# HG changeset patch # User kostya # Date 1200725549 0 # Node ID 953c8efd52986e018a5b8929615f64fcf4056847 # Parent b5702c981fe20eca0c8a4e3c872c38ec3cb68ede Factor out constant addition diff -r b5702c981fe2 -r 953c8efd5298 vc1dsp.c --- a/vc1dsp.c Fri Jan 18 07:04:58 2008 +0000 +++ b/vc1dsp.c Sat Jan 19 06:52:29 2008 +0000 @@ -90,8 +90,8 @@ src = block; dst = block; for(i = 0; i < 8; i++){ - t1 = 12 * (src[0] + src[4]); - t2 = 12 * (src[0] - src[4]); + t1 = 12 * (src[0] + src[4]) + 4; + t2 = 12 * (src[0] - src[4]) + 4; t3 = 16 * src[2] + 6 * src[6]; t4 = 6 * src[2] - 16 * src[6]; @@ -105,14 +105,14 @@ t3 = 9 * src[1] - 16 * src[3] + 4 * src[5] + 15 * src[7]; t4 = 4 * src[1] - 9 * src[3] + 15 * src[5] - 16 * src[7]; - dst[0] = (t5 + t1 + 4) >> 3; - dst[1] = (t6 + t2 + 4) >> 3; - dst[2] = (t7 + t3 + 4) >> 3; - dst[3] = (t8 + t4 + 4) >> 3; - dst[4] = (t8 - t4 + 4) >> 3; - dst[5] = (t7 - t3 + 4) >> 3; - dst[6] = (t6 - t2 + 4) >> 3; - dst[7] = (t5 - t1 + 4) >> 3; + dst[0] = (t5 + t1) >> 3; + dst[1] = (t6 + t2) >> 3; + dst[2] = (t7 + t3) >> 3; + dst[3] = (t8 + t4) >> 3; + dst[4] = (t8 - t4) >> 3; + dst[5] = (t7 - t3) >> 3; + dst[6] = (t6 - t2) >> 3; + dst[7] = (t5 - t1) >> 3; src += 8; dst += 8; @@ -121,8 +121,8 @@ src = block; dst = block; for(i = 0; i < 8; i++){ - t1 = 12 * (src[ 0] + src[32]); - t2 = 12 * (src[ 0] - src[32]); + t1 = 12 * (src[ 0] + src[32]) + 64; + t2 = 12 * (src[ 0] - src[32]) + 64; t3 = 16 * src[16] + 6 * src[48]; t4 = 6 * src[16] - 16 * src[48]; @@ -136,14 +136,14 @@ t3 = 9 * src[ 8] - 16 * src[24] + 4 * src[40] + 15 * src[56]; t4 = 4 * src[ 8] - 9 * src[24] + 15 * src[40] - 16 * src[56]; - dst[ 0] = (t5 + t1 + 64) >> 7; - dst[ 8] = (t6 + t2 + 64) >> 7; - dst[16] = (t7 + t3 + 64) >> 7; - dst[24] = (t8 + t4 + 64) >> 7; - dst[32] = (t8 - t4 + 64 + 1) >> 7; - dst[40] = (t7 - t3 + 64 + 1) >> 7; - dst[48] = (t6 - t2 + 64 + 1) >> 7; - dst[56] = (t5 - t1 + 64 + 1) >> 7; + dst[ 0] = (t5 + t1) >> 7; + dst[ 8] = (t6 + t2) >> 7; + dst[16] = (t7 + t3) >> 7; + dst[24] = (t8 + t4) >> 7; + dst[32] = (t8 - t4 + 1) >> 7; + dst[40] = (t7 - t3 + 1) >> 7; + dst[48] = (t6 - t2 + 1) >> 7; + dst[56] = (t5 - t1 + 1) >> 7; src++; dst++; @@ -162,8 +162,8 @@ src = block; dst = block; for(i = 0; i < 4; i++){ - t1 = 12 * (src[0] + src[4]); - t2 = 12 * (src[0] - src[4]); + t1 = 12 * (src[0] + src[4]) + 4; + t2 = 12 * (src[0] - src[4]) + 4; t3 = 16 * src[2] + 6 * src[6]; t4 = 6 * src[2] - 16 * src[6]; @@ -177,14 +177,14 @@ t3 = 9 * src[1] - 16 * src[3] + 4 * src[5] + 15 * src[7]; t4 = 4 * src[1] - 9 * src[3] + 15 * src[5] - 16 * src[7]; - dst[0] = (t5 + t1 + 4) >> 3; - dst[1] = (t6 + t2 + 4) >> 3; - dst[2] = (t7 + t3 + 4) >> 3; - dst[3] = (t8 + t4 + 4) >> 3; - dst[4] = (t8 - t4 + 4) >> 3; - dst[5] = (t7 - t3 + 4) >> 3; - dst[6] = (t6 - t2 + 4) >> 3; - dst[7] = (t5 - t1 + 4) >> 3; + dst[0] = (t5 + t1) >> 3; + dst[1] = (t6 + t2) >> 3; + dst[2] = (t7 + t3) >> 3; + dst[3] = (t8 + t4) >> 3; + dst[4] = (t8 - t4) >> 3; + dst[5] = (t7 - t3) >> 3; + dst[6] = (t6 - t2) >> 3; + dst[7] = (t5 - t1) >> 3; src += 8; dst += 8; @@ -192,17 +192,17 @@ src = block; for(i = 0; i < 8; i++){ - t1 = 17 * (src[ 0] + src[16]); - t2 = 17 * (src[ 0] - src[16]); + t1 = 17 * (src[ 0] + src[16]) + 64; + t2 = 17 * (src[ 0] - src[16]) + 64; t3 = 22 * src[ 8]; t4 = 22 * src[24]; t5 = 10 * src[ 8]; t6 = 10 * src[24]; - dest[0*linesize] = cm[dest[0*linesize] + ((t1 + t3 + t6 + 64) >> 7)]; - dest[1*linesize] = cm[dest[1*linesize] + ((t2 - t4 + t5 + 64) >> 7)]; - dest[2*linesize] = cm[dest[2*linesize] + ((t2 + t4 - t5 + 64) >> 7)]; - dest[3*linesize] = cm[dest[3*linesize] + ((t1 - t3 - t6 + 64) >> 7)]; + dest[0*linesize] = cm[dest[0*linesize] + ((t1 + t3 + t6) >> 7)]; + dest[1*linesize] = cm[dest[1*linesize] + ((t2 - t4 + t5) >> 7)]; + dest[2*linesize] = cm[dest[2*linesize] + ((t2 + t4 - t5) >> 7)]; + dest[3*linesize] = cm[dest[3*linesize] + ((t1 - t3 - t6) >> 7)]; src ++; dest++; @@ -221,17 +221,17 @@ src = block; dst = block; for(i = 0; i < 8; i++){ - t1 = 17 * (src[0] + src[2]); - t2 = 17 * (src[0] - src[2]); + t1 = 17 * (src[0] + src[2]) + 4; + t2 = 17 * (src[0] - src[2]) + 4; t3 = 22 * src[1]; t4 = 22 * src[3]; t5 = 10 * src[1]; t6 = 10 * src[3]; - dst[0] = (t1 + t3 + t6 + 4) >> 3; - dst[1] = (t2 - t4 + t5 + 4) >> 3; - dst[2] = (t2 + t4 - t5 + 4) >> 3; - dst[3] = (t1 - t3 - t6 + 4) >> 3; + dst[0] = (t1 + t3 + t6) >> 3; + dst[1] = (t2 - t4 + t5) >> 3; + dst[2] = (t2 + t4 - t5) >> 3; + dst[3] = (t1 - t3 - t6) >> 3; src += 8; dst += 8; @@ -239,8 +239,8 @@ src = block; for(i = 0; i < 4; i++){ - t1 = 12 * (src[ 0] + src[32]); - t2 = 12 * (src[ 0] - src[32]); + t1 = 12 * (src[ 0] + src[32]) + 64; + t2 = 12 * (src[ 0] - src[32]) + 64; t3 = 16 * src[16] + 6 * src[48]; t4 = 6 * src[16] - 16 * src[48]; @@ -254,14 +254,14 @@ t3 = 9 * src[ 8] - 16 * src[24] + 4 * src[40] + 15 * src[56]; t4 = 4 * src[ 8] - 9 * src[24] + 15 * src[40] - 16 * src[56]; - dest[0*linesize] = cm[dest[0*linesize] + ((t5 + t1 + 64) >> 7)]; - dest[1*linesize] = cm[dest[1*linesize] + ((t6 + t2 + 64) >> 7)]; - dest[2*linesize] = cm[dest[2*linesize] + ((t7 + t3 + 64) >> 7)]; - dest[3*linesize] = cm[dest[3*linesize] + ((t8 + t4 + 64) >> 7)]; - dest[4*linesize] = cm[dest[4*linesize] + ((t8 - t4 + 64 + 1) >> 7)]; - dest[5*linesize] = cm[dest[5*linesize] + ((t7 - t3 + 64 + 1) >> 7)]; - dest[6*linesize] = cm[dest[6*linesize] + ((t6 - t2 + 64 + 1) >> 7)]; - dest[7*linesize] = cm[dest[7*linesize] + ((t5 - t1 + 64 + 1) >> 7)]; + dest[0*linesize] = cm[dest[0*linesize] + ((t5 + t1) >> 7)]; + dest[1*linesize] = cm[dest[1*linesize] + ((t6 + t2) >> 7)]; + dest[2*linesize] = cm[dest[2*linesize] + ((t7 + t3) >> 7)]; + dest[3*linesize] = cm[dest[3*linesize] + ((t8 + t4) >> 7)]; + dest[4*linesize] = cm[dest[4*linesize] + ((t8 - t4 + 1) >> 7)]; + dest[5*linesize] = cm[dest[5*linesize] + ((t7 - t3 + 1) >> 7)]; + dest[6*linesize] = cm[dest[6*linesize] + ((t6 - t2 + 1) >> 7)]; + dest[7*linesize] = cm[dest[7*linesize] + ((t5 - t1 + 1) >> 7)]; src ++; dest++; @@ -280,17 +280,17 @@ src = block; dst = block; for(i = 0; i < 4; i++){ - t1 = 17 * (src[0] + src[2]); - t2 = 17 * (src[0] - src[2]); + t1 = 17 * (src[0] + src[2]) + 4; + t2 = 17 * (src[0] - src[2]) + 4; t3 = 22 * src[1]; t4 = 22 * src[3]; t5 = 10 * src[1]; t6 = 10 * src[3]; - dst[0] = (t1 + t3 + t6 + 4) >> 3; - dst[1] = (t2 - t4 + t5 + 4) >> 3; - dst[2] = (t2 + t4 - t5 + 4) >> 3; - dst[3] = (t1 - t3 - t6 + 4) >> 3; + dst[0] = (t1 + t3 + t6) >> 3; + dst[1] = (t2 - t4 + t5) >> 3; + dst[2] = (t2 + t4 - t5) >> 3; + dst[3] = (t1 - t3 - t6) >> 3; src += 8; dst += 8; @@ -298,17 +298,17 @@ src = block; for(i = 0; i < 4; i++){ - t1 = 17 * (src[ 0] + src[16]); - t2 = 17 * (src[ 0] - src[16]); + t1 = 17 * (src[ 0] + src[16]) + 64; + t2 = 17 * (src[ 0] - src[16]) + 64; t3 = 22 * src[ 8]; t4 = 22 * src[24]; t5 = 10 * src[ 8]; t6 = 10 * src[24]; - dest[0*linesize] = cm[dest[0*linesize] + ((t1 + t3 + t6 + 64) >> 7)]; - dest[1*linesize] = cm[dest[1*linesize] + ((t2 - t4 + t5 + 64) >> 7)]; - dest[2*linesize] = cm[dest[2*linesize] + ((t2 + t4 - t5 + 64) >> 7)]; - dest[3*linesize] = cm[dest[3*linesize] + ((t1 - t3 - t6 + 64) >> 7)]; + dest[0*linesize] = cm[dest[0*linesize] + ((t1 + t3 + t6) >> 7)]; + dest[1*linesize] = cm[dest[1*linesize] + ((t2 - t4 + t5) >> 7)]; + dest[2*linesize] = cm[dest[2*linesize] + ((t2 + t4 - t5) >> 7)]; + dest[3*linesize] = cm[dest[3*linesize] + ((t1 - t3 - t6) >> 7)]; src ++; dest++;