| 1 |
/* { dg-do compile } */ |
| 2 |
/* { dg-options "-march=rv32gcv_zvfhmin -mabi=ilp32d -mrvv-vector-bits=scalable -ffast-math -fdump-rtl-final" } */ |
| 3 |
|
| 4 |
#include <stdint-gcc.h> |
| 5 |
|
| 6 |
void f0 (_Float16 * __restrict a, _Float16 * __restrict b, int n) |
| 7 |
{ |
| 8 |
for (int i = 0; i < n; i++) |
| 9 |
a[i] = -b[i]; |
| 10 |
} |
| 11 |
|
| 12 |
void f1 (_Float16 * __restrict a, _Float16 * __restrict b, int n) |
| 13 |
{ |
| 14 |
for (int i = 0; i < n; i++) |
| 15 |
a[i] = a[i]+b[i]; |
| 16 |
} |
| 17 |
|
| 18 |
void f2 (_Float16 * __restrict a, _Float16 * __restrict b, int n) |
| 19 |
{ |
| 20 |
for (int i = 0; i < n; i++) |
| 21 |
a[i] = a[i]-b[i]; |
| 22 |
} |
| 23 |
|
| 24 |
void f3 (_Float16 * __restrict a, _Float16 * __restrict b, int n) |
| 25 |
{ |
| 26 |
for (int i = 0; i < n; i++) |
| 27 |
a[i] = a[i]*b[i]; |
| 28 |
} |
| 29 |
|
| 30 |
void f4 (_Float16 * __restrict a, _Float16 * __restrict b, int n) |
| 31 |
{ |
| 32 |
for (int i = 0; i < n; i++) |
| 33 |
a[i] = a[i]/b[i]; |
| 34 |
} |
| 35 |
|
| 36 |
void f6 (_Float16 * __restrict a, _Float16 * __restrict b, int n) |
| 37 |
{ |
| 38 |
for (int i = 0; i < n; i++) |
| 39 |
a[i] = __builtin_fabs (b[i]); |
| 40 |
} |
| 41 |
|
| 42 |
void f7 (_Float16 * __restrict a, _Float16 * __restrict b, int n) |
| 43 |
{ |
| 44 |
for (int i = 0; i < n; i++) |
| 45 |
a[i] = __builtin_sqrtf16 (b[i]); |
| 46 |
} |
| 47 |
|
| 48 |
void f8 (_Float16 * __restrict a, int16_t * __restrict b, int n) |
| 49 |
{ |
| 50 |
for (int i = 0; i < n; i++) |
| 51 |
a[i] = (_Float16) (b[i]); |
| 52 |
} |
| 53 |
|
| 54 |
void f9 (_Float16 * __restrict a, uint16_t * __restrict b, int n) |
| 55 |
{ |
| 56 |
for (int i = 0; i < n; i++) |
| 57 |
a[i] = (_Float16) (b[i]); |
| 58 |
} |
| 59 |
|
| 60 |
void f10 (int16_t * __restrict a, _Float16 * __restrict b, int n) |
| 61 |
{ |
| 62 |
for (int i = 0; i < n; i++) |
| 63 |
a[i] = (int16_t) (b[i]); |
| 64 |
} |
| 65 |
|
| 66 |
void f11 (uint16_t * __restrict a, _Float16 * __restrict b, int n) |
| 67 |
{ |
| 68 |
for (int i = 0; i < n; i++) |
| 69 |
a[i] = (uint16_t) (b[i]); |
| 70 |
} |
| 71 |
|
| 72 |
/* We can't enable FP16 NEG/PLUS/MINUS/MULT/DIV/ABS/SQRTF as well as int/float |
| 73 |
conversion auto-vectorization when -march="*zvfhmin*" because the min |
| 74 |
variant of the extension only provides loads, stores and conversions. |
| 75 |
As we might still vectorize after promotion to float, we need to make |
| 76 |
sure that no vector operations with an HFmode are being generated. */ |
| 77 |
/* { dg-final { scan-rtl-dump-not "plus:VNx\[0-9\]+HF" "final" } } */ |
| 78 |
/* { dg-final { scan-rtl-dump-not "minus:VNx\[0-9\]+HF" "final" } } */ |
| 79 |
/* { dg-final { scan-rtl-dump-not "mult:VNx\[0-9\]+HF" "final" } } */ |
| 80 |
/* { dg-final { scan-rtl-dump-not "div:VNx\[0-9\]+HF" "final" } } */ |
| 81 |
/* { dg-final { scan-rtl-dump-not "neg:VNx\[0-9\]+HF" "final" } } */ |
| 82 |
/* { dg-final { scan-rtl-dump-not "abs:VNx\[0-9\]+HF" "final" } } */ |
| 83 |
/* { dg-final { scan-rtl-dump-not "sqrt:VNx\[0-9\]+HF" "final" } } */ |
| 84 |
/* { dg-final { scan-rtl-dump-not "float:VNx\[0-9\]+HF" "final" } } */ |
| 85 |
/* { dg-final { scan-rtl-dump-not "fix:VNx\[0-9\]+HI\s*.+reg:VNx\[0-9\]+HF" "final" } } */ |