mirror of
https://github.com/holub/mame
synced 2025-05-04 13:33:05 +03:00
0837e7451a
* Sync with bgfx upstream revision b91d0b6 * Sync with bx upstream revision d60912b * Sync with bimg upstream revision bd81f60 * Add astc-codec decoder * Rename VertexDecl to VertexLayout * Rename UniformType enum Int1 to Sampler. * Add NVN stub * Fix unused-const-variable error on macOS * Drop redundant explicit language parameters buildoptions_cpp are only applied to c++ files and buildoptions_objcpp are only applied to objective c++ files. As such, hardcoding -x offers no benefit while preventing overrides (such as one needed by 3rdparty/bgfx/src/renderer_vk.cpp on macOS) from working. * Re-introduce -x c++ in places where C code is compiled as C++ to prevent clang from throwing a warning * Build bgfx as Objective-C++ on macOS It is needed due to included headers * Enable Direct3D12 and Vulkan bgfx rendering backends * Enable building of spirv shaders * Properly escape /c in cmd call * Comment out dx12 bgfx renderer * Honor VERBOSE setting during shaders build * Only invert hlsl shader XYZ_TO_sRGB matrix for opengl * Add spirv shaders * OpenGL ES needs transposed matrix too * Metal needs transposed matrix as well
134 lines
3.4 KiB
C++
134 lines
3.4 KiB
C++
/*
|
|
* Copyright 2010-2019 Branimir Karadzic. All rights reserved.
|
|
* License: https://github.com/bkaradzic/bx#license-bsd-2-clause
|
|
*/
|
|
|
|
#include <bx/allocator.h>
|
|
#include <bx/rng.h>
|
|
#include <bx/simd_t.h>
|
|
#include <bx/timer.h>
|
|
|
|
#include <stdio.h>
|
|
|
|
static void flushCache()
|
|
{
|
|
static uint32_t length = 1 << 26;
|
|
static uint8_t* input = new uint8_t[length];
|
|
static uint8_t* output = new uint8_t[length];
|
|
bx::memCopy(output, input, length);
|
|
}
|
|
|
|
typedef bx::simd128_t (*SimdRsqrtFn)(bx::simd128_t _a);
|
|
|
|
template<SimdRsqrtFn simdRsqrtFn>
|
|
void simd_rsqrt_bench(bx::simd128_t* _dst, bx::simd128_t* _src, uint32_t _numVertices)
|
|
{
|
|
for (uint32_t ii = 0, num = _numVertices/4; ii < num; ++ii)
|
|
{
|
|
bx::simd128_t* ptr = &_src[ii*4];
|
|
bx::simd128_t tmp0 = bx::simd_ld(ptr + 0);
|
|
bx::simd128_t tmp1 = bx::simd_ld(ptr + 1);
|
|
bx::simd128_t tmp2 = bx::simd_ld(ptr + 2);
|
|
bx::simd128_t tmp3 = bx::simd_ld(ptr + 3);
|
|
bx::simd128_t rsqrt0 = simdRsqrtFn(tmp0);
|
|
bx::simd128_t rsqrt1 = simdRsqrtFn(tmp1);
|
|
bx::simd128_t rsqrt2 = simdRsqrtFn(tmp2);
|
|
bx::simd128_t rsqrt3 = simdRsqrtFn(tmp3);
|
|
|
|
ptr = &_dst[ii*4];
|
|
bx::simd_st(ptr + 0, rsqrt0);
|
|
bx::simd_st(ptr + 1, rsqrt1);
|
|
bx::simd_st(ptr + 2, rsqrt2);
|
|
bx::simd_st(ptr + 3, rsqrt3);
|
|
}
|
|
}
|
|
|
|
void simd_bench_pass(bx::simd128_t* _dst, bx::simd128_t* _src, uint32_t _numVertices)
|
|
{
|
|
const uint32_t numIterations = 10;
|
|
|
|
{
|
|
int64_t elapsed = 0;
|
|
for (uint32_t test = 0; test < numIterations; ++test)
|
|
{
|
|
flushCache();
|
|
elapsed += -bx::getHPCounter();
|
|
simd_rsqrt_bench<bx::simd_rsqrt_est>(_dst, _src, _numVertices);
|
|
elapsed += bx::getHPCounter();
|
|
}
|
|
printf(" simd_rsqrt_est: %15f\n", double(elapsed) );
|
|
}
|
|
|
|
{
|
|
int64_t elapsed = 0;
|
|
for (uint32_t test = 0; test < numIterations; ++test)
|
|
{
|
|
flushCache();
|
|
elapsed += -bx::getHPCounter();
|
|
simd_rsqrt_bench<bx::simd_rsqrt_nr>(_dst, _src, _numVertices);
|
|
elapsed += bx::getHPCounter();
|
|
}
|
|
printf(" simd_rsqrt_nr: %15f\n", double(elapsed) );
|
|
}
|
|
|
|
{
|
|
int64_t elapsed = 0;
|
|
for (uint32_t test = 0; test < numIterations; ++test)
|
|
{
|
|
flushCache();
|
|
elapsed += -bx::getHPCounter();
|
|
simd_rsqrt_bench<bx::simd_rsqrt_carmack>(_dst, _src, _numVertices);
|
|
elapsed += bx::getHPCounter();
|
|
}
|
|
printf("simd_rsqrt_carmack: %15f\n", double(elapsed) );
|
|
}
|
|
|
|
{
|
|
int64_t elapsed = 0;
|
|
for (uint32_t test = 0; test < numIterations; ++test)
|
|
{
|
|
flushCache();
|
|
elapsed += -bx::getHPCounter();
|
|
simd_rsqrt_bench<bx::simd_rsqrt>(_dst, _src, _numVertices);
|
|
elapsed += bx::getHPCounter();
|
|
}
|
|
printf(" simd_rsqrt: %15f\n", double(elapsed) );
|
|
}
|
|
}
|
|
|
|
void simd_bench()
|
|
{
|
|
bx::DefaultAllocator allocator;
|
|
bx::RngMwc rng;
|
|
|
|
const uint32_t numVertices = 1024*1024;
|
|
|
|
uint8_t* data = (uint8_t*)BX_ALIGNED_ALLOC(&allocator, 2*numVertices*sizeof(bx::simd128_t), 16);
|
|
bx::simd128_t* src = (bx::simd128_t*)data;
|
|
bx::simd128_t* dst = &src[numVertices];
|
|
|
|
printf("\n -- positive & negative --\n");
|
|
for (uint32_t ii = 0; ii < numVertices; ++ii)
|
|
{
|
|
float* ptr = (float*)&src[ii];
|
|
bx::store(ptr, bx::randUnitSphere(&rng) );
|
|
ptr[3] = 1.0f;
|
|
}
|
|
|
|
simd_bench_pass(dst, src, numVertices);
|
|
|
|
printf("\n -- positive only --\n");
|
|
for (uint32_t ii = 0; ii < numVertices; ++ii)
|
|
{
|
|
float* ptr = (float*)&src[ii];
|
|
ptr[0] = bx::abs(ptr[0]);
|
|
ptr[1] = bx::abs(ptr[1]);
|
|
ptr[2] = bx::abs(ptr[2]);
|
|
ptr[3] = bx::abs(ptr[3]);
|
|
}
|
|
|
|
simd_bench_pass(dst, src, numVertices);
|
|
|
|
BX_ALIGNED_FREE(&allocator, data, 16);
|
|
}
|