Sarek_ir_interp_intrinsics (sarek.Sarek_interp.Sarek_ir_interp

Intrinsic Evaluation Helpers

These helper functions handle specific categories of GPU intrinsics. Split from a monolithic eval_intrinsic for better maintainability and testability. Each helper returns Option to enable clean dispatch logic.

Evaluate GPU thread/block/grid index and dimension intrinsics.

Handles the complete GPU execution model intrinsics:

Thread indices: Position of thread within its block (0 to block_dim-1)
Block indices: Position of block within grid (0 to grid_dim-1)
Dimensions: Size of blocks and grid
Global indices: Thread's global position across entire grid
Global sizes: Total number of threads in each dimension

parameter state
Thread execution state containing all index/dimension values

parameter name
Intrinsic name (e.g. "thread_idx_x", "global_size_y")

returns
Some value if intrinsic matches, None otherwise

Example: eval_gpu_index_intrinsic state "global_idx_x" (* Returns: VInt32 (block_idx_x * block_dim_x + thread_idx_x) *)

val eval_gpu_index_intrinsic : 
  Sarek_ir_interp_value.thread_state ->
  string ->
  Sarek_ir_interp_value.value option

GPU thread/block/grid indices and dimensions

val eval_barrier_intrinsic : string -> Sarek_ir_interp_value.value option

Barrier synchronization intrinsics

Warp operations

The interpreter simulates the warp: Sarek_warp.simulated_width () consecutive threads of the block (linear thread id, x fastest). A shuffle is an exchange through a per-block array in the block's shared table between two barriers, exactly as Native does (Sarek_cpu_runtime_warp), and the source lane is Sarek_warp.source_lane.

val warp_linear_thread_id : Sarek_ir_interp_value.thread_state -> int

val warp_block_size : Sarek_ir_interp_value.thread_state -> int

val warp_exchange_slots : 
  (string, Sarek_ir_interp_value.value array) Stdlib.Hashtbl.t ->
  Sarek_warp.value_type ->
  int ->
  Sarek_ir_interp_value.value array

val warp_check_block : 
  Sarek_ir_interp_value.thread_state ->
  width:int ->
  name:string ->
  unit

val eval_warp_shuffle : 
  (string, Sarek_ir_interp_value.value array) Stdlib.Hashtbl.t ->
  Sarek_ir_interp_value.thread_state ->
  name:string ->
  Sarek_warp.op ->
  Sarek_warp.value_type ->
  Sarek_ir_interp_value.value ->
  Sarek_ir_interp_value.value ->
  Sarek_ir_interp_value.value

val eval_warp_call : 
  (string, Sarek_ir_interp_value.value array) Stdlib.Hashtbl.t ->
  Sarek_ir_interp_value.thread_state ->
  name:string ->
  Sarek_warp.call ->
  Sarek_ir_interp_value.value list ->
  Sarek_ir_interp_value.value

Evaluate a warp operation (warp_shuffle*, warp_size, lane_id). shared is the block's shared table.

val atomic_global_mutex : Stdlib.Mutex.t

val with_atomic_lock : (unit -> 'a) -> 'a

val atomic_index : string -> 'a array -> Sarek_ir_interp_value.value -> int

Global-memory atomics and memory fences.

Atomics on global vectors return the OLD value and update in place. The interpreter models global memory as a shared mutable VArray; fences are no-ops because a single logical memory is already sequentially consistent here. Sufficient for the Sarek_worklist queue pattern (atomic HEAD/TAIL counters + ring slots).

val atomic_arity : string -> int -> 'a -> 'b

val eval_atomic_intrinsic : 
  string ->
  Sarek_ir_interp_value.value list ->
  Sarek_ir_interp_value.value option

val eval_float32_math_intrinsic : 
  string ->
  Sarek_ir_interp_value.value list ->
  Sarek_ir_interp_value.value option

Float32 math intrinsics

val eval_float64_math_intrinsic : 
  string ->
  Sarek_ir_interp_value.value list ->
  Sarek_ir_interp_value.value option

Float64 math intrinsics

val eval_int32_math_intrinsic : 
  string ->
  Sarek_ir_interp_value.value list ->
  Sarek_ir_interp_value.value option

Int32 math intrinsics

val eval_type_conversion_intrinsic : 
  string ->
  Sarek_ir_interp_value.value list ->
  Sarek_ir_interp_value.value option

Type conversion intrinsics