1+ // This program tests vectorized truncates & zero-extends for performance and
2+ // correctness
13#include < iostream>
24#include < memory>
35#include < random>
@@ -11,70 +13,178 @@ static std::mt19937 rng;
1113// Initialize array A with random numbers.
1214template <typename Ty>
1315static void init_data (const std::unique_ptr<Ty[]> &A, unsigned N) {
14- std::uniform_int_distribution<uint64_t > distrib (
15- std::numeric_limits<Ty>::min (), std::numeric_limits<Ty>::max ());
16- for (unsigned i = 0 ; i < N; i++)
17- A[i] = static_cast <Ty>(distrib (rng));
16+ std::uniform_int_distribution<Ty> distrib (std::numeric_limits<Ty>::min (),
17+ std::numeric_limits<Ty>::max ());
18+ for (unsigned I = 0 ; I < N; I++)
19+ A[I] = distrib (rng);
20+ }
21+
22+ // Truncate/Zero-extend elements to create expected results with no
23+ // vectorization
24+ template <typename Ty1, typename Ty2>
25+ static void truncOrZextWithNoVec (const Ty1 *A, Ty2 *B, int Iterations) {
26+ #pragma clang loop vectorize(disable)
27+ for (unsigned I = 0 ; I < Iterations; I++) {
28+ B[I] = A[I];
29+ }
1830}
1931
2032// Truncate/Zero-extend each vector element in a vectorized loop with vectorization width 8
21- template <typename Ty1, typename Ty2> static void truncOrZextVecInLoopWithVW8 (const Ty1 *A, Ty2 *B, int iterations) {
33+ template <typename Ty1, typename Ty2>
34+ static void truncOrZextVecInLoopWithVW8 (const Ty1 *A, Ty2 *B, int Iterations) {
2235#pragma clang loop vectorize_width(8) interleave_count(4)
23- for (unsigned i = 0 ; i < iterations; i++) {
24- B[i] = A[i];
36+ for (unsigned I = 0 ; I < Iterations; I++) {
37+ B[I] = A[I];
38+ }
39+ }
40+
41+ // Truncate/Zero-extend each vector element in a vectorized loop with
42+ // vectorization width 16
43+ template <typename Ty1, typename Ty2>
44+ static void truncOrZextVecInLoopWithVW16 (const Ty1 *A, Ty2 *B, int Iterations) {
45+ #pragma clang loop vectorize_width(16) interleave_count(4)
46+ for (unsigned I = 0 ; I < Iterations; I++) {
47+ B[I] = A[I];
48+ }
49+ }
50+
51+ // Truncate/Zero-extend each vector element in a vectorized loop
52+ template <typename Ty1, typename Ty2>
53+ static void truncOrZextVecInLoop (const Ty1 *A, Ty2 *B, int Iterations) {
54+ #pragma clang loop vectorize(enable)
55+ for (unsigned I = 0 ; I < Iterations; I++) {
56+ B[I] = A[I];
57+ }
58+ }
59+
60+ // Truncate/Zero-extend each vector element while adding in a vectorized loop
61+ // with vectorization width 8
62+ template <typename Ty1, typename Ty2>
63+ static void truncOrZextVecWithAddInLoopWithVW8 (const Ty1 *A, Ty2 *B,
64+ int Iterations) {
65+ #pragma clang loop vectorize_width(8) interleave_count(4)
66+ for (unsigned I = 0 ; I < Iterations; I++) {
67+ B[I] += A[I];
68+ }
69+ }
70+
71+ // Truncate/Zero-extend each vector element while adding in a vectorized loop
72+ // vectorization width 16
73+ template <typename Ty1, typename Ty2>
74+ static void truncOrZextVecWithAddInLoopWithVW16 (const Ty1 *A, Ty2 *B,
75+ int Iterations) {
76+ #pragma clang loop vectorize_width(16) interleave_count(4)
77+ for (unsigned I = 0 ; I < Iterations; I++) {
78+ B[I] += A[I];
79+ }
80+ }
81+
82+ // Truncate/Zero-extend each vector element while adding in a vectorized loop
83+ template <typename Ty1, typename Ty2>
84+ static void truncOrZextVecWithAddInLoop (const Ty1 *A, Ty2 *B, int Iterations) {
85+ #pragma clang loop vectorize(enable)
86+ for (unsigned I = 0 ; I < Iterations; I++) {
87+ B[I] += A[I];
2588 }
2689}
2790
28- template <typename Ty1, typename Ty2> static void __attribute__ ((always_inline))
29- benchForTruncOrZextVecInLoopWithVW8(benchmark::State &state) {
91+ template <typename Ty1, typename Ty2>
92+ static void __attribute__ ((always_inline))
93+ benchForTruncOrZextVecInLoop(benchmark::State &state,
94+ void (*Fn)(const Ty1 *, Ty2 *, int )) {
3095 std::unique_ptr<Ty1[]> A (new Ty1[ITERATIONS]);
3196 std::unique_ptr<Ty2[]> B (new Ty2[ITERATIONS]);
97+ std::unique_ptr<Ty2[]> C (new Ty2[ITERATIONS]);
98+
3299 init_data (A, ITERATIONS);
33- init_data (B, ITERATIONS);
100+
101+ // Check for correctness
102+ truncOrZextWithNoVec (&A[0 ], &C[0 ], ITERATIONS);
103+ Fn (&A[0 ], &B[0 ], ITERATIONS);
104+ for (int I = 0 ; I < ITERATIONS; I++) {
105+ if (B[I] != C[I]) {
106+ std::cerr << " ERROR: Trunc or ZExt operation on "
107+ << " is showing result " " instead of "
108+ << " \n "
109+ exit (1 );
110+ }
111+ }
112+
34113 for (auto _ : state) {
35114 benchmark::DoNotOptimize (B);
36115 benchmark::ClobberMemory ();
37- truncOrZextVecInLoopWithVW8 (&A[0 ], &B[0 ], ITERATIONS);
38- }
39- }
40-
41- // Truncate/Zero-extend each vector element in a vectorized loop
42- template <typename Ty1, typename Ty2> static void truncOrZextVecInLoop (const Ty1 *A, Ty2 *B, int iterations) {
43- #pragma clang loop interleave_count(4)
44- for (unsigned i = 0 ; i < iterations; i++) {
45- B[i] = A[i];
116+ Fn (&A[0 ], &B[0 ], ITERATIONS);
46117 }
47118}
48119
49- template <typename Ty1, typename Ty2> static void __attribute__ ((always_inline))
50- benchForTruncOrZextVecInLoop(benchmark::State &state) {
120+ template <typename Ty1, typename Ty2>
121+ static void __attribute__ ((always_inline))
122+ benchForTruncOrZextVecWithAddInLoop(benchmark::State &state,
123+ void (*Fn)(const Ty1 *, Ty2 *, int )) {
51124 std::unique_ptr<Ty1[]> A (new Ty1[ITERATIONS]);
52125 std::unique_ptr<Ty2[]> B (new Ty2[ITERATIONS]);
126+ std::unique_ptr<Ty2[]> C (new Ty2[ITERATIONS]);
53127 init_data (A, ITERATIONS);
54128 init_data (B, ITERATIONS);
55129 for (auto _ : state) {
56130 benchmark::DoNotOptimize (B);
57131 benchmark::ClobberMemory ();
58- truncOrZextVecInLoop (&A[0 ], &B[0 ], ITERATIONS);
132+ Fn (&A[0 ], &B[0 ], ITERATIONS);
59133 }
60134}
61135
62136// Add vectorized truncate or zero-extend operation benchmarks for different element types
63- #define ADD_BENCHMARK (ty1, ty2 ) \
64- void benchForTruncOrZextVecInLoopWithVW8From_##ty1##_To_##ty2##_(benchmark::State &state) { \
65- benchForTruncOrZextVecInLoopWithVW8<ty1, ty2>(state); \
137+ #define ADD_BENCHMARK (ty1, ty2 ) \
138+ void benchForTruncOrZextVecInLoopWithVW8From_##ty1##_To_##ty2##_( \
139+ benchmark::State &state) { \
140+ benchForTruncOrZextVecInLoop<ty1, ty2>(state, \
141+ &truncOrZextVecInLoopWithVW8); \
66142 } \
67- BENCHMARK (benchForTruncOrZextVecInLoopWithVW8From_##ty1##_To_##ty2##_); \
68- void benchForTruncOrZextVecInLoopFrom_##ty1##_To_##ty2##_(benchmark::State &state) { \
69- benchForTruncOrZextVecInLoop<ty1, ty2>(state); \
143+ BENCHMARK (benchForTruncOrZextVecInLoopWithVW8From_##ty1##_To_##ty2##_); \
144+ void benchForTruncOrZextVecInLoopWithVW16From_##ty1##_To_##ty2##_( \
145+ benchmark::State &state) { \
146+ benchForTruncOrZextVecInLoop<ty1, ty2>(state, \
147+ &truncOrZextVecInLoopWithVW16); \
70148 } \
71- BENCHMARK (benchForTruncOrZextVecInLoopFrom_##ty1##_To_##ty2##_); \
149+ BENCHMARK (benchForTruncOrZextVecInLoopWithVW16From_##ty1##_To_##ty2##_); \
150+ void benchForTruncOrZextVecInLoopFrom_##ty1##_To_##ty2##_( \
151+ benchmark::State &state) { \
152+ benchForTruncOrZextVecInLoop<ty1, ty2>(state, &truncOrZextVecInLoop); \
153+ } \
154+ BENCHMARK (benchForTruncOrZextVecInLoopFrom_##ty1##_To_##ty2##_); \
155+ void benchForTruncOrZextVecWithAddInLoopWithVW8From_##ty1##_To_##ty2##_( \
156+ benchmark::State &state) { \
157+ benchForTruncOrZextVecWithAddInLoop<ty1, ty2>( \
158+ state, &truncOrZextVecWithAddInLoopWithVW8); \
159+ } \
160+ BENCHMARK ( \
161+ benchForTruncOrZextVecWithAddInLoopWithVW8From_##ty1##_To_##ty2##_); \
162+ void benchForTruncOrZextVecWithAddInLoopWithVW16From_##ty1##_To_##ty2##_( \
163+ benchmark::State &state) { \
164+ benchForTruncOrZextVecWithAddInLoop<ty1, ty2>( \
165+ state, &truncOrZextVecWithAddInLoopWithVW16); \
166+ } \
167+ BENCHMARK ( \
168+ benchForTruncOrZextVecWithAddInLoopWithVW16From_##ty1##_To_##ty2##_); \
169+ void benchForTruncOrZextVecWithAddInLoopFrom_##ty1##_To_##ty2##_( \
170+ benchmark::State &state) { \
171+ benchForTruncOrZextVecWithAddInLoop<ty1, ty2>( \
172+ state, &truncOrZextVecWithAddInLoop); \
173+ } \
174+ BENCHMARK (benchForTruncOrZextVecWithAddInLoopFrom_##ty1##_To_##ty2##_);
72175
73176/* Vectorized truncate operations */
74- ADD_BENCHMARK (uint64_t , uint8_t )
75- ADD_BENCHMARK(uint32_t , uint8_t )
76177ADD_BENCHMARK (uint16_t , uint8_t )
77-
178+ ADD_BENCHMARK(uint32_t , uint8_t )
179+ ADD_BENCHMARK(uint64_t , uint8_t )
180+ ADD_BENCHMARK(uint32_t , uint16_t )
181+ ADD_BENCHMARK(uint64_t , uint16_t )
182+ ADD_BENCHMARK(uint64_t , uint32_t )
78183
79184/* Vectorized zero extend operations */
185+ ADD_BENCHMARK(uint8_t , uint16_t )
80186ADD_BENCHMARK(uint8_t , uint32_t )
187+ ADD_BENCHMARK(uint8_t , uint64_t )
188+ ADD_BENCHMARK(uint16_t , uint32_t )
189+ ADD_BENCHMARK(uint16_t , uint64_t )
190+ ADD_BENCHMARK(uint32_t , uint64_t )
0 commit comments