Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Commit b67c019

Browse files
huntergr-armhuntergr-arm
authored
[Vectorizer] Add early exit with stores functional tests (#352)
Some simple early exit tests with stores before and/or after the uncountable exit.
1 parent 27d20d9 commit b67c019

2 files changed

Lines changed: 177 additions & 0 deletions

File tree

‎SingleSource/UnitTests/Vectorizer/early-exit.cpp‎

Lines changed: 171 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,7 @@
11
#include <algorithm>
22
#include <functional>
33
#include <iostream>
4+
#include <numeric>
45

56
/// N is divisible by common VFs. Used to test vector code path without the
67
/// scalar epilogue.
@@ -73,6 +74,68 @@ checkVectorFunction(TestFnTy<Ty> ScalarFn, TestFnTy<Ty> VectorFn,
7374
}
7475
}
7576

77+
78+
// Test function types for two arrays, one used for stores and the other
79+
// for loop control
80+
template <typename Ty>
81+
using InitFnStTy =
82+
std::function<void(Ty *, Ty *)>;
83+
template <typename Ty>
84+
using TestFnStTy =
85+
std::function<Ty(InitFnStTy<Ty>)>;
86+
87+
template <typename Ty>
88+
static void
89+
checkVectorWithStoresFunction(TestFnStTy<Ty> ScalarFn, TestFnStTy<Ty> VectorFn,
90+
TestFnStTy<Ty> ForcedVectorFn, TestFnStTy<Ty> InterleavedFn,
91+
TestFnStTy<Ty> InterleavedOnlyFn, const char *Name) {
92+
std::cout << "Checking " << Name << "\n";
93+
94+
std::array Tests = {std::make_pair(VectorFn, "autovec"),
95+
std::make_pair(ForcedVectorFn, "vector-forced"),
96+
std::make_pair(InterleavedFn, "interleave-forced"),
97+
std::make_pair(InterleavedOnlyFn, "interleave-only")};
98+
99+
// Check finding the target element at all indices between 0 and N.
100+
for (unsigned IdxToFind = 0; IdxToFind < N; ++IdxToFind) {
101+
// Lambda to initialize all array elements to one, except the one to look
102+
// for to zero.
103+
auto Init1 = [IdxToFind](Ty *Data, Ty *Pred) {
104+
std::fill_n(Pred, N, 1);
105+
Pred[IdxToFind] = 0;
106+
std::fill_n(Data, N, 0);
107+
};
108+
109+
// Lambda to initialize all array elements to one, except the one to look
110+
// for and the IdxToFind + 3 to zero.
111+
auto Init2 = [IdxToFind](Ty *Data, Ty *Pred) {
112+
std::fill_n(Pred, N, 1);
113+
Pred[IdxToFind] = 0;
114+
if (IdxToFind + 3 < N)
115+
Pred[IdxToFind + 3] = 0;
116+
std::fill_n(Data, N, 0);
117+
};
118+
119+
auto Reference1 = ScalarFn(Init1);
120+
auto Reference2 = ScalarFn(Init2);
121+
// Run vector functions and check against the scalar result.
122+
for (const auto &[Fn, Name] : Tests) {
123+
auto ToCheck1 = callThroughOptnone(Fn, Init1);
124+
if (Reference1 != ToCheck1) {
125+
std::cerr << "Miscompare for " << Name << ": " << Reference1
126+
<< " != " << ToCheck1 << "\n";
127+
exit(1);
128+
}
129+
auto ToCheck2 = callThroughOptnone(Fn, Init2);
130+
if (Reference2 != ToCheck2) {
131+
std::cerr << "Miscompare for " << Name << ": " << Reference2
132+
<< " != " << ToCheck2 << "\n";
133+
exit(1);
134+
}
135+
}
136+
}
137+
}
138+
76139
// Test function type for three arrays
77140
template <typename Ty>
78141
using TestFnTy3 =
@@ -248,6 +311,34 @@ checkVectorFunctionMulti(TestFnTy3<Ty> ScalarFn, TestFnTy3<Ty> VectorFn,
248311
_Pragma("clang loop vectorize_width(1) interleave_count(4)") Loop \
249312
};
250313

314+
/// Define test functions for single early exits with stores
315+
#define DEFINE_EARLY_EXIT_WITH_STORES(Ty, Init, Data, Pred, Loop) \
316+
auto ScalarFn = [](InitFnStTy<Ty> II) -> Ty { \
317+
Init; \
318+
II(Data, Pred); \
319+
_Pragma("clang loop vectorize(disable) interleave_count(1)") Loop \
320+
}; \
321+
auto VectorFn = [](InitFnStTy<Ty> II) -> Ty { \
322+
Init; \
323+
II(Data, Pred); \
324+
_Pragma("clang loop vectorize(enable)") Loop \
325+
}; \
326+
auto ForcedVectorFn = [](InitFnStTy<Ty> II) -> Ty { \
327+
Init; \
328+
II(Data, Pred); \
329+
_Pragma("clang loop vectorize_width(8) interleave_count(1)") Loop \
330+
}; \
331+
auto InterleavedFn = [](InitFnStTy<Ty> II) -> Ty { \
332+
Init; \
333+
II(Data, Pred); \
334+
_Pragma("clang loop vectorize(enable) interleave_count(4)") Loop \
335+
}; \
336+
auto InterleavedOnlyFn = [](InitFnStTy<Ty> II) -> Ty { \
337+
Init; \
338+
II(Data, Pred); \
339+
_Pragma("clang loop vectorize_width(1) interleave_count(4)") Loop \
340+
};
341+
251342
int main(void) {
252343
{
253344
DEFINE_SCALAR_AND_VECTOR_EARLY_EXIT(
@@ -454,5 +545,85 @@ int main(void) {
454545
"two_early_exits_small_different_values");
455546
}
456547

548+
{
549+
DEFINE_EARLY_EXIT_WITH_STORES(
550+
int,
551+
int Data[N]; int Pred[N];
552+
, Data, Pred, for (unsigned I = 0; I < N; I++) {
553+
Data[I] = I;
554+
if (Pred[I] == 0)
555+
break;
556+
} return std::reduce(Data, Data + N););
557+
checkVectorWithStoresFunction<int>(ScalarFn, VectorFn, ForcedVectorFn,
558+
InterleavedFn, InterleavedOnlyFn, "exit_after_store");
559+
}
560+
561+
{
562+
DEFINE_EARLY_EXIT_WITH_STORES(
563+
int,
564+
int Data[N]; int Pred[N];
565+
, Data, Pred, for (unsigned I = 0; I < N; I++) {
566+
if (Pred[I] == 0)
567+
break;
568+
Data[I] = I;
569+
} return std::reduce(Data, Data + N););
570+
checkVectorWithStoresFunction<int>(ScalarFn, VectorFn, ForcedVectorFn,
571+
InterleavedFn, InterleavedOnlyFn, "exit_before_store");
572+
}
573+
574+
{
575+
DEFINE_EARLY_EXIT_WITH_STORES(
576+
int,
577+
int Data[N]; int Pred[N];
578+
, Data, Pred, for (unsigned I = 0; I < N; I++) {
579+
Data[I] = I;
580+
if (Pred[I] == 0)
581+
break;
582+
Data[I] = I + 3;
583+
} return std::reduce(Data, Data + N););
584+
checkVectorWithStoresFunction<int>(ScalarFn, VectorFn, ForcedVectorFn,
585+
InterleavedFn, InterleavedOnlyFn, "exit_between_stores");
586+
}
587+
588+
{
589+
DEFINE_EARLY_EXIT_WITH_STORES(
590+
int,
591+
int Data[N]; int Pred[N]; unsigned I;
592+
, Data, Pred, for (I = 0; I < N; I++) {
593+
Data[I] = I;
594+
if (Pred[I] == 0)
595+
break;
596+
} return I + std::reduce(Data, Data + N););
597+
checkVectorWithStoresFunction<int>(ScalarFn, VectorFn, ForcedVectorFn,
598+
InterleavedFn, InterleavedOnlyFn,
599+
"exit_after_store_live_out");
600+
}
601+
602+
{
603+
DEFINE_EARLY_EXIT_WITH_STORES(
604+
float,
605+
float Data[N]; float Pred[N];
606+
, Data, Pred, for (unsigned I = 0; I < N; I++) {
607+
Data[I] = I;
608+
if (Pred[I] == 0.0f)
609+
break;
610+
} return std::reduce(Data, Data + N););
611+
checkVectorWithStoresFunction<float>(ScalarFn, VectorFn, ForcedVectorFn,
612+
InterleavedFn, InterleavedOnlyFn, "exit_after_float_store");
613+
}
614+
615+
{
616+
DEFINE_EARLY_EXIT_WITH_STORES(
617+
int, int Data[N]; int Pred[N]; unsigned I;
618+
, Data, Pred, for (I = 0; I < N; I++) {
619+
Data[I] = Data[I] * 3 + 5;
620+
if (Pred[I] == 0)
621+
break;
622+
} return I + std::reduce(Data, Data + N););
623+
checkVectorWithStoresFunction<int>(ScalarFn, VectorFn, ForcedVectorFn,
624+
InterleavedFn, InterleavedOnlyFn,
625+
"exit_after_load_update_store");
626+
}
627+
457628
return 0;
458629
}

‎SingleSource/UnitTests/Vectorizer/early-exit.reference_output‎

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -11,4 +11,10 @@ Checking three_early_exits_odd_trip_count
1111
Checking two_early_exits_small_known_tc
1212
Checking three_early_exits_small_known_tc
1313
Checking two_early_exits_small_different_values
14+
Checking exit_after_store
15+
Checking exit_before_store
16+
Checking exit_between_stores
17+
Checking exit_after_store_live_out
18+
Checking exit_after_float_store
19+
Checking exit_after_load_update_store
1420
exit 0

0 commit comments

Comments
 (0)