diff --git a/MEMO b/MEMO
index 89b67986351d093a7e7dff01479ce72e1df781ae..694b49a929d7d6e25c522fec3552c5f6abd6d288 100644
--- a/MEMO
+++ b/MEMO
@@ -78,7 +78,7 @@ Future ideas
- Check for unordered (no-dependency) writes to the same location
-- String instructions?
+- Vanilla C string instructions?
- Barriers for data exchanged via global vars?
diff --git a/loopy/__init__.py b/loopy/__init__.py
index 0dc0fdf76e7ee38c6b47e97912df0fb1f3127396..eb89d65707a60b1be41f5fbbc57945fc3536f1e5 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -353,6 +353,8 @@ def add_prefetch(kernel, var_name, sweep_inames=[], dim_arg_names=None,
arg = kernel.arg_dict[var_name]
+ # {{{ make parameter names and unification template
+
parameters = []
for i in range(arg.dimensions):
based_on = "%s_dim_%d" % (c_name, i)
@@ -371,8 +373,12 @@ def add_prefetch(kernel, var_name, sweep_inames=[], dim_arg_names=None,
elif len(parameters) == 1:
uni_template = uni_template[var(parameters[0])]
+ # }}}
+
kernel = extract_subst(kernel, rule_name, uni_template, parameters)
+ # {{{ track applied iname rewrites on footprint_subscripts
+
if footprint_subscripts is not None:
if not isinstance(footprint_subscripts, (list, tuple)):
footprint_subscripts = [footprint_subscripts]
@@ -403,6 +409,8 @@ def add_prefetch(kernel, var_name, sweep_inames=[], dim_arg_names=None,
else:
subst_use = rule_name
+ # }}}
+
new_kernel = precompute(kernel, subst_use, arg.dtype, sweep_inames,
new_storage_axis_names=dim_arg_names,
default_tag=default_tag)
diff --git a/loopy/kernel.py b/loopy/kernel.py
index d8d78ece8ce38e8f921cfbe61af27c62b41dc104..f7b2f1590b0ba1e5d434a190a586ee3736dc4039 100644
--- a/loopy/kernel.py
+++ b/loopy/kernel.py
@@ -1690,14 +1690,14 @@ class SetOperationCacheManager:
lower_bound_pw_aff = self.dim_min(set, iname_to_dim[iname][1])
upper_bound_pw_aff = self.dim_max(set, iname_to_dim[iname][1])
- from loopy.isl_helpers import static_max_of_pw_aff, static_min_of_pw_aff
+ from loopy.isl_helpers import static_max_of_pw_aff, static_value_of_pw_aff
from loopy.symbolic import pw_aff_to_expr
size = pw_aff_to_expr(static_max_of_pw_aff(
upper_bound_pw_aff - lower_bound_pw_aff + 1, constants_only=True,
context=context))
base_index = pw_aff_to_expr(
- static_min_of_pw_aff(lower_bound_pw_aff, constants_only=False,
+ static_value_of_pw_aff(lower_bound_pw_aff, constants_only=False,
context=context))
return base_index, size
diff --git a/loopy/reduction.py b/loopy/reduction.py
index 9a09b3c0793c431212cb42e55c6a0b9504dae0fd..9f2b83487144f00cb29a4dd08bae356566ec8359 100644
--- a/loopy/reduction.py
+++ b/loopy/reduction.py
@@ -33,7 +33,7 @@ class ScalarReductionOperation(ReductionOperation):
result = type(self).__name__.replace("ReductionOperation", "").lower()
if self.forced_result_dtype is not None:
- result = "%s<%s>" % (result, str(self.dtype))
+ result = "%s<%s>" % (result, str(self.forced_result_dtype))
return result
diff --git a/test/test_linalg.py b/test/test_linalg.py
index fefdaac15ec87c4c65e89202c951d2b1dbee8f6d..d4db1941ba144f890732510c058fb03e6694279a 100644
--- a/test/test_linalg.py
+++ b/test/test_linalg.py
@@ -221,7 +221,7 @@ def test_plain_matrix_mul(ctx_factory):
kernel_gen = lp.check_kernels(kernel_gen, {})
lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen,
- op_count=vec_size*2*n**3/1e9, op_label="GFlops/s",
+ op_count=[vec_size*2*n**3/1e9], op_label=["GFlops"],
parameters={"n": n}, check_result=check)
@@ -354,7 +354,7 @@ def test_rank_one(ctx_factory):
kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
- op_count=np.dtype(dtype).itemsize*n**2/1e9, op_label="GBytes",
+ op_count=[np.dtype(dtype).itemsize*n**2/1e9], op_label=["GBytes"],
parameters={"n": n})
diff --git a/test/test_loopy.py b/test/test_loopy.py
index 2c447e9cffcad0dfb44176934363479075c852d9..e4644bcf1de36cf5bc4416f2de2f6d998409edc9 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -89,39 +89,41 @@ def test_multi_cse(ctx_factory):
def test_stencil(ctx_factory):
ctx = ctx_factory()
+ # n=32 causes corner case behavior in size calculations for temprorary (a
+ # non-unifiable, two-constant-segments PwAff as the base index)
+
+ n = 256
knl = lp.make_kernel(ctx.devices[0],
- "{[i,j]: 0<= i,j < 32}",
+ "{[i,j]: 0<= i,j < %d}" % n,
[
- "[i] z[i,j] = -2*a[i,j]"
- " + a[i,j-1]"
- " + a[i,j+1]"
- " + a[i-1,j]"
- " + a[i+1,j]"
+ "a_offset(ii, jj) := a[ii+1, jj+1]",
+ "z[i,j] = -2*a_offset(i,j)"
+ " + a_offset(i,j-1)"
+ " + a_offset(i,j+1)"
+ " + a_offset(i-1,j)"
+ " + a_offset(i+1,j)"
],
[
- lp.GlobalArg("a", np.float32, shape=(32,32,)),
- lp.GlobalArg("z", np.float32, shape=(32,32,))
+ lp.GlobalArg("a", np.float32, shape=(n+2,n+2,)),
+ lp.GlobalArg("z", np.float32, shape=(n+2,n+2,))
])
+ ref_knl = knl
def variant_1(knl):
- knl = lp.add_prefetch(knl, "a", [0, 1])
- return knl
-
- def variant_2(knl):
knl = lp.split_dimension(knl, "i", 16, outer_tag="g.1", inner_tag="l.1")
knl = lp.split_dimension(knl, "j", 16, outer_tag="g.0", inner_tag="l.0")
knl = lp.add_prefetch(knl, "a", ["i_inner", "j_inner"])
return knl
- #for variant in [variant_1, variant_2]:
- for variant in [variant_2]:
+ for variant in [variant_1]:
kernel_gen = lp.generate_loop_schedules(variant(knl),
loop_priority=["i_outer", "i_inner_0", "j_0"])
kernel_gen = lp.check_kernels(kernel_gen)
- for knl in kernel_gen:
- print lp.generate_code(knl)
+ lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen,
+ fills_entire_output=False, print_ref_code=True,
+ op_count=[n*n], op_label=["cells"])