summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorYuri Kunde Schlesner <yuriks@yuriks.net>2017-01-27 18:29:10 +0100
committerGitHub <noreply@github.com>2017-01-27 18:29:10 +0100
commitbf14f4be2263b4769e97800b35951717192c2d1c (patch)
tree9c1c47f5a05e9907257f620d8426a0cebaf0cf78
parentSDL: Select audio device (#2403) (diff)
parentVideoCore/Shader: Move entry_point to SetupBatch (diff)
downloadyuzu-bf14f4be2263b4769e97800b35951717192c2d1c.tar
yuzu-bf14f4be2263b4769e97800b35951717192c2d1c.tar.gz
yuzu-bf14f4be2263b4769e97800b35951717192c2d1c.tar.bz2
yuzu-bf14f4be2263b4769e97800b35951717192c2d1c.tar.lz
yuzu-bf14f4be2263b4769e97800b35951717192c2d1c.tar.xz
yuzu-bf14f4be2263b4769e97800b35951717192c2d1c.tar.zst
yuzu-bf14f4be2263b4769e97800b35951717192c2d1c.zip
-rw-r--r--src/citra_qt/debugger/graphics/graphics_vertex_shader.cpp7
-rw-r--r--src/citra_qt/debugger/graphics/graphics_vertex_shader.h1
-rw-r--r--src/video_core/CMakeLists.txt6
-rw-r--r--src/video_core/command_processor.cpp22
-rw-r--r--src/video_core/pica.cpp2
-rw-r--r--src/video_core/shader/shader.cpp102
-rw-r--r--src/video_core/shader/shader.h70
-rw-r--r--src/video_core/shader/shader_interpreter.cpp49
-rw-r--r--src/video_core/shader/shader_interpreter.h26
-rw-r--r--src/video_core/shader/shader_jit_x64.cpp890
-rw-r--r--src/video_core/shader/shader_jit_x64.h115
-rw-r--r--src/video_core/shader/shader_jit_x64_compiler.cpp884
-rw-r--r--src/video_core/shader/shader_jit_x64_compiler.h125
13 files changed, 1189 insertions, 1110 deletions
diff --git a/src/citra_qt/debugger/graphics/graphics_vertex_shader.cpp b/src/citra_qt/debugger/graphics/graphics_vertex_shader.cpp
index ff2e7e363..f37524190 100644
--- a/src/citra_qt/debugger/graphics/graphics_vertex_shader.cpp
+++ b/src/citra_qt/debugger/graphics/graphics_vertex_shader.cpp
@@ -18,7 +18,9 @@
#include "citra_qt/util/util.h"
#include "video_core/pica.h"
#include "video_core/pica_state.h"
+#include "video_core/shader/debug_data.h"
#include "video_core/shader/shader.h"
+#include "video_core/shader/shader_interpreter.h"
using nihstro::OpCode;
using nihstro::Instruction;
@@ -518,8 +520,9 @@ void GraphicsVertexShaderWidget::Reload(bool replace_vertex_data, void* vertex_d
info.labels.insert({entry_point, "main"});
// Generate debug information
- debug_data = Pica::g_state.vs.ProduceDebugInfo(input_vertex, num_attributes, shader_config,
- shader_setup);
+ Pica::Shader::InterpreterEngine shader_engine;
+ shader_engine.SetupBatch(shader_setup, entry_point);
+ debug_data = shader_engine.ProduceDebugInfo(shader_setup, input_vertex, num_attributes);
// Reload widget state
for (int attr = 0; attr < num_attributes; ++attr) {
diff --git a/src/citra_qt/debugger/graphics/graphics_vertex_shader.h b/src/citra_qt/debugger/graphics/graphics_vertex_shader.h
index bedea0bed..3292573f3 100644
--- a/src/citra_qt/debugger/graphics/graphics_vertex_shader.h
+++ b/src/citra_qt/debugger/graphics/graphics_vertex_shader.h
@@ -8,6 +8,7 @@
#include <QTreeView>
#include "citra_qt/debugger/graphics/graphics_breakpoint_observer.h"
#include "nihstro/parser_shbin.h"
+#include "video_core/shader/debug_data.h"
#include "video_core/shader/shader.h"
class QLabel;
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index 6ca319b59..d55b84ce0 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -50,10 +50,12 @@ set(HEADERS
if(ARCHITECTURE_x86_64)
set(SRCS ${SRCS}
- shader/shader_jit_x64.cpp)
+ shader/shader_jit_x64.cpp
+ shader/shader_jit_x64_compiler.cpp)
set(HEADERS ${HEADERS}
- shader/shader_jit_x64.h)
+ shader/shader_jit_x64.h
+ shader/shader_jit_x64_compiler.h)
endif()
create_directory_groups(${SRCS} ${HEADERS})
diff --git a/src/video_core/command_processor.cpp b/src/video_core/command_processor.cpp
index ea58e9f54..eb79974a8 100644
--- a/src/video_core/command_processor.cpp
+++ b/src/video_core/command_processor.cpp
@@ -142,16 +142,18 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
MICROPROFILE_SCOPE(GPU_Drawing);
immediate_attribute_id = 0;
- Shader::UnitState shader_unit;
- g_state.vs.Setup();
+ auto* shader_engine = Shader::GetEngine();
+ shader_engine->SetupBatch(g_state.vs, regs.vs.main_offset);
// Send to vertex shader
if (g_debug_context)
g_debug_context->OnEvent(DebugContext::Event::VertexShaderInvocation,
static_cast<void*>(&immediate_input));
- g_state.vs.Run(shader_unit, immediate_input, regs.vs.num_input_attributes + 1);
- Shader::OutputVertex output_vertex =
- shader_unit.output_registers.ToVertex(regs.vs);
+ Shader::UnitState shader_unit;
+ shader_unit.LoadInputVertex(immediate_input, regs.vs.num_input_attributes + 1);
+ shader_engine->Run(g_state.vs, shader_unit);
+ auto output_vertex = Shader::OutputVertex::FromRegisters(
+ shader_unit.registers.output, regs, regs.vs.output_mask);
// Send to renderer
using Pica::Shader::OutputVertex;
@@ -243,8 +245,10 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
unsigned int vertex_cache_pos = 0;
vertex_cache_ids.fill(-1);
+ auto* shader_engine = Shader::GetEngine();
Shader::UnitState shader_unit;
- g_state.vs.Setup();
+
+ shader_engine->SetupBatch(g_state.vs, regs.vs.main_offset);
for (unsigned int index = 0; index < regs.num_vertices; ++index) {
// Indexed rendering doesn't use the start offset
@@ -283,10 +287,12 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
if (g_debug_context)
g_debug_context->OnEvent(DebugContext::Event::VertexShaderInvocation,
(void*)&input);
- g_state.vs.Run(shader_unit, input, loader.GetNumTotalAttributes());
+ shader_unit.LoadInputVertex(input, loader.GetNumTotalAttributes());
+ shader_engine->Run(g_state.vs, shader_unit);
// Retrieve vertex from register data
- output_vertex = shader_unit.output_registers.ToVertex(regs.vs);
+ output_vertex = Shader::OutputVertex::FromRegisters(shader_unit.registers.output,
+ regs, regs.vs.output_mask);
if (is_indexed) {
vertex_cache[vertex_cache_pos] = output_vertex;
diff --git a/src/video_core/pica.cpp b/src/video_core/pica.cpp
index ce2bd455e..b4a77c632 100644
--- a/src/video_core/pica.cpp
+++ b/src/video_core/pica.cpp
@@ -499,7 +499,7 @@ void Init() {
}
void Shutdown() {
- Shader::ClearCache();
+ Shader::Shutdown();
}
template <typename T>
diff --git a/src/video_core/shader/shader.cpp b/src/video_core/shader/shader.cpp
index 7ae57e619..2da50bd62 100644
--- a/src/video_core/shader/shader.cpp
+++ b/src/video_core/shader/shader.cpp
@@ -2,14 +2,8 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
-#include <atomic>
#include <cmath>
#include <cstring>
-#include <unordered_map>
-#include <utility>
-#include <boost/range/algorithm/fill.hpp>
-#include "common/bit_field.h"
-#include "common/hash.h"
#include "common/logging/log.h"
#include "common/microprofile.h"
#include "video_core/pica.h"
@@ -25,7 +19,8 @@ namespace Pica {
namespace Shader {
-OutputVertex OutputRegisters::ToVertex(const Regs::ShaderConfig& config) const {
+OutputVertex OutputVertex::FromRegisters(Math::Vec4<float24> output_regs[16], const Regs& regs,
+ u32 output_mask) {
// Setup output data
OutputVertex ret;
// TODO(neobrain): Under some circumstances, up to 16 attributes may be output. We need to
@@ -33,13 +28,13 @@ OutputVertex OutputRegisters::ToVertex(const Regs::ShaderConfig& config) const {
unsigned index = 0;
for (unsigned i = 0; i < 7; ++i) {
- if (index >= g_state.regs.vs_output_total)
+ if (index >= regs.vs_output_total)
break;
- if ((config.output_mask & (1 << i)) == 0)
+ if ((output_mask & (1 << i)) == 0)
continue;
- const auto& output_register_map = g_state.regs.vs_output_attributes[index];
+ const auto& output_register_map = regs.vs_output_attributes[index];
u32 semantics[4] = {output_register_map.map_x, output_register_map.map_y,
output_register_map.map_z, output_register_map.map_w};
@@ -47,7 +42,7 @@ OutputVertex OutputRegisters::ToVertex(const Regs::ShaderConfig& config) const {
for (unsigned comp = 0; comp < 4; ++comp) {
float24* out = ((float24*)&ret) + semantics[comp];
if (semantics[comp] != Regs::VSOutputAttributes::INVALID) {
- *out = value[i][comp];
+ *out = output_regs[i][comp];
} else {
// Zero output so that attributes which aren't output won't have denormals in them,
// which would slow us down later.
@@ -76,86 +71,41 @@ OutputVertex OutputRegisters::ToVertex(const Regs::ShaderConfig& config) const {
return ret;
}
-#ifdef ARCHITECTURE_x86_64
-static std::unordered_map<u64, std::unique_ptr<JitShader>> shader_map;
-static const JitShader* jit_shader;
-#endif // ARCHITECTURE_x86_64
+void UnitState::LoadInputVertex(const InputVertex& input, int num_attributes) {
+ // Setup input register table
+ const auto& attribute_register_map = g_state.regs.vs.input_register_map;
+
+ for (int i = 0; i < num_attributes; i++)
+ registers.input[attribute_register_map.GetRegisterForAttribute(i)] = input.attr[i];
+}
+
+MICROPROFILE_DEFINE(GPU_Shader, "GPU", "Shader", MP_RGB(50, 50, 240));
-void ClearCache() {
#ifdef ARCHITECTURE_x86_64
- shader_map.clear();
+static std::unique_ptr<JitX64Engine> jit_engine;
#endif // ARCHITECTURE_x86_64
-}
+static InterpreterEngine interpreter_engine;
-void ShaderSetup::Setup() {
+ShaderEngine* GetEngine() {
#ifdef ARCHITECTURE_x86_64
+ // TODO(yuriks): Re-initialize on each change rather than being persistent
if (VideoCore::g_shader_jit_enabled) {
- u64 cache_key =
- Common::ComputeHash64(&g_state.vs.program_code, sizeof(g_state.vs.program_code)) ^
- Common::ComputeHash64(&g_state.vs.swizzle_data, sizeof(g_state.vs.swizzle_data));
-
- auto iter = shader_map.find(cache_key);
- if (iter != shader_map.end()) {
- jit_shader = iter->second.get();
- } else {
- auto shader = std::make_unique<JitShader>();
- shader->Compile();
- jit_shader = shader.get();
- shader_map[cache_key] = std::move(shader);
+ if (jit_engine == nullptr) {
+ jit_engine = std::make_unique<JitX64Engine>();
}
+ return jit_engine.get();
}
#endif // ARCHITECTURE_x86_64
-}
-
-MICROPROFILE_DEFINE(GPU_Shader, "GPU", "Shader", MP_RGB(50, 50, 240));
-
-void ShaderSetup::Run(UnitState& state, const InputVertex& input, int num_attributes) {
- auto& config = g_state.regs.vs;
- auto& setup = g_state.vs;
-
- MICROPROFILE_SCOPE(GPU_Shader);
- // Setup input register table
- const auto& attribute_register_map = config.input_register_map;
-
- for (int i = 0; i < num_attributes; i++)
- state.registers.input[attribute_register_map.GetRegisterForAttribute(i)] = input.attr[i];
-
- state.conditional_code[0] = false;
- state.conditional_code[1] = false;
+ return &interpreter_engine;
+}
+void Shutdown() {
#ifdef ARCHITECTURE_x86_64
- if (VideoCore::g_shader_jit_enabled) {
- jit_shader->Run(setup, state, config.main_offset);
- } else {
- DebugData<false> dummy_debug_data;
- RunInterpreter(setup, state, dummy_debug_data, config.main_offset);
- }
-#else
- DebugData<false> dummy_debug_data;
- RunInterpreter(setup, state, dummy_debug_data, config.main_offset);
+ jit_engine = nullptr;
#endif // ARCHITECTURE_x86_64
}
-DebugData<true> ShaderSetup::ProduceDebugInfo(const InputVertex& input, int num_attributes,
- const Regs::ShaderConfig& config,
- const ShaderSetup& setup) {
- UnitState state;
- DebugData<true> debug_data;
-
- // Setup input register table
- boost::fill(state.registers.input, Math::Vec4<float24>::AssignToAll(float24::Zero()));
- const auto& attribute_register_map = config.input_register_map;
- for (int i = 0; i < num_attributes; i++)
- state.registers.input[attribute_register_map.GetRegisterForAttribute(i)] = input.attr[i];
-
- state.conditional_code[0] = false;
- state.conditional_code[1] = false;
-
- RunInterpreter(setup, state, debug_data, config.main_offset);
- return debug_data;
-}
-
} // namespace Shader
} // namespace Pica
diff --git a/src/video_core/shader/shader.h b/src/video_core/shader/shader.h
index 2b07759b9..44d9f76c3 100644
--- a/src/video_core/shader/shader.h
+++ b/src/video_core/shader/shader.h
@@ -6,7 +6,6 @@
#include <array>
#include <cstddef>
-#include <memory>
#include <type_traits>
#include <nihstro/shader_bytecode.h>
#include "common/assert.h"
@@ -15,7 +14,6 @@
#include "common/vector_math.h"
#include "video_core/pica.h"
#include "video_core/pica_types.h"
-#include "video_core/shader/debug_data.h"
using nihstro::RegisterType;
using nihstro::SourceRegister;
@@ -75,19 +73,13 @@ struct OutputVertex {
ret.Lerp(factor, v1);
return ret;
}
+
+ static OutputVertex FromRegisters(Math::Vec4<float24> output_regs[16], const Regs& regs,
+ u32 output_mask);
};
static_assert(std::is_pod<OutputVertex>::value, "Structure is not POD");
static_assert(sizeof(OutputVertex) == 32 * sizeof(float), "OutputVertex has invalid size");
-struct OutputRegisters {
- OutputRegisters() = default;
-
- alignas(16) Math::Vec4<float24> value[16];
-
- OutputVertex ToVertex(const Regs::ShaderConfig& config) const;
-};
-static_assert(std::is_pod<OutputRegisters>::value, "Structure is not POD");
-
/**
* This structure contains the state information that needs to be unique for a shader unit. The 3DS
* has four shader units that process shaders in parallel. At the present, Citra only implements a
@@ -100,11 +92,10 @@ struct UnitState {
// required to be 16-byte aligned.
alignas(16) Math::Vec4<float24> input[16];
alignas(16) Math::Vec4<float24> temporary[16];
+ alignas(16) Math::Vec4<float24> output[16];
} registers;
static_assert(std::is_pod<Registers>::value, "Structure is not POD");
- OutputRegisters output_registers;
-
bool conditional_code[2];
// Two Address registers and one loop counter
@@ -130,7 +121,7 @@ struct UnitState {
static size_t OutputOffset(const DestRegister& reg) {
switch (reg.GetRegisterType()) {
case RegisterType::Output:
- return offsetof(UnitState, output_registers.value) +
+ return offsetof(UnitState, registers.output) +
reg.GetIndex() * sizeof(Math::Vec4<float24>);
case RegisterType::Temporary:
@@ -142,13 +133,17 @@ struct UnitState {
return 0;
}
}
-};
-/// Clears the shader cache
-void ClearCache();
+ /**
+ * Loads the unit state with an input vertex.
+ *
+ * @param input Input vertex into the shader
+ * @param num_attributes The number of vertex shader attributes to load
+ */
+ void LoadInputVertex(const InputVertex& input, int num_attributes);
+};
struct ShaderSetup {
-
struct {
// The float uniforms are accessed by the shader JIT using SSE instructions, and are
// therefore required to be 16-byte aligned.
@@ -173,32 +168,37 @@ struct ShaderSetup {
std::array<u32, 1024> program_code;
std::array<u32, 1024> swizzle_data;
+ /// Data private to ShaderEngines
+ struct EngineData {
+ unsigned int entry_point;
+ /// Used by the JIT, points to a compiled shader object.
+ const void* cached_shader = nullptr;
+ } engine_data;
+};
+
+class ShaderEngine {
+public:
+ virtual ~ShaderEngine() = default;
+
/**
* Performs any shader unit setup that only needs to happen once per shader (as opposed to once
* per vertex, which would happen within the `Run` function).
*/
- void Setup();
-
- /**
- * Runs the currently setup shader
- * @param state Shader unit state, must be setup per shader and per shader unit
- * @param input Input vertex into the shader
- * @param num_attributes The number of vertex shader attributes
- */
- void Run(UnitState& state, const InputVertex& input, int num_attributes);
+ virtual void SetupBatch(ShaderSetup& setup, unsigned int entry_point) = 0;
/**
- * Produce debug information based on the given shader and input vertex
- * @param input Input vertex into the shader
- * @param num_attributes The number of vertex shader attributes
- * @param config Configuration object for the shader pipeline
- * @param setup Setup object for the shader pipeline
- * @return Debug information for this shader with regards to the given vertex
+ * Runs the currently setup shader.
+ *
+ * @param setup Shader engine state, must be setup with SetupBatch on each shader change.
+ * @param state Shader unit state, must be setup with input data before each shader invocation.
*/
- DebugData<true> ProduceDebugInfo(const InputVertex& input, int num_attributes,
- const Regs::ShaderConfig& config, const ShaderSetup& setup);
+ virtual void Run(const ShaderSetup& setup, UnitState& state) const = 0;
};
+// TODO(yuriks): Remove and make it non-global state somewhere
+ShaderEngine* GetEngine();
+void Shutdown();
+
} // namespace Shader
} // namespace Pica
diff --git a/src/video_core/shader/shader_interpreter.cpp b/src/video_core/shader/shader_interpreter.cpp
index 20fb9754b..c0c89b857 100644
--- a/src/video_core/shader/shader_interpreter.cpp
+++ b/src/video_core/shader/shader_interpreter.cpp
@@ -7,10 +7,12 @@
#include <cmath>
#include <numeric>
#include <boost/container/static_vector.hpp>
+#include <boost/range/algorithm/fill.hpp>
#include <nihstro/shader_bytecode.h>
#include "common/assert.h"
#include "common/common_types.h"
#include "common/logging/log.h"
+#include "common/microprofile.h"
#include "common/vector_math.h"
#include "video_core/pica_state.h"
#include "video_core/pica_types.h"
@@ -37,12 +39,15 @@ struct CallStackElement {
};
template <bool Debug>
-void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData<Debug>& debug_data,
- unsigned offset) {
+static void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData<Debug>& debug_data,
+ unsigned offset) {
// TODO: Is there a maximal size for this?
boost::container::static_vector<CallStackElement, 16> call_stack;
u32 program_counter = offset;
+ state.conditional_code[0] = false;
+ state.conditional_code[1] = false;
+
auto call = [&program_counter, &call_stack](u32 offset, u32 num_instructions, u32 return_offset,
u8 repeat_count, u8 loop_increment) {
// -1 to make sure when incrementing the PC we end up at the correct offset
@@ -73,9 +78,9 @@ void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData<Debug>
}
};
- const auto& uniforms = g_state.vs.uniforms;
- const auto& swizzle_data = g_state.vs.swizzle_data;
- const auto& program_code = g_state.vs.program_code;
+ const auto& uniforms = setup.uniforms;
+ const auto& swizzle_data = setup.swizzle_data;
+ const auto& program_code = setup.program_code;
// Placeholder for invalid inputs
static float24 dummy_vec4_float24[4];
@@ -170,7 +175,7 @@ void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData<Debug>
float24* dest =
(instr.common.dest.Value() < 0x10)
- ? &state.output_registers.value[instr.common.dest.Value().GetIndex()][0]
+ ? &state.registers.output[instr.common.dest.Value().GetIndex()][0]
: (instr.common.dest.Value() < 0x20)
? &state.registers.temporary[instr.common.dest.Value().GetIndex()][0]
: dummy_vec4_float24;
@@ -513,7 +518,7 @@ void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData<Debug>
float24* dest =
(instr.mad.dest.Value() < 0x10)
- ? &state.output_registers.value[instr.mad.dest.Value().GetIndex()][0]
+ ? &state.registers.output[instr.mad.dest.Value().GetIndex()][0]
: (instr.mad.dest.Value() < 0x20)
? &state.registers.temporary[instr.mad.dest.Value().GetIndex()][0]
: dummy_vec4_float24;
@@ -647,9 +652,33 @@ void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData<Debug>
}
}
-// Explicit instantiation
-template void RunInterpreter(const ShaderSetup&, UnitState&, DebugData<false>&, unsigned offset);
-template void RunInterpreter(const ShaderSetup&, UnitState&, DebugData<true>&, unsigned offset);
+void InterpreterEngine::SetupBatch(ShaderSetup& setup, unsigned int entry_point) {
+ ASSERT(entry_point < 1024);
+ setup.engine_data.entry_point = entry_point;
+}
+
+MICROPROFILE_DECLARE(GPU_Shader);
+
+void InterpreterEngine::Run(const ShaderSetup& setup, UnitState& state) const {
+
+ MICROPROFILE_SCOPE(GPU_Shader);
+
+ DebugData<false> dummy_debug_data;
+ RunInterpreter(setup, state, dummy_debug_data, setup.engine_data.entry_point);
+}
+
+DebugData<true> InterpreterEngine::ProduceDebugInfo(const ShaderSetup& setup,
+ const InputVertex& input,
+ int num_attributes) const {
+ UnitState state;
+ DebugData<true> debug_data;
+
+ // Setup input register table
+ boost::fill(state.registers.input, Math::Vec4<float24>::AssignToAll(float24::Zero()));
+ state.LoadInputVertex(input, num_attributes);
+ RunInterpreter(setup, state, debug_data, setup.engine_data.entry_point);
+ return debug_data;
+}
} // namespace
diff --git a/src/video_core/shader/shader_interpreter.h b/src/video_core/shader/shader_interpreter.h
index d31dcd7a6..d6c0e2d8c 100644
--- a/src/video_core/shader/shader_interpreter.h
+++ b/src/video_core/shader/shader_interpreter.h
@@ -4,18 +4,28 @@
#pragma once
+#include "video_core/shader/debug_data.h"
+#include "video_core/shader/shader.h"
+
namespace Pica {
namespace Shader {
-struct UnitState;
-
-template <bool Debug>
-struct DebugData;
-
-template <bool Debug>
-void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData<Debug>& debug_data,
- unsigned offset);
+class InterpreterEngine final : public ShaderEngine {
+public:
+ void SetupBatch(ShaderSetup& setup, unsigned int entry_point) override;
+ void Run(const ShaderSetup& setup, UnitState& state) const override;
+
+ /**
+ * Produce debug information based on the given shader and input vertex
+ * @param input Input vertex into the shader
+ * @param num_attributes The number of vertex shader attributes
+ * @param config Configuration object for the shader pipeline
+ * @return Debug information for this shader with regards to the given vertex
+ */
+ DebugData<true> ProduceDebugInfo(const ShaderSetup& setup, const InputVertex& input,
+ int num_attributes) const;
+};
} // namespace
diff --git a/src/video_core/shader/shader_jit_x64.cpp b/src/video_core/shader/shader_jit_x64.cpp
index c588b778b..0ee0dd9ef 100644
--- a/src/video_core/shader/shader_jit_x64.cpp
+++ b/src/video_core/shader/shader_jit_x64.cpp
@@ -1,888 +1,48 @@
-// Copyright 2015 Citra Emulator Project
+// Copyright 2016 Citra Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
-#include <algorithm>
-#include <cmath>
-#include <cstdint>
-#include <nihstro/shader_bytecode.h>
-#include <smmintrin.h>
-#include <xmmintrin.h>
-#include "common/assert.h"
-#include "common/logging/log.h"
-#include "common/vector_math.h"
-#include "common/x64/cpu_detect.h"
-#include "common/x64/xbyak_abi.h"
-#include "common/x64/xbyak_util.h"
-#include "video_core/pica_state.h"
-#include "video_core/pica_types.h"
+#include "common/hash.h"
+#include "common/microprofile.h"
#include "video_core/shader/shader.h"
#include "video_core/shader/shader_jit_x64.h"
-
-using namespace Common::X64;
-using namespace Xbyak::util;
-using Xbyak::Label;
-using Xbyak::Reg32;
-using Xbyak::Reg64;
-using Xbyak::Xmm;
+#include "video_core/shader/shader_jit_x64_compiler.h"
namespace Pica {
-
namespace Shader {
-typedef void (JitShader::*JitFunction)(Instruction instr);
-
-const JitFunction instr_table[64] = {
- &JitShader::Compile_ADD, // add
- &JitShader::Compile_DP3, // dp3
- &JitShader::Compile_DP4, // dp4
- &JitShader::Compile_DPH, // dph
- nullptr, // unknown
- &JitShader::Compile_EX2, // ex2
- &JitShader::Compile_LG2, // lg2
- nullptr, // unknown
- &JitShader::Compile_MUL, // mul
- &JitShader::Compile_SGE, // sge
- &JitShader::Compile_SLT, // slt
- &JitShader::Compile_FLR, // flr
- &JitShader::Compile_MAX, // max
- &JitShader::Compile_MIN, // min
- &JitShader::Compile_RCP, // rcp
- &JitShader::Compile_RSQ, // rsq
- nullptr, // unknown
- nullptr, // unknown
- &JitShader::Compile_MOVA, // mova
- &JitShader::Compile_MOV, // mov
- nullptr, // unknown
- nullptr, // unknown
- nullptr, // unknown
- nullptr, // unknown
- &JitShader::Compile_DPH, // dphi
- nullptr, // unknown
- &JitShader::Compile_SGE, // sgei
- &JitShader::Compile_SLT, // slti
- nullptr, // unknown
- nullptr, // unknown
- nullptr, // unknown
- nullptr, // unknown
- nullptr, // unknown
- &JitShader::Compile_NOP, // nop
- &JitShader::Compile_END, // end
- nullptr, // break
- &JitShader::Compile_CALL, // call
- &JitShader::Compile_CALLC, // callc
- &JitShader::Compile_CALLU, // callu
- &JitShader::Compile_IF, // ifu
- &JitShader::Compile_IF, // ifc
- &JitShader::Compile_LOOP, // loop
- nullptr, // emit
- nullptr, // sete
- &JitShader::Compile_JMP, // jmpc
- &JitShader::Compile_JMP, // jmpu
- &JitShader::Compile_CMP, // cmp
- &JitShader::Compile_CMP, // cmp
- &JitShader::Compile_MAD, // madi
- &JitShader::Compile_MAD, // madi
- &JitShader::Compile_MAD, // madi
- &JitShader::Compile_MAD, // madi
- &JitShader::Compile_MAD, // madi
- &JitShader::Compile_MAD, // madi
- &JitShader::Compile_MAD, // madi
- &JitShader::Compile_MAD, // madi
- &JitShader::Compile_MAD, // mad
- &JitShader::Compile_MAD, // mad
- &JitShader::Compile_MAD, // mad
- &JitShader::Compile_MAD, // mad
- &JitShader::Compile_MAD, // mad
- &JitShader::Compile_MAD, // mad
- &JitShader::Compile_MAD, // mad
- &JitShader::Compile_MAD, // mad
-};
-
-// The following is used to alias some commonly used registers. Generally, RAX-RDX and XMM0-XMM3 can
-// be used as scratch registers within a compiler function. The other registers have designated
-// purposes, as documented below:
+JitX64Engine::JitX64Engine() = default;
+JitX64Engine::~JitX64Engine() = default;
-/// Pointer to the uniform memory
-static const Reg64 SETUP = r9;
-/// The two 32-bit VS address offset registers set by the MOVA instruction
-static const Reg64 ADDROFFS_REG_0 = r10;
-static const Reg64 ADDROFFS_REG_1 = r11;
-/// VS loop count register (Multiplied by 16)
-static const Reg32 LOOPCOUNT_REG = r12d;
-/// Current VS loop iteration number (we could probably use LOOPCOUNT_REG, but this quicker)
-static const Reg32 LOOPCOUNT = esi;
-/// Number to increment LOOPCOUNT_REG by on each loop iteration (Multiplied by 16)
-static const Reg32 LOOPINC = edi;
-/// Result of the previous CMP instruction for the X-component comparison
-static const Reg64 COND0 = r13;
-/// Result of the previous CMP instruction for the Y-component comparison
-static const Reg64 COND1 = r14;
-/// Pointer to the UnitState instance for the current VS unit
-static const Reg64 STATE = r15;
-/// SIMD scratch register
-static const Xmm SCRATCH = xmm0;
-/// Loaded with the first swizzled source register, otherwise can be used as a scratch register
-static const Xmm SRC1 = xmm1;
-/// Loaded with the second swizzled source register, otherwise can be used as a scratch register
-static const Xmm SRC2 = xmm2;
-/// Loaded with the third swizzled source register, otherwise can be used as a scratch register
-static const Xmm SRC3 = xmm3;
-/// Additional scratch register
-static const Xmm SCRATCH2 = xmm4;
-/// Constant vector of [1.0f, 1.0f, 1.0f, 1.0f], used to efficiently set a vector to one
-static const Xmm ONE = xmm14;
-/// Constant vector of [-0.f, -0.f, -0.f, -0.f], used to efficiently negate a vector with XOR
-static const Xmm NEGBIT = xmm15;
+void JitX64Engine::SetupBatch(ShaderSetup& setup, unsigned int entry_point) {
+ ASSERT(entry_point < 1024);
+ setup.engine_data.entry_point = entry_point;
-// State registers that must not be modified by external functions calls
-// Scratch registers, e.g., SRC1 and SCRATCH, have to be saved on the side if needed
-static const BitSet32 persistent_regs = BuildRegSet({
- // Pointers to register blocks
- SETUP, STATE,
- // Cached registers
- ADDROFFS_REG_0, ADDROFFS_REG_1, LOOPCOUNT_REG, COND0, COND1,
- // Constants
- ONE, NEGBIT,
-});
+ u64 code_hash = Common::ComputeHash64(&setup.program_code, sizeof(setup.program_code));
+ u64 swizzle_hash = Common::ComputeHash64(&setup.swizzle_data, sizeof(setup.swizzle_data));
-/// Raw constant for the source register selector that indicates no swizzling is performed
-static const u8 NO_SRC_REG_SWIZZLE = 0x1b;
-/// Raw constant for the destination register enable mask that indicates all components are enabled
-static const u8 NO_DEST_REG_MASK = 0xf;
-
-/**
- * Get the vertex shader instruction for a given offset in the current shader program
- * @param offset Offset in the current shader program of the instruction
- * @return Instruction at the specified offset
- */
-static Instruction GetVertexShaderInstruction(size_t offset) {
- return {g_state.vs.program_code[offset]};
-}
-
-static void LogCritical(const char* msg) {
- LOG_CRITICAL(HW_GPU, "%s", msg);
-}
-
-void JitShader::Compile_Assert(bool condition, const char* msg) {
- if (!condition) {
- mov(ABI_PARAM1, reinterpret_cast<size_t>(msg));
- CallFarFunction(*this, LogCritical);
- }
-}
-
-/**
- * Loads and swizzles a source register into the specified XMM register.
- * @param instr VS instruction, used for determining how to load the source register
- * @param src_num Number indicating which source register to load (1 = src1, 2 = src2, 3 = src3)
- * @param src_reg SourceRegister object corresponding to the source register to load
- * @param dest Destination XMM register to store the loaded, swizzled source register
- */
-void JitShader::Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg,
- Xmm dest) {
- Reg64 src_ptr;
- size_t src_offset;
-
- if (src_reg.GetRegisterType() == RegisterType::FloatUniform) {
- src_ptr = SETUP;
- src_offset = ShaderSetup::GetFloatUniformOffset(src_reg.GetIndex());
+ u64 cache_key = code_hash ^ swizzle_hash;
+ auto iter = cache.find(cache_key);
+ if (iter != cache.end()) {
+ setup.engine_data.cached_shader = iter->second.get();
} else {
- src_ptr = STATE;
- src_offset = UnitState::InputOffset(src_reg);
- }
-
- int src_offset_disp = (int)src_offset;
- ASSERT_MSG(src_offset == src_offset_disp, "Source register offset too large for int type");
-
- unsigned operand_desc_id;
-
- const bool is_inverted =
- (0 != (instr.opcode.Value().GetInfo().subtype & OpCode::Info::SrcInversed));
-
- unsigned address_register_index;
- unsigned offset_src;
-
- if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MAD ||
- instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
- operand_desc_id = instr.mad.operand_desc_id;
- offset_src = is_inverted ? 3 : 2;
- address_register_index = instr.mad.address_register_index;
- } else {
- operand_desc_id = instr.common.operand_desc_id;
- offset_src = is_inverted ? 2 : 1;
- address_register_index = instr.common.address_register_index;
- }
-
- if (src_num == offset_src && address_register_index != 0) {
- switch (address_register_index) {
- case 1: // address offset 1
- movaps(dest, xword[src_ptr + ADDROFFS_REG_0 + src_offset_disp]);
- break;
- case 2: // address offset 2
- movaps(dest, xword[src_ptr + ADDROFFS_REG_1 + src_offset_disp]);
- break;
- case 3: // address offset 3
- movaps(dest, xword[src_ptr + LOOPCOUNT_REG.cvt64() + src_offset_disp]);
- break;
- default:
- UNREACHABLE();
- break;
- }
- } else {
- // Load the source
- movaps(dest, xword[src_ptr + src_offset_disp]);
- }
-
- SwizzlePattern swiz = {g_state.vs.swizzle_data[operand_desc_id]};
-
- // Generate instructions for source register swizzling as needed
- u8 sel = swiz.GetRawSelector(src_num);
- if (sel != NO_SRC_REG_SWIZZLE) {
- // Selector component order needs to be reversed for the SHUFPS instruction
- sel = ((sel & 0xc0) >> 6) | ((sel & 3) << 6) | ((sel & 0xc) << 2) | ((sel & 0x30) >> 2);
-
- // Shuffle inputs for swizzle
- shufps(dest, dest, sel);
- }
-
- // If the source register should be negated, flip the negative bit using XOR
- const bool negate[] = {swiz.negate_src1, swiz.negate_src2, swiz.negate_src3};
- if (negate[src_num - 1]) {
- xorps(dest, NEGBIT);
+ auto shader = std::make_unique<JitShader>();
+ shader->Compile(&setup.program_code, &setup.swizzle_data);
+ setup.engine_data.cached_shader = shader.get();
+ cache.emplace_hint(iter, cache_key, std::move(shader));
}
}
-void JitShader::Compile_DestEnable(Instruction instr, Xmm src) {
- DestRegister dest;
- unsigned operand_desc_id;
- if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MAD ||
- instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
- operand_desc_id = instr.mad.operand_desc_id;
- dest = instr.mad.dest.Value();
- } else {
- operand_desc_id = instr.common.operand_desc_id;
- dest = instr.common.dest.Value();
- }
-
- SwizzlePattern swiz = {g_state.vs.swizzle_data[operand_desc_id]};
-
- size_t dest_offset_disp = UnitState::OutputOffset(dest);
-
- // If all components are enabled, write the result to the destination register
- if (swiz.dest_mask == NO_DEST_REG_MASK) {
- // Store dest back to memory
- movaps(xword[STATE + dest_offset_disp], src);
-
- } else {
- // Not all components are enabled, so mask the result when storing to the destination
- // register...
- movaps(SCRATCH, xword[STATE + dest_offset_disp]);
-
- if (Common::GetCPUCaps().sse4_1) {
- u8 mask = ((swiz.dest_mask & 1) << 3) | ((swiz.dest_mask & 8) >> 3) |
- ((swiz.dest_mask & 2) << 1) | ((swiz.dest_mask & 4) >> 1);
- blendps(SCRATCH, src, mask);
- } else {
- movaps(SCRATCH2, src);
- unpckhps(SCRATCH2, SCRATCH); // Unpack X/Y components of source and destination
- unpcklps(SCRATCH, src); // Unpack Z/W components of source and destination
-
- // Compute selector to selectively copy source components to destination for SHUFPS
- // instruction
- u8 sel = ((swiz.DestComponentEnabled(0) ? 1 : 0) << 0) |
- ((swiz.DestComponentEnabled(1) ? 3 : 2) << 2) |
- ((swiz.DestComponentEnabled(2) ? 0 : 1) << 4) |
- ((swiz.DestComponentEnabled(3) ? 2 : 3) << 6);
- shufps(SCRATCH, SCRATCH2, sel);
- }
-
- // Store dest back to memory
- movaps(xword[STATE + dest_offset_disp], SCRATCH);
- }
-}
-
-void JitShader::Compile_SanitizedMul(Xmm src1, Xmm src2, Xmm scratch) {
- movaps(scratch, src1);
- cmpordps(scratch, src2);
-
- mulps(src1, src2);
+MICROPROFILE_DECLARE(GPU_Shader);
- movaps(src2, src1);
- cmpunordps(src2, src2);
+void JitX64Engine::Run(const ShaderSetup& setup, UnitState& state) const {
+ ASSERT(setup.engine_data.cached_shader != nullptr);
- xorps(scratch, src2);
- andps(src1, scratch);
-}
-
-void JitShader::Compile_EvaluateCondition(Instruction instr) {
- // Note: NXOR is used below to check for equality
- switch (instr.flow_control.op) {
- case Instruction::FlowControlType::Or:
- mov(eax, COND0);
- mov(ebx, COND1);
- xor(eax, (instr.flow_control.refx.Value() ^ 1));
- xor(ebx, (instr.flow_control.refy.Value() ^ 1));
- or (eax, ebx);
- break;
-
- case Instruction::FlowControlType::And:
- mov(eax, COND0);
- mov(ebx, COND1);
- xor(eax, (instr.flow_control.refx.Value() ^ 1));
- xor(ebx, (instr.flow_control.refy.Value() ^ 1));
- and(eax, ebx);
- break;
-
- case Instruction::FlowControlType::JustX:
- mov(eax, COND0);
- xor(eax, (instr.flow_control.refx.Value() ^ 1));
- break;
-
- case Instruction::FlowControlType::JustY:
- mov(eax, COND1);
- xor(eax, (instr.flow_control.refy.Value() ^ 1));
- break;
- }
-}
+ MICROPROFILE_SCOPE(GPU_Shader);
-void JitShader::Compile_UniformCondition(Instruction instr) {
- size_t offset = ShaderSetup::GetBoolUniformOffset(instr.flow_control.bool_uniform_id);
- cmp(byte[SETUP + offset], 0);
+ const JitShader* shader = static_cast<const JitShader*>(setup.engine_data.cached_shader);
+ shader->Run(setup, state, setup.engine_data.entry_point);
}
-BitSet32 JitShader::PersistentCallerSavedRegs() {
- return persistent_regs & ABI_ALL_CALLER_SAVED;
-}
-
-void JitShader::Compile_ADD(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
- addps(SRC1, SRC2);
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_DP3(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
-
- Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
-
- movaps(SRC2, SRC1);
- shufps(SRC2, SRC2, _MM_SHUFFLE(1, 1, 1, 1));
-
- movaps(SRC3, SRC1);
- shufps(SRC3, SRC3, _MM_SHUFFLE(2, 2, 2, 2));
-
- shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0));
- addps(SRC1, SRC2);
- addps(SRC1, SRC3);
-
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_DP4(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
-
- Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
-
- movaps(SRC2, SRC1);
- shufps(SRC1, SRC1, _MM_SHUFFLE(2, 3, 0, 1)); // XYZW -> ZWXY
- addps(SRC1, SRC2);
-
- movaps(SRC2, SRC1);
- shufps(SRC1, SRC1, _MM_SHUFFLE(0, 1, 2, 3)); // XYZW -> WZYX
- addps(SRC1, SRC2);
-
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_DPH(Instruction instr) {
- if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::DPHI) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
- } else {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
- }
-
- if (Common::GetCPUCaps().sse4_1) {
- // Set 4th component to 1.0
- blendps(SRC1, ONE, 0b1000);
- } else {
- // Set 4th component to 1.0
- movaps(SCRATCH, SRC1);
- unpckhps(SCRATCH, ONE); // XYZW, 1111 -> Z1__
- unpcklpd(SRC1, SCRATCH); // XYZW, Z1__ -> XYZ1
- }
-
- Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
-
- movaps(SRC2, SRC1);
- shufps(SRC1, SRC1, _MM_SHUFFLE(2, 3, 0, 1)); // XYZW -> ZWXY
- addps(SRC1, SRC2);
-
- movaps(SRC2, SRC1);
- shufps(SRC1, SRC1, _MM_SHUFFLE(0, 1, 2, 3)); // XYZW -> WZYX
- addps(SRC1, SRC2);
-
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_EX2(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- movss(xmm0, SRC1); // ABI_PARAM1
-
- ABI_PushRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
- CallFarFunction(*this, exp2f);
- ABI_PopRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
-
- shufps(xmm0, xmm0, _MM_SHUFFLE(0, 0, 0, 0)); // ABI_RETURN
- movaps(SRC1, xmm0);
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_LG2(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- movss(xmm0, SRC1); // ABI_PARAM1
-
- ABI_PushRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
- CallFarFunction(*this, log2f);
- ABI_PopRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
-
- shufps(xmm0, xmm0, _MM_SHUFFLE(0, 0, 0, 0)); // ABI_RETURN
- movaps(SRC1, xmm0);
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_MUL(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
- Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_SGE(Instruction instr) {
- if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::SGEI) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
- } else {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
- }
-
- cmpleps(SRC2, SRC1);
- andps(SRC2, ONE);
-
- Compile_DestEnable(instr, SRC2);
-}
-
-void JitShader::Compile_SLT(Instruction instr) {
- if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::SLTI) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
- } else {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
- }
-
- cmpltps(SRC1, SRC2);
- andps(SRC1, ONE);
-
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_FLR(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
-
- if (Common::GetCPUCaps().sse4_1) {
- roundps(SRC1, SRC1, _MM_FROUND_FLOOR);
- } else {
- cvttps2dq(SRC1, SRC1);
- cvtdq2ps(SRC1, SRC1);
- }
-
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_MAX(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
- // SSE semantics match PICA200 ones: In case of NaN, SRC2 is returned.
- maxps(SRC1, SRC2);
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_MIN(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
- // SSE semantics match PICA200 ones: In case of NaN, SRC2 is returned.
- minps(SRC1, SRC2);
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_MOVA(Instruction instr) {
- SwizzlePattern swiz = {g_state.vs.swizzle_data[instr.common.operand_desc_id]};
-
- if (!swiz.DestComponentEnabled(0) && !swiz.DestComponentEnabled(1)) {
- return; // NoOp
- }
-
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
-
- // Convert floats to integers using truncation (only care about X and Y components)
- cvttps2dq(SRC1, SRC1);
-
- // Get result
- movq(rax, SRC1);
-
- // Handle destination enable
- if (swiz.DestComponentEnabled(0) && swiz.DestComponentEnabled(1)) {
- // Move and sign-extend low 32 bits
- movsxd(ADDROFFS_REG_0, eax);
-
- // Move and sign-extend high 32 bits
- shr(rax, 32);
- movsxd(ADDROFFS_REG_1, eax);
-
- // Multiply by 16 to be used as an offset later
- shl(ADDROFFS_REG_0, 4);
- shl(ADDROFFS_REG_1, 4);
- } else {
- if (swiz.DestComponentEnabled(0)) {
- // Move and sign-extend low 32 bits
- movsxd(ADDROFFS_REG_0, eax);
-
- // Multiply by 16 to be used as an offset later
- shl(ADDROFFS_REG_0, 4);
- } else if (swiz.DestComponentEnabled(1)) {
- // Move and sign-extend high 32 bits
- shr(rax, 32);
- movsxd(ADDROFFS_REG_1, eax);
-
- // Multiply by 16 to be used as an offset later
- shl(ADDROFFS_REG_1, 4);
- }
- }
-}
-
-void JitShader::Compile_MOV(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_RCP(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
-
- // TODO(bunnei): RCPSS is a pretty rough approximation, this might cause problems if Pica
- // performs this operation more accurately. This should be checked on hardware.
- rcpss(SRC1, SRC1);
- shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
-
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_RSQ(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
-
- // TODO(bunnei): RSQRTSS is a pretty rough approximation, this might cause problems if Pica
- // performs this operation more accurately. This should be checked on hardware.
- rsqrtss(SRC1, SRC1);
- shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
-
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_NOP(Instruction instr) {}
-
-void JitShader::Compile_END(Instruction instr) {
- ABI_PopRegistersAndAdjustStack(*this, ABI_ALL_CALLEE_SAVED, 8);
- ret();
-}
-
-void JitShader::Compile_CALL(Instruction instr) {
- // Push offset of the return
- push(qword, (instr.flow_control.dest_offset + instr.flow_control.num_instructions));
-
- // Call the subroutine
- call(instruction_labels[instr.flow_control.dest_offset]);
-
- // Skip over the return offset that's on the stack
- add(rsp, 8);
-}
-
-void JitShader::Compile_CALLC(Instruction instr) {
- Compile_EvaluateCondition(instr);
- Label b;
- jz(b);
- Compile_CALL(instr);
- L(b);
-}
-
-void JitShader::Compile_CALLU(Instruction instr) {
- Compile_UniformCondition(instr);
- Label b;
- jz(b);
- Compile_CALL(instr);
- L(b);
-}
-
-void JitShader::Compile_CMP(Instruction instr) {
- using Op = Instruction::Common::CompareOpType::Op;
- Op op_x = instr.common.compare_op.x;
- Op op_y = instr.common.compare_op.y;
-
- Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
- Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
-
- // SSE doesn't have greater-than (GT) or greater-equal (GE) comparison operators. You need to
- // emulate them by swapping the lhs and rhs and using LT and LE. NLT and NLE can't be used here
- // because they don't match when used with NaNs.
- static const u8 cmp[] = {CMP_EQ, CMP_NEQ, CMP_LT, CMP_LE, CMP_LT, CMP_LE};
-
- bool invert_op_x = (op_x == Op::GreaterThan || op_x == Op::GreaterEqual);
- Xmm lhs_x = invert_op_x ? SRC2 : SRC1;
- Xmm rhs_x = invert_op_x ? SRC1 : SRC2;
-
- if (op_x == op_y) {
- // Compare X-component and Y-component together
- cmpps(lhs_x, rhs_x, cmp[op_x]);
- movq(COND0, lhs_x);
-
- mov(COND1, COND0);
- } else {
- bool invert_op_y = (op_y == Op::GreaterThan || op_y == Op::GreaterEqual);
- Xmm lhs_y = invert_op_y ? SRC2 : SRC1;
- Xmm rhs_y = invert_op_y ? SRC1 : SRC2;
-
- // Compare X-component
- movaps(SCRATCH, lhs_x);
- cmpss(SCRATCH, rhs_x, cmp[op_x]);
-
- // Compare Y-component
- cmpps(lhs_y, rhs_y, cmp[op_y]);
-
- movq(COND0, SCRATCH);
- movq(COND1, lhs_y);
- }
-
- shr(COND0.cvt32(), 31); // ignores upper 32 bits in source
- shr(COND1, 63);
-}
-
-void JitShader::Compile_MAD(Instruction instr) {
- Compile_SwizzleSrc(instr, 1, instr.mad.src1, SRC1);
-
- if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
- Compile_SwizzleSrc(instr, 2, instr.mad.src2i, SRC2);
- Compile_SwizzleSrc(instr, 3, instr.mad.src3i, SRC3);
- } else {
- Compile_SwizzleSrc(instr, 2, instr.mad.src2, SRC2);
- Compile_SwizzleSrc(instr, 3, instr.mad.src3, SRC3);
- }
-
- Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
- addps(SRC1, SRC3);
-
- Compile_DestEnable(instr, SRC1);
-}
-
-void JitShader::Compile_IF(Instruction instr) {
- Compile_Assert(instr.flow_control.dest_offset >= program_counter,
- "Backwards if-statements not supported");
- Label l_else, l_endif;
-
- // Evaluate the "IF" condition
- if (instr.opcode.Value() == OpCode::Id::IFU) {
- Compile_UniformCondition(instr);
- } else if (instr.opcode.Value() == OpCode::Id::IFC) {
- Compile_EvaluateCondition(instr);
- }
- jz(l_else, T_NEAR);
-
- // Compile the code that corresponds to the condition evaluating as true
- Compile_Block(instr.flow_control.dest_offset);
-
- // If there isn't an "ELSE" condition, we are done here
- if (instr.flow_control.num_instructions == 0) {
- L(l_else);
- return;
- }
-
- jmp(l_endif, T_NEAR);
-
- L(l_else);
- // This code corresponds to the "ELSE" condition
- // Comple the code that corresponds to the condition evaluating as false
- Compile_Block(instr.flow_control.dest_offset + instr.flow_control.num_instructions);
-
- L(l_endif);
-}
-
-void JitShader::Compile_LOOP(Instruction instr) {
- Compile_Assert(instr.flow_control.dest_offset >= program_counter,
- "Backwards loops not supported");
- Compile_Assert(!looping, "Nested loops not supported");
-
- looping = true;
-
- // This decodes the fields from the integer uniform at index instr.flow_control.int_uniform_id.
- // The Y (LOOPCOUNT_REG) and Z (LOOPINC) component are kept multiplied by 16 (Left shifted by
- // 4 bits) to be used as an offset into the 16-byte vector registers later
- size_t offset = ShaderSetup::GetIntUniformOffset(instr.flow_control.int_uniform_id);
- mov(LOOPCOUNT, dword[SETUP + offset]);
- mov(LOOPCOUNT_REG, LOOPCOUNT);
- shr(LOOPCOUNT_REG, 4);
- and(LOOPCOUNT_REG, 0xFF0); // Y-component is the start
- mov(LOOPINC, LOOPCOUNT);
- shr(LOOPINC, 12);
- and(LOOPINC, 0xFF0); // Z-component is the incrementer
- movzx(LOOPCOUNT, LOOPCOUNT.cvt8()); // X-component is iteration count
- add(LOOPCOUNT, 1); // Iteration count is X-component + 1
-
- Label l_loop_start;
- L(l_loop_start);
-
- Compile_Block(instr.flow_control.dest_offset + 1);
-
- add(LOOPCOUNT_REG, LOOPINC); // Increment LOOPCOUNT_REG by Z-component
- sub(LOOPCOUNT, 1); // Increment loop count by 1
- jnz(l_loop_start); // Loop if not equal
-
- looping = false;
-}
-
-void JitShader::Compile_JMP(Instruction instr) {
- if (instr.opcode.Value() == OpCode::Id::JMPC)
- Compile_EvaluateCondition(instr);
- else if (instr.opcode.Value() == OpCode::Id::JMPU)
- Compile_UniformCondition(instr);
- else
- UNREACHABLE();
-
- bool inverted_condition =
- (instr.opcode.Value() == OpCode::Id::JMPU) && (instr.flow_control.num_instructions & 1);
-
- Label& b = instruction_labels[instr.flow_control.dest_offset];
- if (inverted_condition) {
- jz(b, T_NEAR);
- } else {
- jnz(b, T_NEAR);
- }
-}
-
-void JitShader::Compile_Block(unsigned end) {
- while (program_counter < end) {
- Compile_NextInstr();
- }
-}
-
-void JitShader::Compile_Return() {
- // Peek return offset on the stack and check if we're at that offset
- mov(rax, qword[rsp + 8]);
- cmp(eax, (program_counter));
-
- // If so, jump back to before CALL
- Label b;
- jnz(b);
- ret();
- L(b);
-}
-
-void JitShader::Compile_NextInstr() {
- if (std::binary_search(return_offsets.begin(), return_offsets.end(), program_counter)) {
- Compile_Return();
- }
-
- L(instruction_labels[program_counter]);
-
- Instruction instr = GetVertexShaderInstruction(program_counter++);
-
- OpCode::Id opcode = instr.opcode.Value();
- auto instr_func = instr_table[static_cast<unsigned>(opcode)];
-
- if (instr_func) {
- // JIT the instruction!
- ((*this).*instr_func)(instr);
- } else {
- // Unhandled instruction
- LOG_CRITICAL(HW_GPU, "Unhandled instruction: 0x%02x (0x%08x)",
- instr.opcode.Value().EffectiveOpCode(), instr.hex);
- }
-}
-
-void JitShader::FindReturnOffsets() {
- return_offsets.clear();
-
- for (size_t offset = 0; offset < g_state.vs.program_code.size(); ++offset) {
- Instruction instr = GetVertexShaderInstruction(offset);
-
- switch (instr.opcode.Value()) {
- case OpCode::Id::CALL:
- case OpCode::Id::CALLC:
- case OpCode::Id::CALLU:
- return_offsets.push_back(instr.flow_control.dest_offset +
- instr.flow_control.num_instructions);
- break;
- default:
- break;
- }
- }
-
- // Sort for efficient binary search later
- std::sort(return_offsets.begin(), return_offsets.end());
-}
-
-void JitShader::Compile() {
- // Reset flow control state
- program = (CompiledShader*)getCurr();
- program_counter = 0;
- looping = false;
- instruction_labels.fill(Xbyak::Label());
-
- // Find all `CALL` instructions and identify return locations
- FindReturnOffsets();
-
- // The stack pointer is 8 modulo 16 at the entry of a procedure
- ABI_PushRegistersAndAdjustStack(*this, ABI_ALL_CALLEE_SAVED, 8);
-
- mov(SETUP, ABI_PARAM1);
- mov(STATE, ABI_PARAM2);
-
- // Zero address/loop registers
- xor(ADDROFFS_REG_0.cvt32(), ADDROFFS_REG_0.cvt32());
- xor(ADDROFFS_REG_1.cvt32(), ADDROFFS_REG_1.cvt32());
- xor(LOOPCOUNT_REG, LOOPCOUNT_REG);
-
- // Used to set a register to one
- static const __m128 one = {1.f, 1.f, 1.f, 1.f};
- mov(rax, reinterpret_cast<size_t>(&one));
- movaps(ONE, xword[rax]);
-
- // Used to negate registers
- static const __m128 neg = {-0.f, -0.f, -0.f, -0.f};
- mov(rax, reinterpret_cast<size_t>(&neg));
- movaps(NEGBIT, xword[rax]);
-
- // Jump to start of the shader program
- jmp(ABI_PARAM3);
-
- // Compile entire program
- Compile_Block(static_cast<unsigned>(g_state.vs.program_code.size()));
-
- // Free memory that's no longer needed
- return_offsets.clear();
- return_offsets.shrink_to_fit();
-
- ready();
-
- uintptr_t size = reinterpret_cast<uintptr_t>(getCurr()) - reinterpret_cast<uintptr_t>(program);
- ASSERT_MSG(size <= MAX_SHADER_SIZE, "Compiled a shader that exceeds the allocated size!");
- LOG_DEBUG(HW_GPU, "Compiled shader size=%lu", size);
-}
-
-JitShader::JitShader() : Xbyak::CodeGenerator(MAX_SHADER_SIZE) {}
-
} // namespace Shader
-
} // namespace Pica
diff --git a/src/video_core/shader/shader_jit_x64.h b/src/video_core/shader/shader_jit_x64.h
index f37548306..078b2cba5 100644
--- a/src/video_core/shader/shader_jit_x64.h
+++ b/src/video_core/shader/shader_jit_x64.h
@@ -1,121 +1,30 @@
-// Copyright 2015 Citra Emulator Project
+// Copyright 2016 Citra Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#pragma once
-#include <array>
-#include <cstddef>
-#include <utility>
-#include <vector>
-#include <nihstro/shader_bytecode.h>
-#include <xbyak.h>
-#include "common/bit_set.h"
+#include <memory>
+#include <unordered_map>
#include "common/common_types.h"
-#include "common/x64/emitter.h"
#include "video_core/shader/shader.h"
-using nihstro::Instruction;
-using nihstro::OpCode;
-using nihstro::SwizzlePattern;
-
namespace Pica {
-
namespace Shader {
-/// Memory allocated for each compiled shader (64Kb)
-constexpr size_t MAX_SHADER_SIZE = 1024 * 64;
+class JitShader;
-/**
- * This class implements the shader JIT compiler. It recompiles a Pica shader program into x86_64
- * code that can be executed on the host machine directly.
- */
-class JitShader : public Xbyak::CodeGenerator {
+class JitX64Engine final : public ShaderEngine {
public:
- JitShader();
-
- void Run(const ShaderSetup& setup, UnitState& state, unsigned offset) const {
- program(&setup, &state, instruction_labels[offset].getAddress());
- }
-
- void Compile();
+ JitX64Engine();
+ ~JitX64Engine() override;
- void Compile_ADD(Instruction instr);
- void Compile_DP3(Instruction instr);
- void Compile_DP4(Instruction instr);
- void Compile_DPH(Instruction instr);
- void Compile_EX2(Instruction instr);
- void Compile_LG2(Instruction instr);
- void Compile_MUL(Instruction instr);
- void Compile_SGE(Instruction instr);
- void Compile_SLT(Instruction instr);
- void Compile_FLR(Instruction instr);
- void Compile_MAX(Instruction instr);
- void Compile_MIN(Instruction instr);
- void Compile_RCP(Instruction instr);
- void Compile_RSQ(Instruction instr);
- void Compile_MOVA(Instruction instr);
- void Compile_MOV(Instruction instr);
- void Compile_NOP(Instruction instr);
- void Compile_END(Instruction instr);
- void Compile_CALL(Instruction instr);
- void Compile_CALLC(Instruction instr);
- void Compile_CALLU(Instruction instr);
- void Compile_IF(Instruction instr);
- void Compile_LOOP(Instruction instr);
- void Compile_JMP(Instruction instr);
- void Compile_CMP(Instruction instr);
- void Compile_MAD(Instruction instr);
+ void SetupBatch(ShaderSetup& setup, unsigned int entry_point) override;
+ void Run(const ShaderSetup& setup, UnitState& state) const override;
private:
- void Compile_Block(unsigned end);
- void Compile_NextInstr();
-
- void Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg,
- Xbyak::Xmm dest);
- void Compile_DestEnable(Instruction instr, Xbyak::Xmm dest);
-
- /**
- * Compiles a `MUL src1, src2` operation, properly handling the PICA semantics when multiplying
- * zero by inf. Clobbers `src2` and `scratch`.
- */
- void Compile_SanitizedMul(Xbyak::Xmm src1, Xbyak::Xmm src2, Xbyak::Xmm scratch);
-
- void Compile_EvaluateCondition(Instruction instr);
- void Compile_UniformCondition(Instruction instr);
-
- /**
- * Emits the code to conditionally return from a subroutine envoked by the `CALL` instruction.
- */
- void Compile_Return();
-
- BitSet32 PersistentCallerSavedRegs();
-
- /**
- * Assertion evaluated at compile-time, but only triggered if executed at runtime.
- * @param msg Message to be logged if the assertion fails.
- */
- void Compile_Assert(bool condition, const char* msg);
-
- /**
- * Analyzes the entire shader program for `CALL` instructions before emitting any code,
- * identifying the locations where a return needs to be inserted.
- */
- void FindReturnOffsets();
-
- /// Mapping of Pica VS instructions to pointers in the emitted code
- std::array<Xbyak::Label, 1024> instruction_labels;
-
- /// Offsets in code where a return needs to be inserted
- std::vector<unsigned> return_offsets;
-
- unsigned program_counter = 0; ///< Offset of the next instruction to decode
- bool looping = false; ///< True if compiling a loop, used to check for nested loops
-
- using CompiledShader = void(const void* setup, void* state, const u8* start_addr);
- CompiledShader* program = nullptr;
+ std::unordered_map<u64, std::unique_ptr<JitShader>> cache;
};
-} // Shader
-
-} // Pica
+} // namespace Shader
+} // namespace Pica
diff --git a/src/video_core/shader/shader_jit_x64_compiler.cpp b/src/video_core/shader/shader_jit_x64_compiler.cpp
new file mode 100644
index 000000000..49806e8c9
--- /dev/null
+++ b/src/video_core/shader/shader_jit_x64_compiler.cpp
@@ -0,0 +1,884 @@
+// Copyright 2015 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <nihstro/shader_bytecode.h>
+#include <smmintrin.h>
+#include <xmmintrin.h>
+#include "common/assert.h"
+#include "common/logging/log.h"
+#include "common/vector_math.h"
+#include "common/x64/cpu_detect.h"
+#include "common/x64/xbyak_abi.h"
+#include "common/x64/xbyak_util.h"
+#include "video_core/pica_state.h"
+#include "video_core/pica_types.h"
+#include "video_core/shader/shader.h"
+#include "video_core/shader/shader_jit_x64_compiler.h"
+
+using namespace Common::X64;
+using namespace Xbyak::util;
+using Xbyak::Label;
+using Xbyak::Reg32;
+using Xbyak::Reg64;
+using Xbyak::Xmm;
+
+namespace Pica {
+
+namespace Shader {
+
+typedef void (JitShader::*JitFunction)(Instruction instr);
+
+const JitFunction instr_table[64] = {
+ &JitShader::Compile_ADD, // add
+ &JitShader::Compile_DP3, // dp3
+ &JitShader::Compile_DP4, // dp4
+ &JitShader::Compile_DPH, // dph
+ nullptr, // unknown
+ &JitShader::Compile_EX2, // ex2
+ &JitShader::Compile_LG2, // lg2
+ nullptr, // unknown
+ &JitShader::Compile_MUL, // mul
+ &JitShader::Compile_SGE, // sge
+ &JitShader::Compile_SLT, // slt
+ &JitShader::Compile_FLR, // flr
+ &JitShader::Compile_MAX, // max
+ &JitShader::Compile_MIN, // min
+ &JitShader::Compile_RCP, // rcp
+ &JitShader::Compile_RSQ, // rsq
+ nullptr, // unknown
+ nullptr, // unknown
+ &JitShader::Compile_MOVA, // mova
+ &JitShader::Compile_MOV, // mov
+ nullptr, // unknown
+ nullptr, // unknown
+ nullptr, // unknown
+ nullptr, // unknown
+ &JitShader::Compile_DPH, // dphi
+ nullptr, // unknown
+ &JitShader::Compile_SGE, // sgei
+ &JitShader::Compile_SLT, // slti
+ nullptr, // unknown
+ nullptr, // unknown
+ nullptr, // unknown
+ nullptr, // unknown
+ nullptr, // unknown
+ &JitShader::Compile_NOP, // nop
+ &JitShader::Compile_END, // end
+ nullptr, // break
+ &JitShader::Compile_CALL, // call
+ &JitShader::Compile_CALLC, // callc
+ &JitShader::Compile_CALLU, // callu
+ &JitShader::Compile_IF, // ifu
+ &JitShader::Compile_IF, // ifc
+ &JitShader::Compile_LOOP, // loop
+ nullptr, // emit
+ nullptr, // sete
+ &JitShader::Compile_JMP, // jmpc
+ &JitShader::Compile_JMP, // jmpu
+ &JitShader::Compile_CMP, // cmp
+ &JitShader::Compile_CMP, // cmp
+ &JitShader::Compile_MAD, // madi
+ &JitShader::Compile_MAD, // madi
+ &JitShader::Compile_MAD, // madi
+ &JitShader::Compile_MAD, // madi
+ &JitShader::Compile_MAD, // madi
+ &JitShader::Compile_MAD, // madi
+ &JitShader::Compile_MAD, // madi
+ &JitShader::Compile_MAD, // madi
+ &JitShader::Compile_MAD, // mad
+ &JitShader::Compile_MAD, // mad
+ &JitShader::Compile_MAD, // mad
+ &JitShader::Compile_MAD, // mad
+ &JitShader::Compile_MAD, // mad
+ &JitShader::Compile_MAD, // mad
+ &JitShader::Compile_MAD, // mad
+ &JitShader::Compile_MAD, // mad
+};
+
+// The following is used to alias some commonly used registers. Generally, RAX-RDX and XMM0-XMM3 can
+// be used as scratch registers within a compiler function. The other registers have designated
+// purposes, as documented below:
+
+/// Pointer to the uniform memory
+static const Reg64 SETUP = r9;
+/// The two 32-bit VS address offset registers set by the MOVA instruction
+static const Reg64 ADDROFFS_REG_0 = r10;
+static const Reg64 ADDROFFS_REG_1 = r11;
+/// VS loop count register (Multiplied by 16)
+static const Reg32 LOOPCOUNT_REG = r12d;
+/// Current VS loop iteration number (we could probably use LOOPCOUNT_REG, but this quicker)
+static const Reg32 LOOPCOUNT = esi;
+/// Number to increment LOOPCOUNT_REG by on each loop iteration (Multiplied by 16)
+static const Reg32 LOOPINC = edi;
+/// Result of the previous CMP instruction for the X-component comparison
+static const Reg64 COND0 = r13;
+/// Result of the previous CMP instruction for the Y-component comparison
+static const Reg64 COND1 = r14;
+/// Pointer to the UnitState instance for the current VS unit
+static const Reg64 STATE = r15;
+/// SIMD scratch register
+static const Xmm SCRATCH = xmm0;
+/// Loaded with the first swizzled source register, otherwise can be used as a scratch register
+static const Xmm SRC1 = xmm1;
+/// Loaded with the second swizzled source register, otherwise can be used as a scratch register
+static const Xmm SRC2 = xmm2;
+/// Loaded with the third swizzled source register, otherwise can be used as a scratch register
+static const Xmm SRC3 = xmm3;
+/// Additional scratch register
+static const Xmm SCRATCH2 = xmm4;
+/// Constant vector of [1.0f, 1.0f, 1.0f, 1.0f], used to efficiently set a vector to one
+static const Xmm ONE = xmm14;
+/// Constant vector of [-0.f, -0.f, -0.f, -0.f], used to efficiently negate a vector with XOR
+static const Xmm NEGBIT = xmm15;
+
+// State registers that must not be modified by external functions calls
+// Scratch registers, e.g., SRC1 and SCRATCH, have to be saved on the side if needed
+static const BitSet32 persistent_regs = BuildRegSet({
+ // Pointers to register blocks
+ SETUP, STATE,
+ // Cached registers
+ ADDROFFS_REG_0, ADDROFFS_REG_1, LOOPCOUNT_REG, COND0, COND1,
+ // Constants
+ ONE, NEGBIT,
+});
+
+/// Raw constant for the source register selector that indicates no swizzling is performed
+static const u8 NO_SRC_REG_SWIZZLE = 0x1b;
+/// Raw constant for the destination register enable mask that indicates all components are enabled
+static const u8 NO_DEST_REG_MASK = 0xf;
+
+static void LogCritical(const char* msg) {
+ LOG_CRITICAL(HW_GPU, "%s", msg);
+}
+
+void JitShader::Compile_Assert(bool condition, const char* msg) {
+ if (!condition) {
+ mov(ABI_PARAM1, reinterpret_cast<size_t>(msg));
+ CallFarFunction(*this, LogCritical);
+ }
+}
+
+/**
+ * Loads and swizzles a source register into the specified XMM register.
+ * @param instr VS instruction, used for determining how to load the source register
+ * @param src_num Number indicating which source register to load (1 = src1, 2 = src2, 3 = src3)
+ * @param src_reg SourceRegister object corresponding to the source register to load
+ * @param dest Destination XMM register to store the loaded, swizzled source register
+ */
+void JitShader::Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg,
+ Xmm dest) {
+ Reg64 src_ptr;
+ size_t src_offset;
+
+ if (src_reg.GetRegisterType() == RegisterType::FloatUniform) {
+ src_ptr = SETUP;
+ src_offset = ShaderSetup::GetFloatUniformOffset(src_reg.GetIndex());
+ } else {
+ src_ptr = STATE;
+ src_offset = UnitState::InputOffset(src_reg);
+ }
+
+ int src_offset_disp = (int)src_offset;
+ ASSERT_MSG(src_offset == src_offset_disp, "Source register offset too large for int type");
+
+ unsigned operand_desc_id;
+
+ const bool is_inverted =
+ (0 != (instr.opcode.Value().GetInfo().subtype & OpCode::Info::SrcInversed));
+
+ unsigned address_register_index;
+ unsigned offset_src;
+
+ if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MAD ||
+ instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
+ operand_desc_id = instr.mad.operand_desc_id;
+ offset_src = is_inverted ? 3 : 2;
+ address_register_index = instr.mad.address_register_index;
+ } else {
+ operand_desc_id = instr.common.operand_desc_id;
+ offset_src = is_inverted ? 2 : 1;
+ address_register_index = instr.common.address_register_index;
+ }
+
+ if (src_num == offset_src && address_register_index != 0) {
+ switch (address_register_index) {
+ case 1: // address offset 1
+ movaps(dest, xword[src_ptr + ADDROFFS_REG_0 + src_offset_disp]);
+ break;
+ case 2: // address offset 2
+ movaps(dest, xword[src_ptr + ADDROFFS_REG_1 + src_offset_disp]);
+ break;
+ case 3: // address offset 3
+ movaps(dest, xword[src_ptr + LOOPCOUNT_REG.cvt64() + src_offset_disp]);
+ break;
+ default:
+ UNREACHABLE();
+ break;
+ }
+ } else {
+ // Load the source
+ movaps(dest, xword[src_ptr + src_offset_disp]);
+ }
+
+ SwizzlePattern swiz = {(*swizzle_data)[operand_desc_id]};
+
+ // Generate instructions for source register swizzling as needed
+ u8 sel = swiz.GetRawSelector(src_num);
+ if (sel != NO_SRC_REG_SWIZZLE) {
+ // Selector component order needs to be reversed for the SHUFPS instruction
+ sel = ((sel & 0xc0) >> 6) | ((sel & 3) << 6) | ((sel & 0xc) << 2) | ((sel & 0x30) >> 2);
+
+ // Shuffle inputs for swizzle
+ shufps(dest, dest, sel);
+ }
+
+ // If the source register should be negated, flip the negative bit using XOR
+ const bool negate[] = {swiz.negate_src1, swiz.negate_src2, swiz.negate_src3};
+ if (negate[src_num - 1]) {
+ xorps(dest, NEGBIT);
+ }
+}
+
+void JitShader::Compile_DestEnable(Instruction instr, Xmm src) {
+ DestRegister dest;
+ unsigned operand_desc_id;
+ if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MAD ||
+ instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
+ operand_desc_id = instr.mad.operand_desc_id;
+ dest = instr.mad.dest.Value();
+ } else {
+ operand_desc_id = instr.common.operand_desc_id;
+ dest = instr.common.dest.Value();
+ }
+
+ SwizzlePattern swiz = {(*swizzle_data)[operand_desc_id]};
+
+ size_t dest_offset_disp = UnitState::OutputOffset(dest);
+
+ // If all components are enabled, write the result to the destination register
+ if (swiz.dest_mask == NO_DEST_REG_MASK) {
+ // Store dest back to memory
+ movaps(xword[STATE + dest_offset_disp], src);
+
+ } else {
+ // Not all components are enabled, so mask the result when storing to the destination
+ // register...
+ movaps(SCRATCH, xword[STATE + dest_offset_disp]);
+
+ if (Common::GetCPUCaps().sse4_1) {
+ u8 mask = ((swiz.dest_mask & 1) << 3) | ((swiz.dest_mask & 8) >> 3) |
+ ((swiz.dest_mask & 2) << 1) | ((swiz.dest_mask & 4) >> 1);
+ blendps(SCRATCH, src, mask);
+ } else {
+ movaps(SCRATCH2, src);
+ unpckhps(SCRATCH2, SCRATCH); // Unpack X/Y components of source and destination
+ unpcklps(SCRATCH, src); // Unpack Z/W components of source and destination
+
+ // Compute selector to selectively copy source components to destination for SHUFPS
+ // instruction
+ u8 sel = ((swiz.DestComponentEnabled(0) ? 1 : 0) << 0) |
+ ((swiz.DestComponentEnabled(1) ? 3 : 2) << 2) |
+ ((swiz.DestComponentEnabled(2) ? 0 : 1) << 4) |
+ ((swiz.DestComponentEnabled(3) ? 2 : 3) << 6);
+ shufps(SCRATCH, SCRATCH2, sel);
+ }
+
+ // Store dest back to memory
+ movaps(xword[STATE + dest_offset_disp], SCRATCH);
+ }
+}
+
+void JitShader::Compile_SanitizedMul(Xmm src1, Xmm src2, Xmm scratch) {
+ movaps(scratch, src1);
+ cmpordps(scratch, src2);
+
+ mulps(src1, src2);
+
+ movaps(src2, src1);
+ cmpunordps(src2, src2);
+
+ xorps(scratch, src2);
+ andps(src1, scratch);
+}
+
+void JitShader::Compile_EvaluateCondition(Instruction instr) {
+ // Note: NXOR is used below to check for equality
+ switch (instr.flow_control.op) {
+ case Instruction::FlowControlType::Or:
+ mov(eax, COND0);
+ mov(ebx, COND1);
+ xor(eax, (instr.flow_control.refx.Value() ^ 1));
+ xor(ebx, (instr.flow_control.refy.Value() ^ 1));
+ or (eax, ebx);
+ break;
+
+ case Instruction::FlowControlType::And:
+ mov(eax, COND0);
+ mov(ebx, COND1);
+ xor(eax, (instr.flow_control.refx.Value() ^ 1));
+ xor(ebx, (instr.flow_control.refy.Value() ^ 1));
+ and(eax, ebx);
+ break;
+
+ case Instruction::FlowControlType::JustX:
+ mov(eax, COND0);
+ xor(eax, (instr.flow_control.refx.Value() ^ 1));
+ break;
+
+ case Instruction::FlowControlType::JustY:
+ mov(eax, COND1);
+ xor(eax, (instr.flow_control.refy.Value() ^ 1));
+ break;
+ }
+}
+
+void JitShader::Compile_UniformCondition(Instruction instr) {
+ size_t offset = ShaderSetup::GetBoolUniformOffset(instr.flow_control.bool_uniform_id);
+ cmp(byte[SETUP + offset], 0);
+}
+
+BitSet32 JitShader::PersistentCallerSavedRegs() {
+ return persistent_regs & ABI_ALL_CALLER_SAVED;
+}
+
+void JitShader::Compile_ADD(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+ addps(SRC1, SRC2);
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_DP3(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+
+ Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
+
+ movaps(SRC2, SRC1);
+ shufps(SRC2, SRC2, _MM_SHUFFLE(1, 1, 1, 1));
+
+ movaps(SRC3, SRC1);
+ shufps(SRC3, SRC3, _MM_SHUFFLE(2, 2, 2, 2));
+
+ shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0));
+ addps(SRC1, SRC2);
+ addps(SRC1, SRC3);
+
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_DP4(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+
+ Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
+
+ movaps(SRC2, SRC1);
+ shufps(SRC1, SRC1, _MM_SHUFFLE(2, 3, 0, 1)); // XYZW -> ZWXY
+ addps(SRC1, SRC2);
+
+ movaps(SRC2, SRC1);
+ shufps(SRC1, SRC1, _MM_SHUFFLE(0, 1, 2, 3)); // XYZW -> WZYX
+ addps(SRC1, SRC2);
+
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_DPH(Instruction instr) {
+ if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::DPHI) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
+ } else {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+ }
+
+ if (Common::GetCPUCaps().sse4_1) {
+ // Set 4th component to 1.0
+ blendps(SRC1, ONE, 0b1000);
+ } else {
+ // Set 4th component to 1.0
+ movaps(SCRATCH, SRC1);
+ unpckhps(SCRATCH, ONE); // XYZW, 1111 -> Z1__
+ unpcklpd(SRC1, SCRATCH); // XYZW, Z1__ -> XYZ1
+ }
+
+ Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
+
+ movaps(SRC2, SRC1);
+ shufps(SRC1, SRC1, _MM_SHUFFLE(2, 3, 0, 1)); // XYZW -> ZWXY
+ addps(SRC1, SRC2);
+
+ movaps(SRC2, SRC1);
+ shufps(SRC1, SRC1, _MM_SHUFFLE(0, 1, 2, 3)); // XYZW -> WZYX
+ addps(SRC1, SRC2);
+
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_EX2(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ movss(xmm0, SRC1); // ABI_PARAM1
+
+ ABI_PushRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
+ CallFarFunction(*this, exp2f);
+ ABI_PopRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
+
+ shufps(xmm0, xmm0, _MM_SHUFFLE(0, 0, 0, 0)); // ABI_RETURN
+ movaps(SRC1, xmm0);
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_LG2(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ movss(xmm0, SRC1); // ABI_PARAM1
+
+ ABI_PushRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
+ CallFarFunction(*this, log2f);
+ ABI_PopRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
+
+ shufps(xmm0, xmm0, _MM_SHUFFLE(0, 0, 0, 0)); // ABI_RETURN
+ movaps(SRC1, xmm0);
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_MUL(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+ Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_SGE(Instruction instr) {
+ if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::SGEI) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
+ } else {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+ }
+
+ cmpleps(SRC2, SRC1);
+ andps(SRC2, ONE);
+
+ Compile_DestEnable(instr, SRC2);
+}
+
+void JitShader::Compile_SLT(Instruction instr) {
+ if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::SLTI) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
+ } else {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+ }
+
+ cmpltps(SRC1, SRC2);
+ andps(SRC1, ONE);
+
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_FLR(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+
+ if (Common::GetCPUCaps().sse4_1) {
+ roundps(SRC1, SRC1, _MM_FROUND_FLOOR);
+ } else {
+ cvttps2dq(SRC1, SRC1);
+ cvtdq2ps(SRC1, SRC1);
+ }
+
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_MAX(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+ // SSE semantics match PICA200 ones: In case of NaN, SRC2 is returned.
+ maxps(SRC1, SRC2);
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_MIN(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+ // SSE semantics match PICA200 ones: In case of NaN, SRC2 is returned.
+ minps(SRC1, SRC2);
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_MOVA(Instruction instr) {
+ SwizzlePattern swiz = {(*swizzle_data)[instr.common.operand_desc_id]};
+
+ if (!swiz.DestComponentEnabled(0) && !swiz.DestComponentEnabled(1)) {
+ return; // NoOp
+ }
+
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+
+ // Convert floats to integers using truncation (only care about X and Y components)
+ cvttps2dq(SRC1, SRC1);
+
+ // Get result
+ movq(rax, SRC1);
+
+ // Handle destination enable
+ if (swiz.DestComponentEnabled(0) && swiz.DestComponentEnabled(1)) {
+ // Move and sign-extend low 32 bits
+ movsxd(ADDROFFS_REG_0, eax);
+
+ // Move and sign-extend high 32 bits
+ shr(rax, 32);
+ movsxd(ADDROFFS_REG_1, eax);
+
+ // Multiply by 16 to be used as an offset later
+ shl(ADDROFFS_REG_0, 4);
+ shl(ADDROFFS_REG_1, 4);
+ } else {
+ if (swiz.DestComponentEnabled(0)) {
+ // Move and sign-extend low 32 bits
+ movsxd(ADDROFFS_REG_0, eax);
+
+ // Multiply by 16 to be used as an offset later
+ shl(ADDROFFS_REG_0, 4);
+ } else if (swiz.DestComponentEnabled(1)) {
+ // Move and sign-extend high 32 bits
+ shr(rax, 32);
+ movsxd(ADDROFFS_REG_1, eax);
+
+ // Multiply by 16 to be used as an offset later
+ shl(ADDROFFS_REG_1, 4);
+ }
+ }
+}
+
+void JitShader::Compile_MOV(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_RCP(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+
+ // TODO(bunnei): RCPSS is a pretty rough approximation, this might cause problems if Pica
+ // performs this operation more accurately. This should be checked on hardware.
+ rcpss(SRC1, SRC1);
+ shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
+
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_RSQ(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+
+ // TODO(bunnei): RSQRTSS is a pretty rough approximation, this might cause problems if Pica
+ // performs this operation more accurately. This should be checked on hardware.
+ rsqrtss(SRC1, SRC1);
+ shufps(SRC1, SRC1, _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
+
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_NOP(Instruction instr) {}
+
+void JitShader::Compile_END(Instruction instr) {
+ ABI_PopRegistersAndAdjustStack(*this, ABI_ALL_CALLEE_SAVED, 8);
+ ret();
+}
+
+void JitShader::Compile_CALL(Instruction instr) {
+ // Push offset of the return
+ push(qword, (instr.flow_control.dest_offset + instr.flow_control.num_instructions));
+
+ // Call the subroutine
+ call(instruction_labels[instr.flow_control.dest_offset]);
+
+ // Skip over the return offset that's on the stack
+ add(rsp, 8);
+}
+
+void JitShader::Compile_CALLC(Instruction instr) {
+ Compile_EvaluateCondition(instr);
+ Label b;
+ jz(b);
+ Compile_CALL(instr);
+ L(b);
+}
+
+void JitShader::Compile_CALLU(Instruction instr) {
+ Compile_UniformCondition(instr);
+ Label b;
+ jz(b);
+ Compile_CALL(instr);
+ L(b);
+}
+
+void JitShader::Compile_CMP(Instruction instr) {
+ using Op = Instruction::Common::CompareOpType::Op;
+ Op op_x = instr.common.compare_op.x;
+ Op op_y = instr.common.compare_op.y;
+
+ Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+ Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+
+ // SSE doesn't have greater-than (GT) or greater-equal (GE) comparison operators. You need to
+ // emulate them by swapping the lhs and rhs and using LT and LE. NLT and NLE can't be used here
+ // because they don't match when used with NaNs.
+ static const u8 cmp[] = {CMP_EQ, CMP_NEQ, CMP_LT, CMP_LE, CMP_LT, CMP_LE};
+
+ bool invert_op_x = (op_x == Op::GreaterThan || op_x == Op::GreaterEqual);
+ Xmm lhs_x = invert_op_x ? SRC2 : SRC1;
+ Xmm rhs_x = invert_op_x ? SRC1 : SRC2;
+
+ if (op_x == op_y) {
+ // Compare X-component and Y-component together
+ cmpps(lhs_x, rhs_x, cmp[op_x]);
+ movq(COND0, lhs_x);
+
+ mov(COND1, COND0);
+ } else {
+ bool invert_op_y = (op_y == Op::GreaterThan || op_y == Op::GreaterEqual);
+ Xmm lhs_y = invert_op_y ? SRC2 : SRC1;
+ Xmm rhs_y = invert_op_y ? SRC1 : SRC2;
+
+ // Compare X-component
+ movaps(SCRATCH, lhs_x);
+ cmpss(SCRATCH, rhs_x, cmp[op_x]);
+
+ // Compare Y-component
+ cmpps(lhs_y, rhs_y, cmp[op_y]);
+
+ movq(COND0, SCRATCH);
+ movq(COND1, lhs_y);
+ }
+
+ shr(COND0.cvt32(), 31); // ignores upper 32 bits in source
+ shr(COND1, 63);
+}
+
+void JitShader::Compile_MAD(Instruction instr) {
+ Compile_SwizzleSrc(instr, 1, instr.mad.src1, SRC1);
+
+ if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
+ Compile_SwizzleSrc(instr, 2, instr.mad.src2i, SRC2);
+ Compile_SwizzleSrc(instr, 3, instr.mad.src3i, SRC3);
+ } else {
+ Compile_SwizzleSrc(instr, 2, instr.mad.src2, SRC2);
+ Compile_SwizzleSrc(instr, 3, instr.mad.src3, SRC3);
+ }
+
+ Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
+ addps(SRC1, SRC3);
+
+ Compile_DestEnable(instr, SRC1);
+}
+
+void JitShader::Compile_IF(Instruction instr) {
+ Compile_Assert(instr.flow_control.dest_offset >= program_counter,
+ "Backwards if-statements not supported");
+ Label l_else, l_endif;
+
+ // Evaluate the "IF" condition
+ if (instr.opcode.Value() == OpCode::Id::IFU) {
+ Compile_UniformCondition(instr);
+ } else if (instr.opcode.Value() == OpCode::Id::IFC) {
+ Compile_EvaluateCondition(instr);
+ }
+ jz(l_else, T_NEAR);
+
+ // Compile the code that corresponds to the condition evaluating as true
+ Compile_Block(instr.flow_control.dest_offset);
+
+ // If there isn't an "ELSE" condition, we are done here
+ if (instr.flow_control.num_instructions == 0) {
+ L(l_else);
+ return;
+ }
+
+ jmp(l_endif, T_NEAR);
+
+ L(l_else);
+ // This code corresponds to the "ELSE" condition
+ // Comple the code that corresponds to the condition evaluating as false
+ Compile_Block(instr.flow_control.dest_offset + instr.flow_control.num_instructions);
+
+ L(l_endif);
+}
+
+void JitShader::Compile_LOOP(Instruction instr) {
+ Compile_Assert(instr.flow_control.dest_offset >= program_counter,
+ "Backwards loops not supported");
+ Compile_Assert(!looping, "Nested loops not supported");
+
+ looping = true;
+
+ // This decodes the fields from the integer uniform at index instr.flow_control.int_uniform_id.
+ // The Y (LOOPCOUNT_REG) and Z (LOOPINC) component are kept multiplied by 16 (Left shifted by
+ // 4 bits) to be used as an offset into the 16-byte vector registers later
+ size_t offset = ShaderSetup::GetIntUniformOffset(instr.flow_control.int_uniform_id);
+ mov(LOOPCOUNT, dword[SETUP + offset]);
+ mov(LOOPCOUNT_REG, LOOPCOUNT);
+ shr(LOOPCOUNT_REG, 4);
+ and(LOOPCOUNT_REG, 0xFF0); // Y-component is the start
+ mov(LOOPINC, LOOPCOUNT);
+ shr(LOOPINC, 12);
+ and(LOOPINC, 0xFF0); // Z-component is the incrementer
+ movzx(LOOPCOUNT, LOOPCOUNT.cvt8()); // X-component is iteration count
+ add(LOOPCOUNT, 1); // Iteration count is X-component + 1
+
+ Label l_loop_start;
+ L(l_loop_start);
+
+ Compile_Block(instr.flow_control.dest_offset + 1);
+
+ add(LOOPCOUNT_REG, LOOPINC); // Increment LOOPCOUNT_REG by Z-component
+ sub(LOOPCOUNT, 1); // Increment loop count by 1
+ jnz(l_loop_start); // Loop if not equal
+
+ looping = false;
+}
+
+void JitShader::Compile_JMP(Instruction instr) {
+ if (instr.opcode.Value() == OpCode::Id::JMPC)
+ Compile_EvaluateCondition(instr);
+ else if (instr.opcode.Value() == OpCode::Id::JMPU)
+ Compile_UniformCondition(instr);
+ else
+ UNREACHABLE();
+
+ bool inverted_condition =
+ (instr.opcode.Value() == OpCode::Id::JMPU) && (instr.flow_control.num_instructions & 1);
+
+ Label& b = instruction_labels[instr.flow_control.dest_offset];
+ if (inverted_condition) {
+ jz(b, T_NEAR);
+ } else {
+ jnz(b, T_NEAR);
+ }
+}
+
+void JitShader::Compile_Block(unsigned end) {
+ while (program_counter < end) {
+ Compile_NextInstr();
+ }
+}
+
+void JitShader::Compile_Return() {
+ // Peek return offset on the stack and check if we're at that offset
+ mov(rax, qword[rsp + 8]);
+ cmp(eax, (program_counter));
+
+ // If so, jump back to before CALL
+ Label b;
+ jnz(b);
+ ret();
+ L(b);
+}
+
+void JitShader::Compile_NextInstr() {
+ if (std::binary_search(return_offsets.begin(), return_offsets.end(), program_counter)) {
+ Compile_Return();
+ }
+
+ L(instruction_labels[program_counter]);
+
+ Instruction instr = {(*program_code)[program_counter++]};
+
+ OpCode::Id opcode = instr.opcode.Value();
+ auto instr_func = instr_table[static_cast<unsigned>(opcode)];
+
+ if (instr_func) {
+ // JIT the instruction!
+ ((*this).*instr_func)(instr);
+ } else {
+ // Unhandled instruction
+ LOG_CRITICAL(HW_GPU, "Unhandled instruction: 0x%02x (0x%08x)",
+ instr.opcode.Value().EffectiveOpCode(), instr.hex);
+ }
+}
+
+void JitShader::FindReturnOffsets() {
+ return_offsets.clear();
+
+ for (size_t offset = 0; offset < program_code->size(); ++offset) {
+ Instruction instr = {(*program_code)[offset]};
+
+ switch (instr.opcode.Value()) {
+ case OpCode::Id::CALL:
+ case OpCode::Id::CALLC:
+ case OpCode::Id::CALLU:
+ return_offsets.push_back(instr.flow_control.dest_offset +
+ instr.flow_control.num_instructions);
+ break;
+ default:
+ break;
+ }
+ }
+
+ // Sort for efficient binary search later
+ std::sort(return_offsets.begin(), return_offsets.end());
+}
+
+void JitShader::Compile(const std::array<u32, 1024>* program_code_,
+ const std::array<u32, 1024>* swizzle_data_) {
+ program_code = program_code_;
+ swizzle_data = swizzle_data_;
+
+ // Reset flow control state
+ program = (CompiledShader*)getCurr();
+ program_counter = 0;
+ looping = false;
+ instruction_labels.fill(Xbyak::Label());
+
+ // Find all `CALL` instructions and identify return locations
+ FindReturnOffsets();
+
+ // The stack pointer is 8 modulo 16 at the entry of a procedure
+ ABI_PushRegistersAndAdjustStack(*this, ABI_ALL_CALLEE_SAVED, 8);
+
+ mov(SETUP, ABI_PARAM1);
+ mov(STATE, ABI_PARAM2);
+
+ // Zero address/loop registers
+ xor(ADDROFFS_REG_0.cvt32(), ADDROFFS_REG_0.cvt32());
+ xor(ADDROFFS_REG_1.cvt32(), ADDROFFS_REG_1.cvt32());
+ xor(LOOPCOUNT_REG, LOOPCOUNT_REG);
+
+ // Used to set a register to one
+ static const __m128 one = {1.f, 1.f, 1.f, 1.f};
+ mov(rax, reinterpret_cast<size_t>(&one));
+ movaps(ONE, xword[rax]);
+
+ // Used to negate registers
+ static const __m128 neg = {-0.f, -0.f, -0.f, -0.f};
+ mov(rax, reinterpret_cast<size_t>(&neg));
+ movaps(NEGBIT, xword[rax]);
+
+ // Jump to start of the shader program
+ jmp(ABI_PARAM3);
+
+ // Compile entire program
+ Compile_Block(static_cast<unsigned>(program_code->size()));
+
+ // Free memory that's no longer needed
+ program_code = nullptr;
+ swizzle_data = nullptr;
+ return_offsets.clear();
+ return_offsets.shrink_to_fit();
+
+ ready();
+
+ ASSERT_MSG(getSize() <= MAX_SHADER_SIZE, "Compiled a shader that exceeds the allocated size!");
+ LOG_DEBUG(HW_GPU, "Compiled shader size=%lu", getSize());
+}
+
+JitShader::JitShader() : Xbyak::CodeGenerator(MAX_SHADER_SIZE) {}
+
+} // namespace Shader
+
+} // namespace Pica
diff --git a/src/video_core/shader/shader_jit_x64_compiler.h b/src/video_core/shader/shader_jit_x64_compiler.h
new file mode 100644
index 000000000..29e9875ea
--- /dev/null
+++ b/src/video_core/shader/shader_jit_x64_compiler.h
@@ -0,0 +1,125 @@
+// Copyright 2015 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <array>
+#include <cstddef>
+#include <utility>
+#include <vector>
+#include <nihstro/shader_bytecode.h>
+#include <xbyak.h>
+#include "common/bit_set.h"
+#include "common/common_types.h"
+#include "common/x64/emitter.h"
+#include "video_core/shader/shader.h"
+
+using nihstro::Instruction;
+using nihstro::OpCode;
+using nihstro::SwizzlePattern;
+
+namespace Pica {
+
+namespace Shader {
+
+/// Memory allocated for each compiled shader (64Kb)
+constexpr size_t MAX_SHADER_SIZE = 1024 * 64;
+
+/**
+ * This class implements the shader JIT compiler. It recompiles a Pica shader program into x86_64
+ * code that can be executed on the host machine directly.
+ */
+class JitShader : public Xbyak::CodeGenerator {
+public:
+ JitShader();
+
+ void Run(const ShaderSetup& setup, UnitState& state, unsigned offset) const {
+ program(&setup, &state, instruction_labels[offset].getAddress());
+ }
+
+ void Compile(const std::array<u32, 1024>* program_code,
+ const std::array<u32, 1024>* swizzle_data);
+
+ void Compile_ADD(Instruction instr);
+ void Compile_DP3(Instruction instr);
+ void Compile_DP4(Instruction instr);
+ void Compile_DPH(Instruction instr);
+ void Compile_EX2(Instruction instr);
+ void Compile_LG2(Instruction instr);
+ void Compile_MUL(Instruction instr);
+ void Compile_SGE(Instruction instr);
+ void Compile_SLT(Instruction instr);
+ void Compile_FLR(Instruction instr);
+ void Compile_MAX(Instruction instr);
+ void Compile_MIN(Instruction instr);
+ void Compile_RCP(Instruction instr);
+ void Compile_RSQ(Instruction instr);
+ void Compile_MOVA(Instruction instr);
+ void Compile_MOV(Instruction instr);
+ void Compile_NOP(Instruction instr);
+ void Compile_END(Instruction instr);
+ void Compile_CALL(Instruction instr);
+ void Compile_CALLC(Instruction instr);
+ void Compile_CALLU(Instruction instr);
+ void Compile_IF(Instruction instr);
+ void Compile_LOOP(Instruction instr);
+ void Compile_JMP(Instruction instr);
+ void Compile_CMP(Instruction instr);
+ void Compile_MAD(Instruction instr);
+
+private:
+ void Compile_Block(unsigned end);
+ void Compile_NextInstr();
+
+ void Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg,
+ Xbyak::Xmm dest);
+ void Compile_DestEnable(Instruction instr, Xbyak::Xmm dest);
+
+ /**
+ * Compiles a `MUL src1, src2` operation, properly handling the PICA semantics when multiplying
+ * zero by inf. Clobbers `src2` and `scratch`.
+ */
+ void Compile_SanitizedMul(Xbyak::Xmm src1, Xbyak::Xmm src2, Xbyak::Xmm scratch);
+
+ void Compile_EvaluateCondition(Instruction instr);
+ void Compile_UniformCondition(Instruction instr);
+
+ /**
+ * Emits the code to conditionally return from a subroutine envoked by the `CALL` instruction.
+ */
+ void Compile_Return();
+
+ BitSet32 PersistentCallerSavedRegs();
+
+ /**
+ * Assertion evaluated at compile-time, but only triggered if executed at runtime.
+ * @param msg Message to be logged if the assertion fails.
+ */
+ void Compile_Assert(bool condition, const char* msg);
+
+ /**
+ * Analyzes the entire shader program for `CALL` instructions before emitting any code,
+ * identifying the locations where a return needs to be inserted.
+ */
+ void FindReturnOffsets();
+
+ const std::array<u32, 1024>* program_code = nullptr;
+ const std::array<u32, 1024>* swizzle_data = nullptr;
+
+ /// Mapping of Pica VS instructions to pointers in the emitted code
+ std::array<Xbyak::Label, 1024> instruction_labels;
+
+ /// Offsets in code where a return needs to be inserted
+ std::vector<unsigned> return_offsets;
+
+ unsigned program_counter = 0; ///< Offset of the next instruction to decode
+ bool looping = false; ///< True if compiling a loop, used to check for nested loops
+
+ using CompiledShader = void(const void* setup, void* state, const u8* start_addr);
+ CompiledShader* program = nullptr;
+};
+
+} // Shader
+
+} // Pica