CrossTL - Universal Shader Translator

CrossTL is a shader and compute program translator built around CrossGL — an intermediate representation (IR) language that bridges graphics APIs, GPU-compute platforms, and systems-language targets.

Supported Translation Targets

CrossTL provides translation from CrossGL to:

• Metal - Apple's graphics and compute API
• DirectX (HLSL) - Microsoft's graphics API
• OpenGL (GLSL) - Cross-platform graphics
• WebGL (GLSL ES) - Browser graphics through WebGL 2.0 compatible GLSL ES
• WebGPU (WGSL) - Browser and native WebGPU shader output
• Vulkan (SPIRV) - High-performance graphics and compute
• CUDA - NVIDIA parallel computing
• HIP - AMD GPU computing
• Rust - GPU-oriented Rust shader code
• Mojo - Mojo shader/compute modules
• Slang - Real-time shading language
• CrossGL (.cgl) - The IR/interchange format itself

Native source import is available for the bidirectional backends listed in the support matrix. WebGL and WGSL are currently target-only outputs; .webgl.glsl, .wgsl, and .wesl inputs are rejected until dedicated source frontends land.

Backend Readiness: DirectX / Metal / OpenGL

We maintain first-class, bidirectional support for the three cornerstone graphics APIs. Each backend is implemented as both a source (parse/import) and codegen (export) target, so you can round‑trip between native shaders and CrossGL without lossy hops.

• DirectX / HLSL
  • Pipeline coverage: vertex, fragment/pixel, compute, geometry, hull/domain (tessellation), mesh/task, full ray‑tracing stages.
  • Target profile aliases: dx11, dx12, d3d11, and d3d12 resolve to the HLSL emitter for Direct3D deployment planning; final DXBC/DXIL packaging remains a toolchain step.
  • Resource model: cbuffers, register/space bindings, UAV/RW textures & buffers, structured buffers, Interlocked atomics, wave ops, texture/buffer dimension queries.
  • Semantics map to SV_* and user semantics, preserved through CrossGL attributes.
• Metal
  • Stages: vertex, fragment, compute, mesh/object, and ray‑tracing qualifiers.
  • Resource/binding support: argument buffers, [[buffer]], [[texture]], [[sampler]], function constants, packed/simd types, indirect command buffers, payload/hit attributes.
  • Texture methods (sample/load/store/gather/compare) and threadgroup builtins are translated to CrossGL intrinsics.
• OpenGL / GLSL
  • Stages: vertex, fragment, compute with version inference (defaults to #version 450 core when absent).
  • Handles interface blocks, layouts/bindings, sampler & image operations, control flow, structs/arrays, discard, builtins (gl_*) and preprocessor directives.
• Round‑trips verified via converters in crosstl.backend.{DirectX,Metal,GLSL} and registered in translator.source_registry and translator.codegen.registry.

How we validate backend parity

• Targeted unit suites exercise the full feature surface for these three backends (shader stages, bindings, intrinsics, control‑flow, resources, and preprocessor handling).
• Quick check: run pytest tests/test_backend/test_directx tests/test_backend/test_metal tests/test_backend/test_GLSL -q (currently 321 passing tests).
• End‑to‑end translation: translate native HLSL/Metal/GLSL → CrossGL → HLSL/Metal/GLSL/Vulkan to ensure attributes, layouts, and semantics survive round‑trips.

Translation Architecture

CrossTL uses a multi-stage translation pipeline:

1. Lexical Analysis: Tokenization
2. Syntax Analysis: AST generation
3. Semantic Analysis: Type checking and scope resolution
4. IR Generation: Conversion to CrossGL intermediate representation
5. Target Generation: Backend-specific code generation

CrossGL Programming Language Examples

PBR Shader Example

The README demo uses the same current CrossGL source file as the showcase notebook: demos/readme/showcase_pbr.cgl and demos/readme/demo.ipynb. The shader demonstrates a real PBR-style material path: tangent-space normal mapping, GGX normal distribution, Smith geometry, Schlick Fresnel, multiple point lights, texture sampling, tone mapping, and gamma correction.

It uses the current struct-return stage style so backends such as Metal can emit proper stage input/output structs and vertex/fragment entry points.

shader ShowcasePBRShader {
    const float PI = 3.14159265359;
    const float EPSILON = 0.0001;
    const int MAX_LIGHTS = 4;

    struct VertexInput {
        vec3 position;
        vec3 normal;
        vec3 tangent;
        vec2 texCoord;
    }

    struct VertexOutput {
        vec3 worldPosition;
        vec3 worldNormal;
        vec3 worldTangent;
        vec3 worldBitangent;
        vec2 uv;
        vec4 position;
    }

    struct FragmentInput {
        vec3 worldPosition;
        vec3 worldNormal;
        vec3 worldTangent;
        vec3 worldBitangent;
        vec2 uv;
    }

    struct FragmentOutput {
        vec4 color;
    }

    float distributionGGX(vec3 normal, vec3 halfVector, float roughness) {
        float alpha = roughness * roughness;
        float alpha2 = alpha * alpha;
        float ndoth = max(dot(normal, halfVector), 0.0);
        float denom = ndoth * ndoth * (alpha2 - 1.0) + 1.0;
        return alpha2 / max(PI * denom * denom, EPSILON);
    }

    float geometrySchlickGGX(float ndotv, float roughness) {
        float k = (roughness + 1.0) * (roughness + 1.0) / 8.0;
        return ndotv / max(ndotv * (1.0 - k) + k, EPSILON);
    }

    float geometrySmith(vec3 normal, vec3 viewDir, vec3 lightDir, float roughness) {
        float ndotv = max(dot(normal, viewDir), 0.0);
        float ndotl = max(dot(normal, lightDir), 0.0);
        return geometrySchlickGGX(ndotv, roughness) * geometrySchlickGGX(ndotl, roughness);
    }

    vec3 fresnelSchlick(float cosTheta, vec3 baseReflectance) {
        return baseReflectance + (vec3(1.0) - baseReflectance) * pow(max(1.0 - cosTheta, 0.0), 5.0);
    }

    vertex {
        uniform mat4 modelMatrix;
        uniform mat4 viewProjectionMatrix;
        uniform mat3 normalMatrix;

        VertexOutput main(VertexInput input) {
            VertexOutput output;
            vec4 worldPosition = modelMatrix * vec4(input.position, 1.0);
            output.worldPosition = worldPosition.xyz;
            output.worldNormal = normalize(normalMatrix * input.normal);
            output.worldTangent = normalize(normalMatrix * input.tangent);
            output.worldBitangent = normalize(cross(output.worldNormal, output.worldTangent));
            output.uv = input.texCoord;
            output.position = viewProjectionMatrix * worldPosition;
            return output;
        }
    }

    fragment {
        uniform sampler2D albedoMap;
        uniform sampler2D normalMap;
        uniform sampler2D metallicRoughnessMap;
        uniform vec3 cameraPosition;
        uniform vec3 lightPositions[MAX_LIGHTS];
        uniform vec3 lightColors[MAX_LIGHTS];
        uniform int activeLightCount;
        uniform vec3 baseColor;
        uniform float metallicFactor;
        uniform float roughnessFactor;

        vec3 decodeNormal(vec3 encodedNormal, mat3 tangentFrame) {
            vec3 tangentNormal = encodedNormal * 2.0 - 1.0;
            return normalize(tangentFrame * tangentNormal);
        }

        FragmentOutput main(FragmentInput input) {
            FragmentOutput output;
            mat3 tangentFrame = mat3(
                normalize(input.worldTangent),
                normalize(input.worldBitangent),
                normalize(input.worldNormal)
            );
            vec3 normal = decodeNormal(texture(normalMap, input.uv).rgb, tangentFrame);
            vec3 viewDir = normalize(cameraPosition - input.worldPosition);
            vec3 albedo = texture(albedoMap, input.uv).rgb * baseColor;
            vec3 materialSample = texture(metallicRoughnessMap, input.uv).rgb;
            float metallic = clamp(materialSample.b * metallicFactor, 0.0, 1.0);
            float roughness = clamp(materialSample.g * roughnessFactor, 0.04, 1.0);
            vec3 baseReflectance = mix(vec3(0.04), albedo, metallic);
            vec3 radianceSum = vec3(0.0);

            for (int i = 0; i < MAX_LIGHTS; i++) {
                if (i >= activeLightCount) {
                    break;
                }
                vec3 lightVector = lightPositions[i] - input.worldPosition;
                float distanceSq = max(dot(lightVector, lightVector), EPSILON);
                vec3 lightDir = normalize(lightVector);
                vec3 halfVector = normalize(viewDir + lightDir);
                float attenuation = 1.0 / distanceSq;
                vec3 radiance = lightColors[i] * attenuation;

                float normalDistribution = distributionGGX(normal, halfVector, roughness);
                float geometry = geometrySmith(normal, viewDir, lightDir, roughness);
                vec3 fresnel = fresnelSchlick(max(dot(halfVector, viewDir), 0.0), baseReflectance);
                vec3 diffuseShare = (vec3(1.0) - fresnel) * (1.0 - metallic);
                vec3 numerator = normalDistribution * geometry * fresnel;
                float denominator = max(4.0 * max(dot(normal, viewDir), 0.0) * max(dot(normal, lightDir), 0.0), EPSILON);
                vec3 specular = numerator / denominator;
                float ndotl = max(dot(normal, lightDir), 0.0);
                radianceSum += (diffuseShare * albedo / PI + specular) * radiance * ndotl;
            }

            vec3 ambient = albedo * 0.03;
            vec3 color = ambient + radianceSum;
            color = color / (color + vec3(1.0));
            color = pow(color, vec3(1.0 / 2.2));
            output.color = vec4(color, 1.0);
            return output;
        }
    }
}

Run the showcase notebook or translate the source directly:

python -m crosstl translate demos/readme/showcase_pbr.cgl --backend metal --output -
python -m crosstl translate demos/readme/showcase_pbr.cgl --backend directx --output -
python -m crosstl translate demos/readme/showcase_pbr.cgl --backend opengl --output -

Backend Compatibility

All backends support the core CrossGL language (structs, arrays, functions, control flow, texture sampling, shader stages). Some advanced language features have partial backend coverage:

Translation calls targeting unsupported feature/backend combinations raise a clear error with a reason string.

Getting Started

Install CrossTL:

pip install crosstl

Basic Usage

1. Create a CrossGL shader

shader SimpleShader {
    struct VertexInput {
        vec3 position;
        vec2 texCoord;
    }

    struct VertexOutput {
        vec2 uv;
        vec4 position;
    }

    struct FragmentInput {
        vec2 uv;
    }

    struct FragmentOutput {
        vec4 color;
    }

    vertex {
        VertexOutput main(VertexInput input) {
            VertexOutput output;
            output.uv = input.texCoord;
            output.position = vec4(input.position, 1.0);
            return output;
        }
    }

    fragment {
        FragmentOutput main(FragmentInput input) {
            FragmentOutput output;
            output.color = vec4(input.uv, 0.5, 1.0);
            return output;
        }
    }
}

2. Translate to a target

import crosstl

# Translate to any supported backend
metal_code = crosstl.translate('shader.cgl', backend='metal', save_shader='shader.metal')
hlsl_code = crosstl.translate('shader.cgl', backend='directx', save_shader='shader.hlsl')
dx12_hlsl = crosstl.translate('shader.cgl', backend='dx12', save_shader='shader.hlsl')
glsl_code = crosstl.translate('shader.cgl', backend='opengl', save_shader='shader.glsl')

Project audit workflow

For repositories with many shader or GPU source files, start with a scan-only report before writing translated artifacts:

python -m crosstl scan /path/to/repo \
  --target metal \
  --output scan-report.json

python -m crosstl translate-project /path/to/repo \
  --target metal \
  --output-dir crosstl-out \
  --report crosstl-out/portability-report.json

python -m crosstl validate-project \
  crosstl-out/portability-report.json \
  --format text

python -m crosstl inspect-report \
  crosstl-out/portability-report.json \
  --format text

Project reports cover shader and kernel source translation, diagnostics, artifact provenance, optional toolchain checks, and manual migration actions for host/runtime integration. See the project porting guide. The repository also includes pinned open-source porting demos with checked artifacts and platform validator coverage in demos/open-source-porting.

Reverse Translation - Import Existing Code

# Import existing shaders into CrossGL
conversions = [
    ('existing_shader.hlsl', 'unified.cgl'),     # DirectX to CrossGL
    ('gpu_kernel.cu', 'unified.cgl'),            # CUDA to CrossGL
    ('graphics.metal', 'unified.cgl'),           # Metal to CrossGL
]

for source, target in conversions:
    unified_code = crosstl.translate(source, backend='cgl', save_shader=target)
    print(f"Unified {source} into CrossGL: {target}")

Deploying to All Backends

import crosstl
from pathlib import Path

program = 'universal_shader.cgl'

deployment_targets = {
    'metal': '.metal',
    'directx': '.hlsl',
    'dx12': '.hlsl',
    'opengl': '.glsl',
    'vulkan': '.spvasm',
    'rust': '.rs',
    'mojo': '.mojo',
    'cuda': '.cu',
    'hip': '.hip',
    'slang': '.slang',
}

stem = Path(program).stem
for backend, extension in deployment_targets.items():
    output_file = f'deployments/{stem}_{backend}{extension}'
    try:
        crosstl.translate(program, backend=backend, save_shader=output_file)
        print(f"{backend}: {output_file}")
    except Exception as e:
        print(f"{backend}: {str(e)}")

For comprehensive language documentation, visit our Language Reference.

Contributing

CrossGL is a community-driven project. Find out more in our Contributing Guide.

Community

• Twitter
• LinkedIn
• Discord
• YouTube

License

CrossTL is open-source and licensed under the Apache License 2.0.

The CrossGL Team