aboutsummaryrefslogtreecommitdiff
path: root/src/video_core
diff options
context:
space:
mode:
Diffstat (limited to 'src/video_core')
-rw-r--r--src/video_core/engines/maxwell_3d.h2
-rw-r--r--src/video_core/engines/shader_bytecode.h6
-rw-r--r--src/video_core/gpu.cpp2
-rw-r--r--src/video_core/macro_interpreter.cpp2
-rw-r--r--src/video_core/morton.cpp1
-rw-r--r--src/video_core/renderer_opengl/gl_rasterizer_cache.cpp12
-rw-r--r--src/video_core/renderer_opengl/gl_shader_cache.cpp45
-rw-r--r--src/video_core/renderer_opengl/gl_shader_cache.h3
-rw-r--r--src/video_core/renderer_opengl/gl_shader_decompiler.cpp308
-rw-r--r--src/video_core/renderer_opengl/gl_shader_gen.cpp7
-rw-r--r--src/video_core/renderer_opengl/gl_shader_gen.h3
-rw-r--r--src/video_core/renderer_opengl/renderer_opengl.cpp2
-rw-r--r--src/video_core/surface.cpp7
-rw-r--r--src/video_core/textures/decoders.cpp2
14 files changed, 269 insertions, 133 deletions
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 25bb7604a..0faff6fdf 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -164,6 +164,7 @@ public:
return 3;
default:
UNREACHABLE();
+ return 1;
}
}
@@ -871,6 +872,7 @@ public:
return 4;
}
UNREACHABLE();
+ return 1;
}
GPUVAddr StartAddress() const {
diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h
index 5ea094e64..eb703bb5a 100644
--- a/src/video_core/engines/shader_bytecode.h
+++ b/src/video_core/engines/shader_bytecode.h
@@ -575,7 +575,7 @@ union Instruction {
union {
BitField<39, 2, u64> tab5cb8_2;
- BitField<41, 3, u64> tab5c68_1;
+ BitField<41, 3, u64> postfactor;
BitField<44, 2, u64> tab5c68_0;
BitField<48, 1, u64> negate_b;
} fmul;
@@ -609,7 +609,7 @@ union Instruction {
BitField<31, 1, u64> negate_b;
BitField<30, 1, u64> abs_b;
- BitField<47, 2, HalfType> type_b;
+ BitField<28, 2, HalfType> type_b;
BitField<35, 2, HalfType> type_c;
} alu_half;
@@ -1065,6 +1065,7 @@ union Instruction {
LOG_CRITICAL(HW_GPU, "Unhandled texture_info: {}",
static_cast<u32>(texture_info.Value()));
UNREACHABLE();
+ return TextureType::Texture1D;
}
TextureProcessMode GetTextureProcessMode() const {
@@ -1145,6 +1146,7 @@ union Instruction {
LOG_CRITICAL(HW_GPU, "Unhandled texture_info: {}",
static_cast<u32>(texture_info.Value()));
UNREACHABLE();
+ return TextureType::Texture1D;
}
TextureProcessMode GetTextureProcessMode() const {
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 88c45a423..08cf6268f 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -102,6 +102,7 @@ u32 RenderTargetBytesPerPixel(RenderTargetFormat format) {
return 1;
default:
UNIMPLEMENTED_MSG("Unimplemented render target format {}", static_cast<u32>(format));
+ return 1;
}
}
@@ -119,6 +120,7 @@ u32 DepthFormatBytesPerPixel(DepthFormat format) {
return 2;
default:
UNIMPLEMENTED_MSG("Unimplemented Depth format {}", static_cast<u32>(format));
+ return 1;
}
}
diff --git a/src/video_core/macro_interpreter.cpp b/src/video_core/macro_interpreter.cpp
index 9c55e9f1e..64f75db43 100644
--- a/src/video_core/macro_interpreter.cpp
+++ b/src/video_core/macro_interpreter.cpp
@@ -171,6 +171,7 @@ u32 MacroInterpreter::GetALUResult(ALUOperation operation, u32 src_a, u32 src_b)
default:
UNIMPLEMENTED_MSG("Unimplemented ALU operation {}", static_cast<u32>(operation));
+ return 0;
}
}
@@ -268,6 +269,7 @@ bool MacroInterpreter::EvaluateBranchCondition(BranchCondition cond, u32 value)
return value != 0;
}
UNREACHABLE();
+ return true;
}
} // namespace Tegra
diff --git a/src/video_core/morton.cpp b/src/video_core/morton.cpp
index a310491a8..47e76d8fe 100644
--- a/src/video_core/morton.cpp
+++ b/src/video_core/morton.cpp
@@ -192,6 +192,7 @@ static MortonCopyFn GetSwizzleFunction(MortonSwizzleMode mode, Surface::PixelFor
return linear_to_morton_fns[static_cast<std::size_t>(format)];
}
UNREACHABLE();
+ return morton_to_linear_fns[static_cast<std::size_t>(format)];
}
/// 8x8 Z-Order coordinate from 2D coordinates
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
index 5f4cdd119..7ea07631a 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
@@ -101,8 +101,18 @@ std::size_t SurfaceParams::InnerMemorySize(bool force_gl, bool layer_only,
params.srgb_conversion = config.tic.IsSrgbConversionEnabled();
params.pixel_format = PixelFormatFromTextureFormat(config.tic.format, config.tic.r_type.Value(),
params.srgb_conversion);
+
+ if (params.pixel_format == PixelFormat::R16U && config.tsc.depth_compare_enabled) {
+ // Some titles create a 'R16U' (normalized 16-bit) texture with depth_compare enabled,
+ // then attempt to sample from it via a shadow sampler. Convert format to Z16 (which also
+ // causes GetFormatType to properly return 'Depth' below).
+ params.pixel_format = PixelFormat::Z16;
+ }
+
params.component_type = ComponentTypeFromTexture(config.tic.r_type.Value());
params.type = GetFormatType(params.pixel_format);
+ UNIMPLEMENTED_IF(params.type == SurfaceType::ColorTexture && config.tsc.depth_compare_enabled);
+
params.width = Common::AlignUp(config.tic.Width(), GetCompressionFactor(params.pixel_format));
params.height = Common::AlignUp(config.tic.Height(), GetCompressionFactor(params.pixel_format));
params.unaligned_height = config.tic.Height();
@@ -257,7 +267,7 @@ static constexpr std::array<FormatTuple, VideoCore::Surface::MaxPixelFormat> tex
{GL_R8UI, GL_RED_INTEGER, GL_UNSIGNED_BYTE, ComponentType::UInt, false}, // R8UI
{GL_RGBA16F, GL_RGBA, GL_HALF_FLOAT, ComponentType::Float, false}, // RGBA16F
{GL_RGBA16, GL_RGBA, GL_UNSIGNED_SHORT, ComponentType::UNorm, false}, // RGBA16U
- {GL_RGBA16UI, GL_RGBA, GL_UNSIGNED_SHORT, ComponentType::UInt, false}, // RGBA16UI
+ {GL_RGBA16UI, GL_RGBA_INTEGER, GL_UNSIGNED_SHORT, ComponentType::UInt, false}, // RGBA16UI
{GL_R11F_G11F_B10F, GL_RGB, GL_UNSIGNED_INT_10F_11F_11F_REV, ComponentType::Float,
false}, // R11FG11FB10F
{GL_RGBA32UI, GL_RGBA_INTEGER, GL_UNSIGNED_INT, ComponentType::UInt, false}, // RGBA32UI
diff --git a/src/video_core/renderer_opengl/gl_shader_cache.cpp b/src/video_core/renderer_opengl/gl_shader_cache.cpp
index 038b25c75..aea6bf1af 100644
--- a/src/video_core/renderer_opengl/gl_shader_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_cache.cpp
@@ -2,7 +2,9 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
+#include <boost/functional/hash.hpp>
#include "common/assert.h"
+#include "common/hash.h"
#include "core/core.h"
#include "core/memory.h"
#include "video_core/engines/maxwell_3d.h"
@@ -66,14 +68,17 @@ CachedShader::CachedShader(VAddr addr, Maxwell::ShaderProgram program_type)
// stage here.
setup.SetProgramB(GetShaderCode(GetShaderAddress(Maxwell::ShaderProgram::VertexB)));
case Maxwell::ShaderProgram::VertexB:
+ CalculateProperties();
program_result = GLShader::GenerateVertexShader(setup);
gl_type = GL_VERTEX_SHADER;
break;
case Maxwell::ShaderProgram::Geometry:
+ CalculateProperties();
program_result = GLShader::GenerateGeometryShader(setup);
gl_type = GL_GEOMETRY_SHADER;
break;
case Maxwell::ShaderProgram::Fragment:
+ CalculateProperties();
program_result = GLShader::GenerateFragmentShader(setup);
gl_type = GL_FRAGMENT_SHADER;
break;
@@ -140,6 +145,46 @@ GLuint CachedShader::LazyGeometryProgram(OGLProgram& target_program,
return target_program.handle;
};
+static bool IsSchedInstruction(std::size_t offset, std::size_t main_offset) {
+ // sched instructions appear once every 4 instructions.
+ static constexpr std::size_t SchedPeriod = 4;
+ const std::size_t absolute_offset = offset - main_offset;
+ return (absolute_offset % SchedPeriod) == 0;
+}
+
+static std::size_t CalculateProgramSize(const GLShader::ProgramCode& program) {
+ constexpr std::size_t start_offset = 10;
+ std::size_t offset = start_offset;
+ std::size_t size = start_offset * sizeof(u64);
+ while (offset < program.size()) {
+ const u64 inst = program[offset];
+ if (!IsSchedInstruction(offset, start_offset)) {
+ if (inst == 0 || (inst >> 52) == 0x50b) {
+ break;
+ }
+ }
+ size += sizeof(inst);
+ offset++;
+ }
+ return size;
+}
+
+void CachedShader::CalculateProperties() {
+ setup.program.real_size = CalculateProgramSize(setup.program.code);
+ setup.program.real_size_b = 0;
+ setup.program.unique_identifier = Common::CityHash64(
+ reinterpret_cast<const char*>(setup.program.code.data()), setup.program.real_size);
+ if (program_type == Maxwell::ShaderProgram::VertexA) {
+ std::size_t seed = 0;
+ boost::hash_combine(seed, setup.program.unique_identifier);
+ setup.program.real_size_b = CalculateProgramSize(setup.program.code_b);
+ const u64 identifier_b = Common::CityHash64(
+ reinterpret_cast<const char*>(setup.program.code_b.data()), setup.program.real_size_b);
+ boost::hash_combine(seed, identifier_b);
+ setup.program.unique_identifier = static_cast<u64>(seed);
+ }
+}
+
ShaderCacheOpenGL::ShaderCacheOpenGL(RasterizerOpenGL& rasterizer) : RasterizerCache{rasterizer} {}
Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
diff --git a/src/video_core/renderer_opengl/gl_shader_cache.h b/src/video_core/renderer_opengl/gl_shader_cache.h
index 08f470de3..de3671acf 100644
--- a/src/video_core/renderer_opengl/gl_shader_cache.h
+++ b/src/video_core/renderer_opengl/gl_shader_cache.h
@@ -67,6 +67,7 @@ public:
6, "ShaderTrianglesAdjacency");
default:
UNREACHABLE_MSG("Unknown primitive mode.");
+ return LazyGeometryProgram(geometry_programs.points, "points", 1, "ShaderPoints");
}
}
@@ -81,6 +82,8 @@ private:
GLuint LazyGeometryProgram(OGLProgram& target_program, const std::string& glsl_topology,
u32 max_vertices, const std::string& debug_name);
+ void CalculateProperties();
+
VAddr addr;
std::size_t shader_length;
Maxwell::ShaderProgram program_type;
diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
index 4fc09cac6..4e685fa2c 100644
--- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
@@ -364,6 +364,7 @@ public:
return value;
default:
UNREACHABLE_MSG("Unimplemented conversion size: {}", static_cast<u32>(size));
+ return value;
}
}
@@ -626,6 +627,7 @@ public:
return "floatBitsToInt(" + value + ')';
} else {
UNREACHABLE();
+ return value;
}
}
@@ -928,7 +930,7 @@ private:
case Attribute::Index::FrontFacing:
// TODO(Subv): Find out what the values are for the other elements.
ASSERT(stage == Maxwell3D::Regs::ShaderStage::Fragment);
- return "vec4(0, 0, 0, uintBitsToFloat(gl_FrontFacing ? 1 : 0))";
+ return "vec4(0, 0, 0, intBitsToFloat(gl_FrontFacing ? -1 : 0))";
default:
const u32 index{static_cast<u32>(attribute) -
static_cast<u32>(Attribute::Index::Attribute_0)};
@@ -1591,23 +1593,21 @@ private:
process_mode == Tegra::Shader::TextureProcessMode::LL ||
process_mode == Tegra::Shader::TextureProcessMode::LLA;
+ // LOD selection (either via bias or explicit textureLod) not supported in GL for
+ // sampler2DArrayShadow and samplerCubeArrayShadow.
const bool gl_lod_supported = !(
(texture_type == Tegra::Shader::TextureType::Texture2D && is_array && depth_compare) ||
- (texture_type == Tegra::Shader::TextureType::TextureCube && !is_array &&
- depth_compare));
+ (texture_type == Tegra::Shader::TextureType::TextureCube && is_array && depth_compare));
const std::string read_method = lod_needed && gl_lod_supported ? "textureLod(" : "texture(";
std::string texture = read_method + sampler + ", coord";
- if (process_mode != Tegra::Shader::TextureProcessMode::None) {
+ UNIMPLEMENTED_IF(process_mode != Tegra::Shader::TextureProcessMode::None &&
+ !gl_lod_supported);
+
+ if (process_mode != Tegra::Shader::TextureProcessMode::None && gl_lod_supported) {
if (process_mode == Tegra::Shader::TextureProcessMode::LZ) {
- if (gl_lod_supported) {
- texture += ", 0";
- } else {
- // Lod 0 is emulated by a big negative bias
- // in scenarios that are not supported by glsl
- texture += ", -1000";
- }
+ texture += ", 0.0";
} else {
// If present, lod or bias are always stored in the register indexed by the
// gpr20
@@ -1645,15 +1645,15 @@ private:
if (depth_compare && !is_array && texture_type == Tegra::Shader::TextureType::Texture1D) {
coord += ",0.0";
}
+ if (is_array) {
+ coord += ',' + regs.GetRegisterAsInteger(array_register);
+ }
if (depth_compare) {
// Depth is always stored in the register signaled by gpr20
// or in the next register if lod or bias are used
const u64 depth_register = instr.gpr20.Value() + (lod_bias_enabled ? 1 : 0);
coord += ',' + regs.GetRegisterAsFloat(depth_register);
}
- if (is_array) {
- coord += ',' + regs.GetRegisterAsInteger(array_register);
- }
coord += ");";
return std::make_pair(
coord, GetTextureCode(instr, texture_type, process_mode, depth_compare, is_array, 0));
@@ -1681,20 +1681,20 @@ private:
for (size_t i = 0; i < coord_count; ++i) {
const bool last = (i == (coord_count - 1)) && (coord_count > 1);
coord += regs.GetRegisterAsFloat(last ? last_coord_register : coord_register + i);
- if (!last) {
+ if (i < coord_count - 1) {
coord += ',';
}
}
+ if (is_array) {
+ coord += ',' + regs.GetRegisterAsInteger(array_register);
+ }
if (depth_compare) {
// Depth is always stored in the register signaled by gpr20
// or in the next register if lod or bias are used
const u64 depth_register = instr.gpr20.Value() + (lod_bias_enabled ? 1 : 0);
coord += ',' + regs.GetRegisterAsFloat(depth_register);
}
- if (is_array) {
- coord += ',' + regs.GetRegisterAsInteger(array_register);
- }
coord += ");";
return std::make_pair(coord,
@@ -1702,6 +1702,99 @@ private:
is_array, (coord_count > 2 ? 1 : 0)));
}
+ std::pair<std::string, std::string> GetTLD4Code(const Instruction& instr,
+ const Tegra::Shader::TextureType texture_type,
+ const bool depth_compare, const bool is_array) {
+
+ const size_t coord_count = TextureCoordinates(texture_type);
+ const size_t total_coord_count = coord_count + (is_array ? 1 : 0);
+ const size_t total_reg_count = total_coord_count + (depth_compare ? 1 : 0);
+
+ constexpr std::array<const char*, 5> coord_container{
+ {"", "", "vec2 coord = vec2(", "vec3 coord = vec3(", "vec4 coord = vec4("}};
+
+ // If enabled arrays index is always stored in the gpr8 field
+ const u64 array_register = instr.gpr8.Value();
+ // First coordinate index is the gpr8 or gpr8 + 1 when arrays are used
+ const u64 coord_register = array_register + (is_array ? 1 : 0);
+
+ std::string coord = coord_container[total_coord_count];
+ for (size_t i = 0; i < coord_count;) {
+ coord += regs.GetRegisterAsFloat(coord_register + i);
+ ++i;
+ if (i != coord_count) {
+ coord += ',';
+ }
+ }
+
+ if (is_array) {
+ coord += ',' + regs.GetRegisterAsInteger(array_register);
+ }
+ coord += ");";
+
+ const std::string sampler =
+ GetSampler(instr.sampler, texture_type, is_array, depth_compare);
+
+ std::string texture = "textureGather(" + sampler + ", coord, ";
+ if (depth_compare) {
+ // Depth is always stored in the register signaled by gpr20
+ texture += regs.GetRegisterAsFloat(instr.gpr20.Value()) + ')';
+ } else {
+ texture += std::to_string(instr.tld4.component) + ')';
+ }
+ return std::make_pair(coord, texture);
+ }
+
+ std::pair<std::string, std::string> GetTLDSCode(const Instruction& instr,
+ const Tegra::Shader::TextureType texture_type,
+ const bool is_array) {
+
+ const size_t coord_count = TextureCoordinates(texture_type);
+ const size_t total_coord_count = coord_count + (is_array ? 1 : 0);
+ const bool lod_enabled =
+ instr.tlds.GetTextureProcessMode() == Tegra::Shader::TextureProcessMode::LL;
+
+ constexpr std::array<const char*, 4> coord_container{
+ {"", "int coord = (", "ivec2 coord = ivec2(", "ivec3 coord = ivec3("}};
+
+ std::string coord = coord_container[total_coord_count];
+
+ // If enabled arrays index is always stored in the gpr8 field
+ const u64 array_register = instr.gpr8.Value();
+
+ // if is array gpr20 is used
+ const u64 coord_register = is_array ? instr.gpr20.Value() : instr.gpr8.Value();
+
+ const u64 last_coord_register =
+ ((coord_count > 2) || (coord_count == 2 && !lod_enabled)) && !is_array
+ ? static_cast<u64>(instr.gpr20.Value())
+ : coord_register + 1;
+
+ for (size_t i = 0; i < coord_count; ++i) {
+ const bool last = (i == (coord_count - 1)) && (coord_count > 1);
+ coord += regs.GetRegisterAsInteger(last ? last_coord_register : coord_register + i);
+ if (i < coord_count - 1) {
+ coord += ',';
+ }
+ }
+ if (is_array) {
+ coord += ',' + regs.GetRegisterAsInteger(array_register);
+ }
+ coord += ");";
+
+ const std::string sampler = GetSampler(instr.sampler, texture_type, is_array, false);
+
+ std::string texture = "texelFetch(" + sampler + ", coords";
+
+ if (lod_enabled) {
+ // When lod is used always is in grp20
+ texture += ", " + regs.GetRegisterAsInteger(instr.gpr20) + ')';
+ } else {
+ texture += ", 0)";
+ }
+ return std::make_pair(coord, texture);
+ }
+
/**
* Compiles a single instruction from Tegra to GLSL.
* @param offset the offset of the Tegra shader instruction.
@@ -1774,9 +1867,6 @@ private:
UNIMPLEMENTED_IF_MSG(instr.fmul.tab5cb8_2 != 0,
"FMUL tab5cb8_2({}) is not implemented",
instr.fmul.tab5cb8_2.Value());
- UNIMPLEMENTED_IF_MSG(instr.fmul.tab5c68_1 != 0,
- "FMUL tab5cb8_1({}) is not implemented",
- instr.fmul.tab5c68_1.Value());
UNIMPLEMENTED_IF_MSG(
instr.fmul.tab5c68_0 != 1, "FMUL tab5cb8_0({}) is not implemented",
instr.fmul.tab5c68_0
@@ -1786,7 +1876,26 @@ private:
op_b = GetOperandAbsNeg(op_b, false, instr.fmul.negate_b);
- regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " * " + op_b, 1, 1,
+ std::string postfactor_op;
+ if (instr.fmul.postfactor != 0) {
+ s8 postfactor = static_cast<s8>(instr.fmul.postfactor);
+
+ // postfactor encoded as 3-bit 1's complement in instruction,
+ // interpreted with below logic.
+ if (postfactor >= 4) {
+ postfactor = 7 - postfactor;
+ } else {
+ postfactor = 0 - postfactor;
+ }
+
+ if (postfactor > 0) {
+ postfactor_op = " * " + std::to_string(1 << postfactor);
+ } else {
+ postfactor_op = " / " + std::to_string(1 << -postfactor);
+ }
+ }
+
+ regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " * " + op_b + postfactor_op, 1, 1,
instr.alu.saturate_d, 0, true);
break;
}
@@ -1955,6 +2064,8 @@ private:
std::to_string(instr.alu.GetSignedImm20_20())};
default:
UNREACHABLE();
+ return {regs.GetRegisterAsInteger(instr.gpr39, 0, false),
+ std::to_string(instr.alu.GetSignedImm20_20())};
}
}();
const std::string offset = '(' + packed_shift + " & 0xff)";
@@ -2825,9 +2936,6 @@ private:
const Tegra::Shader::TextureType texture_type{instr.tlds.GetTextureType()};
const bool is_array{instr.tlds.IsArrayTexture()};
- ASSERT(texture_type == Tegra::Shader::TextureType::Texture2D);
- ASSERT(is_array == false);
-
UNIMPLEMENTED_IF_MSG(instr.tlds.UsesMiscMode(Tegra::Shader::TextureMiscMode::NODEP),
"NODEP is not implemented");
UNIMPLEMENTED_IF_MSG(instr.tlds.UsesMiscMode(Tegra::Shader::TextureMiscMode::AOFFI),
@@ -2835,54 +2943,16 @@ private:
UNIMPLEMENTED_IF_MSG(instr.tlds.UsesMiscMode(Tegra::Shader::TextureMiscMode::MZ),
"MZ is not implemented");
- u32 extra_op_offset = 0;
-
- ShaderScopedScope scope = shader.Scope();
-
- switch (texture_type) {
- case Tegra::Shader::TextureType::Texture1D: {
- const std::string x = regs.GetRegisterAsInteger(instr.gpr8);
- shader.AddLine("float coords = " + x + ';');
- break;
- }
- case Tegra::Shader::TextureType::Texture2D: {
- UNIMPLEMENTED_IF_MSG(is_array, "Unhandled 2d array texture");
-
- const std::string x = regs.GetRegisterAsInteger(instr.gpr8);
- const std::string y = regs.GetRegisterAsInteger(instr.gpr20);
- // shader.AddLine("ivec2 coords = ivec2(" + x + ", " + y + ");");
- shader.AddLine("ivec2 coords = ivec2(" + x + ", " + y + ");");
- extra_op_offset = 1;
- break;
- }
- default:
- UNIMPLEMENTED_MSG("Unhandled texture type {}", static_cast<u32>(texture_type));
- }
- const std::string sampler =
- GetSampler(instr.sampler, texture_type, is_array, false);
+ const auto [coord, texture] = GetTLDSCode(instr, texture_type, is_array);
- const std::string texture = [&]() {
- switch (instr.tlds.GetTextureProcessMode()) {
- case Tegra::Shader::TextureProcessMode::LZ:
- return "texelFetch(" + sampler + ", coords, 0)";
- case Tegra::Shader::TextureProcessMode::LL:
- shader.AddLine(
- "float lod = " +
- regs.GetRegisterAsInteger(instr.gpr20.Value() + extra_op_offset) + ';');
- return "texelFetch(" + sampler + ", coords, lod)";
- default:
- UNIMPLEMENTED_MSG("Unhandled texture process mode {}",
- static_cast<u32>(instr.tlds.GetTextureProcessMode()));
- return "texelFetch(" + sampler + ", coords, 0)";
- }
- }();
+ const auto scope = shader.Scope();
- WriteTexsInstructionFloat(instr, texture);
+ shader.AddLine(coord);
+ shader.AddLine("vec4 texture_tmp = " + texture + ';');
+ WriteTexsInstructionFloat(instr, "texture_tmp");
break;
}
case OpCode::Id::TLD4: {
- ASSERT(instr.tld4.texture_type == Tegra::Shader::TextureType::Texture2D);
- ASSERT(instr.tld4.array == 0);
UNIMPLEMENTED_IF_MSG(instr.tld4.UsesMiscMode(Tegra::Shader::TextureMiscMode::NODEP),
"NODEP is not implemented");
@@ -2892,56 +2962,29 @@ private:
"NDV is not implemented");
UNIMPLEMENTED_IF_MSG(instr.tld4.UsesMiscMode(Tegra::Shader::TextureMiscMode::PTP),
"PTP is not implemented");
+
+ auto texture_type = instr.tld4.texture_type.Value();
const bool depth_compare =
instr.tld4.UsesMiscMode(Tegra::Shader::TextureMiscMode::DC);
- auto texture_type = instr.tld4.texture_type.Value();
- u32 num_coordinates = TextureCoordinates(texture_type);
- if (depth_compare)
- num_coordinates += 1;
-
- const auto scope = shader.Scope();
+ const bool is_array = instr.tld4.array != 0;
- switch (num_coordinates) {
- case 2: {
- const std::string x = regs.GetRegisterAsFloat(instr.gpr8);
- const std::string y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1);
- shader.AddLine("vec2 coords = vec2(" + x + ", " + y + ");");
- break;
- }
- case 3: {
- const std::string x = regs.GetRegisterAsFloat(instr.gpr8);
- const std::string y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1);
- const std::string z = regs.GetRegisterAsFloat(instr.gpr8.Value() + 2);
- shader.AddLine("vec3 coords = vec3(" + x + ", " + y + ", " + z + ");");
- break;
- }
- default:
- UNIMPLEMENTED_MSG("Unhandled coordinates number {}",
- static_cast<u32>(num_coordinates));
- const std::string x = regs.GetRegisterAsFloat(instr.gpr8);
- const std::string y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1);
- shader.AddLine("vec2 coords = vec2(" + x + ", " + y + ");");
- texture_type = Tegra::Shader::TextureType::Texture2D;
- }
+ const auto [coord, texture] =
+ GetTLD4Code(instr, texture_type, depth_compare, is_array);
- const std::string sampler =
- GetSampler(instr.sampler, texture_type, false, depth_compare);
+ const auto scope = shader.Scope();
- const std::string texture = "textureGather(" + sampler + ", coords, " +
- std::to_string(instr.tld4.component) + ')';
+ shader.AddLine(coord);
+ std::size_t dest_elem{};
- if (depth_compare) {
- regs.SetRegisterToFloat(instr.gpr0, 0, texture, 1, 1, false);
- } else {
- std::size_t dest_elem{};
- for (std::size_t elem = 0; elem < 4; ++elem) {
- if (!instr.tex.IsComponentEnabled(elem)) {
- // Skip disabled components
- continue;
- }
- regs.SetRegisterToFloat(instr.gpr0, elem, texture, 1, 4, false, dest_elem);
- ++dest_elem;
+ shader.AddLine("vec4 texture_tmp = " + texture + ';');
+ for (std::size_t elem = 0; elem < 4; ++elem) {
+ if (!instr.tex.IsComponentEnabled(elem)) {
+ // Skip disabled components
+ continue;
}
+ regs.SetRegisterToFloat(instr.gpr0, elem, "texture_tmp", 1, 4, false,
+ dest_elem);
+ ++dest_elem;
}
break;
}
@@ -2955,28 +2998,31 @@ private:
const auto scope = shader.Scope();
+ std::string coords;
+
const bool depth_compare =
instr.tld4s.UsesMiscMode(Tegra::Shader::TextureMiscMode::DC);
- const std::string op_a = regs.GetRegisterAsFloat(instr.gpr8);
- const std::string op_b = regs.GetRegisterAsFloat(instr.gpr20);
- // TODO(Subv): Figure out how the sampler type is encoded in the TLD4S instruction.
+
const std::string sampler = GetSampler(
instr.sampler, Tegra::Shader::TextureType::Texture2D, false, depth_compare);
- if (depth_compare) {
- // Note: TLD4S coordinate encoding works just like TEXS's
- const std::string op_y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1);
- shader.AddLine("vec3 coords = vec3(" + op_a + ", " + op_y + ", " + op_b + ");");
- } else {
- shader.AddLine("vec2 coords = vec2(" + op_a + ", " + op_b + ");");
- }
- std::string texture = "textureGather(" + sampler + ", coords, " +
- std::to_string(instr.tld4s.component) + ')';
- if (depth_compare) {
- texture = "vec4(" + texture + ')';
- }
+ const std::string op_a = regs.GetRegisterAsFloat(instr.gpr8);
+ coords = "vec2 coords = vec2(" + op_a + ", ";
+ std::string texture = "textureGather(" + sampler + ", coords, ";
- WriteTexsInstructionFloat(instr, texture);
+ if (!depth_compare) {
+ const std::string op_b = regs.GetRegisterAsFloat(instr.gpr20);
+ coords += op_b + ");";
+ texture += std::to_string(instr.tld4s.component) + ')';
+ } else {
+ const std::string op_b = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1);
+ const std::string op_c = regs.GetRegisterAsFloat(instr.gpr20);
+ coords += op_b + ");";
+ texture += op_c + ')';
+ }
+ shader.AddLine(coords);
+ shader.AddLine("vec4 texture_tmp = " + texture + ';');
+ WriteTexsInstructionFloat(instr, "texture_tmp");
break;
}
case OpCode::Id::TXQ: {
@@ -3270,6 +3316,7 @@ private:
return std::to_string(instr.r2p.immediate_mask);
default:
UNREACHABLE();
+ return std::to_string(instr.r2p.immediate_mask);
}
}();
const std::string mask = '(' + regs.GetRegisterAsInteger(instr.gpr8, 0, false) +
@@ -3733,7 +3780,10 @@ private:
}
break;
}
- default: { UNIMPLEMENTED_MSG("Unhandled instruction: {}", opcode->get().GetName()); }
+ default: {
+ UNIMPLEMENTED_MSG("Unhandled instruction: {}", opcode->get().GetName());
+ break;
+ }
}
break;
@@ -3888,4 +3938,4 @@ std::optional<ProgramResult> DecompileProgram(const ProgramCode& program_code, u
return {};
}
-} // namespace OpenGL::GLShader::Decompiler \ No newline at end of file
+} // namespace OpenGL::GLShader::Decompiler
diff --git a/src/video_core/renderer_opengl/gl_shader_gen.cpp b/src/video_core/renderer_opengl/gl_shader_gen.cpp
index 23ed91e27..5d0819dc5 100644
--- a/src/video_core/renderer_opengl/gl_shader_gen.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_gen.cpp
@@ -2,6 +2,7 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
+#include <fmt/format.h>
#include "common/assert.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/renderer_opengl/gl_shader_decompiler.h"
@@ -16,6 +17,8 @@ static constexpr u32 PROGRAM_OFFSET{10};
ProgramResult GenerateVertexShader(const ShaderSetup& setup) {
std::string out = "#version 430 core\n";
out += "#extension GL_ARB_separate_shader_objects : enable\n\n";
+ const std::string id = fmt::format("{:016x}", setup.program.unique_identifier);
+ out += "// Shader Unique Id: VS" + id + "\n\n";
out += Decompiler::GetCommonDeclarations();
out += R"(
@@ -84,6 +87,8 @@ void main() {
ProgramResult GenerateGeometryShader(const ShaderSetup& setup) {
// Version is intentionally skipped in shader generation, it's added by the lazy compilation.
std::string out = "#extension GL_ARB_separate_shader_objects : enable\n\n";
+ const std::string id = fmt::format("{:016x}", setup.program.unique_identifier);
+ out += "// Shader Unique Id: GS" + id + "\n\n";
out += Decompiler::GetCommonDeclarations();
out += "bool exec_geometry();\n";
@@ -117,6 +122,8 @@ void main() {
ProgramResult GenerateFragmentShader(const ShaderSetup& setup) {
std::string out = "#version 430 core\n";
out += "#extension GL_ARB_separate_shader_objects : enable\n\n";
+ const std::string id = fmt::format("{:016x}", setup.program.unique_identifier);
+ out += "// Shader Unique Id: FS" + id + "\n\n";
out += Decompiler::GetCommonDeclarations();
out += "bool exec_fragment();\n";
diff --git a/src/video_core/renderer_opengl/gl_shader_gen.h b/src/video_core/renderer_opengl/gl_shader_gen.h
index 4fa6d7612..fcc20d3b4 100644
--- a/src/video_core/renderer_opengl/gl_shader_gen.h
+++ b/src/video_core/renderer_opengl/gl_shader_gen.h
@@ -177,6 +177,9 @@ struct ShaderSetup {
struct {
ProgramCode code;
ProgramCode code_b; // Used for dual vertex shaders
+ u64 unique_identifier;
+ std::size_t real_size;
+ std::size_t real_size_b;
} program;
/// Used in scenarios where we have a dual vertex shaders
diff --git a/src/video_core/renderer_opengl/renderer_opengl.cpp b/src/video_core/renderer_opengl/renderer_opengl.cpp
index 49a1989e4..235732d86 100644
--- a/src/video_core/renderer_opengl/renderer_opengl.cpp
+++ b/src/video_core/renderer_opengl/renderer_opengl.cpp
@@ -462,6 +462,7 @@ static const char* GetSource(GLenum source) {
RET(OTHER);
default:
UNREACHABLE();
+ return "Unknown source";
}
#undef RET
}
@@ -480,6 +481,7 @@ static const char* GetType(GLenum type) {
RET(MARKER);
default:
UNREACHABLE();
+ return "Unknown type";
}
#undef RET
}
diff --git a/src/video_core/surface.cpp b/src/video_core/surface.cpp
index 9582dd2ca..a97b1562b 100644
--- a/src/video_core/surface.cpp
+++ b/src/video_core/surface.cpp
@@ -65,6 +65,7 @@ PixelFormat PixelFormatFromDepthFormat(Tegra::DepthFormat format) {
default:
LOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
UNREACHABLE();
+ return PixelFormat::S8Z24;
}
}
@@ -141,6 +142,7 @@ PixelFormat PixelFormatFromRenderTargetFormat(Tegra::RenderTargetFormat format)
default:
LOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
UNREACHABLE();
+ return PixelFormat::RGBA8_SRGB;
}
}
@@ -327,6 +329,7 @@ PixelFormat PixelFormatFromTextureFormat(Tegra::Texture::TextureFormat format,
LOG_CRITICAL(HW_GPU, "Unimplemented format={}, component_type={}", static_cast<u32>(format),
static_cast<u32>(component_type));
UNREACHABLE();
+ return PixelFormat::ABGR8U;
}
}
@@ -346,6 +349,7 @@ ComponentType ComponentTypeFromTexture(Tegra::Texture::ComponentType type) {
default:
LOG_CRITICAL(HW_GPU, "Unimplemented component type={}", static_cast<u32>(type));
UNREACHABLE();
+ return ComponentType::UNorm;
}
}
@@ -393,6 +397,7 @@ ComponentType ComponentTypeFromRenderTarget(Tegra::RenderTargetFormat format) {
default:
LOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
UNREACHABLE();
+ return ComponentType::UNorm;
}
}
@@ -403,6 +408,7 @@ PixelFormat PixelFormatFromGPUPixelFormat(Tegra::FramebufferConfig::PixelFormat
default:
LOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
UNREACHABLE();
+ return PixelFormat::ABGR8U;
}
}
@@ -418,6 +424,7 @@ ComponentType ComponentTypeFromDepthFormat(Tegra::DepthFormat format) {
default:
LOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
UNREACHABLE();
+ return ComponentType::UNorm;
}
}
diff --git a/src/video_core/textures/decoders.cpp b/src/video_core/textures/decoders.cpp
index bbae9285f..5db75de22 100644
--- a/src/video_core/textures/decoders.cpp
+++ b/src/video_core/textures/decoders.cpp
@@ -226,7 +226,7 @@ u32 BytesPerPixel(TextureFormat format) {
return 8;
default:
UNIMPLEMENTED_MSG("Format not implemented");
- break;
+ return 1;
}
}
generated by cgit v1.2.3 (git 2.25.1) at 2025-12-31 07:44:57 +0000