From 6fb2849ef8018e50241ac8b607b5f463ba418354 Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Mon, 24 Aug 2015 21:23:59 -0500
Subject: [PATCH] Stats functions: return bare dictionary, export helper
functions
---
doc/reference.rst | 2 ++
doc/tutorial.rst | 2 +-
loopy/__init__.py | 5 +++--
loopy/statistics.py | 46 ++++++++++++++++++++++++++++++++++-----------
4 files changed, 41 insertions(+), 14 deletions(-)
diff --git a/doc/reference.rst b/doc/reference.rst
index e79f17554..9aa9727cd 100644
--- a/doc/reference.rst
+++ b/doc/reference.rst
@@ -537,6 +537,8 @@ Obtaining Kernel Statistics
.. autofunction:: get_gmem_access_poly
+.. autofunction:: sum_mem_access_to_bytes
+
.. autofunction:: get_barrier_poly
.. vim: tw=75:spell
diff --git a/doc/tutorial.rst b/doc/tutorial.rst
index ca22c2a64..9c8026813 100644
--- a/doc/tutorial.rst
+++ b/doc/tutorial.rst
@@ -1264,7 +1264,7 @@ continue using the kernel from the previous example:
>>> from loopy.statistics import get_gmem_access_poly
>>> load_store_map = get_gmem_access_poly(knl)
- >>> print(load_store_map)
+ >>> print(lp.stringify_stats_mapping(load_store_map))
(dtype('float32'), 'uniform', 'load') : [n, m, l] -> { 3 * n * m * l : n >= 1 and m >= 1 and l >= 1 }
(dtype('float32'), 'uniform', 'store') : [n, m, l] -> { n * m * l : n >= 1 and m >= 1 and l >= 1 }
(dtype('float64'), 'uniform', 'load') : [n, m, l] -> { 2 * n * m : n >= 1 and m >= 1 and l >= 1 }
diff --git a/loopy/__init__.py b/loopy/__init__.py
index 8956856d4..77da1e40e 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -64,7 +64,8 @@ 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_gmem_access_poly,
- get_DRAM_access_poly, get_barrier_poly)
+ get_DRAM_access_poly, get_barrier_poly, stringify_stats_mapping,
+ sum_mem_access_to_bytes)
from loopy.codegen import generate_code, generate_body
from loopy.compiled import CompiledKernel
from loopy.options import Options
@@ -105,7 +106,7 @@ __all__ = [
"generate_code", "generate_body",
"get_op_poly", "get_gmem_access_poly", "get_DRAM_access_poly",
- "get_barrier_poly",
+ "get_barrier_poly", "stringify_stats_mapping", "sum_mem_access_to_bytes",
"CompiledKernel",
diff --git a/loopy/statistics.py b/loopy/statistics.py
index d25ea3eac..240bef8eb 100755
--- a/loopy/statistics.py
+++ b/loopy/statistics.py
@@ -76,16 +76,17 @@ class ToCountMap:
except KeyError:
return isl.PwQPolynomial('{ 0 }')
- def __str__(self):
- result = ""
- for key in sorted(self.dict.keys(), key=lambda k: str(k)):
- result += ("%s : %s\n" % (key, self.dict[key]))
- return result
-
def __repr__(self):
return repr(self.dict)
+def stringify_stats_mapping(m):
+ result = ""
+ for key in sorted(m.keys(), key=lambda k: str(k)):
+ result += ("%s : %s\n" % (key, m[key]))
+ return result
+
+
class ExpressionOpCounter(CombineMapper):
def __init__(self, knl):
@@ -447,7 +448,7 @@ def get_op_poly(knl):
domain = (inames_domain.project_out_except(insn_inames, [dim_type.set]))
ops = op_counter(insn.assignee) + op_counter(insn.expression)
op_poly = op_poly + ops*count(knl, domain)
- return op_poly
+ return op_poly.dict
def get_gmem_access_poly(knl): # for now just counting subscripts
@@ -462,10 +463,12 @@ def get_gmem_access_poly(knl): # for now just counting subscripts
- 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
+ - The first string in the map key specifies the global memory
+ access type as
*consecutive*, *nonconsecutive*, or *uniform*.
- - The second string in the map key specifies the DRAM access type as a
+ - The second string in the map key specifies the global memory
+ access type as a
*load*, or a *store*.
- The :class:`islpy.PwQPolynomial` holds the number of DRAM accesses
@@ -514,7 +517,7 @@ def get_gmem_access_poly(knl): # for now just counting subscripts
for key, val in six.iteritems(subs_assignee.dict)))
subs_poly = subs_poly + (subs_expr + subs_assignee)*count(knl, domain)
- return subs_poly
+ return subs_poly.dict
def get_DRAM_access_poly(knl):
@@ -524,6 +527,27 @@ def get_DRAM_access_poly(knl):
return get_gmem_access_poly(knl)
+def sum_mem_access_to_bytes(m):
+ """Sum the mapping returned by :func:`get_gmem_access_poly` to a mapping
+
+ **{(** :class:`string` **,** :class:`string` **)**
+ **:** :class:`islpy.PwQPolynomial` **}**
+
+ i.e., aggregate the transfer numbers for all types into a single byte count.
+ """
+
+ result = {}
+ for (dtype, kind, direction), v in m.items():
+ new_key = (kind, direction)
+ bytes_transferred = int(dtype.itemsize) * v
+ if new_key in result:
+ result[new_key] += bytes_transferred
+ else:
+ result[new_key] = bytes_transferred
+
+ return result
+
+
def get_barrier_poly(knl):
"""Count the number of barriers each thread encounters in a loopy kernel.
@@ -571,4 +595,4 @@ def get_barrier_poly(knl):
else:
barrier_poly += isl.PwQPolynomial('{ 1 }')
- return barrier_poly
+ return barrier_poly.dict
--
GitLab