From dea47d64d394ecd43d52f8cbd7162a0d41048a2e Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Fri, 22 Jul 2011 03:06:54 -0500
Subject: [PATCH] Be smarter about choosing conflict-free lmem layout.
---
examples/matrix-ops.py | 20 +++---
loopy/__init__.py | 140 ++++++++++++++++++++++++++---------------
2 files changed, 99 insertions(+), 61 deletions(-)
diff --git a/examples/matrix-ops.py b/examples/matrix-ops.py
index fcf0a078b..eb1ffb7bf 100644
--- a/examples/matrix-ops.py
+++ b/examples/matrix-ops.py
@@ -136,7 +136,7 @@ def image_matrix_mul(ctx_factory=cl.create_some_context):
knl = lp.split_dimension(knl, "i", 16, outer_tag="g.0", inner_tag="l.1")
knl = lp.split_dimension(knl, "j", 16, outer_tag="g.1", inner_tag="l.0")
knl = lp.split_dimension(knl, "k", 32)
- # slow
+ # slow, but conflict-free
#knl = lp.add_prefetch(knl, 'a', ["i_inner", "k_inner"])
#knl = lp.add_prefetch(knl, 'b', ["j_inner", "k_inner"])
# fast
@@ -164,7 +164,8 @@ def image_matrix_mul(ctx_factory=cl.create_some_context):
return evt
lp.drive_timing_run(kernel_gen, queue, launcher, 2*n**3,
- options=FAST_OPTIONS)
+ options=FAST_OPTIONS + ["-cl-nv-verbose"],
+ force_rebuild=True)
@@ -176,7 +177,7 @@ def fancy_matrix_mul(ctx_factory=cl.create_some_context):
queue = cl.CommandQueue(ctx,
properties=cl.command_queue_properties.PROFILING_ENABLE)
- order = "F"
+ order = "C"
n = 16*40
from pymbolic import var
@@ -188,15 +189,15 @@ def fancy_matrix_mul(ctx_factory=cl.create_some_context):
(c[i, j], a[i, k]*b[k, j])
],
[
- lp.ArrayArg("a", dtype, shape=(n_sym, n_sym), order="F"),
- lp.ArrayArg("b", dtype, shape=(n_sym, n_sym), order="F"),
- lp.ArrayArg("c", dtype, shape=(n_sym, n_sym), order="F"),
+ lp.ArrayArg("a", dtype, shape=(n_sym, n_sym), order=order),
+ lp.ArrayArg("b", dtype, shape=(n_sym, n_sym), order=order),
+ lp.ArrayArg("c", dtype, shape=(n_sym, n_sym), order=order),
lp.ScalarArg("n", np.int32, approximately=1000),
], name="fancy_matmul")
knl = lp.split_dimension(knl, "i", 16, outer_tag="g.0", inner_tag="l.1")
knl = lp.split_dimension(knl, "j", 16, outer_tag="g.1", inner_tag="l.0")
- knl = lp.split_dimension(knl, "k", 19)
+ knl = lp.split_dimension(knl, "k", 16)
knl = lp.add_prefetch(knl, 'a', ["i_inner", "k_inner"])
knl = lp.add_prefetch(knl, 'b', ["k_inner", "j_inner"])
assert knl.get_invalid_reason() is None
@@ -221,7 +222,8 @@ def fancy_matrix_mul(ctx_factory=cl.create_some_context):
return evt
lp.drive_timing_run(kernel_gen, queue, launcher, 2*n**3,
- options=FAST_OPTIONS)
+ options=FAST_OPTIONS + ["-cl-nv-verbose"],
+ force_rebuild=True)
@@ -309,4 +311,4 @@ if __name__ == "__main__":
if len(sys.argv) > 1:
exec(sys.argv[1])
else:
- fancy_matrix_mul()
+ image_matrix_mul()
diff --git a/loopy/__init__.py b/loopy/__init__.py
index 9d4276567..78ee835a2 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -5,9 +5,10 @@ from pytools import Record, memoize_method
from pymbolic.mapper.dependency import DependencyMapper
from pymbolic.mapper.c_code import CCodeMapper
from pymbolic.mapper.stringifier import PREC_NONE
-from pymbolic.mapper import IdentityMapper, CombineMapper, RecursiveMapper
+from pymbolic.mapper import CombineMapper, RecursiveMapper
import pyopencl as cl
+import pyopencl.characterize as cl_char
import islpy as isl
from islpy import dim_type
@@ -24,7 +25,6 @@ register_mpz_with_pymbolic()
# TODO: Try, fix reg. prefetch
# TODO: Divisibility
-# TODO: Reasoning about bank conflicts
# TODO: Functions
# TODO: Common subexpressions
# TODO: Try different kernels
@@ -34,6 +34,7 @@ register_mpz_with_pymbolic()
# TODO: User controllable switch for slab opt
# TODO: User control over schedule
# TODO: Condition hoisting
+# TODO: Separate all-bulk from non-bulk kernels.
# TODO: Custom reductions per red. axis
@@ -143,7 +144,7 @@ def get_bounds_constraints(bset, iname, space=None, admissible_vars=None):
iname_coeff = int(cns.get_coefficient(iname_tp, iname_idx))
if admissible_vars is not None:
- if not (set(cns.get_coefficients_by_name().iterkeys())
+ if not (set(cns.get_coefficients_by_name().iterkeys())
<= admissible_vars):
continue
@@ -1071,7 +1072,12 @@ def insert_register_prefetches(kernel):
# {{{ code generation
+# {{{ support code for AST wrapper objects
+
class GeneratedCode(Record):
+ """Objects of this type are wrapped around ASTs upon
+ return from generation calls to collect information about them.
+ """
__slots__ = ["ast", "num_conditionals"]
def gen_code_block(elements):
@@ -1117,6 +1123,8 @@ def wrap_with(cls, *args):
return GeneratedCode(ast=ast, num_conditionals=num_conditionals)
+# }}}
+
# {{{ C code mapper
class LoopyCCodeMapper(CCodeMapper):
@@ -1157,8 +1165,8 @@ class LoopyCCodeMapper(CCodeMapper):
assert isinstance(expr.index, tuple)
base_access = ("read_imagef(%s, loopy_sampler, (float%d)(%s))"
- % (arg.name, arg.dimensions,
- ", ".join(self.rec(idx, PREC_NONE)
+ % (arg.name, arg.dimensions,
+ ", ".join(self.rec(idx, PREC_NONE)
for idx in expr.index[::-1])))
if arg.dtype == np.float32:
@@ -1191,11 +1199,11 @@ class LoopyCCodeMapper(CCodeMapper):
def map_floor_div(self, expr, prec):
if isinstance(expr.denominator, int) and expr.denominator > 0:
- return ("int_floor_div_pos_b(%s, %s)"
+ return ("int_floor_div_pos_b(%s, %s)"
% (self.rec(expr.numerator, PREC_NONE),
expr.denominator))
else:
- return ("int_floor_div(%s, %s)"
+ return ("int_floor_div(%s, %s)"
% (self.rec(expr.numerator, PREC_NONE),
self.rec(expr.denominator, PREC_NONE)))
@@ -1557,7 +1565,7 @@ def generate_loop_dim_code(cgs, kernel, sched_index,
class TrialRecord(Record):
pass
- if (cgs.try_slab_partition
+ if (cgs.try_slab_partition
and "outer" in iname):
trial_cgs = cgs.copy(try_slab_partition=False)
trials = []
@@ -1681,7 +1689,7 @@ def get_valid_index_vars(kernel, sched_index, exclude_tags=()):
def filter_necessary_constraints(implemented_domain, constraints):
space = implemented_domain.get_dim()
- return [cns
+ return [cns
for cns in constraints
if not implemented_domain.is_subset(
isl.Set.universe(space)
@@ -1772,7 +1780,7 @@ def wrap_in_for_from_constraints(ccm, iname, constraint_bset, stmt):
# }}}
-# {{{ codegen top-level dispatch
+# {{{ loop nest build top-level dispatch
def build_loop_nest(cgs, kernel, sched_index, implemented_domain):
ccm = cgs.c_code_mapper
@@ -1839,49 +1847,80 @@ def build_loop_nest(cgs, kernel, sched_index, implemented_domain):
# }}}
-# {{{ main code generation entrypoint
-
-class CodeGenerationState(Record):
- __slots__ = ["c_code_mapper", "try_slab_partition"]
-
-def generate_code(kernel):
- from cgen import (FunctionBody, FunctionDeclaration,
- POD, Value, ArrayOf, Module, Block,
- Define, Line, Const, LiteralLines, Initializer)
-
- from cgen.opencl import (CLKernel, CLGlobal, CLRequiredWorkGroupSize,
- CLLocal, CLImage)
+# {{{ prefetch preprocessing
- # {{{ assign names, dim storage lengths to prefetches
+def preprocess_prefetch(kernel):
+ """Assign names, dim storage lengths to prefetches.
+ """
all_pf_list = kernel.prefetch.values()
- all_pf_nbytes = [opf.nbytes for opf in all_pf_list]
+ new_prefetch_dict = {}
+ lmem_size = cl_char.usable_local_mem_size(kernel.device)
- new_prefetch = {}
for i_pf, pf in enumerate(kernel.prefetch.itervalues()):
- dim_storage_lengths = [stop-start for start, stop in pf.dim_bounds]
-
+ all_pf_nbytes = [opf.nbytes for opf in all_pf_list]
other_pf_sizes = sum(all_pf_nbytes[:i_pf]+all_pf_nbytes[i_pf+1:])
+ shape = [stop-start for start, stop in pf.dim_bounds]
+ dim_storage_lengths = shape[:]
+
# sizes of all dims except the last one, which we may change
# below to avoid bank conflicts
from pytools import product
other_dim_sizes = (pf.itemsize
* product(dim_storage_lengths[:-1]))
- from pyopencl.characterize import usable_local_mem_size
- if (dim_storage_lengths[-1] % 2 == 0
- and other_pf_sizes+other_dim_sizes*(dim_storage_lengths[-1]+1)
- < usable_local_mem_size(kernel.device)):
- dim_storage_lengths[-1] += 1
+ min_mult = cl_char.local_memory_bank_count(kernel.device)
+ good_incr = None
+ new_dsl = dim_storage_lengths
+
+ for increment in range(dim_storage_lengths[-1]//2):
+
+ test_dsl = dim_storage_lengths[:]
+ test_dsl[-1] = test_dsl[-1] + increment
+ new_mult, why_not = cl_char.why_not_local_access_conflict_free(
+ kernel.device, pf.itemsize,
+ shape, test_dsl)
+
+ # will choose smallest increment 'automatically'
+ if new_mult < min_mult:
+ new_lmem_use = other_pf_sizes+pf.itemsize*product(new_dsl)
+ if new_lmem_use < lmem_size:
+ new_dsl = test_dsl
+ min_mult = new_mult
+ good_incr = increment
+
+ if min_mult != 1:
+ from warnings import warn
+ warn("could not find a conflict-free mem layout "
+ "for prefetch of '%s' "
+ "(currently: %dx conflict, increment: %d)"
+ % (pf.input_vector, min_mult, good_incr),
+ LoopyAdvisory)
+
+ new_pf = pf.copy(dim_storage_lengths=dim_storage_lengths,
+ name="prefetch_%s_%d" % (pf.input_vector, i_pf))
+ new_prefetch_dict[pf.input_vector, pf.index_expr] = new_pf
+ all_pf_list[i_pf] = new_pf
+
+ return kernel.copy(prefetch=new_prefetch_dict)
- new_prefetch[pf.input_vector, pf.index_expr] = \
- pf.copy(dim_storage_lengths=dim_storage_lengths,
- name="prefetch_%s_%d" % (pf.input_vector, i_pf))
+# }}}
- kernel = kernel.copy(prefetch=new_prefetch)
+# {{{ main code generation entrypoint
- # }}}
+class CodeGenerationState(Record):
+ __slots__ = ["c_code_mapper", "try_slab_partition"]
+
+def generate_code(kernel):
+ kernel = preprocess_prefetch(kernel)
+
+ from cgen import (FunctionBody, FunctionDeclaration,
+ POD, Value, ArrayOf, Module, Block,
+ Define, Line, Const, LiteralLines, Initializer)
+
+ from cgen.opencl import (CLKernel, CLGlobal, CLRequiredWorkGroupSize,
+ CLLocal, CLImage)
my_ccm = LoopyCCodeMapper(kernel)
@@ -2075,14 +2114,18 @@ def add_prefetch(kernel, input_access_descr, tags_or_inames, loc_fetch_axes={}):
# }}}
-
-
+# {{{ compiled kernel object
class CompiledKernel:
- def __init__(self, context, kernel, size_args=None, options=[]):
+ def __init__(self, context, kernel, size_args=None, options=[],
+ force_rebuild=False):
self.kernel = kernel
self.code = generate_code(kernel)
+ if force_rebuild:
+ from time import time
+ self.code = "/* %s */\n%s" % (time(), self.code)
+
#from pytools import invoke_editor
#self.code = invoke_editor(self.code)
@@ -2126,20 +2169,11 @@ class CompiledKernel:
self.local_size_func = compile(
lsize_expr, self.size_args)
-
-
-
-# {{{ speed measurement
-
-
# }}}
-
-
-
-# driver ----------------------------------------------------------------------
+# {{{ timing driver
def drive_timing_run(kernel_generator, queue, launch, flop_count=None,
- options=[], print_code=True):
+ options=[], print_code=True, force_rebuild=False):
def time_run(compiled_knl, warmup_rounds=2, timing_rounds=5):
check = True
@@ -2163,7 +2197,8 @@ def drive_timing_run(kernel_generator, queue, launch, flop_count=None,
soln_count = 0
for kernel in kernel_generator:
- compiled = CompiledKernel(queue.context, kernel, options=options)
+ compiled = CompiledKernel(queue.context, kernel, options=options,
+ force_rebuild=force_rebuild)
print "-----------------------------------------------"
print "SOLUTION #%d" % soln_count
@@ -2184,6 +2219,7 @@ def drive_timing_run(kernel_generator, queue, launch, flop_count=None,
print "%d solutions" % soln_count
+# }}}
--
GitLab