gsdx sw x64: add AVX2 implementation for VS

FS was copied from 32 bits (require massive update)
This commit is contained in:
Gregory Hainaut 2016-11-22 19:00:06 +01:00
parent 15220c386a
commit 608bb5ccb2
9 changed files with 3469 additions and 33 deletions

View File

@ -65,11 +65,12 @@ set(GSdxSources
GSDrawingContext.cpp
GSDrawScanline.cpp
GSDrawScanlineCodeGenerator.cpp
GSDrawScanlineCodeGenerator.x64.cpp
GSDrawScanlineCodeGenerator.x64.avx.cpp
GSDrawScanlineCodeGenerator.x64.avx2.cpp
GSDrawScanlineCodeGenerator.x86.cpp
GSDrawScanlineCodeGenerator.x86.avx.cpp
GSDrawScanlineCodeGenerator.x86.avx2.cpp
GSDrawScanlineCodeGenerator.x64.cpp
GSDrawScanlineCodeGenerator.x86.cpp
GSDrawScanlineCodeGenerator.x64.avx.cpp
GSDump.cpp
GSFunctionMap.cpp
GSHwHack.cpp
@ -87,11 +88,12 @@ set(GSdxSources
GSRendererSW.cpp
GSSetting.cpp
GSSetupPrimCodeGenerator.cpp
GSSetupPrimCodeGenerator.x64.cpp
GSSetupPrimCodeGenerator.x64.avx.cpp
GSSetupPrimCodeGenerator.x64.avx2.cpp
GSSetupPrimCodeGenerator.x86.cpp
GSSetupPrimCodeGenerator.x86.avx.cpp
GSSetupPrimCodeGenerator.x86.avx2.cpp
GSSetupPrimCodeGenerator.x64.avx.cpp
GSSetupPrimCodeGenerator.x86.cpp
GSSetupPrimCodeGenerator.x64.cpp
GSShaderOGL.cpp
GSState.cpp
GSTables.cpp

File diff suppressed because it is too large Load Diff

View File

@ -31,7 +31,7 @@ static const int _v = _args + 8;
void GSDrawScanlineCodeGenerator::Generate()
{
//ret(8);
//ret(8);
push(ebx);
push(esi);

View File

@ -77,19 +77,15 @@ GSSetupPrimCodeGenerator::GSSetupPrimCodeGenerator(void* param, uint64 key, void
m_en.c = m_sel.fb && !(m_sel.tfx == TFX_DECAL && m_sel.tcc) ? 1 : 0;
try {
Generate();
} catch (std::exception& e) {
fprintf(stderr, "ERR:GSSetupPrimCodeGenerator %s\n", e.what());
}
}
#if _M_SSE >= 0x501
Generate_AVX2();
#else
void GSSetupPrimCodeGenerator::Generate()
{
if(g_cpu.has(util::Cpu::tAVX))
Generate_AVX();
else
Generate_SSE();
}
#endif
} catch (std::exception& e) {
fprintf(stderr, "ERR:GSSetupPrimCodeGenerator %s\n", e.what());
}
}

View File

@ -35,8 +35,6 @@ class GSSetupPrimCodeGenerator : public GSCodeGenerator
struct {uint32 z:1, f:1, t:1, c:1;} m_en;
void Generate();
#if _M_SSE < 0x501
void Generate_SSE();
void Depth_SSE();
@ -48,9 +46,10 @@ class GSSetupPrimCodeGenerator : public GSCodeGenerator
void Texture_AVX();
void Color_AVX();
#else
void Depth();
void Texture();
void Color();
void Generate_AVX2();
void Depth_AVX2();
void Texture_AVX2();
void Color_AVX2();
#endif
public:

View File

@ -0,0 +1,376 @@
/*
* Copyright (C) 2016-2016 Gregory
* Copyright (C) 2007-2009 Gabest
* http://www.gabest.org
*
* This Program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2, or (at your option)
* any later version.
*
* This Program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with GNU Make; see the file COPYING. If not, write to
* the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA USA.
* http://www.gnu.org/copyleft/gpl.html
*
*/
#include "stdafx.h"
#include "GSSetupPrimCodeGenerator.h"
#include "GSVertexSW.h"
#if _M_SSE >= 0x501 && (defined(_M_AMD64) || defined(_WIN64))
using namespace Xbyak;
#define _rip_local(field) (m_rip ? ptr[rip + &m_local.field] : ptr[t0 + offsetof(GSScanlineLocalData, field)])
#define _rip_local_v(field, offset) (m_rip ? ptr[rip + &m_local.field] : ptr[t0 + offset])
#define _m_shift(i) (Ymm(7+i))
// FIXME windows ?
#define _vertex rcx
void GSSetupPrimCodeGenerator::Generate_AVX2()
{
// Technically we just need the delta < 2GB
m_rip = (size_t)&m_local < 0x80000000 && (size_t)getCurr() < 0x80000000;
#ifdef _WIN64
sub(rsp, 8 + 2 * 16);
vmovdqa(ptr[rsp + 0], ymm6);
vmovdqa(ptr[rsp + 16], ymm7);
#endif
if (!m_rip)
mov(t0, (size_t)&m_local);
if((m_en.z || m_en.f) && m_sel.prim != GS_SPRITE_CLASS || m_en.t || m_en.c && m_sel.iip)
{
mov(rax, (size_t)&m_shift[0]);
for(int i = 0; i < (m_sel.notest ? 2 : 9); i++)
{
vmovaps(_m_shift(i), ptr[rax + i * 32]);
}
}
// ymm7 to ymm 15 = m_shift[i]
Depth_AVX2();
Texture_AVX2();
Color_AVX2();
#ifdef _WIN64
vmovdqa(ymm6, ptr[rsp + 0]);
vmovdqa(ymm7, ptr[rsp + 16]);
add(rsp, 8 + 2 * 16);
#endif
ret();
}
void GSSetupPrimCodeGenerator::Depth_AVX2()
{
if(!m_en.z && !m_en.f)
{
return;
}
if(m_sel.prim != GS_SPRITE_CLASS)
{
const Ymm& dscan_p = ymm6;
// GSVector4 dp8 = dscan.p * GSVector4::broadcast32(&shift[0]);
vbroadcastf128(dscan_p, ptr[a2 + offsetof(GSVertexSW, p)]);
vmulps(ymm1, dscan_p, _m_shift(0));
if(m_en.z)
{
// m_local.d8.p.z = dp8.extract32<2>();
vextractps(_rip_local(d8.p.z), xmm1, 2);
// GSVector8 dz = GSVector8(dscan.p).zzzz();
vshufps(ymm2, dscan_p, dscan_p, _MM_SHUFFLE(2, 2, 2, 2));
for(int i = 0; i < (m_sel.notest ? 1 : 8); i++)
{
// m_local.d[i].z = dz * shift[1 + i];
vmulps(ymm0, ymm2, _m_shift(1 + i));
const size_t variableOffset = offsetof(GSScanlineLocalData, d[0].z) + (i * sizeof(GSScanlineLocalData::d[0]));
vmovaps(_rip_local_v(d[i].z, variableOffset), ymm0);
}
}
if(m_en.f)
{
// m_local.d8.p.f = GSVector4i(dp8).extract32<3>();
// FIXME no truncate ? why ? vcvttps2dq ?
//vcvtps2dq(ymm2, ymm1); // let's guess a typo
vcvttps2dq(ymm2, ymm1);
vpextrd(_rip_local(d8.p.f), xmm2, 3);
// GSVector8 df = GSVector8(dscan.p).wwww();
vshufps(ymm3, dscan_p, dscan_p, _MM_SHUFFLE(3, 3, 3, 3));
for(int i = 0; i < (m_sel.notest ? 1 : 8); i++)
{
// m_local.d[i].f = GSVector8i(df * m_shift[i]).xxzzlh();
vmulps(ymm0, ymm3, _m_shift(1 + i));
vcvttps2dq(ymm0, ymm0);
vpshuflw(ymm0, ymm0, _MM_SHUFFLE(2, 2, 0, 0));
vpshufhw(ymm0, ymm0, _MM_SHUFFLE(2, 2, 0, 0));
const size_t variableOffset = offsetof(GSScanlineLocalData, d[0].f) + (i * sizeof(GSScanlineLocalData::d[0]));
vmovdqa(_rip_local_v(d[i].f, variableOffset), ymm0);
}
}
}
else
{
// GSVector4 p = vertex[index[1]].p;
mov(_vertex.cvt32(), ptr[a1 + sizeof(uint32) * 1]);
shl(_vertex.cvt32(), 6); // * sizeof(GSVertexSW)
add(_vertex, a0);
if(m_en.f)
{
// m_local.p.f = GSVector4i(vertex[index[1]].p).extract32<3>();
vmovaps(xmm0, ptr[_vertex + offsetof(GSVertexSW, p)]);
vcvttps2dq(xmm0, xmm0);
vpextrd(_rip_local(p.f), xmm0, 3);
}
if(m_en.z)
{
// m_local.p.z = vertex[index[1]].t.u32[3]; // uint32 z is bypassed in t.w
mov(eax, ptr[ecx + offsetof(GSVertexSW, t.w)]);
mov(_rip_local(p.z), eax);
}
}
}
void GSSetupPrimCodeGenerator::Texture_AVX2()
{
if(!m_en.t)
{
return;
}
// GSVector8 dt(dscan.t);
vbroadcastf128(ymm0, ptr[a2 + offsetof(GSVertexSW, t)]);
// GSVector8 dt8 = dt * shift[0];
vmulps(ymm1, ymm0, _m_shift(0));
if(m_sel.fst)
{
// m_local.84.stq = GSVector4i(t * 4.0f);
vcvttps2dq(ymm1, ymm1);
vmovdqa(_rip_local(d8.stq), xmm1);
}
else
{
// m_local.d8.stq = t * 4.0f;
vmovaps(_rip_local(d8.stq), xmm1);
}
for(int j = 0, k = m_sel.fst ? 2 : 3; j < k; j++)
{
// GSVector8 dstq = dt.xxxx/yyyy/zzzz();
vshufps(ymm1, ymm0, ymm0, (uint8)_MM_SHUFFLE(j, j, j, j));
for(int i = 0; i < (m_sel.notest ? 1 : 8); i++)
{
// GSVector8 v = dstq * shift[1 + i];
vmulps(ymm2, ymm1, _m_shift(1+i));
if(m_sel.fst)
{
// m_local.d[i].s/t = GSVector8::cast(GSVector8i(v));
vcvttps2dq(ymm2, ymm2);
const size_t variableOffsetS = offsetof(GSScanlineLocalData, d[0].s) + (i * sizeof(GSScanlineLocalData::d[0]));
const size_t variableOffsetT = offsetof(GSScanlineLocalData, d[0].t) + (i * sizeof(GSScanlineLocalData::d[0]));
switch(j)
{
case 0: vmovdqa(_rip_local_v(d[i].s, variableOffsetS), ymm2); break;
case 1: vmovdqa(_rip_local_v(d[i].t, variableOffsetT), ymm2); break;
}
}
else
{
// m_local.d[i].s/t/q = v;
const size_t variableOffsetS = offsetof(GSScanlineLocalData, d[0].s) + (i * sizeof(GSScanlineLocalData::d[0]));
const size_t variableOffsetT = offsetof(GSScanlineLocalData, d[0].t) + (i * sizeof(GSScanlineLocalData::d[0]));
const size_t variableOffsetQ = offsetof(GSScanlineLocalData, d[0].q) + (i * sizeof(GSScanlineLocalData::d[0]));
switch(j)
{
case 0: vmovaps(_rip_local_v(d[i].s, variableOffsetS), ymm2); break;
case 1: vmovaps(_rip_local_v(d[i].t, variableOffsetT), ymm2); break;
case 2: vmovaps(_rip_local_v(d[i].q, variableOffsetQ), ymm2); break;
}
}
}
}
}
void GSSetupPrimCodeGenerator::Color_AVX2()
{
if(!m_en.c)
{
return;
}
if(m_sel.iip)
{
const Ymm& dscan_c = ymm6;
// GSVector8 dc(dscan.c);
vbroadcastf128(dscan_c, ptr[a2 + offsetof(GSVertexSW, c)]);
// m_local.d8.c = GSVector4i(c * 4.0f).xzyw().ps32();
vmulps(ymm1, dscan_c, ymm3);
vcvttps2dq(ymm1, ymm1);
vpshufd(ymm1, ymm1, _MM_SHUFFLE(3, 1, 2, 0));
vpackssdw(ymm1, ymm1);
vmovq(_rip_local(d8.c), xmm1);
// GSVector8 dr = dc.xxxx();
// GSVector8 db = dc.zzzz();
vshufps(ymm2, dscan_c, dscan_c, _MM_SHUFFLE(0, 0, 0, 0));
vshufps(ymm3, dscan_c, dscan_c, _MM_SHUFFLE(2, 2, 2, 2));
for(int i = 0; i < (m_sel.notest ? 1 : 8); i++)
{
// GSVector8i r = GSVector8i(dr * shift[1 + i]).ps32();
vmulps(ymm0, ymm2, _m_shift(1 + i));
vcvttps2dq(ymm0, ymm0);
vpackssdw(ymm0, ymm0);
// GSVector4i b = GSVector8i(db * shift[1 + i]).ps32();
vmulps(ymm1, ymm3, _m_shift(1 + i));
vcvttps2dq(ymm1, ymm1);
vpackssdw(ymm1, ymm1);
// m_local.d[i].rb = r.upl16(b);
vpunpcklwd(ymm0, ymm1);
const size_t variableOffset = offsetof(GSScanlineLocalData, d[0].rb) + (i * sizeof(GSScanlineLocalData::d[0]));
vmovdqa(_rip_local_v(d[i].rb, variableOffset), ymm0);
}
// GSVector8 dg = dc.yyyy();
// GSVector8 da = dc.wwww();
vshufps(ymm2, dscan_c, dscan_c, _MM_SHUFFLE(1, 1, 1, 1));
vshufps(ymm3, dscan_c, dscan_c, _MM_SHUFFLE(3, 3, 3, 3));
for(int i = 0; i < (m_sel.notest ? 1 : 8); i++)
{
// GSVector8i g = GSVector8i(dg * shift[1 + i]).ps32();
vmulps(ymm0, ymm2, _m_shift(1 + i));
vcvttps2dq(ymm0, ymm0);
vpackssdw(ymm0, ymm0);
// GSVector8i a = GSVector8i(da * shift[1 + i]).ps32();
vmulps(ymm1, ymm3, _m_shift(1 + i));
vcvttps2dq(ymm1, ymm1);
vpackssdw(ymm1, ymm1);
// m_local.d[i].ga = g.upl16(a);
vpunpcklwd(ymm0, ymm1);
const size_t variableOffset = offsetof(GSScanlineLocalData, d[0].ga) + (i * sizeof(GSScanlineLocalData::d[0]));
vmovdqa(_rip_local_v(d[i].ga, variableOffset), ymm0);
}
}
else
{
// GSVector4i c = GSVector4i(vertex[index[last].c);
int last = 0;
switch(m_sel.prim)
{
case GS_POINT_CLASS: last = 0; break;
case GS_LINE_CLASS: last = 1; break;
case GS_TRIANGLE_CLASS: last = 2; break;
case GS_SPRITE_CLASS: last = 1; break;
}
if(!(m_sel.prim == GS_SPRITE_CLASS && (m_en.z || m_en.f))) // if this is a sprite, the last vertex was already loaded in Depth()
{
mov(_vertex.cvt32(), ptr[a1 + sizeof(uint32) * last]);
shl(_vertex.cvt32(), 6); // * sizeof(GSVertexSW)
add(_vertex, a0);
}
vbroadcasti128(ymm0, ptr[_vertex + offsetof(GSVertexSW, c)]);
vcvttps2dq(ymm0, ymm0);
// c = c.upl16(c.zwxy());
vpshufd(ymm1, ymm0, _MM_SHUFFLE(1, 0, 3, 2));
vpunpcklwd(ymm0, ymm1);
// if(!tme) c = c.srl16(7);
if(m_sel.tfx == TFX_NONE)
{
vpsrlw(ymm0, 7);
}
// m_local.c.rb = c.xxxx();
// m_local.c.ga = c.zzzz();
vpshufd(ymm1, ymm0, _MM_SHUFFLE(0, 0, 0, 0));
vpshufd(ymm2, ymm0, _MM_SHUFFLE(2, 2, 2, 2));
vmovdqa(_rip_local(c.rb), ymm1);
vmovdqa(_rip_local(c.ga), ymm2);
}
}
#endif

View File

@ -32,7 +32,7 @@ static const int _vertex = _args + 4;
static const int _index = _args + 8;
static const int _dscan = _args + 12;
void GSSetupPrimCodeGenerator::Generate()
void GSSetupPrimCodeGenerator::Generate_AVX2()
{
if((m_en.z || m_en.f) && m_sel.prim != GS_SPRITE_CLASS || m_en.t || m_en.c && m_sel.iip)
{
@ -44,16 +44,16 @@ void GSSetupPrimCodeGenerator::Generate()
}
}
Depth();
Depth_AVX2();
Texture();
Texture_AVX2();
Color();
Color_AVX2();
ret();
}
void GSSetupPrimCodeGenerator::Depth()
void GSSetupPrimCodeGenerator::Depth_AVX2()
{
if(!m_en.z && !m_en.f)
{
@ -149,7 +149,7 @@ void GSSetupPrimCodeGenerator::Depth()
}
}
void GSSetupPrimCodeGenerator::Texture()
void GSSetupPrimCodeGenerator::Texture_AVX2()
{
if(!m_en.t)
{
@ -219,7 +219,7 @@ void GSSetupPrimCodeGenerator::Texture()
}
}
void GSSetupPrimCodeGenerator::Color()
void GSSetupPrimCodeGenerator::Color_AVX2()
{
if(!m_en.c)
{

View File

@ -147,6 +147,7 @@
<ClCompile Include="GSDrawScanline.cpp" />
<ClCompile Include="GSDrawScanlineCodeGenerator.cpp" />
<ClCompile Include="GSDrawScanlineCodeGenerator.x64.avx.cpp" />
<ClCompile Include="GSDrawScanlineCodeGenerator.x64.avx2.cpp" />
<ClCompile Include="GSDrawScanlineCodeGenerator.x64.cpp" />
<ClCompile Include="GSDrawScanlineCodeGenerator.x86.avx.cpp" />
<ClCompile Include="GSDrawScanlineCodeGenerator.x86.avx2.cpp" />
@ -173,6 +174,7 @@
<ClCompile Include="GSSettingsDlg.cpp" />
<ClCompile Include="GSSetupPrimCodeGenerator.cpp" />
<ClCompile Include="GSSetupPrimCodeGenerator.x64.avx.cpp" />
<ClCompile Include="GSSetupPrimCodeGenerator.x64.avx2.cpp" />
<ClCompile Include="GSSetupPrimCodeGenerator.x64.cpp" />
<ClCompile Include="GSSetupPrimCodeGenerator.x86.avx.cpp" />
<ClCompile Include="GSSetupPrimCodeGenerator.x86.avx2.cpp" />

View File

@ -231,6 +231,9 @@
<ClCompile Include="GSDrawScanlineCodeGenerator.x64.avx.cpp">
<Filter>Source Files</Filter>
</ClCompile>
<ClCompile Include="GSDrawScanlineCodeGenerator.x64.avx2.cpp">
<Filter>Source Files</Filter>
</ClCompile>
<ClCompile Include="GSDrawScanlineCodeGenerator.x64.cpp">
<Filter>Source Files</Filter>
</ClCompile>
@ -249,6 +252,9 @@
<ClCompile Include="GSSetupPrimCodeGenerator.x86.avx2.cpp">
<Filter>Source Files</Filter>
</ClCompile>
<ClCompile Include="GSSetupPrimCodeGenerator.x64.avx2.cpp">
<Filter>Source Files</Filter>
</ClCompile>
<ClCompile Include="GSRendererCL.cpp">
<Filter>Source Files</Filter>
</ClCompile>