| 1 | // This file is part of meshoptimizer library; see meshoptimizer.h for version/license details |
|---|---|
| 2 | #include "meshoptimizer.h" |
| 3 | |
| 4 | #include <assert.h> |
| 5 | #include <string.h> |
| 6 | |
| 7 | meshopt_VertexFetchStatistics meshopt_analyzeVertexFetch(const unsigned int* indices, size_t index_count, size_t vertex_count, size_t vertex_size) |
| 8 | { |
| 9 | assert(index_count % 3 == 0); |
| 10 | assert(vertex_size > 0 && vertex_size <= 256); |
| 11 | |
| 12 | meshopt_Allocator allocator; |
| 13 | |
| 14 | meshopt_VertexFetchStatistics result = {}; |
| 15 | |
| 16 | unsigned char* vertex_visited = allocator.allocate<unsigned char>(vertex_count); |
| 17 | memset(vertex_visited, 0, vertex_count); |
| 18 | |
| 19 | const size_t kCacheLine = 64; |
| 20 | const size_t kCacheSize = 128 * 1024; |
| 21 | |
| 22 | // simple direct mapped cache; on typical mesh data this is close to 4-way cache, and this model is a gross approximation anyway |
| 23 | size_t cache[kCacheSize / kCacheLine] = {}; |
| 24 | |
| 25 | for (size_t i = 0; i < index_count; ++i) |
| 26 | { |
| 27 | unsigned int index = indices[i]; |
| 28 | assert(index < vertex_count); |
| 29 | |
| 30 | vertex_visited[index] = 1; |
| 31 | |
| 32 | size_t start_address = index * vertex_size; |
| 33 | size_t end_address = start_address + vertex_size; |
| 34 | |
| 35 | size_t start_tag = start_address / kCacheLine; |
| 36 | size_t end_tag = (end_address + kCacheLine - 1) / kCacheLine; |
| 37 | |
| 38 | assert(start_tag < end_tag); |
| 39 | |
| 40 | for (size_t tag = start_tag; tag < end_tag; ++tag) |
| 41 | { |
| 42 | size_t line = tag % (sizeof(cache) / sizeof(cache[0])); |
| 43 | |
| 44 | // we store +1 since cache is filled with 0 by default |
| 45 | result.bytes_fetched += (cache[line] != tag + 1) * kCacheLine; |
| 46 | cache[line] = tag + 1; |
| 47 | } |
| 48 | } |
| 49 | |
| 50 | size_t unique_vertex_count = 0; |
| 51 | |
| 52 | for (size_t i = 0; i < vertex_count; ++i) |
| 53 | unique_vertex_count += vertex_visited[i]; |
| 54 | |
| 55 | result.overfetch = unique_vertex_count == 0 ? 0 : float(result.bytes_fetched) / float(unique_vertex_count * vertex_size); |
| 56 | |
| 57 | return result; |
| 58 | } |
| 59 |