| 1 | // Copyright 2014 Google Inc. All Rights Reserved. |
| 2 | // |
| 3 | // Use of this source code is governed by a BSD-style license |
| 4 | // that can be found in the COPYING file in the root of the source |
| 5 | // tree. An additional intellectual property rights grant can be found |
| 6 | // in the file PATENTS. All contributing project authors may |
| 7 | // be found in the AUTHORS file in the root of the source tree. |
| 8 | // ----------------------------------------------------------------------------- |
| 9 | // |
| 10 | // Author: Djordje Pesut (djordje.pesut@imgtec.com) |
| 11 | |
| 12 | #include "src/dsp/dsp.h" |
| 13 | |
| 14 | #if defined(WEBP_USE_MIPS32) |
| 15 | |
| 16 | #include "src/enc/cost_enc.h" |
| 17 | |
| 18 | static int GetResidualCost_MIPS32(int ctx0, const VP8Residual* const res) { |
| 19 | int temp0, temp1; |
| 20 | int v_reg, ctx_reg; |
| 21 | int n = res->first; |
| 22 | // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 |
| 23 | int p0 = res->prob[n][ctx0][0]; |
| 24 | CostArrayPtr const costs = res->costs; |
| 25 | const uint16_t* t = costs[n][ctx0]; |
| 26 | // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 |
| 27 | // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll |
| 28 | // be missing during the loop. |
| 29 | int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; |
| 30 | const int16_t* res_coeffs = res->coeffs; |
| 31 | const int res_last = res->last; |
| 32 | const int const_max_level = MAX_VARIABLE_LEVEL; |
| 33 | const int const_2 = 2; |
| 34 | const uint16_t** p_costs = &costs[n][0]; |
| 35 | const size_t inc_p_costs = NUM_CTX * sizeof(*p_costs); |
| 36 | |
| 37 | if (res->last < 0) { |
| 38 | return VP8BitCost(0, p0); |
| 39 | } |
| 40 | |
| 41 | __asm__ volatile ( |
| 42 | ".set push \n\t" |
| 43 | ".set noreorder \n\t" |
| 44 | "subu %[temp1], %[res_last], %[n] \n\t" |
| 45 | "sll %[temp0], %[n], 1 \n\t" |
| 46 | "blez %[temp1], 2f \n\t" |
| 47 | " addu %[res_coeffs], %[res_coeffs], %[temp0] \n\t" |
| 48 | "1: \n\t" |
| 49 | "lh %[v_reg], 0(%[res_coeffs]) \n\t" |
| 50 | "addiu %[n], %[n], 1 \n\t" |
| 51 | "negu %[temp0], %[v_reg] \n\t" |
| 52 | "slti %[temp1], %[v_reg], 0 \n\t" |
| 53 | "movn %[v_reg], %[temp0], %[temp1] \n\t" |
| 54 | "sltiu %[temp0], %[v_reg], 2 \n\t" |
| 55 | "move %[ctx_reg], %[v_reg] \n\t" |
| 56 | "movz %[ctx_reg], %[const_2], %[temp0] \n\t" |
| 57 | "sll %[temp1], %[v_reg], 1 \n\t" |
| 58 | "addu %[temp1], %[temp1], %[VP8LevelFixedCosts] \n\t" |
| 59 | "lhu %[temp1], 0(%[temp1]) \n\t" |
| 60 | "slt %[temp0], %[v_reg], %[const_max_level] \n\t" |
| 61 | "movz %[v_reg], %[const_max_level], %[temp0] \n\t" |
| 62 | "addu %[cost], %[cost], %[temp1] \n\t" |
| 63 | "sll %[v_reg], %[v_reg], 1 \n\t" |
| 64 | "sll %[ctx_reg], %[ctx_reg], 2 \n\t" |
| 65 | "addu %[v_reg], %[v_reg], %[t] \n\t" |
| 66 | "lhu %[temp0], 0(%[v_reg]) \n\t" |
| 67 | "addu %[p_costs], %[p_costs], %[inc_p_costs] \n\t" |
| 68 | "addu %[t], %[p_costs], %[ctx_reg] \n\t" |
| 69 | "addu %[cost], %[cost], %[temp0] \n\t" |
| 70 | "addiu %[res_coeffs], %[res_coeffs], 2 \n\t" |
| 71 | "bne %[n], %[res_last], 1b \n\t" |
| 72 | " lw %[t], 0(%[t]) \n\t" |
| 73 | "2: \n\t" |
| 74 | ".set pop \n\t" |
| 75 | : [cost]"+&r" (cost), [t]"+&r" (t), [n]"+&r" (n), [v_reg]"=&r" (v_reg), |
| 76 | [ctx_reg]"=&r" (ctx_reg), [p_costs]"+&r" (p_costs), [temp0]"=&r" (temp0), |
| 77 | [temp1]"=&r" (temp1), [res_coeffs]"+&r" (res_coeffs) |
| 78 | : [const_2]"r" (const_2), [const_max_level]"r" (const_max_level), |
| 79 | [VP8LevelFixedCosts]"r" (VP8LevelFixedCosts), [res_last]"r" (res_last), |
| 80 | [inc_p_costs]"r" (inc_p_costs) |
| 81 | : "memory" |
| 82 | ); |
| 83 | |
| 84 | // Last coefficient is always non-zero |
| 85 | { |
| 86 | const int v = abs(res->coeffs[n]); |
| 87 | assert(v != 0); |
| 88 | cost += VP8LevelCost(t, v); |
| 89 | if (n < 15) { |
| 90 | const int b = VP8EncBands[n + 1]; |
| 91 | const int ctx = (v == 1) ? 1 : 2; |
| 92 | const int last_p0 = res->prob[b][ctx][0]; |
| 93 | cost += VP8BitCost(0, last_p0); |
| 94 | } |
| 95 | } |
| 96 | return cost; |
| 97 | } |
| 98 | |
| 99 | static void SetResidualCoeffs_MIPS32(const int16_t* const coeffs, |
| 100 | VP8Residual* const res) { |
| 101 | const int16_t* p_coeffs = (int16_t*)coeffs; |
| 102 | int temp0, temp1, temp2, n, n1; |
| 103 | assert(res->first == 0 || coeffs[0] == 0); |
| 104 | |
| 105 | __asm__ volatile ( |
| 106 | ".set push \n\t" |
| 107 | ".set noreorder \n\t" |
| 108 | "addiu %[p_coeffs], %[p_coeffs], 28 \n\t" |
| 109 | "li %[n], 15 \n\t" |
| 110 | "li %[temp2], -1 \n\t" |
| 111 | "0: \n\t" |
| 112 | "ulw %[temp0], 0(%[p_coeffs]) \n\t" |
| 113 | "beqz %[temp0], 1f \n\t" |
| 114 | #if defined(WORDS_BIGENDIAN) |
| 115 | " sll %[temp1], %[temp0], 16 \n\t" |
| 116 | #else |
| 117 | " srl %[temp1], %[temp0], 16 \n\t" |
| 118 | #endif |
| 119 | "addiu %[n1], %[n], -1 \n\t" |
| 120 | "movz %[temp0], %[n1], %[temp1] \n\t" |
| 121 | "movn %[temp0], %[n], %[temp1] \n\t" |
| 122 | "j 2f \n\t" |
| 123 | " addiu %[temp2], %[temp0], 0 \n\t" |
| 124 | "1: \n\t" |
| 125 | "addiu %[n], %[n], -2 \n\t" |
| 126 | "bgtz %[n], 0b \n\t" |
| 127 | " addiu %[p_coeffs], %[p_coeffs], -4 \n\t" |
| 128 | "2: \n\t" |
| 129 | ".set pop \n\t" |
| 130 | : [p_coeffs]"+&r" (p_coeffs), [temp0]"=&r" (temp0), |
| 131 | [temp1]"=&r" (temp1), [temp2]"=&r" (temp2), |
| 132 | [n]"=&r" (n), [n1]"=&r" (n1) |
| 133 | : |
| 134 | : "memory" |
| 135 | ); |
| 136 | res->last = temp2; |
| 137 | res->coeffs = coeffs; |
| 138 | } |
| 139 | |
| 140 | //------------------------------------------------------------------------------ |
| 141 | // Entry point |
| 142 | |
| 143 | extern void VP8EncDspCostInitMIPS32(void); |
| 144 | |
| 145 | WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInitMIPS32(void) { |
| 146 | VP8GetResidualCost = GetResidualCost_MIPS32; |
| 147 | VP8SetResidualCoeffs = SetResidualCoeffs_MIPS32; |
| 148 | } |
| 149 | |
| 150 | #else // !WEBP_USE_MIPS32 |
| 151 | |
| 152 | WEBP_DSP_INIT_STUB(VP8EncDspCostInitMIPS32) |
| 153 | |
| 154 | #endif // WEBP_USE_MIPS32 |
| 155 | |