| 1 | // Copyright 2009-2021 Intel Corporation |
| 2 | // SPDX-License-Identifier: Apache-2.0 |
| 3 | |
| 4 | #pragma once |
| 5 | |
| 6 | #include "../math/math.h" |
| 7 | |
| 8 | /* include SSE wrapper classes */ |
| 9 | #if defined(__SSE__) || defined(__ARM_NEON) |
| 10 | # include "sse.h" |
| 11 | #endif |
| 12 | |
| 13 | /* include AVX wrapper classes */ |
| 14 | #if defined(__AVX__) |
| 15 | # include "avx.h" |
| 16 | #endif |
| 17 | |
| 18 | /* include AVX512 wrapper classes */ |
| 19 | #if defined (__AVX512F__) |
| 20 | # include "avx512.h" |
| 21 | #endif |
| 22 | |
| 23 | namespace embree |
| 24 | { |
| 25 | template <int N> |
| 26 | __forceinline vbool<N> isfinite(const vfloat<N>& v) |
| 27 | { |
| 28 | return (v >= vfloat<N>(-std::numeric_limits<float>::max())) |
| 29 | & (v <= vfloat<N>( std::numeric_limits<float>::max())); |
| 30 | } |
| 31 | |
| 32 | /* foreach unique */ |
| 33 | template<typename vbool, typename vint, typename Closure> |
| 34 | __forceinline void foreach_unique(const vbool& valid0, const vint& vi, const Closure& closure) |
| 35 | { |
| 36 | vbool valid1 = valid0; |
| 37 | while (any(valid1)) { |
| 38 | const int j = int(bsf(movemask(valid1))); |
| 39 | const int i = vi[j]; |
| 40 | const vbool valid2 = valid1 & (i == vi); |
| 41 | valid1 = andn(valid1, valid2); |
| 42 | closure(valid2, i); |
| 43 | } |
| 44 | } |
| 45 | |
| 46 | /* returns the next unique value i in vi and the corresponding valid_i mask */ |
| 47 | template<typename vbool, typename vint> |
| 48 | __forceinline int next_unique(vbool& valid, const vint& vi, /*out*/ vbool& valid_i) |
| 49 | { |
| 50 | assert(any(valid)); |
| 51 | const int j = int(bsf(movemask(valid))); |
| 52 | const int i = vi[j]; |
| 53 | valid_i = valid & (i == vi); |
| 54 | valid = andn(valid, valid_i); |
| 55 | return i; |
| 56 | } |
| 57 | |
| 58 | /* foreach unique index */ |
| 59 | template<typename vbool, typename vint, typename Closure> |
| 60 | __forceinline void foreach_unique_index(const vbool& valid0, const vint& vi, const Closure& closure) |
| 61 | { |
| 62 | vbool valid1 = valid0; |
| 63 | while (any(valid1)) { |
| 64 | const int j = int(bsf(movemask(valid1))); |
| 65 | const int i = vi[j]; |
| 66 | const vbool valid2 = valid1 & (i == vi); |
| 67 | valid1 = andn(valid1, valid2); |
| 68 | closure(valid2, i, j); |
| 69 | } |
| 70 | } |
| 71 | |
| 72 | /* returns the index of the next unique value i in vi and the corresponding valid_i mask */ |
| 73 | template<typename vbool, typename vint> |
| 74 | __forceinline int next_unique_index(vbool& valid, const vint& vi, /*out*/ vbool& valid_i) |
| 75 | { |
| 76 | assert(any(valid)); |
| 77 | const int j = int(bsf(movemask(valid))); |
| 78 | const int i = vi[j]; |
| 79 | valid_i = valid & (i == vi); |
| 80 | valid = andn(valid, valid_i); |
| 81 | return j; |
| 82 | } |
| 83 | |
| 84 | template<typename Closure> |
| 85 | __forceinline void foreach2(int x0, int x1, int y0, int y1, const Closure& closure) |
| 86 | { |
| 87 | __aligned(64) int U[2*VSIZEX]; |
| 88 | __aligned(64) int V[2*VSIZEX]; |
| 89 | int index = 0; |
| 90 | for (int y=y0; y<y1; y++) { |
| 91 | const bool lasty = y+1>=y1; |
| 92 | const vintx vy = y; |
| 93 | for (int x=x0; x<x1; ) { //x+=VSIZEX) { |
| 94 | const bool lastx = x+VSIZEX >= x1; |
| 95 | vintx vx = x+vintx(step); |
| 96 | vintx::storeu(&U[index], vx); |
| 97 | vintx::storeu(&V[index], vy); |
| 98 | const int dx = min(x1-x,VSIZEX); |
| 99 | index += dx; |
| 100 | x += dx; |
| 101 | if (index >= VSIZEX || (lastx && lasty)) { |
| 102 | const vboolx valid = vintx(step) < vintx(index); |
| 103 | closure(valid, vintx::load(U), vintx::load(V)); |
| 104 | x-= max(0, index-VSIZEX); |
| 105 | index = 0; |
| 106 | } |
| 107 | } |
| 108 | } |
| 109 | } |
| 110 | } |
| 111 | |