Merge pull request #1713 from degasus/vertex-loader

virtual vertex loader

1 files changed, 74 insertions, 212 deletions

diff --git a/Source/Core/VideoCommon/VertexLoader.cpp b/Source/Core/VideoCommon/VertexLoader.cpp
index f293564e85..e2a2073ccf 100644
--- a/Source/Core/VideoCommon/VertexLoader.cpp
+++ b/Source/Core/VideoCommon/VertexLoader.cpp
@@ -4,7 +4,6 @@
 
 #include "Common/CommonTypes.h"
 #include "Common/MemoryUtil.h"
-#include "Common/StringUtil.h"
 #include "Common/x64ABI.h"
 #include "Common/x64Emitter.h"
 
@@ -30,97 +29,88 @@
 	#define inline
 #endif
 
-// Matrix components are first in GC format but later in PC format - we need to store it temporarily
-// when decoding each vertex.
-static u8 s_curposmtx = g_main_cp_state.matrix_index_a.PosNormalMtxIdx;
-static u8 s_curtexmtx[8];
-static int s_texmtxwrite = 0;
-static int s_texmtxread = 0;
-
-// Vertex loaders read these. Although the scale ones should be baked into the shader.
-int tcIndex;
-int colIndex;
-int colElements[2];
-// Duplicated (4x and 2x respectively) and used in SSE code in the vertex loader JIT
-GC_ALIGNED128(float posScale[4]);
-GC_ALIGNED64(float tcScale[8][2]);
-
 // This pointer is used as the source/dst for all fixed function loader calls
 u8* g_video_buffer_read_ptr;
 u8* g_vertex_manager_write_ptr;
 
-static const float fractionTable[32] = {
-	1.0f / (1U << 0), 1.0f / (1U << 1), 1.0f / (1U << 2), 1.0f / (1U << 3),
-	1.0f / (1U << 4), 1.0f / (1U << 5), 1.0f / (1U << 6), 1.0f / (1U << 7),
-	1.0f / (1U << 8), 1.0f / (1U << 9), 1.0f / (1U << 10), 1.0f / (1U << 11),
-	1.0f / (1U << 12), 1.0f / (1U << 13), 1.0f / (1U << 14), 1.0f / (1U << 15),
-	1.0f / (1U << 16), 1.0f / (1U << 17), 1.0f / (1U << 18), 1.0f / (1U << 19),
-	1.0f / (1U << 20), 1.0f / (1U << 21), 1.0f / (1U << 22), 1.0f / (1U << 23),
-	1.0f / (1U << 24), 1.0f / (1U << 25), 1.0f / (1U << 26), 1.0f / (1U << 27),
-	1.0f / (1U << 28), 1.0f / (1U << 29), 1.0f / (1U << 30), 1.0f / (1U << 31),
-};
-
 using namespace Gen;
 
-static void LOADERDECL PosMtx_ReadDirect_UByte()
+
+void* VertexLoader::operator new (size_t size)
+{
+	return AllocateAlignedMemory(size, 16);
+}
+
+void VertexLoader::operator delete (void *p)
+{
+	FreeAlignedMemory(p);
+}
+
+static void LOADERDECL PosMtx_ReadDirect_UByte(VertexLoader* loader)
 {
-	BoundingBox::posMtxIdx = s_curposmtx = DataReadU8() & 0x3f;
-	PRIM_LOG("posmtx: %d, ", s_curposmtx);
+	BoundingBox::posMtxIdx = loader->m_curposmtx = DataReadU8() & 0x3f;
+	PRIM_LOG("posmtx: %d, ", loader->m_curposmtx);
 }
 
-static void LOADERDECL PosMtx_Write()
+static void LOADERDECL PosMtx_Write(VertexLoader* loader)
 {
 	// u8, 0, 0, 0
-	DataWrite<u32>(s_curposmtx);
+	DataWrite<u32>(loader->m_curposmtx);
 }
 
-static void LOADERDECL TexMtx_ReadDirect_UByte()
+static void LOADERDECL TexMtx_ReadDirect_UByte(VertexLoader* loader)
 {
-	BoundingBox::texMtxIdx[s_texmtxread] = s_curtexmtx[s_texmtxread] = DataReadU8() & 0x3f;
+	BoundingBox::texMtxIdx[loader->m_texmtxread] = loader->m_curtexmtx[loader->m_texmtxread] = DataReadU8() & 0x3f;
 
-	PRIM_LOG("texmtx%d: %d, ", s_texmtxread, s_curtexmtx[s_texmtxread]);
-	s_texmtxread++;
+	PRIM_LOG("texmtx%d: %d, ", loader->m_texmtxread, loader->m_curtexmtx[loader->m_texmtxread]);
+	loader->m_texmtxread++;
 }
 
-static void LOADERDECL TexMtx_Write_Float()
+static void LOADERDECL TexMtx_Write_Float(VertexLoader* loader)
 {
-	DataWrite(float(s_curtexmtx[s_texmtxwrite++]));
+	DataWrite(float(loader->m_curtexmtx[loader->m_texmtxwrite++]));
 }
 
-static void LOADERDECL TexMtx_Write_Float2()
+static void LOADERDECL TexMtx_Write_Float2(VertexLoader* loader)
 {
 	DataWrite(0.f);
-	DataWrite(float(s_curtexmtx[s_texmtxwrite++]));
+	DataWrite(float(loader->m_curtexmtx[loader->m_texmtxwrite++]));
 }
 
-static void LOADERDECL TexMtx_Write_Float4()
+static void LOADERDECL TexMtx_Write_Float4(VertexLoader* loader)
 {
 #if _M_SSE >= 0x200
-	__m128 output = _mm_cvtsi32_ss(_mm_castsi128_ps(_mm_setzero_si128()), s_curtexmtx[s_texmtxwrite++]);
+	__m128 output = _mm_cvtsi32_ss(_mm_castsi128_ps(_mm_setzero_si128()), loader->m_curtexmtx[loader->m_texmtxwrite++]);
 	_mm_storeu_ps((float*)g_vertex_manager_write_ptr, _mm_shuffle_ps(output, output, 0x45 /* 1, 1, 0, 1 */));
 	g_vertex_manager_write_ptr += sizeof(float) * 4;
 #else
 	DataWrite(0.f);
 	DataWrite(0.f);
-	DataWrite(float(s_curtexmtx[s_texmtxwrite++]));
+	DataWrite(float(loader->m_curtexmtx[loader->m_texmtxwrite++]));
 	// Just to fill out with 0.
 	DataWrite(0.f);
 #endif
 }
 
+static void LOADERDECL SkipVertex(VertexLoader* loader)
+{
+	if (loader->m_vertexSkip)
+	{
+		// reset the output buffer
+		g_vertex_manager_write_ptr -= loader->m_native_vtx_decl.stride;
+
+		loader->m_skippedVertices++;
+	}
+}
+
 VertexLoader::VertexLoader(const TVtxDesc &vtx_desc, const VAT &vtx_attr)
+: VertexLoaderBase(vtx_desc, vtx_attr)
 {
 	m_compiledCode = nullptr;
-	m_numLoadedVertices = 0;
-	m_VertexSize = 0;
-	m_native_vertex_format = nullptr;
 	VertexLoader_Normal::Init();
 	VertexLoader_Position::Init();
 	VertexLoader_TextCoord::Init();
 
-	m_VtxDesc = vtx_desc;
-	SetVAT(vtx_attr);
-
 	#ifdef USE_VERTEX_LOADER_JIT
 	AllocCodeSpace(COMPILED_CODE_SIZE);
 	CompileVertexTranslator();
@@ -130,6 +120,13 @@ VertexLoader::VertexLoader(const TVtxDesc &vtx_desc, const VAT &vtx_attr)
 	CompileVertexTranslator();
 	#endif
 
+	// generate frac factors
+	m_posScale[0] = m_posScale[1] = m_posScale[2] = m_posScale[3] = 1.0f / (1U << m_VtxAttr.PosFrac);
+	for (int i = 0; i < 8; i++)
+		m_tcScale[i][0] = m_tcScale[i][1] = 1.0f / (1U << m_VtxAttr.texCoord[i].Frac);
+
+	for (int i = 0; i < 2; i++)
+		m_colElements[i] = m_VtxAttr.color[i].Elements;
 }
 
 VertexLoader::~VertexLoader()
@@ -150,11 +147,14 @@ void VertexLoader::CompileVertexTranslator()
 
 	m_compiledCode = GetCodePtr();
 	// We only use RAX (caller saved) and RBX (callee saved).
-	ABI_PushRegistersAndAdjustStack({RBX}, 8);
+	ABI_PushRegistersAndAdjustStack({RBX, RBP}, 8);
 
 	// save count
 	MOV(64, R(RBX), R(ABI_PARAM1));
 
+	// save loader
+	MOV(64, R(RBP), R(ABI_PARAM2));
+
 	// Start loop here
 	const u8 *loop_start = GetCodePtr();
 
@@ -162,17 +162,17 @@ void VertexLoader::CompileVertexTranslator()
 	if (m_VtxDesc.Tex0Coord || m_VtxDesc.Tex1Coord || m_VtxDesc.Tex2Coord || m_VtxDesc.Tex3Coord ||
 		m_VtxDesc.Tex4Coord || m_VtxDesc.Tex5Coord || m_VtxDesc.Tex6Coord || m_VtxDesc.Tex7Coord)
 	{
-		WriteSetVariable(32, &tcIndex, Imm32(0));
+		WriteSetVariable(32, &m_tcIndex, Imm32(0));
 	}
 	if (m_VtxDesc.Color0 || m_VtxDesc.Color1)
 	{
-		WriteSetVariable(32, &colIndex, Imm32(0));
+		WriteSetVariable(32, &m_colIndex, Imm32(0));
 	}
 	if (m_VtxDesc.Tex0MatIdx || m_VtxDesc.Tex1MatIdx || m_VtxDesc.Tex2MatIdx || m_VtxDesc.Tex3MatIdx ||
 		m_VtxDesc.Tex4MatIdx || m_VtxDesc.Tex5MatIdx || m_VtxDesc.Tex6MatIdx || m_VtxDesc.Tex7MatIdx)
 	{
-		WriteSetVariable(32, &s_texmtxwrite, Imm32(0));
-		WriteSetVariable(32, &s_texmtxread, Imm32(0));
+		WriteSetVariable(32, &m_texmtxwrite, Imm32(0));
+		WriteSetVariable(32, &m_texmtxread, Imm32(0));
 	}
 #else
 	// Reset pipeline
@@ -404,6 +404,12 @@ void VertexLoader::CompileVertexTranslator()
 		nat_offset += 4;
 	}
 
+	// indexed position formats may skip a the vertex
+	if (m_VtxDesc.Position & 2)
+	{
+		WriteCall(SkipVertex);
+	}
+
 	m_native_components = components;
 	m_native_vtx_decl.stride = nat_offset;
 
@@ -412,7 +418,7 @@ void VertexLoader::CompileVertexTranslator()
 	SUB(64, R(RBX), Imm8(1));
 
 	J_CC(CC_NZ, loop_start);
-	ABI_PopRegistersAndAdjustStack({RBX}, 8);
+	ABI_PopRegistersAndAdjustStack({RBX, RBP}, 8);
 	RET();
 #endif
 }
@@ -420,6 +426,7 @@ void VertexLoader::CompileVertexTranslator()
 void VertexLoader::WriteCall(TPipelineFunction func)
 {
 #ifdef USE_VERTEX_LOADER_JIT
+	MOV(64, R(ABI_PARAM1), R(RBP));
 	ABI_CallFunction((const void*)func);
 #else
 	m_PipelineStages[m_numPipelineStages++] = func;
@@ -444,179 +451,34 @@ void VertexLoader::WriteSetVariable(int bits, void *address, OpArg value)
 }
 #endif
 
-void VertexLoader::SetupRunVertices(const VAT& vat, int primitive, int const count)
+int VertexLoader::RunVertices(int primitive, int count, DataReader src, DataReader dst)
 {
-	m_numLoadedVertices += count;
+	dst.WritePointer(&g_vertex_manager_write_ptr);
+	src.WritePointer(&g_video_buffer_read_ptr);
 
-	// Load position and texcoord scale factors.
-	m_VtxAttr.PosFrac          = vat.g0.PosFrac;
-	m_VtxAttr.texCoord[0].Frac = vat.g0.Tex0Frac;
-	m_VtxAttr.texCoord[1].Frac = vat.g1.Tex1Frac;
-	m_VtxAttr.texCoord[2].Frac = vat.g1.Tex2Frac;
-	m_VtxAttr.texCoord[3].Frac = vat.g1.Tex3Frac;
-	m_VtxAttr.texCoord[4].Frac = vat.g2.Tex4Frac;
-	m_VtxAttr.texCoord[5].Frac = vat.g2.Tex5Frac;
-	m_VtxAttr.texCoord[6].Frac = vat.g2.Tex6Frac;
-	m_VtxAttr.texCoord[7].Frac = vat.g2.Tex7Frac;
-
-	posScale[0] = posScale[1] = posScale[2] = posScale[3] = fractionTable[m_VtxAttr.PosFrac];
-	if (m_native_components & VB_HAS_UVALL)
-		for (int i = 0; i < 8; i++)
-			tcScale[i][0] = tcScale[i][1] = fractionTable[m_VtxAttr.texCoord[i].Frac];
-	for (int i = 0; i < 2; i++)
-		colElements[i] = m_VtxAttr.color[i].Elements;
+	m_numLoadedVertices += count;
+	m_skippedVertices = 0;
 
 	// Prepare bounding box
 	if (!g_ActiveConfig.backend_info.bSupportsBBox)
-		BoundingBox::Prepare(vat, primitive, m_VtxDesc, m_native_vtx_decl);
-}
+		BoundingBox::Prepare(m_vat, primitive, m_VtxDesc, m_native_vtx_decl);
 
-void VertexLoader::ConvertVertices ( int count )
-{
 #ifdef USE_VERTEX_LOADER_JIT
 	if (count > 0)
 	{
-		((void (*)(int))(void*)m_compiledCode)(count);
+		((void (*)(int, VertexLoader* loader))(void*)m_compiledCode)(count, this);
 	}
 #else
 	for (int s = 0; s < count; s++)
 	{
-		tcIndex = 0;
-		colIndex = 0;
-		s_texmtxwrite = s_texmtxread = 0;
+		m_tcIndex = 0;
+		m_colIndex = 0;
+		m_texmtxwrite = m_texmtxread = 0;
 		for (int i = 0; i < m_numPipelineStages; i++)
-			m_PipelineStages[i]();
+			m_PipelineStages[i](this);
 		PRIM_LOG("\n");
 	}
 #endif
-}
-
-int VertexLoader::RunVertices(const VAT& vat, int primitive, int count, DataReader src, DataReader dst)
-{
-	dst.WritePointer(&g_vertex_manager_write_ptr);
-	src.WritePointer(&g_video_buffer_read_ptr);
-	SetupRunVertices(vat, primitive, count);
-	ConvertVertices(count);
-	return count;
-}
-
-void VertexLoader::SetVAT(const VAT& vat)
-{
-	m_VtxAttr.PosElements          = vat.g0.PosElements;
-	m_VtxAttr.PosFormat            = vat.g0.PosFormat;
-	m_VtxAttr.PosFrac              = vat.g0.PosFrac;
-	m_VtxAttr.NormalElements       = vat.g0.NormalElements;
-	m_VtxAttr.NormalFormat         = vat.g0.NormalFormat;
-	m_VtxAttr.color[0].Elements    = vat.g0.Color0Elements;
-	m_VtxAttr.color[0].Comp        = vat.g0.Color0Comp;
-	m_VtxAttr.color[1].Elements    = vat.g0.Color1Elements;
-	m_VtxAttr.color[1].Comp        = vat.g0.Color1Comp;
-	m_VtxAttr.texCoord[0].Elements = vat.g0.Tex0CoordElements;
-	m_VtxAttr.texCoord[0].Format   = vat.g0.Tex0CoordFormat;
-	m_VtxAttr.texCoord[0].Frac     = vat.g0.Tex0Frac;
-	m_VtxAttr.ByteDequant          = vat.g0.ByteDequant;
-	m_VtxAttr.NormalIndex3         = vat.g0.NormalIndex3;
-
-	m_VtxAttr.texCoord[1].Elements = vat.g1.Tex1CoordElements;
-	m_VtxAttr.texCoord[1].Format   = vat.g1.Tex1CoordFormat;
-	m_VtxAttr.texCoord[1].Frac     = vat.g1.Tex1Frac;
-	m_VtxAttr.texCoord[2].Elements = vat.g1.Tex2CoordElements;
-	m_VtxAttr.texCoord[2].Format   = vat.g1.Tex2CoordFormat;
-	m_VtxAttr.texCoord[2].Frac     = vat.g1.Tex2Frac;
-	m_VtxAttr.texCoord[3].Elements = vat.g1.Tex3CoordElements;
-	m_VtxAttr.texCoord[3].Format   = vat.g1.Tex3CoordFormat;
-	m_VtxAttr.texCoord[3].Frac     = vat.g1.Tex3Frac;
-	m_VtxAttr.texCoord[4].Elements = vat.g1.Tex4CoordElements;
-	m_VtxAttr.texCoord[4].Format   = vat.g1.Tex4CoordFormat;
-
-	m_VtxAttr.texCoord[4].Frac     = vat.g2.Tex4Frac;
-	m_VtxAttr.texCoord[5].Elements = vat.g2.Tex5CoordElements;
-	m_VtxAttr.texCoord[5].Format   = vat.g2.Tex5CoordFormat;
-	m_VtxAttr.texCoord[5].Frac     = vat.g2.Tex5Frac;
-	m_VtxAttr.texCoord[6].Elements = vat.g2.Tex6CoordElements;
-	m_VtxAttr.texCoord[6].Format   = vat.g2.Tex6CoordFormat;
-	m_VtxAttr.texCoord[6].Frac     = vat.g2.Tex6Frac;
-	m_VtxAttr.texCoord[7].Elements = vat.g2.Tex7CoordElements;
-	m_VtxAttr.texCoord[7].Format   = vat.g2.Tex7CoordFormat;
-	m_VtxAttr.texCoord[7].Frac     = vat.g2.Tex7Frac;
-
-	if (!m_VtxAttr.ByteDequant)
-	{
-		ERROR_LOG(VIDEO, "ByteDequant is set to zero");
-	}
-};
-
-void VertexLoader::AppendToString(std::string *dest) const
-{
-	dest->reserve(250);
-	static const char *posMode[4] = {
-		"Inv",
-		"Dir",
-		"I8",
-		"I16",
-	};
-	static const char *posFormats[5] = {
-		"u8", "s8", "u16", "s16", "flt",
-	};
-	static const char *colorFormat[8] = {
-		"565",
-		"888",
-		"888x",
-		"4444",
-		"6666",
-		"8888",
-		"Inv",
-		"Inv",
-	};
-
-	dest->append(StringFromFormat("%ib skin: %i P: %i %s-%s ",
-		m_VertexSize, (u32)m_VtxDesc.PosMatIdx,
-		m_VtxAttr.PosElements ? 3 : 2, posMode[m_VtxDesc.Position], posFormats[m_VtxAttr.PosFormat]));
-
-	if (m_VtxDesc.Normal)
-	{
-		dest->append(StringFromFormat("Nrm: %i %s-%s ",
-			m_VtxAttr.NormalElements, posMode[m_VtxDesc.Normal], posFormats[m_VtxAttr.NormalFormat]));
-	}
-
-	u64 color_mode[2] = {m_VtxDesc.Color0, m_VtxDesc.Color1};
-	for (int i = 0; i < 2; i++)
-	{
-		if (color_mode[i])
-		{
-			dest->append(StringFromFormat("C%i: %i %s-%s ", i, m_VtxAttr.color[i].Elements, posMode[color_mode[i]], colorFormat[m_VtxAttr.color[i].Comp]));
-		}
-	}
-	u64 tex_mode[8] = {
-		m_VtxDesc.Tex0Coord, m_VtxDesc.Tex1Coord, m_VtxDesc.Tex2Coord, m_VtxDesc.Tex3Coord,
-		m_VtxDesc.Tex4Coord, m_VtxDesc.Tex5Coord, m_VtxDesc.Tex6Coord, m_VtxDesc.Tex7Coord
-	};
-	for (int i = 0; i < 8; i++)
-	{
-		if (tex_mode[i])
-		{
-			dest->append(StringFromFormat("T%i: %i %s-%s ",
-				i, m_VtxAttr.texCoord[i].Elements, posMode[tex_mode[i]], posFormats[m_VtxAttr.texCoord[i].Format]));
-		}
-	}
-	dest->append(StringFromFormat(" - %i v\n", m_numLoadedVertices));
-}
-
-NativeVertexFormat* VertexLoader::GetNativeVertexFormat()
-{
-	if (m_native_vertex_format)
-		return m_native_vertex_format;
-	auto& native = s_native_vertex_map[m_native_vtx_decl];
-	if (!native)
-	{
-		auto raw_pointer = g_vertex_manager->CreateNativeVertexFormat();
-		native = std::unique_ptr<NativeVertexFormat>(raw_pointer);
-		native->Initialize(m_native_vtx_decl);
-		native->m_components = m_native_components;
-	}
-	m_native_vertex_format = native.get();
-	return native.get();
 
+	return count - m_skippedVertices;
 }
-
-std::unordered_map<PortableVertexDeclaration, std::unique_ptr<NativeVertexFormat>> VertexLoader::s_native_vertex_map;