| 63 | } |
| 64 | |
| 65 | static void computeHistogram(unsigned int (&hist)[1024][3], const unsigned int* data, size_t count) |
| 66 | { |
| 67 | memset(hist, 0, sizeof(hist)); |
| 68 | |
| 69 | // compute 3 10-bit histograms in parallel |
| 70 | for (size_t i = 0; i < count; ++i) |
| 71 | { |
| 72 | unsigned int id = data[i]; |
| 73 | |
| 74 | hist[(id >> 0) & 1023][0]++; |
| 75 | hist[(id >> 10) & 1023][1]++; |
| 76 | hist[(id >> 20) & 1023][2]++; |
| 77 | } |
| 78 | |
| 79 | unsigned int sumx = 0, sumy = 0, sumz = 0; |
| 80 | |
| 81 | // replace histogram data with prefix histogram sums in-place |
| 82 | for (int i = 0; i < 1024; ++i) |
| 83 | { |
| 84 | unsigned int hx = hist[i][0], hy = hist[i][1], hz = hist[i][2]; |
| 85 | |
| 86 | hist[i][0] = sumx; |
| 87 | hist[i][1] = sumy; |
| 88 | hist[i][2] = sumz; |
| 89 | |
| 90 | sumx += hx; |
| 91 | sumy += hy; |
| 92 | sumz += hz; |
| 93 | } |
| 94 | |
| 95 | assert(sumx == count && sumy == count && sumz == count); |
| 96 | } |
| 97 | |
| 98 | static void radixPass(unsigned int* destination, const unsigned int* source, const unsigned int* keys, size_t count, unsigned int (&hist)[1024][3], int pass) |
| 99 | { |
no outgoing calls
no test coverage detected