From 8cb22cf06b8d2f89bd7a60fe32ed96b2afa68081 Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Wed, 12 Apr 2017 18:00:07 +0200
Subject: [PATCH] WIP
---
python/dune/perftool/loopy/target.py | 2 +-
python/dune/perftool/sumfact/basis.py | 9 +-
python/dune/perftool/sumfact/symbolic.py | 16 ++--
python/dune/perftool/sumfact/vectorization.py | 90 ++++++++++++-------
4 files changed, 74 insertions(+), 43 deletions(-)
diff --git a/python/dune/perftool/loopy/target.py b/python/dune/perftool/loopy/target.py
index 48e441cb..ac6b481e 100644
--- a/python/dune/perftool/loopy/target.py
+++ b/python/dune/perftool/loopy/target.py
@@ -116,7 +116,7 @@ class DuneCExpressionToCodeMapper(CExpressionToCodeMapper):
x = self.rec(expr.numerator, PREC_PRODUCT)
return "{} & {}".format(x, n - 1)
else:
- return CExpressionToCodeMapper.map_remainder(expr, enclosing_prec)
+ return CExpressionToCodeMapper.map_remainder(self, expr, enclosing_prec)
class DuneASTBuilder(CASTBuilder):
diff --git a/python/dune/perftool/sumfact/basis.py b/python/dune/perftool/sumfact/basis.py
index 2e84ceac..49e154ec 100644
--- a/python/dune/perftool/sumfact/basis.py
+++ b/python/dune/perftool/sumfact/basis.py
@@ -67,7 +67,7 @@ def pymbolic_coefficient_gradient(element, restriction, component, coeff_func, v
# if the positioning within a SIMD vectors coincides with the index!
direct_indexing_is_possible = True
- buffers = []
+ expressions = []
insn_dep = frozenset()
for i in range(world_dimension()):
# Construct the matrix sequence for this sum factorization
@@ -97,7 +97,7 @@ def pymbolic_coefficient_gradient(element, restriction, component, coeff_func, v
from dune.perftool.sumfact.realization import realize_sum_factorization_kernel
var, insn_dep = realize_sum_factorization_kernel(vsf.copy(insn_dep=vsf.insn_dep.union(insn_dep)))
- buffers.append(var)
+ expressions.append(prim.Subscript(var, vsf.quadrature_index(sf)))
# Check whether we want to return early with something that has the indexing
# already handled! This happens with vectorization when the index coincides
@@ -107,12 +107,11 @@ def pymbolic_coefficient_gradient(element, restriction, component, coeff_func, v
return maybe_wrap_subscript(var, vsf.quadrature_index(sf, visitor.indices)), None
# TODO this should be quite conditional!!!
- for i, buf in enumerate(buffers):
+ for i, expr in enumerate(expressions):
# Write solution from sumfactorization to gradient variable
assignee = prim.Subscript(prim.Variable(name), i)
- expression = prim.Subscript(buf, vsf.quadrature_index(sf))
instruction(assignee=assignee,
- expression=expression,
+ expression=expr,
forced_iname_deps=frozenset(get_backend("quad_inames")()),
forced_iname_deps_is_final=True,
)
diff --git a/python/dune/perftool/sumfact/symbolic.py b/python/dune/perftool/sumfact/symbolic.py
index 3bd7fe09..9a3a1f56 100644
--- a/python/dune/perftool/sumfact/symbolic.py
+++ b/python/dune/perftool/sumfact/symbolic.py
@@ -412,7 +412,10 @@ class VectorizedSumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable)
@property
def padded_indices(self):
- indices = set(range(self.horizontal_width)) - set(range(len(self.kernels)))
+ indices = set(range(self.vector_width))
+ for h in range(len(self.kernels)):
+ for v in range(self.vertical_width):
+ indices.discard(h * self.horizontal_width + v)
return tuple(self.kernels[0].quadrature_index(None) + (i,) for i in indices)
@property
@@ -439,10 +442,11 @@ class VectorizedSumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable)
return True
def horizontal_index(self, sf):
- try:
- return self.kernels.index(sf)
- except ValueError:
- return 0
+ key = tuple(mat.derivative for mat in sf.matrix_sequence)
+ for i, k in enumerate(self.kernels):
+ if tuple(mat.derivative for mat in k.matrix_sequence) == key:
+ return i
+ return 0
def _sliced_expr(self):
quad_inames = quadrature_inames()
@@ -486,7 +490,7 @@ class VectorizedSumfactKernel(SumfactKernelBase, ImmutableRecord, prim.Variable)
sliced = prim.Variable(quad_inames[i])
i = i + 1
- return self.horizontal_index(sf) * self.vertical_width + sliced % self.vertical_width
+ return self.horizontal_index(sf) + prim.Remainder(sliced, self.vertical_width)
def vec_index_tuple(self, sf):
return (self.vec_index(sf),)
diff --git a/python/dune/perftool/sumfact/vectorization.py b/python/dune/perftool/sumfact/vectorization.py
index 7cef7426..b856ccc7 100644
--- a/python/dune/perftool/sumfact/vectorization.py
+++ b/python/dune/perftool/sumfact/vectorization.py
@@ -47,37 +47,51 @@ def no_vectorization(sumfacts):
def vertical_vectorization_strategy(sumfact, depth):
# Assert that this is not already sliced
assert all(mat.slice_size is None for mat in sumfact.matrix_sequence)
+ result = {}
# Determine which of the matrices in the kernel should be sliced
- def determine_slice_direction():
- for i, mat in enumerate(sumfact.matrix_sequence):
+ def determine_slice_direction(sf):
+ for i, mat in enumerate(sf.matrix_sequence):
if mat.quadrature_size % depth == 0:
return i
elif mat.quadrature_size != 1:
raise PerftoolError("Vertical vectorization is not possible!")
- sliced = determine_slice_direction()
-
- kernels = []
- oldtab = sumfact.matrix_sequence[sliced]
- for i in range(depth):
- seq = list(sumfact.matrix_sequence)
- seq[sliced] = oldtab.copy(slice_size=depth,
- slice_index=i)
- kernels.append(sumfact.copy(matrix_sequence=tuple(seq)))
-
- buffer = get_counted_variable("vertical_buffer")
- input = get_counted_variable("vertical_input")
+ if isinstance(sumfact, SumfactKernel):
+ kernels = [sumfact]
+ else:
+ assert isinstance(sumfact, VectorizedSumfactKernel)
+ kernels = sumfact.kernels
+
+ newkernels = []
+ for sf in kernels:
+ sliced = determine_slice_direction(sf)
+ oldtab = sf.matrix_sequence[sliced]
+ for i in range(depth):
+ seq = list(sf.matrix_sequence)
+ seq[sliced] = oldtab.copy(slice_size=depth,
+ slice_index=i)
+ newkernels.append(sf.copy(matrix_sequence=tuple(seq)))
+
+ if isinstance(sumfact, SumfactKernel):
+ buffer = get_counted_variable("vertical_buffer")
+ input = get_counted_variable("vertical_input")
+ result[sumfact] = VectorizedSumfactKernel(kernels=tuple(newkernels),
+ buffer=buffer,
+ input=input,
+ vertical_width=depth,
+ )
+ else:
+ assert isinstance(sumfact, VectorizedSumfactKernel)
+ for sf in kernels:
+ result[sf] = sumfact.copy(kernels=tuple(newkernels),
+ vertical_width=depth,
+ )
- vsf = VectorizedSumfactKernel(kernels=tuple(kernels),
- buffer=buffer,
- input=input,
- vertical_width=depth,
- )
- return {sumfact: vsf}
+ return result
-def horizontal_vectorization_strategy(sumfacts, width):
+def horizontal_vectorization_strategy(sumfacts, width, allow_padding=1):
result = {}
todo = set(sumfacts)
@@ -107,20 +121,31 @@ def horizontal_vectorization_strategy(sumfacts, width):
input = get_counted_variable("joined_input")
for sumf in kernels:
- if len(kernels) in (width, width - 1):
+ if len(kernels) in range(width - allow_padding, width + 1):
result[sumf] = VectorizedSumfactKernel(kernels=kernels,
horizontal_width=width,
buffer=buffer,
input=input,
)
- else:
- result[sumf] = no_vec(sumf)
-
return result
-def diagonal_vectorization_stratget(sumfacts, width):
- return horizontal_vectorization_strategy(sumfacts, width)
+def diagonal_vectorization_strategy(sumfacts, width):
+ if width == 4:
+ horizontal, vertical = 2, 2
+ else:
+ raise NotImplementedError
+
+ result = {}
+
+ horizontal_kernels = horizontal_vectorization_strategy(sumfacts, horizontal, allow_padding=0)
+
+ for sf in horizontal_kernels:
+ vert = vertical_vectorization_strategy(horizontal_kernels[sf], width // horizontal_kernels[sf].horizontal_width)
+ for k in vert:
+ result[k] = vert[k]
+
+ return result
def decide_vectorization_strategy():
@@ -145,11 +170,14 @@ def decide_vectorization_strategy():
width = get_vcl_type_size(np.float64)
sfdict.update(**vertical_vectorization_strategy(sumfact, width))
elif get_option("vectorize_diagonal"):
- width = get_vcl_type_size(np.float64)
- sfdict.update(**diagonal_vectorization_stragegy(sumfact, width))
+ inputkeys = set(sf.input_key for sf in sumfacts)
+ for inputkey in inputkeys:
+ width = get_vcl_type_size(np.float64)
+ sumfact_filter = [sf for sf in sumfacts if sf.input_key == inputkey]
+ sfdict.update(**diagonal_vectorization_strategy(sumfact_filter, width))
else:
sfdict.update(**no_vectorization(sumfacts))
# Register the results
- for old, new in sfdict.items():
- _cache_vectorization_info(old, new)
+ for sf in sumfacts:
+ _cache_vectorization_info(sf, sfdict.get(sf, no_vec(sf)))
--
GitLab