diff --git a/doc/reference.rst b/doc/reference.rst
index c7435bbf7f84570b6672ba1bf42eeab9d495ea56..eb10788cb07d8fb87cee5ce6edc90e11bb9db5c3 100644
--- a/doc/reference.rst
+++ b/doc/reference.rst
@@ -514,4 +514,13 @@ Controlling caching
.. autoclass:: CacheMode
+Obtaining Kernel Statistics
+---------------------------
+
+.. autofunction:: get_op_poly
+
+.. autofunction:: get_DRAM_access_poly
+
+.. autofunction:: get_barrier_poly
+
.. vim: tw=75:spell
diff --git a/loopy/__init__.py b/loopy/__init__.py
index c63aa5f90d6537496ea9fe4ecb11c441070c91b4..f6f611f6538911678ecdd12578ca9ab17d0657f2 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -63,6 +63,7 @@ from loopy.padding import (split_arg_axis, find_padding_multiple,
from loopy.preprocess import (preprocess_kernel, realize_reduction,
infer_unknown_types)
from loopy.schedule import generate_loop_schedules, get_one_scheduled_kernel
+from loopy.statistics import get_op_poly, get_DRAM_access_poly, get_barrier_poly
from loopy.codegen import generate_code, generate_body
from loopy.compiled import CompiledKernel
from loopy.options import Options
@@ -102,6 +103,8 @@ __all__ = [
"generate_loop_schedules", "get_one_scheduled_kernel",
"generate_code", "generate_body",
+ "get_op_poly", "get_DRAM_access_poly", "get_barrier_poly",
+
"CompiledKernel",
"auto_test_vs_ref",
diff --git a/loopy/compiled.py b/loopy/compiled.py
index da659eaba5e3c7c8c99993946da1e7af5bb399bc..a29f357b08e4383c1e6ea1b9b79db8330e8498a0 100644
--- a/loopy/compiled.py
+++ b/loopy/compiled.py
@@ -781,10 +781,11 @@ class _CLKernelInfo(Record):
class CompiledKernel:
def __init__(self, context, kernel):
"""
- :arg kernel: may be a loopy.LoopKernel, a generator returning kernels
- (a warning will be issued if more than one is returned). If the
- kernel has not yet been loop-scheduled, that is done, too, with no
+ :arg kernel: may be a loopy.LoopKernel, a generator returning kernels \
+ (a warning will be issued if more than one is returned). If the \
+ kernel has not yet been loop-scheduled, that is done, too, with no \
specific arguments.
+
"""
self.context = context
diff --git a/loopy/statistics.py b/loopy/statistics.py
index 85ac4c77bdad2573b6a7a033aa0e17221e010d59..5c58d66c8786840c24afa4d9575762b17b96a220 100755
--- a/loopy/statistics.py
+++ b/loopy/statistics.py
@@ -403,8 +403,32 @@ def count(kernel, bset):
return result
-# to evaluate poly: poly.eval_with_dict(dictionary)
def get_op_poly(knl):
+
+ """Count the number of operations in a loopy kernel.
+
+ :parameter knl: A :class:`loopy.LoopKernel` whose operations are to be counted.
+
+ :return: A mapping of **{** :class:`numpy.dtype` \
+ **:** :class:`islpy.PwQPolynomial` **}**.
+
+ - The :class:`islpy.PwQPolynomial` holds the number of operations for \
+ the :class:`numpy.dtype` specified in the key (in terms of \
+ the :class:`loopy.LoopKernel` *inames*).
+
+ Example usage::
+
+ poly = get_op_poly(knl)
+ n = 512
+ m = 256
+ l = 128
+ float32_op_ct = poly.dict[np.dtype(np.float32)].eval_with_dict(
+ {'n': n, 'm': m, 'l': l})
+ float64_op_ct = poly.dict[np.dtype(np.float64)].eval_with_dict(
+ {'n': n, 'm': m, 'l': l})
+
+ """
+
from loopy.preprocess import preprocess_kernel, infer_unknown_types
knl = infer_unknown_types(knl, expect_completion=True)
knl = preprocess_kernel(knl)
@@ -423,6 +447,44 @@ def get_op_poly(knl):
def get_DRAM_access_poly(knl): # for now just counting subscripts
+
+ """Count the number of DRAM accesses in a loopy kernel.
+
+ :parameter knl: A :class:`loopy.LoopKernel` \
+ whose DRAM accesses are to be counted.
+
+ :return: A mapping of **{(** \
+ :class:`numpy.dtype` **,** :class:`string` **,** :class:`string` \
+ **)** **:** :class:`islpy.PwQPolynomial` **}**.
+
+ - The :class:`numpy.dtype` specifies \
+ the type of the data being accessed.
+
+ - The first string in the map key specifies the DRAM access type as \
+ *consecutive*, *nonconsecutive*, or *uniform*.
+
+ - The second string in the map key specifies the DRAM access type as \
+ a *load*, or a *store*.
+
+ - The :class:`islpy.PwQPolynomial` holds the number of DRAM accesses \
+ with the characteristics specified in the key (in terms of the \
+ :class:`loopy.LoopKernel` *inames*).
+
+ Example usage::
+
+ subscript_map = get_DRAM_access_poly(knl)
+ f32_uncoalesced_load = subscript_map.dict[
+ (np.dtype(np.float32), 'nonconsecutive', 'load')
+ ].eval_with_dict({'n': n, 'm': m, 'l': l})
+ f32_coalesced_load = subscript_map.dict[
+ (np.dtype(np.float32), 'consecutive', 'load')
+ ].eval_with_dict({'n': n, 'm': m, 'l': l})
+ f32_coalesced_store = subscript_map.dict[
+ (np.dtype(np.float32), 'consecutive', 'store')
+ ].eval_with_dict({'n': n, 'm': m, 'l': l})
+
+ """
+
from loopy.preprocess import preprocess_kernel, infer_unknown_types
knl = infer_unknown_types(knl, expect_completion=True)
knl = preprocess_kernel(knl)
@@ -448,6 +510,27 @@ def get_DRAM_access_poly(knl): # for now just counting subscripts
def get_barrier_poly(knl):
+
+ """Count the number of barriers in a loopy kernel.
+
+ :parameter knl: A :class:`loopy.LoopKernel` \
+ whose barriers are to be counted.
+
+ :return: An :class:`islpy.PwQPolynomial` holding the number of barrier calls \
+ made (in terms of the :class:`loopy.LoopKernel` *inames*).
+
+ Example usage::
+
+ barrier_poly = get_barrier_poly(knl)
+
+ n = 512
+ m = 256
+ l = 128
+
+ barrier_count = barrier_poly.eval_with_dict({'n': n, 'm': m, 'l': l})
+
+ """
+
from loopy.preprocess import preprocess_kernel, infer_unknown_types
from loopy.schedule import EnterLoop, LeaveLoop, Barrier
from operator import mul