67 lines · cpp
1#include "benchmarks/gpu/LibcGpuBenchmark.h"2#include "benchmarks/gpu/Random.h"3 4#include "hdr/stdint_proxy.h"5#include "src/math/sin.h"6#include "src/math/sinf.h"7 8#if defined(NVPTX_MATH_FOUND) || defined(AMDGPU_MATH_FOUND)9#include "platform.h"10#endif11 12#define BM_RANDOM_INPUT(T, Func, MinExp, MaxExp, N) \13 [](uint32_t call_index) { \14 using namespace LIBC_NAMESPACE::benchmarks; \15 \16 const UniformExponent<T> dist(MinExp, MaxExp); \17 return MathPerf<T>::template run_throughput<N>(Func, dist, call_index); \18 }19 20#define BENCH(T, Name, Func, MinExp, MaxExp) \21 SINGLE_WAVE_BENCHMARK(LlvmLibcSinGpuBenchmark, Name##_1, \22 BM_RANDOM_INPUT(T, Func, MinExp, MaxExp, 1)); \23 SINGLE_WAVE_BENCHMARK(LlvmLibcSinGpuBenchmark, Name##_128, \24 BM_RANDOM_INPUT(T, Func, MinExp, MaxExp, 128)); \25 SINGLE_WAVE_BENCHMARK(LlvmLibcSinGpuBenchmark, Name##_1024, \26 BM_RANDOM_INPUT(T, Func, MinExp, MaxExp, 1024)); \27 SINGLE_WAVE_BENCHMARK(LlvmLibcSinGpuBenchmark, Name##_4096, \28 BM_RANDOM_INPUT(T, Func, MinExp, MaxExp, 4096))29 30BENCH(double, Sin, LIBC_NAMESPACE::sin, -1023, 1023);31BENCH(double, SinTwoPi, LIBC_NAMESPACE::sin, -10, 3);32BENCH(double, SinTwoPow30, LIBC_NAMESPACE::sin, 0, 30);33BENCH(double, SinVeryLarge, LIBC_NAMESPACE::sin, 30, 1000);34 35#ifdef NVPTX_MATH_FOUND36BENCH(double, NvSin, __nv_sin, -1023, 1023);37BENCH(double, NvSinTwoPi, __nv_sin, -10, 3);38BENCH(double, NvSinTwoPow30, __nv_sin, 0, 30);39BENCH(double, NvSinVeryLarge, __nv_sin, 30, 1000);40#endif41 42#ifdef AMDGPU_MATH_FOUND43BENCH(double, AmdSin, __ocml_sin_f64, -1023, 1023);44BENCH(double, AmdSinTwoPi, __ocml_sin_f64, -10, 3);45BENCH(double, AmdSinTwoPow30, __ocml_sin_f64, 0, 30);46BENCH(double, AmdSinVeryLarge, __ocml_sin_f64, 30, 1000);47#endif48 49BENCH(float, Sinf, LIBC_NAMESPACE::sinf, -127, 128);50BENCH(float, SinfTwoPi, LIBC_NAMESPACE::sinf, -10, 3);51BENCH(float, SinfTwoPow30, LIBC_NAMESPACE::sinf, 0, 30);52BENCH(float, SinfVeryLarge, LIBC_NAMESPACE::sinf, 30, 120);53 54#ifdef NVPTX_MATH_FOUND55BENCH(float, NvSinf, __nv_sinf, -127, 128);56BENCH(float, NvSinfTwoPi, __nv_sinf, -10, 3);57BENCH(float, NvSinfTwoPow30, __nv_sinf, 0, 30);58BENCH(float, NvSinfVeryLarge, __nv_sinf, 30, 120);59#endif60 61#ifdef AMDGPU_MATH_FOUND62BENCH(float, AmdSinf, __ocml_sin_f32, -127, 128);63BENCH(float, AmdSinfTwoPi, __ocml_sin_f32, -10, 3);64BENCH(float, AmdSinfTwoPow30, __ocml_sin_f32, 0, 30);65BENCH(float, AmdSinfVeryLarge, __ocml_sin_f32, 30, 120);66#endif67