Skip to content

Instantly share code, notes, and snippets.

@sad1e

sad1e/opt_fsr.fxh

Forked from atyuwen/opt_fsr.fxh
Created April 7, 2025 08:24
Show Gist options
  • Save sad1e/42fb2e59d2201db2651623fba9a15a9f to your computer and use it in GitHub Desktop.
Save sad1e/42fb2e59d2201db2651623fba9a15a9f to your computer and use it in GitHub Desktop.
Download ZIP

Revisions

  1. @atyuwen atyuwen revised this gist Apr 4, 2025. 1 changed file with 134 additions and 0 deletions.
    134 changes: 134 additions & 0 deletions opt_fsr.fxh
    View file
    Original file line number Diff line number Diff line change
    @@ -0,0 +1,134 @@
    //==============================================================================================================================
    // An optimized AMD FSR 1.0 implementation for Mobiles
    // EASU and RCAS are combined in a single pass.
    // Based on https://github.com/GPUOpen-Effects/FidelityFX-FSR/blob/master/ffx-fsr/ffx_fsr1.h
    // Details can be found: https://atyuwen.github.io/posts/optimizing-fsr/
    // Distributed under the MIT License. Copyright (c) 2021 atyuwen.
    // -- FsrEasuSampleH should be implemented by calling shader, like following:
    // AH3 FsrEasuSampleH(AF2 p) { return MyTex.SampleLevel(LinearSampler, p, 0).xyz; }
    //==============================================================================================================================
    void FsrMobile(
    out AH3 pix,
    AF2 ip,
    AF4 con0,
    AF4 con1,
    AF4 con2,
    AF4 con3){
    //------------------------------------------------------------------------------------------------------------------------------
    // Direction is the '+' diff.
    // A
    // B C D
    // E
    AF2 pp=(ip)*(con0.xy)+(con0.zw);
    AF2 tc=(pp+AF2_(0.5))*con1.xy;
    AH3 sC=FsrEasuSampleH(tc);
    #if 0 // Set to 1 to make FSR only affect the screen's central region.
    if (any(abs(tc-0.5)>AF1(0.75/2.0))) {
    pix=sC;
    return;
    }
    #endif
    AH3 sA=FsrEasuSampleH(tc-AF2(0, con1.y));
    AH3 sB=FsrEasuSampleH(tc-AF2(con1.x, 0));
    AH3 sD=FsrEasuSampleH(tc+AF2(con1.x, 0));
    AH3 sE=FsrEasuSampleH(tc+AF2(0, con1.y));
    //------------------------------------------------------------------------------------------------------------------------------
    // Combined RCAS: Min and max of ring.
    AH1 mn4R=min(AMin3H1(sA.r,sB.r,sD.r),sE.r);
    AH1 mn4G=min(AMin3H1(sA.g,sB.g,sD.g),sE.g);
    AH1 mn4B=min(AMin3H1(sA.b,sB.b,sD.b),sE.b);
    AH1 mx4R=max(AMax3H1(sA.r,sB.r,sD.r),sE.r);
    AH1 mx4G=max(AMax3H1(sA.g,sB.g,sD.g),sE.g);
    AH1 mx4B=max(AMax3H1(sA.b,sB.b,sD.b),sE.b);
    // Immediate constants for peak range.
    AH2 peakC=AH2(1.0,-1.0*4.0);
    // Limiters, these need to be high precision RCPs.
    AH1 hitMinR=mn4R*ARcpH1(AH1_(4.0)*mx4R);
    AH1 hitMinG=mn4G*ARcpH1(AH1_(4.0)*mx4G);
    AH1 hitMinB=mn4B*ARcpH1(AH1_(4.0)*mx4B);
    AH1 hitMaxR=(peakC.x-mx4R)*ARcpH1(AH1_(4.0)*mn4R+peakC.y);
    AH1 hitMaxG=(peakC.x-mx4G)*ARcpH1(AH1_(4.0)*mn4G+peakC.y);
    AH1 hitMaxB=(peakC.x-mx4B)*ARcpH1(AH1_(4.0)*mn4B+peakC.y);
    AH1 lobeR=max(-hitMinR,hitMaxR);
    AH1 lobeG=max(-hitMinG,hitMaxG);
    AH1 lobeB=max(-hitMinB,hitMaxB);
    AH1 lobe=max(AH1_(-FSR_RCAS_LIMIT),min(AMax3H1(lobeR,lobeG,lobeB),AH1_(0.0)))*AH1_(con0.x);
    // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes.
    AH1 rcpL=ARcpH1(AH1_(4.0)*lobe+AH1_(1.0));
    AH3 contrast=(lobe*sA+lobe*sB+lobe*sD+lobe*sE)*rcpL;
    //------------------------------------------------------------------------------------------------------------------------------
    AH1 lA=sA.r*AH1_(0.5)+sA.g;
    AH1 lB=sB.r*AH1_(0.5)+sB.g;
    AH1 lC=sC.r*AH1_(0.5)+sC.g;
    AH1 lD=sD.r*AH1_(0.5)+sD.g;
    AH1 lE=sE.r*AH1_(0.5)+sE.g;
    // Then takes magnitude from abs average of both sides of 'C'.
    // Length converts gradient reversal to 0, smoothly to non-reversal at 1, shaped, then adding horz and vert terms.
    AH1 dc=lD-lC;
    AH1 cb=lC-lB;
    AH1 lenX=max(abs(dc),abs(cb));
    lenX=ARcpH1(lenX);
    AH1 dirX=lD-lB;
    lenX=ASatH1(abs(dirX)*lenX);
    lenX*=lenX;
    // Repeat for the y axis.
    AH1 ec=lE-lC;
    AH1 ca=lC-lA;
    AH1 lenY=max(abs(ec),abs(ca));
    lenY=ARcpH1(lenY);
    AH1 dirY=lE-lA;
    lenY=ASatH1(abs(dirY)*lenY);
    AH1 len = lenY * lenY + lenX;
    AH2 dir = AH2(dirX, dirY);
    //------------------------------------------------------------------------------------------------------------------------------
    AH2 dir2=dir*dir;
    AH1 dirR=dir2.x+dir2.y;
    if (dirR<AH1_(1.0/64.0)) {
    pix=contrast+sC*rcpL;
    return;
    }
    dirR=ARsqH1(dirR);
    dir*=AH2_(dirR);
    len=len*AH1_(0.5);
    len*=len;
    AH1 stretch=(dir.x*dir.x+dir.y*dir.y)*ARcpH1(max(abs(dir.x),abs(dir.y)));
    AH2 len2=AH2(AH1_(1.0)+(stretch-AH1_(1.0))*len,AH1_(1.0)+AH1_(-0.5)*len);
    AH1 lob=AH1_(0.5)+AH1_((1.0/4.0-0.04)-0.5)*len;
    AH1 clp=ARcpH1(lob);
    //------------------------------------------------------------------------------------------------------------------------------
    AF2 fp=floor(pp);
    pp-=fp;
    AH2 ppp=AH2(pp);
    AF2 p0=fp*(con1.xy)+(con1.zw);
    AF2 p1=p0+(con2.xy);
    AF2 p2=p0+(con2.zw);
    AF2 p3=p0+(con3.xy);
    p0.y-=con1.w; p3.y+=con1.w;
    AH4 fgcbR=FsrEasuRH(p0);
    AH4 fgcbG=FsrEasuGH(p0);
    AH4 fgcbB=FsrEasuBH(p0);
    AH4 ijfeR=FsrEasuRH(p1);
    AH4 ijfeG=FsrEasuGH(p1);
    AH4 ijfeB=FsrEasuBH(p1);
    AH4 klhgR=FsrEasuRH(p2);
    AH4 klhgG=FsrEasuGH(p2);
    AH4 klhgB=FsrEasuBH(p2);
    AH4 nokjR=FsrEasuRH(p3);
    AH4 nokjG=FsrEasuGH(p3);
    AH4 nokjB=FsrEasuBH(p3);
    //------------------------------------------------------------------------------------------------------------------------------
    // This part is different for FP16, working pairs of taps at a time.
    AH2 pR=AH2_(0.0);
    AH2 pG=AH2_(0.0);
    AH2 pB=AH2_(0.0);
    AH2 pW=AH2_(0.0);
    FsrEasuTapH(pR,pG,pB,pW,AH2( 1.0, 0.0)-ppp.xx,AH2(-1.0,-1.0)-ppp.yy,dir,len2,lob,clp,fgcbR.zw,fgcbG.zw,fgcbB.zw);
    FsrEasuTapH(pR,pG,pB,pW,AH2(-1.0, 0.0)-ppp.xx,AH2( 1.0, 1.0)-ppp.yy,dir,len2,lob,clp,ijfeR.xy,ijfeG.xy,ijfeB.xy);
    FsrEasuTapH(pR,pG,pB,pW,AH2( 0.0,-1.0)-ppp.xx,AH2( 0.0, 0.0)-ppp.yy,dir,len2,lob,clp,ijfeR.zw,ijfeG.zw,ijfeB.zw);
    FsrEasuTapH(pR,pG,pB,pW,AH2( 1.0, 2.0)-ppp.xx,AH2( 1.0, 1.0)-ppp.yy,dir,len2,lob,clp,klhgR.xy,klhgG.xy,klhgB.xy);
    FsrEasuTapH(pR,pG,pB,pW,AH2( 2.0, 1.0)-ppp.xx,AH2( 0.0, 0.0)-ppp.yy,dir,len2,lob,clp,klhgR.zw,klhgG.zw,klhgB.zw);
    FsrEasuTapH(pR,pG,pB,pW,AH2( 0.0, 1.0)-ppp.xx,AH2( 2.0, 2.0)-ppp.yy,dir,len2,lob,clp,nokjR.xy,nokjG.xy,nokjB.xy);
    AH3 aC=AH3(pR.x+pR.y,pG.x+pG.y,pB.x+pB.y);
    AH1 aW=pW.x+pW.y;
    //------------------------------------------------------------------------------------------------------------------------------
    pix=contrast+aC*AH3_(ARcpH1(aW)*rcpL);}
  2. @atyuwen atyuwen renamed this gist Apr 4, 2025. 1 changed file with 0 additions and 0 deletions.
    0 opt_fsr.fxh → opt_fsr_easu.fxh
    View file
    File renamed without changes.
  3. @atyuwen atyuwen revised this gist Sep 13, 2021. 1 changed file with 2 additions and 2 deletions.
    4 changes: 2 additions & 2 deletions opt_fsr.fxh
    View file
    Original file line number Diff line number Diff line change
    @@ -2,9 +2,9 @@
    // An optimized AMD FSR's EASU implementation for Mobiles
    // Based on https://github.com/GPUOpen-Effects/FidelityFX-FSR/blob/master/ffx-fsr/ffx_fsr1.h
    // Details can be found: https://atyuwen.github.io/posts/optimizing-fsr/
    // Distributed under the MIT License.
    // Distributed under the MIT License. Copyright (c) 2021 atyuwen.
    // -- FsrEasuSampleH should be implemented by calling shader, like following:
    // half3 FsrEasuSampleH(float2 p) { return MyTex.SampleLevel(LinearSampler, p, 0).xyz; }
    // AH3 FsrEasuSampleH(AF2 p) { return MyTex.SampleLevel(LinearSampler, p, 0).xyz; }
    //==============================================================================================================================
    void FsrEasuL(
    out AH3 pix,
  4. @atyuwen atyuwen revised this gist Sep 2, 2021. 1 changed file with 3 additions and 3 deletions.
    6 changes: 3 additions & 3 deletions opt_fsr.fxh
    View file
    Original file line number Diff line number Diff line change
    @@ -55,14 +55,14 @@
    pix = sC;
    return;
    }
    dirR=rsqrt(dirR);
    dirR=ARsqH1(dirR);
    dir*=AH2_(dirR);
    len=len*AH1_(0.5);
    len*=len;
    AH1 stretch=(dir.x*dir.x+dir.y*dir.y)*rcp(max(abs(dir.x),abs(dir.y)));
    AH1 stretch=(dir.x*dir.x+dir.y*dir.y)*ARcpH1(max(abs(dir.x),abs(dir.y)));
    AH2 len2=AH2(AH1_(1.0)+(stretch-AH1_(1.0))*len,AH1_(1.0)+AH1_(-0.5)*len);
    AH1 lob=AH1_(0.5)+AH1_((1.0/4.0-0.04)-0.5)*len;
    AH1 clp=rcp(lob);
    AH1 clp=ARcpH1(lob);
    //------------------------------------------------------------------------------------------------------------------------------
    AF2 fp=floor(pp);
    pp-=fp;
  5. @atyuwen atyuwen revised this gist Aug 27, 2021. 1 changed file with 2 additions and 1 deletion.
    3 changes: 2 additions & 1 deletion opt_fsr.fxh
    View file
    Original file line number Diff line number Diff line change
    @@ -1,7 +1,8 @@
    //==============================================================================================================================
    // An optimized AMD FSR's EASU implementation for Mobiles
    // Based on https://github.com/GPUOpen-Effects/FidelityFX-FSR/blob/master/ffx-fsr/ffx_fsr1.h
    // Distributed under the MIT License.
    // Details can be found: https://atyuwen.github.io/posts/optimizing-fsr/
    // Distributed under the MIT License.
    // -- FsrEasuSampleH should be implemented by calling shader, like following:
    // half3 FsrEasuSampleH(float2 p) { return MyTex.SampleLevel(LinearSampler, p, 0).xyz; }
    //==============================================================================================================================
  6. @atyuwen atyuwen revised this gist Aug 27, 2021. 1 changed file with 1 addition and 1 deletion.
    2 changes: 1 addition & 1 deletion opt_fsr.fxh
    View file
    Original file line number Diff line number Diff line change
    @@ -1,6 +1,7 @@
    //==============================================================================================================================
    // An optimized AMD FSR's EASU implementation for Mobiles
    // Based on https://github.com/GPUOpen-Effects/FidelityFX-FSR/blob/master/ffx-fsr/ffx_fsr1.h
    // Distributed under the MIT License.
    // -- FsrEasuSampleH should be implemented by calling shader, like following:
    // half3 FsrEasuSampleH(float2 p) { return MyTex.SampleLevel(LinearSampler, p, 0).xyz; }
    //==============================================================================================================================
    @@ -97,5 +98,4 @@
    AH3 aC=AH3(pR.x+pR.y,pG.x+pG.y,pB.x+pB.y);
    AH1 aW=pW.x+pW.y;
    //------------------------------------------------------------------------------------------------------------------------------
    // Slightly different for FP16 version due to combined min and max.
    pix=aC*AH3_(ARcpH1(aW));}
  7. @atyuwen atyuwen created this gist Aug 25, 2021.
    101 changes: 101 additions & 0 deletions opt_fsr.fxh
    View file
    Original file line number Diff line number Diff line change
    @@ -0,0 +1,101 @@
    //==============================================================================================================================
    // An optimized AMD FSR's EASU implementation for Mobiles
    // Based on https://github.com/GPUOpen-Effects/FidelityFX-FSR/blob/master/ffx-fsr/ffx_fsr1.h
    // -- FsrEasuSampleH should be implemented by calling shader, like following:
    // half3 FsrEasuSampleH(float2 p) { return MyTex.SampleLevel(LinearSampler, p, 0).xyz; }
    //==============================================================================================================================
    void FsrEasuL(
    out AH3 pix,
    AF2 ip,
    AF4 con0,
    AF4 con1,
    AF4 con2,
    AF4 con3){
    //------------------------------------------------------------------------------------------------------------------------------
    // Direction is the '+' diff.
    // A
    // B C D
    // E
    AF2 pp=(ip)*(con0.xy)+(con0.zw);
    AF2 tc=(pp+AF2_(0.5))*con1.xy;
    AH3 sA=FsrEasuSampleH(tc-AF2(0, con1.y));
    AH3 sB=FsrEasuSampleH(tc-AF2(con1.x, 0));
    AH3 sC=FsrEasuSampleH(tc);
    AH3 sD=FsrEasuSampleH(tc+AF2(con1.x, 0));
    AH3 sE=FsrEasuSampleH(tc+AF2(0, con1.y));
    AH1 lA=sA.r*AH1_(0.5)+sA.g;
    AH1 lB=sB.r*AH1_(0.5)+sB.g;
    AH1 lC=sC.r*AH1_(0.5)+sC.g;
    AH1 lD=sD.r*AH1_(0.5)+sD.g;
    AH1 lE=sE.r*AH1_(0.5)+sE.g;
    // Then takes magnitude from abs average of both sides of 'C'.
    // Length converts gradient reversal to 0, smoothly to non-reversal at 1, shaped, then adding horz and vert terms.
    AH1 dc=lD-lC;
    AH1 cb=lC-lB;
    AH1 lenX=max(abs(dc),abs(cb));
    lenX=ARcpH1(lenX);
    AH1 dirX=lD-lB;
    lenX=ASatH1(abs(dirX)*lenX);
    lenX*=lenX;
    // Repeat for the y axis.
    AH1 ec=lE-lC;
    AH1 ca=lC-lA;
    AH1 lenY=max(abs(ec),abs(ca));
    lenY=ARcpH1(lenY);
    AH1 dirY=lE-lA;
    lenY=ASatH1(abs(dirY)*lenY);
    AH1 len = lenY * lenY + lenX;
    AH2 dir = AH2(dirX, dirY);
    //------------------------------------------------------------------------------------------------------------------------------
    AH2 dir2=dir*dir;
    AH1 dirR=dir2.x+dir2.y;
    if (dirR<AH1_(1.0/64.0)) {
    pix = sC;
    return;
    }
    dirR=rsqrt(dirR);
    dir*=AH2_(dirR);
    len=len*AH1_(0.5);
    len*=len;
    AH1 stretch=(dir.x*dir.x+dir.y*dir.y)*rcp(max(abs(dir.x),abs(dir.y)));
    AH2 len2=AH2(AH1_(1.0)+(stretch-AH1_(1.0))*len,AH1_(1.0)+AH1_(-0.5)*len);
    AH1 lob=AH1_(0.5)+AH1_((1.0/4.0-0.04)-0.5)*len;
    AH1 clp=rcp(lob);
    //------------------------------------------------------------------------------------------------------------------------------
    AF2 fp=floor(pp);
    pp-=fp;
    AH2 ppp=AH2(pp);
    AF2 p0=fp*(con1.xy)+(con1.zw);
    AF2 p1=p0+(con2.xy);
    AF2 p2=p0+(con2.zw);
    AF2 p3=p0+(con3.xy);
    p0.y-=con1.w; p3.y+=con1.w;
    AH4 fgcbR=FsrEasuRH(p0);
    AH4 fgcbG=FsrEasuGH(p0);
    AH4 fgcbB=FsrEasuBH(p0);
    AH4 ijfeR=FsrEasuRH(p1);
    AH4 ijfeG=FsrEasuGH(p1);
    AH4 ijfeB=FsrEasuBH(p1);
    AH4 klhgR=FsrEasuRH(p2);
    AH4 klhgG=FsrEasuGH(p2);
    AH4 klhgB=FsrEasuBH(p2);
    AH4 nokjR=FsrEasuRH(p3);
    AH4 nokjG=FsrEasuGH(p3);
    AH4 nokjB=FsrEasuBH(p3);
    //------------------------------------------------------------------------------------------------------------------------------
    // This part is different for FP16, working pairs of taps at a time.
    AH2 pR=AH2_(0.0);
    AH2 pG=AH2_(0.0);
    AH2 pB=AH2_(0.0);
    AH2 pW=AH2_(0.0);
    FsrEasuTapH(pR,pG,pB,pW,AH2( 1.0, 0.0)-ppp.xx,AH2(-1.0,-1.0)-ppp.yy,dir,len2,lob,clp,fgcbR.zw,fgcbG.zw,fgcbB.zw);
    FsrEasuTapH(pR,pG,pB,pW,AH2(-1.0, 0.0)-ppp.xx,AH2( 1.0, 1.0)-ppp.yy,dir,len2,lob,clp,ijfeR.xy,ijfeG.xy,ijfeB.xy);
    FsrEasuTapH(pR,pG,pB,pW,AH2( 0.0,-1.0)-ppp.xx,AH2( 0.0, 0.0)-ppp.yy,dir,len2,lob,clp,ijfeR.zw,ijfeG.zw,ijfeB.zw);
    FsrEasuTapH(pR,pG,pB,pW,AH2( 1.0, 2.0)-ppp.xx,AH2( 1.0, 1.0)-ppp.yy,dir,len2,lob,clp,klhgR.xy,klhgG.xy,klhgB.xy);
    FsrEasuTapH(pR,pG,pB,pW,AH2( 2.0, 1.0)-ppp.xx,AH2( 0.0, 0.0)-ppp.yy,dir,len2,lob,clp,klhgR.zw,klhgG.zw,klhgB.zw);
    FsrEasuTapH(pR,pG,pB,pW,AH2( 0.0, 1.0)-ppp.xx,AH2( 2.0, 2.0)-ppp.yy,dir,len2,lob,clp,nokjR.xy,nokjG.xy,nokjB.xy);
    AH3 aC=AH3(pR.x+pR.y,pG.x+pG.y,pB.x+pB.y);
    AH1 aW=pW.x+pW.y;
    //------------------------------------------------------------------------------------------------------------------------------
    // Slightly different for FP16 version due to combined min and max.
    pix=aC*AH3_(ARcpH1(aW));}