mlir/benchmark/python/common.py

   1 """Common utilities that are useful for all the benchmarks."""
   2 import numpy as np
   3
   4 from mlir import ir
   5 from mlir.dialects import arith
   6 from mlir.dialects import func
   7 from mlir.dialects import memref
   8 from mlir.dialects import scf
   9 from mlir.passmanager import PassManager
  10
  11
  12 def setup_passes(mlir_module):
  13     """Setup pass pipeline parameters for benchmark functions."""
  14     opt = (
  15         "parallelization-strategy=none"
  16     )
  17     pipeline = f"builtin.module(sparsifier{{{opt}}})"
  18     PassManager.parse(pipeline).run(mlir_module.operation)
  19
  20
  21 def create_sparse_np_tensor(dimensions, number_of_elements):
  22     """Constructs a numpy tensor of dimensions `dimensions` that has only a
  23     specific number of nonzero elements, specified by the `number_of_elements`
  24     argument.
  25     """
  26     tensor = np.zeros(dimensions, np.float64)
  27     tensor_indices_list = [
  28         [np.random.randint(0, dimension) for dimension in dimensions]
  29         for _ in range(number_of_elements)
  30     ]
  31     for tensor_indices in tensor_indices_list:
  32         current_tensor = tensor
  33         for tensor_index in tensor_indices[:-1]:
  34             current_tensor = current_tensor[tensor_index]
  35         current_tensor[tensor_indices[-1]] = np.random.uniform(1, 100)
  36     return tensor
  37
  38
  39 def get_kernel_func_from_module(module: ir.Module) -> func.FuncOp:
  40     """Takes an mlir module object and extracts the function object out of it.
  41     This function only works for a module with one region, one block, and one
  42     operation.
  43     """
  44     assert (
  45         len(module.operation.regions) == 1
  46     ), "Expected kernel module to have only one region"
  47     assert (
  48         len(module.operation.regions[0].blocks) == 1
  49     ), "Expected kernel module to have only one block"
  50     assert (
  51         len(module.operation.regions[0].blocks[0].operations) == 1
  52     ), "Expected kernel module to have only one operation"
  53     return module.operation.regions[0].blocks[0].operations[0]
  54
  55
  56 def emit_timer_func() -> func.FuncOp:
  57     """Returns the declaration of nanoTime function. If nanoTime function is
  58     used, the `MLIR_RUNNER_UTILS` and `MLIR_C_RUNNER_UTILS` must be included.
  59     """
  60     i64_type = ir.IntegerType.get_signless(64)
  61     nanoTime = func.FuncOp("nanoTime", ([], [i64_type]), visibility="private")
  62     nanoTime.attributes["llvm.emit_c_interface"] = ir.UnitAttr.get()
  63     return nanoTime
  64
  65
  66 def emit_benchmark_wrapped_main_func(kernel_func, timer_func):
  67     """Takes a function and a timer function, both represented as FuncOp
  68     objects, and returns a new function. This new function wraps the call to
  69     the original function between calls to the timer_func and this wrapping
  70     in turn is executed inside a loop. The loop is executed
  71     len(kernel_func.type.results) times. This function can be used to
  72     create a "time measuring" variant of a function.
  73     """
  74     i64_type = ir.IntegerType.get_signless(64)
  75     memref_of_i64_type = ir.MemRefType.get([ir.ShapedType.get_dynamic_size()], i64_type)
  76     wrapped_func = func.FuncOp(
  77         # Same signature and an extra buffer of indices to save timings.
  78         "main",
  79         (kernel_func.arguments.types + [memref_of_i64_type], kernel_func.type.results),
  80         visibility="public",
  81     )
  82     wrapped_func.attributes["llvm.emit_c_interface"] = ir.UnitAttr.get()
  83
  84     num_results = len(kernel_func.type.results)
  85     with ir.InsertionPoint(wrapped_func.add_entry_block()):
  86         timer_buffer = wrapped_func.arguments[-1]
  87         zero = arith.ConstantOp.create_index(0)
  88         n_iterations = memref.DimOp(timer_buffer, zero)
  89         one = arith.ConstantOp.create_index(1)
  90         iter_args = list(wrapped_func.arguments[-num_results - 1 : -1])
  91         loop = scf.ForOp(zero, n_iterations, one, iter_args)
  92         with ir.InsertionPoint(loop.body):
  93             start = func.CallOp(timer_func, [])
  94             call = func.CallOp(
  95                 kernel_func,
  96                 wrapped_func.arguments[: -num_results - 1] + loop.inner_iter_args,
  97             )
  98             end = func.CallOp(timer_func, [])
  99             time_taken = arith.SubIOp(end, start)
 100             memref.StoreOp(time_taken, timer_buffer, [loop.induction_variable])
 101             scf.YieldOp(list(call.results))
 102         func.ReturnOp(loop)
 103
 104     return wrapped_func