00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021
00027 #include "avcodec.h"
00028 #include "dsputil.h"
00029
00030 #define RV30_LOWPASS(OPNAME, OP) \
00031 static av_unused void OPNAME ## rv30_tpel8_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, const int C1, const int C2){\
00032 const int h=8;\
00033 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
00034 int i;\
00035 for(i=0; i<h; i++)\
00036 {\
00037 OP(dst[0], -(src[-1]+src[2]) + src[0]*C1 + src[1]*C2);\
00038 OP(dst[1], -(src[ 0]+src[3]) + src[1]*C1 + src[2]*C2);\
00039 OP(dst[2], -(src[ 1]+src[4]) + src[2]*C1 + src[3]*C2);\
00040 OP(dst[3], -(src[ 2]+src[5]) + src[3]*C1 + src[4]*C2);\
00041 OP(dst[4], -(src[ 3]+src[6]) + src[4]*C1 + src[5]*C2);\
00042 OP(dst[5], -(src[ 4]+src[7]) + src[5]*C1 + src[6]*C2);\
00043 OP(dst[6], -(src[ 5]+src[8]) + src[6]*C1 + src[7]*C2);\
00044 OP(dst[7], -(src[ 6]+src[9]) + src[7]*C1 + src[8]*C2);\
00045 dst+=dstStride;\
00046 src+=srcStride;\
00047 }\
00048 }\
00049 \
00050 static void OPNAME ## rv30_tpel8_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, const int C1, const int C2){\
00051 const int w=8;\
00052 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
00053 int i;\
00054 for(i=0; i<w; i++)\
00055 {\
00056 const int srcA= src[-1*srcStride];\
00057 const int src0= src[0 *srcStride];\
00058 const int src1= src[1 *srcStride];\
00059 const int src2= src[2 *srcStride];\
00060 const int src3= src[3 *srcStride];\
00061 const int src4= src[4 *srcStride];\
00062 const int src5= src[5 *srcStride];\
00063 const int src6= src[6 *srcStride];\
00064 const int src7= src[7 *srcStride];\
00065 const int src8= src[8 *srcStride];\
00066 const int src9= src[9 *srcStride];\
00067 OP(dst[0*dstStride], -(srcA+src2) + src0*C1 + src1*C2);\
00068 OP(dst[1*dstStride], -(src0+src3) + src1*C1 + src2*C2);\
00069 OP(dst[2*dstStride], -(src1+src4) + src2*C1 + src3*C2);\
00070 OP(dst[3*dstStride], -(src2+src5) + src3*C1 + src4*C2);\
00071 OP(dst[4*dstStride], -(src3+src6) + src4*C1 + src5*C2);\
00072 OP(dst[5*dstStride], -(src4+src7) + src5*C1 + src6*C2);\
00073 OP(dst[6*dstStride], -(src5+src8) + src6*C1 + src7*C2);\
00074 OP(dst[7*dstStride], -(src6+src9) + src7*C1 + src8*C2);\
00075 dst++;\
00076 src++;\
00077 }\
00078 }\
00079 \
00080 static void OPNAME ## rv30_tpel8_h3_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
00081 const int h=8+2;\
00082 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
00083 int i;\
00084 for(i=0; i<h; i++)\
00085 {\
00086 OP(dst[0], 6*src[0]+9*src[1]+src[2]);\
00087 OP(dst[1], 6*src[1]+9*src[2]+src[3]);\
00088 OP(dst[2], 6*src[2]+9*src[3]+src[4]);\
00089 OP(dst[3], 6*src[3]+9*src[4]+src[5]);\
00090 OP(dst[4], 6*src[4]+9*src[5]+src[6]);\
00091 OP(dst[5], 6*src[5]+9*src[6]+src[7]);\
00092 OP(dst[6], 6*src[6]+9*src[7]+src[8]);\
00093 OP(dst[7], 6*src[7]+9*src[8]+src[9]);\
00094 dst+=dstStride;\
00095 src+=srcStride;\
00096 }\
00097 }\
00098 \
00099 static void OPNAME ## rv30_tpel8_v3_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
00100 const int w=8;\
00101 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
00102 int i;\
00103 for(i=0; i<w; i++)\
00104 {\
00105 const int src0= src[0 *srcStride];\
00106 const int src1= src[1 *srcStride];\
00107 const int src2= src[2 *srcStride];\
00108 const int src3= src[3 *srcStride];\
00109 const int src4= src[4 *srcStride];\
00110 const int src5= src[5 *srcStride];\
00111 const int src6= src[6 *srcStride];\
00112 const int src7= src[7 *srcStride];\
00113 const int src8= src[8 *srcStride];\
00114 const int src9= src[9 *srcStride];\
00115 OP(dst[0*dstStride], 6*src0 + 9*src1 + src2);\
00116 OP(dst[1*dstStride], 6*src1 + 9*src2 + src3);\
00117 OP(dst[2*dstStride], 6*src2 + 9*src3 + src4);\
00118 OP(dst[3*dstStride], 6*src3 + 9*src4 + src5);\
00119 OP(dst[4*dstStride], 6*src4 + 9*src5 + src6);\
00120 OP(dst[5*dstStride], 6*src5 + 9*src6 + src7);\
00121 OP(dst[6*dstStride], 6*src6 + 9*src7 + src8);\
00122 OP(dst[7*dstStride], 6*src7 + 9*src8 + src9);\
00123 dst ++;\
00124 src ++;\
00125 }\
00126 }\
00127 \
00128 static void OPNAME ## rv30_tpel8_hv_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
00129 uint8_t half[8*10];\
00130 put_rv30_tpel8_h3_lowpass(half, src, 8, srcStride);\
00131 OPNAME ## rv30_tpel8_v3_lowpass(dst, half, dstStride, 8);\
00132 }\
00133 \
00134 static void OPNAME ## rv30_tpel16_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, const int C1, const int C2){\
00135 OPNAME ## rv30_tpel8_v_lowpass(dst , src , dstStride, srcStride, C1, C2);\
00136 OPNAME ## rv30_tpel8_v_lowpass(dst+8, src+8, dstStride, srcStride, C1, C2);\
00137 src += 8*srcStride;\
00138 dst += 8*dstStride;\
00139 OPNAME ## rv30_tpel8_v_lowpass(dst , src , dstStride, srcStride, C1, C2);\
00140 OPNAME ## rv30_tpel8_v_lowpass(dst+8, src+8, dstStride, srcStride, C1, C2);\
00141 }\
00142 \
00143 static void OPNAME ## rv30_tpel16_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, const int C1, const int C2){\
00144 OPNAME ## rv30_tpel8_h_lowpass(dst , src , dstStride, srcStride, C1, C2);\
00145 OPNAME ## rv30_tpel8_h_lowpass(dst+8, src+8, dstStride, srcStride, C1, C2);\
00146 src += 8*srcStride;\
00147 dst += 8*dstStride;\
00148 OPNAME ## rv30_tpel8_h_lowpass(dst , src , dstStride, srcStride, C1, C2);\
00149 OPNAME ## rv30_tpel8_h_lowpass(dst+8, src+8, dstStride, srcStride, C1, C2);\
00150 }\
00151 \
00152 static void OPNAME ## rv30_tpel16_hv_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
00153 OPNAME ## rv30_tpel8_hv_lowpass(dst , src , dstStride, srcStride);\
00154 OPNAME ## rv30_tpel8_hv_lowpass(dst+8, src+8, dstStride, srcStride);\
00155 src += 8*srcStride;\
00156 dst += 8*dstStride;\
00157 OPNAME ## rv30_tpel8_hv_lowpass(dst , src , dstStride, srcStride);\
00158 OPNAME ## rv30_tpel8_hv_lowpass(dst+8, src+8, dstStride, srcStride);\
00159 }\
00160 \
00161
00162 #define RV30_MC(OPNAME, SIZE) \
00163 static void OPNAME ## rv30_tpel ## SIZE ## _mc10_c(uint8_t *dst, uint8_t *src, int stride){\
00164 OPNAME ## rv30_tpel ## SIZE ## _h_lowpass(dst, src, stride, stride, 12, 6);\
00165 }\
00166 \
00167 static void OPNAME ## rv30_tpel ## SIZE ## _mc20_c(uint8_t *dst, uint8_t *src, int stride){\
00168 OPNAME ## rv30_tpel ## SIZE ## _h_lowpass(dst, src, stride, stride, 6, 12);\
00169 }\
00170 \
00171 static void OPNAME ## rv30_tpel ## SIZE ## _mc01_c(uint8_t *dst, uint8_t *src, int stride){\
00172 OPNAME ## rv30_tpel ## SIZE ## _v_lowpass(dst, src, stride, stride, 12, 6);\
00173 }\
00174 \
00175 static void OPNAME ## rv30_tpel ## SIZE ## _mc02_c(uint8_t *dst, uint8_t *src, int stride){\
00176 OPNAME ## rv30_tpel ## SIZE ## _v_lowpass(dst, src, stride, stride, 6, 12);\
00177 }\
00178 \
00179 static void OPNAME ## rv30_tpel ## SIZE ## _mc11_c(uint8_t *dst, uint8_t *src, int stride){\
00180 uint8_t half[SIZE*SIZE];\
00181 put_rv30_tpel ## SIZE ## _h_lowpass(half, src, SIZE, stride, 12, 6);\
00182 OPNAME ## rv30_tpel ## SIZE ## _v_lowpass(dst, src, stride, stride, 12, 6);\
00183 }\
00184 \
00185 static void OPNAME ## rv30_tpel ## SIZE ## _mc12_c(uint8_t *dst, uint8_t *src, int stride){\
00186 uint8_t half[SIZE*SIZE];\
00187 put_rv30_tpel ## SIZE ## _h_lowpass(half, src, SIZE, stride, 12, 6);\
00188 OPNAME ## rv30_tpel ## SIZE ## _v_lowpass(dst, src, stride, stride, 6, 12);\
00189 }\
00190 \
00191 static void OPNAME ## rv30_tpel ## SIZE ## _mc21_c(uint8_t *dst, uint8_t *src, int stride){\
00192 uint8_t half[SIZE*SIZE];\
00193 put_rv30_tpel ## SIZE ## _h_lowpass(half, src, SIZE, stride, 6, 12);\
00194 OPNAME ## rv30_tpel ## SIZE ## _v_lowpass(dst, src, stride, stride, 12, 6);\
00195 }\
00196 \
00197 static void OPNAME ## rv30_tpel ## SIZE ## _mc22_c(uint8_t *dst, uint8_t *src, int stride){\
00198 OPNAME ## rv30_tpel ## SIZE ## _hv_lowpass(dst, src, stride, stride);\
00199 }\
00200 \
00201
00202 #define op_avg(a, b) a = (((a)+cm[((b) + 8)>>4]+1)>>1)
00203 #define op_put(a, b) a = cm[((b) + 8)>>4]
00204
00205 RV30_LOWPASS(put_ , op_put)
00206 RV30_LOWPASS(avg_ , op_avg)
00207 RV30_MC(put_, 8)
00208 RV30_MC(put_, 16)
00209 RV30_MC(avg_, 8)
00210 RV30_MC(avg_, 16)
00211
00212 void ff_rv30dsp_init(DSPContext* c, AVCodecContext *avctx) {
00213 c->put_rv30_tpel_pixels_tab[0][ 0] = c->put_h264_qpel_pixels_tab[0][0];
00214 c->put_rv30_tpel_pixels_tab[0][ 1] = put_rv30_tpel16_mc10_c;
00215 c->put_rv30_tpel_pixels_tab[0][ 2] = put_rv30_tpel16_mc20_c;
00216 c->put_rv30_tpel_pixels_tab[0][ 4] = put_rv30_tpel16_mc01_c;
00217 c->put_rv30_tpel_pixels_tab[0][ 5] = put_rv30_tpel16_mc11_c;
00218 c->put_rv30_tpel_pixels_tab[0][ 6] = put_rv30_tpel16_mc21_c;
00219 c->put_rv30_tpel_pixels_tab[0][ 8] = put_rv30_tpel16_mc02_c;
00220 c->put_rv30_tpel_pixels_tab[0][ 9] = put_rv30_tpel16_mc12_c;
00221 c->put_rv30_tpel_pixels_tab[0][10] = put_rv30_tpel16_mc22_c;
00222 c->avg_rv30_tpel_pixels_tab[0][ 0] = c->avg_h264_qpel_pixels_tab[0][0];
00223 c->avg_rv30_tpel_pixels_tab[0][ 1] = avg_rv30_tpel16_mc10_c;
00224 c->avg_rv30_tpel_pixels_tab[0][ 2] = avg_rv30_tpel16_mc20_c;
00225 c->avg_rv30_tpel_pixels_tab[0][ 4] = avg_rv30_tpel16_mc01_c;
00226 c->avg_rv30_tpel_pixels_tab[0][ 5] = avg_rv30_tpel16_mc11_c;
00227 c->avg_rv30_tpel_pixels_tab[0][ 6] = avg_rv30_tpel16_mc21_c;
00228 c->avg_rv30_tpel_pixels_tab[0][ 8] = avg_rv30_tpel16_mc02_c;
00229 c->avg_rv30_tpel_pixels_tab[0][ 9] = avg_rv30_tpel16_mc12_c;
00230 c->avg_rv30_tpel_pixels_tab[0][10] = avg_rv30_tpel16_mc22_c;
00231 c->put_rv30_tpel_pixels_tab[1][ 0] = c->put_h264_qpel_pixels_tab[1][0];
00232 c->put_rv30_tpel_pixels_tab[1][ 1] = put_rv30_tpel8_mc10_c;
00233 c->put_rv30_tpel_pixels_tab[1][ 2] = put_rv30_tpel8_mc20_c;
00234 c->put_rv30_tpel_pixels_tab[1][ 4] = put_rv30_tpel8_mc01_c;
00235 c->put_rv30_tpel_pixels_tab[1][ 5] = put_rv30_tpel8_mc11_c;
00236 c->put_rv30_tpel_pixels_tab[1][ 6] = put_rv30_tpel8_mc21_c;
00237 c->put_rv30_tpel_pixels_tab[1][ 8] = put_rv30_tpel8_mc02_c;
00238 c->put_rv30_tpel_pixels_tab[1][ 9] = put_rv30_tpel8_mc12_c;
00239 c->put_rv30_tpel_pixels_tab[1][10] = put_rv30_tpel8_mc22_c;
00240 c->avg_rv30_tpel_pixels_tab[1][ 0] = c->avg_h264_qpel_pixels_tab[1][0];
00241 c->avg_rv30_tpel_pixels_tab[1][ 1] = avg_rv30_tpel8_mc10_c;
00242 c->avg_rv30_tpel_pixels_tab[1][ 2] = avg_rv30_tpel8_mc20_c;
00243 c->avg_rv30_tpel_pixels_tab[1][ 4] = avg_rv30_tpel8_mc01_c;
00244 c->avg_rv30_tpel_pixels_tab[1][ 5] = avg_rv30_tpel8_mc11_c;
00245 c->avg_rv30_tpel_pixels_tab[1][ 6] = avg_rv30_tpel8_mc21_c;
00246 c->avg_rv30_tpel_pixels_tab[1][ 8] = avg_rv30_tpel8_mc02_c;
00247 c->avg_rv30_tpel_pixels_tab[1][ 9] = avg_rv30_tpel8_mc12_c;
00248 c->avg_rv30_tpel_pixels_tab[1][10] = avg_rv30_tpel8_mc22_c;
00249 }