);
   procedure UpBuildVol(last:PsrOpBlock);
   procedure emit_loop(adr:TSrcAdr);
  end;
@@ -209,6 +211,16 @@ begin
  end;
 end;
 
+function TEmit_SOPP.IsUnknow(Adr:TSrcAdr):Boolean;
+var
+ pLabel:PsrLabel;
+begin
+ pLabel:=FindLabel(Adr);
+ Assert(pLabel<>nil);
+
+ Result:=pLabel^.IsType(ltUnknow);
+end;
+
 procedure TEmit_SOPP._emit_S_BRANCH_COND(pSlot:PsrRegSlot;n:Boolean);
 var
  c_adr,b_adr:TSrcAdr;
@@ -245,6 +257,32 @@ begin
 
 end;
 
+procedure TEmit_SOPP.emit_block_unknow(adr:TSrcAdr);
+var
+ c_adr:TSrcAdr;
+ e_adr:TSrcAdr;
+ pOpChild:PsrOpBlock;
+ Info:TsrBlockInfo;
+begin
+ Info:=Default(TsrBlockInfo);
+
+ c_adr:=FCursor.Adr;                      //get current
+ SetPtr(adr.get_pc,btAdr);                //set new
+ e_adr:=FCursor.pCode^.FTop.pELabel^.Adr; //get end of code
+ SetPtr(c_adr.get_pc,btMain);             //ret current
+
+ Info.b_adr:=adr;
+ Info.e_adr:=e_adr;
+ Info.bType:=btAdr;
+
+ //down group
+ pOpChild:=AllocBlockOp;
+ pOpChild^.SetInfo(Info);
+ PushBlockOp(line,pOpChild,nil);
+
+ SetPtr(adr.get_pc,btAdr);
+end;
+
 procedure TEmit_SOPP._emit_S_BRANCH;
 var
  c_adr,b_adr:TSrcAdr;
@@ -254,7 +292,10 @@ begin
  b_adr:=c_adr;
  b_adr.Offdw:=get_branch_offset(FSPI);
 
-
+ if IsUnknow(b_adr) then
+ begin
+  emit_block_unknow(b_adr);
+ end else
  if (SmallInt(FSPI.SOPP.SIMM)<0) then //up
  begin  //continue?
   if not IsBegLoop(b_adr) then Assert(false,'Unknow');

spirv/emit_vop1.pas

@@ -17,18 +17,10 @@ type
  TEmit_VOP1=object(TEmitOp)
   procedure _emit_VOP1;
   procedure _emit_V_MOV_B32;
-  procedure _emit_V_CVT_F32_I32;
+  procedure _emit_V_CVT(OpId:DWORD;dst_type,src_type:TsrDataType);
-  procedure _emit_V_CVT_U32_F32;
-  procedure _emit_V_CVT_I32_F32;
-  procedure _emit_V_CVT_F32_U32;
   procedure _emit_V_CVT_OFF_F32_I4;
-  procedure _emit_V_SQRT_F32;
+  procedure _emit_V_EXT_F32(OpId:DWORD);
-  procedure _emit_V_RSQ_F32;
-  procedure _emit_V_FLOOR_F32;
-  procedure _emit_V_FRACT_F32;
-  procedure _emit_V_EXP_F32;
   procedure _emit_V_RCP_F32;
-  procedure _emit_V_SIN_F32;
  end;
 
 implementation
@@ -43,44 +35,14 @@ begin
  _MakeCopy(dst,src);
 end;
 
-procedure TEmit_VOP1._emit_V_CVT_F32_I32;
+procedure TEmit_VOP1._emit_V_CVT(OpId:DWORD;dst_type,src_type:TsrDataType);
 Var
  dst:PsrRegSlot;
  src:PsrRegNode;
 begin
  dst:=FRegsStory.get_vdst8(FSPI.VOP1.VDST);
- src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtInt32);
+ src:=fetch_ssrc9(FSPI.VOP1.SRC0,src_type);
- emit_Op1(Op.OpConvertSToF,dtFloat32,dst,src);
+ emit_Op1(OpId,dst_type,dst,src);
-end;
-
-procedure TEmit_VOP1._emit_V_CVT_I32_F32;
-Var
- dst:PsrRegSlot;
- src:PsrRegNode;
-begin
- dst:=FRegsStory.get_vdst8(FSPI.VOP1.VDST);
- src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtFloat32);
- emit_Op1(Op.OpConvertFToS,dtInt32,dst,src);
-end;
-
-procedure TEmit_VOP1._emit_V_CVT_U32_F32;
-Var
- dst:PsrRegSlot;
- src:PsrRegNode;
-begin
- dst:=FRegsStory.get_vdst8(FSPI.VOP1.VDST);
- src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtFloat32);
- emit_Op1(Op.OpConvertFToU,dtUInt32,dst,src);
-end;
-
-procedure TEmit_VOP1._emit_V_CVT_F32_U32;
-Var
- dst:PsrRegSlot;
- src:PsrRegNode;
-begin
- dst:=FRegsStory.get_vdst8(FSPI.VOP1.VDST);
- src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtUInt32);
- emit_Op1(Op.OpConvertUToF,dtFloat32,dst,src);
 end;
 
 //V_CVT_OFF_F32_I4
@@ -119,54 +81,14 @@ begin
  emit_OpFDiv(dst,subf,num_16);
 end;
 
-procedure TEmit_VOP1._emit_V_SQRT_F32;
+procedure TEmit_VOP1._emit_V_EXT_F32(OpId:DWORD);
 Var
  dst:PsrRegSlot;
  src:PsrRegNode;
 begin
  dst:=FRegsStory.get_vdst8(FSPI.VOP1.VDST);
  src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtFloat32);
- emit_OpSqrt(dst,src);
+ emit_OpExt1(OpId,dtFloat32,dst,src);
-end;
-
-procedure TEmit_VOP1._emit_V_RSQ_F32;
-Var
- dst:PsrRegSlot;
- src:PsrRegNode;
-begin
- dst:=FRegsStory.get_vdst8(FSPI.VOP1.VDST);
- src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtFloat32);
- emit_OpInverseSqrt(dst,src);
-end;
-
-procedure TEmit_VOP1._emit_V_FLOOR_F32;
-Var
- dst:PsrRegSlot;
- src:PsrRegNode;
-begin
- dst:=FRegsStory.get_vdst8(FSPI.VOP1.VDST);
- src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtFloat32);
- emit_OpFloor(dst,src);
-end;
-
-procedure TEmit_VOP1._emit_V_FRACT_F32;
-Var
- dst:PsrRegSlot;
- src:PsrRegNode;
-begin
- dst:=FRegsStory.get_vdst8(FSPI.VOP1.VDST);
- src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtFloat32);
- emit_OpFract(dst,src);
-end;
-
-procedure TEmit_VOP1._emit_V_EXP_F32;
-Var
- dst:PsrRegSlot;
- src:PsrRegNode;
-begin
- dst:=FRegsStory.get_vdst8(FSPI.VOP1.VDST);
- src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtFloat32);
- emit_OpExp2(dst,src);
 end;
 
 procedure TEmit_VOP1._emit_V_RCP_F32;
@@ -183,86 +105,48 @@ begin
  emit_OpFDiv(dst,one,src);
 end;
 
-procedure TEmit_VOP1._emit_V_SIN_F32;
-Var
- dst:PsrRegSlot;
- src:PsrRegNode;
-begin
- dst:=FRegsStory.get_vdst8(FSPI.VOP1.VDST);
- src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtFloat32);
- emit_OpSin(dst,src);
-end;
-
 procedure TEmit_VOP1._emit_VOP1;
 begin
 
  Case FSPI.VOP1.OP of
 
+  V_NOP:;
+
   V_MOV_B32:
     begin
      _emit_V_MOV_B32;
     end;
 
-  V_CVT_F32_I32:
+  V_CVT_F32_I32: _emit_V_CVT(Op.OpConvertSToF,dtFloat32,dtInt32);
-    begin
+  V_CVT_F32_U32: _emit_V_CVT(Op.OpConvertUToF,dtFloat32,dtUInt32);
-     _emit_V_CVT_F32_I32;
+  V_CVT_U32_F32: _emit_V_CVT(Op.OpConvertFToU,dtUInt32 ,dtFloat32);
-    end;
+  V_CVT_I32_F32: _emit_V_CVT(Op.OpConvertFToS,dtInt32  ,dtFloat32);
-
-  V_CVT_I32_F32:
-    begin
-     _emit_V_CVT_I32_F32;
-    end;
-
-  V_CVT_U32_F32:
-    begin
-     _emit_V_CVT_U32_F32;
-    end;
-
-  V_CVT_F32_U32:
-    begin
-     _emit_V_CVT_F32_U32;
-    end;
 
   V_CVT_OFF_F32_I4:
     begin
      _emit_V_CVT_OFF_F32_I4;
     end;
 
-  V_SQRT_F32:
+  V_FRACT_F32: _emit_V_EXT_F32(GlslOp.Fract);
-    begin
+  V_TRUNC_F32: _emit_V_EXT_F32(GlslOp.Trunc);
-     _emit_V_SQRT_F32;
+  V_CEIL_F32 : _emit_V_EXT_F32(GlslOp.Ceil);
-    end;
 
-  V_RSQ_F32:
+  V_FLOOR_F32: _emit_V_EXT_F32(GlslOp.Floor);
-    begin
+  V_EXP_F32  : _emit_V_EXT_F32(GlslOp.Exp2);
-     _emit_V_RSQ_F32;
+  V_LOG_F32  : _emit_V_EXT_F32(GlslOp.Log2);
-    end;
 
-  V_FLOOR_F32:
+  V_RSQ_F32  : _emit_V_EXT_F32(GlslOp.InverseSqrt);
-    begin
-     _emit_V_FLOOR_F32;
-    end;
 
-  V_FRACT_F32:
+  V_SQRT_F32 : _emit_V_EXT_F32(GlslOp.Sqrt);
-    begin
-     _emit_V_FRACT_F32;
-    end;
 
-  V_EXP_F32:
+  V_SIN_F32  : _emit_V_EXT_F32(GlslOp.Sin);
-    begin
+  V_COS_F32  : _emit_V_EXT_F32(GlslOp.Cos);
-     _emit_V_EXP_F32;
-    end;
 
   V_RCP_F32:
     begin
      _emit_V_RCP_F32;
     end;
 
-  V_SIN_F32:
-    begin
-     _emit_V_SIN_F32;
-    end;
-
   else
    Assert(false,'VOP1?'+IntToStr(FSPI.VOP1.OP));
  end;

spirv/emit_vop3.pas

@@ -50,7 +50,7 @@ type
   procedure _emit_V_FMA_F32;
   procedure _emit_V_CUBE(OpId:DWORD);
   procedure _emit_V_MOV_B32;
-  procedure _emit_V_SQRT_F32;
+  procedure _emit_V_EXT_F32(OpId:DWORD);
   procedure _emit_V_RCP_F32;
  end;
 
@@ -681,7 +681,7 @@ begin
 
 end;
 
-procedure TEmit_VOP3._emit_V_SQRT_F32;
+procedure TEmit_VOP3._emit_V_EXT_F32(OpId:DWORD);
 Var
  dst:PsrRegSlot;
  src:PsrRegNode;
@@ -693,7 +693,7 @@ begin
  _emit_src_abs_bit(@src,1);
  _emit_src_neg_bit(@src,1);
 
- emit_OpSqrt(dst,src);
+ emit_OpExt1(OpId,dtFloat32,dst,src);
 
  _emit_dst_omod_f(dst);
  _emit_dst_clamp_f(dst);
@@ -868,6 +868,8 @@ end;
 procedure TEmit_VOP3._emit_VOP3a;
 begin
 
+ //VOP2 analog
+
  Case FSPI.VOP3a.OP of
 
   256+V_CNDMASK_B32:
@@ -915,6 +917,8 @@ begin
      _emit_V_MAC_F32;
     end;
 
+  //VOP3 only
+
   V_MUL_LO_I32:
     begin
      _emit_V_MUL_LO_I32;
@@ -970,15 +974,29 @@ begin
   V_CUBETC_F32:_emit_V_CUBE(OpCUBETC);
   V_CUBEMA_F32:_emit_V_CUBE(OpCUBEMA);
 
+  //VOP1 analog
+
+  384+V_NOP:;
+
   384+V_MOV_B32:
    begin
     _emit_V_MOV_B32;
    end;
 
-  384+V_SQRT_F32:
+  384+V_FRACT_F32: _emit_V_EXT_F32(GlslOp.Fract);
-   begin
+  384+V_TRUNC_F32: _emit_V_EXT_F32(GlslOp.Trunc);
-    _emit_V_SQRT_F32;
+  384+V_CEIL_F32 : _emit_V_EXT_F32(GlslOp.Ceil);
-   end;
+
+  384+V_FLOOR_F32: _emit_V_EXT_F32(GlslOp.Floor);
+  384+V_EXP_F32  : _emit_V_EXT_F32(GlslOp.Exp2);
+  384+V_LOG_F32  : _emit_V_EXT_F32(GlslOp.Log2);
+
+  384+V_RSQ_F32  : _emit_V_EXT_F32(GlslOp.InverseSqrt);
+
+  384+V_SQRT_F32 : _emit_V_EXT_F32(GlslOp.Sqrt);
+
+  384+V_SIN_F32  : _emit_V_EXT_F32(GlslOp.Sin);
+  384+V_COS_F32  : _emit_V_EXT_F32(GlslOp.Cos);
 
   384+V_RCP_F32:
    begin