| 1 |
#include <cmath> |
| 2 |
#include <iostream> |
| 3 |
#include <memory> |
| 4 |
#include <stdexcept> |
| 5 |
|
| 6 |
#include <hip/hip_fp16.h> |
| 7 |
#include <hip/hip_runtime.h> |
| 8 |
#include <math.h> |
| 9 |
|
| 10 |
namespace { |
| 11 |
template <class T, class F> |
| 12 |
__global__ void global_entry_point(F f, T* out) |
| 13 |
{ |
| 14 |
*out = f(); |
| 15 |
} |
| 16 |
|
| 17 |
template <class T, class F> |
| 18 |
bool verify(F f, T expected) |
| 19 |
{ |
| 20 |
std::unique_ptr<T> cpu_T(new T); |
| 21 |
T* gpu_T = nullptr; |
| 22 |
bool result = true; |
| 23 |
{ |
| 24 |
hipError_t err = hipMalloc((void**)&gpu_T, sizeof(T)); |
| 25 |
if (err != hipSuccess) { |
| 26 |
std::cerr << "hipMalloc failed: " << hipGetErrorString(err) << std::endl; |
| 27 |
result = false; |
| 28 |
} |
| 29 |
} |
| 30 |
hipLaunchKernelGGL(global_entry_point, 1, 1, 0, 0, f, gpu_T); |
| 31 |
{ |
| 32 |
hipError_t err = |
| 33 |
hipMemcpy(cpu_T.get(), gpu_T, sizeof(T), hipMemcpyDeviceToHost); |
| 34 |
if (err != hipSuccess) { |
| 35 |
std::cerr << "hipMemcpy failed: " << hipGetErrorString(err) << std::endl; |
| 36 |
result = false; |
| 37 |
} |
| 38 |
} |
| 39 |
result = hipFree(gpu_T) == hipSuccess && result; |
| 40 |
result = *cpu_T == expected && result; |
| 41 |
return result; |
| 42 |
} |
| 43 |
} |
| 44 |
|
| 45 |
int main(int argc, char** argv) |
| 46 |
{ |
| 47 |
bool valid = verify([] __device__() { return std::round(1.4f); }, 1.0f); |
| 48 |
valid &= verify([] __device__() { return max<_Float16>(1.0f, 2.0f); }, 2.0f); |
| 49 |
valid &= verify([] __device__() { return min<_Float16>(1.0f, 2.0f); }, 1.0f); |
| 50 |
|
| 51 |
if (valid) { |
| 52 |
return 0; |
| 53 |
} else { |
| 54 |
return 1; |
| 55 |
} |
| 56 |
} |