Mercurial > libavcodec.hg
comparison h264idct.c @ 8457:fa07932f2c89 libavcodec
flatten an array, since gcc fails at optimizing multidimensional arrays
h264_idct8_add_c: 780 -> 735 cycles on conroe
author | lorenm |
---|---|
date | Thu, 25 Dec 2008 01:20:37 +0000 |
parents | de2509cf3c44 |
children | 5b7d5a33a3e2 |
comparison
equal
deleted
inserted
replaced
8456:30f2b80242b2 | 8457:fa07932f2c89 |
---|---|
70 idct_internal(dst, block, stride, 8, 3, 0); | 70 idct_internal(dst, block, stride, 8, 3, 0); |
71 } | 71 } |
72 | 72 |
73 void ff_h264_idct8_add_c(uint8_t *dst, DCTELEM *block, int stride){ | 73 void ff_h264_idct8_add_c(uint8_t *dst, DCTELEM *block, int stride){ |
74 int i; | 74 int i; |
75 DCTELEM (*src)[8] = (DCTELEM(*)[8])block; | |
76 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP; | 75 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP; |
77 | 76 |
78 block[0] += 32; | 77 block[0] += 32; |
79 | 78 |
80 for( i = 0; i < 8; i++ ) | 79 for( i = 0; i < 8; i++ ) |
81 { | 80 { |
82 const int a0 = src[i][0] + src[i][4]; | 81 const int a0 = block[0+i*8] + block[4+i*8]; |
83 const int a2 = src[i][0] - src[i][4]; | 82 const int a2 = block[0+i*8] - block[4+i*8]; |
84 const int a4 = (src[i][2]>>1) - src[i][6]; | 83 const int a4 = (block[2+i*8]>>1) - block[6+i*8]; |
85 const int a6 = (src[i][6]>>1) + src[i][2]; | 84 const int a6 = (block[6+i*8]>>1) + block[2+i*8]; |
86 | 85 |
87 const int b0 = a0 + a6; | 86 const int b0 = a0 + a6; |
88 const int b2 = a2 + a4; | 87 const int b2 = a2 + a4; |
89 const int b4 = a2 - a4; | 88 const int b4 = a2 - a4; |
90 const int b6 = a0 - a6; | 89 const int b6 = a0 - a6; |
91 | 90 |
92 const int a1 = -src[i][3] + src[i][5] - src[i][7] - (src[i][7]>>1); | 91 const int a1 = -block[3+i*8] + block[5+i*8] - block[7+i*8] - (block[7+i*8]>>1); |
93 const int a3 = src[i][1] + src[i][7] - src[i][3] - (src[i][3]>>1); | 92 const int a3 = block[1+i*8] + block[7+i*8] - block[3+i*8] - (block[3+i*8]>>1); |
94 const int a5 = -src[i][1] + src[i][7] + src[i][5] + (src[i][5]>>1); | 93 const int a5 = -block[1+i*8] + block[7+i*8] + block[5+i*8] + (block[5+i*8]>>1); |
95 const int a7 = src[i][3] + src[i][5] + src[i][1] + (src[i][1]>>1); | 94 const int a7 = block[3+i*8] + block[5+i*8] + block[1+i*8] + (block[1+i*8]>>1); |
96 | 95 |
97 const int b1 = (a7>>2) + a1; | 96 const int b1 = (a7>>2) + a1; |
98 const int b3 = a3 + (a5>>2); | 97 const int b3 = a3 + (a5>>2); |
99 const int b5 = (a3>>2) - a5; | 98 const int b5 = (a3>>2) - a5; |
100 const int b7 = a7 - (a1>>2); | 99 const int b7 = a7 - (a1>>2); |
101 | 100 |
102 src[i][0] = b0 + b7; | 101 block[0+i*8] = b0 + b7; |
103 src[i][7] = b0 - b7; | 102 block[7+i*8] = b0 - b7; |
104 src[i][1] = b2 + b5; | 103 block[1+i*8] = b2 + b5; |
105 src[i][6] = b2 - b5; | 104 block[6+i*8] = b2 - b5; |
106 src[i][2] = b4 + b3; | 105 block[2+i*8] = b4 + b3; |
107 src[i][5] = b4 - b3; | 106 block[5+i*8] = b4 - b3; |
108 src[i][3] = b6 + b1; | 107 block[3+i*8] = b6 + b1; |
109 src[i][4] = b6 - b1; | 108 block[4+i*8] = b6 - b1; |
110 } | 109 } |
111 for( i = 0; i < 8; i++ ) | 110 for( i = 0; i < 8; i++ ) |
112 { | 111 { |
113 const int a0 = src[0][i] + src[4][i]; | 112 const int a0 = block[i+0*8] + block[i+4*8]; |
114 const int a2 = src[0][i] - src[4][i]; | 113 const int a2 = block[i+0*8] - block[i+4*8]; |
115 const int a4 = (src[2][i]>>1) - src[6][i]; | 114 const int a4 = (block[i+2*8]>>1) - block[i+6*8]; |
116 const int a6 = (src[6][i]>>1) + src[2][i]; | 115 const int a6 = (block[i+6*8]>>1) + block[i+2*8]; |
117 | 116 |
118 const int b0 = a0 + a6; | 117 const int b0 = a0 + a6; |
119 const int b2 = a2 + a4; | 118 const int b2 = a2 + a4; |
120 const int b4 = a2 - a4; | 119 const int b4 = a2 - a4; |
121 const int b6 = a0 - a6; | 120 const int b6 = a0 - a6; |
122 | 121 |
123 const int a1 = -src[3][i] + src[5][i] - src[7][i] - (src[7][i]>>1); | 122 const int a1 = -block[i+3*8] + block[i+5*8] - block[i+7*8] - (block[i+7*8]>>1); |
124 const int a3 = src[1][i] + src[7][i] - src[3][i] - (src[3][i]>>1); | 123 const int a3 = block[i+1*8] + block[i+7*8] - block[i+3*8] - (block[i+3*8]>>1); |
125 const int a5 = -src[1][i] + src[7][i] + src[5][i] + (src[5][i]>>1); | 124 const int a5 = -block[i+1*8] + block[i+7*8] + block[i+5*8] + (block[i+5*8]>>1); |
126 const int a7 = src[3][i] + src[5][i] + src[1][i] + (src[1][i]>>1); | 125 const int a7 = block[i+3*8] + block[i+5*8] + block[i+1*8] + (block[i+1*8]>>1); |
127 | 126 |
128 const int b1 = (a7>>2) + a1; | 127 const int b1 = (a7>>2) + a1; |
129 const int b3 = a3 + (a5>>2); | 128 const int b3 = a3 + (a5>>2); |
130 const int b5 = (a3>>2) - a5; | 129 const int b5 = (a3>>2) - a5; |
131 const int b7 = a7 - (a1>>2); | 130 const int b7 = a7 - (a1>>2); |