| 1 |
/* { dg-do run } */ |
| 2 |
/* { dg-options "-O3 --save-temps -ffast-math" } */ |
| 3 |
|
| 4 |
#include <arm_neon.h> |
| 5 |
|
| 6 |
extern void abort (void); |
| 7 |
|
| 8 |
extern float fabsf (float); |
| 9 |
extern double fabs (double); |
| 10 |
extern int isnan (double); |
| 11 |
extern float fmaxf (float, float); |
| 12 |
extern float fminf (float, float); |
| 13 |
extern double fmax (double, double); |
| 14 |
extern double fmin (double, double); |
| 15 |
|
| 16 |
#define NUM_TESTS 16 |
| 17 |
#define DELTA 0.000001 |
| 18 |
#define NAN (0.0 / 0.0) |
| 19 |
|
| 20 |
float input_float32[] = {0.1f, -0.1f, 0.4f, 10.3f, |
| 21 |
200.0f, -800.0f, -13.0f, -0.5f, |
| 22 |
NAN, -870.0f, 10.4f, 310.11f, |
| 23 |
0.0f, -865.0f, -2213.0f, -1.5f}; |
| 24 |
|
| 25 |
double input_float64[] = {0.1, -0.1, 0.4, 10.3, |
| 26 |
200.0, -800.0, -13.0, -0.5, |
| 27 |
NAN, -870.0, 10.4, 310.11, |
| 28 |
0.0, -865.0, -2213.0, -1.5}; |
| 29 |
|
| 30 |
#define EQUALF(a, b) (fabsf (a - b) < DELTA) |
| 31 |
#define EQUALD(a, b) (fabs (a - b) < DELTA) |
| 32 |
|
| 33 |
/* Floating point 'unordered' variants. */ |
| 34 |
|
| 35 |
#undef TEST |
| 36 |
#define TEST(MAXMIN, CMP_OP, SUFFIX, Q, TYPE, LANES, FLOAT) \ |
| 37 |
int \ |
| 38 |
test_v##MAXMIN##v##SUFFIX##_##TYPE##x##LANES##_t (void) \ |
| 39 |
{ \ |
| 40 |
int i, j; \ |
| 41 |
int moves = (NUM_TESTS - LANES) + 1; \ |
| 42 |
TYPE##_t out_l[NUM_TESTS]; \ |
| 43 |
TYPE##_t out_v[NUM_TESTS]; \ |
| 44 |
\ |
| 45 |
/* Calculate linearly. */ \ |
| 46 |
for (i = 0; i < moves; i++) \ |
| 47 |
{ \ |
| 48 |
out_l[i] = input_##TYPE[i]; \ |
| 49 |
for (j = 0; j < LANES; j++) \ |
| 50 |
{ \ |
| 51 |
if (isnan (out_l[i])) \ |
| 52 |
continue; \ |
| 53 |
if (isnan (input_##TYPE[i + j]) \ |
| 54 |
|| input_##TYPE[i + j] CMP_OP out_l[i]) \ |
| 55 |
out_l[i] = input_##TYPE[i + j]; \ |
| 56 |
} \ |
| 57 |
} \ |
| 58 |
\ |
| 59 |
/* Calculate using vector reduction intrinsics. */ \ |
| 60 |
for (i = 0; i < moves; i++) \ |
| 61 |
{ \ |
| 62 |
TYPE##x##LANES##_t t1 = vld1##Q##_##SUFFIX (input_##TYPE + i); \ |
| 63 |
out_v[i] = v##MAXMIN##v##Q##_##SUFFIX (t1); \ |
| 64 |
} \ |
| 65 |
\ |
| 66 |
/* Compare. */ \ |
| 67 |
for (i = 0; i < moves; i++) \ |
| 68 |
{ \ |
| 69 |
if (!EQUAL##FLOAT (out_v[i], out_l[i]) \ |
| 70 |
&& !(isnan (out_v[i]) && isnan (out_l[i]))) \ |
| 71 |
return 0; \ |
| 72 |
} \ |
| 73 |
return 1; \ |
| 74 |
} |
| 75 |
|
| 76 |
#define BUILD_VARIANTS(TYPE, STYPE, W32, W64, F) \ |
| 77 |
TEST (max, >, STYPE, , TYPE, W32, F) \ |
| 78 |
TEST (max, >, STYPE, q, TYPE, W64, F) \ |
| 79 |
TEST (min, <, STYPE, , TYPE, W32, F) \ |
| 80 |
TEST (min, <, STYPE, q, TYPE, W64, F) |
| 81 |
|
| 82 |
BUILD_VARIANTS (float32, f32, 2, 4, F) |
| 83 |
/* { dg-final { scan-assembler "fmaxp\\ts\[0-9\]+, v\[0-9\]+\.2s" } } */ |
| 84 |
/* { dg-final { scan-assembler "fminp\\ts\[0-9\]+, v\[0-9\]+\.2s" } } */ |
| 85 |
/* { dg-final { scan-assembler "fmaxv\\ts\[0-9\]+, v\[0-9\]+\.4s" } } */ |
| 86 |
/* { dg-final { scan-assembler "fminv\\ts\[0-9\]+, v\[0-9\]+\.4s" } } */ |
| 87 |
TEST (max, >, f64, q, float64, 2, D) |
| 88 |
/* { dg-final { scan-assembler "fmaxp\\td\[0-9\]+, v\[0-9\]+\.2d" } } */ |
| 89 |
TEST (min, <, f64, q, float64, 2, D) |
| 90 |
/* { dg-final { scan-assembler "fminp\\td\[0-9\]+, v\[0-9\]+\.2d" } } */ |
| 91 |
|
| 92 |
/* Floating point 'nm' variants. */ |
| 93 |
|
| 94 |
#undef TEST |
| 95 |
#define TEST(MAXMIN, F, SUFFIX, Q, TYPE, LANES, FLOAT) \ |
| 96 |
int \ |
| 97 |
test_v##MAXMIN##nmv##SUFFIX##_##TYPE##x##LANES##_t (void) \ |
| 98 |
{ \ |
| 99 |
int i, j; \ |
| 100 |
int moves = (NUM_TESTS - LANES) + 1; \ |
| 101 |
TYPE##_t out_l[NUM_TESTS]; \ |
| 102 |
TYPE##_t out_v[NUM_TESTS]; \ |
| 103 |
\ |
| 104 |
/* Calculate linearly. */ \ |
| 105 |
for (i = 0; i < moves; i++) \ |
| 106 |
{ \ |
| 107 |
out_l[i] = input_##TYPE[i]; \ |
| 108 |
for (j = 0; j < LANES; j++) \ |
| 109 |
out_l[i] = f##MAXMIN##F (input_##TYPE[i + j], out_l[i]); \ |
| 110 |
} \ |
| 111 |
\ |
| 112 |
/* Calculate using vector reduction intrinsics. */ \ |
| 113 |
for (i = 0; i < moves; i++) \ |
| 114 |
{ \ |
| 115 |
TYPE##x##LANES##_t t1 = vld1##Q##_##SUFFIX (input_##TYPE + i); \ |
| 116 |
out_v[i] = v##MAXMIN##nmv##Q##_##SUFFIX (t1); \ |
| 117 |
} \ |
| 118 |
\ |
| 119 |
/* Compare. */ \ |
| 120 |
for (i = 0; i < moves; i++) \ |
| 121 |
{ \ |
| 122 |
if (!EQUAL##FLOAT (out_v[i], out_l[i])) \ |
| 123 |
return 0; \ |
| 124 |
} \ |
| 125 |
return 1; \ |
| 126 |
} |
| 127 |
|
| 128 |
TEST (max, f, f32, , float32, 2, D) |
| 129 |
/* { dg-final { scan-assembler "fmaxnmp\\ts\[0-9\]+, v\[0-9\]+\.2s" } } */ |
| 130 |
TEST (min, f, f32, , float32, 2, D) |
| 131 |
/* { dg-final { scan-assembler "fminnmp\\ts\[0-9\]+, v\[0-9\]+\.2s" } } */ |
| 132 |
TEST (max, f, f32, q, float32, 4, D) |
| 133 |
/* { dg-final { scan-assembler "fmaxnmv\\ts\[0-9\]+, v\[0-9\]+\.4s" } } */ |
| 134 |
TEST (min, f, f32, q, float32, 4, D) |
| 135 |
/* { dg-final { scan-assembler "fminnmv\\ts\[0-9\]+, v\[0-9\]+\.4s" } } */ |
| 136 |
TEST (max, , f64, q, float64, 2, D) |
| 137 |
/* { dg-final { scan-assembler "fmaxnmp\\td\[0-9\]+, v\[0-9\]+\.2d" } } */ |
| 138 |
TEST (min, , f64, q, float64, 2, D) |
| 139 |
/* { dg-final { scan-assembler "fminnmp\\td\[0-9\]+, v\[0-9\]+\.2d" } } */ |
| 140 |
|
| 141 |
#undef TEST |
| 142 |
#define TEST(MAXMIN, CMP_OP, SUFFIX, Q, TYPE, LANES, FLOAT) \ |
| 143 |
{ \ |
| 144 |
if (!test_v##MAXMIN##v##SUFFIX##_##TYPE##x##LANES##_t ()) \ |
| 145 |
abort (); \ |
| 146 |
} |
| 147 |
|
| 148 |
int |
| 149 |
main (int argc, char **argv) |
| 150 |
{ |
| 151 |
BUILD_VARIANTS (float32, f32, 2, 4, F) |
| 152 |
TEST (max, >, f64, q, float64, 2, D) |
| 153 |
TEST (min, <, f64, q, float64, 2, D) |
| 154 |
|
| 155 |
#undef TEST |
| 156 |
#define TEST(MAXMIN, CMP_OP, SUFFIX, Q, TYPE, LANES, FLOAT) \ |
| 157 |
{ \ |
| 158 |
if (!test_v##MAXMIN##nmv##SUFFIX##_##TYPE##x##LANES##_t ()) \ |
| 159 |
abort (); \ |
| 160 |
} |
| 161 |
|
| 162 |
BUILD_VARIANTS (float32, f32, 2, 4, F) |
| 163 |
TEST (max, >, f64, q, float64, 2, D) |
| 164 |
TEST (min, <, f64, q, float64, 2, D) |
| 165 |
|
| 166 |
return 0; |
| 167 |
} |
| 168 |
|