noahbackus/godot
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 2 Wiki Activity

Update embree to 3.13.5

Browse Source
This commit is contained in:
DeeJayLSP
2022-11-24 11:45:59 -03:00
committed by Rémi Verschelde
parent f16c5b564b
commit 5e4158eb48
110 changed files with 9963 additions and 6494 deletions
Download Patch File Download Diff File Expand all files Collapse all files
thirdparty/embree/common/simd/vint4_sse2.h
Vendored
+77 -23
View File
@@ -106,7 +106,14 @@ namespace embree
#endif
#if defined(__SSE4_1__)
#if defined(__aarch64__)
static __forceinline vint4 load(const unsigned char* ptr) {
return _mm_load4epu8_epi32(((__m128i*)ptr));
}
static __forceinline vint4 loadu(const unsigned char* ptr) {
return _mm_load4epu8_epi32(((__m128i*)ptr));
}
#elif defined(__SSE4_1__)
static __forceinline vint4 load(const unsigned char* ptr) {
return _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr));
}
@@ -127,7 +134,9 @@ namespace embree
#endif
static __forceinline vint4 load(const unsigned short* ptr) {
#if defined (__SSE4_1__)
#if defined(__aarch64__)
return __m128i(vmovl_u16(vld1_u16(ptr)));
#elif defined (__SSE4_1__)
return _mm_cvtepu16_epi32(_mm_loadu_si128((__m128i*)ptr));
#else
return vint4(ptr[0],ptr[1],ptr[2],ptr[3]);
@@ -135,7 +144,12 @@ namespace embree
}
static __forceinline void store(unsigned char* ptr, const vint4& v) {
#if defined(__SSE4_1__)
#if defined(__aarch64__)
int32x4_t x = v;
uint16x4_t y = vqmovn_u32(uint32x4_t(x));
uint8x8_t z = vqmovn_u16(vcombine_u16(y, y));
vst1_lane_u32((uint32_t *)ptr,uint32x2_t(z), 0);
#elif defined(__SSE4_1__)
__m128i x = v;
x = _mm_packus_epi32(x, x);
x = _mm_packus_epi16(x, x);
@@ -147,20 +161,26 @@ namespace embree
}
static __forceinline void store(unsigned short* ptr, const vint4& v) {
#if defined(__aarch64__)
uint32x4_t x = uint32x4_t(v.v);
uint16x4_t y = vqmovn_u32(x);
vst1_u16(ptr, y);
#else
for (size_t i=0;i<4;i++)
ptr[i] = (unsigned short)v[i];
#endif
}
static __forceinline vint4 load_nt(void* ptr) {
#if defined(__SSE4_1__)
return _mm_stream_load_si128((__m128i*)ptr);
#if defined(__aarch64__) || defined(__SSE4_1__)
return _mm_stream_load_si128((__m128i*)ptr);
#else
return _mm_load_si128((__m128i*)ptr);
#endif
}
static __forceinline void store_nt(void* ptr, const vint4& v) {
#if defined(__SSE4_1__)
#if !defined(__aarch64__) && defined(__SSE4_1__)
_mm_stream_ps((float*)ptr, _mm_castsi128_ps(v));
#else
_mm_store_si128((__m128i*)ptr,v);
@@ -169,7 +189,7 @@ namespace embree
template<int scale = 4>
static __forceinline vint4 gather(const int* ptr, const vint4& index) {
#if defined(__AVX2__)
#if defined(__AVX2__) && !defined(__aarch64__)
return _mm_i32gather_epi32(ptr, index, scale);
#else
return vint4(
@@ -185,7 +205,7 @@ namespace embree
vint4 r = zero;
#if defined(__AVX512VL__)
return _mm_mmask_i32gather_epi32(r, mask, index, ptr, scale);
#elif defined(__AVX2__)
#elif defined(__AVX2__) && !defined(__aarch64__)
return _mm_mask_i32gather_epi32(r, ptr, index, mask, scale);
#else
if (likely(mask[0])) r[0] = *(int*)(((char*)ptr)+scale*index[0]);
@@ -222,7 +242,7 @@ namespace embree
#endif
}
#if defined(__x86_64__)
#if defined(__x86_64__) || defined(__aarch64__)
static __forceinline vint4 broadcast64(long long a) { return _mm_set1_epi64x(a); }
#endif
@@ -236,6 +256,8 @@ namespace embree
friend __forceinline vint4 select(const vboolf4& m, const vint4& t, const vint4& f) {
#if defined(__AVX512VL__)
return _mm_mask_blend_epi32(m, (__m128i)f, (__m128i)t);
#elif defined(__aarch64__)
return _mm_castps_si128(_mm_blendv_ps((__m128)f.v,(__m128) t.v, (__m128)m.v));
#elif defined(__SSE4_1__)
return _mm_castps_si128(_mm_blendv_ps(_mm_castsi128_ps(f), _mm_castsi128_ps(t), m));
#else
@@ -256,7 +278,9 @@ namespace embree
__forceinline vint4 operator +(const vint4& a) { return a; }
__forceinline vint4 operator -(const vint4& a) { return _mm_sub_epi32(_mm_setzero_si128(), a); }
#if defined(__SSSE3__)
#if defined(__aarch64__)
__forceinline vint4 abs(const vint4& a) { return vabsq_s32(a.v); }
#elif defined(__SSSE3__)
__forceinline vint4 abs(const vint4& a) { return _mm_abs_epi32(a); }
#endif
@@ -272,7 +296,7 @@ namespace embree
__forceinline vint4 operator -(const vint4& a, int b) { return a - vint4(b); }
__forceinline vint4 operator -(int a, const vint4& b) { return vint4(a) - b; }
#if defined(__SSE4_1__)
#if (defined(__aarch64__)) || defined(__SSE4_1__)
__forceinline vint4 operator *(const vint4& a, const vint4& b) { return _mm_mullo_epi32(a, b); }
#else
__forceinline vint4 operator *(const vint4& a, const vint4& b) { return vint4(a[0]*b[0],a[1]*b[1],a[2]*b[2],a[3]*b[3]); }
@@ -292,8 +316,8 @@ namespace embree
__forceinline vint4 operator ^(const vint4& a, int b) { return a ^ vint4(b); }
__forceinline vint4 operator ^(int a, const vint4& b) { return vint4(a) ^ b; }
__forceinline vint4 operator <<(const vint4& a, int n) { return _mm_slli_epi32(a, n); }
__forceinline vint4 operator >>(const vint4& a, int n) { return _mm_srai_epi32(a, n); }
__forceinline vint4 operator <<(const vint4& a, const int n) { return _mm_slli_epi32(a, n); }
__forceinline vint4 operator >>(const vint4& a, const int n) { return _mm_srai_epi32(a, n); }
__forceinline vint4 sll (const vint4& a, int b) { return _mm_slli_epi32(a, b); }
__forceinline vint4 sra (const vint4& a, int b) { return _mm_srai_epi32(a, b); }
@@ -309,7 +333,7 @@ namespace embree
__forceinline vint4& operator -=(vint4& a, const vint4& b) { return a = a - b; }
__forceinline vint4& operator -=(vint4& a, int b) { return a = a - b; }
#if defined(__SSE4_1__)
#if (defined(__aarch64__)) || defined(__SSE4_1__)
__forceinline vint4& operator *=(vint4& a, const vint4& b) { return a = a * b; }
__forceinline vint4& operator *=(vint4& a, int b) { return a = a * b; }
#endif
@@ -393,7 +417,7 @@ namespace embree
#endif
}
#if defined(__SSE4_1__)
#if defined(__aarch64__) || defined(__SSE4_1__)
__forceinline vint4 min(const vint4& a, const vint4& b) { return _mm_min_epi32(a, b); }
__forceinline vint4 max(const vint4& a, const vint4& b) { return _mm_max_epi32(a, b); }
@@ -417,6 +441,16 @@ namespace embree
__forceinline vint4 unpacklo(const vint4& a, const vint4& b) { return _mm_castps_si128(_mm_unpacklo_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b))); }
__forceinline vint4 unpackhi(const vint4& a, const vint4& b) { return _mm_castps_si128(_mm_unpackhi_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b))); }
#if defined(__aarch64__)
template<int i0, int i1, int i2, int i3>
__forceinline vint4 shuffle(const vint4& v) {
return vreinterpretq_s32_u8(vqtbl1q_u8( (uint8x16_t)v.v, _MN_SHUFFLE(i0, i1, i2, i3)));
}
template<int i0, int i1, int i2, int i3>
__forceinline vint4 shuffle(const vint4& a, const vint4& b) {
return vreinterpretq_s32_u8(vqtbl2q_u8( (uint8x16x2_t){(uint8x16_t)a.v, (uint8x16_t)b.v}, _MF_SHUFFLE(i0, i1, i2, i3)));
}
#else
template<int i0, int i1, int i2, int i3>
__forceinline vint4 shuffle(const vint4& v) {
return _mm_shuffle_epi32(v, _MM_SHUFFLE(i3, i2, i1, i0));
@@ -426,7 +460,7 @@ namespace embree
__forceinline vint4 shuffle(const vint4& a, const vint4& b) {
return _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), _MM_SHUFFLE(i3, i2, i1, i0)));
}
#endif
#if defined(__SSE3__)
template<> __forceinline vint4 shuffle<0, 0, 2, 2>(const vint4& v) { return _mm_castps_si128(_mm_moveldup_ps(_mm_castsi128_ps(v))); }
template<> __forceinline vint4 shuffle<1, 1, 3, 3>(const vint4& v) { return _mm_castps_si128(_mm_movehdup_ps(_mm_castsi128_ps(v))); }
@@ -438,7 +472,7 @@ namespace embree
return shuffle<i,i,i,i>(v);
}
#if defined(__SSE4_1__)
#if defined(__SSE4_1__) && !defined(__aarch64__)
template<int src> __forceinline int extract(const vint4& b) { return _mm_extract_epi32(b, src); }
template<int dst> __forceinline vint4 insert(const vint4& a, const int b) { return _mm_insert_epi32(a, b, dst); }
#else
@@ -446,18 +480,27 @@ namespace embree
template<int dst> __forceinline vint4 insert(const vint4& a, int b) { vint4 c = a; c[dst&3] = b; return c; }
#endif
template<> __forceinline int extract<0>(const vint4& b) { return _mm_cvtsi128_si32(b); }
__forceinline int toScalar(const vint4& v) { return _mm_cvtsi128_si32(v); }
__forceinline size_t toSizeT(const vint4& v) {
#if defined(__aarch64__)
__forceinline size_t toSizeT(const vint4& v) {
uint64x2_t x = uint64x2_t(v.v);
return x[0];
}
#else
__forceinline size_t toSizeT(const vint4& v) {
#if defined(__WIN32__) && !defined(__X86_64__) // win32 workaround
return toScalar(v);
#elif defined(__ARM_NEON)
// FIXME(LTE): Do we need a swap(i.e. use lane 1)?
return vgetq_lane_u64(*(reinterpret_cast<const uint64x2_t *>(&v)), 0);
#else
return _mm_cvtsi128_si64(v);
#endif
}
#endif
#if defined(__AVX512VL__)
@@ -475,7 +518,17 @@ namespace embree
/// Reductions
////////////////////////////////////////////////////////////////////////////////
#if defined(__SSE4_1__)
#if defined(__aarch64__) || defined(__SSE4_1__)
#if defined(__aarch64__)
__forceinline vint4 vreduce_min(const vint4& v) { int h = vminvq_s32(v); return vdupq_n_s32(h); }
__forceinline vint4 vreduce_max(const vint4& v) { int h = vmaxvq_s32(v); return vdupq_n_s32(h); }
__forceinline vint4 vreduce_add(const vint4& v) { int h = vaddvq_s32(v); return vdupq_n_s32(h); }
__forceinline int reduce_min(const vint4& v) { return vminvq_s32(v); }
__forceinline int reduce_max(const vint4& v) { return vmaxvq_s32(v); }
__forceinline int reduce_add(const vint4& v) { return vaddvq_s32(v); }
#else
__forceinline vint4 vreduce_min(const vint4& v) { vint4 h = min(shuffle<1,0,3,2>(v),v); return min(shuffle<2,3,0,1>(h),h); }
__forceinline vint4 vreduce_max(const vint4& v) { vint4 h = max(shuffle<1,0,3,2>(v),v); return max(shuffle<2,3,0,1>(h),h); }
__forceinline vint4 vreduce_add(const vint4& v) { vint4 h = shuffle<1,0,3,2>(v) + v ; return shuffle<2,3,0,1>(h) + h ; }
@@ -483,6 +536,7 @@ namespace embree
__forceinline int reduce_min(const vint4& v) { return toScalar(vreduce_min(v)); }
__forceinline int reduce_max(const vint4& v) { return toScalar(vreduce_max(v)); }
__forceinline int reduce_add(const vint4& v) { return toScalar(vreduce_add(v)); }
#endif
__forceinline size_t select_min(const vint4& v) { return bsf(movemask(v == vreduce_min(v))); }
__forceinline size_t select_max(const vint4& v) { return bsf(movemask(v == vreduce_max(v))); }
@@ -502,7 +556,7 @@ namespace embree
/// Sorting networks
////////////////////////////////////////////////////////////////////////////////
#if defined(__SSE4_1__)
#if (defined(__aarch64__)) || defined(__SSE4_1__)
__forceinline vint4 usort_ascending(const vint4& v)
{