Merge cf49b058f4 into b15b5e4728

XenosRecomp/CMakeLists.txt

@@ -4,10 +4,16 @@ if (WIN32)
     option(XENOS_RECOMP_DXIL "Generate DXIL shader cache" ON)
 endif()
 
+if (APPLE)
+    option(XENOS_RECOMP_AIR "Generate Metal AIR shader cache" ON)
+endif()
+
 set(SMOLV_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../thirdparty/smol-v/source")
 
-add_executable(XenosRecomp 
+add_executable(XenosRecomp
     constant_table.h
+    air_compiler.cpp
+    air_compiler.h
     dxc_compiler.cpp
     dxc_compiler.h
     main.cpp
@@ -51,3 +57,7 @@ if (XENOS_RECOMP_DXIL)
     target_compile_definitions(XenosRecomp PRIVATE XENOS_RECOMP_DXIL)
     target_link_libraries(XenosRecomp PRIVATE Microsoft::DXIL)
 endif()
+
+if (XENOS_RECOMP_AIR)
+    target_compile_definitions(XenosRecomp PRIVATE XENOS_RECOMP_AIR)
+endif()

XenosRecomp/air_compiler.cpp

@@ -0,0 +1,109 @@
+#include "air_compiler.h"
+
+#include <iostream>
+#include <spawn.h>
+#include <unistd.h>
+#include <sys/wait.h>
+#include <cstdio>
+
+std::vector<uint8_t> AirCompiler::compile(const std::string& shaderSource) {
+    // First, generate AIR from shader source
+    std::string inputFile = ".metal";
+    int tmpFD = makeTemporaryFile(inputFile);
+    write(tmpFD, shaderSource.data(), shaderSource.size());
+    close(tmpFD);
+
+    std::string irFile = ".ir";
+    tmpFD = makeTemporaryFile(irFile);
+    close(tmpFD);
+
+    pid_t pid;
+    char* airArgv[] = { "xcrun", "-sdk", "macosx", "metal", "-o", irFile.data(), "-c", inputFile.data(), "-D__air__", "-DUNLEASHED_RECOMP", "-Wno-unused-variable", "-frecord-sources", "-gline-tables-only", nullptr };
+
+    if (posix_spawn(&pid, "/usr/bin/xcrun", nullptr, nullptr, airArgv, nullptr) != 0) {
+        unlink(inputFile.data());
+        unlink(irFile.data());
+        fmt::println("Failed to spawn AIR xcrun process");
+        exit(1);
+    }
+
+    int status;
+
+    if (waitpid(pid, &status, 0) == -1) {
+        unlink(inputFile.data());
+        unlink(irFile.data());
+        fmt::println("Failed to wait AIR xcrun process");
+        exit(1);
+    }
+
+    if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
+        unlink(inputFile.data());
+        unlink(irFile.data());
+        fmt::println("AIR xcrun exited with code {}", WEXITSTATUS(status));
+        fmt::println("{}", shaderSource);
+        exit(1);
+    }
+
+    unlink(inputFile.data());
+
+    // Now we need to turn the AIR into a .metallib
+    std::string libFile = ".metallib";
+    tmpFD = makeTemporaryFile(libFile);
+    close(tmpFD);
+
+    char* libArgv[] = { "xcrun", "-sdk", "macosx", "metallib", "-o", libFile.data(), irFile.data(), nullptr };
+
+    if (posix_spawn(&pid, "/usr/bin/xcrun", nullptr, nullptr, libArgv, nullptr) != 0) {
+        unlink(irFile.data());
+        unlink(libFile.data());
+        fmt::println("Failed to spawn .metallib xcrun process");
+        exit(1);
+    }
+
+    if (waitpid(pid, &status, 0) == -1) {
+        unlink(irFile.data());
+        unlink(libFile.data());
+        fmt::println("Failed to wait .metallib xcrun process");
+        exit(1);
+    }
+
+    if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
+        unlink(irFile.data());
+        unlink(libFile.data());
+        fmt::println(".metallib exited with code {}", WEXITSTATUS(status));
+        exit(1);
+    }
+
+    // Read from built .metallib
+    FILE* file = fopen(libFile.data(), "rb");
+    fseek(file, 0, SEEK_END);
+    size_t fileSize = ftell(file);
+    fseek(file, 0, SEEK_SET);
+    auto data = std::vector<uint8_t>(fileSize);
+    fread(data.data(), 1, fileSize, file);
+    fclose(file);
+
+    // Cleanup temporary files
+    unlink(irFile.data());
+    unlink(libFile.data());
+
+    return data;
+}
+
+int AirCompiler::makeTemporaryFile(std::string &extension) {
+    const std::string path = "/tmp/xenos_metal_XXXXXX";
+
+    size_t size = path.size() + extension.size() + 1;
+    char fullTemplate[size];
+    snprintf(fullTemplate, size, "%s%s", path.c_str(), extension.c_str());
+
+    int tmpFD = mkstemps(fullTemplate, extension.size());
+    if (tmpFD == -1) {
+        fmt::println("Failed to open temporary file, \"{}\"!", std::string(fullTemplate));
+        unlink(fullTemplate);
+        exit(1);
+    }
+
+    extension = fullTemplate;
+    return tmpFD;
+}

XenosRecomp/air_compiler.h

@@ -0,0 +1,9 @@
+#pragma once
+
+struct AirCompiler
+{
+    static std::vector<uint8_t> compile(const std::string& shaderSource);
+
+private:
+    static int makeTemporaryFile(std::string& extension);
+};

XenosRecomp/dxc_compiler.cpp

@@ -34,6 +34,11 @@ IDxcBlob* DxcCompiler::compile(const std::string& shaderSource, bool compilePixe
             target = L"-T vs_6_0";
     }
 
+    if (!compileLibrary)
+    {
+        args[argCount++] = L"-E shaderMain";
+    }
+
     args[argCount++] = target;
     args[argCount++] = L"-HV 2021";
     args[argCount++] = L"-all-resources-bound";

XenosRecomp/main.cpp

@@ -1,6 +1,7 @@
 #include "shader.h"
 #include "shader_recompiler.h"
 #include "dxc_compiler.h"
+#include "air_compiler.h"
 
 static std::unique_ptr<uint8_t[]> readAllBytes(const char* filePath, size_t& fileSize)
 {
@@ -26,6 +27,7 @@ struct RecompiledShader
     uint8_t* data = nullptr;
     IDxcBlob* dxil = nullptr;
     std::vector<uint8_t> spirv;
+    std::vector<uint8_t> air;
     uint32_t specConstantsMask = 0;
 };
 
@@ -133,6 +135,10 @@ int main(int argc, char** argv)
                 assert(*(reinterpret_cast<uint32_t *>(shader.dxil->GetBufferPointer()) + 1) != 0 && "DXIL was not signed properly!");
 #endif
 
+#ifdef XENOS_RECOMP_AIR
+                shader.air = AirCompiler::compile(recompiler.out);
+#endif
+
                 IDxcBlob* spirv = dxcCompiler.compile(recompiler.out, recompiler.isPixelShader, false, true);
                 assert(spirv != nullptr);
 
@@ -154,18 +160,24 @@ int main(int argc, char** argv)
 
         std::vector<uint8_t> dxil;
         std::vector<uint8_t> spirv;
+        std::vector<uint8_t> air;
 
         for (auto& [hash, shader] : shaders)
         {
-            f.println("\t{{ 0x{:X}, {}, {}, {}, {}, {} }},",
-                hash, dxil.size(), (shader.dxil != nullptr) ? shader.dxil->GetBufferSize() : 0, spirv.size(), shader.spirv.size(), shader.specConstantsMask);
+            f.println("\t{{ 0x{:X}, {}, {}, {}, {}, {}, {}, {} }},",
+                hash, dxil.size(), (shader.dxil != nullptr) ? shader.dxil->GetBufferSize() : 0,
+                spirv.size(), shader.spirv.size(), air.size(), shader.air.size(), shader.specConstantsMask);
 
             if (shader.dxil != nullptr)
             {
                 dxil.insert(dxil.end(), reinterpret_cast<uint8_t *>(shader.dxil->GetBufferPointer()),
                     reinterpret_cast<uint8_t *>(shader.dxil->GetBufferPointer()) + shader.dxil->GetBufferSize());
             }
-            
+
+#ifdef XENOS_RECOMP_AIR
+            air.insert(air.end(), shader.air.begin(), shader.air.end());
+#endif
+
             spirv.insert(spirv.end(), shader.spirv.begin(), shader.spirv.end());
         }
 
@@ -189,6 +201,22 @@ int main(int argc, char** argv)
         f.println("const size_t g_dxilCacheDecompressedSize = {};", dxil.size());
 #endif
 
+#ifdef XENOS_RECOMP_AIR
+        fmt::println("Compressing AIR cache...");
+
+        std::vector<uint8_t> airCompressed(ZSTD_compressBound(air.size()));
+        airCompressed.resize(ZSTD_compress(airCompressed.data(), airCompressed.size(), air.data(), air.size(), level));
+
+        f.print("const uint8_t g_compressedAirCache[] = {{");
+
+        for (auto data : airCompressed)
+            f.print("{},", data);
+
+        f.println("}};");
+        f.println("const size_t g_airCacheCompressedSize = {};", airCompressed.size());
+        f.println("const size_t g_airCacheDecompressedSize = {};", air.size());
+#endif
+
         fmt::println("Compressing SPIRV cache...");
 
         std::vector<uint8_t> spirvCompressed(ZSTD_compressBound(spirv.size()));

XenosRecomp/shader_common.h

@@ -10,10 +10,11 @@
     #define SPEC_CONSTANT_REVERSE_Z         (1 << 4)
 #endif
 
-#if !defined(__cplusplus) || defined(__INTELLISENSE__)
+#if defined(__air__) || !defined(__cplusplus) || defined(__INTELLISENSE__)
 
-#define FLT_MIN asfloat(0xff7fffff)
+#ifndef __air__
 #define FLT_MAX asfloat(0x7f7fffff)
+#endif
 
 #ifdef __spirv__
 
@@ -34,6 +35,31 @@ struct PushConstants
 
 #define g_SpecConstants() g_SpecConstants
 
+#elif __air__
+
+#include <metal_stdlib>
+
+using namespace metal;
+
+constant uint G_SPEC_CONSTANTS [[function_constant(0)]];
+constant uint G_SPEC_CONSTANTS_VAL = is_function_constant_defined(G_SPEC_CONSTANTS) ? G_SPEC_CONSTANTS : 0;
+
+uint g_SpecConstants()
+{
+    return G_SPEC_CONSTANTS_VAL;
+}
+
+struct PushConstants
+{
+    ulong VertexShaderConstants;
+    ulong PixelShaderConstants;
+    ulong SharedConstants;
+};
+
+#define g_Booleans (*(reinterpret_cast<device uint*>(g_PushConstants.SharedConstants + 256)))
+#define g_SwappedTexcoords (*(reinterpret_cast<device uint*>(g_PushConstants.SharedConstants + 260)))
+#define g_AlphaThreshold (*(reinterpret_cast<device float*>(g_PushConstants.SharedConstants + 264)))
+
 #else
 
 #define DEFINE_SHARED_CONSTANTS() \
@@ -45,6 +71,56 @@ uint g_SpecConstants();
 
 #endif
 
+#ifdef __air__
+
+struct Texture2DDescriptorHeap
+{
+    array<texture2d<float>, 1> g [[id(0)]];
+};
+
+struct Texture3DDescriptorHeap
+{
+    array<texture3d<float>, 1> g [[id(0)]];
+};
+
+struct TextureCubeDescriptorHeap
+{
+    array<texturecube<float>, 1> g [[id(0)]];
+};
+
+struct SamplerDescriptorHeap
+{
+    array<sampler, 1> g [[id(0)]];
+};
+
+uint2 getTexture2DDimensions(texture2d<float> texture)
+{
+    return uint2(texture.get_width(), texture.get_height());
+}
+
+float4 tfetch2D(constant Texture2DDescriptorHeap& textureHeap,
+                constant SamplerDescriptorHeap& samplerHeap,
+                uint resourceDescriptorIndex,
+                uint samplerDescriptorIndex,
+                float2 texCoord, float2 offset)
+{
+    texture2d<float> texture = textureHeap.g[resourceDescriptorIndex];
+    sampler sampler = samplerHeap.g[samplerDescriptorIndex];
+    return texture.sample(sampler, texCoord + offset / (float2)getTexture2DDimensions(texture));
+}
+
+float2 getWeights2D(constant Texture2DDescriptorHeap& textureHeap,
+                    constant SamplerDescriptorHeap& samplerHeap,
+                    uint resourceDescriptorIndex,
+                    uint samplerDescriptorIndex,
+                    float2 texCoord, float2 offset)
+{
+    texture2d<float> texture = textureHeap.g[resourceDescriptorIndex];
+    return select(fract(texCoord * (float2)getTexture2DDimensions(texture) + offset - 0.5), 0.0, isnan(texCoord));
+}
+
+#else
+
 Texture2D<float4> g_Texture2DDescriptorHeap[] : register(t0, space0);
 Texture3D<float4> g_Texture3DDescriptorHeap[] : register(t0, space1);
 TextureCube<float4> g_TextureCubeDescriptorHeap[] : register(t0, space2);
@@ -69,6 +145,46 @@ float2 getWeights2D(uint resourceDescriptorIndex, uint samplerDescriptorIndex, f
     return select(isnan(texCoord), 0.0, frac(texCoord * getTexture2DDimensions(texture) + offset - 0.5));
 }
 
+#endif
+
+#ifdef __air__
+#define selectWrapper(a, b, c) select(c, b, a)
+#else
+#define selectWrapper(a, b, c) select(a, b, c)
+#endif
+
+#ifdef __air__
+#define frac(X) fract(X)
+
+template<typename T>
+void clip(T a)
+{
+    if (a < 0.0) {
+        discard_fragment();
+    }
+}
+
+template<typename T>
+float rcp(T a)
+{
+    return 1.0 / a;
+}
+
+template<typename T>
+float4x4 mul(T a, T b)
+{
+    return b * a;
+}
+#endif
+
+#ifdef __air__
+#define UNROLL
+#define BRANCH
+#else
+#define UNROLL [unroll]
+#define BRANCH [branch]
+#endif
+
 float w0(float a)
 {
     return (1.0f / 6.0f) * (a * (a * (-a + 3.0f) - 3.0f) + 1.0f);
@@ -109,6 +225,74 @@ float h1(float a)
     return 1.0f + w3(a) / (w2(a) + w3(a)) + 0.5f;
 }
 
+struct CubeMapData
+{
+    float3 cubeMapDirections[2];
+    uint cubeMapIndex;
+};
+
+#ifdef __air__
+
+float4 tfetch2DBicubic(constant Texture2DDescriptorHeap& textureHeap,
+                       constant SamplerDescriptorHeap& samplerHeap,
+                       uint resourceDescriptorIndex,
+                       uint samplerDescriptorIndex,
+                       float2 texCoord, float2 offset)
+{
+    texture2d<float> texture = textureHeap.g[resourceDescriptorIndex];
+    sampler sampler = samplerHeap.g[samplerDescriptorIndex];
+    uint2 dimensions = getTexture2DDimensions(texture);
+
+    float x = texCoord.x * dimensions.x + offset.x;
+    float y = texCoord.y * dimensions.y + offset.y;
+
+    x -= 0.5f;
+    y -= 0.5f;
+    float px = floor(x);
+    float py = floor(y);
+    float fx = x - px;
+    float fy = y - py;
+
+    float g0x = g0(fx);
+    float g1x = g1(fx);
+    float h0x = h0(fx);
+    float h1x = h1(fx);
+    float h0y = h0(fy);
+    float h1y = h1(fy);
+
+    float4 r =
+        g0(fy) * (g0x * texture.sample(sampler, float2(px + h0x, py + h0y) / float2(dimensions)) +
+              g1x * texture.sample(sampler, float2(px + h1x, py + h0y) / float2(dimensions))) +
+        g1(fy) * (g0x * texture.sample(sampler, float2(px + h0x, py + h1y) / float2(dimensions)) +
+              g1x * texture.sample(sampler, float2(px + h1x, py + h1y) / float2(dimensions)));
+
+    return r;
+}
+
+float4 tfetch3D(constant Texture3DDescriptorHeap& textureHeap,
+                constant SamplerDescriptorHeap& samplerHeap,
+                uint resourceDescriptorIndex,
+                uint samplerDescriptorIndex,
+                float3 texCoord)
+{
+    texture3d<float> texture = textureHeap.g[resourceDescriptorIndex];
+    sampler sampler = samplerHeap.g[samplerDescriptorIndex];
+    return texture.sample(sampler, texCoord);
+}
+
+float4 tfetchCube(constant TextureCubeDescriptorHeap& textureHeap,
+                  constant SamplerDescriptorHeap& samplerHeap,
+                  uint resourceDescriptorIndex,
+                  uint samplerDescriptorIndex,
+                  float3 texCoord, thread CubeMapData* cubeMapData)
+{
+    texturecube<float> texture = textureHeap.g[resourceDescriptorIndex];
+    sampler sampler = samplerHeap.g[samplerDescriptorIndex];
+    return texture.sample(sampler, cubeMapData->cubeMapDirections[(uint)texCoord.z]);
+}
+
+#else
+
 float4 tfetch2DBicubic(uint resourceDescriptorIndex, uint samplerDescriptorIndex, float2 texCoord, float2 offset)
 {
     Texture2D<float4> texture = g_Texture2DDescriptorHeap[resourceDescriptorIndex];
@@ -146,17 +330,13 @@ float4 tfetch3D(uint resourceDescriptorIndex, uint samplerDescriptorIndex, float
     return g_Texture3DDescriptorHeap[resourceDescriptorIndex].Sample(g_SamplerDescriptorHeap[samplerDescriptorIndex], texCoord);
 }
 
-struct CubeMapData
-{
-    float3 cubeMapDirections[2];
-    uint cubeMapIndex;
-};
-
 float4 tfetchCube(uint resourceDescriptorIndex, uint samplerDescriptorIndex, float3 texCoord, inout CubeMapData cubeMapData)
 {
     return g_TextureCubeDescriptorHeap[resourceDescriptorIndex].Sample(g_SamplerDescriptorHeap[samplerDescriptorIndex], cubeMapData.cubeMapDirections[texCoord.z]);
 }
 
+#endif
+
 float4 tfetchR11G11B10(uint4 value)
 {
     if (g_SpecConstants() & SPEC_CONSTANT_R11G11B10_NORMAL)
@@ -169,7 +349,11 @@ float4 tfetchR11G11B10(uint4 value)
     }
     else
     {
+#ifdef __air__
+        return as_type<float4>(value);
+#else
         return asfloat(value);
+#endif
     }
 }
 
@@ -178,6 +362,19 @@ float4 tfetchTexcoord(uint swappedTexcoords, float4 value, uint semanticIndex)
     return (swappedTexcoords & (1ull << semanticIndex)) != 0 ? value.yxwz : value;
 }
 
+#ifdef __air__
+
+float4 cube(float4 value, thread CubeMapData* cubeMapData)
+{
+    uint index = cubeMapData->cubeMapIndex;
+    cubeMapData->cubeMapDirections[index] = value.xyz;
+    ++cubeMapData->cubeMapIndex;
+
+    return float4(0.0, 0.0, 0.0, index);
+}
+
+#else
+
 float4 cube(float4 value, inout CubeMapData cubeMapData)
 {
     uint index = cubeMapData.cubeMapIndex;
@@ -187,6 +384,8 @@ float4 cube(float4 value, inout CubeMapData cubeMapData)
     return float4(0.0, 0.0, 0.0, index);
 }
 
+#endif
+
 float4 dst(float4 src0, float4 src1)
 {
     float4 dest;
@@ -202,15 +401,34 @@ float4 max4(float4 src0)
     return max(max(src0.x, src0.y), max(src0.z, src0.w));
 }
 
+#ifdef __air__
+
+float2 getPixelCoord(constant Texture2DDescriptorHeap& textureHeap,
+                     uint resourceDescriptorIndex,
+                     float2 texCoord)
+{
+    texture2d<float> texture = textureHeap.g[resourceDescriptorIndex];
+    return (float2)getTexture2DDimensions(texture) * texCoord;
+}
+
+#else
+
 float2 getPixelCoord(uint resourceDescriptorIndex, float2 texCoord)
 {
     return getTexture2DDimensions(g_Texture2DDescriptorHeap[resourceDescriptorIndex]) * texCoord;
 }
 
+#endif
+
 float computeMipLevel(float2 pixelCoord)
 {
+#ifdef __air__
+    float2 dx = dfdx(pixelCoord);
+    float2 dy = dfdy(pixelCoord);
+#else
     float2 dx = ddx(pixelCoord);
     float2 dy = ddy(pixelCoord);
+#endif
     float deltaMaxSqr = max(dot(dx, dx), dot(dy, dy));
     return max(0.0, 0.5 * log2(deltaMaxSqr));
 }

XenosRecomp/shader_recompiler.h

@@ -45,7 +45,7 @@ struct ShaderRecompiler : StringBuffer
             out += '\t';
     }
 
-    void printDstSwizzle(uint32_t dstSwizzle, bool operand);
+    uint32_t printDstSwizzle(uint32_t dstSwizzle, bool operand);
     void printDstSwizzle01(uint32_t dstRegister, uint32_t dstSwizzle);
 
     void recompile(const VertexFetchInstruction& instr, uint32_t address);