prifak/yuzu
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'master' of https://github.com/yuzu-emu/yuzu

Browse Source
This commit is contained in:
Jarek Syrylak
2018-07-05 19:33:38 +01:00
parent 22c008ea20 762bf6a522
commit 1ba63bd4d7
7 changed files with 86 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/video_core/engines/shader_bytecode.h
View File

@@ -328,6 +328,19 @@ union Instruction {
BitField<49, 3, PredCondition> cond;
} isetp;
union {
BitField<0, 3, u64> pred0;
BitField<3, 3, u64> pred3;
BitField<12, 3, u64> pred12;
BitField<15, 1, u64> neg_pred12;
BitField<24, 2, PredOperation> cond;
BitField<29, 3, u64> pred29;
BitField<32, 1, u64> neg_pred29;
BitField<39, 3, u64> pred39;
BitField<42, 1, u64> neg_pred39;
BitField<45, 2, PredOperation> op;
} psetp;
union {
BitField<39, 3, u64> pred39;
BitField<42, 1, u64> neg_pred;
@@ -438,6 +451,8 @@ public:
enum class Id {
KIL,
SSY,
SYNC,
DEPBAR,
BFE_C,
BFE_R,
BFE_IMM,
@@ -534,6 +549,7 @@ public:
Shift,
Ffma,
Flow,
Synch,
Memory,
FloatSet,
FloatSetPredicate,
@@ -638,10 +654,12 @@ private:
INST("111000110011----", Id::KIL, Type::Flow, "KIL"),
INST("111000101001----", Id::SSY, Type::Flow, "SSY"),
INST("111000100100----", Id::BRA, Type::Flow, "BRA"),
INST("1111000011110---", Id::DEPBAR, Type::Synch, "DEPBAR"),
INST("1111000011111---", Id::SYNC, Type::Synch, "SYNC"),
INST("1110111111011---", Id::LD_A, Type::Memory, "LD_A"),
INST("1110111110010---", Id::LD_C, Type::Memory, "LD_C"),
INST("1110111111110---", Id::ST_A, Type::Memory, "ST_A"),
INST("1100000000111---", Id::TEX, Type::Memory, "TEX"),
INST("110000----111---", Id::TEX, Type::Memory, "TEX"),
INST("1101111101001---", Id::TEXQ, Type::Memory, "TEXQ"),
INST("1101100---------", Id::TEXS, Type::Memory, "TEXS"),
INST("1101101---------", Id::TLDS, Type::Memory, "TLDS"),

5
src/video_core/renderer_opengl/gl_rasterizer.cpp
View File

@@ -686,7 +686,10 @@ u32 RasterizerOpenGL::SetupTextures(Maxwell::ShaderStage stage, GLuint program,
// Bind the uniform to the sampler.
GLint uniform = glGetUniformLocation(program, entry.GetName().c_str());
ASSERT(uniform != -1);
if (uniform == -1) {
continue;
}
glProgramUniform1i(program, uniform, current_bindpoint);
const auto texture = maxwell3d.GetStageTexture(entry.GetStage(), entry.GetOffset());

6
src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
View File

@@ -108,7 +108,8 @@ static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_form
{GL_DEPTH24_STENCIL8, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, ComponentType::UNorm,
false}, // Z24S8
{GL_DEPTH24_STENCIL8, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, ComponentType::UNorm,
false}, // S8Z24
false}, // S8Z24
{GL_DEPTH_COMPONENT32F, GL_DEPTH_COMPONENT, GL_FLOAT, ComponentType::Float, false}, // Z32F
}};
static const FormatTuple& GetFormatTuple(PixelFormat pixel_format, ComponentType component_type) {
@@ -191,7 +192,7 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr),
MortonCopy<true, PixelFormat::DXT1>, MortonCopy<true, PixelFormat::DXT23>,
MortonCopy<true, PixelFormat::DXT45>, MortonCopy<true, PixelFormat::DXN1>,
MortonCopy<true, PixelFormat::ASTC_2D_4X4>, MortonCopy<true, PixelFormat::Z24S8>,
MortonCopy<true, PixelFormat::S8Z24>,
MortonCopy<true, PixelFormat::S8Z24>, MortonCopy<true, PixelFormat::Z32F>,
};
static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr),
@@ -213,6 +214,7 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr),
MortonCopy<false, PixelFormat::ABGR8>,
MortonCopy<false, PixelFormat::Z24S8>,
MortonCopy<false, PixelFormat::S8Z24>,
MortonCopy<false, PixelFormat::Z32F>,
};
// Allocate an uninitialized texture of appropriate size and format for the surface

9
src/video_core/renderer_opengl/gl_rasterizer_cache.h
View File

@@ -42,6 +42,7 @@ struct SurfaceParams {
// DepthStencil formats
Z24S8 = 13,
S8Z24 = 14,
Z32F = 15,
MaxDepthStencilFormat,
@@ -94,6 +95,7 @@ struct SurfaceParams {
4, // ASTC_2D_4X4
1, // Z24S8
1, // S8Z24
1, // Z32F
}};
ASSERT(static_cast<size_t>(format) < compression_factor_table.size());
@@ -120,6 +122,7 @@ struct SurfaceParams {
32, // ASTC_2D_4X4
32, // Z24S8
32, // S8Z24
32, // Z32F
}};
ASSERT(static_cast<size_t>(format) < bpp_table.size());
@@ -135,6 +138,8 @@ struct SurfaceParams {
return PixelFormat::S8Z24;
case Tegra::DepthFormat::Z24_S8_UNORM:
return PixelFormat::Z24S8;
case Tegra::DepthFormat::Z32_FLOAT:
return PixelFormat::Z32F;
default:
LOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
UNREACHABLE();
@@ -235,6 +240,8 @@ struct SurfaceParams {
return Tegra::DepthFormat::S8_Z24_UNORM;
case PixelFormat::Z24S8:
return Tegra::DepthFormat::Z24_S8_UNORM;
case PixelFormat::Z32F:
return Tegra::DepthFormat::Z32_FLOAT;
default:
UNREACHABLE();
}
@@ -284,6 +291,8 @@ struct SurfaceParams {
case Tegra::DepthFormat::S8_Z24_UNORM:
case Tegra::DepthFormat::Z24_S8_UNORM:
return ComponentType::UNorm;
case Tegra::DepthFormat::Z32_FLOAT:
return ComponentType::Float;
default:
LOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
UNREACHABLE();

40
src/video_core/renderer_opengl/gl_shader_decompiler.cpp
View File

@@ -1213,6 +1213,9 @@ private:
switch (instr.conversion.f2f.rounding) {
case Tegra::Shader::F2fRoundingOp::None:
break;
case Tegra::Shader::F2fRoundingOp::Round:
op_a = "roundEven(" + op_a + ')';
break;
case Tegra::Shader::F2fRoundingOp::Floor:
op_a = "floor(" + op_a + ')';
break;
@@ -1477,6 +1480,36 @@ private:
}
break;
}
case OpCode::Type::PredicateSetPredicate: {
std::string op_a =
GetPredicateCondition(instr.psetp.pred12, instr.psetp.neg_pred12 != 0);
std::string op_b =
GetPredicateCondition(instr.psetp.pred29, instr.psetp.neg_pred29 != 0);
using Tegra::Shader::Pred;
// We can't use the constant predicate as destination.
ASSERT(instr.psetp.pred3 != static_cast<u64>(Pred::UnusedIndex));
std::string second_pred =
GetPredicateCondition(instr.psetp.pred39, instr.psetp.neg_pred39 != 0);
std::string combiner = GetPredicateCombiner(instr.psetp.op);
std::string predicate =
'(' + op_a + ") " + GetPredicateCombiner(instr.psetp.cond) + " (" + op_b + ')';
// Set the primary predicate to the result of Predicate OP SecondPredicate
SetPredicate(instr.psetp.pred3,
'(' + predicate + ") " + combiner + " (" + second_pred + ')');
if (instr.psetp.pred0 != static_cast<u64>(Pred::UnusedIndex)) {
// Set the secondary predicate to the result of !Predicate OP SecondPredicate,
// if enabled
SetPredicate(instr.psetp.pred0,
"!(" + predicate + ") " + combiner + " (" + second_pred + ')');
}
break;
}
case OpCode::Type::FloatSet: {
std::string op_a = instr.fset.neg_a ? "-" : "";
op_a += regs.GetRegisterAsFloat(instr.gpr8);
@@ -1599,6 +1632,13 @@ private:
// can ignore this when generating GLSL code.
break;
}
case OpCode::Id::DEPBAR:
case OpCode::Id::SYNC: {
// TODO(Subv): Find out if we actually have to care about these instructions or if
// the GLSL compiler takes care of that for us.
LOG_WARNING(HW_GPU, "DEPBAR/SYNC instruction is stubbed");
break;
}
default: {
LOG_CRITICAL(HW_GPU, "Unhandled instruction: {}", opcode->GetName());
UNREACHABLE();

8
src/video_core/renderer_opengl/maxwell_to_gl.h
View File

@@ -29,6 +29,10 @@ inline GLenum VertexType(Maxwell::VertexAttribute attrib) {
switch (attrib.size) {
case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
return GL_UNSIGNED_BYTE;
case Maxwell::VertexAttribute::Size::Size_16_16:
return GL_UNSIGNED_SHORT;
case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
return GL_UNSIGNED_INT_2_10_10_10_REV;
}
LOG_CRITICAL(Render_OpenGL, "Unimplemented vertex size={}", attrib.SizeString());
@@ -41,6 +45,10 @@ inline GLenum VertexType(Maxwell::VertexAttribute attrib) {
switch (attrib.size) {
case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
return GL_BYTE;
case Maxwell::VertexAttribute::Size::Size_16_16:
return GL_SHORT;
case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
return GL_INT_2_10_10_10_REV;
}
LOG_CRITICAL(Render_OpenGL, "Unimplemented vertex size={}", attrib.SizeString());

2
src/video_core/textures/decoders.cpp
View File

@@ -78,6 +78,7 @@ static u32 DepthBytesPerPixel(DepthFormat format) {
switch (format) {
case DepthFormat::S8_Z24_UNORM:
case DepthFormat::Z24_S8_UNORM:
case DepthFormat::Z32_FLOAT:
return 4;
default:
UNIMPLEMENTED_MSG("Format not implemented");
@@ -132,6 +133,7 @@ std::vector<u8> UnswizzleDepthTexture(VAddr address, DepthFormat format, u32 wid
switch (format) {
case DepthFormat::S8_Z24_UNORM:
case DepthFormat::Z24_S8_UNORM:
case DepthFormat::Z32_FLOAT:
CopySwizzledData(width, height, bytes_per_pixel, bytes_per_pixel, data,
unswizzled_data.data(), true, block_height);
break;