mirror of https://github.com/PCSX2/pcsx2.git
xpad: compatible with pcsx
GSdx: minor optimizations, +crcs git-svn-id: http://pcsx2.googlecode.com/svn/trunk@496 96395faa-99c1-11dd-bbfe-3dabce05a288
This commit is contained in:
gabest11
parent
c987d23b9f
commit
d14d09cab3
|
|
@ -65,6 +65,8 @@ CRC::Game CRC::m_games[] =
|
|||
{0x98D4BC93, MetalGearSolid3, EU, false},
|
||||
{0x86BC3040, MetalGearSolid3, US, false}, //Subsistance disc 1
|
||||
{0x0481AD8A, MetalGearSolid3, JP, false},
|
||||
{0x79ED26AD, MetalGearSolid3, EU, false},
|
||||
{0x5E31EA42, MetalGearSolid3, EU, false},
|
||||
{0x278722BF, DBZBT2, US, false},
|
||||
{0xFE961D28, DBZBT2, US, false},
|
||||
{0x0393B6BE, DBZBT2, EU, false},
|
||||
|
|
@ -110,6 +112,8 @@ CRC::Game CRC::m_games[] =
|
|||
{0xA61A4C6D, GodOfWar, Unknown, false},
|
||||
{0xE23D532B, GodOfWar, Unknown, false},
|
||||
{0x2F123FD8, GodOfWar2, RU, false},
|
||||
{0x2F123FD8, GodOfWar2, US, false},
|
||||
{0x44A8A22A, GodOfWar2, EU, false},
|
||||
{0x5D482F18, JackieChanAdv, Unknown, false},
|
||||
{0xf0a6d880, HarvestMoon, US, true},
|
||||
{0x75c01a04, NamcoXCapcom, US, false},
|
||||
|
|
|
|||
|
|
@ -19,7 +19,7 @@
|
|||
*
|
||||
*/
|
||||
|
||||
// TODO: x64
|
||||
// TODO: x64 (use the extra regs to avoid spills of zs, zd, uf, vf, rb, ga and keep a few constants in the last two like aref or afix)
|
||||
|
||||
#include "StdAfx.h"
|
||||
#include "GSDrawScanlineCodeGenerator.h"
|
||||
|
|
@ -48,29 +48,28 @@ void GSDrawScanlineCodeGenerator::Generate()
|
|||
|
||||
align(16);
|
||||
|
||||
L("@@");
|
||||
L("loop");
|
||||
|
||||
// ecx = steps
|
||||
// esi = fzbr
|
||||
// edi = fzbc
|
||||
// xmm0 = z/zi
|
||||
// xmm2 = s/si (tme)
|
||||
// xmm3 = t/ti (tme)
|
||||
// xmm4 = q (tme)
|
||||
// xmm2 = u (tme)
|
||||
// xmm3 = v (tme)
|
||||
// xmm5 = rb (!tme)
|
||||
// xmm6 = ga (!tme)
|
||||
// xmm7 = test
|
||||
|
||||
TestZ(xmm1, m_env.sel.tfx != TFX_NONE ? xmm6 : xmm4);
|
||||
bool tme = m_env.sel.tfx != TFX_NONE;
|
||||
|
||||
TestZ(tme ? xmm5 : xmm2, tme ? xmm6 : xmm3);
|
||||
|
||||
// ecx = steps
|
||||
// esi = fzbr
|
||||
// edi = fzbc
|
||||
// ebp = za
|
||||
// - xmm0
|
||||
// xmm2 = s/si (tme)
|
||||
// xmm3 = t/ti (tme)
|
||||
// xmm4 = q (tme)
|
||||
// xmm2 = u (tme)
|
||||
// xmm3 = v (tme)
|
||||
// xmm5 = rb (!tme)
|
||||
// xmm6 = ga (!tme)
|
||||
// xmm7 = test
|
||||
|
|
@ -99,12 +98,12 @@ L("@@");
|
|||
// xmm6 = ga
|
||||
// xmm7 = test
|
||||
|
||||
if(m_env.sel.fb)
|
||||
if(m_env.sel.fwrite)
|
||||
{
|
||||
movdqa(xmm3, xmmword[&m_env.fm]);
|
||||
}
|
||||
|
||||
if(m_env.sel.zb)
|
||||
if(m_env.sel.zwrite)
|
||||
{
|
||||
movdqa(xmm4, xmmword[&m_env.zm]);
|
||||
}
|
||||
|
|
@ -175,51 +174,39 @@ L("@@");
|
|||
// fm |= test;
|
||||
// zm |= test;
|
||||
|
||||
if(m_env.sel.fb || m_env.sel.ztest)
|
||||
if(m_env.sel.fwrite)
|
||||
{
|
||||
por(xmm3, xmm7);
|
||||
}
|
||||
|
||||
if(m_env.sel.zb)
|
||||
if(m_env.sel.zwrite)
|
||||
{
|
||||
por(xmm4, xmm7);
|
||||
}
|
||||
|
||||
// ebx = fa
|
||||
// ecx = steps
|
||||
// esi = fzbr
|
||||
// edi = fzbc
|
||||
// ebp = za
|
||||
// xmm2 = fd
|
||||
// xmm3 = fm
|
||||
// xmm4 = zm
|
||||
// xmm5 = rb
|
||||
// xmm6 = ga
|
||||
|
||||
// int fzm = ~(fm == GSVector4i::xffffffff()).ps32(zm == GSVector4i::xffffffff()).ps32().mask();
|
||||
|
||||
if(m_env.sel.fb && m_env.sel.zb)
|
||||
{
|
||||
pcmpeqd(xmm0, xmm0);
|
||||
movdqa(xmm1, xmm0);
|
||||
pcmpeqd(xmm0, xmm3);
|
||||
pcmpeqd(xmm1, xmm4);
|
||||
}
|
||||
else if(m_env.sel.fb)
|
||||
{
|
||||
pcmpeqd(xmm0, xmm0);
|
||||
pcmpeqd(xmm0, xmm3);
|
||||
pxor(xmm1, xmm1);
|
||||
}
|
||||
else if(m_env.sel.zb)
|
||||
{
|
||||
pxor(xmm0, xmm0);
|
||||
pcmpeqd(xmm1, xmm1);
|
||||
|
||||
if(m_env.sel.fwrite && m_env.sel.zwrite)
|
||||
{
|
||||
movdqa(xmm0, xmm1);
|
||||
pcmpeqd(xmm1, xmm3);
|
||||
pcmpeqd(xmm0, xmm4);
|
||||
packssdw(xmm1, xmm0);
|
||||
}
|
||||
else if(m_env.sel.fwrite)
|
||||
{
|
||||
pcmpeqd(xmm1, xmm3);
|
||||
packssdw(xmm1, xmm1);
|
||||
}
|
||||
else if(m_env.sel.zwrite)
|
||||
{
|
||||
pcmpeqd(xmm1, xmm4);
|
||||
packssdw(xmm1, xmm1);
|
||||
}
|
||||
|
||||
packssdw(xmm0, xmm1);
|
||||
pmovmskb(edx, xmm0);
|
||||
pmovmskb(edx, xmm1);
|
||||
not(edx);
|
||||
|
||||
// ebx = fa
|
||||
|
|
@ -267,13 +254,13 @@ L("step");
|
|||
// if(steps <= 0) break;
|
||||
|
||||
test(ecx, ecx);
|
||||
jle("@f", T_NEAR);
|
||||
jle("exit", T_NEAR);
|
||||
|
||||
Step();
|
||||
|
||||
jmp("@b", T_NEAR);
|
||||
jmp("loop", T_NEAR);
|
||||
|
||||
L("@@");
|
||||
L("exit");
|
||||
|
||||
pop(ebp);
|
||||
pop(edi);
|
||||
|
|
@ -435,6 +422,10 @@ void GSDrawScanlineCodeGenerator::Init(int params)
|
|||
movaps(xmmword[&m_env.temp.s], xmm2);
|
||||
movaps(xmmword[&m_env.temp.t], xmm3);
|
||||
movaps(xmmword[&m_env.temp.q], xmm4);
|
||||
|
||||
rcpps(xmm4, xmm4);
|
||||
mulps(xmm2, xmm4);
|
||||
mulps(xmm3, xmm4);
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -564,6 +555,10 @@ void GSDrawScanlineCodeGenerator::Step()
|
|||
movaps(xmmword[&m_env.temp.s], xmm2);
|
||||
movaps(xmmword[&m_env.temp.t], xmm3);
|
||||
movaps(xmmword[&m_env.temp.q], xmm4);
|
||||
|
||||
rcpps(xmm4, xmm4);
|
||||
mulps(xmm2, xmm4);
|
||||
mulps(xmm3, xmm4);
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -620,12 +615,14 @@ void GSDrawScanlineCodeGenerator::TestZ(const Xmm& temp1, const Xmm& temp2)
|
|||
mov(ebp, dword[esi + 4]);
|
||||
add(ebp, dword[edi + 4]);
|
||||
|
||||
// GSVector4i zs = sprite ? zi : (GSVector4i(z * 0.5f) << 1) | (GSVector4i(z) & GSVector4i::x00000001());
|
||||
// GSVector4i zs = zi;
|
||||
|
||||
if(!m_env.sel.sprite)
|
||||
{
|
||||
if(m_env.sel.zoverflow)
|
||||
{
|
||||
// zs = (GSVector4i(z * 0.5f) << 1) | (GSVector4i(z) & GSVector4i::x00000001());
|
||||
|
||||
static float half = 0.5f;
|
||||
|
||||
movss(temp1, dword[&half]);
|
||||
|
|
@ -643,6 +640,8 @@ void GSDrawScanlineCodeGenerator::TestZ(const Xmm& temp1, const Xmm& temp2)
|
|||
}
|
||||
else
|
||||
{
|
||||
// zs = GSVector4i(z);
|
||||
|
||||
cvttps2dq(xmm0, xmm0);
|
||||
}
|
||||
|
||||
|
|
@ -654,52 +653,52 @@ void GSDrawScanlineCodeGenerator::TestZ(const Xmm& temp1, const Xmm& temp2)
|
|||
|
||||
if(m_env.sel.ztest)
|
||||
{
|
||||
ReadPixel(temp1, ebp);
|
||||
ReadPixel(xmm1, ebp);
|
||||
|
||||
if(m_env.sel.zwrite && m_env.sel.zpsm < 2)
|
||||
{
|
||||
movdqa(xmmword[&m_env.temp.zd], temp1);
|
||||
movdqa(xmmword[&m_env.temp.zd], xmm1);
|
||||
}
|
||||
|
||||
// zd &= 0xffffffff >> m_env.sel.zpsm * 8;
|
||||
|
||||
if(m_env.sel.zpsm)
|
||||
{
|
||||
pslld(temp1, m_env.sel.zpsm * 8);
|
||||
psrld(temp1, m_env.sel.zpsm * 8);
|
||||
pslld(xmm1, m_env.sel.zpsm * 8);
|
||||
psrld(xmm1, m_env.sel.zpsm * 8);
|
||||
}
|
||||
|
||||
if(m_env.sel.zoverflow || m_env.sel.zpsm == 0)
|
||||
{
|
||||
// GSVector4i o = GSVector4i::x80000000();
|
||||
|
||||
pcmpeqd(temp2, temp2);
|
||||
pslld(temp2, 31);
|
||||
pcmpeqd(xmm4, xmm4);
|
||||
pslld(xmm4, 31);
|
||||
|
||||
// GSVector4i zso = zs - o;
|
||||
|
||||
psubd(xmm0, temp2);
|
||||
psubd(xmm0, xmm4);
|
||||
|
||||
// GSVector4i zdo = zd - o;
|
||||
|
||||
psubd(temp1, temp2);
|
||||
psubd(xmm1, xmm4);
|
||||
}
|
||||
|
||||
switch(m_env.sel.ztst)
|
||||
{
|
||||
case ZTST_GEQUAL:
|
||||
// test |= zso < zdo;
|
||||
pcmpgtd(temp1, xmm0);
|
||||
por(xmm7, temp1);
|
||||
pcmpgtd(xmm1, xmm0);
|
||||
por(xmm7, xmm1);
|
||||
break;
|
||||
|
||||
case ZTST_GREATER:
|
||||
// test |= zso <= zdo;
|
||||
movdqa(temp2, temp1);
|
||||
pcmpgtd(temp1, xmm0);
|
||||
por(xmm7, temp1);
|
||||
pcmpeqd(temp2, xmm0);
|
||||
por(xmm7, temp1);
|
||||
movdqa(xmm4, xmm1);
|
||||
pcmpgtd(xmm1, xmm0);
|
||||
por(xmm7, xmm1);
|
||||
pcmpeqd(xmm4, xmm0);
|
||||
por(xmm7, xmm1);
|
||||
break;
|
||||
}
|
||||
|
||||
|
|
@ -717,19 +716,10 @@ void GSDrawScanlineCodeGenerator::SampleTexture()
|
|||
mov(ebx, dword[&m_env.tex]);
|
||||
|
||||
// ebx = tex
|
||||
// edx = clut
|
||||
|
||||
if(!m_env.sel.fst)
|
||||
{
|
||||
// GSVector4 w = q.rcp();
|
||||
|
||||
rcpps(xmm4, xmm4);
|
||||
|
||||
// u = GSVector4i(s * w);
|
||||
// v = GSVector4i(t * w);
|
||||
|
||||
mulps(xmm2, xmm4);
|
||||
mulps(xmm3, xmm4);
|
||||
// TODO: move these into Init/Step too?
|
||||
|
||||
cvttps2dq(xmm2, xmm2);
|
||||
cvttps2dq(xmm3, xmm3);
|
||||
|
|
@ -1000,19 +990,37 @@ void GSDrawScanlineCodeGenerator::Wrap(const Xmm& uv)
|
|||
{
|
||||
// xmm0, xmm1, xmm4, xmm5, xmm6 = free
|
||||
|
||||
if(m_env.sel.wms == m_env.sel.wmt)
|
||||
int wms_clamp = ((m_env.sel.wms + 1) >> 1) & 1;
|
||||
int wmt_clamp = ((m_env.sel.wmt + 1) >> 1) & 1;
|
||||
|
||||
int region = ((m_env.sel.wms | m_env.sel.wmt) >> 1) & 1;
|
||||
|
||||
if(wms_clamp == wmt_clamp)
|
||||
{
|
||||
if(m_env.sel.wms)
|
||||
if(wms_clamp)
|
||||
{
|
||||
if(region)
|
||||
{
|
||||
pmaxsw(uv, xmmword[&m_env.t.min]);
|
||||
}
|
||||
else
|
||||
{
|
||||
pxor(xmm0, xmm0);
|
||||
pmaxsw(uv, xmm0);
|
||||
}
|
||||
|
||||
pminsw(uv, xmmword[&m_env.t.max]);
|
||||
}
|
||||
else
|
||||
{
|
||||
pand(uv, xmmword[&m_env.t.min]);
|
||||
|
||||
if(region)
|
||||
{
|
||||
por(uv, xmmword[&m_env.t.max]);
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
movdqa(xmm1, uv);
|
||||
|
|
@ -1028,7 +1036,11 @@ void GSDrawScanlineCodeGenerator::Wrap(const Xmm& uv)
|
|||
// GSVector4i repeat = (t & m_env.t.min) | m_env.t.max;
|
||||
|
||||
pand(xmm1, xmm4);
|
||||
|
||||
if(region)
|
||||
{
|
||||
por(xmm1, xmm5);
|
||||
}
|
||||
|
||||
// clamp.blend8(repeat, m_env.t.mask);
|
||||
|
||||
|
|
@ -1041,26 +1053,50 @@ void GSDrawScanlineCodeGenerator::Wrap(const Xmm& uv0, const Xmm& uv1)
|
|||
{
|
||||
// xmm0, xmm1, xmm4, xmm5, xmm6 = free
|
||||
|
||||
if(m_env.sel.wms == m_env.sel.wmt)
|
||||
int wms_clamp = ((m_env.sel.wms + 1) >> 1) & 1;
|
||||
int wmt_clamp = ((m_env.sel.wmt + 1) >> 1) & 1;
|
||||
|
||||
int region = ((m_env.sel.wms | m_env.sel.wmt) >> 1) & 1;
|
||||
|
||||
if(wms_clamp == wmt_clamp)
|
||||
{
|
||||
if(wms_clamp)
|
||||
{
|
||||
if(region)
|
||||
{
|
||||
movdqa(xmm4, xmmword[&m_env.t.min]);
|
||||
|
||||
pmaxsw(uv0, xmm4);
|
||||
pmaxsw(uv1, xmm4);
|
||||
}
|
||||
else
|
||||
{
|
||||
pxor(xmm0, xmm0);
|
||||
pmaxsw(uv0, xmm0);
|
||||
pmaxsw(uv1, xmm0);
|
||||
}
|
||||
|
||||
movdqa(xmm5, xmmword[&m_env.t.max]);
|
||||
|
||||
if(m_env.sel.wms)
|
||||
{
|
||||
pmaxsw(uv0, xmm4);
|
||||
pminsw(uv0, xmm5);
|
||||
pmaxsw(uv1, xmm4);
|
||||
pminsw(uv1, xmm5);
|
||||
}
|
||||
else
|
||||
{
|
||||
movdqa(xmm4, xmmword[&m_env.t.min]);
|
||||
|
||||
pand(uv0, xmm4);
|
||||
por(uv0, xmm5);
|
||||
pand(uv1, xmm4);
|
||||
|
||||
if(region)
|
||||
{
|
||||
movdqa(xmm5, xmmword[&m_env.t.max]);
|
||||
|
||||
por(uv0, xmm5);
|
||||
por(uv1, xmm5);
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
movdqa(xmm1, uv0);
|
||||
|
|
@ -1072,16 +1108,20 @@ void GSDrawScanlineCodeGenerator::Wrap(const Xmm& uv0, const Xmm& uv1)
|
|||
// GSVector4i clamp = t.sat_i16(m_env.t.min, m_env.t.max);
|
||||
|
||||
pmaxsw(uv0, xmm4);
|
||||
pminsw(uv0, xmm5);
|
||||
pmaxsw(uv1, xmm4);
|
||||
pminsw(uv0, xmm5);
|
||||
pminsw(uv1, xmm5);
|
||||
|
||||
// GSVector4i repeat = (t & m_env.t.min) | m_env.t.max;
|
||||
|
||||
pand(xmm1, xmm4);
|
||||
por(xmm1, xmm5);
|
||||
pand(xmm6, xmm4);
|
||||
|
||||
if(region)
|
||||
{
|
||||
por(xmm1, xmm5);
|
||||
por(xmm6, xmm5);
|
||||
}
|
||||
|
||||
// clamp.blend8(repeat, m_env.t.mask);
|
||||
|
||||
|
|
@ -1453,6 +1493,8 @@ void GSDrawScanlineCodeGenerator::WriteZBuf()
|
|||
|
||||
movdqa(xmm1, xmmword[!m_env.sel.sprite ? &m_env.temp.zs : &m_env.p.z]);
|
||||
|
||||
bool fast = false;
|
||||
|
||||
if(m_env.sel.ztest && m_env.sel.zpsm < 2)
|
||||
{
|
||||
// zs = zs.blend8(zd, zm);
|
||||
|
|
@ -1461,46 +1503,10 @@ void GSDrawScanlineCodeGenerator::WriteZBuf()
|
|||
movdqa(xmm7, xmmword[&m_env.temp.zd]);
|
||||
blend8(xmm1, xmm7);
|
||||
|
||||
// if(fzm & 0x30) GSVector4i::storel(&vm16[addr + 0], zs);
|
||||
// if(fzm & 0xc0) GSVector4i::storeh(&vm16[addr + 8], zs);
|
||||
|
||||
test(dh, 0x0f);
|
||||
je("wz30");
|
||||
movq(qword[ebp * 2 + (size_t)m_env.vm], xmm1);
|
||||
L("wz30");
|
||||
|
||||
test(dh, 0xf0);
|
||||
je("wzc0");
|
||||
movhps(qword[ebp * 2 + (size_t)m_env.vm + 8 * 2], xmm1);
|
||||
L("wzc0");
|
||||
fast = true;
|
||||
}
|
||||
else
|
||||
{
|
||||
// if(fzm & 0x10) WritePixel(zpsm, &vm16[addr + 0], zs.extract32<0>());
|
||||
// if(fzm & 0x20) WritePixel(zpsm, &vm16[addr + 2], zs.extract32<1>());
|
||||
// if(fzm & 0x40) WritePixel(zpsm, &vm16[addr + 8], zs.extract32<2>());
|
||||
// if(fzm & 0x80) WritePixel(zpsm, &vm16[addr + 10], zs.extract32<3>());
|
||||
|
||||
test(dh, 0x03);
|
||||
je("wz10");
|
||||
WritePixel(xmm1, xmm0, ebp, 0, m_env.sel.zpsm);
|
||||
L("wz10");
|
||||
|
||||
test(dh, 0x0c);
|
||||
je("wz20");
|
||||
WritePixel(xmm1, xmm0, ebp, 1, m_env.sel.zpsm);
|
||||
L("wz20");
|
||||
|
||||
test(dh, 0x30);
|
||||
je("wz40");
|
||||
WritePixel(xmm1, xmm0, ebp, 2, m_env.sel.zpsm);
|
||||
L("wz40");
|
||||
|
||||
test(dh, 0xc0);
|
||||
je("wz80");
|
||||
WritePixel(xmm1, xmm0, ebp, 3, m_env.sel.zpsm);
|
||||
L("wz80");
|
||||
}
|
||||
WritePixel(xmm1, xmm0, ebp, dh, fast, m_env.sel.zpsm);
|
||||
}
|
||||
|
||||
void GSDrawScanlineCodeGenerator::AlphaBlend()
|
||||
|
|
@ -1538,13 +1544,12 @@ void GSDrawScanlineCodeGenerator::AlphaBlend()
|
|||
movdqa(xmm1, xmm2);
|
||||
movdqa(xmm4, xmm2);
|
||||
|
||||
mov(eax, 0x1f);
|
||||
movd(xmm7, eax);
|
||||
pshufd(xmm7, xmm7, _MM_SHUFFLE(0, 0, 0, 0)); // 0x001f
|
||||
pcmpeqd(xmm7, xmm7);
|
||||
psrld(xmm7, 27); // 0x0000001f
|
||||
pand(xmm0, xmm7);
|
||||
pslld(xmm0, 3);
|
||||
|
||||
psllw(xmm7, 10); // 0x7c00
|
||||
pslld(xmm7, 10); // 0x00007c00
|
||||
pand(xmm4, xmm7);
|
||||
pslld(xmm4, 9);
|
||||
|
||||
|
|
@ -1552,11 +1557,11 @@ void GSDrawScanlineCodeGenerator::AlphaBlend()
|
|||
|
||||
movdqa(xmm4, xmm1);
|
||||
|
||||
psrlw(xmm7, 5); // 0x03e0
|
||||
psrld(xmm7, 5); // 0x000003e0
|
||||
pand(xmm1, xmm7);
|
||||
psrld(xmm1, 2);
|
||||
|
||||
psllw(xmm7, 10); // 0x8000
|
||||
psllw(xmm7, 10); // 0x00008000
|
||||
pand(xmm4, xmm7);
|
||||
pslld(xmm4, 8);
|
||||
|
||||
|
|
@ -1598,7 +1603,7 @@ void GSDrawScanlineCodeGenerator::AlphaBlend()
|
|||
|
||||
if(!(m_env.sel.fpsm == 1 && m_env.sel.abec == 1))
|
||||
{
|
||||
// GSVector4i a = abec < 2 ? c[abec * 2 + 1].yywwlh().sll16(7) : m_env.afix2;
|
||||
// GSVector4i a = abec < 2 ? c[abec * 2 + 1].yywwlh().sll16(7) : m_env.afix;
|
||||
|
||||
switch(m_env.sel.abec)
|
||||
{
|
||||
|
|
@ -1728,9 +1733,12 @@ void GSDrawScanlineCodeGenerator::AlphaBlend()
|
|||
blend8r(xmm6, xmm4);
|
||||
}
|
||||
else
|
||||
{
|
||||
if(m_env.sel.fpsm != 1) // TODO: fpsm == 0 && fm == 0xffxxxxxx
|
||||
{
|
||||
mix16(xmm6, xmm4, xmm7);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void GSDrawScanlineCodeGenerator::WriteFrame(int params)
|
||||
|
|
@ -1817,48 +1825,9 @@ void GSDrawScanlineCodeGenerator::WriteFrame(int params)
|
|||
blend(xmm5, xmm2, xmm3); // TODO: could be skipped in certain cases, depending on fpsm and fm
|
||||
}
|
||||
|
||||
if(m_env.sel.rfb && m_env.sel.fpsm < 2)
|
||||
{
|
||||
// if(fzm & 0x03) GSVector4i::storel(&vm16[addr + 0], fs);
|
||||
// if(fzm & 0x0c) GSVector4i::storeh(&vm16[addr + 8], fs);
|
||||
bool fast = m_env.sel.rfb && m_env.sel.fpsm < 2;
|
||||
|
||||
test(dl, 0x0f);
|
||||
je("wf03");
|
||||
movq(qword[ebx * 2 + (size_t)m_env.vm], xmm5);
|
||||
L("wf03");
|
||||
|
||||
test(dl, 0xf0);
|
||||
je("wf0c");
|
||||
movhps(qword[ebx * 2 + (size_t)m_env.vm + 8 * 2], xmm5);
|
||||
L("wf0c");
|
||||
}
|
||||
else
|
||||
{
|
||||
// if(fzm & 0x01) WritePixel(fpsm, &vm16[addr + 0], fs.extract32<0>());
|
||||
// if(fzm & 0x02) WritePixel(fpsm, &vm16[addr + 2], fs.extract32<1>());
|
||||
// if(fzm & 0x04) WritePixel(fpsm, &vm16[addr + 8], fs.extract32<2>());
|
||||
// if(fzm & 0x08) WritePixel(fpsm, &vm16[addr + 10], fs.extract32<3>());
|
||||
|
||||
test(dl, 0x03);
|
||||
je("wf01");
|
||||
WritePixel(xmm5, xmm0, ebx, 0, m_env.sel.fpsm);
|
||||
L("wf01");
|
||||
|
||||
test(dl, 0x0c);
|
||||
je("wf02");
|
||||
WritePixel(xmm5, xmm0, ebx, 1, m_env.sel.fpsm);
|
||||
L("wf02");
|
||||
|
||||
test(dl, 0x30);
|
||||
je("wf04");
|
||||
WritePixel(xmm5, xmm0, ebx, 2, m_env.sel.fpsm);
|
||||
L("wf04");
|
||||
|
||||
test(dl, 0xc0);
|
||||
je("wf08");
|
||||
WritePixel(xmm5, xmm0, ebx, 3, m_env.sel.fpsm);
|
||||
L("wf08");
|
||||
}
|
||||
WritePixel(xmm5, xmm0, ebx, dl, fast, m_env.sel.fpsm);
|
||||
}
|
||||
|
||||
void GSDrawScanlineCodeGenerator::ReadPixel(const Xmm& dst, const Reg32& addr)
|
||||
|
|
@ -1867,6 +1836,52 @@ void GSDrawScanlineCodeGenerator::ReadPixel(const Xmm& dst, const Reg32& addr)
|
|||
movhps(dst, qword[addr * 2 + (size_t)m_env.vm + 8 * 2]);
|
||||
}
|
||||
|
||||
void GSDrawScanlineCodeGenerator::WritePixel(const Xmm& src, const Xmm& temp, const Reg32& addr, const Reg8& mask, bool fast, int psm)
|
||||
{
|
||||
if(fast)
|
||||
{
|
||||
// if(fzm & 0x0f) GSVector4i::storel(&vm16[addr + 0], fs);
|
||||
// if(fzm & 0xf0) GSVector4i::storeh(&vm16[addr + 8], fs);
|
||||
|
||||
test(mask, 0x0f);
|
||||
je("@f");
|
||||
movq(qword[addr * 2 + (size_t)m_env.vm], src);
|
||||
L("@@");
|
||||
|
||||
test(mask, 0xf0);
|
||||
je("@f");
|
||||
movhps(qword[addr * 2 + (size_t)m_env.vm + 8 * 2], src);
|
||||
L("@@");
|
||||
}
|
||||
else
|
||||
{
|
||||
// if(fzm & 0x03) WritePixel(fpsm, &vm16[addr + 0], fs.extract32<0>());
|
||||
// if(fzm & 0x0c) WritePixel(fpsm, &vm16[addr + 2], fs.extract32<1>());
|
||||
// if(fzm & 0x30) WritePixel(fpsm, &vm16[addr + 8], fs.extract32<2>());
|
||||
// if(fzm & 0xc0) WritePixel(fpsm, &vm16[addr + 10], fs.extract32<3>());
|
||||
|
||||
test(mask, 0x03);
|
||||
je("@f");
|
||||
WritePixel(src, temp, addr, 0, psm);
|
||||
L("@@");
|
||||
|
||||
test(mask, 0x0c);
|
||||
je("@f");
|
||||
WritePixel(src, temp, addr, 1, psm);
|
||||
L("@@");
|
||||
|
||||
test(mask, 0x30);
|
||||
je("@f");
|
||||
WritePixel(src, temp, addr, 2, psm);
|
||||
L("@@");
|
||||
|
||||
test(mask, 0xc0);
|
||||
je("@f");
|
||||
WritePixel(src, temp, addr, 3, psm);
|
||||
L("@@");
|
||||
}
|
||||
}
|
||||
|
||||
void GSDrawScanlineCodeGenerator::WritePixel(const Xmm& src, const Xmm& temp, const Reg32& addr, uint8 i, int psm)
|
||||
{
|
||||
static const int offsets[4] = {0, 2, 8, 10};
|
||||
|
|
|
|||
|
|
@ -56,6 +56,7 @@ class GSDrawScanlineCodeGenerator : public CodeGenerator
|
|||
void WriteFrame(int params);
|
||||
|
||||
void ReadPixel(const Xmm& dst, const Reg32& addr);
|
||||
void WritePixel(const Xmm& src, const Xmm& temp, const Reg32& addr, const Reg8& mask, bool fast, int psm);
|
||||
void WritePixel(const Xmm& src, const Xmm& temp, const Reg32& addr, uint8 i, int psm);
|
||||
void ReadTexel(const Xmm& dst, const Xmm& addr, const Xmm& temp1, const Xmm& temp2);
|
||||
void ReadTexel(const Xmm& dst, const Xmm& addr, uint8 i);
|
||||
|
|
|
|||
|
|
@ -19,6 +19,8 @@
|
|||
*
|
||||
*/
|
||||
|
||||
// TODO: skip scissor test when vtrace.p.min/max is inside the region
|
||||
|
||||
#include "StdAfx.h"
|
||||
#include "GSRasterizer.h"
|
||||
|
||||
|
|
@ -39,8 +41,8 @@ void GSRasterizer::Draw(const GSRasterizerData* data)
|
|||
m_dsf.sl = NULL;
|
||||
m_dsf.sr = NULL;
|
||||
m_dsf.sp = NULL;
|
||||
m_dsf.ssl = NULL;
|
||||
m_dsf.ssp = NULL;
|
||||
m_dsf.ssl = NULL;
|
||||
m_dsf.ssp = NULL;
|
||||
|
||||
m_ds->BeginDraw(data, &m_dsf);
|
||||
|
||||
|
|
|
|||
|
|
@ -310,8 +310,8 @@ protected:
|
|||
p.sel.fst = PRIM->FST;
|
||||
p.sel.ltf = context->TEX1.IsLinear();
|
||||
p.sel.tlu = GSLocalMemory::m_psm[context->TEX0.PSM].pal > 0;
|
||||
p.sel.wms = ((context->CLAMP.WMS + 1) >> 1) & 1;
|
||||
p.sel.wmt = ((context->CLAMP.WMT + 1) >> 1) & 1;
|
||||
p.sel.wms = context->CLAMP.WMS;
|
||||
p.sel.wmt = context->CLAMP.WMT;
|
||||
|
||||
if(p.sel.iip == 0 && p.sel.tfx == TFX_MODULATE && p.sel.tcc)
|
||||
{
|
||||
|
|
|
|||
|
|
@ -53,13 +53,13 @@ union GSScanlineSelector
|
|||
DWORD ftest:1; // 33
|
||||
DWORD zwrite:1; // 34
|
||||
DWORD ztest:1; // 35
|
||||
DWORD wms:1; // 36 (0: repeat, 1: clamp)
|
||||
DWORD wmt:1; // 37
|
||||
DWORD datm:1; // 38
|
||||
DWORD colclamp:1; // 39
|
||||
DWORD fba:1; // 40
|
||||
DWORD dthe:1; // 41
|
||||
DWORD zoverflow:1; // 42 (z max >= 0x80000000)
|
||||
DWORD wms:2; // 36
|
||||
DWORD wmt:2; // 38
|
||||
DWORD datm:1; // 40
|
||||
DWORD colclamp:1; // 41
|
||||
DWORD fba:1; // 42
|
||||
DWORD dthe:1; // 43
|
||||
DWORD zoverflow:1; // 44 (z max >= 0x80000000)
|
||||
};
|
||||
|
||||
struct
|
||||
|
|
|
|||
|
|
@ -73,11 +73,34 @@ BOOL xpadApp::InitInstance()
|
|||
|
||||
//
|
||||
|
||||
#define PS2E_LT_PAD 0x02
|
||||
#define PS2E_PAD_VERSION 0x0002
|
||||
static struct
|
||||
{
|
||||
static const UINT32 revision = 0;
|
||||
static const UINT32 build = 2;
|
||||
static const UINT32 minor = 0;
|
||||
} s_ver;
|
||||
|
||||
static bool s_ps2 = false;
|
||||
|
||||
EXPORT_C_(UINT32) PSEgetLibType()
|
||||
{
|
||||
return PSE_LT_PAD;
|
||||
}
|
||||
|
||||
EXPORT_C_(char*) PSEgetLibName()
|
||||
{
|
||||
return "XPad";
|
||||
}
|
||||
|
||||
EXPORT_C_(UINT32) PSEgetLibVersion()
|
||||
{
|
||||
return (s_ver.minor << 16) | (s_ver.revision << 8) | s_ver.build;
|
||||
}
|
||||
|
||||
EXPORT_C_(UINT32) PS2EgetLibType()
|
||||
{
|
||||
s_ps2 = true;
|
||||
|
||||
return PS2E_LT_PAD;
|
||||
}
|
||||
|
||||
|
|
@ -88,11 +111,7 @@ EXPORT_C_(char*) PS2EgetLibName()
|
|||
|
||||
EXPORT_C_(UINT32) PS2EgetLibVersion2(UINT32 type)
|
||||
{
|
||||
const UINT32 revision = 0;
|
||||
const UINT32 build = 1;
|
||||
const UINT32 minor = 0;
|
||||
|
||||
return (build << 0) | (revision << 8) | (PS2E_PAD_VERSION << 16) | (minor << 24);
|
||||
return (s_ver.build << 0) | (s_ver.revision << 8) | (PS2E_PAD_VERSION << 16) | (s_ver.minor << 24);
|
||||
}
|
||||
|
||||
//
|
||||
|
|
@ -159,7 +178,7 @@ public:
|
|||
: m_pad(pad)
|
||||
, m_connected(false)
|
||||
, m_ds2native(false)
|
||||
, m_analog(false)
|
||||
, m_analog(!s_ps2) // defaults to analog off for ps2
|
||||
, m_locked(false)
|
||||
, m_vibration(true)
|
||||
, m_small(0)
|
||||
|
|
@ -523,19 +542,7 @@ public:
|
|||
ret = (this->*m_handler)(m_index - 3, value);
|
||||
break;
|
||||
}
|
||||
/*
|
||||
//if(m_cmd != 'B')
|
||||
{
|
||||
static FILE* log = NULL;
|
||||
if(log == NULL)
|
||||
{
|
||||
log = fopen("c:\\log.txt", "w");
|
||||
}
|
||||
if(m_index == 1) fprintf(log, "\n");
|
||||
fprintf(log, "*** %02x %d %02x %02x\n", m_cmd, m_index - 1, value, ret);
|
||||
fflush(log);
|
||||
}
|
||||
*/
|
||||
|
||||
return ret;
|
||||
}
|
||||
|
||||
|
|
@ -587,10 +594,13 @@ EXPORT_C_(UINT32) PADopen(void* pDsp)
|
|||
XInputEnable(TRUE);
|
||||
|
||||
if(s_nRefs == 0)
|
||||
{
|
||||
if(s_ps2)
|
||||
{
|
||||
s_hWnd = *(HWND*)pDsp;
|
||||
s_GSWndProc = (WNDPROC)SetWindowLongPtr(s_hWnd, GWLP_WNDPROC, (LPARAM)PADwndProc);
|
||||
}
|
||||
}
|
||||
|
||||
s_nRefs++;
|
||||
|
||||
|
|
@ -602,9 +612,12 @@ EXPORT_C PADclose()
|
|||
s_nRefs--;
|
||||
|
||||
if(s_nRefs == 0)
|
||||
{
|
||||
if(s_ps2)
|
||||
{
|
||||
SetWindowLongPtr(s_hWnd, GWLP_WNDPROC, (LPARAM)s_GSWndProc);
|
||||
}
|
||||
}
|
||||
|
||||
XInputEnable(FALSE);
|
||||
}
|
||||
|
|
@ -626,6 +639,31 @@ EXPORT_C_(BYTE) PADpoll(BYTE value)
|
|||
return s_padplugin.Poll(value);
|
||||
}
|
||||
|
||||
EXPORT_C_(UINT32) PADreadPort(int port, PadDataS* data)
|
||||
{
|
||||
PADstartPoll(port + 1);
|
||||
|
||||
data->type = PADpoll('B') >> 4;
|
||||
PADpoll(0);
|
||||
data->status = PADpoll(0) | (PADpoll(0) << 8);
|
||||
data->rightJoyX = data->moveX = PADpoll(0);
|
||||
data->rightJoyY = data->moveY = PADpoll(0);
|
||||
data->leftJoyX = PADpoll(0);
|
||||
data->leftJoyY = PADpoll(0);
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
EXPORT_C_(UINT32) PADreadPort1(PadDataS* ppds)
|
||||
{
|
||||
return PADreadPort(0, ppds);
|
||||
}
|
||||
|
||||
EXPORT_C_(UINT32) PADreadPort2(PadDataS* ppds)
|
||||
{
|
||||
return PADreadPort(1, ppds);
|
||||
}
|
||||
|
||||
EXPORT_C_(KeyEvent*) PADkeyEvent()
|
||||
{
|
||||
static KeyEvent event;
|
||||
|
|
|
|||
|
|
@ -4,6 +4,9 @@ LIBRARY "xpad"
|
|||
|
||||
EXPORTS
|
||||
; Explicit exports can go here
|
||||
PSEgetLibType
|
||||
PSEgetLibName
|
||||
PSEgetLibVersion
|
||||
PS2EgetLibType
|
||||
PS2EgetLibName
|
||||
PS2EgetLibVersion2
|
||||
|
|
@ -18,3 +21,6 @@ EXPORTS
|
|||
PADconfigure
|
||||
PADtest
|
||||
PADabout
|
||||
PADreadPort
|
||||
PADreadPort1
|
||||
PADreadPort2
|
||||
|
|
|
|||
|
|
@ -36,6 +36,49 @@ public:
|
|||
DECLARE_MESSAGE_MAP()
|
||||
};
|
||||
|
||||
// ps1
|
||||
|
||||
#define PSE_LT_PAD 8
|
||||
|
||||
// MOUSE SCPH-1030
|
||||
#define PSE_PAD_TYPE_MOUSE 1
|
||||
// NEGCON - 16 button analog controller SLPH-00001
|
||||
#define PSE_PAD_TYPE_NEGCON 2
|
||||
// GUN CONTROLLER - gun controller SLPH-00014 from Konami
|
||||
#define PSE_PAD_TYPE_GUN 3
|
||||
// STANDARD PAD SCPH-1080, SCPH-1150
|
||||
#define PSE_PAD_TYPE_STANDARD 4
|
||||
// ANALOG JOYSTICK SCPH-1110
|
||||
#define PSE_PAD_TYPE_ANALOGJOY 5
|
||||
// GUNCON - gun controller SLPH-00034 from Namco
|
||||
#define PSE_PAD_TYPE_GUNCON 6
|
||||
// ANALOG CONTROLLER SCPH-1150
|
||||
#define PSE_PAD_TYPE_ANALOGPAD 7
|
||||
|
||||
struct PadDataS
|
||||
{
|
||||
// controler type - fill it withe predefined values above
|
||||
BYTE type;
|
||||
|
||||
// status of buttons - every controller fills this field
|
||||
WORD status;
|
||||
|
||||
// for analog pad fill those next 4 bytes
|
||||
// values are analog in range 0-255 where 128 is center position
|
||||
BYTE rightJoyX, rightJoyY, leftJoyX, leftJoyY;
|
||||
|
||||
// for mouse fill those next 2 bytes
|
||||
// values are in range -128 - 127
|
||||
BYTE moveX, moveY;
|
||||
|
||||
BYTE reserved[91];
|
||||
} ;
|
||||
|
||||
// ps2
|
||||
|
||||
#define PS2E_LT_PAD 0x02
|
||||
#define PS2E_PAD_VERSION 0x0002
|
||||
|
||||
struct KeyEvent
|
||||
{
|
||||
UINT32 key;
|
||||
|
|
@ -44,4 +87,3 @@ struct KeyEvent
|
|||
|
||||
#define KEYPRESS 1
|
||||
#define KEYRELEASE 2
|
||||
|
||||
|
|
|
|||
|
|
@ -788,7 +788,6 @@
|
|||
</Configuration>
|
||||
<Configuration
|
||||
Name="Debug SSE4|Win32"
|
||||
IntermediateDirectory="$(ConfigurationName)"
|
||||
ConfigurationType="2"
|
||||
InheritedPropertySheets=".\vsprops\common.vsprops;.\vsprops\debug.vsprops;.\vsprops\sse2.vsprops;.\vsprops\sse4.vsprops"
|
||||
UseOfMFC="1"
|
||||
|
|
|
|||
Loading…
Reference in New Issue