|
| virtual std::string | schedule_value () const =0 |
| | Schedule value string identifying this backend's GPU schedule ("CUDA"/"ROCM").
|
| virtual codegen::TargetType | target_type () const =0 |
| | Target type used for instrumentation info.
|
| virtual std::unique_ptr< codegen::LanguageExtension > | create_device_language_extension () const =0 |
| | Create the backend device-code language extension (CUDA/HIP).
|
| virtual bool | is_device_pointer_storage (const types::StorageType &storage) const =0 |
| | True if storage designates a device-resident generic pointer for this backend.
|
| virtual std::string | kernel_file_extension () const =0 |
| | File extension for the kernel translation unit ("cu"/"rocm.cpp").
|
| virtual void | emit_kernel_includes (codegen::CodeSnippetFactory &library_snippet_factory) const =0 |
| | Emit the backend include directives required by the generated kernel.
|
| virtual void | emit_library_preamble (codegen::PrettyPrinter &library_stream) const =0 |
| | Emit any extra preamble at the top of the kernel translation unit.
|
| virtual void | emit_kernel_call (codegen::PrettyPrinter &main_stream, const std::string &kernel_name, symbolic::Expression &num_blocks, symbolic::Expression &block_size, std::vector< std::string > &arguments_device)=0 |
| | Emit the host-side kernel launch.
|
| bool | is_nested_in_gpu_kernel () |
| | True if this Reduce is nested inside an enclosing GPU kernel (an ancestor Map with this backend's GPU schedule).
|
| void | dispatch_header (codegen::PrettyPrinter &stream, const std::string &kernel_name, std::vector< std::string > &arguments_declaration) |
| void | dispatch_reduce_core (codegen::CodeSnippetFactory &library_snippet_factory, codegen::PrettyPrinter &library_stream, GPUDimension dimension, bool declare_scope_variables, std::vector< std::string > &scope_variables) |
| | Emit the privatize + grid-stride reduce loop + atomic merge into the given (device) stream, parallelized on dimension.
|
| void | dispatch_atomic_merge (codegen::PrettyPrinter &library_stream, codegen::CodeSnippetFactory &library_snippet_factory, const structured_control_flow::ReductionInfo &reduction) |
| | Emit the device-side helper(s) and the atomic merge statement that combines a thread's private partial into the global accumulator.
|
| virtual bool | begin_node (PrettyPrinter &stream) |
| virtual void | end_node (PrettyPrinter &stream, bool has_declaration) |
Atomics-based GPU code generator for structured_control_flow::Reduce.
This is the target-agnostic core shared by the CUDA and ROCm reduce dispatchers. A Reduce loop carries one or more associative/commutative reductions (structured_control_flow::ReductionInfo) into a loop-invariant accumulator; this dispatcher handles a Reduce whose own schedule is a GPU schedule, i.e. the reduction dimension is parallelized across threads on its configured GPU dimension.
Correctness strategy ("atomics baseline"):
- The reduction dimension is parallelized with a grid-stride loop on the schedule's dimension: thread t starts at init + t*stride and advances by num_threads*stride. This is correct for any launch geometry (even a single thread), so the lowering composes whether the Reduce is top-level or nested inside a GPU map.
- For every reduction the accumulator is privatized: a thread-local partial is declared and initialized to the operator's identity, and the accumulator container is shadowed to point at that private storage so the in-body combine (acc = acc OP x) accumulates into the register.
- After the loop, each thread merges its private partial into the real device-resident accumulator with an atomic combine (native atomicAdd where available, otherwise a compare-and-swap loop). Associativity and commutativity (guaranteed by the Reduce node) make the merge order-free.
Subclasses supply only the backend-specific policy (language extension, device-pointer storage, kernel launch syntax and includes).
1.15.0