ShuriZma
/
pcsx2
mirror of https://github.com/PCSX2/pcsx2.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

microVU: Accurate CLIP

This commit is contained in:
TellowKrinkle 2025-02-17 23:26:39 -06:00
parent c0e69ff9ab
commit 57876cbe5a
2 changed files with 19 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
pcsx2/x86/microVU_Misc.h
View File

@ -21,6 +21,7 @@ struct mVU_Globals
u32 signbit [4] = __four(0x80000000); u32 signbit [4] = __four(0x80000000);
u32 minvals [4] = __four(0xff7fffff); u32 minvals [4] = __four(0xff7fffff);
u32 maxvals [4] = __four(0x7f7fffff); u32 maxvals [4] = __four(0x7f7fffff);
u32 exponent[4] = __four(0x7f800000);
u32 one [4] = __four(0x3f800000); u32 one [4] = __four(0x3f800000);
u32 Pi4 [4] = __four(0x3f490fdb); u32 Pi4 [4] = __four(0x3f490fdb);
u32 T1 [4] = __four(0x3f7ffff5); u32 T1 [4] = __four(0x3f7ffff5);

36
pcsx2/x86/microVU_Upper.inl
View File

@ -544,36 +544,36 @@ mVUop(mVU_CLIP)
const xmm& Fs = mVU.regAlloc->allocReg(_Fs_, 0, 0xf); const xmm& Fs = mVU.regAlloc->allocReg(_Fs_, 0, 0xf);
const xmm& Ft = mVU.regAlloc->allocReg(_Ft_, 0, 0x1); const xmm& Ft = mVU.regAlloc->allocReg(_Ft_, 0, 0x1);
const xmm& t1 = mVU.regAlloc->allocReg(); const xmm& t1 = mVU.regAlloc->allocReg();
const xmm& t2 = mVU.regAlloc->allocReg();
mVUunpack_xyzw(Ft, Ft, 0); mVUunpack_xyzw(Ft, Ft, 0);
mVUallocCFLAGa(mVU, gprT1, cFLAG.lastWrite); mVUallocCFLAGa(mVU, gprT1, cFLAG.lastWrite);
xSHL(gprT1, 6); xSHL(gprT1, 6);
xAND.PS(Ft, ptr128[mVUglob.absclip]); xMOVAPS (t1, ptr128[mVUglob.exponent]);
xMOVAPS(t1, Ft); xPAND (t1, Fs);
xPOR(t1, ptr128[mVUglob.signbit]); xPXOR (t2, t2);
xPCMP.EQD(t1, t2); // Denormal check
xPANDN (t1, Fs); // If denormal, set to zero, which can't be greater than any nonnegative denormal in Ft
xPAND (Ft, ptr128[mVUglob.absclip]);
xCMPNLE.PS(t1, Fs); // -w, -z, -y, -x xMOVAPS (Fs, ptr128[mVUglob.signbit]);
xCMPLT.PS(Ft, Fs); // +w, +z, +y, +x xPXOR (Fs, t1); // Negate
xPCMP.GTD(t1, Ft); // +w, +z, +y, +x
xPCMP.GTD(Fs, Ft); // -w, -z, -y, -x
xMOVAPS(Fs, Ft); // Fs = +w, +z, +y, +x xPBLEND.W (Fs, t1, 0x55); // Squish together
xUNPCK.LPS(Ft, t1); // Ft = -y,+y,-x,+x xPACK.SSWB(Fs, Fs); // Convert u16 to u8
xUNPCK.HPS(Fs, t1); // Fs = -w,+w,-z,+z xPMOVMSKB (gprT2, Fs); // Get bitmask
xAND (gprT2, 0x3f); // Mask unused stuff
xMOVMSKPS(gprT2, Fs); // -w,+w,-z,+z xAND (gprT1, 0xffffff);
xAND(gprT2, 0x3); xOR (gprT1, gprT2);
xSHL(gprT2, 4);
xOR(gprT1, gprT2);
xMOVMSKPS(gprT2, Ft); // -y,+y,-x,+x
xAND(gprT2, 0xf);
xOR(gprT1, gprT2);
xAND(gprT1, 0xffffff);
mVUallocCFLAGb(mVU, gprT1, cFLAG.write); mVUallocCFLAGb(mVU, gprT1, cFLAG.write);
mVU.regAlloc->clearNeeded(Fs); mVU.regAlloc->clearNeeded(Fs);
mVU.regAlloc->clearNeeded(Ft); mVU.regAlloc->clearNeeded(Ft);
mVU.regAlloc->clearNeeded(t1); mVU.regAlloc->clearNeeded(t1);
mVU.regAlloc->clearNeeded(t2);
mVU.profiler.EmitOp(opCLIP); mVU.profiler.EmitOp(opCLIP);
} }
pass3 pass3