Method compile_kernel

rightnow_cli/compiler.py:92–214

Compile a CUDA kernel and return compilation results.

(self, kernel_data: Dict[str, Any])

            return ["-gencode=arch=compute_70,code=sm_70"]
    
    def compile_kernel(self, kernel_data: Dict[str, Any]) -> Dict[str, Any]:
        """Compile a CUDA kernel and return compilation results."""
        kernel_code = kernel_data.get("code", "")
        if not kernel_code:
            raise ValueError("No kernel code provided")
        
        with tempfile.TemporaryDirectory() as temp_dir:
            temp_path = Path(temp_dir)
            
            cuda_file = temp_path / "kernel.cu"
            ptx_file = temp_path / "kernel.ptx"
            cubin_file = temp_path / "kernel.cubin"
            
            full_code = self._prepare_kernel_code(kernel_code)
            
            with open(cuda_file, 'w') as f:
                f.write(full_code)
            
            compile_flags = [
                "-ptx",
                "-O3",
                "-std=c++14",
                "-use_fast_math",
                "-lineinfo",
                f"-o={ptx_file}"
            ] + self.arch_flags

            constraints = kernel_data.get("constraints", {})
            if "max_registers" in constraints:
                compile_flags.append(f"-maxrregcount={constraints['max_registers']}")

            # CRITICAL FIX: Add -ccbin flag on Windows to explicitly specify x64 compiler
            # This prevents cudafe++ ACCESS_VIOLATION errors
            ccbin_path = None
            if platform.system() == "Windows":
                try:
                    from .utils.detection import ToolchainDetector
                    detector = ToolchainDetector()
                    compiler_info = detector.detect_cpp_compiler()

                    if compiler_info.available and compiler_info.path:
                        cl_dir = str(Path(compiler_info.path).parent)

                        # Ensure x64 (CUDA 12+ requirement)
                        if 'x86' in cl_dir.lower() and 'x64' not in cl_dir.lower():
                            x64_dir = cl_dir.replace('x86', 'x64').replace('X86', 'X64').replace('Hostx86', 'Hostx64')
                            if Path(x64_dir, 'cl.exe').exists():
                                cl_dir = x64_dir

                        # Add -ccbin flag to explicitly specify compiler
                        compile_flags.insert(0, f"-ccbin={cl_dir}")
                        ccbin_path = cl_dir
                except Exception:
                    pass

            try:
                console.print(f"[cyan]Compiling kernel for {kernel_data.get('operation', 'unknown')}...[/cyan]")