47 lines · cpp
1#include "benchmarks/gpu/LibcGpuBenchmark.h"2 3#include "hdr/stdint_proxy.h"4#include "src/math/atan2.h"5 6#if defined(NVPTX_MATH_FOUND) || defined(AMDGPU_MATH_FOUND)7#include "platform.h"8#endif9 10#define BM_RANDOM_INPUTS(T, Func, MinExp, MaxExp, N) \11 [](uint32_t call_index) { \12 using namespace LIBC_NAMESPACE::benchmarks; \13 \14 const UniformExponent<T> dist(MinExp, MaxExp); \15 return MathPerf<T>::template run_throughput<N>(Func, dist, dist, \16 call_index); \17 }18 19#define BENCH(T, Name, Func, MinExp, MaxExp) \20 SINGLE_WAVE_BENCHMARK(LlvmLibcAtan2GpuBenchmark, Name##_1, \21 BM_RANDOM_INPUTS(T, Func, MinExp, MaxExp, 1)); \22 SINGLE_WAVE_BENCHMARK(LlvmLibcAtan2GpuBenchmark, Name##_128, \23 BM_RANDOM_INPUTS(T, Func, MinExp, MaxExp, 128)); \24 SINGLE_WAVE_BENCHMARK(LlvmLibcAtan2GpuBenchmark, Name##_1024, \25 BM_RANDOM_INPUTS(T, Func, MinExp, MaxExp, 1024)); \26 SINGLE_WAVE_BENCHMARK(LlvmLibcAtan2GpuBenchmark, Name##_4096, \27 BM_RANDOM_INPUTS(T, Func, MinExp, MaxExp, 4096))28 29BENCH(double, Atan2, LIBC_NAMESPACE::atan2, -1023, 1023);30BENCH(double, Atan2TwoPi, LIBC_NAMESPACE::atan2, -10, 3);31BENCH(double, Atan2TwoPow30, LIBC_NAMESPACE::atan2, 0, 30);32BENCH(double, Atan2Large, LIBC_NAMESPACE::atan2, 30, 1000);33 34#ifdef NVPTX_MATH_FOUND35BENCH(double, NvAtan2, __nv_atan2, -1023, 1023);36BENCH(double, NvAtan2TwoPi, __nv_atan2, -10, 3);37BENCH(double, NvAtan2TwoPow30, __nv_atan2, 0, 30);38BENCH(double, NvAtan2Large, __nv_atan2, 30, 1000);39#endif40 41#ifdef AMDGPU_MATH_FOUND42BENCH(double, AmdAtan2, __ocml_atan2_f64, -1023, 1023);43BENCH(double, AmdAtan2TwoPi, __ocml_atan2_f64, -10, 3);44BENCH(double, AmdAtan2TwoPow30, __ocml_atan2_f64, 0, 30);45BENCH(double, AmdAtan2Large, __ocml_atan2_f64, 30, 1000);46#endif47