// File: huffman_codes.cpp // See Copyright Notice and license at the end of include/lzham.h #include "lzham_core.h" #include "lzham_huffman_codes.h" namespace lzham { struct sym_freq { uint m_freq; uint16 m_left; uint16 m_right; inline bool operator< (const sym_freq& other) const { return m_freq > other.m_freq; } }; static inline sym_freq* radix_sort_syms(uint num_syms, sym_freq* syms0, sym_freq* syms1) { const uint cMaxPasses = 2; uint hist[256 * cMaxPasses]; memset(hist, 0, sizeof(hist[0]) * 256 * cMaxPasses); { sym_freq* p = syms0; sym_freq* q = syms0 + (num_syms >> 1) * 2; for ( ; p != q; p += 2) { const uint freq0 = p[0].m_freq; const uint freq1 = p[1].m_freq; hist[ freq0 & 0xFF]++; hist[256 + ((freq0 >> 8) & 0xFF)]++; hist[ freq1 & 0xFF]++; hist[256 + ((freq1 >> 8) & 0xFF)]++; } if (num_syms & 1) { const uint freq = p->m_freq; hist[ freq & 0xFF]++; hist[256 + ((freq >> 8) & 0xFF)]++; } } sym_freq* pCur_syms = syms0; sym_freq* pNew_syms = syms1; const uint total_passes = (hist[256] == num_syms) ? 1 : cMaxPasses; for (uint pass = 0; pass < total_passes; pass++) { const uint* pHist = &hist[pass << 8]; uint offsets[256]; uint cur_ofs = 0; for (uint i = 0; i < 256; i += 2) { offsets[i] = cur_ofs; cur_ofs += pHist[i]; offsets[i+1] = cur_ofs; cur_ofs += pHist[i+1]; } const uint pass_shift = pass << 3; sym_freq* p = pCur_syms; sym_freq* q = pCur_syms + (num_syms >> 1) * 2; for ( ; p != q; p += 2) { uint c0 = p[0].m_freq; uint c1 = p[1].m_freq; if (pass) { c0 >>= 8; c1 >>= 8; } c0 &= 0xFF; c1 &= 0xFF; if (c0 == c1) { uint dst_offset0 = offsets[c0]; offsets[c0] = dst_offset0 + 2; pNew_syms[dst_offset0] = p[0]; pNew_syms[dst_offset0 + 1] = p[1]; } else { uint dst_offset0 = offsets[c0]++; uint dst_offset1 = offsets[c1]++; pNew_syms[dst_offset0] = p[0]; pNew_syms[dst_offset1] = p[1]; } } if (num_syms & 1) { uint c = ((p->m_freq) >> pass_shift) & 0xFF; uint dst_offset = offsets[c]; offsets[c] = dst_offset + 1; pNew_syms[dst_offset] = *p; } sym_freq* t = pCur_syms; pCur_syms = pNew_syms; pNew_syms = t; } #if LZHAM_ASSERTS_ENABLED uint prev_freq = 0; for (uint i = 0; i < num_syms; i++) { LZHAM_ASSERT(!(pCur_syms[i].m_freq < prev_freq)); prev_freq = pCur_syms[i].m_freq; } #endif return pCur_syms; } struct huffman_work_tables { enum { cMaxInternalNodes = cHuffmanMaxSupportedSyms }; sym_freq syms0[cHuffmanMaxSupportedSyms + 1 + cMaxInternalNodes]; sym_freq syms1[cHuffmanMaxSupportedSyms + 1 + cMaxInternalNodes]; #if !USE_CALCULATE_MINIMUM_REDUNDANCY uint16 queue[cMaxInternalNodes]; #endif }; uint get_generate_huffman_codes_table_size() { return sizeof(huffman_work_tables); } #define USE_CALCULATE_MINIMUM_REDUNDANCY 1 #if USE_CALCULATE_MINIMUM_REDUNDANCY /* calculate_minimum_redundancy() written by Alistair Moffat, alistair@cs.mu.oz.au, Jyrki Katajainen, jyrki@diku.dk November 1996. */ static void calculate_minimum_redundancy(int A[], int n) { int root; /* next root node to be used */ int leaf; /* next leaf to be used */ int next; /* next value to be assigned */ int avbl; /* number of available nodes */ int used; /* number of internal nodes */ int dpth; /* current depth of leaves */ /* check for pathological cases */ if (n==0) { return; } if (n==1) { A[0] = 0; return; } /* first pass, left to right, setting parent pointers */ A[0] += A[1]; root = 0; leaf = 2; for (next=1; next < n-1; next++) { /* select first item for a pairing */ if (leaf>=n || A[root]=n || (root=0; next--) A[next] = A[A[next]]+1; /* third pass, right to left, setting leaf depths */ avbl = 1; used = dpth = 0; root = n-2; next = n-1; while (avbl>0) { while (root>=0 && A[root]==dpth) { used++; root--; } while (avbl>used) { A[next--] = dpth; avbl--; } avbl = 2*used; dpth++; used = 0; } } #endif bool generate_huffman_codes(void* pContext, uint num_syms, const uint16* pFreq, uint8* pCodesizes, uint& max_code_size, uint& total_freq_ret) { if ((!num_syms) || (num_syms > cHuffmanMaxSupportedSyms)) return false; huffman_work_tables& state = *static_cast(pContext);; uint max_freq = 0; uint total_freq = 0; uint num_used_syms = 0; for (uint i = 0; i < num_syms; i++) { uint freq = pFreq[i]; if (!freq) pCodesizes[i] = 0; else { total_freq += freq; max_freq = math::maximum(max_freq, freq); sym_freq& sf = state.syms0[num_used_syms]; sf.m_left = (uint16)i; sf.m_right = UINT16_MAX; sf.m_freq = freq; num_used_syms++; } } total_freq_ret = total_freq; if (num_used_syms == 1) { pCodesizes[state.syms0[0].m_left] = 1; return true; } sym_freq* syms = radix_sort_syms(num_used_syms, state.syms0, state.syms1); #if USE_CALCULATE_MINIMUM_REDUNDANCY int x[cHuffmanMaxSupportedSyms]; for (uint i = 0; i < num_used_syms; i++) x[i] = syms[i].m_freq; calculate_minimum_redundancy(x, num_used_syms); uint max_len = 0; for (uint i = 0; i < num_used_syms; i++) { uint len = x[i]; max_len = math::maximum(len, max_len); pCodesizes[syms[i].m_left] = static_cast(len); } max_code_size = max_len; #else // Computes Huffman codelengths in linear time. More readable than calculate_minimum_redundancy(), and approximately the same speed, but not in-place. // Dummy node sym_freq& sf = state.syms0[num_used_syms]; sf.m_left = UINT16_MAX; sf.m_right = UINT16_MAX; sf.m_freq = UINT_MAX; uint next_internal_node = num_used_syms + 1; uint queue_front = 0; uint queue_end = 0; uint next_lowest_sym = 0; uint num_nodes_remaining = num_used_syms; do { uint left_freq = syms[next_lowest_sym].m_freq; uint left_child = next_lowest_sym; if ((queue_end > queue_front) && (syms[state.queue[queue_front]].m_freq < left_freq)) { left_child = state.queue[queue_front]; left_freq = syms[left_child].m_freq; queue_front++; } else next_lowest_sym++; uint right_freq = syms[next_lowest_sym].m_freq; uint right_child = next_lowest_sym; if ((queue_end > queue_front) && (syms[state.queue[queue_front]].m_freq < right_freq)) { right_child = state.queue[queue_front]; right_freq = syms[right_child].m_freq; queue_front++; } else next_lowest_sym++; LZHAM_ASSERT(next_internal_node < huffman_work_tables::cMaxInternalNodes); const uint internal_node_index = next_internal_node; next_internal_node++; syms[internal_node_index].m_freq = left_freq + right_freq; syms[internal_node_index].m_left = static_cast(left_child); syms[internal_node_index].m_right = static_cast(right_child); LZHAM_ASSERT(queue_end < huffman_work_tables::cMaxInternalNodes); state.queue[queue_end] = static_cast(internal_node_index); queue_end++; num_nodes_remaining--; } while (num_nodes_remaining > 1); LZHAM_ASSERT(next_lowest_sym == num_used_syms); LZHAM_ASSERT((queue_end - queue_front) == 1); uint cur_node_index = state.queue[queue_front]; uint32* pStack = (syms == state.syms0) ? (uint32*)state.syms1 : (uint32*)state.syms0; uint32* pStack_top = pStack; uint max_level = 0; for ( ; ; ) { uint level = cur_node_index >> 16; uint node_index = cur_node_index & 0xFFFF; uint left_child = syms[node_index].m_left; uint right_child = syms[node_index].m_right; uint next_level = (cur_node_index + 0x10000) & 0xFFFF0000; if (left_child < num_used_syms) { max_level = math::maximum(max_level, level); pCodesizes[syms[left_child].m_left] = static_cast(level + 1); if (right_child < num_used_syms) { pCodesizes[syms[right_child].m_left] = static_cast(level + 1); if (pStack == pStack_top) break; cur_node_index = *--pStack; } else { cur_node_index = next_level | right_child; } } else { if (right_child < num_used_syms) { max_level = math::maximum(max_level, level); pCodesizes[syms[right_child].m_left] = static_cast(level + 1); cur_node_index = next_level | left_child; } else { *pStack++ = next_level | left_child; cur_node_index = next_level | right_child; } } } max_code_size = max_level + 1; #endif return true; } } // namespace lzham