C++及系统软件技术大会演讲稿合辑_部分26.pdfVIP

Vec3f struct Vec3f { __m128 xyz; }; void cross(Vec3f r, const Vec3f a, const Vec3f b) { const f128 A = _mm_shuffle_ps(a.xyz, a.xyz, _MM_SHUFFLE(3, 0, 2, 1)); const f128 B = _mm_shuffle_ps(b.xyz, b.xyz, _MM_SHUFFLE(3, 1, 0, 2)); const f128 C = _mm_shuffle_ps(a.xyz, a.xyz, _MM_SHUFFLE(3, 1, 0, 2)); const f128 D = _mm_shuffle_ps(b.xyz, b.xyz, _MM_SHUFFLE(3, 0, 2, 1)); r.xyz = _mm_fmsub_ps(A, B, _mm_mul_ps(C, D)); } Vec3f Structure of Array (SOA) struct Vec3f { __m256 x; __m256 y; __m256 z; }; void cross(Vec3f __restrict r, const Vec3f __restrict a, const Vec3f __restrict b) { r.x = _mm256_mul_ps(a.z, b.y); r.y = _mm256_mul_ps(a.x, b.z); r.z = _mm256_mul_ps(a.y, b.x); r.x = _mm256_fmsub_ps(a.y, b.z, r.x); r.y = _mm256_fmsub_ps(a.z, b.x, r.y); r.z = _mm256_fmsub_ps(a.x, b.y, r.z); } SIMD-at-Insomniac-Games-How 向量归一化 void normalize(Vec3f v) { __m256 len = _mm256_sqrt_ps(dot(v, v)); v.x = _mm256_div_ps(v.x, len); v.y = _mm256_div_ps(v.y, len); v.z = _mm256_div_ps(v.z, len); } 用乘法代替除法 void normalize(Vec3f v) { __m256 len = _mm256_sqrt_ps(dot(v, v)); len = _mm256_div_ps(_mm256_set1_ps(1.0f), len); v.x = _mm256_mul_ps(v.x, len); v.y = _mm256_mul_ps(v.y, len); v.z = _mm256_mul_ps(v.z, len); } rcp指令 • 除法仍然是最慢的计算 • 在Haswell架构上，vdivps的latency是17-21 ，throughput是13。 • vrcpps近似计算倒数指令 • 在Haswell架构上，latency是7 ，throughput是2。 void normalize(Vec3f v) { __m256 len = _mm256_sqrt_ps(dot(v, v)); len = _mm256_rcp_ps(len); v.x = _mm256_mul_ps(v.x, len); v.y = _mm256_mul_ps(v.y, len); v.z = _mm256_mul_ps(v.z, len); } 完全使用rcp代替div ？ rcp误差 • The maximum relative error for this approximation is less than 1.5*2^-12 • _mm256_rcp_ps • rcp(1.0f) = 0.999756。 • normalize()用于模型表面法线归一化 • 几何相交测试 • _mm256_rsqrt_ps std::vector void makeNumbers(std::vectorint values) { f