Yaakov/phase 2 b a

CMakeLists.txt

@@ -3,8 +3,15 @@ cmake_minimum_required(VERSION 3.13.4)
 include(CheckCXXSourceCompiles)
 
 set(POLYGEIST_ENABLE_CUDA 0 CACHE BOOL "Enable CUDA frontend and backend")
+set(POLYGEIST_ENABLE_CUDA_SYNTAX_ONLY 0 CACHE BOOL "Enable CUDA syntax parsing without requiring CUDA toolkit")
 set(POLYGEIST_ENABLE_ROCM 0 CACHE BOOL "Enable ROCM backend")
 
+# If CUDA_SYNTAX_ONLY is enabled, set flag for frontend-only mode
+if(POLYGEIST_ENABLE_CUDA_SYNTAX_ONLY)
+  set(POLYGEIST_CUDA_FRONTEND_ONLY 1)
+  message(STATUS "CUDA syntax-only mode enabled (no CUDA toolkit required)")
+endif()
+
 set(POLYGEIST_ENABLE_POLYMER 0 CACHE BOOL "Enable Polymer")
 set(POLYGEIST_POLYMER_ENABLE_ISL 0 CACHE BOOL "Enable Polymer isl")
 set(POLYGEIST_POLYMER_ENABLE_PLUTO 0 CACHE BOOL "Enable Polymer pluto")

include/polygeist/Passes/Passes.h

@@ -79,6 +79,9 @@ createGpuSerializeToHsacoPass(StringRef arch, StringRef features,
 void registerGpuSerializeToCubinPass();
 void registerGpuSerializeToHsacoPass();
 
+std::unique_ptr<Pass> createConvertGPUToVortexPass();
+std::unique_ptr<Pass> createGenerateVortexMainPass();
+
 void populateForBreakToWhilePatterns(RewritePatternSet &patterns);
 } // namespace polygeist
 } // namespace mlir

include/polygeist/Passes/Passes.td

@@ -300,4 +300,52 @@ def ConvertPolygeistToLLVM : Pass<"convert-polygeist-to-llvm", "mlir::ModuleOp">
   ];
 }
 
+def ConvertGPUToVortex : Pass<"convert-gpu-to-vortex", "ModuleOp"> {
+  let summary = "Lower GPU dialect operations to Vortex RISC-V intrinsics";
+  let description = [{
+    This pass converts GPU dialect operations to LLVM dialect with Vortex-specific
+    intrinsics. It lowers operations like gpu.thread_id, gpu.block_id to RISC-V
+    CSR reads via inline assembly, preparing the code for the Vortex GPGPU backend.
+
+    Example:
+      %tid = gpu.thread_id x
+    becomes:
+      %tid = llvm.inline_asm "csrr $0, 0xCC0" : () -> i32
+
+    The pass automatically emits JSON metadata files describing kernel argument
+    layouts for runtime argument marshaling. Files are written to the current
+    working directory as <kernel_name>.meta.json.
+  }];
+  let constructor = "mlir::polygeist::createConvertGPUToVortexPass()";
+  let dependentDialects = [
+    "LLVM::LLVMDialect",
+    "gpu::GPUDialect",
+  ];
+}
+
+def GenerateVortexMain : Pass<"generate-vortex-main", "ModuleOp"> {
+  let summary = "Generate Vortex main() wrapper for kernel execution";
+  let description = [{
+    This pass generates the Vortex-specific main() entry point and kernel_body
+    wrapper function. It should run AFTER gpu-to-llvm lowering has converted
+    gpu.func to llvm.func.
+
+    The generated code:
+    1. main() function that:
+       - Reads kernel args from VX_CSR_MSCRATCH (0x340) via inline assembly
+       - Calls vx_spawn_threads() with grid dimensions and kernel callback
+    2. kernel_body() wrapper that:
+       - Takes void* args pointer
+       - Unpacks individual arguments from the struct
+       - Calls the original lowered kernel function
+
+    This matches the Vortex kernel execution model where kernels are launched
+    via vx_spawn_threads() with a callback function.
+  }];
+  let constructor = "mlir::polygeist::createGenerateVortexMainPass()";
+  let dependentDialects = [
+    "LLVM::LLVMDialect",
+  ];
+}
+
 #endif // POLYGEIST_PASSES

lib/polygeist/ExecutionEngine/CMakeLists.txt

@@ -1,5 +1,11 @@
 # TODO we do not support cross compilation currently
 
+# Skip execution engine entirely if syntax-only mode
+if(POLYGEIST_CUDA_FRONTEND_ONLY)
+  message(STATUS "Skipping CUDA execution engine (syntax-only mode)")
+  return()
+endif()
+
 if(POLYGEIST_ENABLE_CUDA)
   find_package(CUDA)
   enable_language(CUDA)

lib/polygeist/Passes/CMakeLists.txt

@@ -18,6 +18,8 @@ add_mlir_dialect_library(MLIRPolygeistTransforms
   InnerSerialization.cpp
   ForBreakToWhile.cpp
   ConvertParallelToGPU.cpp
+  ConvertGPUToVortex.cpp
+  GenerateVortexMain.cpp
   SerializeToCubin.cpp
   SerializeToHsaco.cpp
   ParallelLoopUnroll.cpp
@@ -80,7 +82,8 @@ target_compile_definitions(obj.MLIRPolygeistTransforms
   POLYGEIST_PGO_DATA_DIR_ENV_VAR="${POLYGEIST_PGO_DATA_DIR_ENV_VAR}"
 )
 
-if(POLYGEIST_ENABLE_CUDA)
+# Only require CUDA toolkit if full CUDA support (not syntax-only mode)
+if(POLYGEIST_ENABLE_CUDA AND NOT POLYGEIST_CUDA_FRONTEND_ONLY)
   find_package(CUDA)
   enable_language(CUDA)