Skip to content

Commit

Permalink
GS:SW: Use unaligned loads to reduce constant size on AVX2
Browse files Browse the repository at this point in the history 
Allows more instructions to use 1-byte offsets
  • Loading branch information
@TellowKrinkle
TellowKrinkle committed Oct 7, 2024
1 parent 0d434d6 commit eb0b341
Show file tree
Hide file tree
Showing 5 changed files with 100 additions and 78 deletions.
30 changes: 16 additions & 14 deletions pcsx2/GS/Renderers/SW/GSDrawScanline.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -207,10 +207,11 @@ void GSDrawScanline::CSetupPrim(const GSVertexSW* vertex, const u16* index, cons
constexpr int vlen = sizeof(VectorF) / sizeof(float);

#if _M_SSE >= 0x501
const GSVector8* shift = (GSVector8*)g_const_256b.m_shift;
const GSVector4 step_shift = GSVector4::broadcast32(&shift[0]);
auto load_shift = [](int i) { return GSVector8::load<false>(&g_const_256b.m_shift[8 - i]); };
const GSVector4 step_shift = GSVector4::broadcast32(&g_const_256b.m_shift[0]);
#else
const GSVector4* shift = (GSVector4*)g_const_128b.m_shift;
static const GSVector4* shift = reinterpret_cast<const GSVector4*>(g_const_128b.m_shift);
auto load_shift = [](int i) { return shift[1 + i]; };
const GSVector4 step_shift = shift[0];
#endif

Expand All @@ -234,22 +235,23 @@ void GSDrawScanline::CSetupPrim(const GSVertexSW* vertex, const u16* index, cons

for (int i = 0; i < vlen; i++)
{
local.d[i].f = VectorI(df * shift[1 + i]).xxzzlh();
local.d[i].f = VectorI(df * load_shift(i)).xxzzlh();
}
}

if (has_z && !sel.zequal)
{
const GSVector4 dz = GSVector4::broadcast64(&dscan.p.z);
const VectorF dzf(static_cast<float>(dscan.p.F64[1]));
#if _M_SSE >= 0x501
GSVector4::storel(&local.d8.p.z, dz.mul64(GSVector4::f32to64(shift)));
double dz = dscan.p.F64[1] * g_const_256b.m_shift[0];
memcpy(&local.d8.p.z, &dz, sizeof(dz));
#else
const GSVector4 dz = GSVector4::broadcast64(&dscan.p.z);
local.d4.z = dz.mul64(GSVector4::f32to64(shift));
#endif
for (int i = 0; i < vlen; i++)
{
local.d[i].z = dzf * shift[i + 1];
local.d[i].z = dzf * load_shift(i);
}
}
}
Expand Down Expand Up @@ -297,7 +299,7 @@ void GSDrawScanline::CSetupPrim(const GSVertexSW* vertex, const u16* index, cons

for (int i = 0; i < vlen; i++)
{
VectorF v = dstq * shift[1 + i];
VectorF v = dstq * load_shift(i);

if (sel.fst)
{
Expand Down Expand Up @@ -336,8 +338,8 @@ void GSDrawScanline::CSetupPrim(const GSVertexSW* vertex, const u16* index, cons

for (int i = 0; i < vlen; i++)
{
VectorI r = VectorI(dr * shift[1 + i]).ps32();
VectorI b = VectorI(db * shift[1 + i]).ps32();
VectorI r = VectorI(dr * load_shift(i)).ps32();
VectorI b = VectorI(db * load_shift(i)).ps32();

local.d[i].rb = r.upl16(b);
}
Expand All @@ -347,8 +349,8 @@ void GSDrawScanline::CSetupPrim(const GSVertexSW* vertex, const u16* index, cons

for (int i = 0; i < vlen; i++)
{
VectorI g = VectorI(dg * shift[1 + i]).ps32();
VectorI a = VectorI(da * shift[1 + i]).ps32();
VectorI g = VectorI(dg * load_shift(i)).ps32();
VectorI a = VectorI(da * load_shift(i)).ps32();

local.d[i].ga = g.upl16(a);
}
Expand Down Expand Up @@ -515,7 +517,7 @@ __ri void GSDrawScanline::CDrawScanline(int pixels, int left, int top, const GSV
steps = pixels + skip - vlen;
left -= skip;
#if _M_SSE >= 0x501
test = GSVector8i::i8to32(g_const_256b.m_test[skip]) | GSVector8i::i8to32(g_const_256b.m_test[15 + (steps & (steps >> 31))]);
test = GSVector8i::i8to32(&g_const_256b.m_test[16 - skip]) | GSVector8i::i8to32(&g_const_256b.m_test[0 - (steps & (steps >> 31))]);
#else
test = const_test[skip] | const_test[7 + (steps & (steps >> 31))];
#endif
Expand Down Expand Up @@ -1756,7 +1758,7 @@ __ri void GSDrawScanline::CDrawScanline(int pixels, int left, int top, const GSV
if (!sel.notest)
{
#if _M_SSE >= 0x501
test = GSVector8i::i8to32(g_const_256b.m_test[15 + (steps & (steps >> 31))]);
test = GSVector8i::i8to32(&g_const_256b.m_test[0 - (steps & (steps >> 31))]);
#else
test = const_test[7 + (steps & (steps >> 31))];
#endif
Expand Down
80 changes: 56 additions & 24 deletions pcsx2/GS/Renderers/SW/GSDrawScanlineCodeGenerator.all.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -661,25 +661,29 @@ void GSDrawScanlineCodeGenerator::Init()

lea(a0.cvt32(), ptr[a0 + a1 - vecints]);

// GSVector4i test = m_test[skip] | m_test[7 + (steps & (steps >> 31))];

mov(eax, a0.cvt32());
sar(eax, 31); // GH: 31 to extract the sign of the register
and_(eax, a0.cvt32());
if (isXmm)
shl(eax, 4); // * sizeof(m_test[0])
cdqe();

if (isXmm)
{
// GSVector4i test = m_test[skip] | m_test[7 + (steps & (steps >> 31))];
mov(eax, a0.cvt32());
sar(eax, 31); // GH: 31 to extract the sign of the register
and_(eax, a0.cvt32());
shl(eax, 4); // * sizeof(m_test[0])
cdqe();
shl(a1.cvt32(), 4); // * sizeof(m_test[0])
movdqa(_test, ptr[a1 + _m_const + offsetof(GSScanlineConstantData128B, m_test[0])]);
por(_test, ptr[rax + _m_const + offsetof(GSScanlineConstantData128B, m_test[7])]);
}
else
{
pmovsxbd(_test, ptr[a1 * 8 + _m_const + offsetof(GSScanlineConstantData256B, m_test[0])]);
pmovsxbd(xym0, ptr[rax * 8 + _m_const + offsetof(GSScanlineConstantData256B, m_test[15])]);
// GSVector8i test = loadu(&m_test[16 - skip]) | loadu(&m_test[steps >= 0 ? 0 : -steps]);
mov(eax, a1.cvt32());
neg(rax); // rax = -skip
pmovsxbd(_test, ptr[rax + _m_const + offsetof(GSScanlineConstantData256B, m_test[16])]);
xor_(t0.cvt32(), t0.cvt32());
mov(eax, a0.cvt32());
neg(eax); // eax = -steps
cmovs(eax, t0.cvt32()); // if (eax < 0) eax = 0
pmovsxbd(xym0, ptr[rax + _m_const + offsetof(GSScanlineConstantData256B, m_test[0])]);
por(_test, xym0);
shl(a1.cvt32(), 5); // * sizeof(m_test[0])
}
Expand Down Expand Up @@ -922,7 +926,7 @@ void GSDrawScanlineCodeGenerator::Init()
/// Inputs: a0=steps, t0=fza_offset
/// Outputs[x86]: xym0=z xym2=s, xym3=t, xym4=q, xym5=rb, xym6=ga, xym7=test
/// Destroys[x86]: all
/// Destroys[x64]: xym0, xym1, xym2, xym3
/// Destroys[x64]: xym0, xym1, xym2, xym3, t2
void GSDrawScanlineCodeGenerator::Step()
{
// steps -= 4;
Expand Down Expand Up @@ -1048,19 +1052,22 @@ void GSDrawScanlineCodeGenerator::Step()

if (!m_sel.notest)
{
#if USING_XMM
// test = m_test[7 + (steps & (steps >> 31))];

mov(eax, a0.cvt32());
sar(eax, 31); // GH: 31 to extract the sign of the register
and_(eax, a0.cvt32());
if (isXmm)
shl(eax, 4);
shl(eax, 4);
cdqe();

#if USING_XMM
movdqa(_test, ptr[rax + _m_const + offsetof(GSScanlineConstantData128B, m_test[7])]);
#else
pmovsxbd(_test, ptr[rax * 8 + _m_const + offsetof(GSScanlineConstantData256B, m_test[15])]);
// test = loadu(&m_test[steps >= 0 ? 0 : -steps]);
xor_(t2.cvt32(), t2.cvt32());
mov(eax, a0.cvt32());
neg(eax); // eax = -steps
cmovs(eax, t2.cvt32()); // if (eax < 0) eax = 0;
pmovsxbd(_test, ptr[rax + _m_const + offsetof(GSScanlineConstantData256B, m_test[0])]);
#endif
}
}
Expand Down Expand Up @@ -1655,29 +1662,54 @@ void GSDrawScanlineCodeGenerator::SampleTextureLOD()
pslld(xym4, 9);
psrld(xym4, 9);

auto log2_coeff = [this](int i) -> Address
#if USING_YMM
auto load_log2_coeff = [this](const XYm& reg, int i)
{
ptr[_m_const + log2_coeff_offset(i)];
vbroadcastss(reg, ptr[_m_const + log2_coeff_offset(i)]);
};
auto log2_coeff = [this, &load_log2_coeff](int i)
{
load_log2_coeff(xym6, i);
return xym6;
};
#else
auto log2_coeff = [this](int i) -> Operand
{
return ptr[_m_const + log2_coeff_offset(i)];
};
auto load_log2_coeff = [this, &log2_coeff](const XYm& reg, int i)
{
movaps(reg, log2_coeff(i));
};
#endif

orps(xym4, log2_coeff(3));
load_log2_coeff(xym1, 3);
orps(xym4, xym1);

// xym4 = mant(q) | 1.0f

if (hasFMA)
{
movaps(xym5, log2_coeff(0)); // c0
load_log2_coeff(xym5, 0); // c0
vfmadd213ps(xym5, xym4, log2_coeff(1)); // c0 * xym4 + c1
vfmadd213ps(xym5, xym4, log2_coeff(2)); // (c0 * xym4 + c1) * xym4 + c2
subps(xym4, log2_coeff(3)); // xym4 - 1.0f
subps(xym4, xym1); // xym4 - 1.0f
vfmadd213ps(xym4, xym5, xym0); // ((c0 * xym4 + c1) * xym4 + c2) * (xym4 - 1.0f) + xym0
}
else
{
THREEARG(mulps, xym5, xym4, log2_coeff(0));
if (hasAVX)
{
vmulps(xym5, xym4, log2_coeff(0));
}
else
{
load_log2_coeff(xym5, 0);
mulps(xym5, xym4);
}
addps(xym5, log2_coeff(1));
mulps(xym5, xym4);
subps(xym4, log2_coeff(3));
subps(xym4, xym1);
addps(xym5, log2_coeff(2));
mulps(xym4, xym5);
addps(xym4, xym0);
Expand Down
4 changes: 4 additions & 0 deletions pcsx2/GS/Renderers/SW/GSNewCodeGenerator.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -204,10 +204,12 @@ class GSNewCodeGenerator
FORWARD_OO_OI(or_)
FORWARD_OO_OI(sub)
FORWARD_OO_OI(xor_)
FORWARD(2, BASE, cmovs, const Reg&, const Operand&)
FORWARD(2, BASE, lea, const Reg&, const Address&)
FORWARD(2, BASE, mov, const Operand&, size_t)
FORWARD(2, BASE, mov, ARGS_OO)
FORWARD(2, BASE, movzx, const Reg&, const Operand&)
FORWARD(1, BASE, neg, const Operand&)
FORWARD(1, BASE, not_, const Operand&)
FORWARD(1, BASE, pop, const Operand&)
FORWARD(1, BASE, push, const Operand&)
Expand Down Expand Up @@ -243,6 +245,8 @@ class GSNewCodeGenerator
AFORWARD(2, minps, ARGS_XO)
SFORWARD(2, movaps, ARGS_XO)
SFORWARD(2, movaps, const Address&, const Xmm&)
SFORWARD(2, movups, ARGS_XO)
SFORWARD(2, movups, const Address&, const Xmm&)
SFORWARD(2, movd, const Address&, const Xmm&)
SFORWARD(2, movd, const Reg32&, const Xmm&)
SFORWARD(2, movd, const Xmm&, const Address&)
Expand Down
43 changes: 11 additions & 32 deletions pcsx2/GS/Renderers/SW/GSScanlineEnvironment.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -256,46 +256,25 @@ namespace GSScanlineConstantData
// Constant shared by all threads (to reduce cache miss)
struct alignas(64) GSScanlineConstantData256B
{
alignas(32) u8 m_test[16][8] = {
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},
{0xff, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},
{0xff, 0xff, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},
{0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, 0x00},
{0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00},
{0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00},
{0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00},
{0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00},
{0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff},
{0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff},
{0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff},
{0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff},
{0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff},
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xff},
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff},
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},
// All AVX processors support unaligned access with little to no penalty as long as you don't cross a cache line.
// Take advantage of that to store single vectors that we index with single-element alignment
alignas(32) u8 m_test[24] = {
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
};
alignas(32) float m_shift[9][8] = {
{ 8.0f , 8.0f , 8.0f , 8.0f , 8.0f , 8.0f , 8.0f , 8.0f},
{ 0.0f , 1.0f , 2.0f , 3.0f , 4.0f , 5.0f , 6.0f , 7.0f},
{ -1.0f , 0.0f , 1.0f , 2.0f , 3.0f , 4.0f , 5.0f , 6.0f},
{ -2.0f , -1.0f , 0.0f , 1.0f , 2.0f , 3.0f , 4.0f , 5.0f},
{ -3.0f , -2.0f , -1.0f , 0.0f , 1.0f , 2.0f , 3.0f , 4.0f},
{ -4.0f , -3.0f , -2.0f , -1.0f , 0.0f , 1.0f , 2.0f , 3.0f},
{ -5.0f , -4.0f , -3.0f , -2.0f , -1.0f , 0.0f , 1.0f , 2.0f},
{ -6.0f , -5.0f , -4.0f , -3.0f , -2.0f , -1.0f , 0.0f , 1.0f},
{ -7.0f , -6.0f , -5.0f , -4.0f , -3.0f , -2.0f , -1.0f , 0.0f},
float m_log2_coef[4] = {};
alignas(64) float m_shift[16] = {
8.0f, -7.0f, -6.0f, -5.0f, -4.0f, -3.0f, -2.0f, -1.0f,
0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f,
};
alignas(32) float m_log2_coef[4][8] = {};

constexpr GSScanlineConstantData256B()
{
using namespace GSScanlineConstantData;
for (size_t n = 0; n < std::size(log2_coef); ++n)
{
for (size_t i = 0; i < 8; ++i)
{
m_log2_coef[n][i] = log2_coef[n];
}
m_log2_coef[n] = log2_coef[n];
}
}
};
Expand Down
21 changes: 13 additions & 8 deletions pcsx2/GS/Renderers/SW/GSSetupPrimCodeGenerator.all.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -110,7 +110,12 @@ void GSSetupPrimCodeGenerator::Generate()

for (int i = 0; i < (m_sel.notest ? 2 : many_regs ? 9 : 5); i++)
{
movaps(XYm(3 + i), ptr[rax + i * vecsize]);
if (isXmm)
movaps(XYm(3 + i), ptr[rax + i * vecsize]);
else if (i == 0)
vbroadcastss(xym3, ptr[rax]);
else
movups(XYm(3 + i), ptr[rax + (9 - i) * sizeof(float)]);
}
}

Expand Down Expand Up @@ -253,7 +258,7 @@ void GSSetupPrimCodeGenerator::Depth_YMM()
if (i < 4 || many_regs)
vmulps(ymm0, Ymm(4 + i), ymm1);
else
vmulps(ymm0, ymm1, ptr[g_const_256b.m_shift[i + 1]]);
vmulps(ymm0, ymm1, ptr[&g_const_256b.m_shift[8 - i]]);
cvttps2dq(ymm0, ymm0);
pshuflw(ymm0, ymm0, _MM_SHUFFLE(2, 2, 0, 0));
pshufhw(ymm0, ymm0, _MM_SHUFFLE(2, 2, 0, 0));
Expand Down Expand Up @@ -281,7 +286,7 @@ void GSSetupPrimCodeGenerator::Depth_YMM()
if (i < 4 || many_regs)
vmulps(ymm1, Ymm(4 + i), ymm0);
else
vmulps(ymm1, ymm0, ptr[g_const_256b.m_shift[i + 1]]);
vmulps(ymm1, ymm0, ptr[&g_const_256b.m_shift[8 - i]]);
movaps(_rip_local_di(i, z), ymm1);
}
}
Expand Down Expand Up @@ -356,7 +361,7 @@ void GSSetupPrimCodeGenerator::Texture()
if (i < 4 || many_regs)
THREEARG(mulps, xym2, XYm(4 + i), xym1);
else
vmulps(ymm2, ymm1, ptr[g_const_256b.m_shift[i + 1]]);
vmulps(ymm2, ymm1, ptr[&g_const_256b.m_shift[8 - i]]);

if (m_sel.fst)
{
Expand Down Expand Up @@ -424,7 +429,7 @@ void GSSetupPrimCodeGenerator::Color()
if (i < 4 || many_regs)
THREEARG(mulps, xym0, XYm(4 + i), xym2);
else
vmulps(ymm0, ymm2, ptr[g_const_256b.m_shift[i + 1]]);
vmulps(ymm0, ymm2, ptr[&g_const_256b.m_shift[8 - i]]);
cvttps2dq(xym0, xym0);
packssdw(xym0, xym0);

Expand All @@ -433,7 +438,7 @@ void GSSetupPrimCodeGenerator::Color()
if (i < 4 || many_regs)
THREEARG(mulps, xym1, XYm(4 + i), xym3);
else
vmulps(ymm1, ymm3, ptr[g_const_256b.m_shift[i + 1]]);
vmulps(ymm1, ymm3, ptr[&g_const_256b.m_shift[8 - i]]);
cvttps2dq(xym1, xym1);
packssdw(xym1, xym1);

Expand All @@ -460,7 +465,7 @@ void GSSetupPrimCodeGenerator::Color()
if (i < 4 || many_regs)
THREEARG(mulps, xym0, XYm(4 + i), xym2);
else
vmulps(ymm0, ymm2, ptr[g_const_256b.m_shift[i + 1]]);
vmulps(ymm0, ymm2, ptr[&g_const_256b.m_shift[8 - i]]);
cvttps2dq(xym0, xym0);
packssdw(xym0, xym0);

Expand All @@ -469,7 +474,7 @@ void GSSetupPrimCodeGenerator::Color()
if (i < 4 || many_regs)
THREEARG(mulps, xym1, XYm(4 + i), xym3);
else
vmulps(ymm1, ymm3, ptr[g_const_256b.m_shift[i + 1]]);
vmulps(ymm1, ymm3, ptr[&g_const_256b.m_shift[8 - i]]);
cvttps2dq(xym1, xym1);
packssdw(xym1, xym1);

Expand Down

0 comments on commit eb0b341

Please sign in to comment.