From 6622a32d5b402973703f1d74f7e56158ec3b7e66 Mon Sep 17 00:00:00 2001
From: James Stevens <jdsteve2@illinois.edu>
Date: Mon, 26 Oct 2015 15:30:25 -0500
Subject: [PATCH] moved reg counter to perf model
---
doc/reference.rst | 2 -
loopy/__init__.py | 7 +-
loopy/statistics.py | 198 ----------------------------------------
test/test_statistics.py | 90 ------------------
4 files changed, 3 insertions(+), 294 deletions(-)
diff --git a/doc/reference.rst b/doc/reference.rst
index 9dad1182c..59ab3c986 100644
--- a/doc/reference.rst
+++ b/doc/reference.rst
@@ -548,6 +548,4 @@ Obtaining Kernel Statistics
.. autofunction:: get_barrier_poly
-.. autofunction:: estimate_regs_per_thread
-
.. vim: tw=75:spell
diff --git a/loopy/__init__.py b/loopy/__init__.py
index 2df476e1f..7073108b2 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -64,8 +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, estimate_regs_per_thread,
- stringify_stats_mapping, sum_mem_access_to_bytes)
+ 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
@@ -106,8 +106,7 @@ __all__ = [
"generate_code", "generate_body",
"get_op_poly", "get_gmem_access_poly", "get_DRAM_access_poly",
- "get_barrier_poly", "estimate_regs_per_thread", "stringify_stats_mapping",
- "sum_mem_access_to_bytes",
+ "get_barrier_poly", "stringify_stats_mapping", "sum_mem_access_to_bytes",
"CompiledKernel",
diff --git a/loopy/statistics.py b/loopy/statistics.py
index ae441255b..834f48207 100755
--- a/loopy/statistics.py
+++ b/loopy/statistics.py
@@ -380,156 +380,6 @@ class GlobalSubscriptCounter(CombineMapper):
"map_slice not implemented.")
-class RegisterUsageEstimator(CombineMapper):
-
- def __init__(self, knl):
- self.knl = knl
- from loopy.expression import TypeInferenceMapper
- self.type_inf = TypeInferenceMapper(knl)
- self.vars_found = []
- self.subs_found = []
-
- def combine(self, values):
- return sum(values)
-
- def forget_prev_vars(self):
- del self.vars_found[:]
-
- def forget_prev_subs(self):
- del self.subs_found[:]
-
- def map_constant(self, expr):
- return 0
-
- def map_variable(self, expr):
- name = expr.name
- if expr in self.vars_found:
- return 0
-
- self.vars_found.append(expr)
- if name in self.knl.temporary_variables:
- if self.knl.temporary_variables[name].is_local:
- return 0
- else:
- return 1
- elif name in self.knl.all_inames():
- from loopy.kernel.data import AxisTag, VectorizeTag, UnrollTag
- tag = self.knl.iname_to_tag.get(name)
- if (tag is None or not(isinstance(tag, AxisTag)
- or isinstance(tag, VectorizeTag)
- or isinstance(tag, UnrollTag))):
- return 1
- else:
- return 0
- else:
- return 1
-
- map_tagged_variable = map_variable
-
- def map_call(self, expr):
- return self.rec(expr.parameters)
-
- def map_subscript(self, expr):
- name = expr.aggregate.name # name of array
-
- if name in self.knl.arg_dict:
- # not a temporary variable
- array = self.knl.arg_dict[name]
- elif self.knl.temporary_variables[name].is_local:
- # temp var is in shared mem
- return 0 + self.rec(expr.index)
- elif (expr.index, expr.aggregate) in self.subs_found:
- # temp var is NOT shared, but already counted
- return 0 + self.rec(expr.index)
- else:
- # temp var is NOT shared and NOT already counted
- self.subs_found.append((expr.index, expr.aggregate))
- return 1 + self.rec(expr.index)
-
- # expr is not a temporary variable
-
- if not isinstance(array, lp.GlobalArg):
- # This array is not in global memory, and is not a temporary variable
- # TODO how should we count arrays in const/texture mem? ImageArg?
- # Ignore for now
- return self.rec(expr.index)
-
- # this is a global mem access
- if (expr.index, expr.aggregate) in self.subs_found:
- return 0 + self.rec(expr.index)
- else:
- self.subs_found.append((expr.index, expr.aggregate))
- return 1 + self.rec(expr.index)
-
- def map_sum(self, expr):
- assert expr.children
- return sum(self.rec(child) for child in expr.children)
-
- map_product = map_sum
-
- def map_quotient(self, expr, *args):
- return self.rec(expr.numerator) + self.rec(expr.denominator)
-
- map_floor_div = map_quotient
- map_remainder = map_quotient
-
- def map_power(self, expr):
- return self.rec(expr.base) + self.rec(expr.exponent)
-
- def map_left_shift(self, expr):
- return self.rec(expr.shiftee)+self.rec(expr.shift)
-
- map_right_shift = map_left_shift
-
- def map_bitwise_not(self, expr):
- return self.rec(expr.child)
-
- def map_bitwise_or(self, expr):
- return sum(self.rec(child) for child in expr.children)
-
- map_bitwise_xor = map_bitwise_or
- map_bitwise_and = map_bitwise_or
-
- def map_comparison(self, expr):
- return self.rec(expr.left)+self.rec(expr.right)
-
- map_logical_not = map_bitwise_not
- map_logical_or = map_bitwise_or
- map_logical_and = map_logical_or
-
- def map_if(self, expr):
- warnings.warn("RegisterUsageEstimator counting register usage as "
- "sum of if-statement branches.")
- return self.rec(expr.condition) + self.rec(expr.then) + self.rec(expr.else_)
-
- def map_if_positive(self, expr):
- warnings.warn("RegisterUsageEstimator counting register usage as "
- "sum of if_pos-statement branches.")
- return self.rec(expr.criterion) + self.rec(expr.then) + self.rec(expr.else_)
-
- map_min = map_bitwise_or
- map_max = map_min
-
- def map_common_subexpression(self, expr):
- raise NotImplementedError("GlobalSubscriptCounter encountered "
- "common_subexpression, "
- "map_common_subexpression not implemented.")
-
- def map_substitution(self, expr):
- raise NotImplementedError("GlobalSubscriptCounter encountered "
- "substitution, "
- "map_substitution not implemented.")
-
- def map_derivative(self, expr):
- raise NotImplementedError("GlobalSubscriptCounter encountered "
- "derivative, "
- "map_derivative not implemented.")
-
- def map_slice(self, expr):
- raise NotImplementedError("GlobalSubscriptCounter encountered slice, "
- "map_slice not implemented.")
-
-
def count(kernel, bset):
try:
return bset.card()
@@ -756,51 +606,3 @@ def get_barrier_poly(knl):
return barrier_poly
-
-def estimate_regs_per_thread(knl):
-
- """Estimate registers per thread usage by a loopy kernel.
-
- :parameter knl: A :class:`loopy.LoopKernel` whose reg usage will be estimated.
-
- :return: An :class:`integer` holding an estimate for the number of registers
- used per thread. This number will most likely be too low, but will
- hopefully be consistantly too low by the same constant factor.
-
- """
-
- from loopy.preprocess import preprocess_kernel, infer_unknown_types
- from loopy.schedule import EnterLoop, LeaveLoop, Barrier, RunInstruction # noqa
- knl = infer_unknown_types(knl, expect_completion=True)
- knl = preprocess_kernel(knl)
- knl = lp.get_one_scheduled_kernel(knl)
- max_regs = 0
- block_reg_totals = [0]
- # counters to track nested sets of previously used iname+index combinations
- reg_counters = [RegisterUsageEstimator(knl)]
-
- for sched_item in knl.schedule:
- if isinstance(sched_item, EnterLoop):
- block_reg_totals.append(0)
- # start a new estimator
- reg_counters.append(RegisterUsageEstimator(knl))
-
- elif isinstance(sched_item, LeaveLoop):
- if block_reg_totals[-1] > max_regs:
- max_regs = block_reg_totals[-1]
- # pop to resume previous total
- block_reg_totals.pop()
- reg_counters.pop()
-
- elif isinstance(sched_item, RunInstruction):
- insn = knl.id_to_insn[sched_item.insn_id]
- block_reg_totals[-1] += reg_counters[-1](insn.assignee) + \
- reg_counters[-1](insn.expression)
-
- # finished looping, check outer block
- if block_reg_totals[-1] > max_regs:
- max_regs = block_reg_totals[-1]
-
- return max_regs
-
-
diff --git a/test/test_statistics.py b/test/test_statistics.py
index 5cd6a7781..0dffe5c35 100644
--- a/test/test_statistics.py
+++ b/test/test_statistics.py
@@ -28,7 +28,6 @@ from pyopencl.tools import ( # noqa
as pytest_generate_tests)
import loopy as lp
from loopy.statistics import get_op_poly, get_gmem_access_poly, get_barrier_poly
-from loopy.statistics import estimate_regs_per_thread
import numpy as np
@@ -541,92 +540,6 @@ def test_barrier_counter_barriers():
assert barrier_count == 50*10*2
-def test_reg_counter_basic():
-
- knl = lp.make_kernel(
- "[n,m,l] -> {[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
- [
- """
- c[i, j, k] = a[i,j,k]*b[i,j,k]/3.0+a[i,j,k]
- e[i, k+1] = g[i,k]*h[i,k+1]
- """
- ],
- name="basic", assumptions="n,m,l >= 1")
-
- knl = lp.add_and_infer_dtypes(knl,
- dict(a=np.float32, b=np.float32, g=np.float64, h=np.float64))
- regs = estimate_regs_per_thread(knl)
- assert regs == 6
-
-
-def test_reg_counter_reduction():
-
- knl = lp.make_kernel(
- "{[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
- [
- "c[i, j] = sum(k, a[i, k]*b[k, j])"
- ],
- name="matmul_serial", assumptions="n,m,l >= 1")
-
- knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32, b=np.float32))
- regs = estimate_regs_per_thread(knl)
- assert regs == 6
-
-
-def test_reg_counter_logic():
-
- knl = lp.make_kernel(
- "{[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
- [
- """
- e[i,k] = if(not(k<l-2) and k>6 or k/2==l, g[i,k]*2, g[i,k]+h[i,k]/2)
- """
- ],
- name="logic", assumptions="n,m,l >= 1")
-
- knl = lp.add_and_infer_dtypes(knl, dict(g=np.float32, h=np.float64))
- regs = estimate_regs_per_thread(knl)
- assert regs == 6
-
-
-def test_reg_counter_specialops():
-
- knl = lp.make_kernel(
- "{[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
- [
- """
- c[i, j, k] = (2*a[i,j,k])%(2+b[i,j,k]/3.0)
- e[i, k] = (1+g[i,k])**(1+h[i,k+1])
- """
- ],
- name="specialops", assumptions="n,m,l >= 1")
-
- knl = lp.add_and_infer_dtypes(knl,
- dict(a=np.float32, b=np.float32, g=np.float64, h=np.float64))
- regs = estimate_regs_per_thread(knl)
- assert regs == 6
-
-
-def test_reg_counter_bitwise():
-
- knl = lp.make_kernel(
- "{[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
- [
- """
- c[i, j, k] = (a[i,j,k] | 1) + (b[i,j,k] & 1)
- e[i, k] = (g[i,k] ^ k)*(~h[i,k+1]) + (g[i, k] << (h[i,k] >> k))
- """
- ],
- name="bitwise", assumptions="n,m,l >= 1")
-
- knl = lp.add_and_infer_dtypes(
- knl, dict(
- a=np.int32, b=np.int32,
- g=np.int64, h=np.int64))
- regs = estimate_regs_per_thread(knl)
- assert regs == 6
-
-
def test_all_counters_parallel_matmul():
knl = lp.make_kernel(
@@ -681,9 +594,6 @@ def test_all_counters_parallel_matmul():
assert f32coal == n*l
- regs = estimate_regs_per_thread(knl)
- assert regs == 4
-
if __name__ == "__main__":
if len(sys.argv) > 1:
--
GitLab