Emulator-Archive/fpPS4
1
0
Fork 0
mirror of https://github.com/red-prig/fpPS4.git synced 2024-11-26 16:10:25 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Update shader recompiler

Browse Source
This commit is contained in:
Pavel 2022-11-15 15:54:47 +03:00
parent 47d666c76c
commit 8e6c7c3d0b
18 changed files with 604 additions and 190 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

92
chip/ps4_pssl.pas
View File

@ -788,12 +788,15 @@ const
IMAGE_LOAD_PCK_SGN =$03;
IMAGE_LOAD_MIP_PCK =$04;
IMAGE_LOAD_MIP_PCK_SGN=$05;
IMAGE_STORE =$08;
IMAGE_STORE_MIP =$09;
IMAGE_STORE_PCK =$0A;
IMAGE_STORE_MIP_PCK =$0B;
IMAGE_GET_RESINFO =$0E;
IMAGE_ATOMIC_SWAP =$0F;
IMAGE_STORE_PCK =$0a;
IMAGE_STORE_MIP_PCK =$0b;
IMAGE_GET_RESINFO =$0e;
IMAGE_ATOMIC_SWAP =$0f;
IMAGE_ATOMIC_CMPSWAP =$10;
IMAGE_ATOMIC_ADD =$11;
IMAGE_ATOMIC_SUB =$12;
@ -803,12 +806,13 @@ const
IMAGE_ATOMIC_UMAX =$17;
IMAGE_ATOMIC_AND =$18;
IMAGE_ATOMIC_OR =$19;
IMAGE_ATOMIC_XOR =$1A;
IMAGE_ATOMIC_INC =$1B;
IMAGE_ATOMIC_DEC =$1C;
IMAGE_ATOMIC_FCMPSWAP =$1D;
IMAGE_ATOMIC_FMIN =$1E;
IMAGE_ATOMIC_FMAX =$1F;
IMAGE_ATOMIC_XOR =$1a;
IMAGE_ATOMIC_INC =$1b;
IMAGE_ATOMIC_DEC =$1c;
IMAGE_ATOMIC_FCMPSWAP =$1d;
IMAGE_ATOMIC_FMIN =$1e;
IMAGE_ATOMIC_FMAX =$1f;
IMAGE_SAMPLE =$20;
IMAGE_SAMPLE_CL =$21;
IMAGE_SAMPLE_D =$22;
@ -819,12 +823,12 @@ const
IMAGE_SAMPLE_LZ =$27;
IMAGE_SAMPLE_C =$28;
IMAGE_SAMPLE_C_CL =$29;
IMAGE_SAMPLE_C_D =$2A;
IMAGE_SAMPLE_C_D_CL =$2B;
IMAGE_SAMPLE_C_L =$2C;
IMAGE_SAMPLE_C_B =$2D;
IMAGE_SAMPLE_C_B_CL =$2E;
IMAGE_SAMPLE_C_LZ =$2F;
IMAGE_SAMPLE_C_D =$2a;
IMAGE_SAMPLE_C_D_CL =$2b;
IMAGE_SAMPLE_C_L =$2c;
IMAGE_SAMPLE_C_B =$2d;
IMAGE_SAMPLE_C_B_CL =$2e;
IMAGE_SAMPLE_C_LZ =$2f;
IMAGE_SAMPLE_O =$30;
IMAGE_SAMPLE_CL_O =$31;
IMAGE_SAMPLE_D_O =$32;
@ -835,24 +839,25 @@ const
IMAGE_SAMPLE_LZ_O =$37;
IMAGE_SAMPLE_C_O =$38;
IMAGE_SAMPLE_C_CL_O =$39;
IMAGE_SAMPLE_C_D_O =$3A;
IMAGE_SAMPLE_C_D_CL_O =$3B;
IMAGE_SAMPLE_C_L_O =$3C;
IMAGE_SAMPLE_C_B_O =$3D;
IMAGE_SAMPLE_C_B_CL_O =$3E;
IMAGE_SAMPLE_C_LZ_O =$3F;
IMAGE_SAMPLE_C_D_O =$3a;
IMAGE_SAMPLE_C_D_CL_O =$3b;
IMAGE_SAMPLE_C_L_O =$3c;
IMAGE_SAMPLE_C_B_O =$3d;
IMAGE_SAMPLE_C_B_CL_O =$3e;
IMAGE_SAMPLE_C_LZ_O =$3f;
IMAGE_GATHER4 =$40;
IMAGE_GATHER4_CL =$41;
IMAGE_GATHER4_L =$42;
IMAGE_GATHER4_B =$43;
IMAGE_GATHER4_B_CL =$44;
IMAGE_GATHER4_LZ =$45;
IMAGE_GATHER4_C =$46;
IMAGE_GATHER4_C_CL =$47;
IMAGE_GATHER4_C_L =$4C;
IMAGE_GATHER4_C_B =$4D;
IMAGE_GATHER4_C_B_CL =$4E;
IMAGE_GATHER4_C_LZ =$4F;
IMAGE_GATHER4_L =$44;
IMAGE_GATHER4_B =$45;
IMAGE_GATHER4_B_CL =$46;
IMAGE_GATHER4_LZ =$47;
IMAGE_GATHER4_C =$48;
IMAGE_GATHER4_C_CL =$49;
IMAGE_GATHER4_C_L =$4c;
IMAGE_GATHER4_C_B =$4d;
IMAGE_GATHER4_C_B_CL =$4e;
IMAGE_GATHER4_C_LZ =$4f;
IMAGE_GATHER4_O =$50;
IMAGE_GATHER4_CL_O =$51;
IMAGE_GATHER4_L_O =$54;
@ -861,19 +866,22 @@ const
IMAGE_GATHER4_LZ_O =$57;
IMAGE_GATHER4_C_O =$58;
IMAGE_GATHER4_C_CL_O =$59;
IMAGE_GATHER4_C_L_O =$5C;
IMAGE_GATHER4_C_B_O =$5D;
IMAGE_GATHER4_C_B_CL_O=$5E;
IMAGE_GATHER4_C_LZ_O =$5F;
IMAGE_GATHER4_C_L_O =$5c;
IMAGE_GATHER4_C_B_O =$5d;
IMAGE_GATHER4_C_B_CL_O=$5e;
IMAGE_GATHER4_C_LZ_O =$5f;
IMAGE_GET_LOD =$60;
IMAGE_SAMPLE_CD =$68;
IMAGE_SAMPLE_CD_CL =$69;
IMAGE_SAMPLE_C_CD =$6A;
IMAGE_SAMPLE_C_CD_CL =$6B;
IMAGE_SAMPLE_CD_O =$6C;
IMAGE_SAMPLE_CD_CL_O =$6D;
IMAGE_SAMPLE_C_CD_O =$6E;
IMAGE_SAMPLE_C_CD_CL_O=$6F;
IMAGE_SAMPLE_C_CD =$6a;
IMAGE_SAMPLE_C_CD_CL =$6b;
IMAGE_SAMPLE_CD_O =$6c;
IMAGE_SAMPLE_CD_CL_O =$6d;
IMAGE_SAMPLE_C_CD_O =$6e;
IMAGE_SAMPLE_C_CD_CL_O=$6f;
//DS
DS_ADD_U32 =$00;

63
spirv/SprvEmit.pas
View File

@ -35,9 +35,21 @@ type
Constructor Create;
procedure SetUserData(pData:Pointer);
Procedure InitVs(RSRC2:TSPI_SHADER_PGM_RSRC2_VS;instance:Byte);
Procedure InitPs(RSRC2:TSPI_SHADER_PGM_RSRC2_PS;ENA:TSPI_PS_INPUT_ENA);
Procedure InitCs(RSRC2:TCOMPUTE_PGM_RSRC2;NTX:TCOMPUTE_NUM_THREAD_X;NTY:TCOMPUTE_NUM_THREAD_Y;NTZ:TCOMPUTE_NUM_THREAD_Z);
procedure FillGPR(VGPRS,USER_SGPR,SGPRS:Byte);
Procedure InitVs(RSRC1:TSPI_SHADER_PGM_RSRC1_VS;
RSRC2:TSPI_SHADER_PGM_RSRC2_VS;
instance:Byte);
Procedure InitPs(RSRC1:TSPI_SHADER_PGM_RSRC1_PS;
RSRC2:TSPI_SHADER_PGM_RSRC2_PS;
ENA:TSPI_PS_INPUT_ENA);
Procedure InitCs(RSRC1:TCOMPUTE_PGM_RSRC1;
RSRC2:TCOMPUTE_PGM_RSRC2;
NTX:TCOMPUTE_NUM_THREAD_X;
NTY:TCOMPUTE_NUM_THREAD_Y;
NTZ:TCOMPUTE_NUM_THREAD_Z);
procedure emit_spi; override;
@ -127,7 +139,32 @@ begin
DataLayoutList.SetUserData(pData);
end;
Procedure TSprvEmit.InitVs(RSRC2:TSPI_SHADER_PGM_RSRC2_VS;instance:Byte);
procedure TSprvEmit.FillGPR(VGPRS,USER_SGPR,SGPRS:Byte);
var
p:Byte;
begin
if (VGPRS<>0) then
For p:=0 to VGPRS-1 do
begin
if (RegsStory.VGRP[p].current=nil) then
begin
SetConst_i(@RegsStory.VGRP[p],dtUnknow,0);
end;
end;
if (SGPRS<>0) then
For p:=USER_SGPR to (SGPRS+USER_SGPR)-1 do
begin
if (RegsStory.SGRP[p].current=nil) then
begin
SetConst_i(@RegsStory.SGRP[p],dtUnknow,0);
end;
end;
end;
Procedure TSprvEmit.InitVs(RSRC1:TSPI_SHADER_PGM_RSRC1_VS;
RSRC2:TSPI_SHADER_PGM_RSRC2_VS;
instance:Byte);
var
p:Byte;
begin
@ -232,11 +269,14 @@ begin
AddInput(@RegsStory.VGRP[p],dtUint32,itVInstance,0);
end;
AddCapability(Capability.Shader);
FillGPR(RSRC1.VGPRS,RSRC2.USER_SGPR,RSRC1.SGPRS);
AddCapability(Capability.Shader);
end;
Procedure TSprvEmit.InitPs(RSRC2:TSPI_SHADER_PGM_RSRC2_PS;ENA:TSPI_PS_INPUT_ENA);
Procedure TSprvEmit.InitPs(RSRC1:TSPI_SHADER_PGM_RSRC1_PS;
RSRC2:TSPI_SHADER_PGM_RSRC2_PS;
ENA:TSPI_PS_INPUT_ENA);
var
p:Byte;
begin
@ -385,11 +425,16 @@ begin
//Per-pixel fixed point position Y[31:16], X[15:0]
end;
AddCapability(Capability.Shader);
FillGPR(RSRC1.VGPRS,RSRC2.USER_SGPR,RSRC1.SGPRS);
AddCapability(Capability.Shader);
end;
Procedure TSprvEmit.InitCs(RSRC2:TCOMPUTE_PGM_RSRC2;NTX:TCOMPUTE_NUM_THREAD_X;NTY:TCOMPUTE_NUM_THREAD_Y;NTZ:TCOMPUTE_NUM_THREAD_Z);
Procedure TSprvEmit.InitCs(RSRC1:TCOMPUTE_PGM_RSRC1;
RSRC2:TCOMPUTE_PGM_RSRC2;
NTX:TCOMPUTE_NUM_THREAD_X;
NTY:TCOMPUTE_NUM_THREAD_Y;
NTZ:TCOMPUTE_NUM_THREAD_Z);
var
p:Byte;
begin
@ -472,6 +517,8 @@ begin
if (FLocalSize.y=0) then FLocalSize.y:=1;
if (FLocalSize.z=0) then FLocalSize.z:=1;
FillGPR(RSRC1.VGPRS,RSRC2.USER_SGPR,RSRC1.SGPRS);
AddCapability(Capability.Shader);
end;

299
spirv/emit_mimg.pas
View File

@ -18,15 +18,28 @@ uses
emit_fetch;
type
TsrImageMods=Set Of (imMinLod,imGrad,imLod,imBiasLod,imZeroLod,imDref,imOffset);
TImgSampleParam=object
roffset:DWORD;
mods:TsrImageMods;
img_op:Integer;
coord,pcf,bias,lod,min_lod,offset:PsrRegNode;
end;
TEmit_MIMG=class(TEmitFetch)
procedure emit_MIMG;
procedure DistribDmask(dst:PsrRegNode;info:PsrImageInfo);
procedure DistribDmask(DMASK:Byte;dst:PsrRegNode;info:PsrImageInfo);
function GatherDmask(telem:TsrDataType):PsrRegNode;
Function GatherCoord_f(var offset:DWORD;dim_id:Byte):PsrRegNode;
Function GatherCoord_u(var offset:DWORD;dim_id:Byte):PsrRegNode;
Function Gather_value(var offset:DWORD;rtype:TsrDataType):PsrRegNode;
Function Gather_packed_offset(var offset:DWORD;dim:Byte):PsrRegNode;
procedure Gather_sample_param(var p:TImgSampleParam;info:PsrImageInfo);
procedure add_sample_op(var p:TImgSampleParam;node:PSpirvOp);
procedure emit_image_sample(Tgrp:PsrNode;info:PsrImageInfo);
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);
end;
@ -374,9 +387,6 @@ begin
Result.tinfo.Format :=GetImageFormat(PT);
end;
type
TsrImageMods=Set Of (imMinLod,imGrad,imLod,imBiasLod,imZeroLod,imDref,imOffset);
function GetImageMods(OP:Byte):TsrImageMods;
begin
Case OP of
@ -449,14 +459,40 @@ begin
end;
end;
procedure TEmit_MIMG.DistribDmask(dst:PsrRegNode;info:PsrImageInfo); //result
function GetImgOpStr(i:Integer):RawByteString;
procedure t(o:Integer;const v:RawByteString); inline;
begin
if ((o and i)<>0) then
begin
if (Result<>'') then
begin
Result:=Result+'|'+v;
end else
begin
Result:=v;
end;
end;
end;
begin
Result:='';
t(ImageOperands.Bias ,'Bias');
t(ImageOperands.Lod ,'Lod');
t(ImageOperands.Grad ,'Grad');
t(ImageOperands.ConstOffset ,'ConstOffset');
t(ImageOperands.Sample ,'Sample');
t(ImageOperands.MinLod ,'MinLod');
end;
procedure TEmit_MIMG.DistribDmask(DMASK:Byte;dst:PsrRegNode;info:PsrImageInfo); //result
var
pSlot:PsrRegSlot;
i,d:Byte;
begin
d:=0;
For i:=0 to 3 do
if Byte(FSPI.MIMG.DMASK).TestBit(i) then
if DMASK.TestBit(i) then
begin
pSlot:=get_vdst8(FSPI.MIMG.VDATA+d);
Inc(d);
@ -630,6 +666,93 @@ begin
end;
procedure TEmit_MIMG.Gather_sample_param(var p:TImgSampleParam;info:PsrImageInfo);
begin
p.roffset:=0;
p.mods:=GetImageMods(FSPI.MIMG.OP);
p.img_op:=0;
//gather
if (imOffset in p.mods) then
begin
p.offset:=Gather_packed_offset(p.roffset,GetDimCount(info^.tinfo.Dim));
p.img_op:=p.img_op or ImageOperands.ConstOffset;
end;
if (imBiasLod in p.mods) then
begin
p.bias:=Gather_value(p.roffset,dtFloat32);
p.img_op:=p.img_op or ImageOperands.Bias;
end;
if (imDref in p.mods) then
begin
p.pcf:=Gather_value(p.roffset,dtFloat32);
end;
if (imGrad in p.mods) then
begin
Assert(false,'TODO imGrad');
end;
p.coord:=GatherCoord_f(p.roffset,info^.tinfo.Dim);
if (imLod in p.mods) then
begin
p.lod:=Gather_value(p.roffset,dtFloat32);
p.img_op:=p.img_op or ImageOperands.Lod;
end else
if (imZeroLod in p.mods) then
begin
p.lod:=NewReg_s(dtFloat32,0);
p.img_op:=p.img_op or ImageOperands.Lod;
end;
if (imMinLod in p.mods) then
begin
p.min_lod:=Gather_value(p.roffset,dtFloat32);
p.img_op:=p.img_op or ImageOperands.MinLod;
end;
//gather
end;
procedure TEmit_MIMG.add_sample_op(var p:TImgSampleParam;node:PSpirvOp);
begin
//ImageOperands
if (p.img_op<>0) then
begin
node^.AddLiteral(p.img_op,GetImgOpStr(p.img_op));
if ((p.img_op and ImageOperands.Bias)<>0) then
begin
node^.AddParam(p.bias);
end;
if ((p.img_op and ImageOperands.Lod)<>0) then
begin
node^.AddParam(p.lod);
end;
if ((p.img_op and ImageOperands.Grad)<>0) then
begin
Assert(false,'TODO imGrad');
end;
if ((p.img_op and ImageOperands.ConstOffset)<>0) then
begin
node^.AddParam(p.offset);
end;
if ((p.img_op and ImageOperands.MinLod)<>0) then
begin
node^.AddParam(p.min_lod);
AddCapability(Capability.MinLod);
end;
end;
//ImageOperands
end;
procedure TEmit_MIMG.emit_image_sample(Tgrp:PsrNode;info:PsrImageInfo);
var
src:array[0..3] of PsrRegSlot;
@ -637,12 +760,11 @@ var
pLayout:PsrDataLayout;
Sgrp:PsrNode;
dst,cmb,coord,pcf,lod,offset:PsrRegNode;
dst,cmb:PsrRegNode;
roffset:DWORD;
param:TImgSampleParam;
node:PSpirvOp;
begin
if not get_srsrc(FSPI.MIMG.SSAMP,4,@src) then Assert(false);
@ -659,74 +781,101 @@ begin
dst:=NewReg(info^.dtype.AsVector(4));
end;
roffset:=0;
//gather
param:=Default(TImgSampleParam);
Gather_sample_param(param,info);
//gather
Case FSPI.MIMG.OP of
IMAGE_SAMPLE:
begin
coord:=GatherCoord_f(roffset,info^.tinfo.Dim);
//OpImage
if (FExecutionModel=ExecutionModel.Fragment) then
begin
node:=OpImageSampleImplicitLod(line,cmb,dst,coord);
end else
begin
node:=OpImageSampleExplicitLod(line,cmb,dst,coord);
end;
if (imDref in param.mods) then
begin
if (FExecutionModel=ExecutionModel.Fragment) and
((param.img_op and ImageOperands.Lod)=0) then
begin
node:=OpImageSampleDrefImplicitLod(line,cmb,dst,param.coord,param.pcf);
end else
begin
node:=OpImageSampleDrefExplicitLod(line,cmb,dst,param.coord,param.pcf);
end;
end else
begin
if (FExecutionModel=ExecutionModel.Fragment) and
((param.img_op and ImageOperands.Lod)=0) then
begin
node:=OpImageSampleImplicitLod(line,cmb,dst,param.coord);
end else
begin
node:=OpImageSampleExplicitLod(line,cmb,dst,param.coord);
end;
end;
//OpImage
end;
//ImageOperands
add_sample_op(param,node);
//ImageOperands
IMAGE_SAMPLE_LZ:
begin
coord:=GatherCoord_f(roffset,info^.tinfo.Dim);
DistribDmask(FSPI.MIMG.DMASK,dst,info);
end;
node:=OpImageSampleExplicitLod(line,cmb,dst,coord);
procedure TEmit_MIMG.emit_image_sample_gather(Tgrp:PsrNode;info:PsrImageInfo);
var
src:array[0..3] of PsrRegSlot;
node^.AddLiteral(ImageOperands.Lod,'Lod');
pLayout:PsrDataLayout;
Sgrp:PsrNode;
//0
lod:=NewReg_s(dtFloat32,0);
node^.AddParam(lod);
end;
dst,cmb:PsrRegNode;
IMAGE_SAMPLE_LZ_O:
begin
offset:=Gather_packed_offset(roffset,GetDimCount(info^.tinfo.Dim));
param:TImgSampleParam;
id:Byte;
coord:=GatherCoord_f(roffset,info^.tinfo.Dim);
node:PSpirvOp;
begin
if not get_srsrc(FSPI.MIMG.SSAMP,4,@src) then Assert(false);
node:=OpImageSampleExplicitLod(line,cmb,dst,coord);
pLayout:=GroupingSharp(src,rtSSharp4);
Sgrp:=FetchSampler(pLayout);
node^.AddLiteral(ImageOperands.Lod or ImageOperands.ConstOffset,'Lod|ConstOffset');
cmb:=OpSampledImage(line,Tgrp,Sgrp,info^.dtype,info^.tinfo);
//0
lod:=NewReg_s(dtFloat32,0);
node^.AddParam(lod);
//1
node^.AddParam(offset);
end;
IMAGE_SAMPLE_C_LZ:
begin
pcf:=Gather_value(roffset,dtFloat32);
coord:=GatherCoord_f(roffset,info^.tinfo.Dim);
node:=OpImageSampleDrefExplicitLod(line,cmb,dst,coord,pcf);
node^.AddLiteral(ImageOperands.Lod,'Lod');
//0
lod:=NewReg_s(dtFloat32,0);
node^.AddParam(lod);
end;
else
Assert(false,'MIMG?'+IntToStr(FSPI.MIMG.OP));
if (info^.tinfo.Depth=1) then
begin
dst:=NewReg(info^.dtype);
end else
begin
dst:=NewReg(info^.dtype.AsVector(4));
end;
DistribDmask(dst,info);
//gather
param:=Default(TImgSampleParam);
Gather_sample_param(param,info);
//gather
id:=BsrByte(FSPI.MIMG.DMASK);
//OpImage
if (imDref in param.mods) then
begin
Assert(id=0,'gather pcf with non red component');
node:=OpImageDrefGather(line,cmb,dst,param.coord,param.pcf);
end else
begin
node:=OpImageGather(line,cmb,dst,param.coord,id);
end;
//OpImage
//OpExtension "SPV_AMD_texture_gather_bias_lod"
//Capability.ImageGatherBiasLodAMD
//VK_AMD_texture_gather_bias_lod
begin
param.img_op:=param.img_op and (not (ImageOperands.Bias or ImageOperands.Lod))
end;
//ImageOperands
add_sample_op(param,node);
//ImageOperands
DistribDmask(15,dst,info);
end;
procedure TEmit_MIMG.emit_image_load(Tgrp:PsrNode;info:PsrImageInfo);
@ -750,8 +899,7 @@ begin
if (info^.tinfo.MS<>0) then //fragid T# 2D MSAA
begin
smp:=fetch_vsrc8(FSPI.MIMG.VADDR+roffset,dtUint32);
Inc(roffset);
smp:=Gather_value(roffset,dtUint32);
node^.AddLiteral(ImageOperands.Sample,'Sample');
node^.AddParam(smp);
@ -762,8 +910,7 @@ begin
coord:=GatherCoord_u(roffset,info^.tinfo.Dim);
node:=OpImageFetch(line,Tgrp,dst,coord);
lod:=fetch_vsrc8(FSPI.MIMG.VADDR+roffset,dtUint32);
Inc(roffset);
lod:=Gather_value(roffset,dtUint32);
node^.AddLiteral(ImageOperands.Lod,'Lod');
node^.AddParam(lod);
@ -772,7 +919,7 @@ begin
Assert(false,'MIMG?'+IntToStr(FSPI.MIMG.OP));
end;
DistribDmask(dst,info);
DistribDmask(FSPI.MIMG.DMASK,dst,info);
end;
procedure TEmit_MIMG.emit_image_store(Tgrp:PsrNode;info:PsrImageInfo);
@ -795,8 +942,7 @@ begin
if (info^.tinfo.MS<>0) then //fragid T# 2D MSAA
begin
smp:=fetch_vsrc8(FSPI.MIMG.VADDR+roffset,dtUint32);
Inc(roffset);
smp:=Gather_value(roffset,dtUint32);
node^.AddLiteral(ImageOperands.Sample,'Sample');
node^.AddParam(smp);
@ -807,8 +953,7 @@ begin
coord:=GatherCoord_u(roffset,info^.tinfo.Dim);
node:=OpImageWrite(line,Tgrp,coord,dst);
lod:=fetch_vsrc8(FSPI.MIMG.VADDR+roffset,dtUint32);
Inc(roffset);
lod:=Gather_value(roffset,dtUint32);
node^.AddLiteral(ImageOperands.Lod,'Lod');
node^.AddParam(lod);
@ -862,6 +1007,16 @@ begin
emit_image_sample(Tgrp,@info);
end;
IMAGE_GATHER4..IMAGE_GATHER4_C_LZ_O: //sampled gather
begin
Assert(FSPI.MIMG.UNRM=0,'FSPI.MIMG.UNRM');
info.tinfo.Sampled:=1;
Tgrp:=FetchImage(pLayout,info.dtype,info.tinfo);
emit_image_sample_gather(Tgrp,@info);
end;
IMAGE_LOAD..IMAGE_LOAD_MIP_PCK_SGN: //loaded
begin
info.tinfo.Sampled:=1;

59
spirv/emit_op.pas
View File

@ -136,14 +136,16 @@ type
procedure OpLogicalAnd(dst:PsrRegSlot;src0,src1:PsrRegNode);
//
function OpNotTo(src:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
function OpBitwiseOrTo(src0,src1:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
function OpBitwiseAndTo(src0,src1:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
function OpBitwiseAndTo(src0:PsrRegNode;src1:QWORD;ppLine:PPspirvOp=nil):PsrRegNode;
function OpOrTo(src0,src1:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
function OpAndTo(src0,src1:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
function OpAndTo(src0:PsrRegNode;src1:QWORD;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;
function OpImageSampleDrefExplicitLod(pLine:PspirvOp;img:PsrNode;dst,coord,pcf:PsrRegNode):PSpirvOp;
function OpImageGather(pLine:PspirvOp;img:PsrNode;dst,coord:PsrRegNode;id:Byte):PSpirvOp;
function OpImageDrefGather(pLine:PspirvOp;img:PsrNode;dst,coord,pcf:PsrRegNode):PSpirvOp;
function OpImageFetch(pLine:PspirvOp;Tgrp:PsrNode;dst,coord:PsrRegNode):PSpirvOp;
function OpImageRead(pLine:PspirvOp;Tgrp:PsrNode;dst,idx:PsrRegNode):PspirvOp;
function OpImageWrite(pLine:PspirvOp;Tgrp:PsrNode;idx,src:PsrRegNode):PspirvOp;
@ -1180,22 +1182,22 @@ begin
_set_line(ppLine,_Op1(_get_line(ppLine),Op.OpNot,Result,src)); //post type
end;
function TEmitOp.OpBitwiseOrTo(src0,src1:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
function TEmitOp.OpOrTo(src0,src1:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
begin
Result:=NewReg(dtUnknow);
_set_line(ppLine,_Op2(_get_line(ppLine),Op.OpBitwiseOr,Result,src0,src1)); //post type
end;
function TEmitOp.OpBitwiseAndTo(src0,src1:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
function TEmitOp.OpAndTo(src0,src1:PsrRegNode;ppLine:PPspirvOp=nil):PsrRegNode;
begin
Result:=NewReg(dtUnknow);
_set_line(ppLine,_Op2(_get_line(ppLine),Op.OpBitwiseAnd,Result,src0,src1)); //post type
end;
function TEmitOp.OpBitwiseAndTo(src0:PsrRegNode;src1:QWORD;ppLine:PPspirvOp=nil):PsrRegNode;
function TEmitOp.OpAndTo(src0:PsrRegNode;src1:QWORD;ppLine:PPspirvOp=nil):PsrRegNode;
begin
if (src0=nil) then Exit(src0);
Result:=OpBitwiseAndTo(src0,NewReg_q(src0^.dtype,src1,ppLine),ppLine);
Result:=OpAndTo(src0,NewReg_q(src0^.dtype,src1,ppLine),ppLine);
end;
//
@ -1265,6 +1267,49 @@ begin
Result:=node;
end;
function TEmitOp.OpImageGather(pLine:PspirvOp;img:PsrNode;dst,coord:PsrRegNode;id:Byte):PSpirvOp;
Var
node:PSpirvOp;
comp:PsrRegNode;
begin
Case id of
0..3:;
else
Assert(False);
end;
comp:=NewReg_i(dtUint32,id);
node:=AddSpirvOp(pLine,Op.OpImageGather); //need first
node^.pType:=TypeList.Fetch(dst^.dtype);
node^.pDst:=dst;
node^.AddParam(img); //Sampled Image
node^.AddParam(coord); //Coordinate
node^.AddParam(comp); //Component
//Image Operands
Result:=node;
end;
function TEmitOp.OpImageDrefGather(pLine:PspirvOp;img:PsrNode;dst,coord,pcf:PsrRegNode):PSpirvOp;
Var
node:PSpirvOp;
begin
node:=AddSpirvOp(pLine,Op.OpImageDrefGather); //need first
node^.pType:=TypeList.Fetch(dst^.dtype);
node^.pDst:=dst;
node^.AddParam(img); //Sampled Image
node^.AddParam(coord); //Coordinate
node^.AddParam(pcf); //Dref
//Image Operands
Result:=node;
end;
function TEmitOp.OpImageFetch(pLine:PspirvOp;Tgrp:PsrNode;dst,coord:PsrRegNode):PSpirvOp;
Var
node:PSpirvOp;

2
spirv/emit_post.pas
View File

@ -852,6 +852,7 @@ begin
end;
end;
node^.sType:=sType;
node^.pType:=pType;
if node^.IsTop then
@ -862,6 +863,7 @@ begin
end else
begin
node^.sType:=nil;
node^.pType:=TypeList.Fetch(node^.dtype);
end;

10
spirv/emit_post_op.pas
View File

@ -1745,7 +1745,7 @@ begin
data[1]:=(1 shl count)-1;
num_31:=NewReg_q(dtUInt32,data[1],@Node);
//
rsl:=OpBitwiseAndTo(rsl,num_31,@node);
rsl:=OpAndTo(rsl,num_31,@node);
rsl^.PrepType(ord(dtUInt32));
end;
@ -1777,7 +1777,7 @@ begin
num_31:=NewReg_q(dtUInt32,31,@Node);
//
rIndex:=node^.ParamNode(1)^.AsReg; //orig
rIndex:=OpBitwiseAndTo(rIndex,num_31,@node);
rIndex:=OpAndTo(rIndex,num_31,@node);
rIndex^.PrepType(ord(dtUInt32));
end;
@ -1802,7 +1802,7 @@ begin
end;
//
rCount:=node^.ParamNode(2)^.AsReg; //orig
rCount:=OpBitwiseAndTo(rCount,num_31,@node);
rCount:=OpAndTo(rCount,num_31,@node);
rCount^.PrepType(ord(dtUInt32));
end;
@ -1868,13 +1868,13 @@ begin
src[0]:=node^.ParamNode(1)^.AsReg;
src[1]:=node^.ParamNode(2)^.AsReg;
src[0]:=OpBitwiseAndTo(src[0],bitmsk,@node);
src[0]:=OpAndTo(src[0],bitmsk,@node);
src[0]^.PrepType(ord(dtUInt32));
bitmsk:=OpNotTo(bitmsk,@node);
bitmsk^.PrepType(ord(dtUInt32));
src[1]:=OpBitwiseAndTo(src[1],bitmsk,@node);
src[1]:=OpAndTo(src[1],bitmsk,@node);
src[1]^.PrepType(ord(dtUInt32));
_Op2(node,Op.OpBitwiseOr,dst,src[0],src[1]);

69
spirv/emit_sop2.pas
View File

@ -27,7 +27,9 @@ type
procedure emit_S_AND_B64;
procedure emit_S_ANDN2_B64;
procedure emit_S_OR_B64;
procedure emit_S_XOR_B64;
procedure emit_S_ORN2_B64;
procedure emit_S_NAND_B64;
procedure emit_S_NOR_B64;
procedure emit_S_CSELECT_B32;
procedure emit_S_CSELECT_B64;
@ -117,7 +119,7 @@ begin
src[0]:=fetch_ssrc9(FSPI.SOP2.SSRC0,dtUInt32);
src[1]:=fetch_ssrc9(FSPI.SOP2.SSRC1,dtUInt32);
src[1]:=OpBitwiseAndTo(src[1],31);
src[1]:=OpAndTo(src[1],31);
src[1]^.PrepType(ord(dtUInt32));
Op2(Op.OpShiftLeftLogical,src[0]^.dtype,dst,src[0],src[1]);
@ -135,7 +137,7 @@ begin
src[0]:=fetch_ssrc9(FSPI.SOP2.SSRC0,dtUInt32);
src[1]:=fetch_ssrc9(FSPI.SOP2.SSRC1,dtUInt32);
src[1]:=OpBitwiseAndTo(src[1],31);
src[1]:=OpAndTo(src[1],31);
src[1]^.PrepType(ord(dtUInt32));
Op2(Op.OpShiftRightLogical,src[0]^.dtype,dst,src[0],src[1]);
@ -143,7 +145,7 @@ begin
OpISccNotZero(dst^.current); //SCC = (sdst.u != 0)
end;
procedure TEmit_SOP2.emit_S_AND_B32;
procedure TEmit_SOP2.emit_S_AND_B32; //sdst = (ssrc0 & ssrc1); SCC = (sdst != 0)
Var
dst:PsrRegSlot;
src:array[0..1] of PsrRegNode;
@ -158,7 +160,7 @@ begin
OpISccNotZero(dst^.current); //SCC = (sdst.u != 0)
end;
procedure TEmit_SOP2.emit_S_AND_B64; //SCC = (sdst[2] != 0)
procedure TEmit_SOP2.emit_S_AND_B64; //sdst[2] = (ssrc0[2] & ssrc1[2]); SCC = (sdst[2] != 0)
Var
dst:array[0..1] of PsrRegSlot;
src0,src1,src2:array[0..1] of PsrRegNode;
@ -177,7 +179,7 @@ begin
OpLogicalOr(get_scc,src2[0],src2[1]); //implict cast (int != 0)
end;
procedure TEmit_SOP2.emit_S_ANDN2_B64; //SCC = (sdst[2] != 0)
procedure TEmit_SOP2.emit_S_ANDN2_B64; //sdst[2] = (ssrc0[2] & ~ssrc1[2]); SCC = (sdst[2] != 0)
Var
dst:array[0..1] of PsrRegSlot;
src0,src1,src2:array[0..1] of PsrRegNode;
@ -199,7 +201,7 @@ begin
OpLogicalOr(get_scc,src2[0],src2[1]); //implict cast (int != 0)
end;
procedure TEmit_SOP2.emit_S_OR_B64; //SCC = (sdst[2] != 0)
procedure TEmit_SOP2.emit_S_OR_B64; //sdst[2] = (ssrc0[2] | ssrc1[2]); SCC = (sdst[2] != 0)
Var
dst:array[0..1] of PsrRegSlot;
src0,src1,src2:array[0..1] of PsrRegNode;
@ -218,7 +220,26 @@ begin
OpLogicalOr(get_scc,src2[0],src2[1]); //implict cast (int != 0)
end;
procedure TEmit_SOP2.emit_S_ORN2_B64; //SCC = (sdst[2] != 0)
procedure TEmit_SOP2.emit_S_XOR_B64; //sdst[2] = (ssrc0[2] ^ ssrc1[2]); SCC = (sdst[2] != 0)
Var
dst:array[0..1] of PsrRegSlot;
src0,src1,src2:array[0..1] of PsrRegNode;
begin
if not get_sdst7_pair(FSPI.SOP2.SDST,@dst) then Assert(False);
if not fetch_ssrc9_pair(FSPI.SOP2.SSRC0,@src0,dtUInt32) then Assert(False);
if not fetch_ssrc9_pair(FSPI.SOP2.SSRC1,@src1,dtUInt32) then Assert(False);
OpBitwiseXor(dst[0],src0[0],src1[0]);
OpBitwiseXor(dst[1],src0[1],src1[1]);
src2[0]:=dst[0]^.current;
src2[1]:=dst[1]^.current;
OpLogicalOr(get_scc,src2[0],src2[1]); //implict cast (int != 0)
end;
procedure TEmit_SOP2.emit_S_ORN2_B64; //sdst[2] = (ssrc0[2] | ~ssrc1[2]); SCC = (sdst[2] != 0)
Var
dst:array[0..1] of PsrRegSlot;
src0,src1,src2:array[0..1] of PsrRegNode;
@ -240,7 +261,7 @@ begin
OpLogicalOr(get_scc,src2[0],src2[1]); //implict cast (int != 0)
end;
procedure TEmit_SOP2.emit_S_NOR_B64; //SCC = (sdst[2] != 0)
procedure TEmit_SOP2.emit_S_NAND_B64; //sdst[2] = ~(ssrc0[2] & ssrc1[2]); SCC = (sdst[2] != 0)
Var
dst:array[0..1] of PsrRegSlot;
src0,src1,src2:array[0..1] of PsrRegNode;
@ -250,8 +271,30 @@ begin
if not fetch_ssrc9_pair(FSPI.SOP2.SSRC0,@src0,dtUInt32) then Assert(False);
if not fetch_ssrc9_pair(FSPI.SOP2.SSRC1,@src1,dtUInt32) then Assert(False);
src2[0]:=OpBitwiseOrTo(src0[0],src1[0]);
src2[1]:=OpBitwiseOrTo(src0[1],src1[1]);
src2[0]:=OpAndTo(src0[0],src1[0]);
src2[1]:=OpAndTo(src0[1],src1[1]);
OpNot(dst[0],src2[0]);
OpNot(dst[1],src2[1]);
src2[0]:=dst[0]^.current;
src2[1]:=dst[1]^.current;
OpLogicalOr(get_scc,src2[0],src2[1]); //implict cast (int != 0)
end;
procedure TEmit_SOP2.emit_S_NOR_B64; //sdst[2] = ~(ssrc0[2] | ssrc1[2]); SCC = (sdst[2] != 0)
Var
dst:array[0..1] of PsrRegSlot;
src0,src1,src2:array[0..1] of PsrRegNode;
begin
if not get_sdst7_pair(FSPI.SOP2.SDST,@dst) then Assert(False);
if not fetch_ssrc9_pair(FSPI.SOP2.SSRC0,@src0,dtUInt32) then Assert(False);
if not fetch_ssrc9_pair(FSPI.SOP2.SSRC1,@src1,dtUInt32) then Assert(False);
src2[0]:=OpOrTo(src0[0],src1[0]);
src2[1]:=OpOrTo(src0[1],src1[1]);
OpNot(dst[0],src2[0]);
OpNot(dst[1],src2[1]);
@ -307,7 +350,7 @@ begin
src[0]:=fetch_ssrc9(FSPI.SOP2.SSRC0,dtUint32);
src[1]:=fetch_ssrc9(FSPI.SOP2.SSRC1,dtUint32);
offset:=OpBitwiseAndTo(src[1],31);
offset:=OpAndTo(src[1],31);
count :=OpShrTo(src[1],16);
Op3(Op.OpBitFieldUExtract,dtUInt32,dst,src[0],offset,count);
@ -337,8 +380,12 @@ begin
S_OR_B64: emit_S_OR_B64;
S_XOR_B64: emit_S_XOR_B64;
S_ORN2_B64: emit_S_ORN2_B64;
S_NAND_B64: emit_S_NAND_B64;
S_NOR_B64: emit_S_NOR_B64;
S_CSELECT_B32: emit_S_CSELECT_B32;

3
spirv/emit_vbuf_chain.pas
View File

@ -228,7 +228,8 @@ begin
adr.stride :=PV^.stride;
adr.align :=info.GetAlignSize;
adr.fsize :=info.GetSizeFormat;
adr.csize :=Min(info.GetElemSize*info.count,adr.fsize);
adr.csize :=adr.fsize;
//adr.csize :=Min(info.GetElemSize*info.count,adr.fsize);
adr.ioffset:=FSPI.MUBUF.OFFSET;
if (adr.stride=0) then adr.stride:=1;

4
spirv/emit_vop1.pas
View File

@ -97,7 +97,7 @@ begin
dst:=get_vdst8(FSPI.VOP1.VDST);
src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtUInt32);
src:=OpBitwiseAndTo(src,15);
src:=OpAndTo(src,15);
src^.PrepType(ord(dtInt32));
src:=OpISubTo(src,8);
@ -116,7 +116,7 @@ begin
dst:=get_vdst8(FSPI.VOP1.VDST);
src:=fetch_ssrc9(FSPI.VOP1.SRC0,dtUInt32);
src:=OpBitwiseAndTo(src,$FF);
src:=OpAndTo(src,$FF);
src^.PrepType(ord(dtUInt32));
Op1(Op.OpConvertUToF,dtFloat32,dst,src);

12
spirv/emit_vop2.pas
View File

@ -124,7 +124,7 @@ begin
src[0]:=fetch_ssrc9(FSPI.VOP2.SRC0 ,rtype);
src[1]:=fetch_vsrc8(FSPI.VOP2.VSRC1,dtUint32);
src[1]:=OpBitwiseAndTo(src[1],31);
src[1]:=OpAndTo(src[1],31);
src[1]^.PrepType(ord(dtUInt32));
Op2(OpId,src[0]^.dtype,dst,src[0],src[1]);
@ -140,7 +140,7 @@ begin
src[0]:=fetch_ssrc9(FSPI.VOP2.SRC0 ,dtUint32);
src[1]:=fetch_vsrc8(FSPI.VOP2.VSRC1,rtype);
src[0]:=OpBitwiseAndTo(src[0],31);
src[0]:=OpAndTo(src[0],31);
src[0]^.PrepType(ord(dtUInt32));
Op2(OpId,src[1]^.dtype,dst,src[1],src[0]);
@ -275,10 +275,10 @@ begin
bit24:=NewReg_q(dtUInt32,$FFFFFF);
src[0]:=OpBitwiseAndTo(src[0],bit24);
src[0]:=OpAndTo(src[0],bit24);
src[0]^.PrepType(ord(dtInt32));
src[1]:=OpBitwiseAndTo(src[1],bit24);
src[1]:=OpAndTo(src[1],bit24);
src[1]^.PrepType(ord(dtInt32));
OpIMul(dst,src[0],src[1]);
@ -297,10 +297,10 @@ begin
bit24:=NewReg_q(dtUInt32,$FFFFFF);
src[0]:=OpBitwiseAndTo(src[0],bit24);
src[0]:=OpAndTo(src[0],bit24);
src[0]^.PrepType(ord(dtUInt32));
src[1]:=OpBitwiseAndTo(src[1],bit24);
src[1]:=OpAndTo(src[1],bit24);
src[1]^.PrepType(ord(dtUInt32));
OpIMul(dst,src[0],src[1]);

18
spirv/emit_vop3.pas
View File

@ -339,10 +339,10 @@ begin
bit24:=NewReg_q(dtUInt32,$FFFFFF);
src[0]:=OpBitwiseAndTo(src[0],bit24);
src[0]:=OpAndTo(src[0],bit24);
src[0]^.PrepType(ord(dtInt32));
src[1]:=OpBitwiseAndTo(src[1],bit24);
src[1]:=OpAndTo(src[1],bit24);
src[1]^.PrepType(ord(dtInt32));
OpIMul(dst,src[0],src[1]);
@ -366,10 +366,10 @@ begin
bit24:=NewReg_q(dtUInt32,$FFFFFF);
src[0]:=OpBitwiseAndTo(src[0],bit24);
src[0]:=OpAndTo(src[0],bit24);
src[0]^.PrepType(ord(dtUInt32));
src[1]:=OpBitwiseAndTo(src[1],bit24);
src[1]:=OpAndTo(src[1],bit24);
src[1]^.PrepType(ord(dtUInt32));
OpIMul(dst,src[0],src[1]);
@ -530,10 +530,10 @@ begin
bit24:=NewReg_q(dtUInt32,$FFFFFF);
src[0]:=OpBitwiseAndTo(src[0],bit24);
src[0]:=OpAndTo(src[0],bit24);
src[0]^.PrepType(ord(dtInt32));
src[1]:=OpBitwiseAndTo(src[1],bit24);
src[1]:=OpAndTo(src[1],bit24);
src[1]^.PrepType(ord(dtInt32));
OpFmaI32(dst,src[0],src[1],src[2]);
@ -558,10 +558,10 @@ begin
bit24:=NewReg_q(dtUInt32,$FFFFFF);
src[0]:=OpBitwiseAndTo(src[0],bit24);
src[0]:=OpAndTo(src[0],bit24);
src[0]^.PrepType(ord(dtUInt32));
src[1]:=OpBitwiseAndTo(src[1],bit24);
src[1]:=OpAndTo(src[1],bit24);
src[1]^.PrepType(ord(dtUInt32));
OpFmaU32(dst,src[0],src[1],src[2]);
@ -966,7 +966,7 @@ begin
src[1]:=fetch_ssrc9(FSPI.VOP3b.SRC1,dtUInt32);
src[2]:=fetch_ssrc9(FSPI.VOP3b.SRC2,dtUInt32);
src[2]:=OpBitwiseAndTo(src[2],1);
src[2]:=OpAndTo(src[2],1);
src[2]^.PrepType(ord(dtUInt32));
OpIAddExt(dst,car,src[0],src[1]); //src0+src1

21
spirv/pssl-spirv.lpr
View File

@ -122,6 +122,18 @@ begin
Blocks.Clear;
end;
procedure _print_hex(addr:Pointer;size:DWORD);
var
i:DWORD;
begin
For i:=0 to size-1 do
begin
if (i<>0) and ((i mod 16)=0) then Writeln;
Write(HexStr(PByte(addr)[i],2));
end;
if (i<>0) and ((i mod 16)<>0) then Writeln;
end;
procedure load_dump(const fname:RawByteString);
var
M:TMemoryStream;
@ -219,6 +231,8 @@ begin
addr:=Align(addr,8);
end;
//_print_hex(addr,size);
Case W.REG of
mmCOMPUTE_PGM_LO:
begin
@ -356,7 +370,7 @@ begin
if cfg.FPrintInfo then
Writeln('USGPR:',GPU_REGS.PS.RSRC2.USER_SGPR,' VGPRS:',GPU_REGS.PS.RSRC1.VGPRS,' SGPRS:',GPU_REGS.PS.RSRC1.SGPRS);
SprvEmit.InitPs(GPU_REGS.PS.RSRC2,GPU_REGS.PS.INPUT_ENA);
SprvEmit.InitPs(GPU_REGS.PS.RSRC1,GPU_REGS.PS.RSRC2,GPU_REGS.PS.INPUT_ENA);
SprvEmit.SetUserData(@GPU_REGS.PS.USER_DATA);
end;
kShaderTypeVsVs:
@ -364,7 +378,7 @@ begin
if cfg.FPrintInfo then
Writeln('USGPR:',GPU_REGS.VS.RSRC2.USER_SGPR,' VGPRS:',GPU_REGS.VS.RSRC1.VGPRS,' SGPRS:',GPU_REGS.VS.RSRC1.SGPRS);
SprvEmit.InitVs(GPU_REGS.VS.RSRC2,GPU_REGS.VGT_NUM_INSTANCES);
SprvEmit.InitVs(GPU_REGS.VS.RSRC1,GPU_REGS.VS.RSRC2,GPU_REGS.VGT_NUM_INSTANCES);
SprvEmit.SetUserData(@GPU_REGS.VS.USER_DATA);
end;
kShaderTypeCs:
@ -372,9 +386,8 @@ begin
if cfg.FPrintInfo then
Writeln('USGPR:',GPU_REGS.CS.RSRC2.USER_SGPR,' VGPRS:',GPU_REGS.CS.RSRC1.VGPRS,' SGPRS:',GPU_REGS.CS.RSRC1.SGPRS);
SprvEmit.InitCs(GPU_REGS.CS.RSRC2,GPU_REGS.CS.NUM_THREAD_X,GPU_REGS.CS.NUM_THREAD_Y,GPU_REGS.CS.NUM_THREAD_Z);
SprvEmit.InitCs(GPU_REGS.CS.RSRC1,GPU_REGS.CS.RSRC2,GPU_REGS.CS.NUM_THREAD_X,GPU_REGS.CS.NUM_THREAD_Y,GPU_REGS.CS.NUM_THREAD_Z);
SprvEmit.SetUserData(@GPU_REGS.CS.USER_DATA);
end;
else

40
spirv/srBuffer.pas
View File

@ -53,7 +53,9 @@ type
FCount:PtrUint;
dtype:TsrDataType;
pType:PsrType;
sType,pType:PsrType;
FID:Integer; //alloc late
//
@ -72,8 +74,8 @@ type
function FetchValue(_offset,_size:PtrUint;_dtype:TsrDataType):TFieldFetchValue;
function FetchRuntimeArray(_offset,_stride:PtrUint):TFieldFetchValue;
function IsStructUsedRuntimeArray:Boolean;
function IsStructNotUsed:Boolean; inline;
function IsTop:Boolean; inline;
function IsStructNotUsed:Boolean;
function IsTop:Boolean;
function GetStructDecorate:DWORD;
procedure UpdateSize;
function GetSize:PtrUint;
@ -262,6 +264,8 @@ begin
_stride:=_dtype.Child.BitSize div 8;
end;
Assert(_size=(_dtype.BitSize div 8));
node:=Find_le(_offset);
if (node<>nil) then
begin
@ -410,12 +414,25 @@ begin
end;
end;
function TsrField.IsStructNotUsed:Boolean; inline;
function TsrField.IsStructNotUsed:Boolean;
var
child:PsrField;
begin
Result:=(FCount<=1) and (pParent<>nil);
Result:=False;
if IsTop then Exit;
if (FCount>1) then Exit;
child:=First;
if (child=nil) then Exit;
if (child^.offset<>0) then Exit;
if (child^.size<>stride) then Exit;
Result:=True;
end;
function TsrField.IsTop:Boolean; inline;
function TsrField.IsTop:Boolean;
begin
Result:=(pParent=nil);
end;
@ -886,7 +903,7 @@ begin
pVar:=node^.pVar;
if (pVar<>nil) and node^.IsUsed then
begin
pDebugInfoList^.OpSourceExtension(node^.GetString);
pDebugInfoList^.OpSource(node^.GetString);
end;
node:=Next(node);
end;
@ -1022,18 +1039,19 @@ var
node:PsrField;
SD:DWORD;
begin
if (pField^.dtype<>dtTypeStruct) then Exit;
if (pField^.pType=nil) then Exit;
if (pField^.sType=nil) then Exit;
if (pField^.sType^.dtype<>dtTypeStruct) then Exit;
pDecorateList:=FEmit.GetDecorateList;
SD:=pField^.GetStructDecorate;
if (SD<>DWORD(-1)) then
begin
pDecorateList^.OpDecorate(pField^.pType,SD,0);
pDecorateList^.OpDecorate(pField^.sType,SD,0);
end;
node:=pField^.First;
While (node<>nil) do
begin
pDecorateList^.OpMemberDecorate(pField^.pType,node^.FID,node^.offset);
pDecorateList^.OpMember(pField^.sType,node^.FID,node^.offset);
node:=pField^.Next(node);
end;
end;

88
spirv/srDecorate.pas
View File

@ -6,6 +6,7 @@ interface
uses
spirv,
ginodes,
srNode,
srOp;
@ -16,21 +17,42 @@ type
function emit_glsl_ext:PSpirvOp;
end;
PsrDecorate=^TsrDecorate;
TsrDecorate=object
private
pLeft,pRight:PsrDecorate;
//----
key:packed record
data:PsrNode;
param:array[0..2] of DWORD;
end;
function c(n1,n2:PsrDecorate):Integer; static;
public
node:PSpirvOp;
end;
PsrDecorateList=^TsrDecorateList;
TsrDecorateList=object(TsrOpBlockCustom)
type
TNodeFetch=specialize TNodeFetch<PsrDecorate,TsrDecorate>;
var
FNTree:TNodeFetch;
function Fetch(data:PsrNode;param1,param2,param3:DWORD):PsrDecorate;
procedure OpDecorate(Data:PsrNode;dec_id,param:DWORD);
procedure OpMemberDecorate(Data:PsrNode;index,offset:DWORD);
procedure OpMember (Data:PsrNode;index,offset:DWORD);
end;
PsrDebugInfoList=^TsrDebugInfoList;
TsrDebugInfoList=object(TsrOpBlockCustom)
procedure OpSourceExtension(const n:RawByteString);
procedure OpName(Data:PsrNode;const n:RawByteString);
procedure OpSource(const n:RawByteString);
procedure OpName (Data:PsrNode;const n:RawByteString);
function OpString(const n:RawByteString):PsrNode;
end;
implementation
//
function TsrHeaderList.emit_glsl_ext:PSpirvOp;
begin
if (FGLSL_std_450=nil) then
@ -42,13 +64,57 @@ begin
Result:=FGLSL_std_450;
end;
//
function TsrDecorate.c(n1,n2:PsrDecorate):Integer;
var
i:Byte;
begin
//first data
Result:=Integer(n1^.key.data>n2^.key.data)-Integer(n1^.key.data<n2^.key.data);
if (Result<>0) then Exit;
//param[i]
For i:=0 to 2 do
begin
Result:=Integer(n1^.key.param[i]>n2^.key.param[i])-Integer(n1^.key.param[i]<n2^.key.param[i]);
if (Result<>0) then Exit;
end;
end;
//
function TsrDecorateList.Fetch(data:PsrNode;param1,param2,param3:DWORD):PsrDecorate;
var
node:TsrDecorate;
begin
node:=Default(TsrDecorate);
node.key.data:=data;
node.key.param[0]:=param1;
node.key.param[1]:=param2;
node.key.param[2]:=param3;
Result:=FNTree.Find(@node);
if (Result=nil) then
begin
Result:=Emit.Alloc(SizeOf(TsrDecorate));
Move(node,Result^,SizeOf(TsrDecorate));
FNTree.Insert(Result);
end;
end;
procedure TsrDecorateList.OpDecorate(Data:PsrNode;dec_id,param:DWORD);
var
deco:PsrDecorate;
node:PSpirvOp;
begin
deco:=Fetch(Data,Op.OpDecorate,dec_id,param);
if (deco^.node<>nil) then Exit;
node:=AddSpirvOp(Op.OpDecorate);
node^.AddParam(Data);
node^.AddLiteral(dec_id,Decoration.GetStr(dec_id));
Case dec_id of
Decoration.BuiltIn:
node^.AddLiteral(param,BuiltIn.GetStr(param));
@ -63,20 +129,30 @@ begin
else;
end;
deco^.node:=node;
end;
procedure TsrDecorateList.OpMemberDecorate(Data:PsrNode;index,offset:DWORD);
procedure TsrDecorateList.OpMember(Data:PsrNode;index,offset:DWORD);
var
deco:PsrDecorate;
node:PSpirvOp;
begin
deco:=Fetch(Data,Op.OpMemberDecorate,index,offset);
if (deco^.node<>nil) then Exit;
node:=AddSpirvOp(Op.OpMemberDecorate);
node^.AddParam(Data);
node^.AddLiteral(index);
node^.AddLiteral(Decoration.Offset,Decoration.GetStr(Decoration.Offset));
node^.AddLiteral(offset);
deco^.node:=node;
end;
procedure TsrDebugInfoList.OpSourceExtension(const n:RawByteString);
//
procedure TsrDebugInfoList.OpSource(const n:RawByteString);
var
node:PSpirvOp;
begin
@ -103,5 +179,7 @@ begin
Result:=node^.pDst;
end;
//
end.

4
spirv/srLayout.pas
View File

@ -582,7 +582,7 @@ begin
node:=First;
While (node<>nil) do
begin
pDebugInfoList^.OpSourceExtension(node^.GetString);
pDebugInfoList^.OpSource(node^.GetString);
node:=Next(node);
end;
end;
@ -604,7 +604,7 @@ begin
block:=pHeap^.FindByPtr(node^.pData);
if (block<>nil) then
begin
pDebugInfoList^.OpSourceExtension(node^.GetFuncString(block^.Size));
pDebugInfoList^.OpSource(node^.GetFuncString(block^.Size));
end;
end;
node:=Next(node);

2
spirv/srUniform.pas
View File

@ -442,7 +442,7 @@ begin
pVar:=node^.pVar;
if (pVar<>nil) and node^.IsUsed then
begin
pDebugInfoList^.OpSourceExtension(node^.GetString);
pDebugInfoList^.OpSource(node^.GetString);
end;
node:=Next(node);
end;

2
spirv/srVertLayout.pas
View File

@ -179,7 +179,7 @@ begin
pVar:=node^.pVar;
if (pVar<>nil) and node^.IsUsed then
begin
FDebugInfo^.OpSourceExtension(node^.GetString);
FDebugInfo^.OpSource(node^.GetString);
end;
node:=Next(node);
end;

6
vulkan/vShaderManager.pas
View File

@ -231,17 +231,17 @@ begin
case FStage of
vShaderStagePs :
begin
SprvEmit.InitPs(GPU_REGS.SPI.PS.RSRC2,GPU_REGS.SPI.PS.INPUT_ENA);
SprvEmit.InitPs(GPU_REGS.SPI.PS.RSRC1,GPU_REGS.SPI.PS.RSRC2,GPU_REGS.SPI.PS.INPUT_ENA);
SprvEmit.SetUserData(@GPU_REGS.SPI.PS.USER_DATA);
end;
vShaderStageVs:
begin
SprvEmit.InitVs(GPU_REGS.SPI.VS.RSRC2,GPU_REGS.VGT_NUM_INSTANCES);
SprvEmit.InitVs(GPU_REGS.SPI.VS.RSRC1,GPU_REGS.SPI.VS.RSRC2,GPU_REGS.VGT_NUM_INSTANCES);
SprvEmit.SetUserData(@GPU_REGS.SPI.VS.USER_DATA);
end;
vShaderStageCs:
begin
SprvEmit.InitCs(GPU_REGS.SPI.CS.RSRC2,GPU_REGS.SPI.CS.NUM_THREAD_X,GPU_REGS.SPI.CS.NUM_THREAD_Y,GPU_REGS.SPI.CS.NUM_THREAD_Z);
SprvEmit.InitCs(GPU_REGS.SPI.CS.RSRC1,GPU_REGS.SPI.CS.RSRC2,GPU_REGS.SPI.CS.NUM_THREAD_X,GPU_REGS.SPI.CS.NUM_THREAD_Y,GPU_REGS.SPI.CS.NUM_THREAD_Z);
SprvEmit.SetUserData(@GPU_REGS.SPI.CS.USER_DATA);
end;